Relay across any mix of links, one online answer for every client, first claim owns a hostname

A tunnel no longer depends on the kind of link under it. The relay hub keeps
one forwarding table per network instead of one per protocol, the envelope
names the end-to-end protocol, and a neighbour reached over several kinds of
transport is sent to over the best open one. A member in the middle needs no
protocol in common with either end. Which protocol a pair runs is settled by
the lower endpoint id's announced order, so both ends agree without asking.
A direct link that comes up later is shorter than the path through others, so
traffic moves to it with the same logical link.

The agent now answers "is it online" once: a live control session or any data
path. Members and peers report how they are reached (direct over a named
transport, relayed through a named member, or no data path yet), and `tsng
status` and the tray both read that instead of a session list and a WireGuard
handshake respectively.

A hostname belongs to whoever claimed it first. A claim carries the time its
author first claimed the name, signed with the record, and replicas order by
(time, id) so every member names the same holder. A later claimant keeps
working by address, is not found by the name, and is told so in its status,
`tsng status`, the tray and a HostnameTaken event.

Wire change: control and data ALPN 5, control PROTOCOL_VERSION 5, envelope
version 2, signing domain v3, state schema 7, local control protocol 18.
Upgrade every member together.

Co-Authored-By: Claude <noreply@anthropic.com>

.cargo/config.toml

@@ -0,0 +1,5 @@
+# Windows: link the C runtime statically, so the agent and the tray start on a
+# machine that has no Visual C++ redistributable installed. The installer ships
+# nothing but our own files and Wintun.
+[target.x86_64-pc-windows-msvc]
+rustflags = ["-C", "target-feature=+crt-static"]

.gitea/workflows/release.yml

@@ -0,0 +1,350 @@
+name: Release
+
+# Build the agent for every supported target on a version tag, leave each build
+# as a run artifact, and then — only after every target has built — attach the
+# archives to the release. Linux is a fully static musl binary (built with
+# cargo-zigbuild) so one artifact runs on any distro and can be repackaged
+# however a distro prefers.
+#
+# Targeted at Gitea Actions: artifacts use @v3 (the v4 backend is not available
+# on Gitea/GHES), and the release is filled through Gitea's own API with curl
+# rather than `gh`, which speaks only to github.com.
+on:
+  push:
+    tags: ["v*"]
+
+permissions:
+  contents: write
+
+env:
+  CARGO_TERM_COLOR: always
+  CARGO_NET_RETRY: "10"
+
+jobs:
+  build:
+    name: ${{ matrix.asset }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-musl
+            asset: tsunagi-linux-x86_64
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: amd64
+            pkg_arch: x86_64
+            gnu_target: x86_64-unknown-linux-gnu
+          - os: ubuntu-latest
+            target: aarch64-unknown-linux-musl
+            asset: tsunagi-linux-aarch64
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: arm64
+            pkg_arch: aarch64
+          # 32-bit ARM with hardware float (ARMv7, `v7l`): single-board and edge
+          # devices. Statically linked like the others.
+          - os: ubuntu-latest
+            target: armv7-unknown-linux-musleabihf
+            asset: tsunagi-linux-armv7
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: armhf
+            pkg_arch: armv7h
+          - os: macos-latest
+            target: aarch64-apple-darwin
+            asset: tsunagi-macos-aarch64
+            archive: tar.gz
+            zig: false
+          - os: windows-latest
+            target: x86_64-pc-windows-msvc
+            asset: tsunagi-windows-x86_64
+            archive: zip
+            zig: false
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          targets: ${{ matrix.target }}
+
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: ${{ matrix.target }}
+
+      # aws-lc-sys (iroh's rustls backend) and the bundled sqlite are C, and the
+      # musl targets have no prebuilt bindings, so bindgen needs libclang and the
+      # C build needs cmake. zig is the cross C compiler cargo-zigbuild drives.
+      - name: Install musl build and packaging tooling
+        if: matrix.zig
+        run: sudo apt-get update && sudo apt-get install -y --no-install-recommends cmake clang libclang-dev libarchive-tools zstd
+      - name: Install zig
+        if: matrix.zig
+        uses: mlugg/setup-zig@v2
+        with:
+          # Pin a real, archived Zig release. Left unpinned the action resolves
+          # to the newest version it knows of (a yet-unreleased 0.16.0-dev),
+          # whose tarball 404s on ziglang.org while the community mirrors flake.
+          version: 0.14.1
+      - name: Install cargo-zigbuild
+        if: matrix.zig
+        uses: taiki-e/install-action@v2
+        with:
+          tool: cargo-zigbuild
+
+      - name: Build
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [[ "${{ matrix.zig }}" == "true" ]]; then
+            # zig's linker ignores the `-O1` GNU-ld optimization flag Rust passes
+            # and says so on stderr, which rustc now surfaces as a linker_messages
+            # warning. It is cosmetic — the binary is correct — so quiet just that
+            # one lint to keep the musl build log clean. A real link error still
+            # fails the build.
+            export RUSTFLAGS="${RUSTFLAGS:-} -A linker_messages"
+            cargo zigbuild --release --locked --target "${{ matrix.target }}" --bin tsng
+          else
+            cargo build --release --locked --target "${{ matrix.target }}" --bin tsng
+          fi
+
+      # The tray GUI. On Linux it is built against glibc (an old baseline for
+      # portability), not musl: a GUI must dlopen libGL/X11/Wayland at runtime,
+      # which a static musl binary cannot. macOS and Windows build it natively.
+      - name: Build the tray GUI
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [[ "${{ runner.os }}" == "Linux" ]]; then
+            if [[ -n "${{ matrix.gnu_target }}" ]]; then
+              # tray-icon links GTK and libappindicator through pkg-config, which
+              # has no cross story under zig, so the Linux GUI is built natively
+              # (x86_64 only) against the runner's glibc.
+              sudo apt-get update
+              sudo apt-get install -y --no-install-recommends \
+                libgtk-3-dev libayatana-appindicator3-dev libxdo-dev
+              cargo build --release --locked \
+                --target "${{ matrix.gnu_target }}" --bin tsunagi-tray
+            else
+              echo "no GUI build for ${{ matrix.target }}"
+            fi
+          else
+            cargo build --release --locked --target "${{ matrix.target }}" --bin tsunagi-tray
+          fi
+
+      - name: Package
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          version="${GITHUB_REF_NAME#v}"
+          stage="dist/pkg/${{ matrix.asset }}"
+          mkdir -p "$stage"
+
+          bin="tsng"
+          [[ "${{ runner.os }}" == "Windows" ]] && bin="tsng.exe"
+          cp "target/${{ matrix.target }}/release/${bin}" "$stage/"
+
+          # The tray GUI alongside it, where it was built (Linux: x86_64 only).
+          tray="tsunagi-tray"
+          [[ "${{ runner.os }}" == "Windows" ]] && tray="tsunagi-tray.exe"
+          if [[ "${{ runner.os }}" == "Linux" ]]; then
+            if [[ -n "${{ matrix.gnu_target }}" ]]; then
+              cp "target/${{ matrix.gnu_target }}/release/${tray}" "$stage/"
+            fi
+          else
+            cp "target/${{ matrix.target }}/release/${tray}" "$stage/"
+          fi
+
+          # The top-level README, plus the per-OS install notes and integration
+          # files, so the archive carries what the platform needs to set up.
+          cp README.md "$stage/"
+          case "${{ runner.os }}" in
+            Linux)
+              cp dist/linux/tsunagi.service dist/linux/install.sh \
+                 dist/linux/50-tsunagi-resolved.rules "$stage/"
+              cp dist/linux/README.md "$stage/INSTALL.md"
+              ;;
+            macOS)
+              cp dist/macos/cy.hexor.tsunagi.plist "$stage/"
+              cp dist/macos/README.md "$stage/INSTALL.md"
+              # The tray as an application bundle, for the Homebrew cask.
+              bash dist/macos/build-app.sh \
+                "target/${{ matrix.target }}/release/tsunagi-tray" \
+                "$version" "$stage/Tsunagi.app"
+              ;;
+            Windows)
+              cp dist/windows/README.md "$stage/INSTALL.md"
+              ;;
+          esac
+
+          archive="${{ matrix.asset }}-${version}.${{ matrix.archive }}"
+          if [[ "${{ matrix.archive }}" == "zip" ]]; then
+            (cd dist/pkg && 7z a "../../${archive}" "${{ matrix.asset }}")
+          else
+            tar -C dist/pkg -czf "${archive}" "${{ matrix.asset }}"
+          fi
+
+      # The Windows installer: the agent as a service, Wintun, the tray and the
+      # PATH entry in one setup program. Wintun is the signed library from
+      # wintun.net, pinned by hash so a changed download fails the build.
+      - name: Build the Windows installer
+        if: runner.os == 'Windows'
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          $version = $env:GITHUB_REF_NAME.TrimStart('v')
+          Invoke-WebRequest -Uri https://www.wintun.net/builds/wintun-0.14.1.zip -OutFile wintun.zip
+          $hash = (Get-FileHash wintun.zip -Algorithm SHA256).Hash
+          if ($hash -ne '07C256185D6EE3652E09FA55C0B673E2624B565E02C4B9091C79CA7D2F24EF51') {
+            throw "wintun.zip has an unexpected hash: $hash"
+          }
+          Expand-Archive wintun.zip -DestinationPath wintun-dl
+          choco install innosetup -y --no-progress
+          $iscc = Join-Path ${env:ProgramFiles(x86)} 'Inno Setup 6\ISCC.exe'
+          $stage = Join-Path $PWD 'dist/pkg/${{ matrix.asset }}'
+          & $iscc "/DAppVersion=$version" "/DSourceDir=$stage" `
+            "/DWintunDir=$(Join-Path $PWD 'wintun-dl/wintun/bin/amd64')" `
+            "/O$PWD" dist/windows/tsunagi.iss
+          if ($LASTEXITCODE -ne 0) { throw "ISCC failed with $LASTEXITCODE" }
+
+      # A dependency-free .deb for the x86_64 and aarch64 Linux builds: it drops
+      # the static binary, the systemd service and the polkit rule. Built with
+      # dpkg-deb (present on the runner); see dist/deb/build-deb.sh.
+      - name: Build the Debian package
+        if: matrix.deb
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          bash dist/deb/build-deb.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.deb_arch }}" \
+            "$version" \
+            "tsunagi_${version}_${{ matrix.deb_arch }}.deb"
+
+      # A dependency-free Arch package (.pkg.tar.zst), assembled with bsdtar and
+      # zstd (no makepkg); same binary, service and polkit rule. See
+      # dist/arch/build-pkg.sh.
+      - name: Build the Arch package
+        if: matrix.pkg_arch
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          pkgver="${version//-/_}"
+          bash dist/arch/build-pkg.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.pkg_arch }}" \
+            "$version" \
+            "tsunagi-${pkgver}-1-${{ matrix.pkg_arch }}.pkg.tar.zst"
+
+      # The GUI variants: the same agent plus the tray binary and a desktop
+      # launcher, named tsunagi-gui (conflicting with the headless packages).
+      - name: Build the GUI packages
+        if: matrix.gnu_target
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          pkgver="${version//-/_}"
+          tray="target/${{ matrix.gnu_target }}/release/tsunagi-tray"
+          bash dist/deb/build-deb.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.deb_arch }}" "$version" \
+            "tsunagi-gui_${version}_${{ matrix.deb_arch }}.deb" "$tray"
+          bash dist/arch/build-pkg.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.pkg_arch }}" "$version" \
+            "tsunagi-gui-${pkgver}-1-${{ matrix.pkg_arch }}.pkg.tar.zst" "$tray"
+
+      # @v3 on purpose: Gitea does not support the v4 artifact backend. This is
+      # what makes the archive (and the .deb, where built) show up under the run
+      # and hands it to `publish`.
+      - name: Upload the build outputs as run artifacts
+        uses: actions/upload-artifact@v3
+        with:
+          name: ${{ matrix.asset }}
+          path: |
+            ${{ matrix.asset }}-*.*
+            *.deb
+            *.pkg.tar.zst
+            tsunagi-setup-*.exe
+          if-no-files-found: error
+
+  publish:
+    name: Publish release
+    needs: build
+    runs-on: ubuntu-latest
+    steps:
+      - name: Download every build artifact
+        uses: actions/download-artifact@v3
+        with:
+          path: artifacts
+
+      - name: Show what was collected
+        run: find artifacts -type f
+
+      - name: Create the release and upload the archives
+        shell: bash
+        env:
+          # Gitea injects this with access to the current repository. If the
+          # release calls below return 403, the repo/runner is configured to
+          # hand Actions a read-only token — grant it write access to releases.
+          TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -u
+
+          tag="${GITHUB_REF_NAME}"
+          # Gitea sets GITHUB_API_URL to its own API root (…/api/v1); fall back
+          # to building it from the server URL if it is absent.
+          base="${GITHUB_API_URL:-${GITHUB_SERVER_URL}/api/v1}"
+          api="${base}/repos/${GITHUB_REPOSITORY}"
+          auth="Authorization: token ${TOKEN}"
+          echo "release API: $api"
+
+          prerelease=false
+          case "$tag" in *-*) prerelease=true ;; esac
+
+          # The release id for this tag, if it already exists.
+          code=$(curl -sS -o resp.json -w '%{http_code}' -H "$auth" "$api/releases/tags/$tag" || true)
+          echo "get release by tag: HTTP $code"
+          id=$(grep -o '"id":[0-9]*' resp.json 2>/dev/null | head -n1 | cut -d: -f2 || true)
+
+          if [ -z "$id" ]; then
+            code=$(curl -sS -o resp.json -w '%{http_code}' -X POST -H "$auth" \
+              -H 'Content-Type: application/json' \
+              -d "{\"tag_name\":\"$tag\",\"name\":\"tsunagi $tag\",\"prerelease\":$prerelease}" \
+              "$api/releases" || true)
+            echo "create release: HTTP $code"
+            id=$(grep -o '"id":[0-9]*' resp.json 2>/dev/null | head -n1 | cut -d: -f2 || true)
+            if [ -z "$id" ]; then
+              echo "could not create the release; response was:" >&2
+              cat resp.json >&2 || true
+              exit 1
+            fi
+          fi
+          echo "release id: $id"
+
+          # download-artifact@v3 nests each artifact in a directory named after
+          # its upload, with the single archive inside. The read loop runs in
+          # this shell (process substitution, not a pipe) so a failed upload
+          # actually fails the job.
+          rc=0
+          while IFS= read -r -d '' f; do
+            name=$(basename "$f")
+            code=$(curl -sS -o up.json -w '%{http_code}' -X POST -H "$auth" \
+              -F "attachment=@${f}" \
+              "$api/releases/$id/assets?name=${name}" || true)
+            echo "upload $name: HTTP $code"
+            if [ "$code" -lt 200 ] || [ "$code" -ge 300 ]; then
+              cat up.json >&2 || true
+              rc=1
+            fi
+          done < <(find artifacts -type f \( -name '*.tar.gz' -o -name '*.zip' -o -name '*.deb' -o -name '*.pkg.tar.zst' -o -name 'tsunagi-setup-*.exe' \) -print0)
+          exit "$rc"

.github/workflows/release.yml

@@ -0,0 +1,322 @@
+name: Release
+
+# GitHub version of the release pipeline. The build half mirrors the Gitea one
+# in .gitea/workflows/release.yml (static musl Linux via cargo-zigbuild, native
+# macOS and Windows); only the publish half differs: it uses the current
+# artifact actions (Node 24) and fills the release with `gh`.
+#
+# Every job is gated on `github.server_url == 'https://github.com'`, so this file
+# runs only on GitHub. Gitea runs its own copy in .gitea/workflows (which pins
+# the older @v3 artifact actions it supports); should a Gitea instance also read
+# .github/workflows, the guard skips these jobs there so they never fail on the
+# artifact backend or `gh`.
+on:
+  push:
+    tags: ["v*"]
+
+permissions:
+  contents: write
+
+env:
+  CARGO_TERM_COLOR: always
+  CARGO_NET_RETRY: "10"
+
+jobs:
+  build:
+    name: ${{ matrix.asset }}
+    if: ${{ github.server_url == 'https://github.com' }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-musl
+            asset: tsunagi-linux-x86_64
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: amd64
+            pkg_arch: x86_64
+            gnu_target: x86_64-unknown-linux-gnu
+          - os: ubuntu-latest
+            target: aarch64-unknown-linux-musl
+            asset: tsunagi-linux-aarch64
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: arm64
+            pkg_arch: aarch64
+          # 32-bit ARM with hardware float (ARMv7, `v7l`): single-board and edge
+          # devices. Statically linked like the others.
+          - os: ubuntu-latest
+            target: armv7-unknown-linux-musleabihf
+            asset: tsunagi-linux-armv7
+            archive: tar.gz
+            zig: true
+            deb: true
+            deb_arch: armhf
+            pkg_arch: armv7h
+          - os: macos-latest
+            target: aarch64-apple-darwin
+            asset: tsunagi-macos-aarch64
+            archive: tar.gz
+            zig: false
+          - os: windows-latest
+            target: x86_64-pc-windows-msvc
+            asset: tsunagi-windows-x86_64
+            archive: zip
+            zig: false
+
+    steps:
+      - uses: actions/checkout@v5
+
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          targets: ${{ matrix.target }}
+
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: ${{ matrix.target }}
+
+      # aws-lc-sys (iroh's rustls backend) and the bundled sqlite are C, and the
+      # musl targets have no prebuilt bindings, so bindgen needs libclang and the
+      # C build needs cmake. zig is the cross C compiler cargo-zigbuild drives.
+      - name: Install musl build and packaging tooling
+        if: matrix.zig
+        run: sudo apt-get update && sudo apt-get install -y --no-install-recommends cmake clang libclang-dev libarchive-tools zstd
+      - name: Install zig
+        if: matrix.zig
+        uses: mlugg/setup-zig@v2
+        with:
+          # Pin a real, archived Zig release. Left unpinned the action resolves
+          # to the newest version it knows of (a yet-unreleased 0.16.0-dev),
+          # whose tarball 404s on ziglang.org while the community mirrors flake.
+          version: 0.14.1
+      - name: Install cargo-zigbuild
+        if: matrix.zig
+        uses: taiki-e/install-action@v2
+        with:
+          tool: cargo-zigbuild
+
+      - name: Build
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [[ "${{ matrix.zig }}" == "true" ]]; then
+            # zig's linker ignores the `-O1` GNU-ld optimization flag Rust passes
+            # and says so on stderr, which rustc now surfaces as a linker_messages
+            # warning. It is cosmetic — the binary is correct — so quiet just that
+            # one lint to keep the musl build log clean. A real link error still
+            # fails the build.
+            export RUSTFLAGS="${RUSTFLAGS:-} -A linker_messages"
+            cargo zigbuild --release --locked --target "${{ matrix.target }}" --bin tsng
+          else
+            cargo build --release --locked --target "${{ matrix.target }}" --bin tsng
+          fi
+
+      # The tray GUI. On Linux it is built against glibc (an old baseline for
+      # portability), not musl: a GUI must dlopen libGL/X11/Wayland at runtime,
+      # which a static musl binary cannot. macOS and Windows build it natively.
+      - name: Build the tray GUI
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [[ "${{ runner.os }}" == "Linux" ]]; then
+            if [[ -n "${{ matrix.gnu_target }}" ]]; then
+              # tray-icon links GTK and libappindicator through pkg-config, which
+              # has no cross story under zig, so the Linux GUI is built natively
+              # (x86_64 only) against the runner's glibc.
+              sudo apt-get update
+              sudo apt-get install -y --no-install-recommends \
+                libgtk-3-dev libayatana-appindicator3-dev libxdo-dev
+              cargo build --release --locked \
+                --target "${{ matrix.gnu_target }}" --bin tsunagi-tray
+            else
+              echo "no GUI build for ${{ matrix.target }}"
+            fi
+          else
+            cargo build --release --locked --target "${{ matrix.target }}" --bin tsunagi-tray
+          fi
+
+      - name: Package
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          version="${GITHUB_REF_NAME#v}"
+          stage="dist/pkg/${{ matrix.asset }}"
+          mkdir -p "$stage"
+
+          bin="tsng"
+          [[ "${{ runner.os }}" == "Windows" ]] && bin="tsng.exe"
+          cp "target/${{ matrix.target }}/release/${bin}" "$stage/"
+
+          # The tray GUI alongside it, where it was built (Linux: x86_64 only).
+          tray="tsunagi-tray"
+          [[ "${{ runner.os }}" == "Windows" ]] && tray="tsunagi-tray.exe"
+          if [[ "${{ runner.os }}" == "Linux" ]]; then
+            if [[ -n "${{ matrix.gnu_target }}" ]]; then
+              cp "target/${{ matrix.gnu_target }}/release/${tray}" "$stage/"
+            fi
+          else
+            cp "target/${{ matrix.target }}/release/${tray}" "$stage/"
+          fi
+
+          # The top-level README, plus the per-OS install notes and integration
+          # files, so the archive carries what the platform needs to set up.
+          cp README.md "$stage/"
+          case "${{ runner.os }}" in
+            Linux)
+              cp dist/linux/tsunagi.service dist/linux/install.sh \
+                 dist/linux/50-tsunagi-resolved.rules "$stage/"
+              cp dist/linux/README.md "$stage/INSTALL.md"
+              ;;
+            macOS)
+              cp dist/macos/cy.hexor.tsunagi.plist "$stage/"
+              cp dist/macos/README.md "$stage/INSTALL.md"
+              # The tray as an application bundle, for the Homebrew cask.
+              bash dist/macos/build-app.sh \
+                "target/${{ matrix.target }}/release/tsunagi-tray" \
+                "$version" "$stage/Tsunagi.app"
+              ;;
+            Windows)
+              cp dist/windows/README.md "$stage/INSTALL.md"
+              ;;
+          esac
+
+          archive="${{ matrix.asset }}-${version}.${{ matrix.archive }}"
+          if [[ "${{ matrix.archive }}" == "zip" ]]; then
+            (cd dist/pkg && 7z a "../../${archive}" "${{ matrix.asset }}")
+          else
+            tar -C dist/pkg -czf "${archive}" "${{ matrix.asset }}"
+          fi
+
+      # The Windows installer: the agent as a service, Wintun, the tray and the
+      # PATH entry in one setup program. Wintun is the signed library from
+      # wintun.net, pinned by hash so a changed download fails the build.
+      - name: Build the Windows installer
+        if: runner.os == 'Windows'
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          $version = $env:GITHUB_REF_NAME.TrimStart('v')
+          Invoke-WebRequest -Uri https://www.wintun.net/builds/wintun-0.14.1.zip -OutFile wintun.zip
+          $hash = (Get-FileHash wintun.zip -Algorithm SHA256).Hash
+          if ($hash -ne '07C256185D6EE3652E09FA55C0B673E2624B565E02C4B9091C79CA7D2F24EF51') {
+            throw "wintun.zip has an unexpected hash: $hash"
+          }
+          Expand-Archive wintun.zip -DestinationPath wintun-dl
+          choco install innosetup -y --no-progress
+          $iscc = Join-Path ${env:ProgramFiles(x86)} 'Inno Setup 6\ISCC.exe'
+          $stage = Join-Path $PWD 'dist/pkg/${{ matrix.asset }}'
+          & $iscc "/DAppVersion=$version" "/DSourceDir=$stage" `
+            "/DWintunDir=$(Join-Path $PWD 'wintun-dl/wintun/bin/amd64')" `
+            "/O$PWD" dist/windows/tsunagi.iss
+          if ($LASTEXITCODE -ne 0) { throw "ISCC failed with $LASTEXITCODE" }
+
+      # A dependency-free .deb for the x86_64 and aarch64 Linux builds: it drops
+      # the static binary, the systemd service and the polkit rule. Built with
+      # dpkg-deb (present on the runner); see dist/deb/build-deb.sh.
+      - name: Build the Debian package
+        if: matrix.deb
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          bash dist/deb/build-deb.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.deb_arch }}" \
+            "$version" \
+            "tsunagi_${version}_${{ matrix.deb_arch }}.deb"
+
+      # A dependency-free Arch package (.pkg.tar.zst), assembled with bsdtar and
+      # zstd (no makepkg); same binary, service and polkit rule. See
+      # dist/arch/build-pkg.sh.
+      - name: Build the Arch package
+        if: matrix.pkg_arch
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          pkgver="${version//-/_}"
+          bash dist/arch/build-pkg.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.pkg_arch }}" \
+            "$version" \
+            "tsunagi-${pkgver}-1-${{ matrix.pkg_arch }}.pkg.tar.zst"
+
+      # The GUI variants: the same agent plus the tray binary and a desktop
+      # launcher, named tsunagi-gui (conflicting with the headless packages).
+      - name: Build the GUI packages
+        if: matrix.gnu_target
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME#v}"
+          pkgver="${version//-/_}"
+          tray="target/${{ matrix.gnu_target }}/release/tsunagi-tray"
+          bash dist/deb/build-deb.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.deb_arch }}" "$version" \
+            "tsunagi-gui_${version}_${{ matrix.deb_arch }}.deb" "$tray"
+          bash dist/arch/build-pkg.sh \
+            "target/${{ matrix.target }}/release/tsng" \
+            "${{ matrix.pkg_arch }}" "$version" \
+            "tsunagi-gui-${pkgver}-1-${{ matrix.pkg_arch }}.pkg.tar.zst" "$tray"
+
+      - name: Upload the build outputs as run artifacts
+        uses: actions/upload-artifact@v7
+        with:
+          name: ${{ matrix.asset }}
+          path: |
+            ${{ matrix.asset }}-*.*
+            *.deb
+            *.pkg.tar.zst
+            tsunagi-setup-*.exe
+          if-no-files-found: error
+
+  publish:
+    name: Publish release
+    needs: build
+    if: ${{ github.server_url == 'https://github.com' }}
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - name: Download every build artifact
+        uses: actions/download-artifact@v8
+        with:
+          path: artifacts
+
+      - name: Show what was collected
+        run: find artifacts -type f
+
+      - name: Create the release and upload the archives
+        shell: bash
+        env:
+          GH_TOKEN: ${{ github.token }}
+          TAG: ${{ github.ref_name }}
+        run: |
+          set -euo pipefail
+
+          # download-artifact nests each artifact under its own directory, so
+          # the archives are artifacts/<asset>/<file>.
+          mapfile -t files < <(find artifacts -type f \( -name '*.tar.gz' -o -name '*.zip' -o -name '*.deb' -o -name '*.pkg.tar.zst' -o -name 'tsunagi-setup-*.exe' \))
+          printf 'asset: %s\n' "${files[@]}"
+
+          prerelease=()
+          case "$TAG" in *-*) prerelease=(--prerelease) ;; esac
+
+          # Create the release on the first run; on a re-run, add/replace assets
+          # on the release that is already there.
+          if gh release view "$TAG" --repo "$GITHUB_REPOSITORY" >/dev/null 2>&1; then
+            gh release upload "$TAG" "${files[@]}" --repo "$GITHUB_REPOSITORY" --clobber
+          else
+            gh release create "$TAG" "${files[@]}" \
+              --repo "$GITHUB_REPOSITORY" \
+              --title "tsunagi $TAG" \
+              --generate-notes \
+              --verify-tag \
+              "${prerelease[@]}"
+          fi

AGENTS.md

@@ -8,28 +8,74 @@ before touching the code. Do not restate the full design here; follow the links.
 A proof-of-concept agent library for small private mesh networks. A user
 supplies a network name and one shared secret; agents derive the same network
 space independently, find each other, prove membership and exchange control
-messages. Scope and non-scope are in [README.md](README.md).
+messages. What it does and does not do is in [docs/architecture.md](docs/architecture.md) and [docs/threat-model.md](docs/threat-model.md).
 
 There is no central server, no owner, no registration and no majority vote.
 Design accordingly: a majority is not a root of trust.
 
+## User experience
+
+User convenience is a core requirement. The CLI should feel logical and
+predictable to an experienced administrator: use consistent selectors and
+flags, sensible defaults, and actionable errors. Let users refer to configured
+networks by their exact name as well as by id or a unique id prefix; do not
+make them copy an opaque identifier when the stored name identifies one network.
+When a selector is ambiguous, list the matching names and ids and change
+nothing. Never guess which resource a destructive command should affect.
+Keep help text and examples in sync with behavior, and preserve existing
+scripts when improving the interactive workflow.
+
 ## Architectural boundaries
 
 Keep these separate. Crossing them is the main thing to review for.
 
 - **Control plane vs data plane.** The separation is **logical, not physical**.
-  The control protocol in `src/proto/` knows nothing about packets, and
+  The control protocol in `crates/tsunagi-core/src/proto/` knows nothing about
-  `src/dataplane/` knows nothing about the control protocol; either can be
+  packets, and a protocol crate knows nothing about the control protocol;
-  replaced on its own. Both may ride on iroh — refusing to would throw away
+  either can be replaced on its own. Both may ride on iroh — refusing to
-  iroh's NAT traversal and force the data plane to reimplement it. They use
+  would throw away iroh's NAT traversal and force the data plane to
-  different ALPNs and different connections, so a busy or broken data plane
+  reimplement it. They use different ALPNs and different connections, so a
-  cannot disturb control traffic.
+  busy or broken data plane cannot disturb control traffic.
+- **One agent, one interface.** It belongs to the system level, along with
+  the addresses on it and the decision of whose packet is whose. Several
+  protocols may be carrying traffic at once and none of them owns the thing
+  they carry it for. A protocol is handed a routed packet and hands back a
+  decrypted one; it never creates an interface and never picks an address.
+- **One state directory, one identity, one live agent — many networks.** A
+  network is added to the agent that is already running, never by starting a
+  second one on the same directory. A second agent is a second identity and
+  is isolated: its own directories, its own interface, its own runtime. Do
+  not add anything that lets two agents share a directory, and do not make a
+  network's lifetime depend on the process that happened to start it.
+- **A protocol is a separate crate with its own version.** The version peers
+  compare is the *wire* version, never the software version: two peers on
+  different releases work together for as long as the bytes between them
+  have not changed. Nothing negotiated may be derived from anything that
+  moves with a release.
+- **A protocol never sees the network secret.** It proves who is at the other
+  end of a tunnel; proving membership of a network stays in the core, which
+  is why the authenticated transport does too.
+- **A relay carries bytes, never packets.** A datagram passed on for two
+  other peers goes link in, link out: it is never written to the middle's
+  interface or decrypted there. Multihop routing has a bounded hop limit;
+  equal-cost next hops are chosen per flow. Reachability is first-hand and
+  volatile: each authenticated member advertises its own direct links, of
+  whatever kind. A link is a way to reach a neighbour and a protocol is an
+  end-to-end tunnel to a peer; the two meet only in the envelope, so a
+  tunnel never depends on which transports carry it. Build routing tables on topology changes, never per packet. Transit
+  must not acquire a routing mutex or wait for a protocol/TUN reader.
+- **Broadcast fanout belongs at local IP ingress.** Participation is local to
+  each network, enabled by default, persisted and advertised to authenticated
+  peers. Resolve one source domain and target only its willing members. Remote
+  delivery never triggers another fanout. Physical LAN exports must extend
+  explicit ingress and source/destination admission policies; never bypass
+  ownership checks or add broadcast flooding to encrypted transit.
 - **Plugins never learn reachability.** An `IpPlugin` is handed a `PacketLink`
   per peer and moves datagrams over it. Addresses, hole punching and relays
-  belong to `src/dataplane/transport/`. A plugin announcement says *who*, never
+  belong to `crates/tsunagi-core/src/dataplane/transport/`. A plugin announcement says *who*, never
   *where*.
-- **The core never parses a plugin payload.** See `src/dataplane/mod.rs`. Only
+- **The core never parses a plugin payload.** See `crates/tsunagi-core/src/dataplane/mod.rs`. Only
-  `src/dataplane/wireguard/announcement.rs` interprets WireGuard payloads, and
+  `crates/tsunagi-wg-quic/src/announcement.rs` interprets `wg-quic` payloads, and
   only after bounding every field. A data plane failure must never stop the
   control plane.
 - **Derived, not claimed.** A peer's overlay address is derived from its public
@@ -37,7 +83,7 @@ Keep these separate. Crossing them is the main thing to review for.
   inbound packets are dropped unless their source is the address derived for
   the peer that sent them. Never trust an address a peer announces.
 - **Signed state is the only durable agreement.** A fact that must survive a
-  participant being away goes in `src/state/` as a record signed by its
+  participant being away goes in `crates/tsunagi-core/src/state/` as a record signed by its
   author, never in a session. Merging is deterministic, an older version never
   rolls back a newer one, and absence from a snapshot is not deletion. Never
   add a vote or a quorum: a majority is not a trust root here, and it would
@@ -46,6 +92,12 @@ Keep these separate. Crossing them is the main thing to review for.
   transaction, **before** the record is announced.
 - **Plugins own their system objects.** A plugin creates and removes its own
   interface and nothing else. Never touch routing, DNS or firewall settings.
+  The one exception is the system level, not a plugin: while a network's
+  broadcast is on, `overlay::hostrules` may install exactly a
+  `255.255.255.255/32` route and an inbound UDP firewall allowance (overlay range
+  only) for the interface the agent owns. Both are tagged `tsunagi:<interface>`,
+  reversible, removed on broadcast off, leave and shutdown, and computed from
+  local state, never from a remote announcement.
 - **Device identity vs network identity.** The iroh endpoint id is the device's
   public key. `NetworkId` is derived from name + secret only. Never conflate
   them, and never let one change the other.
@@ -65,7 +117,7 @@ Keep these separate. Crossing them is the main thing to review for.
    normal path *and* the important failures. See [docs/testing.md](docs/testing.md).
    Do not add tests for getters or to move a coverage number.
 2. **Never change `IDENTITY_SCHEME`, the derivation labels, or the transcript
-   encoding** in `src/identity/network.rs` and `src/proto/handshake.rs` without
+   encoding** in `crates/tsunagi-core/src/identity/network.rs` and `crates/tsunagi-core/src/proto/handshake.rs` without
    treating it as an incompatible protocol change. Bumping the crate version or
    the control protocol version must not change an existing `NetworkId`.
 3. **Secrets never leak.** Not into logs, not into `Debug`, not into status
@@ -76,7 +128,7 @@ Keep these separate. Crossing them is the main thing to review for.
    the library; tests opt out explicitly at the top of each file.
 5. **Bounds before allocation.** Frame lengths are checked against
    `Limits::max_frame_len` before a buffer is allocated. Every string, list and
-   queue has a limit in `src/config.rs`.
+   queue has a limit in `crates/tsunagi-core/src/config.rs`.
 6. **Failure is contained.** A bad signature, wrong secret, malformed packet or
    unknown version rejects one message or one session. It never stops another
    network and never stops the agent. There is no irreversible global error
@@ -93,25 +145,35 @@ Keep these separate. Crossing them is the main thing to review for.
    up shutdown. Wind tasks down with a grace period and then abort.
 10. **Nothing from a remote announcement becomes a shell command, a filesystem
     path or an OS setting.** The agent never touches interfaces or OS settings
-    it did not create.
+    it did not create (the tagged broadcast host rules above are for its own
+    interface).
 
 ## Where things live
 
 | path                | responsibility |
 |---------------------|----------------|
-| `src/identity/`     | device identity; deterministic network space identity and derived keys |
+| `crates/tsunagi-core/src/identity/`     | device identity; deterministic network space identity and derived keys |
-| `src/storage/`      | `state.sqlite`, `cache.sqlite`, directory ownership lock |
+| `crates/tsunagi-core/src/storage/`      | `state.sqlite`, `cache.sqlite`, directory ownership lock |
-| `src/discovery.rs`  | candidate sources; test and static backends |
+| `crates/tsunagi-core/src/discovery.rs`  | candidate sources; test and static backends |
-| `src/proto/`        | framing, message formats, membership handshake |
+| `crates/tsunagi-core/src/proto/`        | framing, message formats, membership handshake |
-| `src/net.rs`        | iroh endpoint adapter and observability snapshots |
+| `crates/tsunagi-core/src/net.rs`        | iroh endpoint adapter and observability snapshots |
-| `src/agent/`        | agent lifecycle, per-network runtimes, sessions, events, status |
+| `crates/tsunagi-core/src/agent/`        | agent lifecycle, per-network runtimes, sessions, events, status |
-| `src/state/`        | signed records that outlive a session, their merge rules and address allocation |
+| `crates/tsunagi-core/src/state/`        | signed records that outlive a session, their merge rules and address allocation |
-| `src/dataplane/`    | the plugin contract, the packet transport, and the WireGuard plugin |
+| `crates/tsunagi-core/src/overlay/`   | the one interface an agent owns: provisioning, the TUN, routing, source checks, broadcast host rules |
-| `src/bin/tsunagi.rs`| the command line agent; the only place that owns a runtime, a logger and signals |
+| `crates/tsunagi-core/src/dataplane/` | the protocol contract and the authenticated packet transport |
+| `crates/tsunagi-core/src/dns/`       | the DNS view of a network: zone, server, resolver publication |
+| `crates/tsunagi-wg-quic/`       | the `wg-quic` protocol: its keys, its announcement, its tunnels |
+| `crates/tsunagi/` | the command line agent, built as `tsng`; the only place that owns a runtime, a logger and signals |
 | `tests/`            | integration tests; `tests/common/` is the shared harness |
 
 Add abstractions only at real substitution or testing boundaries. Do not add a
-trait per struct. Prefer one crate with clear modules over many small crates.
+trait per struct.
+
+**The system level and its plugins are separate crates**, so that boundary is
+checked by the compiler and not by discipline: a plugin can reach only what
+`tsunagi` makes public, and carries its own version. Everything else stays
+one crate with clear modules — do not split further without a reason of that
+kind.
 
 ## Testing rules
 
@@ -133,7 +195,26 @@ trait per struct. Prefer one crate with clear modules over many small crates.
 - Running several library instances in one process is not a test of several
   system processes; do not describe it as one.
 
-## Before you open a change
+## Commit every completed change
+
+Agents must create a Git commit before finishing any task that changes this
+repository. Do not leave the user to make the commit. Inspect recent commit
+messages and follow their style: a short English subject stating the change,
+then a useful body explaining the problem, final behavior, architectural
+decisions, compatibility/migration effects, and checks actually performed.
+Write enough context for a future maintainer or LLM to understand why the code
+exists. Record failed or skipped checks honestly; never claim unrun tests.
+
+Before committing, inspect the diff and stage only this task's changes. Preserve
+unrelated user changes and never commit secrets or generated build artifacts.
+If Git has no effective author name or email, set only the missing repository-
+local values to `AB` and `ab@hexor.cy`. Do not override an existing identity or
+change global Git configuration. Report the resulting commit hash.
+
+Version bumps, release tags, and pushing commits/tags belong to the user unless
+they explicitly ask the agent to do them. Read-only tasks need no empty commit.
+
+## Required validation
 
 ```bash
 cargo fmt --all -- --check

Cargo.toml

@@ -1,94 +1,40 @@
-[package]
+# The system level and its plugins are separate crates, so the boundary
-name = "tsunagi"
+# between them is checked by the compiler rather than by discipline: a plugin
-version = "0.1.0"
+# can reach only what the core makes public, and carries its own version.
+[workspace]
+resolver = "3"
+members = ["crates/*"]
+exclude = ["vendor/mainline"]
+
+# Mainline 8.0.0 treats Windows UDP read timeouts as socket failures.
+# Keep the small receive-loop correction reproducible until an upstream fix.
+[patch.crates-io]
+mainline = { path = "vendor/mainline" }
+
+[workspace.package]
 edition = "2024"
 rust-version = "1.91"
-license = "MIT OR Apache-2.0"
+license = "WTFPL"
-description = "Proof-of-concept library for small private mesh networks: persistent agent identity, deterministic network spaces, iroh-based control plane."
+repository = "https://github.com/house-of-vanity/tsunagi"
-repository = "https://github.com/tsunagi-net/tsunagi"
-readme = "README.md"
-keywords = ["mesh", "p2p", "iroh", "networking"]
-categories = ["network-programming"]
 
-[features]
+[workspace.lints.rust]
-default = ["cli"]
-# The `tsunagi` command line binary. Library users can opt out.
-cli = ["dep:clap", "dep:anstream", "dep:anstyle", "dep:tracing-subscriber", "tokio/signal", "tun-device", "dns-publish"]
-# A real TUN device, so the WireGuard plugin can carry actual IP traffic.
-# Needs CAP_NET_ADMIN at run time; without it the plugin still runs and its
-# in-memory device can be used for tests.
-tun-device = ["dep:tun", "dep:rtnetlink", "dep:caps", "dep:futures-util"]
-# Telling the operating system where to send its questions. Linux only for
-# now; the zone and the server work without it.
-dns-publish = ["dep:zbus"]
-
-[[bin]]
-name = "tsunagi"
-path = "src/bin/tsunagi.rs"
-required-features = ["cli"]
-
-[dependencies]
-clap = { version = "4.5", features = ["derive", "env"], optional = true }
-# Already in the tree through clap. `anstream` strips the escapes when stdout
-# is not a terminal and turns on virtual terminal processing on Windows, so
-# colour is never written where it would show up as rubbish.
-anstream = { version = "1.0", optional = true }
-anstyle = { version = "1.0", optional = true }
-tracing-subscriber = { version = "0.3", features = ["env-filter"], optional = true }
-iroh = { version = "1.2", default-features = false, features = ["tls-ring"] }
-tokio = { version = "1.53", features = ["rt", "rt-multi-thread", "sync", "time", "macros"] }
-rusqlite = { version = "0.40", features = ["bundled"] }
-hkdf = "0.13"
-hmac = "0.13"
-sha2 = "0.11"
-subtle = "2.6"
-zeroize = { version = "1.9", features = ["derive"] }
-rand = "0.10"
-serde = { version = "1.0", features = ["derive"] }
-postcard = { version = "1.1", default-features = false, features = ["use-std"] }
-# Already in the tree through iroh. A packet codec, not a DNS server: the
-# zone logic is ours and a full server framework would be a large dependency
-# for answering A records from memory.
-simple-dns = "0.12"
-# The IANA top-level domain list, compiled in: one function, no
-# dependencies, no network. Used only to warn that a zone name shadows a
-# real public domain, never to refuse one.
-tld = "2.40"
-data-encoding = "2.11"
-hex = "0.4"
-thiserror = "2.0"
-tracing = "0.1"
-fs4 = { version = "1.1", features = ["sync"] }
-directories = "6.0"
-# The real system hostname, without a libc call of our own: this crate is
-# forbidden `unsafe` and will not make one.
-gethostname = "1.1"
-netwatch = "0.19.3"
-bytes = "1.12.1"
-boringtun = { version = "0.7.1", default-features = false }
-tun = { version = "0.8", features = ["async"], optional = true }
-
-# Linux-only interface provisioning. `rtnetlink` configures the interface in
-# process, so no `ip` invocation is ever needed; `caps` keeps CAP_NET_ADMIN
-# out of the effective set except during the moments it is used.
-[target.'cfg(target_os = "linux")'.dependencies]
-rtnetlink = { version = "0.23", optional = true }
-# Pure Rust D-Bus, no libdbus to link against. `tokio` rather than the
-# default reactor, because the agent brings its own.
-zbus = { version = "5.19", default-features = false, features = ["tokio"], optional = true }
-caps = { version = "0.5", optional = true }
-futures-util = { version = "0.3", default-features = false, optional = true }
-
-[dev-dependencies]
-tokio = { version = "1.53", features = ["rt", "rt-multi-thread", "sync", "time", "macros", "process"] }
-tempfile = "3.24"
-tracing-subscriber = { version = "0.3", features = ["env-filter"] }
-
-[lints.rust]
 missing_docs = "warn"
-unsafe_code = "forbid"
+unsafe_code = "deny"
 
-[lints.clippy]
+[workspace.lints.clippy]
 unwrap_used = "warn"
 expect_used = "warn"
 panic = "warn"
+
+[workspace.dependencies]
+iroh = { version = "1.2", default-features = false, features = ["tls-ring", "portmapper"] }
+tokio = { version = "1.53", features = ["rt", "rt-multi-thread", "sync", "time", "macros"] }
+serde = { version = "1.0", features = ["derive"] }
+postcard = { version = "1.1", default-features = false, features = ["use-std"] }
+bytes = "1.12.1"
+thiserror = "2.0"
+tracing = "0.1"
+hex = "0.4"
+data-encoding = "2.11"
+tempfile = "3.24"
+tracing-subscriber = { version = "0.3", features = ["env-filter"] }

LICENSE

@@ -0,0 +1,13 @@
+            DO WHAT THE FUCK YOU WANT TO PUBLIC LICENSE
+                    Version 2, December 2004
+
+ Copyright (C) 2004 Sam Hocevar <sam@hocevar.net>
+
+ Everyone is permitted to copy and distribute verbatim or modified
+ copies of this license document, and changing it is allowed as long
+ as the name is changed.
+
+            DO WHAT THE FUCK YOU WANT TO PUBLIC LICENSE
+   TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
+
+  0. You just DO WHAT THE FUCK YOU WANT TO.

README.md

@@ -1,457 +1,184 @@
-# tsunagi
+<p align="center">
+  <img src="dist/linux/tsunagi.svg" width="96" alt="">
+</p>
 
-A proof-of-concept Rust library for **small private mesh networks** — a handful
+<h1 align="center">tsunagi</h1>
-of friends, home machines, a few servers. Units to dozens of participants, not
-thousands.
 
-The end user configures exactly two things:
+<p align="center">
+  <b>A mesh VPN with no server.</b><br>
+  Pick a network name, share a secret, and your machines are on one private network.
+</p>
 
-```text
+<p align="center">
-network_name
+  <a href="https://github.com/house-of-vanity/tsunagi/releases/latest"><img alt="release" src="https://img.shields.io/github/v/release/house-of-vanity/tsunagi?include_prereleases"></a>
-secret          # one shared secret; "password" and "secret" mean the same value
+  <a href="https://github.com/house-of-vanity/tsunagi/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/house-of-vanity/tsunagi/actions/workflows/ci.yml/badge.svg"></a>
+  <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-WTFPL-blue"></a>
+</p>
+
+```console
+laptop$ tsng join -n home
+joined `home`
+  secret  tsn1u7c…
+
+server$ tsng join -n home -s tsn1u7c…
+server$ tsng id hostname server
+
+laptop$ ssh server.home
 ```
 
-From those, every agent independently derives the same network space. There is
+No account, no coordination server, no admin console, no keys to approve.
-no central server, no network owner with special powers, no registration and no
+Anyone who knows the **name** and the **secret** is in. Nobody else can find
-majority vote. Anyone who knows the parameters can join; nobody has to trust
+the network, let alone enter it.
-anybody else.
 
-## What this proof of concept actually does
+## Why
 
-A working library with **real iroh connections** and integration tests:
+Tailscale, Headscale and ZeroTier are great, but each needs a control server:
+theirs, or one you have to run and keep alive. tsunagi doesn't. Machines find
+each other through the public BitTorrent DHT, connect directly (hole punching,
+with relay fallback), and keep the network state among themselves.
 
-- persistent device identity stored in SQLite, stable across restarts;
+|                        | Tailscale          | Headscale        | ZeroTier            | **tsunagi**         |
-- several independent networks at once in one agent;
+| ---------------------- | ------------------ | ---------------- | ------------------- | ------------------- |
-- deterministic network identity derived from name + secret;
+| Control server         | hosted by vendor   | you run it       | hosted or self-run  | **none**            |
-- candidates supplied by a replaceable discovery component;
+| Account / sign-in      | yes                | yes              | yes                 | **no**              |
-- real iroh connections plus an explicit mutual proof of network membership;
+| To join a network      | sign in, get approved | pre-auth key  | network ID, approval | **name + secret**  |
-- a small versioned control protocol: handshake, hostname/capability
+| Direct connections     | yes                | yes              | yes                 | yes                 |
-  announcement, ping/pong;
+| Exit nodes             | yes                | yes              | via routes          | yes                 |
-- automatic reconnect with bounded exponential backoff and jitter;
-- status snapshots, an event stream and honest diagnostics;
-- configuration restored after a restart;
-- correct behaviour when the disposable cache is missing or corrupt;
-- a **WireGuard data plane**, in userspace: its own key per network, an IPv6
-  overlay with deterministically derived addresses and optional IPv4, real
-  tunnels carried over iroh, and address ownership enforced rather than
-  believed;
-- a **command line agent**, `tsunagi`, with a local control socket.
 
-### What it deliberately does **not** do
+## Features
 
-Not implemented, and not pretended to be: Mainline DHT, DNS, routing through
+- **Two things to know:** a network name and a secret. A secret is generated for you when you make a network.
-intermediate participants, a full CRDT, dynamically loaded plugins, a system
+- **Direct, encrypted links** between members (WireGuard cryptography), with relay fallback behind strict NATs.
-service, a complete CLI, or a local control socket. Snapshot synchronisation
+- **Names for machines:** `server.home` just works, even while `server` is switched off. A name belongs to whoever claimed it first; a later device with the same name keeps working by address and is told so.
-and signed revocations are designed for but not implemented — see
+- **Exit nodes:** send all your internet traffic through any member that offers one.
-[docs/sync-model.md](docs/sync-model.md). The WireGuard plugin's own limits,
+- **Multi-hop:** members that can't reach each other directly are routed through a member that can, over whatever kind of link each hop has. A member reached that way is online, and shown as relayed; traffic moves to a direct link by itself the moment one comes up.
-including that its system backend is Linux-only, are in
+- **LAN games:** UDP broadcasts are relayed, so "LAN" lobbies show up across the mesh.
-[docs/wireguard.md](docs/wireguard.md#limits-and-future-work).
+- **Several networks at once** on one device.
+- **Tray app** for Windows, macOS and Linux; a command line for everything.
 
-**Control and data are separated logically, not physically.** Both ride on
+## Install
-iroh, on different ALPNs and different connections, so the data plane inherits
-iroh's hole punching and relay fallback instead of reimplementing them — while
-the control protocol still knows nothing about packets and can keep a different
-transport underneath it later. Filtering user traffic remains the operating
-system's and the user's responsibility, not this library's.
 
-## Requirements
+Grab a build from the [latest release](https://github.com/house-of-vanity/tsunagi/releases/latest).
 
-- Rust 1.91 or newer (iroh 1.2 requires it), edition 2024. Pinned dependencies
+| Platform              | How                                                                              |
-  in `Cargo.lock`.
+| --------------------- | -------------------------------------------------------------------------------- |
-- No internet, no DHT, no public relay, no administrator rights and no changes
+| **Windows**           | Run `tsunagi-setup-<version>-x86_64.exe`. Installs the service, the tray and `tsng`. |
-  to OS network settings are needed to build or test.
+| **macOS**             | `brew tap house-of-vanity/tap && brew install --cask tsunagi-gui`<br>CLI only: `brew install tsunagi` |
-- WireGuard runs in userspace (boringtun): **no kernel module and no `wg`
+| **Debian / Ubuntu**   | `sudo apt install ./tsunagi-gui_<version>_amd64.deb` (`tsunagi_…` for no tray)   |
-  tool**. Only creating a real network interface needs `CAP_NET_ADMIN`, and
+| **Arch**              | `sudo pacman -U tsunagi-gui-<version>-1-x86_64.pkg.tar.zst`                      |
-  `--no-tun` skips even that.
+| **NixOS**             | [Flake and module](#nixos) below                                                 |
+| **Any other Linux**   | Unpack the tarball, then `sudo ./install.sh && sudo systemctl enable --now tsunagi` |
 
-## Trying it on two machines
+Linux packages run the agent as an unprivileged service user. Add yourself to
+the `tsunagi` group (`sudo usermod -aG tsunagi $USER`, then log in again) to
+use `tsng` and the tray without `sudo`.
 
-On the first machine:
+## Quick start
 
-```bash
+```sh
-cargo build --release
+tsng join -n home                 # first machine: makes the network and prints a secret
-./target/release/tsunagi id secret generate   # prints tsn1...; share it privately
+tsng join -n home -s tsn1…        # every other machine
-./target/release/tsunagi status          # this device, the agent, and this host
+tsng status                       # who is online, over which path
-
-./target/release/tsunagi up --network lab --secret "$SECRET" --wireguard
 ```
 
-It prints its endpoint id and then waits. On the second machine, pass that id:
+Every member gets an address in `10.13.37.0/24` and a name like `<host>.home`.
+Machines are called `tsunagi-<id>` until you rename them: `tsng id hostname laptop`.
 
-```bash
+### Exit nodes
-./target/release/tsunagi up --network lab --secret "$SECRET" --wireguard \
+
-  --peer <endpoint-id-from-the-first-machine>
+```sh
+tsng network exit-node home on    # on the machine that should be the exit
+tsng exit-node                    # on another one: who offers an exit
+tsng exit-node laptop             # send everything through it
+tsng exit-node off
 ```
 
-Within a few seconds both print something like:
+The tray does the same with a checkbox. An exit node sees everything you send
+through it, so use only members you trust.
+[Details and per-OS notes](docs/exit-node.md).
 
-```text
+## Platforms
-  + peer b47c958462 connected over Direct rtt=Some(4.5ms)
-  + data link to b47c958462 for wireguard: Direct via Ip(…), datagram 1382
 
---- status ---
+| Platform                  | Build        | CLI `tsng` | Tray | Exit node: offer / use | Packages                            |
-control: 1 peer(s), 0 dial failure(s), 0 handshake failure(s)
+| ------------------------- | ------------ | :--------: | :--: | :--------------------: | ----------------------------------- |
-wireguard: tsunkkcp43lmdje on fd15:1d9e:fa21:f201:…/64 mtu 1280, 1/1 tunnel(s) established
+| Linux x86_64              | release      |     ✔      |  ✔   |        ✔ / ✔           | `.deb`, `.pkg.tar.zst`, tarball, Nix |
-  4jO4kx9Z fd15:1d9e:fa21:f201:… handshake 3s ago tx=0 rx=0 dropped=0 path=Direct via Ip(…)
+| Linux aarch64             | release      |     ✔      |  –¹  |        ✔ / ✔           | `.deb`, `.pkg.tar.zst`, tarball, Nix |
-```
+| Linux armv7               | release      |     ✔      |  –¹  |        ✔ / ✔           | `.deb`, `.pkg.tar.zst`, tarball      |
+| macOS, Apple silicon      | release      |     ✔      |  ✔   |        ✔ / ✔²          | Homebrew, tarball                    |
+| macOS, Intel              | from source  |     ✔³     |  ✔³  |        ✔ / ✔²³         | –                                    |
+| Windows x86_64            | release      |     ✔      |  ✔   |        ✔ / ✔²          | installer, zip                       |
 
-`1/1 tunnel(s) established` means a real WireGuard handshake completed.
+<sup>¹ Not in the release; builds from source (and through Nix).
+² Newer and less battle-tested than on Linux.
+³ Builds from source, not covered by CI.</sup>
 
-## Checking that it works
+The Linux CLI is a single static binary: it runs on any distribution.
 
-From another shell on either machine:
+## NixOS
 
-```bash
+The repository is a flake with a package and a module.
-tsunagi status
-```
 
-```text
+```nix
-endpoint  7d76ccbbc21bf30767e14422c0494740a2cecc02aa9f82d5b8d57bdae350e7fc
+# flake.nix
-hostname  tsunagi-7d76ccbbc2
+{
-bound     0.0.0.0:41641
+  inputs.tsunagi.url = "github:house-of-vanity/tsunagi";
 
-network lab (z2o4qwrvnj3zb6st2aoqg4abf342j662q2ujttqrsmz22argk2ba)  active
+  outputs = { nixpkgs, tsunagi, ... }: {
-  peer b345d5271b  tsunagi-b345d5271b  Direct  rtt 24ms
+    nixosConfigurations.myhost = nixpkgs.lib.nixosSystem {
-  overlay tsunz2o4qwrvnj3 fd09:…:c1c6/64 and 100.110.49.177 mtu 1280  1/1 tunnel(s) up
+      modules = [
-    SDsEb/WF  fd09:…:c4b / 100.65.243.53  handshake 4s ago  tx 0 rx 0  Direct via Ip(…)
+        tsunagi.nixosModules.default
-```
+        {
-
+          services.tsunagi = {
-`1/1 tunnel(s) up` and a recent handshake mean the tunnel is live. Then send
+            enable = true;
-real traffic to the peer's overlay address:
+            tray.enable = true;      # optional, for desktops
-
+            users = [ "alice" ];     # may use tsng and the tray without sudo
-```bash
+          };
-ping6 fd09:…:c4b        # or
-ping  100.65.243.53
-```
-
-`tx` and `rx` in the status should start moving.
-
-IPv6 works out of the box: each member's address is derived from the network
-id and collides with essentially nothing.
-
-**IPv4 addresses are allocated and then remembered.** The default range is
-`10.13.37.0/24`; the first member to join settles it and later members adopt
-what they find, so `--ipv4-range` only matters for whoever starts the network:
-
-```bash
-tsunagi up --network lab --secret "$SECRET" --wireguard --ipv4-range 10.44.0.0/16
-tsunagi up --network lab --secret "$SECRET" --wireguard --ipv4-range none   # IPv6 only
-```
-
-An address is claimed with a record signed by that member's persistent device
-key, stored, and merged between every replica. A member that disappears for a
-month comes back to the same address, because the claim outlived the session.
-No vote is involved — see
-[docs/wireguard.md](docs/wireguard.md#ipv4-allocated-signed-and-kept) and
-[docs/sync-model.md](docs/sync-model.md).
-
-Because the address is allocated at run time rather than derived, it is not
-known until the agent has started and agreed with its peers. The agent then
-assigns it to the interface itself.
-
-## Names
-
-`--dns` serves a local DNS zone for the network's members, so they can be
-reached by name instead of by address:
-
-```bash
-tsunagi up --network lab --secret "$SECRET" --wireguard --dns
-dig @10.13.37.69 -p 5354 music.lab
-```
-
-Names come from signed state, which is the point: **a member that is
-switched off still resolves**, because its claim outlived the session. IPv4
-only — the IPv6 overlay address derives from a key that travels in live
-announcements, so it cannot be answered for a member that is away, and
-answering for some members and not others depending on who is online is
-worse than not answering.
-
-The zone is the network name unless `--dns-zone` says otherwise. It is
-yours to choose, so a name that shadows a real public domain is reported and
-then used: `--dns-zone ru` warns that every public `.ru` name becomes
-unreachable from this host, and then does it. `.internal` is reserved for
-exactly this and is never mentioned.
-
-On Linux the agent tells systemd-resolved to send questions for that suffix
-here, over D-Bus, scoped to the overlay interface and as a *routing* domain
-so it never becomes the resolver for anything else. resolved drops the whole
-setting when the interface goes, and the interface goes with the agent.
-
-That last step needs permission that `CAP_NET_ADMIN` does not give:
-systemd-resolved asks polkit, and polkit decides by user. Running as a
-system service is enough; so is a polkit rule granting this user the
-`org.freedesktop.resolve1.set-*` actions. **Without it the server still
-runs** — `tsunagi status` prints where it is listening and the exact `dig`
-line — so the automatic part is missing, not the feature.
-
-The server is authoritative for its zone and nothing else. No recursion, no
-forwarding, no cache: pointing a resolver at it can never make it a route to
-the outside.
-
-## Privileges
-
-On Linux the agent **manages its own overlay interface**. It creates the TUN
-interface, sets the MTU, brings it up and assigns both overlay addresses, all
-over netlink in process — no `ip` invocation, no shell, nothing that a
-remote peer could influence.
-
-That needs `CAP_NET_ADMIN`, granted once:
-
-```bash
-sudo setcap cap_net_admin+p /usr/local/bin/tsunagi
-```
-
-`+p` rather than `+ep`: the capability is then *permitted* but not
-*effective*, and the agent raises it only around the handful of netlink calls
-that need it — a few milliseconds at startup, and again if its address
-allocation changes. Everything else, including every byte from the network,
-is handled with it lowered. `+ep` works too; the agent lowers it on the way
-in.
-
-`tsunagi status` says which of these applies on the host it runs on, along
-with what the agent is doing. It grades each finding: **ok** for what works,
-**warn** for what the agent runs without and you can fix from the line it
-prints, **FAIL** for what it cannot work around. The words carry the grade as
-well as the colour, so the report reads the same piped to a file or on a
-terminal without colour, and it honours `NO_COLOR`.
-
-Members are listed online first, then the ones that are away. A member that
-is away is reported plainly rather than flagged: in a mesh of laptops it is
-the ordinary condition, not a fault. The signed state is what makes that
-sayable — it remembers who belongs while they are gone, so the report can say
-"offline, 10.13.37.99 still reserved for it" instead of leaving a
-dial-failure counter to imply it. Counters are history and are never graded:
-a peer that left and came back should not leave the report looking broken.
-
-`id` is the other half: it shows what this device is — its signing key, the
-name it answers to, and the secret of every network it has joined — and
-changes those. Every item takes the same shape, so there is nothing to
-remember: name it to see it, name it with a value to change it.
-
-```
-tsunagi id                      everything about this device
-tsunagi id hostname             the name it answers to
-tsunagi id hostname mango       change it
-tsunagi id key                  the key it signs with
-tsunagi id key rotate           replace that key
-tsunagi id secret               the secret of each joined network
-tsunagi id secret generate      a fresh secret for a network that does not exist yet
-```
-
-Secrets appear in `id`, which is where you go to ask for one, and never in
-`status`, in a log, in a `Debug` rendering or in anything sent to a peer.
-
-The name is part of the signed state, so changing it revokes the previous
-one: there is one record per author, a new version replaces the whole claim,
-and no replica can keep the old name standing. Changing it while the agent
-runs goes through the agent, which republishes and tells its peers straight
-away.
-
-Replacing the signing key makes this device a different member, and it loses
-the address and name the old key held — nothing can sign on a retired key's
-behalf, and by design there is no authority that could overrule an author. So
-the outgoing key signs a release for every network on its way out, which
-frees them for whoever wants them next, and the whole thing commits at once.
-
-`status` and `id` both prefer a running agent, which is live and
-authoritative, and fall back to reading the state store when there is none.
-Reading takes no directory lock, so neither has to wait for the agent it is
-asking about — nor does either need one to be running.
-
-### It cleans up after itself
-
-The interface is tied to an open file descriptor and is deliberately **not**
-made persistent, so the kernel removes it when the agent exits — on a clean
-shutdown, on a panic, on `SIGKILL`, on power loss alike. Keeping it is what
-would take an action; removing it is the default.
-
-If something is left behind anyway — an interface made by an older version's
-manual recipe, or one from a run killed in the instant between creating it and
-recording it — the next start **replaces it**, along with any stale addresses
-it carried. Two things are never touched:
-
-* an interface that is not a TUN, because the name colliding with somebody's
-  bridge is not a reason to destroy the bridge;
-* a TUN that another process is holding open, because that is a working
-  overlay belonging to somebody else — most likely a second agent on this
-  host, which should be given a different `--wg-prefix`.
-
-Both of those refuse with an explanation rather than guessing.
-
-Two settings the manual recipe used to need are gone with it.
-`keep_addr_on_down` existed only because an interface nobody held open lost
-carrier and had its IPv6 addresses flushed, and `nodad` only because duplicate
-address detection can never finish without carrier. An interface held open for
-its whole life has carrier for its whole life.
-
-### The MTU is 1280
-
-That is the minimum IPv6 requires (RFC 8200), and Linux enforces it by
-disabling IPv6 outright on an interface below it — the per-device
-`/proc/sys/net/ipv6` entries vanish and adding an address fails with
-`Invalid argument`. A smaller MTU cannot work at all, so the agent refuses one
-rather than letting it fail later. See
-[docs/wireguard.md](docs/wireguard.md#mtu) for the ceiling that pushes back
-from the other side.
-
-### Summary
-
-| approach | agent runs as | notes |
-|---|---|---|
-| `setcap cap_net_admin+p` | ordinary user, one capability | recommended: nothing to prepare, nothing left behind. Lost on every rebuild or copy of the binary. |
-| systemd service | `User=`, `AmbientCapabilities=CAP_NET_ADMIN` | the same, for an installed service |
-| `sudo tsunagi up` | root | everything works, nothing is isolated |
-| `--no-tun` | ordinary user, no capabilities | tunnels run and handshake, traffic never reaches the OS |
-
-**Not implemented yet.** macOS and Windows have no provisioner: both need
-real platform work — `utun` and `SystemConfiguration` on one, the IP Helper
-API and a Wintun adapter on the other. There the agent says so and `--no-tun`
-is the way to run it; the control plane and the tunnels are unaffected. The
-decision logic that says *what* to change is shared and tested on every
-platform, so only the execution is left to write.
-
-## Checks
-
-```bash
-cargo fmt --all -- --check
-cargo clippy --locked --workspace --all-targets -- -D warnings
-cargo test --locked --workspace --all-targets
-```
-
-The whole suite runs offline on loopback. Set `TSUNAGI_TEST_LOG=tsunagi=debug`
-to see agent logs while a test runs.
-
-There are also two runnable demos, which are demos and not substitutes for the
-tests:
-
-```bash
-cargo run --example two_agents       # control plane only
-cargo run --example wireguard_mesh   # a WireGuard overlay carrying a real packet
-```
-
-Both run with no privileges and change nothing on the host.
-
-## Usage
-
-```rust,no_run
-use std::sync::Arc;
-use tsunagi::config::{AgentConfig, StoragePaths, TransportPolicy};
-use tsunagi::discovery::SharedMemoryDiscovery;
-use tsunagi::identity::{NetworkName, NetworkSecret};
-use tsunagi::proto::ControlMessage;
-use tsunagi::{Agent, Result};
-
-// The library never starts a runtime, installs a logger, handles signals,
-// forks, or calls process::exit. The binary owns all of that.
-#[tokio::main]
-async fn main() -> Result<()> {
-    let config = AgentConfig::new(StoragePaths::user_default()?)
-        .with_transport(TransportPolicy::N0Defaults)
-        .with_discovery(Arc::new(SharedMemoryDiscovery::new()));
-
-    let agent = Agent::spawn(config).await?;
-
-    let name = NetworkName::new("kitchen-table")?;
-    let secret = NetworkSecret::generate();       // 32 random bytes
-    println!("share this: {}", secret.encode().as_str());
-
-    let network = agent.join_network(&name, &secret).await?;
-
-    let mut events = agent.subscribe();
-    tokio::spawn(async move {
-        while let Ok(event) = events.recv().await {
-            println!("{event:?}");
         }
-    });
+      ];
-
+    };
-    for peer in agent.network_status(network).await?.connected_peers() {
+  };
-        agent
-            .send(network, peer, ControlMessage::Ping { seq: 1, payload: vec![] })
-            .await?;
-    }
-
-    agent.shutdown().await;
-    Ok(())
 }
 ```
 
-`TransportPolicy::LocalOnly` is the default, so a plain `AgentConfig::new` never
+Just trying it: `nix run github:house-of-vanity/tsunagi#tsunagi-cli -- --help`.
-reaches the internet by accident. Opt into `DirectOnly` or `N0Defaults`
+`packages.<system>.tsunagi` (with the tray) and `tsunagi-cli` are also exported,
-explicitly.
+along with `overlays.default`.
 
-## How peers find each other
+## Build from source
 
-Two different lookups are involved, and only one of them is this project's:
+Rust 1.91 or newer.
 
-**1. Resolving one endpoint's address — iroh's, and it works today.**
+```sh
-With `--transport relay` or `--transport direct`, iroh publishes a signed
+cargo build --release --bin tsng               # the agent and CLI
-record of this endpoint's addresses, keyed by its endpoint id, to the public
+cargo build --release --bin tsunagi-tray       # the tray (Linux needs GTK 3 and libayatana-appindicator)
-service run by Number 0 — "n0", the company behind iroh — at `dns.iroh.link`,
+```
-over pkarr and DNS, and resolves other endpoints the same way. That is why `--peer <endpoint-id>` works with no address attached:
-iroh looks it up. None of that code is ours.
 
-**2. Finding who is in a network — ours, and it is still manual.**
+Windows needs [Wintun](https://www.wintun.net) next to `tsng.exe`.
-`NetworkDiscovery` maps a secret-derived `DiscoveryKey` to a set of *candidate*
+Per-platform notes: [Linux](dist/linux/README.md), [Arch](dist/arch/README.md),
-members. Two backends exist: `StaticBootstrap` (what `--peer` feeds) and an
+[macOS](dist/macos/README.md), [Windows](dist/windows/README.md).
-in-memory one for tests. The planned Mainline DHT backend, which would let
-members find each other from the network secret alone, is **not implemented**.
-So today you bootstrap by passing one peer's id; after that the mesh is
-whatever those agents reach.
 
-What this means in practice:
+## How it works
 
-- With `relay` or `direct`, **your endpoint id and IP addresses are published
+- Your name and secret derive a network identity. Members publish signed
-  to a public third-party service** (Number 0's, unless you change it). They are not secret, and the network secret is
+  contact records to the public Mainline DHT under it, and find each other there.
-  never published, but an observer of that service learns that your endpoint
+- Every connection proves knowledge of the secret before it's accepted.
-  exists and where it is. `--transport local` publishes nothing.
+  Traffic between members is end-to-end encrypted, relays included.
-- A relay, when one is needed, sees the volume and timing of your traffic — not
+- Addresses and names are claimed with each device's own key and replicated
-  its contents. The default relays are Number 0's, in the US, EU and
+  between members. There is no one in charge, so nothing to be taken down.
-  Asia-Pacific.
 
-## Storage
+The secret is the only key, so make it a good one (the generated ones are).
-
+The [threat model](docs/threat-model.md) says what this does and doesn't protect against.
-Two physically separate SQLite files, placed wherever the library's
-configuration says (`StoragePaths`). A future system service supplies its own
-paths; tests always use temporary directories.
-
-| file           | holds                                            | when damaged            |
-|----------------|--------------------------------------------------|-------------------------|
-| `state.sqlite` | device identity, network configuration, hostname | clear error, never reset |
-| `cache.sqlite` | address hints and other recoverable data          | discarded and recreated  |
-
-The WireGuard plugin keeps its own keys in its own `wireguard.sqlite`, wherever
-its configuration points, because plugin keys are neither the iroh identity nor
-the network secret.
-
-The command line agent puts everything under the platform's per-user
-directories by default; `--state-dir` and `--cache-dir` override them.
-
-One state directory belongs to one live agent, enforced with a real OS file
-lock rather than an existence check.
 
 ## Documentation
 
-- [docs/architecture.md](docs/architecture.md) — module boundaries and runtime.
+[Architecture](docs/architecture.md) ·
-- [docs/wireguard.md](docs/wireguard.md) — the WireGuard plugin: overlay
+[Protocol](docs/protocol.md) ·
-  addressing, announcements, backends, reconciliation.
+[WireGuard data plane](docs/wireguard.md) ·
-- [docs/protocol.md](docs/protocol.md) — identity derivation, framing, handshake.
+[Routing](docs/routing.md) ·
-- [docs/sync-model.md](docs/sync-model.md) — the planned signed-state model and
+[Exit nodes](docs/exit-node.md) ·
-  what is deliberately not built yet.
+[Threat model](docs/threat-model.md) ·
-- [docs/threat-model.md](docs/threat-model.md) — threat model and known limits.
+[Contributing](AGENTS.md)
-- [docs/testing.md](docs/testing.md) — what the suite covers and what it does not.
-- [AGENTS.md](AGENTS.md) — rules for anyone (human or agent) changing this repo.
-
-## Security in one paragraph
-
-Membership is proved by an HMAC over a transcript keyed by a value derived from
-the shared secret, bound to the specific iroh connection through the TLS
-exporter, to the network id, to both endpoint identities and to distinct role
-labels. This targets high-entropy secrets: there is no PAKE here, so a short
-human passphrase is guessable offline by anyone who can reach the handshake.
-Anyone who knows the secret is a full participant and can create many
-identities. Read [docs/threat-model.md](docs/threat-model.md) before relying on
-any of this.
 
 ## Licence
 
-MIT OR Apache-2.0.
+[WTFPL](LICENSE): do what the fuck you want to.

crates/tsunagi-core/Cargo.toml

@@ -0,0 +1,113 @@
+[package]
+name = "tsunagi-core"
+version = "0.1.0"
+description = "Small private mesh networks: persistent agent identity, deterministic network spaces, an iroh control plane, signed state and a local DNS view."
+readme = "../../README.md"
+keywords = ["mesh", "p2p", "iroh", "networking"]
+categories = ["network-programming"]
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[lib]
+name = "tsunagi"
+
+[features]
+default = []
+# A real TUN device, so a plugin can carry actual IP traffic. On Linux it
+# needs CAP_NET_ADMIN; on Windows the process must run elevated and
+# `wintun.dll` must be reachable. Without either the in-memory device serves
+# tests. The platform-specific dependencies are declared per target, so
+# enabling this only pulls what the target it is built for actually uses.
+tun-device = [
+    "dep:tun",
+    "dep:rtnetlink",
+    "dep:caps",
+    "dep:futures-util",
+]
+# Telling the operating system where to send its DNS questions. systemd-resolved
+# on Linux, the Name Resolution Policy Table on Windows; the zone and the
+# server work without it either way.
+dns-publish = ["dep:zbus"]
+# The test harness in `testing`, for this crate's own tests and for a
+# protocol crate's. Not on by default: it is only of use to a test.
+testing = ["dep:tempfile", "dep:tracing-subscriber"]
+
+[dependencies]
+arc-swap = "1.9"
+iroh.workspace = true
+# `net` for the local control interface: a Unix socket on Unix, a named pipe
+# on Windows. Both live behind the same `net` feature.
+tokio = { workspace = true, features = ["net"] }
+tokio-stream = "0.1"
+serde.workspace = true
+postcard.workspace = true
+bytes.workspace = true
+thiserror.workspace = true
+tracing.workspace = true
+hex.workspace = true
+data-encoding.workspace = true
+rusqlite = { version = "0.40", features = ["bundled"] }
+hkdf = "0.13"
+hmac = "0.13"
+sha2 = "0.11"
+subtle = "2.6"
+zeroize = { version = "1.9", features = ["derive"] }
+rand = "0.10"
+# Already in the tree through iroh. A packet codec, not a DNS server: the
+# zone logic is ours and a full server framework would be a large dependency
+# for answering A records from memory.
+simple-dns = "0.12"
+# The IANA top-level domain list, compiled in: one function, no
+# dependencies, no network. Used only to warn that a zone name shadows a
+# real public domain, never to refuse one.
+tld = "2.40"
+fs4 = { version = "1.1", features = ["sync"] }
+directories = "6.0"
+# The real system hostname, without a libc call of our own: this crate is
+# forbidden `unsafe` and will not make one.
+gethostname = "1.1"
+netwatch = "0.19.3"
+netdev = "0.46"
+mainline = "8.0"
+futures-lite = "2.6"
+tun = { version = "0.8", features = ["async"], optional = true }
+# Only for the `testing` harness, which a protocol crate's tests use too.
+tempfile = { version = "3.24", optional = true }
+tracing-subscriber = { version = "0.3", features = ["env-filter"], optional = true }
+
+# Linux-only interface provisioning. `rtnetlink` configures the interface in
+# process, so no `ip` invocation is ever needed; `caps` keeps CAP_NET_ADMIN
+# out of the effective set except during the moments it is used.
+[target.'cfg(target_os = "linux")'.dependencies]
+rtnetlink = { version = "0.23", optional = true }
+# Pure Rust D-Bus, no libdbus to link against. `tokio` rather than the
+# default reactor, because the agent brings its own.
+zbus = { version = "5.19", default-features = false, features = ["tokio"], optional = true }
+caps = { version = "0.5", optional = true }
+futures-util = { version = "0.3", default-features = false, optional = true }
+
+# Windows-only interface provisioning. The Wintun adapter itself is created
+# and driven through the `tun` dependency, whose safe wrapper is the only way
+# in — loading `wintun.dll` is an `unsafe` call this crate forbids. `netdev`
+# (already in the tree through `netwatch`) reads the interface table through a
+# safe wrapper, which is how the provisioner observes an adapter by name
+# before deciding what to change. Creating an adapter needs the process to be
+# elevated and `wintun.dll` to be reachable at run time.
+[target.'cfg(target_os = "windows")'.dependencies]
+windows-sys = { version = "0.61", features = [
+    "Win32_Security",
+    "Win32_Security_Authorization",
+    "Win32_Foundation",
+] }
+
+[dev-dependencies]
+# Its own tests use the harness it publishes.
+tsunagi-core = { path = ".", features = ["testing"] }
+tokio = { version = "1.53", features = ["rt", "rt-multi-thread", "sync", "time", "macros", "process", "test-util"] }
+tempfile.workspace = true
+tracing-subscriber.workspace = true
+
+[lints]
+workspace = true

crates/tsunagi-core/src/agent/events.rs

@@ -129,6 +129,19 @@ pub enum Event {
         /// Why it is not up.
         reason: String,
     },
+    /// The name this agent claimed was claimed first by another member.
+    ///
+    /// Nothing stops working: this agent keeps its address and every
+    /// connection. It just is not found by that name until it is renamed or
+    /// the member that holds the name lets it go.
+    HostnameTaken {
+        /// The network.
+        network: NetworkId,
+        /// The name that is taken.
+        name: String,
+        /// The member that holds it.
+        holder: EndpointId,
+    },
     /// An IP plugin reported an error. Never fatal.
     PluginError {
         /// The network the call was scoped to.

crates/tsunagi-core/src/agent/mod.rs

@@ -26,8 +26,8 @@ mod status;
 
 pub use events::Event;
 pub use status::{
-    AgentStatus, CandidateStatus, MemberStatus, NetworkMetrics, NetworkState, NetworkStatus,
+    AgentStatus, CandidateStatus, DataPath, HostnameConflict, MemberStatus, NetworkMetrics,
-    PeerStatus,
+    NetworkState, NetworkStatus, OverlayStatus, PeerStatus,
 };
 
 use std::collections::HashMap;
@@ -37,23 +37,29 @@ use iroh::{EndpointAddr, EndpointId};
 use tokio::sync::{RwLock, broadcast, mpsc, oneshot};
 use tokio::task::JoinHandle;
 
+use crate::BoxFuture;
 use crate::config::{AgentConfig, Limits};
 use crate::dataplane::transport::PacketTransport;
 use crate::dataplane::transport::iroh_link::{IrohTransport, TransportContext};
-use crate::dataplane::{PluginContext, PluginRequest};
+use crate::dataplane::{PacketSink, PluginContext, PluginRequest};
 use crate::error::{Error, Result};
 use crate::identity::{DeviceIdentity, NetworkId, NetworkKeys, NetworkName, NetworkSecret};
 use crate::net::EndpointAdapter;
+use crate::overlay::PacketCarrier;
 use crate::proto::handshake;
 use crate::proto::message::ControlMessage;
+use crate::state::Ipv4Range;
 use crate::storage::{CacheOutcome, Storage};
 
+use crate::task::{TASK_GRACE, wind_down};
 use network::{InboundSession, NetCommand, NetworkHandle, RuntimeParams};
 use shutdown::Shutdown;
 
 /// Summary of a configured network, whether or not it is running.
 #[derive(Debug, Clone)]
 pub struct ConfiguredNetwork {
+    /// Saved local broadcast participation.
+    pub broadcast: bool,
     /// Public network identifier.
     pub network_id: NetworkId,
     /// Network name.
@@ -87,6 +93,10 @@ struct Inner {
     accept_task: std::sync::Mutex<Option<JoinHandle<()>>>,
     plugin_task: std::sync::Mutex<Option<JoinHandle<()>>>,
     transport: std::sync::OnceLock<Arc<dyn PacketTransport>>,
+    /// Who holds which overlay address, across every network.
+    routes: Arc<crate::overlay::RoutingTable>,
+    /// The one interface, when the agent was given a way to make one.
+    interface: std::sync::OnceLock<Arc<crate::overlay::Interface>>,
 }
 
 impl Inner {
@@ -107,11 +117,69 @@ impl Inner {
     }
 }
 
+/// Routes an outbound packet to whichever protocol can carry it.
+///
+/// Holds a weak reference: the interface belongs to the agent, and a strong
+/// one here would keep the agent alive for as long as its own interface.
+struct Carrier(Weak<Inner>);
+
+impl std::fmt::Debug for Carrier {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("Carrier")
+    }
+}
+
+impl std::fmt::Debug for Sink {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("Sink")
+    }
+}
+
+impl PacketCarrier for Carrier {
+    fn carry(&self, route: crate::overlay::Route, packet: bytes::Bytes) -> bool {
+        let Some(inner) = self.0.upgrade() else {
+            return false;
+        };
+        // Offered to each protocol in turn. With one configured this is a
+        // single call; the shape is what allows several to be live at once,
+        // each carrying the peers it has a link to.
+        inner
+            .config
+            .plugins
+            .iter()
+            .any(|plugin| plugin.carry(route.network, route.peer, packet.clone()))
+    }
+}
+
+/// Writes a packet a protocol decrypted to the interface.
+struct Sink(Weak<Inner>);
+
+impl PacketSink for Sink {
+    fn deliver<'a>(
+        &'a self,
+        network: NetworkId,
+        peer: EndpointId,
+        packet: bytes::Bytes,
+    ) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let Some(inner) = self.0.upgrade() else {
+                return;
+            };
+            let Some(interface) = inner.interface.get() else {
+                return;
+            };
+            // The rejection is already counted on the interface; there is
+            // nothing useful to do with it here.
+            let _ = interface.deliver(network, peer, packet).await;
+        })
+    }
+}
+
 /// Answers the data plane transport's questions about the agent.
 ///
-/// Holds a weak reference on purpose: the transport lives inside the agent, so
+/// Holds a weak reference for the same reason as [`Carrier`]: the transport
-/// a strong one would be a cycle and the agent — with its open databases and
+/// lives inside the agent, so a strong one would be a cycle and the agent —
-/// its directory lock — would never be released.
+/// with its open databases and its directory lock — would never be released.
 #[derive(Debug)]
 struct TransportCtx(Weak<Inner>);
 
@@ -178,17 +246,54 @@ impl Agent {
             accept_task: std::sync::Mutex::new(None),
             plugin_task: std::sync::Mutex::new(None),
             transport: std::sync::OnceLock::new(),
+            routes: Arc::new(crate::overlay::RoutingTable::new()),
+            interface: std::sync::OnceLock::new(),
             config,
         });
 
+        // One interface for the agent, if it was given a way to make one.
+        // Failing to is reported and not fatal: the control plane works, and
+        // so do the protocols, they just have nowhere to put packets.
+        if let Some(factory) = inner.config.tun_factory.clone() {
+            let carrier = Arc::new(Carrier(Arc::downgrade(&inner))) as Arc<dyn PacketCarrier>;
+            match crate::overlay::Interface::start(
+                factory,
+                inner.config.interface_name.clone(),
+                inner.config.interface_mtu,
+                Arc::clone(&inner.routes),
+                carrier,
+            )
+            .await
+            {
+                Ok(interface) => {
+                    let _ = inner.interface.set(Arc::new(interface));
+                }
+                Err(err) => {
+                    let _ = inner.events.send(Event::PluginError {
+                        network: NetworkId::from_bytes([0u8; 32]),
+                        protocol: "overlay".into(),
+                        reason: err.to_string(),
+                    });
+                }
+            }
+        }
+
         // The data plane rides on iroh too, which is where it gets hole
         // punching and relay fallback from. It is a separate ALPN and a
         // separate connection, so the two planes stay independent.
-        let transport: Arc<dyn PacketTransport> = Arc::new(IrohTransport::new(
+        let iroh_transport: Arc<dyn PacketTransport> = Arc::new(IrohTransport::new(
             inner.adapter.clone(),
             Arc::clone(&inner.limits),
             Arc::new(TransportCtx(Arc::downgrade(&inner))) as Arc<dyn TransportContext>,
         ));
+        let transport: Arc<dyn PacketTransport> = if inner.config.custom_transports.is_empty() {
+            iroh_transport
+        } else {
+            Arc::new(crate::dataplane::MultiPacketTransport::new(
+                iroh_transport,
+                inner.config.custom_transports.clone(),
+            ))
+        };
         let _ = inner.transport.set(transport);
 
         if let CacheOutcome::Reset(reason) = inner.storage.cache_outcome().clone() {
@@ -205,7 +310,11 @@ impl Agent {
         // bound; overflow drops the request rather than stalling the plugin.
         if !inner.config.plugins.is_empty() {
             let (plugin_tx, plugin_rx) = mpsc::channel(64);
-            let context = PluginContext::new(plugin_tx, inner.identity.endpoint_id());
+            let sink = inner
+                .interface
+                .get()
+                .map(|_| Arc::new(Sink(Arc::downgrade(&inner))) as Arc<dyn PacketSink>);
+            let context = PluginContext::new(plugin_tx, inner.identity.endpoint_id(), sink);
             for plugin in &inner.config.plugins {
                 plugin.attach(context.clone());
             }
@@ -245,6 +354,138 @@ impl Agent {
         self.inner.adapter.loopback_addr()
     }
 
+    /// Delivers an inbound data plane link to the matching network runtime.
+    pub async fn install_inbound_link(
+        &self,
+        inbound: crate::dataplane::transport::InboundLink,
+    ) -> Result<()> {
+        let network = inbound.network;
+        let sender = {
+            let networks = self.inner.networks.read().await;
+            networks.get(&network).map(|handle| handle.commands.clone())
+        };
+        let Some(sender) = sender else {
+            return Err(Error::NetworkNotActive(network));
+        };
+        sender
+            .send(NetCommand::InboundLink(Box::new(inbound)))
+            .await
+            .map_err(|_| Error::NetworkNotActive(network))?;
+        Ok(())
+    }
+
+    /// Reserves the configured overlay range for a network, if it can.
+    ///
+    /// Done here, where activations are serialised, rather than inside the
+    /// runtime: whether a range is free depends on what the other networks
+    /// took, and deciding that in a task would make the answer depend on
+    /// which task ran first.
+    ///
+    /// `None` means this agent will not propose a range for that network and
+    /// waits to adopt whatever it settles on. One agent has one interface, so
+    /// proposing a range it could not route would be worse than having none:
+    /// the lowest author's range wins, and the collision would spread.
+    async fn reserve_range(&self, network: NetworkId) -> RangePlan {
+        let reserve = |wanted: Ipv4Range| {
+            let reservation = crate::overlay::NetworkRoutes {
+                exit: Default::default(),
+                broadcast: Default::default(),
+                range: Some(wanted),
+                local: None,
+                peers: Vec::new(),
+            };
+            self.inner.routes.set_network(network, reservation).is_ok()
+        };
+
+        // If this network already has signed records in storage, check if its
+        // members have already agreed on a range. If so, preserve and reserve
+        // THAT range rather than preemptively seizing the default configured range.
+        let stored_records = self
+            .inner
+            .storage
+            .signed_records(network)
+            .await
+            .unwrap_or_default();
+        let mut state_set = crate::state::StateSet::new();
+        state_set.merge_all(network, stored_records);
+        if let Some(agreed) = state_set.agreed_range() {
+            if reserve(agreed) {
+                return RangePlan {
+                    propose: if Some(agreed) == self.inner.config.overlay_ipv4_range {
+                        Some(agreed)
+                    } else {
+                        None
+                    },
+                    ..RangePlan::default()
+                };
+            } else {
+                tracing::warn!(
+                    %network,
+                    range = %agreed,
+                    "agreed range for this network overlaps another network already active on this interface"
+                );
+                return RangePlan {
+                    conflict: Some(agreed),
+                    ..RangePlan::default()
+                };
+            }
+        }
+
+        let Some(configured) = self.inner.config.overlay_ipv4_range else {
+            return RangePlan::default();
+        };
+
+        // The configured range first, so a network a device has always had
+        // keeps the addresses it has always had.
+        if reserve(configured) {
+            return RangePlan {
+                propose: Some(configured),
+                ..RangePlan::default()
+            };
+        }
+
+        // A second network on the same agent cannot have it — one agent,
+        // one interface — so it falls back to the range derived from its
+        // own id, which every one of its members derives identically
+        // without being told.
+        let derived = crate::state::derived_ipv4_range(network);
+        if reserve(derived) {
+            tracing::info!(
+                %network,
+                range = %derived,
+                "another network here holds the configured range; this one uses the \
+                 range derived from its id unless its members settled another"
+            );
+            return RangePlan {
+                fallback: Some(derived),
+                conflict: Some(configured),
+                ..RangePlan::default()
+            };
+        }
+
+        tracing::info!(%network, "not proposing a range for this network");
+        RangePlan {
+            conflict: Some(configured),
+            ..RangePlan::default()
+        }
+    }
+
+    /// The overlay interface this agent owns, when it has one.
+    ///
+    /// One agent, one interface, so this is not asked per network: a packet
+    /// on it may belong to any of them.
+    pub fn overlay(&self) -> Option<OverlayStatus> {
+        let interface = self.inner.interface.get()?;
+        Some(OverlayStatus {
+            interface: interface.name().to_string(),
+            on_host: interface.on_host(),
+            mtu: interface.mtu(),
+            addresses: interface.wanted_addresses(),
+            counters: interface.counters(),
+            broadcast_rules: interface.broadcast_rules(),
+        })
+    }
+
     /// The hostname announced to peers.
     pub fn hostname(&self) -> String {
         self.inner.read_hostname()
@@ -315,13 +556,47 @@ impl Agent {
         &self,
         name: &NetworkName,
         secret: &NetworkSecret,
+    ) -> Result<NetworkId> {
+        self.join_network_with_broadcast(name, secret, None).await
+    }
+
+    /// Joins with an explicit broadcast choice; absent preserves saved policy.
+    pub async fn join_network_with_broadcast(
+        &self,
+        name: &NetworkName,
+        secret: &NetworkSecret,
+        broadcast: Option<bool>,
     ) -> Result<NetworkId> {
         let keys = NetworkKeys::derive(name, secret);
         let network_id = keys.network_id();
+
+        // A network's identity is its name *and* its secret, so the same
+        // name with a different secret is a different network — and one that
+        // looks identical in anything that shows a name. Almost always a
+        // mistyped secret, so it is said out loud rather than left to be
+        // discovered as an empty network sitting beside a working one.
+        if let Ok(configured) = self.inner.storage.list_networks().await {
+            for other in configured {
+                if other.name == *name && other.network_id != network_id {
+                    tracing::warn!(
+                        "`{name}` is already configured with a different secret, as {}. \
+                         Joining with this one adds a second network under the same name; \
+                         they share nothing. Check the secret, or use \
+                         `tsng network secret` \
+                         to see which is which.",
+                        other.network_id
+                    );
+                }
+            }
+        }
+
         self.inner
             .storage
             .upsert_network(network_id, name.clone(), secret.clone(), true)
             .await?;
+        if let Some(enabled) = broadcast {
+            self.set_broadcast(network_id, enabled).await?;
+        }
         match self.activate_with_keys(keys).await {
             // Already a member of exactly this network space: nothing to do.
             Ok(()) | Err(Error::NetworkAlreadyActive(_)) => Ok(network_id),
@@ -329,6 +604,87 @@ impl Agent {
         }
     }
 
+    /// Offers this agent as an exit node in a network, or stops, now and
+    /// after restart. Off unless somebody turns it on.
+    pub async fn set_exit_offer(&self, network_id: NetworkId, enabled: bool) -> Result<()> {
+        self.inner
+            .storage
+            .set_exit_node(network_id, enabled)
+            .await?;
+        if self.is_active(network_id).await {
+            let (reply, receive) = oneshot::channel();
+            self.command(network_id, NetCommand::SetExitOffer { enabled, reply })
+                .await?;
+            receive.await.map_err(|_| Error::Stopped)?;
+        }
+        Ok(())
+    }
+
+    /// Sends this device's internet traffic through a member of a network,
+    /// or back the ordinary way with `None`.
+    ///
+    /// A device has one default route, so choosing in one network clears the
+    /// choice in the others. The member has to be connected and offering
+    /// to be an exit node. Remembered across restarts.
+    pub async fn set_exit_via(&self, network_id: NetworkId, via: Option<EndpointId>) -> Result<()> {
+        let active = self.is_active(network_id).await;
+        if via.is_some() && !active {
+            return Err(Error::NetworkNotActive(network_id));
+        }
+        if active {
+            let (reply, receive) = oneshot::channel();
+            self.command(network_id, NetCommand::SetExitVia { via, reply })
+                .await?;
+            receive
+                .await
+                .map_err(|_| Error::Stopped)?
+                .map_err(|reason| Error::Rejected { reason })?;
+        }
+        self.inner
+            .storage
+            .set_exit_via(network_id, via.map(|peer| *peer.as_bytes()))
+            .await?;
+        if via.is_some() {
+            // One default route: whatever the other networks were doing, they
+            // are not any more.
+            let others: Vec<NetworkId> = self
+                .inner
+                .networks
+                .read()
+                .await
+                .keys()
+                .copied()
+                .filter(|other| *other != network_id)
+                .collect();
+            for other in others {
+                let (reply, receive) = oneshot::channel();
+                if self
+                    .command(other, NetCommand::SetExitVia { via: None, reply })
+                    .await
+                    .is_ok()
+                {
+                    let _ = receive.await;
+                }
+            }
+        }
+        Ok(())
+    }
+
+    /// Changes a network's local broadcast policy now and after restart.
+    pub async fn set_broadcast(&self, network_id: NetworkId, enabled: bool) -> Result<()> {
+        self.inner
+            .storage
+            .set_broadcast(network_id, enabled)
+            .await?;
+        if self.is_active(network_id).await {
+            let (reply, receive) = oneshot::channel();
+            self.command(network_id, NetCommand::SetBroadcast { enabled, reply })
+                .await?;
+            receive.await.map_err(|_| Error::Stopped)?;
+        }
+        Ok(())
+    }
+
     /// Activates a configured network that is currently inactive.
     ///
     /// Fails with [`Error::NetworkAlreadyActive`] if it is already running.
@@ -357,20 +713,59 @@ impl Agent {
             return Err(Error::NetworkAlreadyActive(network_id));
         }
 
+        let range = self.reserve_range(network_id).await;
+        let mut backends = Vec::new();
+        if let Some(discovery) = &self.inner.config.discovery {
+            backends.push(discovery.clone());
+        }
+        if let Some(dht) = &self.inner.config.dht {
+            backends.push(dht.for_network(&keys));
+        }
+        let discovery = if backends.is_empty() {
+            None
+        } else {
+            Some(
+                Arc::new(crate::discovery::CompositeDiscovery::new(backends))
+                    as Arc<dyn crate::discovery::NetworkDiscovery>,
+            )
+        };
+        let stored = self
+            .inner
+            .storage
+            .list_networks()
+            .await?
+            .into_iter()
+            .find(|stored| stored.network_id == network_id);
+        let broadcast = stored.as_ref().is_none_or(|stored| stored.broadcast);
+        let exit_offer = stored.as_ref().is_some_and(|stored| stored.exit_node);
+        let exit_via = stored
+            .as_ref()
+            .and_then(|stored| stored.exit_via)
+            .and_then(|bytes| EndpointId::from_bytes(&bytes).ok());
         let handle = network::spawn(RuntimeParams {
+            broadcast,
+            exit_offer,
+            exit_via,
             keys,
             adapter: self.inner.adapter.clone(),
             storage: self.inner.storage.clone(),
             events: self.inner.events.clone(),
             limits: Arc::clone(&self.inner.limits),
             reconnect: self.inner.config.reconnect.clone(),
-            discovery: self.inner.config.discovery.clone(),
+            discovery,
             discovery_interval: self.inner.config.discovery_interval,
+            discovery_policy: self.inner.config.discovery_policy.clone(),
             plugins: self.inner.config.plugins.clone(),
+            routes: Arc::clone(&self.inner.routes),
+            interface: self.inner.interface.get().cloned(),
             hostname: self.inner.read_hostname(),
             transport: self.inner.transport.get().cloned(),
             device_secret: self.inner.identity.signing_key(),
-            ipv4_range: self.inner.config.overlay_ipv4_range,
+            #[cfg(feature = "testing")]
+            unreachable_data_peers: Arc::clone(&self.inner.config.unreachable_data_peers),
+            ipv4_range: range.propose,
+            ipv4_fallback: range.fallback,
+            range_conflict: range.conflict,
         });
         networks.insert(network_id, handle);
         drop(networks);
@@ -401,17 +796,83 @@ impl Agent {
         for plugin in &self.inner.config.plugins {
             plugin.on_network_deactivated(network_id);
         }
+        // The network's addresses come off the interface; the interface
+        // itself stays, because the agent owns it and other networks may
+        // still be using it.
+        self.inner.routes.remove_network(network_id);
+        if let Some(interface) = self.inner.interface.get() {
+            interface
+                .set_underlay(network_id, Vec::new(), Vec::new())
+                .await;
+        }
+        if let Some(interface) = self.inner.interface.get()
+            && let Err(err) = interface.sync_addresses().await
+        {
+            tracing::warn!(%err, "cannot update the overlay addresses");
+        }
+        if let Some(interface) = self.inner.interface.get() {
+            interface.sync_broadcast_rules().await;
+            interface.sync_exit_rules().await;
+        }
         self.inner.storage.set_auto_start(network_id, false).await?;
         Ok(())
     }
 
     /// Deactivates a network if it is running and removes it from the state
     /// store together with its cached hints.
+    ///
+    /// Local and silent: nobody is told. [`Agent::leave_network`] is the one
+    /// that says goodbye first, and is what a person means by leaving.
     pub async fn forget_network(&self, network_id: NetworkId) -> Result<()> {
         if self.is_active(network_id).await {
             self.deactivate_network(network_id).await?;
         }
-        self.inner.storage.remove_network(network_id).await
+        self.inner.storage.remove_network(network_id).await?;
+        // Whatever a protocol kept for this network goes with it. The
+        // agent cannot know what that is, only that there is no longer any
+        // reason to hold it.
+        for plugin in &self.inner.config.plugins {
+            plugin.on_network_forgotten(network_id);
+        }
+        Ok(())
+    }
+
+    /// Leaves a network: gives up what was claimed, then forgets it.
+    ///
+    /// The order matters and cannot be improved on. A signed `Release` goes
+    /// out first, while there are still sessions to carry it, so the address
+    /// and name this agent held are freed for somebody else rather than
+    /// staying reserved to a member that has gone. Only then is the network
+    /// deactivated and removed.
+    ///
+    /// Reaching every member is not on offer and never could be: a member
+    /// that is away hears the tombstone from the ones that were here, the
+    /// same way it hears everything else. Leaving with nobody connected
+    /// tells nobody, and says so in the outcome rather than pretending.
+    ///
+    /// The author's version counter is deliberately **kept**. Rejoining the
+    /// same network with the same device key must continue from a higher
+    /// version than the release, or every replica would treat the new claim
+    /// as stale and ignore it.
+    pub async fn leave_network(&self, network_id: NetworkId) -> Result<LeaveOutcome> {
+        let mut outcome = LeaveOutcome::default();
+        if self.is_active(network_id).await {
+            let (reply_tx, reply_rx) = oneshot::channel();
+            self.command(network_id, NetCommand::Release { reply: reply_tx })
+                .await?;
+            if let Ok(peers) = reply_rx.await {
+                outcome.announced = true;
+                outcome.peers_told = peers;
+            }
+            // The tombstone is queued on each session, not yet written to
+            // the wire. A short pause is the difference between peers
+            // hearing it now and hearing it from somebody else much later.
+            if outcome.peers_told > 0 {
+                tokio::time::sleep(RELEASE_FLUSH).await;
+            }
+        }
+        self.forget_network(network_id).await?;
+        Ok(outcome)
     }
 
     /// Whether a network is currently running.
@@ -429,6 +890,7 @@ impl Agent {
             .await?
             .into_iter()
             .map(|stored| ConfiguredNetwork {
+                broadcast: stored.broadcast,
                 network_id: stored.network_id,
                 name: stored.name,
                 auto_start: stored.auto_start,
@@ -503,16 +965,27 @@ impl Agent {
             }
             let keys = NetworkKeys::derive(&stored.name, &stored.secret);
             networks.push(NetworkStatus {
+                broadcast: stored.broadcast,
+                exit_offer: stored.exit_node,
+                exit_via: stored
+                    .exit_via
+                    .and_then(|bytes| EndpointId::from_bytes(&bytes).ok()),
+                exit_via_online: false,
+                exit_rules: Default::default(),
                 descriptor: keys.descriptor(),
                 name: stored.name,
                 network_id: stored.network_id,
                 state: NetworkState::Inactive,
                 peers: Vec::new(),
                 candidates: Vec::new(),
-                // An inactive network has no runtime to ask; the roster comes
+                // An inactive network has no runtime to ask; the roster and
-                // from one. Empty, not invented.
+                // the range come from one. Empty, not invented.
                 members: Vec::new(),
+                hostname_conflict: None,
+                range: None,
+                range_conflict: None,
                 metrics: NetworkMetrics::default(),
+                relay: Default::default(),
             });
         }
 
@@ -537,7 +1010,8 @@ impl Agent {
         self.command(network_id, NetCommand::Reannounce).await
     }
 
-    /// Asks one network to re-run discovery and re-evaluate dials right now.
+    /// Asks one network to re-evaluate known peers and dials right now.
+    /// External lookup still follows its connected/isolation policy.
     ///
     /// Call this when the host's network environment changed. Platform wake-up
     /// notifications can be wired to it later.
@@ -545,7 +1019,7 @@ impl Agent {
         self.command(network_id, NetCommand::Recheck).await
     }
 
-    /// Asks every running network to re-run discovery right now.
+    /// Asks every running network to re-evaluate known peers right now.
     pub async fn recheck(&self) {
         let senders: Vec<mpsc::Sender<NetCommand>> = self
             .inner
@@ -574,15 +1048,35 @@ impl Agent {
             handle.stop().await;
         }
 
-        self.inner.adapter.close().await;
+        if let Some(dht) = &self.inner.config.dht {
+            dht.shutdown().await;
+        }
 
-        for handle in [&self.inner.accept_task, &self.inner.plugin_task] {
+        // iroh's own close waits for peers to acknowledge; a peer that has
+        // gone silent must not decide how long that takes.
+        if tokio::time::timeout(TASK_GRACE, self.inner.adapter.close())
+            .await
+            .is_err()
+        {
+            tracing::warn!("the endpoint did not close in time; abandoning it");
+        }
+
+        for (what, handle) in [
+            ("inbound connections", &self.inner.accept_task),
+            ("plugin requests", &self.inner.plugin_task),
+        ] {
             let task = handle.lock().ok().and_then(|mut guard| guard.take());
             if let Some(task) = task {
-                let _ = task.await;
+                wind_down(task, TASK_GRACE, what).await;
             }
         }
 
+        // The interface goes with the agent that owns it. Before the
+        // plugins, so nothing is still trying to write to it.
+        if let Some(interface) = self.inner.interface.get() {
+            interface.remove().await;
+        }
+
         // Plugins remove whatever system objects they created. A plugin that
         // misbehaves here must not hold up the agent, so this is bounded.
         for plugin in &self.inner.config.plugins {
@@ -613,6 +1107,41 @@ impl Agent {
     }
 }
 
+/// What a network may propose as its overlay range, and what it cannot have.
+#[derive(Debug, Clone, Copy, Default)]
+struct RangePlan {
+    /// Proposed straight away: the configured range, when it is free here.
+    propose: Option<Ipv4Range>,
+    /// Proposed after a moment, when the configured range is another
+    /// network's: the range derived from this network's own id.
+    fallback: Option<Ipv4Range>,
+    /// The configured range, when this network cannot have it.
+    conflict: Option<Ipv4Range>,
+}
+
+/// How long a release is given to reach the sessions it was queued on.
+///
+/// Short: it is one small message on a connection that is already open, and
+/// the alternative to waiting at all is tearing the sessions down underneath
+/// it.
+const RELEASE_FLUSH: std::time::Duration = std::time::Duration::from_millis(300);
+
+/// What happened when an agent left a network.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct LeaveOutcome {
+    /// Whether a signed release was published at all.
+    ///
+    /// `false` when the network was not running: there was nothing to
+    /// publish it from, so this was a local removal only.
+    pub announced: bool,
+    /// How many connected peers it was sent to.
+    ///
+    /// Zero with `announced` true means the tombstone is in this agent's
+    /// own state and nowhere else, and it is leaving with it — so nobody
+    /// will learn of it.
+    pub peers_told: usize,
+}
+
 /// How long each plugin gets to tear itself down during agent shutdown.
 const PLUGIN_SHUTDOWN_GRACE: std::time::Duration = std::time::Duration::from_secs(10);
 
@@ -680,6 +1209,26 @@ async fn plugin_request_loop(weak: Weak<Inner>, mut requests: mpsc::Receiver<Plu
                     }
                 }
             }
+            PluginRequest::ProtocolFailure {
+                network,
+                peer,
+                protocol,
+                reason,
+            } => {
+                let sender = {
+                    let networks = inner.networks.read().await;
+                    networks.get(&network).map(|handle| handle.commands.clone())
+                };
+                if let Some(sender) = sender {
+                    let _ = sender
+                        .send(NetCommand::ProtocolFailure {
+                            peer,
+                            protocol,
+                            reason,
+                        })
+                        .await;
+                }
+            }
         }
     }
 }
@@ -825,8 +1374,23 @@ async fn handle_inbound_data(inner: Arc<Inner>, conn: iroh::endpoint::Connection
         conn.close(5u32.into(), b"data plane not ready");
         return;
     };
-    // Downcasting is avoided by keeping the accept side on the concrete type.
+    let iroh_transport =
-    let Some(iroh_transport) = transport.as_ref().as_any().downcast_ref::<IrohTransport>() else {
+        if let Some(iroh) = transport.as_ref().as_any().downcast_ref::<IrohTransport>() {
+            Some(iroh)
+        } else if let Some(multi) = transport
+            .as_ref()
+            .as_any()
+            .downcast_ref::<crate::dataplane::MultiPacketTransport>()
+        {
+            multi
+                .default_transport()
+                .as_ref()
+                .as_any()
+                .downcast_ref::<IrohTransport>()
+        } else {
+            None
+        };
+    let Some(iroh_transport) = iroh_transport else {
         conn.close(5u32.into(), b"unsupported data transport");
         return;
     };

crates/tsunagi-core/src/agent/session.rs

@@ -73,17 +73,21 @@ pub(crate) struct Session {
     pub(crate) outbound: mpsc::Sender<Vec<u8>>,
     pub(crate) hostname: Option<String>,
     pub(crate) capabilities: Vec<PluginCapability>,
+    pub(crate) broadcast: bool,
+    /// The peer offers itself as an exit node in this network.
+    pub(crate) exit_node: bool,
     pub(crate) messages_sent: u64,
     pub(crate) messages_received: u64,
     pub(crate) bytes_sent: u64,
     pub(crate) bytes_received: u64,
     reader: JoinHandle<()>,
     writer: JoinHandle<()>,
+    path_monitor: JoinHandle<()>,
     shutdown: Shutdown,
 }
 
 impl Session {
-    /// Signals both tasks to stop and waits for them, with a bounded grace
+    /// Signals all tasks to stop and waits for them, with a bounded grace
     /// period.
     ///
     /// A peer that stops reading must not be able to hold up shutdown, so the
@@ -93,6 +97,7 @@ impl Session {
             conn,
             reader,
             writer,
+            path_monitor,
             shutdown,
             ..
         } = self;
@@ -102,23 +107,27 @@ impl Session {
 
         let reader_abort = reader.abort_handle();
         let writer_abort = writer.abort_handle();
+        let path_abort = path_monitor.abort_handle();
         let joined = tokio::time::timeout(STOP_GRACE, async move {
             let _ = reader.await;
             let _ = writer.await;
+            let _ = path_monitor.await;
         })
         .await;
         if joined.is_err() {
             tracing::debug!("session tasks did not wind down in time; aborting them");
             reader_abort.abort();
             writer_abort.abort();
+            path_abort.abort();
         }
     }
 
-    /// Aborts both tasks without waiting. Used on replacement.
+    /// Aborts all tasks without waiting. Used on replacement.
     pub(crate) fn abort(&self) {
         self.shutdown.trigger();
         self.reader.abort();
         self.writer.abort();
+        self.path_monitor.abort();
     }
 }
 
@@ -154,6 +163,52 @@ pub(crate) fn spawn(
         parent_shutdown.clone(),
     ));
 
+    let path_conn = conn.clone();
+    let path_shutdown = shutdown.clone();
+    let path_parent = parent_shutdown.clone();
+    let path_monitor = tokio::spawn(async move {
+        use tokio_stream::StreamExt;
+        let mut events = path_conn.path_events();
+        loop {
+            tokio::select! {
+                biased;
+                _ = path_shutdown.wait() => break,
+                _ = path_parent.wait() => break,
+                ev = events.next() => {
+                    match ev {
+                        Some(iroh::endpoint::PathEvent::Selected { remote_addr, local_addr, .. }) => {
+                            tracing::debug!(
+                                peer = %peer.fmt_short(),
+                                remote = %remote_addr,
+                                local = ?local_addr,
+                                "Connection path selected (direct / relay / hole punch switch)"
+                            );
+                        }
+                        Some(iroh::endpoint::PathEvent::Opened { remote_addr, .. }) => {
+                            tracing::debug!(
+                                peer = %peer.fmt_short(),
+                                remote = %remote_addr,
+                                "Network path opened to peer"
+                            );
+                        }
+                        Some(iroh::endpoint::PathEvent::Closed { remote_addr, .. }) => {
+                            tracing::debug!(
+                                peer = %peer.fmt_short(),
+                                remote = %remote_addr,
+                                "Network path closed for peer"
+                            );
+                        }
+                        Some(iroh::endpoint::PathEvent::Lagged { missed, .. }) => {
+                            tracing::trace!(peer = %peer.fmt_short(), missed, "Path events lagged");
+                        }
+                        Some(_) => {}
+                        None => break,
+                    }
+                }
+            }
+        }
+    });
+
     let reader = tokio::spawn(reader_task(
         id,
         network_id,
@@ -174,12 +229,15 @@ pub(crate) fn spawn(
         outbound: outbound_tx,
         hostname: None,
         capabilities: Vec::new(),
+        broadcast: false,
+        exit_node: false,
         messages_sent: 0,
         messages_received: 0,
         bytes_sent: 0,
         bytes_received: 0,
         reader,
         writer,
+        path_monitor,
         shutdown,
     }
 }

crates/tsunagi-core/src/agent/shutdown.rs

@@ -51,3 +51,25 @@ impl Shutdown {
         // The sender is gone, which for our purposes means "stop".
     }
 }
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[tokio::test]
+    async fn waiting_resolves_once_triggered_and_stays_resolved() {
+        let shutdown = Shutdown::new();
+        assert!(!shutdown.is_triggered());
+        let waiter = shutdown.clone();
+        let waiting = tokio::spawn(async move { waiter.wait().await });
+        shutdown.trigger();
+        waiting.await.unwrap();
+        assert!(shutdown.is_triggered());
+        // Triggering twice is not an error, and a token cloned afterwards
+        // sees the trigger that already happened.
+        shutdown.trigger();
+        shutdown.clone().wait().await;
+    }
+}

crates/tsunagi-core/src/agent/status.rs

@@ -36,6 +36,66 @@ pub struct CandidateStatus {
     pub consecutive_failures: u32,
 }
 
+/// The overlay interface, as the agent sees it.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct OverlayStatus {
+    /// The interface name the operating system gave.
+    pub interface: String,
+    /// Whether that interface exists on the host.
+    ///
+    /// `false` with an in-memory device: the name is real to the agent and
+    /// to nothing else. Reported, because everything that would configure
+    /// the operating system for this interface must not when it is this.
+    pub on_host: bool,
+    /// Its MTU.
+    pub mtu: u32,
+    /// The addresses it should be carrying.
+    pub addresses: Vec<crate::overlay::Cidr>,
+    /// What has happened on it.
+    pub counters: crate::overlay::Counters,
+    /// The route and firewall allowance installed for LAN broadcast, when
+    /// some network has broadcast on and the interface is on the host.
+    /// `None` is "none installed", not "unknown".
+    pub broadcast_rules: Option<crate::overlay::BroadcastRulesReport>,
+}
+
+/// How packets reach a member right now.
+///
+/// Being reachable is one thing and being reachable well is another, and
+/// the second is never a reason to call a member away: a member that can only
+/// be reached through somebody else is online, over a longer path.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum DataPath {
+    /// Nothing carries packets to it yet. The member is connected for
+    /// control only, which is the state of a data path being set up.
+    None,
+    /// A direct link, of the named kind of transport.
+    Direct {
+        /// The kind of transport carrying it, e.g. `wg`.
+        transport: String,
+    },
+    /// Through other members.
+    Relayed {
+        /// Number of links on the way, at least two.
+        hops: u8,
+        /// The member the first of them goes to.
+        via: EndpointId,
+    },
+}
+
+/// A name that a member claimed and somebody else holds.
+///
+/// A name belongs to whoever claimed it first. The member that claimed it
+/// later is still a member and still reachable by its address; it is just
+/// not found by that name, on any member of the network.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct HostnameConflict {
+    /// The name that is taken.
+    pub name: String,
+    /// The member that holds it.
+    pub holder: EndpointId,
+}
+
 /// A member the signed state knows about, connected or not.
 ///
 /// This is the durable roster: it comes from signed records, so a member that
@@ -57,11 +117,29 @@ pub struct MemberStatus {
     /// Signed, so it is still known while the member is away — which is what
     /// lets an absent member be named rather than shown as a bare id.
     pub hostname: Option<String>,
+    /// The last name this agent saw it announce, whether or not it is here.
+    ///
+    /// Not signed, so it is a convenience for naming a member that is away
+    /// and never a claim: the zone is served from `hostname`.
+    pub last_hostname: Option<String>,
+    /// The name it claimed, when somebody else claimed it first.
+    pub hostname_conflict: Option<HostnameConflict>,
+    /// Whether this agent can reach it now, by any path: a live control
+    /// session, or a data path through other members. The one answer to "is
+    /// it online" for every client, so that none of them has to work it out
+    /// from a different part of the picture. This agent counts as online.
+    pub online: bool,
+    /// How packets reach it. [`DataPath::None`] while it is offline.
+    pub data_path: DataPath,
 }
 
 /// Status of one authenticated session.
 #[derive(Debug, Clone)]
 pub struct PeerStatus {
+    /// This authenticated peer accepts network broadcasts.
+    pub broadcast: bool,
+    /// This authenticated peer offers itself as an exit node.
+    pub exit_node: bool,
     /// Authenticated endpoint id.
     pub endpoint_id: EndpointId,
     /// Which side this agent played in the handshake.
@@ -70,6 +148,12 @@ pub struct PeerStatus {
     ///
     /// A mutable binding, not an identity.
     pub hostname: Option<String>,
+    /// Protocols agreed with this peer: both sides have them at the same
+    /// version.
+    ///
+    /// Empty means no data plane with that peer, which leaves the control
+    /// plane working — messages and state still flow.
+    pub protocols: Vec<String>,
     /// Capabilities the peer announced. Payloads stay opaque.
     pub capabilities: Vec<PluginCapability>,
     /// How long the session has been up.
@@ -90,6 +174,8 @@ pub struct PeerStatus {
     pub control_bytes_sent: u64,
     /// Control payload bytes read from this session.
     pub control_bytes_received: u64,
+    /// How packets reach this peer right now.
+    pub data_path: DataPath,
 }
 
 /// Counters scoped to one logical network.
@@ -126,6 +212,16 @@ pub struct NetworkMetrics {
 /// Status of one network.
 #[derive(Debug, Clone)]
 pub struct NetworkStatus {
+    /// Local broadcast participation in this network.
+    pub broadcast: bool,
+    /// This agent offers itself as an exit node in this network.
+    pub exit_offer: bool,
+    /// The member this device sends all its internet traffic through.
+    pub exit_via: Option<EndpointId>,
+    /// Whether that member is connected and still offering to be one.
+    pub exit_via_online: bool,
+    /// What the host rules for the exit settings are doing.
+    pub exit_rules: crate::overlay::ExitRulesReport,
     /// Immutable deterministic description of the network space.
     pub descriptor: NetworkDescriptor,
     /// Network name, for convenience.
@@ -140,8 +236,23 @@ pub struct NetworkStatus {
     pub candidates: Vec<CandidateStatus>,
     /// Members the signed state knows about, whether connected or not.
     pub members: Vec<MemberStatus>,
+    /// This agent's own name, when another member claimed it first and so
+    /// holds it. This agent keeps working by address; nobody finds it by
+    /// this name until it is renamed or the other member lets it go.
+    pub hostname_conflict: Option<HostnameConflict>,
+    /// The overlay range this network uses, once it has one.
+    pub range: Option<crate::state::Ipv4Range>,
+    /// The range it could not have, because another network on this agent
+    /// already holds it.
+    ///
+    /// One agent has one interface, so an address belongs to one network.
+    /// This network waits to adopt whatever its members settle on instead of
+    /// proposing something it could not route.
+    pub range_conflict: Option<crate::state::Ipv4Range>,
     /// Per-network counters.
     pub metrics: NetworkMetrics,
+    /// What has gone through a peer in the middle, in both directions.
+    pub relay: crate::dataplane::relay::RelayCounters,
 }
 
 impl NetworkStatus {

crates/tsunagi-core/src/config.rs

@@ -11,12 +11,42 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use crate::dataplane::SharedPlugin;
-use crate::discovery::NetworkDiscovery;
+use crate::discovery::{MainlineDiscovery, NetworkDiscovery};
-use crate::error::{Error, Result};
+use crate::error::Result;
 
 /// Qualifier/organisation/application triple used for platform directories.
+///
+/// Unused on macOS, where the default paths are fixed system locations rather
+/// than per-user directories derived from this.
+#[cfg_attr(target_os = "macos", allow(dead_code))]
 const APP_NAME: &str = "tsunagi";
 
+/// Rendezvous slots per network; this bounds a discovery sample, not membership.
+pub const DHT_SLOTS: u8 = 16;
+/// BEP44 permits 1000 bencoded bytes; a 996-byte string has a four-byte prefix.
+pub const DHT_MAX_VALUE: usize = 996;
+/// Maximum direct addresses in one rendezvous record.
+pub const DHT_MAX_ADDRS: usize = 8;
+/// Maximum relay URL bytes in a rendezvous record.
+pub const DHT_MAX_RELAY_LEN: usize = 512;
+/// Application freshness, independent of storage nodes' own expiration.
+pub const DHT_RECORD_TTL: Duration = Duration::from_secs(900);
+/// Clock skew tolerated when reading a recently published record.
+pub const DHT_CLOCK_SKEW: Duration = Duration::from_secs(120);
+
+/// Maximum reassembled data-transport payload, independent of the path MTU.
+pub const MAX_DATA_DATAGRAM: usize = 64 * 1024;
+/// Maximum incomplete datagrams held by one authenticated data link.
+pub const MAX_DATA_ASSEMBLIES: usize = 64;
+/// Maximum allocated payload bytes for incomplete datagrams on one data link.
+pub const MAX_DATA_REASSEMBLY_BYTES: usize = 256 * 1024;
+/// Maximum fragments of one logical datagram, including after a path MTU change.
+pub const MAX_DATA_FRAGMENTS: usize = 128;
+/// Time allowed to assemble a datagram; loss never stalls later datagrams.
+pub const DATA_REASSEMBLY_TTL: Duration = Duration::from_secs(5);
+/// Completed or discarded packet IDs retained to ignore late duplicate fragments.
+pub const DATA_RECENT_IDS: usize = 128;
+
 /// Where the two stores live.
 ///
 /// The mandatory state and the disposable cache are separate both logically and
@@ -48,17 +78,34 @@ impl StoragePaths {
         }
     }
 
-    /// The per-user platform directories.
+    /// The default platform directories.
     ///
-    /// A future system service can supply its own paths instead.
+    /// Per user on Linux and Windows. On macOS the agent runs as a root
+    /// LaunchDaemon (creating a utun interface needs root, and `$HOME` under
+    /// `sudo` points at the wrong place), so the stores live in the system
+    /// locations instead — `/var/db/tsunagi` for the mandatory state and a
+    /// disposable cache beside it. A caller that wants somewhere else passes
+    /// explicit paths to [`StoragePaths::new`] or [`StoragePaths::under`].
     pub fn user_default() -> Result<Self> {
-        let dirs = directories::ProjectDirs::from("", "", APP_NAME).ok_or_else(|| {
+        #[cfg(target_os = "macos")]
-            Error::Storage("no valid home directory for platform config paths".into())
+        {
-        })?;
+            Ok(Self {
-        Ok(Self {
+                state_dir: PathBuf::from("/var/db/tsunagi"),
-            state_dir: dirs.data_dir().to_path_buf(),
+                cache_dir: PathBuf::from("/var/db/tsunagi/cache"),
-            cache_dir: dirs.cache_dir().to_path_buf(),
+            })
-        })
+        }
+        #[cfg(not(target_os = "macos"))]
+        {
+            let dirs = directories::ProjectDirs::from("", "", APP_NAME).ok_or_else(|| {
+                crate::error::Error::Storage(
+                    "no valid home directory for platform config paths".into(),
+                )
+            })?;
+            Ok(Self {
+                state_dir: dirs.data_dir().to_path_buf(),
+                cache_dir: dirs.cache_dir().to_path_buf(),
+            })
+        }
     }
 
     /// Path of the mandatory state database.
@@ -137,6 +184,8 @@ pub struct Limits {
     pub write_timeout: Duration,
     /// Maximum simultaneous outbound dials per network.
     pub max_concurrent_dials: usize,
+    /// Maximum unverified discovery candidates retained per network.
+    pub max_discovery_candidates: usize,
     /// Maximum simultaneous authenticated sessions per network.
     pub max_sessions_per_network: usize,
     /// Maximum simultaneous inbound connections being handshaken.
@@ -163,6 +212,7 @@ impl Default for Limits {
             dial_timeout: Duration::from_secs(10),
             write_timeout: Duration::from_secs(30),
             max_concurrent_dials: 8,
+            max_discovery_candidates: 16,
             max_sessions_per_network: 64,
             max_inbound_handshakes: 32,
             session_send_queue: 64,
@@ -212,6 +262,33 @@ impl ReconnectPolicy {
     }
 }
 
+/// Scheduling of candidate lookup and independent self-publication.
+#[derive(Debug, Clone)]
+pub struct DiscoveryPolicy {
+    /// Mean interval between successful self-publications (20% jitter).
+    pub publish_interval: Duration,
+    /// Initial delay between unsuccessful bootstrap lookups.
+    pub lookup_interval: Duration,
+    /// Maximum delay between bootstrap lookups.
+    pub max_lookup_interval: Duration,
+    /// Continuous isolation before a previously connected network searches again.
+    pub reconnect_delay: Duration,
+    /// Deadline for one backend operation.
+    pub request_timeout: Duration,
+}
+
+impl Default for DiscoveryPolicy {
+    fn default() -> Self {
+        Self {
+            publish_interval: Duration::from_secs(300),
+            lookup_interval: Duration::from_secs(5),
+            max_lookup_interval: Duration::from_secs(30),
+            reconnect_delay: Duration::from_secs(60),
+            request_timeout: Duration::from_secs(60),
+        }
+    }
+}
+
 /// Everything needed to start an [`crate::Agent`].
 #[derive(Clone)]
 pub struct AgentConfig {
@@ -223,13 +300,24 @@ pub struct AgentConfig {
     pub bind_addrs: Vec<SocketAddr>,
     /// How much external connectivity machinery the endpoint may use.
     pub transport: TransportPolicy,
+    /// Real QUIC transport settings for constrained-path integration tests.
+    #[cfg(feature = "testing")]
+    pub test_quic_transport: Option<iroh::endpoint::QuicTransportConfig>,
+    /// Peers with no direct data path, for tests. See
+    /// [`AgentConfig::with_unreachable_data_peers`].
+    #[cfg(feature = "testing")]
+    pub unreachable_data_peers: Arc<std::sync::Mutex<std::collections::HashSet<iroh::EndpointId>>>,
     /// Hostname announced to peers. `None` keeps whatever the state store holds,
     /// falling back to the OS hostname and finally to a short endpoint id.
     pub hostname: Option<String>,
-    /// Discovery backend. `None` disables discovery-driven dialling; static
+    /// Additional discovery backend, composed with Mainline when it is enabled.
-    /// bootstrap candidates still work.
     pub discovery: Option<Arc<dyn NetworkDiscovery>>,
-    /// How often each active network re-runs discovery and re-evaluates dials.
+    /// Optional Mainline client, shared by this agent's networks. The CLI enables
+    /// it by default except in local-only mode; library callers opt in explicitly.
+    pub dht: Option<MainlineDiscovery>,
+    /// Lookup, publication and recovery scheduling.
+    pub discovery_policy: DiscoveryPolicy,
+    /// How often each active network maintains known peers and observes address changes.
     pub discovery_interval: Duration,
     /// Bounds applied to network input.
     pub limits: Limits,
@@ -237,6 +325,20 @@ pub struct AgentConfig {
     pub reconnect: ReconnectPolicy,
     /// IP plugins whose capabilities are announced and dispatched.
     pub plugins: Vec<SharedPlugin>,
+    /// Custom packet transports registered per protocol.
+    pub custom_transports:
+        std::collections::HashMap<String, Arc<dyn crate::dataplane::PacketTransport>>,
+    /// Where the overlay interface comes from, when the agent should have one.
+    ///
+    /// One agent, one interface: it belongs here rather than to a protocol,
+    /// because every protocol carries traffic for the same addresses on it.
+    /// `None` means no interface — the protocols still run and their packets
+    /// are discarded.
+    pub tun_factory: Option<Arc<dyn crate::overlay::TunFactory>>,
+    /// The name to ask the operating system for.
+    pub interface_name: String,
+    /// The interface MTU.
+    pub interface_mtu: u32,
     /// The IPv4 overlay range this agent proposes.
     ///
     /// Addresses are allocated from it and recorded in signed state, so a
@@ -253,12 +355,24 @@ impl AgentConfig {
             paths,
             bind_addrs: Vec::new(),
             transport: TransportPolicy::default(),
+            #[cfg(feature = "testing")]
+            test_quic_transport: None,
+            #[cfg(feature = "testing")]
+            unreachable_data_peers: Arc::new(std::sync::Mutex::new(
+                std::collections::HashSet::new(),
+            )),
             hostname: None,
             discovery: None,
+            dht: None,
+            discovery_policy: DiscoveryPolicy::default(),
             discovery_interval: Duration::from_secs(5),
             limits: Limits::default(),
             reconnect: ReconnectPolicy::default(),
             plugins: Vec::new(),
+            custom_transports: std::collections::HashMap::new(),
+            tun_factory: None,
+            interface_name: "tsun0".to_string(),
+            interface_mtu: 1280,
             overlay_ipv4_range: Some(crate::state::DEFAULT_IPV4_RANGE),
         }
     }
@@ -278,6 +392,21 @@ impl AgentConfig {
         self
     }
 
+    /// Peers this agent will not open or accept a **direct** data link with.
+    ///
+    /// Test-only, and behind the `testing` feature so a release build has
+    /// no such switch. The one arrangement that cannot be made on a single
+    /// host is two agents that both reach a third and not each other,
+    /// which is exactly the case a relay exists for.
+    #[cfg(feature = "testing")]
+    pub fn with_unreachable_data_peers(
+        mut self,
+        peers: Arc<std::sync::Mutex<std::collections::HashSet<iroh::EndpointId>>>,
+    ) -> Self {
+        self.unreachable_data_peers = peers;
+        self
+    }
+
     /// Sets the transport policy.
     pub fn with_transport(mut self, transport: TransportPolicy) -> Self {
         self.transport = transport;
@@ -290,7 +419,19 @@ impl AgentConfig {
         self
     }
 
-    /// Sets how often discovery runs.
+    /// Adds Mainline discovery alongside the configured candidate backend.
+    pub fn with_dht(mut self, dht: MainlineDiscovery) -> Self {
+        self.dht = Some(dht);
+        self
+    }
+
+    /// Sets lookup, publication and isolation recovery timing.
+    pub fn with_discovery_policy(mut self, policy: DiscoveryPolicy) -> Self {
+        self.discovery_policy = policy;
+        self
+    }
+
+    /// Sets the known-peer maintenance and endpoint-address observation interval.
     pub fn with_discovery_interval(mut self, interval: Duration) -> Self {
         self.discovery_interval = interval;
         self
@@ -314,6 +455,29 @@ impl AgentConfig {
         self
     }
 
+    /// Registers a custom packet transport for a specific protocol.
+    pub fn with_custom_transport(
+        mut self,
+        protocol: impl Into<String>,
+        transport: Arc<dyn crate::dataplane::PacketTransport>,
+    ) -> Self {
+        self.custom_transports.insert(protocol.into(), transport);
+        self
+    }
+
+    /// Gives the agent an overlay interface, from this factory.
+    pub fn with_interface(
+        mut self,
+        factory: Arc<dyn crate::overlay::TunFactory>,
+        name: impl Into<String>,
+        mtu: u32,
+    ) -> Self {
+        self.tun_factory = Some(factory);
+        self.interface_name = name.into();
+        self.interface_mtu = mtu;
+        self
+    }
+
     /// Replaces the limits.
     pub fn with_limits(mut self, limits: Limits) -> Self {
         self.limits = limits;
@@ -335,6 +499,8 @@ impl std::fmt::Debug for AgentConfig {
             .field("transport", &self.transport)
             .field("hostname", &self.hostname)
             .field("discovery", &self.discovery.as_ref().map(|d| d.name()))
+            .field("dht", &self.dht.is_some())
+            .field("discovery_policy", &self.discovery_policy)
             .field("discovery_interval", &self.discovery_interval)
             .field("limits", &self.limits)
             .field("reconnect", &self.reconnect)
@@ -349,3 +515,5 @@ impl std::fmt::Debug for AgentConfig {
             .finish()
     }
 }
+/// Maximum number of transport links a routed datagram may traverse.
+pub const ROUTING_HOP_LIMIT: u8 = 16;

crates/tsunagi-core/src/dataplane/mod.rs

@@ -19,8 +19,10 @@
 //! A data plane failure never stops the daemon: errors returned here are
 //! recorded and surfaced, the control plane keeps running.
 
+pub mod relay;
+pub mod routing;
 pub mod transport;
-pub mod wireguard;
+pub mod tunnel;
 
 use std::sync::Arc;
 
@@ -30,7 +32,10 @@ use tokio::sync::mpsc;
 use crate::BoxFuture;
 use crate::identity::NetworkId;
 
-pub use transport::{PacketLink, PacketTransport, SharedLink, TransportError};
+pub use transport::{
+    MultiPacketTransport, PacketLink, PacketTransport, SharedLink, TransportError,
+};
+pub use tunnel::{GenericTunnelPlugin, TunnelCodec};
 
 /// Maximum length of a plugin protocol identifier.
 pub const MAX_PROTOCOL_ID_LEN: usize = 32;
@@ -42,7 +47,7 @@ pub const MAX_PROTOCOL_ID_LEN: usize = 32;
 /// chooses to do so must validate it itself.
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub struct PluginCapability {
-    /// Protocol identifier, e.g. `wireguard`. Bounded by [`MAX_PROTOCOL_ID_LEN`].
+    /// Protocol identifier, e.g. `wg-quic`. Bounded by [`MAX_PROTOCOL_ID_LEN`].
     pub protocol: String,
     /// Version of the plugin's announcement format.
     pub version: u16,
@@ -52,6 +57,51 @@ pub struct PluginCapability {
     pub data: Vec<u8>,
 }
 
+/// Where a protocol hands the packets it has decrypted.
+///
+/// A protocol proves *who* sent a packet; it does not know what that member
+/// is entitled to say, because entitlement is an address claim the system
+/// level holds. So a decrypted packet goes here rather than straight to an
+/// interface, and is checked on the way.
+pub trait PacketSink: Send + Sync + 'static {
+    /// Hands over one packet, attributed to the peer whose tunnel decrypted
+    /// it.
+    fn deliver<'a>(
+        &'a self,
+        network: NetworkId,
+        peer: EndpointId,
+        packet: bytes::Bytes,
+    ) -> BoxFuture<'a, ()>;
+}
+
+/// A sink that drops everything, for a protocol running without an interface.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct DiscardPackets;
+
+impl PacketSink for DiscardPackets {
+    fn deliver<'a>(
+        &'a self,
+        _network: NetworkId,
+        _peer: EndpointId,
+        _packet: bytes::Bytes,
+    ) -> BoxFuture<'a, ()> {
+        Box::pin(async move {})
+    }
+}
+
+/// One setting a protocol accepts.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct ProtocolOption {
+    /// The key, as written in `key=value`.
+    pub key: &'static str,
+    /// A word for the value, for the help line: `BYTES`, `NAME`.
+    pub value: &'static str,
+    /// What it does.
+    pub help: &'static str,
+    /// What happens when it is not given.
+    pub default: Option<&'static str>,
+}
+
 /// Errors a plugin may return. They are recorded, never fatal for the agent.
 #[derive(Debug, thiserror::Error)]
 #[non_exhaustive]
@@ -69,7 +119,7 @@ pub enum PluginError {
 
 /// A request a plugin makes of the agent that owns it.
 #[derive(Debug)]
-pub(crate) enum PluginRequest {
+pub enum PluginRequest {
     /// Re-send this agent's announcement to every peer of a network.
     Reannounce(NetworkId),
     /// Surface a plugin error on the agent's event stream.
@@ -81,6 +131,17 @@ pub(crate) enum PluginRequest {
         /// Human readable reason, free of secrets.
         reason: String,
     },
+    /// A plugin reported that a data link or tunnel for a peer failed.
+    ProtocolFailure {
+        /// Network the failure occurred in.
+        network: NetworkId,
+        /// Peer whose protocol failed.
+        peer: EndpointId,
+        /// Protocol that failed.
+        protocol: String,
+        /// Human readable reason.
+        reason: String,
+    },
 }
 
 /// The agent-side handle a plugin is given when it is attached.
@@ -95,13 +156,32 @@ pub(crate) enum PluginRequest {
 pub struct PluginContext {
     sender: Option<mpsc::Sender<PluginRequest>>,
     local: Option<EndpointId>,
+    /// Where decrypted packets go. Absent when nothing is carrying traffic,
+    /// in which case a protocol still runs and its packets are discarded.
+    sink: Option<Arc<dyn PacketSink>>,
 }
 
 impl PluginContext {
-    pub(crate) fn new(sender: mpsc::Sender<PluginRequest>, local: EndpointId) -> Self {
+    pub(crate) fn new(
+        sender: mpsc::Sender<PluginRequest>,
+        local: EndpointId,
+        sink: Option<Arc<dyn PacketSink>>,
+    ) -> Self {
         Self {
             sender: Some(sender),
             local: Some(local),
+            sink,
+        }
+    }
+
+    /// Where to hand a decrypted packet.
+    ///
+    /// Always answers: with no interface to write to, the packets are
+    /// discarded, which is what `--no-tun` means and is not an error.
+    pub fn packet_sink(&self) -> Arc<dyn PacketSink> {
+        match &self.sink {
+            Some(sink) => Arc::clone(sink),
+            None => Arc::new(DiscardPackets),
         }
     }
 
@@ -110,6 +190,30 @@ impl PluginContext {
         Self {
             sender: None,
             local: None,
+            sink: None,
+        }
+    }
+
+    /// A context configured with a custom packet sink.
+    pub fn with_sink(sink: Arc<dyn PacketSink>) -> Self {
+        Self {
+            sender: None,
+            local: None,
+            sink: Some(sink),
+        }
+    }
+
+    /// A context configured with a channel for plugin requests and custom sink, for tests.
+    #[cfg(feature = "testing")]
+    pub fn for_test(
+        sender: mpsc::Sender<PluginRequest>,
+        local: EndpointId,
+        sink: Option<Arc<dyn PacketSink>>,
+    ) -> Self {
+        Self {
+            sender: Some(sender),
+            local: Some(local),
+            sink,
         }
     }
 
@@ -154,6 +258,22 @@ impl PluginContext {
             reason: reason.into(),
         });
     }
+
+    /// Reports that data plane connectivity or a tunnel for a peer has failed.
+    pub fn report_peer_protocol_failure(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        protocol: impl Into<String>,
+        reason: impl Into<String>,
+    ) {
+        self.send(PluginRequest::ProtocolFailure {
+            network,
+            peer,
+            protocol: protocol.into(),
+            reason: reason.into(),
+        });
+    }
 }
 
 impl std::fmt::Debug for PluginContext {
@@ -170,11 +290,34 @@ impl std::fmt::Debug for PluginContext {
 /// Implementations must be cheap and non-blocking: the agent calls them from
 /// its runtime tasks. Anything slow belongs in the plugin's own tasks.
 pub trait IpPlugin: Send + Sync + std::fmt::Debug + 'static {
-    /// Stable protocol identifier, e.g. `wireguard`.
+    /// Stable protocol identifier, e.g. `wg-quic`.
     ///
     /// Must be non-empty and at most [`MAX_PROTOCOL_ID_LEN`] bytes.
     fn protocol_id(&self) -> &str;
 
+    /// The wire version of this protocol.
+    ///
+    /// Two peers carry traffic for each other only when they have the same
+    /// protocol at the same version. There is no negotiating a middle
+    /// ground: a protocol either speaks the same words at both ends or it
+    /// does not, and pretending otherwise produces a link that fails later
+    /// and less clearly.
+    ///
+    /// **Not the software version.** A plugin crate has its own version and
+    /// it is nobody else's business: two peers on different builds work
+    /// together for as long as the wire between them has not changed. This
+    /// number moves only when the bytes do, so it must never be derived from
+    /// `CARGO_PKG_VERSION` or anything else that moves with a release.
+    fn protocol_version(&self) -> u16;
+
+    /// The settings this protocol accepts, and what they mean.
+    ///
+    /// Declared rather than documented elsewhere, so the agent can list them
+    /// without knowing anything about the protocol.
+    fn options(&self) -> &'static [ProtocolOption] {
+        &[]
+    }
+
     /// Called once, when the agent starts, before any network is activated.
     ///
     /// The plugin keeps the context to ask for re-announcements and to report
@@ -224,6 +367,20 @@ pub trait IpPlugin: Send + Sync + std::fmt::Debug + 'static {
         let _ = (network, range, allocations);
     }
 
+    /// Carries one packet to a peer, encrypting it however this protocol
+    /// does.
+    ///
+    /// `false` when it cannot right now — no link, no tunnel, not this
+    /// protocol's peer — which the caller reports rather than treats as an
+    /// error. The packet came off the one interface the agent owns, and
+    /// which protocol takes it is settled by asking.
+    ///
+    /// The default carries nothing, which is right for a plugin that only
+    /// announces something.
+    fn carry(&self, _network: NetworkId, _peer: EndpointId, _packet: bytes::Bytes) -> bool {
+        false
+    }
+
     /// A data plane link to a peer is available for this plugin's protocol.
     ///
     /// The plugin moves its packets over this link and never learns how the
@@ -243,6 +400,19 @@ pub trait IpPlugin: Send + Sync + std::fmt::Debug + 'static {
     /// The plugin is expected to remove whatever it created for that network.
     fn on_network_deactivated(&self, network: NetworkId);
 
+    /// Called when this agent has left a network for good.
+    ///
+    /// Deactivation is temporary and keeps everything ready for next time;
+    /// this is the other one. Whatever the plugin holds *durably* for that
+    /// network — a key of its own, a file, a record — goes now, because the
+    /// agent is no longer a member and keeping it is keeping a secret for a
+    /// network it cannot rejoin without being told the secret again.
+    ///
+    /// Always preceded by [`IpPlugin::on_network_deactivated`], so this is
+    /// only about what outlives a session. Defaulted to nothing, for a
+    /// plugin that stores nothing.
+    fn on_network_forgotten(&self, _network: NetworkId) {}
+
     /// Called once when the agent shuts down.
     ///
     /// The plugin removes the system objects it created and stops its tasks.
@@ -290,6 +460,10 @@ impl IpPlugin for TestCapabilityPlugin {
         &self.protocol
     }
 
+    fn protocol_version(&self) -> u16 {
+        1
+    }
+
     fn local_capability(
         &self,
         _network: NetworkId,

crates/tsunagi-core/src/dataplane/relay.rs

@@ -0,0 +1,638 @@
+//! Userspace forwarding. Topology updates publish immutable tables; the packet
+//! path takes no routing mutex, walks no graph and never inspects plugin bytes.
+//! Each transport reader forwards transit immediately, independently of the
+//! destination protocol's receive task. Only local delivery enters its inbox.
+
+use std::collections::{HashMap, HashSet};
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
+
+use arc_swap::ArcSwap;
+use bytes::Bytes;
+use iroh::EndpointId;
+use tokio::sync::{Notify, mpsc, watch};
+use tokio::task::JoinHandle;
+
+use super::routing::{FlowId, PeerId, RoutingTable, envelope, mix64};
+use super::transport::{PacketLink, SharedLink, TransportError};
+use crate::BoxFuture;
+use crate::config::{MAX_DATA_DATAGRAM, ROUTING_HOP_LIMIT};
+use crate::identity::NetworkId;
+
+/// Fixed routing header, subtracted from the logical transport MTU.
+pub const RELAY_OVERHEAD: usize = envelope::HEADER;
+const INBOX: usize = 256;
+
+/// Forwarding counters, shared by all protocols of one network.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct RelayCounters {
+    /// Datagrams originated here and sent through another member.
+    pub sent_via: u64,
+    /// Transit datagrams handed to the next transport link.
+    pub forwarded: u64,
+    /// Local datagrams received through another member.
+    pub received_via: u64,
+    /// Missing or failed next hop.
+    pub dropped_no_link: u64,
+    /// Malformed envelope or unknown member.
+    pub dropped_unknown: u64,
+    /// Exhausted hop limit, including loops during topology convergence.
+    pub dropped_hop_limit: u64,
+    /// Full local protocol inbox. Transit does not use this queue.
+    pub dropped_congested: u64,
+}
+
+#[derive(Debug, Default)]
+struct Tally {
+    sent_via: AtomicU64,
+    forwarded: AtomicU64,
+    received_via: AtomicU64,
+    dropped_no_link: AtomicU64,
+    dropped_unknown: AtomicU64,
+    dropped_hop_limit: AtomicU64,
+    dropped_congested: AtomicU64,
+}
+
+impl Tally {
+    fn snapshot(&self) -> RelayCounters {
+        RelayCounters {
+            sent_via: self.sent_via.load(Ordering::Relaxed),
+            forwarded: self.forwarded.load(Ordering::Relaxed),
+            received_via: self.received_via.load(Ordering::Relaxed),
+            dropped_no_link: self.dropped_no_link.load(Ordering::Relaxed),
+            dropped_unknown: self.dropped_unknown.load(Ordering::Relaxed),
+            dropped_hop_limit: self.dropped_hop_limit.load(Ordering::Relaxed),
+            dropped_congested: self.dropped_congested.load(Ordering::Relaxed),
+        }
+    }
+}
+
+/// The links to one neighbour, best first. Which transport carries a frame
+/// is decided here and nowhere else: the end-to-end protocol is in the
+/// envelope and never depends on it.
+#[derive(Debug)]
+struct NextHop {
+    peer: PeerId,
+    links: Box<[SharedLink]>,
+}
+
+impl NextHop {
+    fn usable(&self) -> bool {
+        self.links.iter().any(|link| !link.is_closed())
+    }
+
+    fn first_open(&self) -> Option<&SharedLink> {
+        self.links.iter().find(|link| !link.is_closed())
+    }
+
+    /// Hands a frame to the best open link that accepts it. A link that
+    /// cannot carry the frame (too large for its path, closed a moment ago)
+    /// leaves it to the next one; the frame is only copied when there is
+    /// another to try.
+    fn send(&self, frame: Bytes) -> Result<(), TransportError> {
+        let mut open = self.links.iter().filter(|link| !link.is_closed());
+        let Some(mut current) = open.next() else {
+            return Err(TransportError::Closed);
+        };
+        for next in open {
+            if current.send(frame.clone()).is_ok() {
+                return Ok(());
+            }
+            current = next;
+        }
+        current.send(frame)
+    }
+}
+
+#[derive(Debug)]
+struct ForwardRoute {
+    hops: u8,
+    next: Box<[NextHop]>,
+}
+
+impl ForwardRoute {
+    fn select(&self, source: &PeerId, flow: FlowId) -> Option<&NextHop> {
+        if self.next.len() == 1 {
+            return self.next.first().filter(|hop| hop.usable());
+        }
+        let seed = u64::from_le_bytes(source[..8].try_into().ok()?);
+        let start = mix64(flow ^ seed) as usize % self.next.len();
+        // Stable fallback if a link closed just before its table was replaced.
+        (0..self.next.len())
+            .map(|i| &self.next[(start + i) % self.next.len()])
+            .find(|hop| hop.usable())
+    }
+}
+
+#[derive(Debug, Default)]
+struct ForwardingTable {
+    routes: HashMap<PeerId, ForwardRoute>,
+    /// Local delivery, by who sent it and which end-to-end protocol it is.
+    inboxes: HashMap<(PeerId, u64), mpsc::Sender<Bytes>>,
+    members: HashSet<PeerId>,
+}
+
+#[derive(Debug)]
+struct Plane {
+    local: PeerId,
+    table: ArcSwap<ForwardingTable>,
+    tally: Arc<Tally>,
+}
+
+impl Plane {
+    /// Synchronous hot path: one snapshot, bounded header decode, table lookup,
+    /// and transport send. An exclusively owned buffer is modified in place.
+    fn receive(&self, incoming: PeerId, frame: Bytes) {
+        let Some(header) = envelope::decode(&frame) else {
+            self.tally.dropped_unknown.fetch_add(1, Ordering::Relaxed);
+            return;
+        };
+        let table = self.table.load();
+        if !table.members.contains(&header.source) || header.source == self.local {
+            self.tally.dropped_unknown.fetch_add(1, Ordering::Relaxed);
+            return;
+        }
+        if header.destination == self.local {
+            let Some(inbox) = table.inboxes.get(&(header.source, header.protocol)) else {
+                self.tally.dropped_unknown.fetch_add(1, Ordering::Relaxed);
+                return;
+            };
+            if inbox.try_send(frame.slice(RELAY_OVERHEAD..)).is_err() {
+                self.tally.dropped_congested.fetch_add(1, Ordering::Relaxed);
+            } else if incoming != header.source {
+                self.tally.received_via.fetch_add(1, Ordering::Relaxed);
+            }
+            return;
+        }
+        if header.remaining <= 1 {
+            self.tally.dropped_hop_limit.fetch_add(1, Ordering::Relaxed);
+            return;
+        }
+        let hop = table
+            .routes
+            .get(&header.destination)
+            .and_then(|route| route.select(&header.source, header.flow));
+        let Some(hop) = hop else {
+            self.tally.dropped_no_link.fetch_add(1, Ordering::Relaxed);
+            return;
+        };
+        if hop.send(envelope::decrement(frame)).is_ok() {
+            self.tally.forwarded.fetch_add(1, Ordering::Relaxed);
+        } else {
+            self.tally.dropped_no_link.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+}
+
+/// One physical link to a neighbour, of whatever kind of transport.
+#[derive(Debug)]
+struct RawLink {
+    kind: String,
+    link: SharedLink,
+    reader: JoinHandle<()>,
+}
+
+/// How a destination is reached right now.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct RouteSummary {
+    /// Number of transport links to it. One is a direct link.
+    pub hops: u8,
+    /// The neighbour the first link goes to. For a direct link that is the
+    /// destination itself.
+    pub via: EndpointId,
+}
+
+#[derive(Debug)]
+struct State {
+    plane: Arc<Plane>,
+    /// Transport kinds, best first. A neighbour with several links is
+    /// reached over the first of these it has.
+    priority: Vec<String>,
+    graph: HashMap<PeerId, Vec<PeerId>>,
+    members: HashSet<PeerId>,
+    raw: HashMap<PeerId, Vec<RawLink>>,
+    links: HashMap<(PeerId, String), Arc<PeerLink>>,
+}
+
+impl State {
+    fn rank(&self, kind: &str) -> usize {
+        self.priority
+            .iter()
+            .position(|known| known == kind)
+            .unwrap_or(usize::MAX)
+    }
+
+    /// A neighbour's links, best kind first.
+    fn ordered(&self, peer: &PeerId) -> Vec<&RawLink> {
+        let mut links: Vec<&RawLink> = self
+            .raw
+            .get(peer)
+            .map(|links| links.iter().collect())
+            .unwrap_or_default();
+        links.sort_by_key(|raw| (self.rank(&raw.kind), raw.kind.clone()));
+        links
+    }
+
+    fn publish(&self) {
+        let mut graph = self.graph.clone();
+        let mut neighbours: Vec<PeerId> = self
+            .raw
+            .iter()
+            .filter(|(_, links)| links.iter().any(|raw| !raw.link.is_closed()))
+            .map(|(peer, _)| *peer)
+            .collect();
+        neighbours.sort_unstable();
+        graph.insert(self.plane.local, neighbours);
+        let routing = RoutingTable::build(self.plane.local, &graph, ROUTING_HOP_LIMIT);
+        let routes = routing
+            .iter()
+            .filter_map(|(peer, route)| {
+                let next: Box<[_]> = route
+                    .next_hops
+                    .iter()
+                    .filter_map(|id| {
+                        let links: Box<[_]> = self
+                            .ordered(id)
+                            .into_iter()
+                            .map(|raw| raw.link.clone())
+                            .collect();
+                        (!links.is_empty()).then_some(NextHop { peer: *id, links })
+                    })
+                    .collect();
+                (!next.is_empty()).then_some((
+                    *peer,
+                    ForwardRoute {
+                        hops: route.hops,
+                        next,
+                    },
+                ))
+            })
+            .collect();
+        let mut members = self.members.clone();
+        members.extend(self.raw.keys().copied());
+        self.plane.table.store(Arc::new(ForwardingTable {
+            routes,
+            members,
+            inboxes: self
+                .links
+                .iter()
+                .map(|((peer, _), link)| ((*peer, link.tag), link.inbox_tx.clone()))
+                .collect(),
+        }));
+    }
+
+    fn stop_readers(&mut self) {
+        for raw in self.raw.values().flatten() {
+            raw.reader.abort();
+        }
+    }
+}
+
+impl Drop for State {
+    fn drop(&mut self) {
+        self.stop_readers();
+        for link in self.links.values() {
+            link.close();
+        }
+        self.plane.table.store(Arc::default());
+    }
+}
+
+/// One network's transport links and its published forwarding table.
+///
+/// There is one table for the whole network, whatever protocols run over it.
+/// A link is a way to reach a neighbour; a protocol is an end-to-end channel
+/// to a peer. The two meet only in the envelope, which names the protocol, so
+/// a tunnel keeps working over any mix of transports and a middle member
+/// needs no protocol in common with either end.
+#[derive(Debug)]
+pub struct RelayHub {
+    network: NetworkId,
+    state: Mutex<State>,
+    tally: Arc<Tally>,
+    changed: Arc<Notify>,
+}
+
+impl RelayHub {
+    /// Creates the router at this network's local endpoint.
+    pub fn new(network: NetworkId, local: EndpointId) -> Arc<Self> {
+        let tally = Arc::new(Tally::default());
+        Arc::new(Self {
+            network,
+            state: Mutex::new(State {
+                plane: Arc::new(Plane {
+                    local: *local.as_bytes(),
+                    table: ArcSwap::from_pointee(ForwardingTable::default()),
+                    tally: tally.clone(),
+                }),
+                priority: Vec::new(),
+                graph: HashMap::new(),
+                members: HashSet::new(),
+                raw: HashMap::new(),
+                links: HashMap::new(),
+            }),
+            tally,
+            changed: Arc::default(),
+        })
+    }
+
+    fn state(&self) -> std::sync::MutexGuard<'_, State> {
+        self.state.lock().unwrap_or_else(|error| error.into_inner())
+    }
+
+    /// Sets which kinds of transport are preferred, best first, when a
+    /// neighbour can be reached over more than one.
+    pub fn set_priority(&self, kinds: Vec<String>) {
+        let mut state = self.state();
+        if state.priority != kinds {
+            state.priority = kinds;
+            state.publish();
+        }
+    }
+
+    /// Wakes the control loop promptly when a transport reader ends.
+    pub async fn changed(&self) {
+        self.changed.notified().await;
+    }
+
+    /// Current counters.
+    pub fn counters(&self) -> RelayCounters {
+        self.tally.snapshot()
+    }
+
+    /// Publishes authenticated topology. Refreshing identical state is cheap.
+    ///
+    /// Every row of `graph` is one member's own account of who it has a link
+    /// to; what kind of link it is does not matter here.
+    pub fn set_topology(&self, graph: HashMap<PeerId, Vec<PeerId>>, members: HashSet<PeerId>) {
+        let mut state = self.state();
+        if state.graph != graph || state.members != members {
+            state.graph = graph;
+            state.members = members;
+            state.publish();
+        }
+    }
+
+    /// Whether the current table has a path to this destination.
+    pub fn reachable(&self, peer: EndpointId) -> bool {
+        self.state()
+            .plane
+            .table
+            .load()
+            .routes
+            .contains_key(peer.as_bytes())
+    }
+
+    /// How a destination is reached right now, if it is.
+    pub fn route(&self, peer: EndpointId) -> Option<RouteSummary> {
+        let table = self.state().plane.table.load_full();
+        let route = table.routes.get(peer.as_bytes())?;
+        let via = EndpointId::from_bytes(&route.next.first()?.peer).ok()?;
+        Some(RouteSummary {
+            hops: route.hops,
+            via,
+        })
+    }
+
+    /// The kinds of transport with an open direct link to a peer, best
+    /// first.
+    pub fn direct_kinds(&self, peer: EndpointId) -> Vec<String> {
+        let state = self.state();
+        state
+            .ordered(peer.as_bytes())
+            .into_iter()
+            .filter(|raw| !raw.link.is_closed())
+            .map(|raw| raw.kind.clone())
+            .collect()
+    }
+
+    /// Stable end-to-end channel for a protocol, preserved across topology
+    /// and transport changes.
+    pub fn link(&self, peer: EndpointId, protocol: &str) -> Arc<PeerLink> {
+        let mut state = self.state();
+        let key = (*peer.as_bytes(), protocol.to_owned());
+        if let Some(link) = state.links.get(&key) {
+            return link.clone();
+        }
+        let (inbox_tx, inbox_rx) = mpsc::channel(INBOX);
+        let (closed, _) = watch::channel(false);
+        let link = Arc::new(PeerLink {
+            network: self.network,
+            peer,
+            tag: envelope::protocol_tag(protocol),
+            plane: state.plane.clone(),
+            inbox_tx,
+            inbox_rx: tokio::sync::Mutex::new(inbox_rx),
+            closed,
+        });
+        state.links.insert(key, link.clone());
+        state.publish();
+        link
+    }
+
+    /// Whether the protocol already owns its end-to-end channel to a peer.
+    pub fn has_link(&self, peer: EndpointId, protocol: &str) -> bool {
+        self.state()
+            .links
+            .contains_key(&(*peer.as_bytes(), protocol.to_owned()))
+    }
+
+    /// Installs a raw transport link and starts its independent ingress reader.
+    ///
+    /// A neighbour may have one link per kind of transport; installing a
+    /// second of the same kind replaces the first.
+    pub fn set_direct(&self, peer: EndpointId, kind: &str, raw: SharedLink) {
+        if raw.network() != self.network || raw.peer() != peer {
+            return;
+        }
+        let mut state = self.state();
+        let id = *peer.as_bytes();
+        let plane = state.plane.clone();
+        let changed = self.changed.clone();
+        let reader_link = raw.clone();
+        let reader = tokio::spawn(async move {
+            let mut batch = 0;
+            while let Some(frame) = reader_link.recv().await {
+                plane.receive(id, frame);
+                batch += 1;
+                if batch == 64 {
+                    tokio::task::yield_now().await;
+                    batch = 0;
+                }
+            }
+            changed.notify_one();
+        });
+        let links = state.raw.entry(id).or_default();
+        if let Some(position) = links.iter().position(|known| known.kind == kind) {
+            links.remove(position).reader.abort();
+        }
+        links.push(RawLink {
+            kind: kind.to_owned(),
+            link: raw,
+            reader,
+        });
+        state.publish();
+    }
+
+    /// Removes one physical link without tearing down end-to-end protocol state.
+    pub fn clear_direct(&self, peer: EndpointId, kind: &str) {
+        let mut state = self.state();
+        let mut changed = false;
+        if let Some(links) = state.raw.get_mut(peer.as_bytes()) {
+            if let Some(position) = links.iter().position(|known| known.kind == kind) {
+                links.remove(position).reader.abort();
+                changed = true;
+            }
+            if links.is_empty() {
+                state.raw.remove(peer.as_bytes());
+            }
+        }
+        if changed {
+            state.publish();
+        }
+    }
+
+    /// Revokes all links and topology involving a departed authenticated peer.
+    pub fn remove_peer(&self, peer: EndpointId) {
+        let mut state = self.state();
+        if let Some(links) = state.raw.remove(peer.as_bytes()) {
+            for raw in links {
+                raw.reader.abort();
+            }
+        }
+        let gone: Vec<_> = state
+            .links
+            .keys()
+            .filter(|(id, _)| id == peer.as_bytes())
+            .cloned()
+            .collect();
+        for key in gone {
+            if let Some(link) = state.links.remove(&key) {
+                link.close();
+            }
+        }
+        state.members.remove(peer.as_bytes());
+        state.graph.remove(peer.as_bytes());
+        for neighbors in state.graph.values_mut() {
+            neighbors.retain(|id| id != peer.as_bytes());
+        }
+        state.publish();
+    }
+
+    /// Stops every reader and releases all transport handles.
+    pub fn close(&self) {
+        let mut state = self.state();
+        state.stop_readers();
+        state.raw.clear();
+        for link in state.links.values() {
+            link.close();
+        }
+        state.links.clear();
+        state.graph.clear();
+        state.members.clear();
+        state.plane.table.store(Arc::default());
+    }
+}
+
+/// A protocol's end-to-end channel, independent of its current next hop.
+#[derive(Debug)]
+pub struct PeerLink {
+    network: NetworkId,
+    peer: EndpointId,
+    tag: u64,
+    plane: Arc<Plane>,
+    inbox_tx: mpsc::Sender<Bytes>,
+    inbox_rx: tokio::sync::Mutex<mpsc::Receiver<Bytes>>,
+    closed: watch::Sender<bool>,
+}
+
+impl PeerLink {
+    fn close(&self) {
+        self.closed.send_replace(true);
+    }
+}
+
+impl PacketLink for PeerLink {
+    fn network(&self) -> NetworkId {
+        self.network
+    }
+    fn peer(&self) -> EndpointId {
+        self.peer
+    }
+    fn max_datagram_size(&self) -> usize {
+        MAX_DATA_DATAGRAM - RELAY_OVERHEAD
+    }
+    fn send(&self, payload: Bytes) -> Result<(), TransportError> {
+        self.send_flow(payload, 0)
+    }
+    fn send_flow(&self, payload: Bytes, flow: FlowId) -> Result<(), TransportError> {
+        if self.is_closed() {
+            return Err(TransportError::Closed);
+        }
+        if payload.len() > self.max_datagram_size() {
+            return Err(TransportError::TooLarge {
+                size: payload.len(),
+                limit: self.max_datagram_size(),
+            });
+        }
+        let table = self.plane.table.load();
+        let route = table
+            .routes
+            .get(self.peer.as_bytes())
+            .ok_or(TransportError::Closed)?;
+        let hop = route
+            .select(&self.plane.local, flow)
+            .ok_or(TransportError::Closed)?;
+        hop.send(envelope::encode(
+            self.plane.local,
+            *self.peer.as_bytes(),
+            flow,
+            self.tag,
+            &payload,
+        ))?;
+        if route.hops > 1 {
+            self.plane.tally.sent_via.fetch_add(1, Ordering::Relaxed);
+        }
+        Ok(())
+    }
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+        Box::pin(async move {
+            let mut closed = self.closed.subscribe();
+            if *closed.borrow_and_update() {
+                return None;
+            }
+            let mut inbox = self.inbox_rx.lock().await;
+            tokio::select! { biased; _ = closed.changed() => None, frame = inbox.recv() => frame }
+        })
+    }
+    fn closed(&self) -> BoxFuture<'_, ()> {
+        Box::pin(async move {
+            let mut closed = self.closed.subscribe();
+            if !*closed.borrow_and_update() {
+                let _ = closed.changed().await;
+            }
+        })
+    }
+    fn is_closed(&self) -> bool {
+        *self.closed.borrow()
+    }
+    fn path_description(&self) -> String {
+        let table = self.plane.table.load();
+        match table.routes.get(self.peer.as_bytes()) {
+            Some(route) if route.hops == 1 => route.next[0]
+                .first_open()
+                .map_or_else(|| "unreachable".into(), |link| link.path_description()),
+            Some(route) => format!(
+                "relay {} hops via {} ({} equal paths)",
+                route.hops,
+                hex::encode(route.next[0].peer),
+                route.next.len()
+            ),
+            None => "unreachable".into(),
+        }
+    }
+}
+
+#[cfg(test)]
+#[path = "relay_tests.rs"]
+mod tests;

crates/tsunagi-core/src/dataplane/relay_tests.rs

@@ -0,0 +1,418 @@
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+use super::*;
+use crate::identity::{NetworkKeys, NetworkName, NetworkSecret};
+use std::sync::atomic::AtomicBool;
+
+fn network() -> NetworkId {
+    NetworkKeys::derive(
+        &NetworkName::new("routing-tests").unwrap(),
+        &NetworkSecret::from_bytes([3; 32]).unwrap(),
+    )
+    .network_id()
+}
+fn peer(seed: u8) -> EndpointId {
+    iroh::SecretKey::from_bytes(&[seed; 32]).public()
+}
+fn id(seed: u8) -> PeerId {
+    *peer(seed).as_bytes()
+}
+fn tag() -> u64 {
+    envelope::protocol_tag("ip")
+}
+
+#[derive(Debug)]
+struct Wire {
+    peer: EndpointId,
+    sent: Mutex<Vec<Bytes>>,
+    feed: mpsc::Sender<Bytes>,
+    inbound: tokio::sync::Mutex<mpsc::Receiver<Bytes>>,
+    dead: AtomicBool,
+    count: AtomicU64,
+    record: bool,
+}
+impl Wire {
+    fn new(seed: u8, record: bool) -> Arc<Self> {
+        let (feed, inbound) = mpsc::channel(32);
+        Arc::new(Self {
+            peer: peer(seed),
+            sent: Mutex::default(),
+            feed,
+            inbound: tokio::sync::Mutex::new(inbound),
+            dead: AtomicBool::new(false),
+            count: AtomicU64::new(0),
+            record,
+        })
+    }
+    fn take(&self) -> Vec<Bytes> {
+        std::mem::take(&mut self.sent.lock().unwrap())
+    }
+}
+impl PacketLink for Wire {
+    fn network(&self) -> NetworkId {
+        network()
+    }
+    fn peer(&self) -> EndpointId {
+        self.peer
+    }
+    fn max_datagram_size(&self) -> usize {
+        MAX_DATA_DATAGRAM
+    }
+    fn send(&self, frame: Bytes) -> Result<(), TransportError> {
+        if self.is_closed() {
+            return Err(TransportError::Closed);
+        }
+        self.count.fetch_add(1, Ordering::Relaxed);
+        if self.record {
+            self.sent.lock().unwrap().push(frame);
+        }
+        Ok(())
+    }
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+        Box::pin(async move { self.inbound.lock().await.recv().await })
+    }
+    fn closed(&self) -> BoxFuture<'_, ()> {
+        Box::pin(std::future::pending())
+    }
+    fn is_closed(&self) -> bool {
+        self.dead.load(Ordering::Relaxed)
+    }
+    fn path_description(&self) -> String {
+        "test-direct".into()
+    }
+}
+
+fn topology(hub: &RelayHub, edges: &[(u8, &[u8])], members: &[u8]) {
+    hub.set_topology(
+        edges
+            .iter()
+            .map(|(a, bs)| (id(*a), bs.iter().map(|b| id(*b)).collect()))
+            .collect(),
+        members.iter().map(|p| id(*p)).collect(),
+    );
+}
+
+#[tokio::test]
+async fn ecmp_is_per_flow_direct_wins_and_existing_tunnel_survives_changes() {
+    let hub = RelayHub::new(network(), peer(1));
+    let b = Wire::new(2, true);
+    let c = Wire::new(3, true);
+    hub.set_direct(peer(2), "ip", b.clone());
+    hub.set_direct(peer(3), "ip", c.clone());
+    topology(&hub, &[(2, &[4]), (3, &[4])], &[1, 2, 3, 4]);
+    let link = hub.link(peer(4), "ip");
+    let mut chosen = HashMap::new();
+    for _ in 0..8 {
+        for flow in 0..100 {
+            link.send_flow(Bytes::from_static(b"ciphertext"), flow)
+                .unwrap();
+        }
+        for (hop, wire) in [(2, &b), (3, &c)] {
+            for frame in wire.take() {
+                let header = envelope::decode(&frame).unwrap();
+                assert_eq!(header.source, id(1));
+                assert_eq!(header.destination, id(4));
+                if let Some(old) = chosen.insert(header.flow, hop) {
+                    assert_eq!(old, hop);
+                }
+            }
+        }
+    }
+    assert_eq!(
+        chosen.values().copied().collect::<HashSet<_>>(),
+        HashSet::from([2, 3])
+    );
+    let direct = Wire::new(4, true);
+    hub.set_direct(peer(4), "ip", direct.clone());
+    link.send_flow(Bytes::from_static(b"direct"), 8).unwrap();
+    assert_eq!(direct.take().len(), 1);
+    assert!(b.take().is_empty() && c.take().is_empty());
+    hub.clear_direct(peer(4), "ip");
+    b.dead.store(true, Ordering::Relaxed);
+    for flow in 0..100 {
+        link.send_flow(Bytes::from_static(b"fallback"), flow)
+            .unwrap();
+    }
+    assert_eq!(c.take().len(), 100);
+    hub.clear_direct(peer(2), "ip");
+    topology(&hub, &[(3, &[])], &[1, 2, 3, 4]);
+    assert!(link.send(Bytes::new()).is_err());
+    assert!(!link.is_closed());
+    assert!(Arc::ptr_eq(&link, &hub.link(peer(4), "ip")));
+    hub.remove_peer(peer(4));
+    assert!(link.is_closed());
+    tokio::time::timeout(std::time::Duration::from_secs(1), link.closed())
+        .await
+        .unwrap();
+    assert!(link.recv().await.is_none());
+    hub.close();
+}
+
+#[tokio::test]
+async fn transport_reader_forwards_without_any_plugin_reader_and_isolates_protocols() {
+    let hub = RelayHub::new(network(), peer(2));
+    let a = Wire::new(1, true);
+    let c = Wire::new(3, true);
+    hub.set_direct(peer(1), "ip", a.clone());
+    hub.set_direct(peer(3), "ip", c.clone());
+    topology(&hub, &[(3, &[4])], &[1, 2, 3, 4]);
+    let frame = envelope::encode(id(1), id(4), 91, tag(), b"end-to-end encrypted");
+    let pointer = frame.as_ptr();
+    a.feed.send(frame).await.unwrap();
+    tokio::time::timeout(std::time::Duration::from_secs(1), async {
+        while hub.counters().forwarded == 0 {
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .unwrap();
+    let sent = c.take();
+    assert_eq!(sent[0].as_ptr(), pointer);
+    assert_eq!(
+        envelope::decode(&sent[0]).unwrap().remaining,
+        ROUTING_HOP_LIMIT - 1
+    );
+    assert_eq!(&sent[0][RELAY_OVERHEAD..], b"end-to-end encrypted");
+    assert_eq!(hub.counters().forwarded, 1);
+    hub.close();
+}
+
+#[tokio::test]
+async fn a_frame_is_delivered_only_to_the_protocol_it_names() {
+    let hub = RelayHub::new(network(), peer(2));
+    hub.set_direct(peer(1), "ip", Wire::new(1, true));
+    let ip = hub.link(peer(1), "ip");
+    let other = hub.link(peer(1), "other-ip");
+    let plane = ip.plane.clone();
+    plane.receive(id(1), envelope::encode(id(1), id(2), 0, tag(), b"for ip"));
+    plane.receive(
+        id(1),
+        envelope::encode(
+            id(1),
+            id(2),
+            0,
+            envelope::protocol_tag("other-ip"),
+            b"for other",
+        ),
+    );
+    plane.receive(
+        id(1),
+        envelope::encode(
+            id(1),
+            id(2),
+            0,
+            envelope::protocol_tag("nobody-runs-this"),
+            b"for nobody",
+        ),
+    );
+    assert_eq!(ip.recv().await.unwrap(), Bytes::from_static(b"for ip"));
+    assert_eq!(
+        other.recv().await.unwrap(),
+        Bytes::from_static(b"for other")
+    );
+    assert_eq!(hub.counters().dropped_unknown, 1);
+    hub.close();
+}
+
+/// A middle member joins two neighbours over different transports, and the
+/// two ends share a protocol the middle has never heard of.
+#[tokio::test]
+async fn a_tunnel_crosses_a_middle_member_whose_links_are_of_different_kinds() {
+    let a = RelayHub::new(network(), peer(1));
+    let c = RelayHub::new(network(), peer(3));
+    let b = RelayHub::new(network(), peer(2));
+    let a_to_c = Wire::new(3, true);
+    let c_to_a = Wire::new(1, true);
+    let c_to_b = Wire::new(2, true);
+    let b_to_c = Wire::new(3, true);
+    a.set_direct(peer(3), "wg", a_to_c.clone());
+    c.set_direct(peer(1), "wg", c_to_a);
+    c.set_direct(peer(2), "wg-quic", c_to_b.clone());
+    b.set_direct(peer(3), "wg-quic", b_to_c.clone());
+    topology(&a, &[(3, &[1, 2])], &[1, 2, 3]);
+    topology(&b, &[(3, &[1, 2])], &[1, 2, 3]);
+    topology(&c, &[(1, &[3]), (2, &[3])], &[1, 2, 3]);
+
+    let outgoing = a.link(peer(2), "tunnel");
+    let incoming = b.link(peer(1), "tunnel");
+    outgoing
+        .send_flow(Bytes::from_static(b"sealed"), 5)
+        .unwrap();
+    let frame = a_to_c.take().remove(0);
+
+    // The middle forwards it onto its other kind of link without looking
+    // inside.
+    c.state().plane.receive(id(1), frame);
+    let frame = c_to_b.take().remove(0);
+    assert_eq!(c.counters().forwarded, 1);
+
+    b.state().plane.receive(id(3), frame);
+    assert_eq!(
+        incoming.recv().await.unwrap(),
+        Bytes::from_static(b"sealed")
+    );
+    assert_eq!(b.counters().received_via, 1);
+    assert_eq!(
+        b.route(peer(1)),
+        Some(RouteSummary {
+            hops: 2,
+            via: peer(3)
+        })
+    );
+    a.close();
+    b.close();
+    c.close();
+}
+
+#[tokio::test]
+async fn a_neighbour_is_reached_over_the_preferred_kind_and_falls_back_when_it_is_lost() {
+    let hub = RelayHub::new(network(), peer(1));
+    hub.set_priority(vec!["wg".into(), "tcp-tls".into(), "wg-quic".into()]);
+    let slow = Wire::new(2, true);
+    let fast = Wire::new(2, true);
+    hub.set_direct(peer(2), "wg-quic", slow.clone());
+    let link = hub.link(peer(2), "tunnel");
+    link.send(Bytes::from_static(b"one")).unwrap();
+    assert_eq!(slow.take().len(), 1);
+    assert_eq!(hub.direct_kinds(peer(2)), ["wg-quic"]);
+
+    // A better kind appears: traffic moves to it with nothing to redo.
+    hub.set_direct(peer(2), "wg", fast.clone());
+    assert_eq!(hub.direct_kinds(peer(2)), ["wg", "wg-quic"]);
+    link.send(Bytes::from_static(b"two")).unwrap();
+    assert_eq!(fast.take().len(), 1);
+    assert!(slow.take().is_empty());
+
+    // It dies: the worse one carries on, and the logical link is untouched.
+    fast.dead.store(true, Ordering::Relaxed);
+    link.send(Bytes::from_static(b"three")).unwrap();
+    assert_eq!(slow.take().len(), 1);
+    assert!(!link.is_closed());
+    assert_eq!(
+        hub.route(peer(2)),
+        Some(RouteSummary {
+            hops: 1,
+            via: peer(2)
+        })
+    );
+
+    hub.clear_direct(peer(2), "wg-quic");
+    hub.clear_direct(peer(2), "wg");
+    assert!(!hub.reachable(peer(2)));
+    assert!(Arc::ptr_eq(&link, &hub.link(peer(2), "tunnel")));
+    hub.close();
+}
+
+#[tokio::test]
+async fn local_delivery_is_bounded_and_malformed_or_unknown_sources_are_dropped() {
+    let hub = RelayHub::new(network(), peer(2));
+    let a = Wire::new(1, true);
+    hub.set_direct(peer(1), "ip", a);
+    let link = hub.link(peer(1), "ip");
+    let plane = link.plane.clone();
+    for _ in 0..INBOX + 1 {
+        plane.receive(id(1), envelope::encode(id(1), id(2), 0, tag(), b"ip"));
+    }
+    assert_eq!(hub.counters().dropped_congested, 1);
+    for _ in 0..INBOX {
+        assert_eq!(link.recv().await.unwrap(), Bytes::from_static(b"ip"));
+    }
+    plane.receive(id(1), Bytes::from_static(b"broken"));
+    plane.receive(
+        id(1),
+        envelope::encode(id(9), id(2), 0, tag(), b"unknown member"),
+    );
+    assert_eq!(hub.counters().dropped_unknown, 2);
+    plane.receive(
+        id(1),
+        envelope::encode(id(1), id(9), 0, tag(), b"unknown destination"),
+    );
+    assert_eq!(hub.counters().dropped_no_link, 1);
+    hub.close();
+}
+
+#[tokio::test]
+async fn inconsistent_tables_cannot_loop_beyond_the_hop_limit() {
+    let a = RelayHub::new(network(), peer(1));
+    let b = RelayHub::new(network(), peer(2));
+    let ab = Wire::new(2, true);
+    let ba = Wire::new(1, true);
+    a.set_direct(peer(2), "ip", ab.clone());
+    b.set_direct(peer(1), "ip", ba.clone());
+    // Deliberately inconsistent snapshots while updates are in flight.
+    topology(&a, &[(2, &[4])], &[1, 2, 3, 4]);
+    topology(&b, &[(1, &[4])], &[1, 2, 3, 4]);
+    let ap = a.link(peer(4), "ip").plane.clone();
+    let bp = b.link(peer(4), "ip").plane.clone();
+    let mut frame = envelope::encode(id(3), id(4), 7, tag(), b"transit");
+    for turn in 0..ROUTING_HOP_LIMIT {
+        let (plane, incoming, wire) = if turn % 2 == 0 {
+            (&ap, id(2), &ab)
+        } else {
+            (&bp, id(1), &ba)
+        };
+        plane.receive(incoming, frame);
+        let sent = wire.take();
+        if turn == ROUTING_HOP_LIMIT - 1 {
+            assert!(sent.is_empty());
+            break;
+        }
+        assert_eq!(sent.len(), 1);
+        frame = sent.into_iter().next().unwrap();
+    }
+    assert_eq!(
+        a.counters().forwarded + b.counters().forwarded,
+        u64::from(ROUTING_HOP_LIMIT - 1)
+    );
+    assert_eq!(
+        a.counters().dropped_hop_limit + b.counters().dropped_hop_limit,
+        1
+    );
+    a.close();
+    b.close();
+}
+
+/// Measures the actual synchronous transit routine, including header checks,
+/// snapshot lookup, ECMP, in-place TTL update and mock transport submission.
+/// Packet construction, encryption and socket I/O are outside this measurement.
+#[tokio::test]
+#[ignore = "manual release-mode forwarding microbenchmark"]
+async fn forwarding_benchmark() {
+    let hub = RelayHub::new(network(), peer(2));
+    let wire = Wire::new(3, false);
+    hub.set_direct(peer(3), "ip", wire.clone());
+    let alternate = Wire::new(5, false);
+    hub.set_direct(peer(5), "ip", alternate.clone());
+    topology(&hub, &[(3, &[4]), (5, &[4])], &[1, 2, 3, 4, 5]);
+    let plane = hub.link(peer(4), "ip").plane.clone();
+    let source = id(1);
+    let destination = id(4);
+    let incoming = id(1);
+    for size in [64, 1280] {
+        let mut elapsed = std::time::Duration::ZERO;
+        const COUNT: usize = 10000;
+        const BATCHES: usize = 100;
+        for _ in 0..BATCHES {
+            let frames: Vec<_> = (0..COUNT)
+                .map(|flow| {
+                    envelope::encode(source, destination, flow as u64, tag(), &vec![42; size])
+                })
+                .collect();
+            let start = std::time::Instant::now();
+            for frame in frames {
+                plane.receive(incoming, std::hint::black_box(frame));
+            }
+            elapsed += start.elapsed();
+        }
+        let count = (COUNT * BATCHES) as f64;
+        println!(
+            "transit {size} B: {:.1} ns/packet, {:.2} Mpps (1M packets, two equal paths)",
+            elapsed.as_nanos() as f64 / count,
+            count / elapsed.as_secs_f64() / 1e6
+        );
+    }
+    assert_eq!(
+        wire.count.load(Ordering::Relaxed) + alternate.count.load(Ordering::Relaxed),
+        2000000
+    );
+    hub.close();
+}

crates/tsunagi-core/src/dataplane/routing/envelope.rs

@@ -0,0 +1,104 @@
+//! Fixed header. Transit changes one byte in an owned receive buffer.
+use super::{FlowId, PeerId};
+use crate::config::{MAX_DATA_DATAGRAM, ROUTING_HOP_LIMIT};
+use bytes::{BufMut, Bytes, BytesMut};
+
+pub(crate) const HEADER: usize = 2 + 32 + 32 + 8 + 8;
+const VERSION: u8 = 2;
+
+#[derive(Debug, PartialEq, Eq)]
+pub(crate) struct Header {
+    pub source: PeerId,
+    pub destination: PeerId,
+    pub remaining: u8,
+    pub flow: FlowId,
+    /// Which end-to-end protocol the payload belongs to, whatever links carry it.
+    pub protocol: u64,
+}
+
+/// A stable 64-bit tag for a protocol id (FNV-1a), the same on every member.
+pub(crate) fn protocol_tag(name: &str) -> u64 {
+    name.bytes().fold(0xcbf2_9ce4_8422_2325, |hash, byte| {
+        (hash ^ u64::from(byte)).wrapping_mul(0x0000_0100_0000_01b3)
+    })
+}
+
+pub(crate) fn decode(bytes: &[u8]) -> Option<Header> {
+    if bytes.len() < HEADER
+        || bytes.len() > MAX_DATA_DATAGRAM
+        || bytes[0] != VERSION
+        || bytes[1] == 0
+        || bytes[1] > ROUTING_HOP_LIMIT
+    {
+        return None;
+    }
+    Some(Header {
+        source: bytes[2..34].try_into().ok()?,
+        destination: bytes[34..66].try_into().ok()?,
+        remaining: bytes[1],
+        flow: u64::from_be_bytes(bytes[66..74].try_into().ok()?),
+        protocol: u64::from_be_bytes(bytes[74..82].try_into().ok()?),
+    })
+}
+
+pub(crate) fn encode(
+    source: PeerId,
+    destination: PeerId,
+    flow: FlowId,
+    protocol: u64,
+    payload: &[u8],
+) -> Bytes {
+    let mut bytes = BytesMut::with_capacity(HEADER + payload.len());
+    bytes.put_u8(VERSION);
+    bytes.put_u8(ROUTING_HOP_LIMIT);
+    bytes.extend_from_slice(&source);
+    bytes.extend_from_slice(&destination);
+    bytes.put_u64(flow);
+    bytes.put_u64(protocol);
+    bytes.extend_from_slice(payload);
+    bytes.freeze()
+}
+
+pub(crate) fn decrement(bytes: Bytes) -> Bytes {
+    let mut bytes = match bytes.try_into_mut() {
+        Ok(bytes) => bytes,
+        Err(bytes) => BytesMut::from(bytes.as_ref()),
+    };
+    // Called only after decode and after excluding remaining <= 1.
+    bytes[1] -= 1;
+    bytes.freeze()
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    #[test]
+    fn transit_reuses_the_buffer_and_rejects_invalid_headers() {
+        let frame = encode([1; 32], [2; 32], 17, protocol_tag("wg"), &[42; 1280]);
+        let address = frame.as_ptr();
+        let frame = decrement(frame);
+        assert_eq!(
+            frame.as_ptr(),
+            address,
+            "no payload copy for an owned buffer"
+        );
+        let header = decode(&frame).unwrap();
+        assert_eq!(header.remaining, ROUTING_HOP_LIMIT - 1);
+        assert_eq!(header.protocol, protocol_tag("wg"));
+        assert_ne!(protocol_tag("wg"), protocol_tag("wg-quic"));
+        assert_eq!(&frame[HEADER..], &[42; 1280]);
+        for length in 0..HEADER {
+            assert!(decode(&frame[..length]).is_none());
+        }
+        for ttl in [0, ROUTING_HOP_LIMIT + 1] {
+            let mut bad = frame.to_vec();
+            bad[1] = ttl;
+            assert!(decode(&bad).is_none());
+        }
+        let mut bad = frame.to_vec();
+        bad[0] = 99;
+        assert!(decode(&bad).is_none());
+        assert!(decode(&vec![1; MAX_DATA_DATAGRAM + 1]).is_none());
+    }
+}

crates/tsunagi-core/src/dataplane/routing/flow.rs

@@ -0,0 +1,125 @@
+//! Optional IP flow classifier for IP plugins, called before encryption.
+//! The router itself treats this identifier and the payload as opaque.
+
+use super::FlowId;
+
+fn hash(parts: &[&[u8]]) -> FlowId {
+    let mut value = 0xcbf29ce484222325u64;
+    for part in parts {
+        for byte in *part {
+            value = (value ^ u64::from(*byte)).wrapping_mul(0x100000001b3);
+        }
+    }
+    value
+}
+
+/// Stable directional IP flow hash, independent of TCP sequence numbers,
+/// checksums and packet contents. Fragmented traffic uses addresses/protocol
+/// instead of ports: even non-initial fragments and subsequent fragmented
+/// datagrams stay together. This deliberately coalesces fragmented flows.
+/// Malformed packets get a coarse hash; parsing is bounded and never allocates.
+pub fn ip_flow(packet: &[u8]) -> FlowId {
+    match packet.first().map(|b| b >> 4) {
+        Some(4) if packet.len() >= 20 => {
+            let addresses = &packet[12..20];
+            let protocol = &packet[9..10];
+            let offset = usize::from(packet[0] & 15) * 4;
+            if u16::from_be_bytes([packet[6], packet[7]]) & 0x3fff != 0 {
+                return hash(&[addresses, protocol]);
+            }
+            if offset >= 20
+                && matches!(packet[9], 6 | 17)
+                && let Some(ports) = packet.get(offset..offset + 4)
+            {
+                return hash(&[addresses, protocol, ports]);
+            }
+            hash(&[addresses, protocol])
+        }
+        Some(6) if packet.len() >= 40 => {
+            let addresses = &packet[8..40];
+            let mut protocol = packet[6];
+            let mut offset = 40;
+            for _ in 0..8 {
+                match protocol {
+                    44 => {
+                        if let Some(fragment) = packet.get(offset..offset + 8) {
+                            return hash(&[addresses, &fragment[..1]]);
+                        }
+                        break;
+                    }
+                    0 | 43 | 60 | 51 => {
+                        let Some(header) = packet.get(offset..offset + 2) else {
+                            break;
+                        };
+                        let size = if protocol == 51 {
+                            (usize::from(header[1]) + 2) * 4
+                        } else {
+                            (usize::from(header[1]) + 1) * 8
+                        };
+                        protocol = header[0];
+                        offset += size;
+                    }
+                    6 | 17 => {
+                        if let Some(ports) = packet.get(offset..offset + 4) {
+                            return hash(&[addresses, &[protocol], ports]);
+                        }
+                        break;
+                    }
+                    _ => break,
+                }
+            }
+            hash(&[addresses, &[protocol]])
+        }
+        _ => 0,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn tcp_flow_ignores_packet_contents_but_distinguishes_ports() {
+        let mut packet = [0u8; 80];
+        packet[0] = 0x45;
+        packet[9] = 6;
+        packet[12..24].copy_from_slice(&[10, 0, 0, 1, 10, 0, 0, 2, 0, 22, 1, 2]);
+        let flow = ip_flow(&packet);
+        packet[4] = 123; // IP id changes for each unfragmented packet.
+        packet[24..].fill(47);
+        assert_eq!(flow, ip_flow(&packet));
+        packet[23] = 3;
+        assert_ne!(flow, ip_flow(&packet));
+        packet[6] = 0x20;
+        let fragment_flow = ip_flow(&packet);
+        packet[6] = 0;
+        packet[7] = 1;
+        packet[20..].fill(99);
+        assert_eq!(fragment_flow, ip_flow(&packet));
+        packet[4] = 97;
+        assert_eq!(fragment_flow, ip_flow(&packet));
+    }
+
+    #[test]
+    fn ipv6_extensions_and_fragments_are_bounded_and_stable() {
+        let mut packet = [0u8; 80];
+        packet[0] = 0x60;
+        packet[6] = 0;
+        packet[40] = 17;
+        packet[48..52].copy_from_slice(&[1, 2, 3, 4]);
+        let flow = ip_flow(&packet);
+        packet[52..].fill(5);
+        assert_eq!(flow, ip_flow(&packet));
+        packet[50] = 9;
+        assert_ne!(flow, ip_flow(&packet));
+        packet[6] = 44;
+        packet[40] = 6;
+        let flow = ip_flow(&packet);
+        packet[42] = 32;
+        packet[48..].fill(3);
+        assert_eq!(flow, ip_flow(&packet));
+        for len in 0..packet.len() {
+            let _ = ip_flow(&packet[..len]);
+        }
+    }
+}

crates/tsunagi-core/src/dataplane/routing/mod.rs

@@ -0,0 +1,131 @@
+//! Transport-independent routing over opaque identities. Tables are built on
+//! topology changes, never on the packet path. No WireGuard or iroh APIs here.
+pub(crate) mod envelope;
+pub mod flow;
+
+use std::collections::{BTreeSet, HashMap, VecDeque};
+
+/// Opaque authenticated identity; the transport adapter validates its keys.
+pub type PeerId = [u8; 32];
+/// Stable flow identifier, carried unchanged between transit routers.
+pub type FlowId = u64;
+
+/// A shortest route with all equal-cost first hops in stable order.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Route {
+    /// Number of transport links to the destination.
+    pub hops: u8,
+    /// Equal-cost first hops. Direct destinations have one entry.
+    pub next_hops: Box<[PeerId]>,
+}
+
+/// Immutable shortest-path table, replaced when connectivity changes.
+#[derive(Debug, Clone, Default)]
+pub struct RoutingTable {
+    routes: HashMap<PeerId, Route>,
+}
+
+impl RoutingTable {
+    /// Computes shortest directed paths bounded by `hop_limit`.
+    /// The caller supplies only authenticated, fresh, compatible edges.
+    pub fn build(local: PeerId, graph: &HashMap<PeerId, Vec<PeerId>>, hop_limit: u8) -> Self {
+        let mut distances = HashMap::from([(local, 0u8)]);
+        let mut first_hops: HashMap<PeerId, BTreeSet<PeerId>> = HashMap::new();
+        let mut pending = VecDeque::from([local]);
+        while let Some(node) = pending.pop_front() {
+            let distance = distances[&node];
+            if distance >= hop_limit {
+                continue;
+            }
+            let inherited = first_hops.get(&node).cloned().unwrap_or_default();
+            for &next in graph.get(&node).into_iter().flatten() {
+                let next_distance = distance + 1;
+                match distances.get(&next) {
+                    Some(&known) if known < next_distance => continue,
+                    None => {
+                        distances.insert(next, next_distance);
+                        pending.push_back(next);
+                    }
+                    _ => {}
+                }
+                let hops = first_hops.entry(next).or_default();
+                if node == local {
+                    hops.insert(next);
+                } else {
+                    hops.extend(inherited.iter().copied());
+                }
+            }
+        }
+        Self {
+            routes: first_hops
+                .into_iter()
+                .map(|(peer, hops)| {
+                    (
+                        peer,
+                        Route {
+                            hops: distances[&peer],
+                            next_hops: hops.into_iter().collect(),
+                        },
+                    )
+                })
+                .collect(),
+        }
+    }
+    /// Precomputed route to a destination.
+    pub fn get(&self, destination: &PeerId) -> Option<&Route> {
+        self.routes.get(destination)
+    }
+    /// All reachable destinations, excluding ourselves.
+    pub fn iter(&self) -> impl Iterator<Item = (&PeerId, &Route)> {
+        self.routes.iter()
+    }
+}
+
+/// Stable ECMP mixing, without random state or hashing the payload per hop.
+pub(crate) fn mix64(mut value: u64) -> u64 {
+    value ^= value >> 30;
+    value = value.wrapping_mul(0xbf58476d1ce4e5b9);
+    value ^= value >> 27;
+    value = value.wrapping_mul(0x94d049bb133111eb);
+    value ^ (value >> 31)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    fn p(n: u8) -> PeerId {
+        [n; 32]
+    }
+    #[test]
+    fn shortest_paths_keep_equal_cost_hops_and_prefer_direct_links() {
+        let mut graph = HashMap::from([
+            (p(1), vec![p(3), p(2)]),
+            (p(2), vec![p(4)]),
+            (p(3), vec![p(4)]),
+            (p(4), vec![p(5)]),
+            (p(5), vec![p(1)]),
+        ]);
+        let routes = RoutingTable::build(p(1), &graph, 16);
+        assert_eq!(routes.get(&p(5)).unwrap().hops, 3);
+        assert_eq!(&*routes.get(&p(5)).unwrap().next_hops, &[p(2), p(3)]);
+        assert!(routes.get(&p(1)).is_none());
+        assert!(routes.get(&p(9)).is_none());
+        graph.get_mut(&p(1)).unwrap().push(p(5));
+        assert_eq!(
+            RoutingTable::build(p(1), &graph, 16).get(&p(5)).unwrap(),
+            &Route {
+                hops: 1,
+                next_hops: Box::new([p(5)])
+            }
+        );
+    }
+    #[test]
+    fn topology_removal_and_hop_budget_remove_impossible_routes() {
+        let mut graph = HashMap::from([(p(1), vec![p(2)]), (p(2), vec![p(3)]), (p(3), vec![p(4)])]);
+        assert!(RoutingTable::build(p(1), &graph, 2).get(&p(4)).is_none());
+        assert!(RoutingTable::build(p(1), &graph, 3).get(&p(4)).is_some());
+        graph.remove(&p(2));
+        assert!(RoutingTable::build(p(1), &graph, 16).get(&p(4)).is_none());
+    }
+}

crates/tsunagi-core/src/dataplane/transport/fragments.rs

@@ -0,0 +1,234 @@
+//! Bounded fragmentation of opaque transport payloads, below peer relaying.
+//!
+//! Data ALPN v3: u64 packet ID, u32 total length, u32 byte offset (big endian),
+//! followed by payload. Each QUIC connection owns its IDs and reassembly state.
+//! There is no retransmission: one missing fragment loses one datagram only.
+
+use std::collections::{HashMap, VecDeque};
+
+use bytes::{BufMut, Bytes, BytesMut};
+use tokio::time::Instant;
+
+use crate::config::{
+    DATA_REASSEMBLY_TTL, DATA_RECENT_IDS, MAX_DATA_ASSEMBLIES, MAX_DATA_DATAGRAM,
+    MAX_DATA_FRAGMENTS, MAX_DATA_REASSEMBLY_BYTES,
+};
+
+pub(super) const HEADER: usize = 16;
+
+pub(super) fn encode(id: u64, total: usize, offset: usize, payload: &[u8]) -> Bytes {
+    let mut frame = BytesMut::with_capacity(HEADER + payload.len());
+    frame.put_u64(id);
+    frame.put_u32(total as u32);
+    frame.put_u32(offset as u32);
+    frame.extend_from_slice(payload);
+    frame.freeze()
+}
+
+#[derive(Debug)]
+struct Assembly {
+    started: Instant,
+    bytes: Vec<u8>,
+    ranges: Vec<(usize, usize)>,
+    received: usize,
+}
+
+#[derive(Debug, Default)]
+pub(super) struct Reassembler {
+    packets: HashMap<u64, Assembly>,
+    buffered: usize,
+    recent: VecDeque<u64>,
+}
+
+impl Reassembler {
+    fn remember(&mut self, id: u64) {
+        if self.recent.len() == DATA_RECENT_IDS {
+            self.recent.pop_front();
+        }
+        self.recent.push_back(id);
+    }
+
+    fn remove(&mut self, id: u64) -> Option<Assembly> {
+        let packet = self.packets.remove(&id)?;
+        self.buffered -= packet.bytes.len();
+        self.remember(id);
+        Some(packet)
+    }
+
+    pub(super) fn expire(&mut self, now: Instant) {
+        let expired: Vec<_> = self
+            .packets
+            .iter()
+            .filter(|(_, p)| now.duration_since(p.started) >= DATA_REASSEMBLY_TTL)
+            .map(|(id, _)| *id)
+            .collect();
+        for id in expired {
+            self.remove(id);
+        }
+    }
+
+    pub(super) fn push(&mut self, frame: Bytes, now: Instant) -> Option<Bytes> {
+        self.expire(now);
+        let id = u64::from_be_bytes(frame.get(..8)?.try_into().ok()?);
+        let total = u32::from_be_bytes(frame.get(8..12)?.try_into().ok()?) as usize;
+        let offset = u32::from_be_bytes(frame.get(12..HEADER)?.try_into().ok()?) as usize;
+        let payload = frame.get(HEADER..)?;
+        let end = offset.checked_add(payload.len())?;
+        if total > MAX_DATA_DATAGRAM
+            || end > total
+            || (payload.is_empty() && total != 0)
+            || self.recent.contains(&id)
+        {
+            return None;
+        }
+        if !self.packets.contains_key(&id) {
+            if offset == 0 && end == total {
+                self.remember(id);
+                return Some(frame.slice(HEADER..));
+            }
+            // The length is validated before allocating. Drop the oldest
+            // incomplete packet when the per-link byte or entry budget is full.
+            while self.packets.len() >= MAX_DATA_ASSEMBLIES
+                || self.buffered + total > MAX_DATA_REASSEMBLY_BYTES
+            {
+                let oldest = self
+                    .packets
+                    .iter()
+                    .min_by_key(|(_, p)| p.started)
+                    .map(|(id, _)| *id)?;
+                self.remove(oldest);
+            }
+            self.packets.insert(
+                id,
+                Assembly {
+                    started: now,
+                    bytes: vec![0; total],
+                    ranges: Vec::new(),
+                    received: 0,
+                },
+            );
+            self.buffered += total;
+        }
+        let packet = self.packets.get_mut(&id)?;
+        if packet.bytes.len() != total {
+            self.remove(id);
+            return None;
+        }
+        for &(start, stop) in &packet.ranges {
+            if start == offset && stop == end && packet.bytes[offset..end] == *payload {
+                return None; // a duplicate does not reset the expiry time
+            }
+            if offset < stop && end > start {
+                self.remove(id); // conflicting or overlapping fragments
+                return None;
+            }
+        }
+        if packet.ranges.len() == MAX_DATA_FRAGMENTS {
+            self.remove(id);
+            return None;
+        }
+        packet.bytes[offset..end].copy_from_slice(payload);
+        packet.ranges.push((offset, end));
+        packet.received += payload.len();
+        if packet.received == total {
+            return self.remove(id).map(|p| Bytes::from(p.bytes));
+        }
+        None
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+
+    #[test]
+    fn reordered_duplicates_and_changing_fragment_sizes_reassemble_once() {
+        let mut rx = Reassembler::default();
+        let now = Instant::now();
+        let packet: Vec<u8> = (0..9000).map(|i| (i % 251) as u8).collect();
+        let mut parts = Vec::new();
+        let mut offset = 0;
+        for size in [1100, 700, 1100, 1100, 700, 1100, 1100, 1100, 1000] {
+            parts.push(encode(
+                7,
+                packet.len(),
+                offset,
+                &packet[offset..offset + size],
+            ));
+            offset += size;
+        }
+        assert_eq!(offset, packet.len());
+        for part in parts[1..].iter().rev() {
+            assert!(rx.push(part.clone(), now).is_none());
+            assert!(rx.push(part.clone(), now).is_none());
+        }
+        assert_eq!(rx.push(parts[0].clone(), now).unwrap().as_ref(), packet);
+        assert!(rx.push(parts[0].clone(), now).is_none());
+        assert_eq!(rx.buffered, 0);
+        assert!(rx.packets.is_empty());
+    }
+
+    #[test]
+    fn a_missing_fragment_never_blocks_the_next_packet_and_expires() {
+        let mut rx = Reassembler::default();
+        let now = Instant::now();
+        assert!(rx.push(encode(1, 1280, 0, &[1; 800]), now).is_none());
+        assert_eq!(
+            rx.push(encode(2, 4, 0, b"next"), now).unwrap(),
+            &b"next"[..]
+        );
+        rx.expire(now + DATA_REASSEMBLY_TTL);
+        assert!(rx.packets.is_empty());
+        assert_eq!(rx.buffered, 0);
+        assert!(
+            rx.push(encode(1, 1280, 800, &[1; 480]), now + DATA_REASSEMBLY_TTL)
+                .is_none()
+        );
+    }
+
+    #[test]
+    fn malformed_conflicting_and_oversized_fragments_are_bounded() {
+        let now = Instant::now();
+        let mut rx = Reassembler::default();
+        for len in 0..HEADER {
+            assert!(rx.push(Bytes::from(vec![0xff; len]), now).is_none());
+        }
+        assert!(
+            rx.push(encode(1, MAX_DATA_DATAGRAM + 1, 0, b"x"), now)
+                .is_none()
+        );
+        assert!(rx.push(encode(2, 100, 100, b"x"), now).is_none());
+        assert!(rx.push(encode(3, 100, 0, b""), now).is_none());
+        assert_eq!(rx.buffered, 0);
+        for (id, bad) in [
+            (4, encode(4, 100, 0, b"different")),
+            (5, encode(5, 100, 3, b"overlap")),
+            (6, encode(6, 101, 9, b"changed total")),
+        ] {
+            assert!(rx.push(encode(id, 100, 0, b"123456789"), now).is_none());
+            assert!(rx.push(bad, now).is_none());
+            assert_eq!(rx.buffered, 0);
+        }
+        for id in 100..400 {
+            assert!(
+                rx.push(encode(id, MAX_DATA_DATAGRAM, 0, b"x"), now)
+                    .is_none()
+            );
+            assert!(rx.buffered <= MAX_DATA_REASSEMBLY_BYTES);
+            assert!(rx.packets.len() <= MAX_DATA_ASSEMBLIES);
+            assert!(rx.recent.len() <= DATA_RECENT_IDS);
+        }
+    }
+
+    #[test]
+    fn excessive_fragment_counts_are_discarded() {
+        let mut rx = Reassembler::default();
+        let now = Instant::now();
+        for offset in 0..=MAX_DATA_FRAGMENTS {
+            assert!(rx.push(encode(1, 1000, offset, b"x"), now).is_none());
+        }
+        assert!(rx.packets.is_empty());
+        assert_eq!(rx.buffered, 0);
+    }
+}

crates/tsunagi-core/src/dataplane/transport/iroh_link.rs

@@ -20,13 +20,14 @@
 
 use std::collections::HashMap;
 use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
 
 use bytes::Bytes;
 use iroh::EndpointId;
 use iroh::endpoint::{Connection, RecvStream, SendStream};
 
 use crate::BoxFuture;
-use crate::config::Limits;
+use crate::config::{DATA_REASSEMBLY_TTL, Limits, MAX_DATA_DATAGRAM, MAX_DATA_FRAGMENTS};
 use crate::error::ProtocolError;
 use crate::identity::{NetworkId, NetworkKeys};
 use crate::net::EndpointAdapter;
@@ -36,6 +37,7 @@ use crate::proto::message::{
 };
 use crate::proto::{read_frame, write_frame};
 
+use super::fragments::{self, Reassembler};
 use super::{InboundLink, PacketLink, PacketTransport, SharedLink, TransportError};
 
 /// What the iroh transport needs from the agent.
@@ -61,6 +63,8 @@ pub struct IrohLink {
     peer: EndpointId,
     conn: Connection,
     max_datagram: usize,
+    next_packet: AtomicU64,
+    reassembly: tokio::sync::Mutex<Reassembler>,
     // Kept alive so the peer sees the channel as open; the connection closes
     // when the link is dropped.
     _send: tokio::sync::Mutex<SendStream>,
@@ -76,14 +80,16 @@ impl IrohLink {
         send: SendStream,
         recv: RecvStream,
     ) -> Self {
-        let local_limit = conn.max_datagram_size().unwrap_or(0);
+        // The negotiated limit is for reassembled payloads, never a snapshot
-        // Both ends must agree, so the smaller limit wins.
+        // of the initial path (which may still be a relay or another VPN).
-        let max_datagram = local_limit.min(peer_limit);
+        let max_datagram = MAX_DATA_DATAGRAM.min(peer_limit);
         Self {
             network,
             peer,
             conn,
             max_datagram,
+            next_packet: AtomicU64::new(0),
+            reassembly: tokio::sync::Mutex::new(Reassembler::default()),
             _send: tokio::sync::Mutex::new(send),
             _recv: tokio::sync::Mutex::new(recv),
         }
@@ -110,17 +116,65 @@ impl PacketLink for IrohLink {
                 limit: self.max_datagram,
             });
         }
-        self.conn.send_datagram(payload).map_err(|err| {
+        let id = self.next_packet.fetch_add(1, Ordering::Relaxed);
-            use iroh::endpoint::SendDatagramError;
+        let mut offset: usize = 0;
-            match err {
+        let mut attempts = 0;
-                SendDatagramError::ConnectionLost(_) => TransportError::Closed,
+        // Re-read the current path's capacity for every fragment. A migration
-                other => TransportError::Other(other.to_string()),
+        // can shrink it even between two send_datagram calls.
+        for _ in 0..MAX_DATA_FRAGMENTS {
+            loop {
+                let capacity = self
+                    .conn
+                    .max_datagram_size()
+                    .unwrap_or(0)
+                    .checked_sub(fragments::HEADER)
+                    .filter(|n| *n > 0)
+                    .ok_or_else(|| {
+                        TransportError::Other("QUIC path cannot carry data fragments".into())
+                    })?;
+                let end = offset.saturating_add(capacity).min(payload.len());
+                let frame = fragments::encode(id, payload.len(), offset, &payload[offset..end]);
+                match self.conn.send_datagram(frame) {
+                    Ok(()) => {
+                        offset = end;
+                        break;
+                    }
+                    Err(iroh::endpoint::SendDatagramError::TooLarge) if attempts < 3 => {
+                        attempts += 1;
+                    }
+                    Err(iroh::endpoint::SendDatagramError::ConnectionLost(_)) => {
+                        return Err(TransportError::Closed);
+                    }
+                    Err(err) => return Err(TransportError::Other(err.to_string())),
+                }
             }
-        })
+            if offset == payload.len() {
+                return Ok(());
+            }
+        }
+        Err(TransportError::Other(
+            "QUIC path needs too many fragments".into(),
+        ))
     }
 
     fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
-        Box::pin(async move { self.conn.read_datagram().await.ok() })
+        Box::pin(async move {
+            let mut reassembly = self.reassembly.lock().await;
+            loop {
+                match tokio::time::timeout(DATA_REASSEMBLY_TTL, self.conn.read_datagram()).await {
+                    Ok(Ok(frame)) => {
+                        if let Some(packet) = reassembly.push(frame, tokio::time::Instant::now()) {
+                            return Some(packet);
+                        }
+                    }
+                    Ok(Err(_)) => {
+                        *reassembly = Reassembler::default();
+                        return None;
+                    }
+                    Err(_) => reassembly.expire(tokio::time::Instant::now()),
+                }
+            }
+        })
     }
 
     fn closed(&self) -> BoxFuture<'_, ()> {
@@ -208,10 +262,14 @@ impl IrohTransport {
         }
 
         let serves = self.lookup.serves(outcome.network_id, &open.protocol).await;
-        let max_datagram = conn.max_datagram_size().unwrap_or(0);
+        if open.max_datagram == 0 || conn.max_datagram_size().is_none() {
+            return Err(TransportError::Other(
+                "data channel has no datagram support".into(),
+            ));
+        }
         let ack = DataOpenAck {
             accepted: serves,
-            max_datagram: max_datagram as u32,
+            max_datagram: MAX_DATA_DATAGRAM as u32,
         };
         write_frame(
             &mut send,
@@ -226,7 +284,14 @@ impl IrohTransport {
             return Err(TransportError::Declined(open.protocol));
         }
 
-        let link = IrohLink::new(outcome.network_id, peer, conn, usize::MAX, send, recv);
+        let link = IrohLink::new(
+            outcome.network_id,
+            peer,
+            conn,
+            open.max_datagram as usize,
+            send,
+            recv,
+        );
         Ok(InboundLink {
             network: outcome.network_id,
             peer,
@@ -289,6 +354,7 @@ impl PacketTransport for IrohTransport {
 
             let open = DataOpen {
                 protocol: protocol.to_string(),
+                max_datagram: MAX_DATA_DATAGRAM as u32,
             };
             write_frame(
                 &mut send,
@@ -310,6 +376,12 @@ impl PacketTransport for IrohTransport {
                 return Err(TransportError::Declined(protocol.to_string()));
             }
 
+            if ack.max_datagram == 0 || conn.max_datagram_size().is_none() {
+                return Err(TransportError::Other(
+                    "data channel has no datagram support".into(),
+                ));
+            }
+
             let link = IrohLink::new(network, peer, conn, ack.max_datagram as usize, send, recv);
             Ok(Arc::new(link) as SharedLink)
         })

crates/tsunagi-core/src/dataplane/transport/mod.rs

@@ -29,8 +29,11 @@
 //! encrypted by the transport, and scoped to exactly one network, one peer and
 //! one plugin protocol.
 
+mod fragments;
 pub mod iroh_link;
 
+use std::sync::Arc;
+
 use bytes::Bytes;
 use iroh::EndpointId;
 
@@ -77,8 +80,8 @@ pub trait PacketLink: Send + Sync + std::fmt::Debug + 'static {
 
     /// The largest datagram this link can carry, in bytes.
     ///
-    /// A plugin must size its own packets to fit, because there is no
+    /// This is the logical payload limit. A transport can fragment underneath
-    /// fragmentation here.
+    /// it so path MTU changes do not force a plugin to resize the host interface.
     fn max_datagram_size(&self) -> usize;
 
     /// Sends one datagram.
@@ -87,6 +90,16 @@ pub trait PacketLink: Send + Sync + std::fmt::Debug + 'static {
     /// handed to the transport, nothing more.
     fn send(&self, payload: Bytes) -> Result<(), TransportError>;
 
+    /// Sends a datagram with an opaque stable flow identifier. Plugins derive
+    /// this before encryption; transports that do not route can ignore it.
+    fn send_flow(
+        &self,
+        payload: Bytes,
+        _flow: super::routing::FlowId,
+    ) -> Result<(), TransportError> {
+        self.send(payload)
+    }
+
     /// Receives the next datagram, or `None` once the link is finished.
     fn recv(&self) -> BoxFuture<'_, Option<Bytes>>;
 
@@ -144,3 +157,51 @@ pub trait PacketTransport: Send + Sync + std::fmt::Debug + 'static {
         protocol: &'a str,
     ) -> BoxFuture<'a, Result<SharedLink, TransportError>>;
 }
+
+/// A composite packet transport that dispatches requests by protocol id.
+#[derive(Debug)]
+pub struct MultiPacketTransport {
+    default: Arc<dyn PacketTransport>,
+    by_protocol: std::collections::HashMap<String, Arc<dyn PacketTransport>>,
+}
+
+impl MultiPacketTransport {
+    /// Creates a new composite transport with a default fallback transport.
+    pub fn new(
+        default: Arc<dyn PacketTransport>,
+        by_protocol: std::collections::HashMap<String, Arc<dyn PacketTransport>>,
+    ) -> Self {
+        Self {
+            default,
+            by_protocol,
+        }
+    }
+
+    /// Access the default underlying transport.
+    pub fn default_transport(&self) -> &Arc<dyn PacketTransport> {
+        &self.default
+    }
+}
+
+impl PacketTransport for MultiPacketTransport {
+    fn name(&self) -> &str {
+        "multi"
+    }
+
+    fn as_any(&self) -> &dyn std::any::Any {
+        self
+    }
+
+    fn open<'a>(
+        &'a self,
+        network: NetworkId,
+        peer: EndpointId,
+        protocol: &'a str,
+    ) -> BoxFuture<'a, Result<SharedLink, TransportError>> {
+        if let Some(transport) = self.by_protocol.get(protocol) {
+            transport.open(network, peer, protocol)
+        } else {
+            self.default.open(network, peer, protocol)
+        }
+    }
+}

crates/tsunagi-core/src/dataplane/tunnel.rs

@@ -0,0 +1,380 @@
+//! Generic harness for IP tunnel plugins.
+//!
+//! Implementing [`super::IpPlugin`] directly requires managing per-peer background
+//! tasks, synchronization, packet queues, and graceful shutdown.
+//!
+//! [`TunnelCodec`] provides a simplified interface where a plugin author only
+//! defines cryptographic transformation and handshake payloads.
+//! [`GenericTunnelPlugin`] adapts any [`TunnelCodec`] into a full [`super::IpPlugin`].
+
+use std::collections::HashMap;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex, RwLock};
+
+use bytes::Bytes;
+use iroh::EndpointId;
+use tokio::task::JoinHandle;
+
+use super::{IpPlugin, PluginCapability, PluginContext, PluginError, ProtocolOption, SharedLink};
+use crate::BoxFuture;
+use crate::identity::NetworkId;
+
+/// The simplified contract for an IP tunnel protocol implementation.
+///
+/// Implementors only handle key announcements, link lifecycle events,
+/// and packet encrypt/decrypt. Concurrency and task management are handled
+/// by [`GenericTunnelPlugin`].
+pub trait TunnelCodec: Send + Sync + 'static {
+    /// Protocol identifier, e.g. `tcp-tls`.
+    fn protocol_id(&self) -> &str;
+
+    /// Protocol wire version.
+    fn protocol_version(&self) -> u16;
+
+    /// Options accepted by the protocol.
+    fn options(&self) -> &'static [ProtocolOption] {
+        &[]
+    }
+
+    /// Generates local capability announcement data for a network.
+    fn local_capability_data(&self, network: NetworkId) -> Result<Vec<u8>, PluginError>;
+
+    /// Handles a peer's announced capability payload.
+    fn on_peer_capability_data(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        data: &[u8],
+    ) -> Result<(), PluginError>;
+
+    /// Called when an authenticated link to a peer is established.
+    fn on_peer_link_up(&self, network: NetworkId, peer: EndpointId, link: &SharedLink) {
+        let _ = (network, peer, link);
+    }
+
+    /// Called when a peer goes away.
+    fn on_peer_down(&self, network: NetworkId, peer: EndpointId) {
+        let _ = (network, peer);
+    }
+
+    /// Encrypts an outbound IP packet into a payload to send over the link.
+    fn encrypt(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        packet: &[u8],
+    ) -> Result<Bytes, PluginError>;
+
+    /// Decrypts an inbound payload from the link into an IP packet.
+    fn decrypt(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        payload: &[u8],
+    ) -> Result<Bytes, PluginError>;
+
+    /// Called during shutdown.
+    fn shutdown<'a>(&'a self) -> BoxFuture<'a, ()> {
+        Box::pin(async {})
+    }
+}
+
+struct TunnelPeerState {
+    link: SharedLink,
+    tx_packets: AtomicU64,
+    rx_packets: AtomicU64,
+    tx_bytes: AtomicU64,
+    rx_bytes: AtomicU64,
+    dropped: AtomicU64,
+    established_at: std::time::Instant,
+}
+
+/// Statistics and status for one active peer in [`GenericTunnelPlugin`].
+#[derive(Debug, Clone)]
+pub struct TunnelPeerReport {
+    /// Peer endpoint identity.
+    pub peer: EndpointId,
+    /// Protocol identifier.
+    pub protocol: String,
+    /// Human-readable path description.
+    pub path: String,
+    /// Packets successfully encrypted and sent.
+    pub tx_packets: u64,
+    /// Packets successfully received and decrypted.
+    pub rx_packets: u64,
+    /// Plaintext bytes successfully encrypted and sent.
+    pub tx_bytes: u64,
+    /// Plaintext bytes successfully received and decrypted.
+    pub rx_bytes: u64,
+    /// Packets dropped due to encryption/decryption/send failures.
+    pub dropped: u64,
+    /// Uptime in seconds since the data link was established.
+    pub uptime_secs: u64,
+}
+
+/// Generic wrapper that turns a [`TunnelCodec`] into a full [`IpPlugin`].
+pub struct GenericTunnelPlugin<C: TunnelCodec> {
+    codec: Arc<C>,
+    context: Mutex<Option<PluginContext>>,
+    links: RwLock<HashMap<(NetworkId, EndpointId), Arc<TunnelPeerState>>>,
+    tasks: Mutex<HashMap<(NetworkId, EndpointId), JoinHandle<()>>>,
+}
+
+impl<C: TunnelCodec> GenericTunnelPlugin<C> {
+    /// Wraps `codec` into an [`IpPlugin`].
+    pub fn new(codec: C) -> Self {
+        Self {
+            codec: Arc::new(codec),
+            context: Mutex::new(None),
+            links: RwLock::new(HashMap::new()),
+            tasks: Mutex::new(HashMap::new()),
+        }
+    }
+
+    /// Returns a reference to the underlying codec.
+    pub fn codec(&self) -> &C {
+        &self.codec
+    }
+
+    /// Returns the active peers with established data links for a network.
+    pub fn active_peers(&self, network: NetworkId) -> Vec<TunnelPeerReport> {
+        let guard = match self.links.read() {
+            Ok(g) => g,
+            Err(p) => p.into_inner(),
+        };
+        guard
+            .iter()
+            .filter(|((net, _), _)| *net == network)
+            .map(|((_, peer), state)| TunnelPeerReport {
+                peer: *peer,
+                protocol: self.codec.protocol_id().to_string(),
+                path: state.link.path_description(),
+                tx_packets: state.tx_packets.load(Ordering::Relaxed),
+                rx_packets: state.rx_packets.load(Ordering::Relaxed),
+                tx_bytes: state.tx_bytes.load(Ordering::Relaxed),
+                rx_bytes: state.rx_bytes.load(Ordering::Relaxed),
+                dropped: state.dropped.load(Ordering::Relaxed),
+                uptime_secs: state.established_at.elapsed().as_secs(),
+            })
+            .collect()
+    }
+}
+
+impl<C: TunnelCodec> std::fmt::Debug for GenericTunnelPlugin<C> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("GenericTunnelPlugin")
+            .field("protocol_id", &self.codec.protocol_id())
+            .field("protocol_version", &self.codec.protocol_version())
+            .finish()
+    }
+}
+
+impl<C: TunnelCodec> IpPlugin for GenericTunnelPlugin<C> {
+    fn protocol_id(&self) -> &str {
+        self.codec.protocol_id()
+    }
+
+    fn protocol_version(&self) -> u16 {
+        self.codec.protocol_version()
+    }
+
+    fn options(&self) -> &'static [ProtocolOption] {
+        self.codec.options()
+    }
+
+    fn attach(&self, context: PluginContext) {
+        if let Ok(mut guard) = self.context.lock() {
+            *guard = Some(context);
+        }
+    }
+
+    fn local_capability(
+        &self,
+        network: NetworkId,
+    ) -> Result<Option<PluginCapability>, PluginError> {
+        let data = self.codec.local_capability_data(network)?;
+        Ok(Some(PluginCapability {
+            protocol: self.codec.protocol_id().to_string(),
+            version: self.codec.protocol_version(),
+            enabled: true,
+            data,
+        }))
+    }
+
+    fn on_peer_capability(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        capability: &PluginCapability,
+    ) -> Result<(), PluginError> {
+        if capability.protocol != self.codec.protocol_id() {
+            return Ok(());
+        }
+        self.codec
+            .on_peer_capability_data(network, peer, &capability.data)
+    }
+
+    fn carry(&self, network: NetworkId, peer: EndpointId, packet: bytes::Bytes) -> bool {
+        let state = {
+            let guard = match self.links.read() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard.get(&(network, peer)).cloned()
+        };
+
+        let Some(state) = state else {
+            return false;
+        };
+
+        match self.codec.encrypt(network, peer, &packet) {
+            Ok(ciphertext) => {
+                if state.link.send(ciphertext).is_ok() {
+                    state.tx_packets.fetch_add(1, Ordering::Relaxed);
+                    state
+                        .tx_bytes
+                        .fetch_add(packet.len() as u64, Ordering::Relaxed);
+                    true
+                } else {
+                    state.dropped.fetch_add(1, Ordering::Relaxed);
+                    false
+                }
+            }
+            Err(err) => {
+                state.dropped.fetch_add(1, Ordering::Relaxed);
+                tracing::debug!(%peer, %err, "tunnel encrypt failed");
+                false
+            }
+        }
+    }
+
+    fn on_peer_link(&self, network: NetworkId, peer: EndpointId, link: SharedLink) {
+        let state = Arc::new(TunnelPeerState {
+            link: Arc::clone(&link),
+            tx_packets: AtomicU64::new(0),
+            rx_packets: AtomicU64::new(0),
+            tx_bytes: AtomicU64::new(0),
+            rx_bytes: AtomicU64::new(0),
+            dropped: AtomicU64::new(0),
+            established_at: std::time::Instant::now(),
+        });
+        {
+            let mut guard = match self.links.write() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard.insert((network, peer), Arc::clone(&state));
+        }
+
+        self.codec.on_peer_link_up(network, peer, &link);
+
+        let codec = Arc::clone(&self.codec);
+        let sink = match self.context.lock() {
+            Ok(g) => g.as_ref().map(|ctx| ctx.packet_sink()),
+            Err(p) => p.into_inner().as_ref().map(|ctx| ctx.packet_sink()),
+        };
+
+        let rx_link = Arc::clone(&link);
+        let rx_state = Arc::clone(&state);
+        let handle = tokio::spawn(async move {
+            while let Some(payload) = rx_link.recv().await {
+                match codec.decrypt(network, peer, &payload) {
+                    Ok(packet) => {
+                        rx_state.rx_packets.fetch_add(1, Ordering::Relaxed);
+                        rx_state
+                            .rx_bytes
+                            .fetch_add(packet.len() as u64, Ordering::Relaxed);
+                        if let Some(ref sink) = sink {
+                            sink.deliver(network, peer, packet).await;
+                        }
+                    }
+                    Err(err) => {
+                        rx_state.dropped.fetch_add(1, Ordering::Relaxed);
+                        tracing::debug!(%peer, %err, "tunnel decrypt failed");
+                    }
+                }
+            }
+        });
+
+        let mut guard = match self.tasks.lock() {
+            Ok(g) => g,
+            Err(p) => p.into_inner(),
+        };
+        if let Some(old) = guard.insert((network, peer), handle) {
+            old.abort();
+        }
+    }
+
+    fn on_peer_gone(&self, network: NetworkId, peer: EndpointId) {
+        {
+            let mut guard = match self.tasks.lock() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            if let Some(task) = guard.remove(&(network, peer)) {
+                task.abort();
+            }
+        }
+        {
+            let mut guard = match self.links.write() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard.remove(&(network, peer));
+        }
+        self.codec.on_peer_down(network, peer);
+    }
+
+    fn on_network_deactivated(&self, network: NetworkId) {
+        let keys_to_remove: Vec<(NetworkId, EndpointId)> = {
+            let guard = match self.tasks.lock() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard
+                .keys()
+                .filter(|(net, _)| *net == network)
+                .copied()
+                .collect()
+        };
+
+        {
+            let mut guard = match self.tasks.lock() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            for key in &keys_to_remove {
+                if let Some(task) = guard.remove(key) {
+                    task.abort();
+                }
+            }
+        }
+
+        {
+            let mut guard = match self.links.write() {
+                Ok(g) => g,
+                Err(p) => p.into_inner(),
+            };
+            guard.retain(|(net, _), _| *net != network);
+        }
+    }
+
+    fn shutdown<'a>(&'a self) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let mut tasks = Vec::new();
+            {
+                let mut guard = match self.tasks.lock() {
+                    Ok(g) => g,
+                    Err(p) => p.into_inner(),
+                };
+                for (_, task) in guard.drain() {
+                    tasks.push(task);
+                }
+            }
+            for task in tasks {
+                task.abort();
+            }
+            self.codec.shutdown().await;
+        })
+    }
+}

crates/tsunagi-core/src/discovery.rs

@@ -12,8 +12,7 @@
 //!
 //! * *Finding members of a network* — [`NetworkDiscovery::resolve`], keyed by
 //!   the secret-derived [`DiscoveryKey`]. That is what lives here, and today
-//!   it is [`StaticBootstrap`] plus a test backend; a DHT backend is future
+//!   it includes [`StaticBootstrap`], [`MainlineDiscovery`] and a test backend.
-//!   work.
 //! * *Resolving the address of one iroh endpoint* — **iroh's job, not ours**.
 //!   With [`crate::config::TransportPolicy::N0Defaults`] or `DirectOnly`, iroh
 //!   publishes and resolves endpoint addresses through Number 0's public
@@ -23,14 +22,18 @@
 //! No empty result ever proves a network is empty. It only means "nobody found
 //! yet".
 //!
-//! Mainline DHT discovery is future work and is not implemented here.
+//! Mainline rendezvous is opt-in for library callers; the command line enables it.
+
+mod mainline;
+pub(crate) mod worker;
+pub use mainline::MainlineDiscovery;
 
 use std::collections::HashMap;
 use std::sync::{Arc, Mutex};
 
 use iroh::{EndpointAddr, EndpointId};
 
-use crate::error::Result;
+use crate::error::{Error, Result};
 use crate::identity::DiscoveryKey;
 
 pub use crate::BoxFuture;
@@ -44,6 +47,18 @@ pub enum CandidateSource {
     Discovery,
     /// An address hint restored from the disposable cache.
     Cache,
+    /// An author of a signed record: somebody who belongs to this network,
+    /// learned from state that reached us through anybody.
+    ///
+    /// It carries no address of its own — the endpoint's own discovery has
+    /// to resolve it — but knowing that a member exists is what turns a
+    /// star around whoever was named on the command line into a mesh.
+    Member,
+    /// Passed on by a member we are talking to.
+    ///
+    /// A candidate like any other: an introduction is not a vouching, and
+    /// membership is still decided by the handshake.
+    Introduced,
 }
 
 /// An unverified candidate peer.
@@ -91,6 +106,23 @@ pub trait NetworkDiscovery: Send + Sync + std::fmt::Debug + 'static {
 
     /// Returns the candidates currently known for `key`.
     fn resolve<'a>(&'a self, key: DiscoveryKey) -> BoxFuture<'a, Result<Vec<Candidate>>>;
+
+    /// Delivers candidates as they arrive. The default adapts a batch backend;
+    /// remote backends can override this so the first dial need not wait for all lookups.
+    fn resolve_into<'a>(
+        &'a self,
+        key: DiscoveryKey,
+        candidates: tokio::sync::mpsc::Sender<Candidate>,
+    ) -> BoxFuture<'a, Result<()>> {
+        Box::pin(async move {
+            for candidate in self.resolve(key).await? {
+                if candidates.send(candidate).await.is_err() {
+                    break;
+                }
+            }
+            Ok(())
+        })
+    }
 }
 
 /// A statically configured list of bootstrap candidates.
@@ -261,13 +293,25 @@ impl NetworkDiscovery for CompositeDiscovery {
 
     fn publish<'a>(&'a self, key: DiscoveryKey, addr: EndpointAddr) -> BoxFuture<'a, Result<()>> {
         Box::pin(async move {
+            let mut tasks = tokio::task::JoinSet::new();
             for backend in &self.backends {
-                // One failing backend must not stop the others.
+                let backend = backend.clone();
-                if let Err(err) = backend.publish(key, addr.clone()).await {
+                let addr = addr.clone();
-                    tracing::debug!(backend = backend.name(), %err, "publish failed");
+                tasks.spawn(async move { backend.publish(key, addr).await });
+            }
+            let mut failed = false;
+            while let Some(result) = tasks.join_next().await {
+                if !matches!(result, Ok(Ok(()))) {
+                    failed = true;
                 }
             }
-            Ok(())
+            if failed {
+                Err(Error::Discovery(
+                    "a discovery publication failed; will retry".into(),
+                ))
+            } else {
+                Ok(())
+            }
         })
     }
 
@@ -277,11 +321,12 @@ impl NetworkDiscovery for CompositeDiscovery {
         endpoint: EndpointId,
     ) -> BoxFuture<'a, Result<()>> {
         Box::pin(async move {
+            let mut tasks = tokio::task::JoinSet::new();
             for backend in &self.backends {
-                if let Err(err) = backend.unpublish(key, endpoint).await {
+                let backend = backend.clone();
-                    tracing::debug!(backend = backend.name(), %err, "unpublish failed");
+                tasks.spawn(async move { backend.unpublish(key, endpoint).await });
-                }
             }
+            while tasks.join_next().await.is_some() {}
             Ok(())
         })
     }
@@ -300,4 +345,25 @@ impl NetworkDiscovery for CompositeDiscovery {
             Ok(out)
         })
     }
+
+    fn resolve_into<'a>(
+        &'a self,
+        key: DiscoveryKey,
+        candidates: tokio::sync::mpsc::Sender<Candidate>,
+    ) -> BoxFuture<'a, Result<()>> {
+        Box::pin(async move {
+            let mut tasks = tokio::task::JoinSet::new();
+            for backend in &self.backends {
+                let backend = backend.clone();
+                let candidates = candidates.clone();
+                tasks.spawn(async move {
+                    if let Err(err) = backend.resolve_into(key, candidates).await {
+                        tracing::debug!(backend = backend.name(), %err, "resolve failed");
+                    }
+                });
+            }
+            while tasks.join_next().await.is_some() {}
+            Ok(())
+        })
+    }
 }

crates/tsunagi-core/src/discovery/mainline.rs

@@ -0,0 +1,516 @@
+//! Mainline BEP44 rendezvous. Slots are a changing sample, not a membership list.
+
+use ::mainline::{Dht, MutableItem, SigningKey, async_dht::AsyncDht};
+use futures_lite::StreamExt;
+use iroh::{EndpointAddr, EndpointId};
+use sha2::{Digest, Sha256};
+use std::collections::{HashSet, VecDeque};
+use std::net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
+use tokio::sync::{Mutex as AsyncMutex, mpsc};
+use tokio::task::JoinSet;
+use tokio::time::timeout;
+use zeroize::Zeroizing;
+
+use super::{BoxFuture, Candidate, CandidateSource, NetworkDiscovery};
+use crate::config::{
+    DHT_CLOCK_SKEW, DHT_MAX_ADDRS, DHT_MAX_RELAY_LEN, DHT_MAX_VALUE, DHT_RECORD_TTL, DHT_SLOTS,
+};
+use crate::error::{Error, Result};
+use crate::identity::{DiscoveryKey, NetworkKeys};
+
+const QUERY_TIMEOUT: Duration = Duration::from_secs(8);
+const MAGIC: &[u8; 5] = b"TSND\x01";
+
+/// One lazily started Mainline client, shared by all networks on one agent.
+/// No socket is opened until an active network publishes or looks up candidates.
+/// Agent shutdown closes this client and all its clones; a new agent needs a new client.
+#[derive(Debug, Clone, Default)]
+pub struct MainlineDiscovery {
+    client: Arc<AsyncMutex<ClientState>>,
+    allow_loopback: bool,
+}
+
+#[derive(Debug, Default)]
+enum ClientState {
+    #[default]
+    Pending,
+    Ready(AsyncDht),
+    Stopped,
+}
+
+impl MainlineDiscovery {
+    /// Uses a caller-supplied Mainline node instead of public bootstrap defaults.
+    pub fn from_dht(dht: Dht) -> Self {
+        Self {
+            client: Arc::new(AsyncMutex::new(ClientState::Ready(dht.as_async()))),
+            allow_loopback: false,
+        }
+    }
+
+    /// Creates a loopback-only client for a local Mainline Testnet.
+    #[cfg(feature = "testing")]
+    pub fn local_testnet(bootstrap: &[String]) -> Result<Self> {
+        // Reject anything that could resolve or send outside loopback.
+        if bootstrap.is_empty()
+            || bootstrap.iter().any(|s| {
+                s.parse::<SocketAddr>()
+                    .map_or(true, |a| !a.ip().is_loopback())
+            })
+        {
+            return Err(Error::Discovery(
+                "test bootstrap must contain loopback sockets".into(),
+            ));
+        }
+        let dht = Dht::builder()
+            .bootstrap(bootstrap)
+            .bind_address(Ipv4Addr::LOCALHOST)
+            .port(0)
+            .request_timeout(Duration::from_millis(200))
+            .build()
+            .map_err(|_| Error::Discovery("cannot bind local DHT".into()))?;
+        Ok(Self {
+            allow_loopback: true,
+            ..Self::from_dht(dht)
+        })
+    }
+
+    /// Binds this client's rendezvous backend to one network's independent keys.
+    pub fn for_network(&self, keys: &NetworkKeys) -> Arc<dyn NetworkDiscovery> {
+        Arc::new(NetworkDht {
+            client: self.clone(),
+            key: keys.discovery_key(),
+            seed: Zeroizing::new(*keys.dht_write_key()),
+            observed: Mutex::new(VecDeque::new()),
+        })
+    }
+
+    async fn client(&self) -> Result<AsyncDht> {
+        let mut client = self.client.lock().await;
+        match &*client {
+            ClientState::Ready(dht) => return Ok(dht.clone()),
+            ClientState::Stopped => return Err(Error::Discovery("DHT client is stopped".into())),
+            ClientState::Pending => {}
+        }
+        // Construction may resolve bootstrap hostnames. Keep it off Tokio's
+        // executor; a construction failure is retried on the next operation.
+        let dht = tokio::task::spawn_blocking(|| Dht::builder().port(0).build())
+            .await
+            .map_err(|_| Error::Discovery("DHT startup task failed".into()))?
+            .map(Dht::as_async)
+            .map_err(|_| Error::Discovery("cannot start DHT client".into()))?;
+        *client = ClientState::Ready(dht.clone());
+        Ok(dht)
+    }
+
+    pub(crate) async fn shutdown(&self) {
+        *self.client.lock().await = ClientState::Stopped;
+    }
+}
+
+struct NetworkDht {
+    client: MainlineDiscovery,
+    key: DiscoveryKey,
+    seed: Zeroizing<[u8; 32]>,
+    // Records encountered while publishing. Only an explicit lookup consumes
+    // these hints: publication never causes dials on a connected network.
+    observed: Mutex<VecDeque<Vec<u8>>>,
+}
+
+impl std::fmt::Debug for NetworkDht {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("NetworkDht(<redacted>)")
+    }
+}
+
+fn salt(slot: u8) -> Vec<u8> {
+    let mut salt = b"tsunagi-rendezvous-v1".to_vec();
+    salt.push(slot);
+    salt
+}
+
+fn slots_for(id: EndpointId) -> [u8; 2] {
+    let hash = Sha256::digest(id.as_bytes());
+    let first = hash[0] % DHT_SLOTS;
+    [first, (first + 1 + hash[1] % (DHT_SLOTS - 1)) % DHT_SLOTS]
+}
+
+fn now() -> Result<u64> {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_secs())
+        .map_err(|_| Error::Discovery("system clock predates Unix epoch".into()))
+}
+
+impl NetworkDht {
+    async fn publish_slot(&self, dht: &AsyncDht, slot: u8, value: &[u8]) -> Result<()> {
+        let salt = salt(slot);
+        let signer = SigningKey::from_bytes(&self.seed);
+        let public = signer.verifying_key().to_bytes();
+        for attempt in 0..3 {
+            let recent = timeout(
+                QUERY_TIMEOUT,
+                dht.get_mutable_most_recent(&public, Some(&salt)),
+            )
+            .await
+            .map_err(|_| Error::Discovery("DHT read timed out".into()))?;
+            if let Some(item) = &recent
+                && item.value().len() <= DHT_MAX_VALUE
+                && let Ok(mut observed) = self.observed.lock()
+            {
+                if observed.len() == DHT_SLOTS as usize {
+                    observed.pop_front();
+                }
+                observed.push_back(item.value().to_vec());
+            }
+            let seq = recent
+                .as_ref()
+                .map_or(0, MutableItem::seq)
+                .checked_add(1)
+                .ok_or_else(|| Error::Discovery("DHT sequence exhausted".into()))?;
+            let item = MutableItem::new(signer.clone(), value, seq, Some(&salt));
+            match timeout(
+                QUERY_TIMEOUT,
+                dht.put_mutable(item, recent.as_ref().map(MutableItem::seq)),
+            )
+            .await
+            {
+                Ok(Ok(outcome)) if outcome.stored_at > 0 => return Ok(()),
+                _ if attempt < 2 => {
+                    // CAS is per storage node, not a global lock. Read again on
+                    // conflicts; another writer's entry is an acceptable sample.
+                    tokio::time::sleep(Duration::from_millis(rand::random_range(30..150))).await;
+                }
+                _ => break,
+            }
+        }
+        Err(Error::Discovery(
+            "DHT publication was not acknowledged".into(),
+        ))
+    }
+}
+
+impl NetworkDiscovery for NetworkDht {
+    fn name(&self) -> &str {
+        "mainline"
+    }
+
+    fn publish<'a>(&'a self, key: DiscoveryKey, addr: EndpointAddr) -> BoxFuture<'a, Result<()>> {
+        Box::pin(async move {
+            if key != self.key {
+                return Err(Error::Discovery("wrong rendezvous network".into()));
+            }
+            let value = encode_record(&addr, now()?, self.client.allow_loopback)?;
+            let dht = self.client.client().await?;
+            let [first, second] = slots_for(addr.id);
+            let (a, b) = tokio::join!(
+                self.publish_slot(&dht, first, &value),
+                self.publish_slot(&dht, second, &value)
+            );
+            a.and(b)
+        })
+    }
+
+    fn unpublish<'a>(&'a self, _: DiscoveryKey, _: EndpointId) -> BoxFuture<'a, Result<()>> {
+        // Stopping publication lets freshness expire. Deleting a shared slot
+        // could erase a different writer, and BEP44 has no delete operation.
+        Box::pin(async { Ok(()) })
+    }
+
+    fn resolve<'a>(&'a self, key: DiscoveryKey) -> BoxFuture<'a, Result<Vec<Candidate>>> {
+        Box::pin(async move {
+            let (tx, mut rx) = mpsc::channel(DHT_SLOTS as usize);
+            let collect = async move {
+                let mut found = Vec::new();
+                while let Some(candidate) = rx.recv().await {
+                    found.push(candidate);
+                }
+                found
+            };
+            let (result, found) = tokio::join!(self.resolve_into(key, tx), collect);
+            result?;
+            Ok(found)
+        })
+    }
+
+    fn resolve_into<'a>(
+        &'a self,
+        key: DiscoveryKey,
+        candidates: mpsc::Sender<Candidate>,
+    ) -> BoxFuture<'a, Result<()>> {
+        Box::pin(async move {
+            if key != self.key {
+                return Err(Error::Discovery("wrong rendezvous network".into()));
+            }
+            let dht = self.client.client().await?;
+            let public = SigningKey::from_bytes(&self.seed)
+                .verifying_key()
+                .to_bytes();
+            let observed = self
+                .observed
+                .lock()
+                .map(|mut values| std::mem::take(&mut *values))
+                .unwrap_or_default();
+            let mut seen = HashSet::new();
+            for value in observed {
+                if let Some(addr) = decode_record(&value, now()?, self.client.allow_loopback)
+                    && seen.insert(addr.id)
+                    && candidates
+                        .send(Candidate::new(addr, CandidateSource::Discovery))
+                        .await
+                        .is_err()
+                {
+                    return Ok(());
+                }
+            }
+            let mut slots: Vec<_> = (0..DHT_SLOTS).collect();
+            for i in (1..slots.len()).rev() {
+                slots.swap(i, rand::random_range(0..=i));
+            }
+            let mut queries = JoinSet::new();
+            loop {
+                if seen.len() >= DHT_SLOTS as usize {
+                    break;
+                }
+                while queries.len() < 4 {
+                    let Some(slot) = slots.pop() else {
+                        break;
+                    };
+                    let dht = dht.clone();
+                    let allow_loopback = self.client.allow_loopback;
+                    queries.spawn(async move {
+                        timeout(QUERY_TIMEOUT, async move {
+                            let salt = salt(slot);
+                            let target = MutableItem::target_from_key(&public, Some(&salt));
+                            let mut items = dht.get_mutable(&public, Some(&salt), None);
+                            while let Some(item) = items.next().await {
+                                // Mainline verifies the signature; also bind the
+                                // returned item to the key and slot we requested.
+                                if item.key() != &public || item.target() != &target {
+                                    continue;
+                                }
+                                if let Some(addr) =
+                                    decode_record(item.value(), now().ok()?, allow_loopback)
+                                {
+                                    return Some(addr);
+                                }
+                            }
+                            None
+                        })
+                        .await
+                        .ok()
+                        .flatten()
+                    });
+                }
+                let Some(result) = queries.join_next().await else {
+                    break;
+                };
+                if let Ok(Some(addr)) = result
+                    && seen.insert(addr.id)
+                    && candidates
+                        .send(Candidate::new(addr, CandidateSource::Discovery))
+                        .await
+                        .is_err()
+                {
+                    break;
+                }
+            }
+            Ok(())
+        })
+    }
+}
+
+fn usable(addr: &SocketAddr, allow_loopback: bool) -> bool {
+    addr.port() != 0
+        && !addr.ip().is_unspecified()
+        && !addr.ip().is_multicast()
+        && (allow_loopback || !addr.ip().is_loopback())
+        && match addr.ip() {
+            IpAddr::V4(ip) => !ip.is_broadcast() && !ip.is_link_local(),
+            IpAddr::V6(ip) => !ip.is_unicast_link_local(),
+        }
+}
+
+fn encode_record(addr: &EndpointAddr, timestamp: u64, allow_loopback: bool) -> Result<Vec<u8>> {
+    let ips: Vec<_> = addr
+        .ip_addrs()
+        .filter(|ip| usable(ip, allow_loopback))
+        .take(DHT_MAX_ADDRS)
+        .collect();
+    let relay = addr
+        .relay_urls()
+        .next()
+        .map(ToString::to_string)
+        .unwrap_or_default();
+    if relay.len() > DHT_MAX_RELAY_LEN || (ips.is_empty() && relay.is_empty()) {
+        return Err(Error::Discovery("no encodable endpoint address yet".into()));
+    }
+    let mut bytes = Vec::with_capacity(DHT_MAX_VALUE);
+    bytes.extend_from_slice(MAGIC);
+    bytes.extend_from_slice(&timestamp.to_be_bytes());
+    bytes.extend_from_slice(addr.id.as_bytes());
+    bytes.extend_from_slice(&(relay.len() as u16).to_be_bytes());
+    bytes.extend_from_slice(relay.as_bytes());
+    bytes.push(ips.len() as u8);
+    for ip in ips {
+        match ip.ip() {
+            IpAddr::V4(v) => {
+                bytes.push(4);
+                bytes.extend_from_slice(&v.octets());
+            }
+            IpAddr::V6(v) => {
+                bytes.push(6);
+                bytes.extend_from_slice(&v.octets());
+            }
+        }
+        bytes.extend_from_slice(&ip.port().to_be_bytes());
+    }
+    if bytes.len() > DHT_MAX_VALUE {
+        return Err(Error::Discovery("DHT record too large".into()));
+    }
+    Ok(bytes)
+}
+
+fn take<'a>(input: &mut &'a [u8], n: usize) -> Option<&'a [u8]> {
+    let (head, rest) = input.split_at_checked(n)?;
+    *input = rest;
+    Some(head)
+}
+
+fn decode_record(mut bytes: &[u8], now: u64, allow_loopback: bool) -> Option<EndpointAddr> {
+    if bytes.len() > DHT_MAX_VALUE || take(&mut bytes, MAGIC.len())? != MAGIC {
+        return None;
+    }
+    let timestamp = u64::from_be_bytes(take(&mut bytes, 8)?.try_into().ok()?);
+    if timestamp > now.saturating_add(DHT_CLOCK_SKEW.as_secs())
+        || now.saturating_sub(timestamp) >= DHT_RECORD_TTL.as_secs()
+    {
+        return None;
+    }
+    let id = EndpointId::from_bytes(take(&mut bytes, 32)?.try_into().ok()?).ok()?;
+    let mut addr = EndpointAddr::new(id);
+    let len = u16::from_be_bytes(take(&mut bytes, 2)?.try_into().ok()?) as usize;
+    if len > DHT_MAX_RELAY_LEN {
+        return None;
+    }
+    let relay = std::str::from_utf8(take(&mut bytes, len)?).ok()?;
+    if !relay.is_empty() {
+        addr = addr.with_relay_url(relay.parse().ok()?);
+    }
+    let count = *take(&mut bytes, 1)?.first()? as usize;
+    if count > DHT_MAX_ADDRS {
+        return None;
+    }
+    for _ in 0..count {
+        let ip = match *take(&mut bytes, 1)?.first()? {
+            4 => IpAddr::V4(Ipv4Addr::from(
+                <[u8; 4]>::try_from(take(&mut bytes, 4)?).ok()?,
+            )),
+            6 => IpAddr::V6(Ipv6Addr::from(
+                <[u8; 16]>::try_from(take(&mut bytes, 16)?).ok()?,
+            )),
+            _ => return None,
+        };
+        let port = u16::from_be_bytes(take(&mut bytes, 2)?.try_into().ok()?);
+        let socket = SocketAddr::new(ip, port);
+        if usable(&socket, allow_loopback) {
+            addr = addr.with_ip_addr(socket);
+        }
+    }
+    (bytes.is_empty() && !addr.addrs.is_empty()).then_some(addr)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use crate::testing::network;
+
+    fn address() -> EndpointAddr {
+        EndpointAddr::new(iroh::SecretKey::generate().public())
+            .with_ip_addr("127.0.0.1:12345".parse().unwrap())
+    }
+
+    #[test]
+    fn records_reject_expiration_future_dates_truncation_and_oversize() {
+        let addr = address();
+        let bytes = encode_record(&addr, 1_000, true).unwrap();
+        assert_eq!(decode_record(&bytes, 1_100, true).unwrap(), addr);
+        assert!(decode_record(&bytes, 1_901, true).is_none());
+        assert!(decode_record(&bytes, 100, true).is_none());
+        for n in 0..bytes.len() {
+            assert!(decode_record(&bytes[..n], 1_100, true).is_none());
+        }
+        assert!(decode_record(&vec![0; DHT_MAX_VALUE + 1], 1_100, true).is_none());
+        assert!(encode_record(&addr, 1_000, false).is_err());
+        assert!(decode_record(&bytes, 1_100, false).is_none());
+    }
+
+    #[tokio::test]
+    async fn a_stopped_custom_client_never_falls_back_to_public_bootstrap() {
+        let net = mainline::Testnet::builder(3).build().unwrap();
+        let client = MainlineDiscovery::local_testnet(&net.bootstrap).unwrap();
+        let (name, secret) = network("mainline-shutdown");
+        let keys = NetworkKeys::derive(&name, &secret);
+        let backend = client.for_network(&keys);
+        client.shutdown().await;
+        assert!(backend.resolve(keys.discovery_key()).await.is_err());
+    }
+
+    #[tokio::test]
+    async fn concurrent_writers_share_slots_and_other_secrets_find_nothing() {
+        let net = mainline::Testnet::builder(5).build().unwrap();
+        let client_a = MainlineDiscovery::local_testnet(&net.bootstrap).unwrap();
+        let client_b = MainlineDiscovery::local_testnet(&net.bootstrap).unwrap();
+        let (name, secret) = network("mainline-slots");
+        let keys = NetworkKeys::derive(&name, &secret);
+        let a = client_a.for_network(&keys);
+        let b = client_b.for_network(&keys);
+        let first = address();
+        let second = loop {
+            let next = address();
+            if slots_for(next.id) == slots_for(first.id) {
+                break next;
+            }
+        };
+        let key = keys.discovery_key();
+        let (ra, rb) = tokio::join!(
+            a.publish(key, first.clone()),
+            b.publish(key, second.clone())
+        );
+        assert!(ra.is_ok() || rb.is_ok());
+        let (found_a, found_b) = tokio::join!(a.resolve(key), b.resolve(key));
+        let found_a = found_a.unwrap();
+        let found_b = found_b.unwrap();
+        assert!(
+            found_a.iter().any(|c| c.addr == second) || found_b.iter().any(|c| c.addr == first),
+            "at least one writer must discover the other despite colliding in both slots"
+        );
+        assert!(
+            found_a
+                .iter()
+                .chain(&found_b)
+                .all(|c| c.addr == first || c.addr == second)
+        );
+        let other = NetworkKeys::derive(&name, &crate::identity::NetworkSecret::generate());
+        assert!(
+            client_a
+                .for_network(&other)
+                .resolve(other.discovery_key())
+                .await
+                .unwrap()
+                .is_empty()
+        );
+        // Restarting the publisher must read seq from DHT instead of starting at 1.
+        let restarted = client_a.for_network(&keys);
+        restarted.publish(key, first.clone()).await.unwrap();
+        assert!(
+            b.resolve(key)
+                .await
+                .unwrap()
+                .iter()
+                .any(|c| c.addr == first)
+        );
+    }
+}

crates/tsunagi-core/src/discovery/worker.rs

@@ -0,0 +1,251 @@
+//! Per-network discovery tasks. Dropping the owner cancels all backend futures.
+
+use std::sync::Arc;
+use std::time::Duration;
+use tokio::sync::{mpsc, watch};
+use tokio::task::JoinSet;
+use tokio::time::{Instant, timeout};
+
+use super::{Candidate, NetworkDiscovery};
+use crate::config::DiscoveryPolicy;
+use crate::identity::DiscoveryKey;
+use crate::net::EndpointAdapter;
+
+pub(crate) struct DiscoveryWorker {
+    connected: watch::Sender<Connectivity>,
+    tasks: JoinSet<()>,
+    backend: Arc<dyn NetworkDiscovery>,
+    key: DiscoveryKey,
+    endpoint: iroh::EndpointId,
+}
+
+#[derive(Clone, Copy)]
+enum Connectivity {
+    Initial,
+    Connected,
+    Isolated(Instant),
+}
+
+impl DiscoveryWorker {
+    pub(crate) fn spawn(
+        backend: Arc<dyn NetworkDiscovery>,
+        key: DiscoveryKey,
+        adapter: EndpointAdapter,
+        policy: DiscoveryPolicy,
+        address_check: Duration,
+        candidates: mpsc::Sender<Candidate>,
+    ) -> Self {
+        let (connected, receiver) = watch::channel(Connectivity::Initial);
+        let mut tasks = JoinSet::new();
+        tasks.spawn(publish_loop(
+            backend.clone(),
+            key,
+            adapter.clone(),
+            policy.clone(),
+            address_check,
+        ));
+        tasks.spawn(lookup_loop(
+            backend.clone(),
+            key,
+            policy,
+            receiver,
+            candidates,
+        ));
+        Self {
+            connected,
+            tasks,
+            backend,
+            key,
+            endpoint: adapter.endpoint_id(),
+        }
+    }
+
+    pub(crate) fn set_connected(&self, connected: bool) {
+        self.connected.send_if_modified(|current| {
+            if matches!(*current, Connectivity::Connected) == connected {
+                return false;
+            }
+            *current = if connected {
+                Connectivity::Connected
+            } else {
+                Connectivity::Isolated(Instant::now())
+            };
+            true
+        });
+    }
+
+    pub(crate) async fn stop(mut self) {
+        self.tasks.abort_all();
+        while self.tasks.join_next().await.is_some() {}
+        // BEP44 has no deletion. Memory/static backends can withdraw promptly;
+        // an unresponsive backend must not extend agent shutdown.
+        let _ = timeout(
+            Duration::from_millis(250),
+            self.backend.unpublish(self.key, self.endpoint),
+        )
+        .await;
+    }
+}
+
+fn jitter(delay: Duration) -> Duration {
+    delay
+        .mul_f64(0.8 + 0.4 * rand::random::<f64>())
+        .max(Duration::from_millis(10))
+}
+
+async fn publish_loop(
+    backend: Arc<dyn NetworkDiscovery>,
+    key: DiscoveryKey,
+    adapter: EndpointAdapter,
+    policy: DiscoveryPolicy,
+    address_check: Duration,
+) {
+    let mut previous = None;
+    let mut due = Instant::now();
+    let mut ticker = tokio::time::interval(address_check.max(Duration::from_millis(10)));
+    ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
+    loop {
+        ticker.tick().await;
+        let mut addr = adapter.addr();
+        addr.addrs.extend(adapter.loopback_addr().addrs);
+        if previous.as_ref() != Some(&addr) {
+            previous = Some(addr.clone());
+            due = Instant::now();
+        }
+        if Instant::now() < due {
+            continue;
+        }
+        let result = timeout(policy.request_timeout, backend.publish(key, addr.clone())).await;
+        match result {
+            Ok(Ok(())) => {
+                due = Instant::now() + jitter(policy.publish_interval);
+            }
+            result => {
+                tracing::debug!(?result, "discovery publication failed; will retry");
+                due = Instant::now() + jitter(policy.lookup_interval);
+            }
+        }
+    }
+}
+
+async fn lookup_loop(
+    backend: Arc<dyn NetworkDiscovery>,
+    key: DiscoveryKey,
+    policy: DiscoveryPolicy,
+    mut connected: watch::Receiver<Connectivity>,
+    candidates: mpsc::Sender<Candidate>,
+) {
+    let mut last_isolation = None;
+    let mut due = Instant::now();
+    let mut delay = policy.lookup_interval;
+    loop {
+        let state = *connected.borrow_and_update();
+        if matches!(state, Connectivity::Connected) {
+            if connected.changed().await.is_err() {
+                break;
+            }
+            continue;
+        }
+        if let Connectivity::Isolated(since) = state
+            && last_isolation != Some(since)
+        {
+            // Preserve the transition time even if a rapid reconnect/disconnect
+            // overwrote a watch value before this task had a chance to run.
+            due = since + policy.reconnect_delay;
+            delay = policy.lookup_interval;
+            last_isolation = Some(since);
+        }
+        tokio::select! {
+            biased;
+            changed = connected.changed() => {
+                if changed.is_err() { break; }
+                continue;
+            }
+            _ = tokio::time::sleep_until(due) => {}
+        }
+        tokio::select! {
+            biased;
+            changed = connected.changed() => {
+                if changed.is_err() { break; }
+                continue;
+            }
+            result = timeout(policy.request_timeout, backend.resolve_into(key, candidates.clone())) => {
+                if !matches!(result, Ok(Ok(()))) {
+                    tracing::debug!(?result, "discovery lookup failed; will retry");
+                }
+            }
+        }
+        due = Instant::now() + jitter(delay);
+        delay = delay.saturating_mul(2).min(policy.max_lookup_interval);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use crate::Result;
+    use crate::discovery::BoxFuture;
+    use std::sync::atomic::{AtomicUsize, Ordering};
+
+    #[derive(Debug, Default)]
+    struct Reads(AtomicUsize);
+    impl NetworkDiscovery for Reads {
+        fn name(&self) -> &str {
+            "reads"
+        }
+        fn publish<'a>(
+            &'a self,
+            _: DiscoveryKey,
+            _: iroh::EndpointAddr,
+        ) -> BoxFuture<'a, Result<()>> {
+            Box::pin(async { Ok(()) })
+        }
+        fn unpublish<'a>(
+            &'a self,
+            _: DiscoveryKey,
+            _: iroh::EndpointId,
+        ) -> BoxFuture<'a, Result<()>> {
+            Box::pin(async { Ok(()) })
+        }
+        fn resolve<'a>(&'a self, _: DiscoveryKey) -> BoxFuture<'a, Result<Vec<Candidate>>> {
+            self.0.fetch_add(1, Ordering::SeqCst);
+            Box::pin(async { Ok(Vec::new()) })
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn recovery_requires_a_full_minute_and_reconnection_restarts_that_minute() {
+        let reads = Arc::new(Reads::default());
+        let (tx, rx) = watch::channel(Connectivity::Connected);
+        let (candidates, _receiver) = mpsc::channel(16);
+        let mut tasks = JoinSet::new();
+        tasks.spawn(lookup_loop(
+            reads.clone(),
+            DiscoveryKey::from_bytes([0; 32]),
+            DiscoveryPolicy::default(),
+            rx,
+            candidates,
+        ));
+        tokio::task::yield_now().await;
+        tx.send(Connectivity::Isolated(Instant::now())).unwrap();
+        tokio::task::yield_now().await;
+        tokio::time::advance(Duration::from_secs(59)).await;
+        tokio::task::yield_now().await;
+        assert_eq!(reads.0.load(Ordering::SeqCst), 0);
+        tx.send(Connectivity::Connected).unwrap();
+        tx.send(Connectivity::Isolated(Instant::now())).unwrap();
+        tokio::task::yield_now().await;
+        tokio::time::advance(Duration::from_secs(59)).await;
+        tokio::task::yield_now().await;
+        assert_eq!(reads.0.load(Ordering::SeqCst), 0);
+        tokio::time::advance(Duration::from_secs(1)).await;
+        tokio::task::yield_now().await;
+        assert_eq!(reads.0.load(Ordering::SeqCst), 1);
+        tx.send(Connectivity::Connected).unwrap();
+        tokio::task::yield_now().await;
+        tokio::time::advance(Duration::from_secs(600)).await;
+        tokio::task::yield_now().await;
+        assert_eq!(reads.0.load(Ordering::SeqCst), 1);
+    }
+}

crates/tsunagi-core/src/dns/mod.rs

@@ -0,0 +1,84 @@
+//! A DNS view of the overlay.
+//!
+//! The names and addresses of a network, served to the host that runs the
+//! agent, so members can be reached by name. Answered from signed state, so a
+//! member that is switched off still resolves.
+//!
+//! Three parts, kept apart on purpose:
+//!
+//! * [`zone`] decides what the answer is. Pure, and knows nothing about
+//!   packets or sockets.
+//! * [`server`] puts that on the wire.
+//! * `publish` tells the operating system where to send its questions,
+//!   which is the only part that differs between platforms.
+
+pub mod publish;
+pub mod server;
+pub mod zone;
+
+pub use publish::{DnsPublisher, PublishError, Published};
+
+use std::net::{Ipv4Addr, Ipv6Addr, SocketAddr};
+
+/// Where the server listens, one list per address family.
+///
+/// Loopback, and only loopback. The zones are a view *for the host running
+/// the agent*: it is that host's resolver that is pointed at them. Binding
+/// an overlay address instead would put them in front of the whole mesh —
+/// and an agent in several networks would then answer one network's
+/// questions about another's names, which is precisely what separate
+/// networks are for.
+///
+/// Both families, and independently: a question arrives over whichever one
+/// the resolver happens to use, and on a host with one of them switched off
+/// the other must still answer. So neither list failing says anything about
+/// the other, and a listener is opened from each.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ListenPlan {
+    /// IPv4 candidates, best first.
+    pub v4: Vec<SocketAddr>,
+    /// IPv6 candidates, best first.
+    pub v6: Vec<SocketAddr>,
+}
+
+impl ListenPlan {
+    /// The two lists, so a caller can bind one listener from each.
+    pub fn families(&self) -> [&[SocketAddr]; 2] {
+        [&self.v4, &self.v6]
+    }
+}
+
+/// Builds the listen plan for one port.
+pub fn listen_plan(port: u16) -> ListenPlan {
+    ListenPlan {
+        v4: vec![SocketAddr::from((Ipv4Addr::LOCALHOST, port))],
+        v6: vec![SocketAddr::from((Ipv6Addr::LOCALHOST, port))],
+    }
+}
+
+pub use server::{DnsServer, SharedZone};
+pub use zone::{Answer, Query, Zone, ZoneError, ZoneName, Zones};
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[test]
+    fn both_families_are_planned_for_and_neither_is_reachable_off_this_host() {
+        let plan = listen_plan(5354);
+        assert_eq!(
+            plan.v4,
+            vec!["127.0.0.1:5354".parse::<SocketAddr>().unwrap()]
+        );
+        assert_eq!(plan.v6, vec!["[::1]:5354".parse::<SocketAddr>().unwrap()]);
+        assert_eq!(plan.families().len(), 2);
+        // Nothing here is reachable from another member: the zones are a
+        // view for this host, and an agent in two networks must not answer
+        // one network's questions about the other's names.
+        for address in plan.families().concat() {
+            assert!(address.ip().is_loopback(), "{address} is not loopback");
+        }
+    }
+}

crates/tsunagi-core/src/dns/publish/macos.rs

@@ -0,0 +1,333 @@
+//! Telling the macOS resolver to send some questions here, through
+//! `/etc/resolver`.
+//!
+//! macOS reads one file per domain out of `/etc/resolver`: a file named after
+//! a domain, containing the `nameserver` and `port` to send that domain's
+//! questions to, makes the system resolver a *scoped* resolver for exactly
+//! that suffix and nothing else (see `resolver(5)`). That is the same contract
+//! the systemd-resolved and NRPT publishers arrange on the other platforms:
+//! route these suffixes to this server, and claim nothing else.
+//!
+//! # Why a port, and so no privileged socket
+//!
+//! A resolver file carries a `port`, so the server does not have to sit on 53
+//! and the agent needs nothing like `CAP_NET_BIND_SERVICE`. It can serve on a
+//! high port and still be reached for its suffixes.
+//!
+//! # Tagged, reversible, and never clobbering
+//!
+//! Every file this publisher writes begins with a marker line, so it only ever
+//! replaces or removes files it wrote itself: a pre-existing `/etc/resolver`
+//! entry a user set up by hand is left untouched. The files are removed when
+//! the resolver is turned off and on shutdown; a crash leaves them behind, and
+//! the next apply for the same domains overwrites them.
+//!
+//! # Privilege
+//!
+//! Writing under `/etc/resolver` needs root. A permission error is reported as
+//! a refusal — the kind of failure a person has to act on — rather than a
+//! transient one, so the agent does not retry it at the ordinary pace.
+
+use std::collections::BTreeSet;
+use std::net::IpAddr;
+use std::path::{Path, PathBuf};
+use std::sync::Mutex;
+
+use crate::BoxFuture;
+
+use super::{DnsPublisher, PublishError, Published};
+
+/// The directory macOS reads scoped resolvers from.
+const RESOLVER_DIR: &str = "/etc/resolver";
+
+/// The first line of every file this publisher writes, so it recognises its
+/// own and never touches a file a user created.
+const MARKER: &str = "# managed by tsunagi — do not edit";
+
+/// Writes and removes `/etc/resolver` files for the overlay zones.
+#[derive(Debug)]
+pub struct ResolverDirPublisher {
+    /// The resolver directory. A field rather than a constant so the tests can
+    /// point it at a temporary directory and never touch the real one.
+    dir: PathBuf,
+    /// The domain files written by the last apply, so a later apply can remove
+    /// the ones no longer wanted and `revert` can remove them all.
+    written: Mutex<BTreeSet<String>>,
+}
+
+impl Default for ResolverDirPublisher {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl ResolverDirPublisher {
+    /// A publisher writing to the real `/etc/resolver`.
+    pub fn new() -> Self {
+        Self::in_dir(RESOLVER_DIR)
+    }
+
+    /// A publisher writing to a specific directory. For the tests.
+    fn in_dir(dir: impl Into<PathBuf>) -> Self {
+        Self {
+            dir: dir.into(),
+            written: Mutex::new(BTreeSet::new()),
+        }
+    }
+
+    fn apply_inner(&self, published: &Published) -> Result<(), PublishError> {
+        let desired = desired_files(published)?;
+
+        // The directory may not exist yet; creating it needs root.
+        std::fs::create_dir_all(&self.dir).map_err(|err| classify(&self.dir, err))?;
+
+        // Remove the files from a previous apply that are no longer wanted.
+        let previous = lock(&self.written).clone();
+        for name in previous.difference(&keys(&desired)) {
+            remove_if_ours(&self.dir.join(name));
+        }
+
+        for (name, body) in &desired {
+            let path = self.dir.join(name);
+            // Only ever overwrite a file that is already ours, or a name that
+            // is free. A resolver a user set up by hand is left in place and
+            // reported, never clobbered.
+            if path.exists() && !is_ours(&path) {
+                return Err(PublishError::Failed(format!(
+                    "{} already exists and was not created by tsunagi; \
+                     remove it or rename the network to publish its resolver",
+                    path.display()
+                )));
+            }
+            std::fs::write(&path, body).map_err(|err| classify(&path, err))?;
+        }
+
+        *lock(&self.written) = keys(&desired);
+        Ok(())
+    }
+
+    fn revert_inner(&self) {
+        let written = std::mem::take(&mut *lock(&self.written));
+        for name in &written {
+            remove_if_ours(&self.dir.join(name));
+        }
+    }
+}
+
+impl DnsPublisher for ResolverDirPublisher {
+    fn name(&self) -> &str {
+        "resolver-dir"
+    }
+
+    fn apply<'a>(&'a self, published: &'a Published) -> BoxFuture<'a, Result<(), PublishError>> {
+        // A handful of small file writes, short enough to run in place rather
+        // than hop to the blocking pool.
+        Box::pin(async move { self.apply_inner(published) })
+    }
+
+    fn revert(&self) -> BoxFuture<'_, Result<(), PublishError>> {
+        Box::pin(async move {
+            self.revert_inner();
+            Ok(())
+        })
+    }
+}
+
+/// The resolver files a publication implies: a map of file name to contents.
+fn desired_files(published: &Published) -> Result<Vec<(String, String)>, PublishError> {
+    if published.servers.is_empty() {
+        return Err(PublishError::Failed(
+            "no server address to point the resolver at".to_string(),
+        ));
+    }
+    let body = file_contents(published);
+    let mut files = Vec::new();
+    for domain in &published.domains {
+        let name = resolver_file_name(domain)?;
+        files.push((name, body.clone()));
+    }
+    Ok(files)
+}
+
+/// The contents of one resolver file: the marker, every server, one port.
+fn file_contents(published: &Published) -> String {
+    let mut out = String::from(MARKER);
+    out.push('\n');
+    for server in &published.servers {
+        let ip = match server.ip() {
+            IpAddr::V4(v4) => v4.to_string(),
+            IpAddr::V6(v6) => v6.to_string(),
+        };
+        out.push_str("nameserver ");
+        out.push_str(&ip);
+        out.push('\n');
+    }
+    // One port for the file; every server listens on the same one. The `port`
+    // directive is what lets the server stay off 53.
+    if let Some(first) = published.servers.first() {
+        out.push_str("port ");
+        out.push_str(&first.port().to_string());
+        out.push('\n');
+    }
+    out
+}
+
+/// The file name for a domain, rejecting anything that is not a plain label
+/// path component: a resolver file name must never escape the directory.
+fn resolver_file_name(domain: &str) -> Result<String, PublishError> {
+    let trimmed = domain.trim_end_matches('.');
+    if trimmed.is_empty()
+        || trimmed.contains('/')
+        || trimmed.contains('\\')
+        || trimmed.contains(std::path::MAIN_SEPARATOR)
+        || trimmed == "."
+        || trimmed == ".."
+        || trimmed.starts_with('.')
+    {
+        return Err(PublishError::Failed(format!(
+            "cannot make a resolver file name for the domain {domain:?}"
+        )));
+    }
+    Ok(trimmed.to_string())
+}
+
+/// Whether a file is one this publisher wrote, by its marker first line.
+fn is_ours(path: &Path) -> bool {
+    match std::fs::read_to_string(path) {
+        Ok(text) => text.lines().next() == Some(MARKER),
+        Err(_) => false,
+    }
+}
+
+/// Removes a file, but only if it is one of ours. Best effort.
+fn remove_if_ours(path: &Path) {
+    if is_ours(path)
+        && let Err(err) = std::fs::remove_file(path)
+        && err.kind() != std::io::ErrorKind::NotFound
+    {
+        tracing::debug!(path = %path.display(), %err, "cannot remove the resolver file");
+    }
+}
+
+/// The set of file names in a desired-files list.
+fn keys(files: &[(String, String)]) -> BTreeSet<String> {
+    files.iter().map(|(name, _)| name.clone()).collect()
+}
+
+/// Turns a filesystem error into the right kind of publish error: a permission
+/// problem is a refusal a person must act on, anything else is a plain failure.
+fn classify(path: &Path, err: std::io::Error) -> PublishError {
+    if err.kind() == std::io::ErrorKind::PermissionDenied {
+        PublishError::Refused(format!(
+            "writing {} needs root; start the agent with `sudo` or as a root LaunchDaemon",
+            path.display()
+        ))
+    } else {
+        PublishError::Failed(format!("cannot write {}: {err}", path.display()))
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::net::SocketAddr;
+
+    use super::*;
+
+    fn published(domains: &[&str]) -> Published {
+        Published {
+            interface: "utun5".into(),
+            servers: vec![
+                "10.13.37.1:5353".parse::<SocketAddr>().unwrap(),
+                "[fd00::1]:5353".parse::<SocketAddr>().unwrap(),
+            ],
+            domains: domains.iter().map(|d| d.to_string()).collect(),
+        }
+    }
+
+    #[test]
+    fn a_file_lists_every_server_and_one_port_under_the_marker() {
+        let body = file_contents(&published(&["mynet."]));
+        let mut lines = body.lines();
+        assert_eq!(lines.next(), Some(MARKER));
+        assert_eq!(lines.next(), Some("nameserver 10.13.37.1"));
+        assert_eq!(lines.next(), Some("nameserver fd00::1"));
+        assert_eq!(lines.next(), Some("port 5353"));
+        assert_eq!(lines.next(), None);
+    }
+
+    #[test]
+    fn a_domain_becomes_a_plain_file_name_without_its_trailing_dot() {
+        assert_eq!(resolver_file_name("mynet.").unwrap(), "mynet");
+        assert_eq!(resolver_file_name("my.net").unwrap(), "my.net");
+    }
+
+    #[test]
+    fn a_domain_that_would_escape_the_directory_is_refused() {
+        assert!(resolver_file_name("../etc/passwd").is_err());
+        assert!(resolver_file_name("a/b").is_err());
+        assert!(resolver_file_name("").is_err());
+        assert!(resolver_file_name(".").is_err());
+        assert!(resolver_file_name(".hidden").is_err());
+    }
+
+    #[tokio::test]
+    async fn apply_writes_a_file_per_domain_and_revert_removes_them() {
+        let dir = tempfile::tempdir().unwrap();
+        let publisher = ResolverDirPublisher::in_dir(dir.path());
+
+        publisher
+            .apply(&published(&["mynet.", "other."]))
+            .await
+            .unwrap();
+        assert!(dir.path().join("mynet").is_file());
+        assert!(dir.path().join("other").is_file());
+
+        publisher.revert().await.unwrap();
+        assert!(!dir.path().join("mynet").exists());
+        assert!(!dir.path().join("other").exists());
+    }
+
+    #[tokio::test]
+    async fn re_applying_removes_a_domain_that_is_no_longer_wanted() {
+        let dir = tempfile::tempdir().unwrap();
+        let publisher = ResolverDirPublisher::in_dir(dir.path());
+
+        publisher.apply(&published(&["a.", "b."])).await.unwrap();
+        publisher.apply(&published(&["a."])).await.unwrap();
+
+        assert!(dir.path().join("a").is_file());
+        assert!(
+            !dir.path().join("b").exists(),
+            "the dropped domain's file is removed"
+        );
+    }
+
+    #[tokio::test]
+    async fn a_file_a_user_created_is_never_clobbered_or_removed() {
+        let dir = tempfile::tempdir().unwrap();
+        let foreign = dir.path().join("mynet");
+        std::fs::write(&foreign, "nameserver 9.9.9.9\n").unwrap();
+
+        let publisher = ResolverDirPublisher::in_dir(dir.path());
+        let err = publisher.apply(&published(&["mynet."])).await.unwrap_err();
+        assert!(matches!(err, PublishError::Failed(_)), "{err:?}");
+        // Untouched.
+        assert_eq!(
+            std::fs::read_to_string(&foreign).unwrap(),
+            "nameserver 9.9.9.9\n"
+        );
+
+        // And revert leaves a foreign file alone even if it is in our set.
+        publisher.revert().await.unwrap();
+        assert!(foreign.is_file());
+    }
+}

crates/tsunagi-core/src/dns/publish/mock.rs

@@ -74,7 +74,10 @@ mod tests {
     fn published() -> Published {
         Published {
             interface: "tsundemo".into(),
-            server: "10.13.37.69:5354".parse().unwrap(),
+            servers: vec![
+                "10.13.37.69:5354".parse().unwrap(),
+                "[::1]:5354".parse().unwrap(),
+            ],
             domains: vec!["lab".into()],
         }
     }

crates/tsunagi-core/src/dns/publish/mod.rs

@@ -23,6 +23,16 @@ mod resolved;
 #[cfg(all(feature = "dns-publish", target_os = "linux"))]
 pub use resolved::ResolvedPublisher;
 
+#[cfg(all(feature = "dns-publish", target_os = "windows"))]
+mod windows;
+#[cfg(all(feature = "dns-publish", target_os = "windows"))]
+pub use windows::NrptPublisher;
+
+#[cfg(all(feature = "dns-publish", target_os = "macos"))]
+mod macos;
+#[cfg(all(feature = "dns-publish", target_os = "macos"))]
+pub use macos::ResolverDirPublisher;
+
 mod unsupported;
 pub use unsupported::UnsupportedPublisher;
 
@@ -32,10 +42,16 @@ pub struct Published {
     /// The interface questions should be sent through.
     ///
     /// The server listens on an overlay address, which is only reachable
-    /// over the overlay interface, so the two travel together.
+    /// over the overlay interface, so the two travel together. It must be an
+    /// interface this agent created: configuring one it did not is
+    /// configuring somebody else's.
     pub interface: String,
-    /// Where the server is listening.
+    /// Every address the server is listening on, best first.
-    pub server: SocketAddr,
+    ///
+    /// One per address family, so a resolver reaches it over whichever it
+    /// uses. All of them are published together and none is preferred by
+    /// this: which one a resolver picks is its business.
+    pub servers: Vec<SocketAddr>,
     /// The suffixes that belong to this server.
     ///
     /// Routing suffixes only: they say *which questions* come here, never
@@ -62,20 +78,54 @@ pub enum PublishError {
 }
 
 impl PublishError {
-    /// What the user can do about it, when there is something.
+    /// Whether waiting will fix it.
+    ///
+    /// A refusal will not change on its own — somebody has to grant
+    /// permission — so retrying it at the pace of everything else is just
+    /// noise. Anything else might be a service still starting.
+    pub fn needs_a_human(&self) -> bool {
+        matches!(self, PublishError::Refused(_))
+    }
+
+    /// One line for a status table.
     pub fn remedy(&self) -> Option<&'static str> {
         match self {
             PublishError::Refused(_) => Some(
-                "systemd-resolved asks polkit before accepting this, and polkit \
+                "grant this user the `org.freedesktop.resolve1.set-*` actions in \
-                 decides by user. Run the agent as a system service, or install a \
+                 /etc/polkit-1/rules.d, or run the agent as a system service",
-                 polkit rule allowing this user the `org.freedesktop.resolve1.set-*` \
-                 actions.",
             ),
             _ => None,
         }
     }
 }
 
+/// The polkit rule that lets this user configure the resolver.
+///
+/// Printed in full rather than described, because the point of this feature
+/// is that the user has as little to do as possible, and "write a polkit
+/// rule" is a great deal more work than pasting one.
+///
+/// It grants exactly the four actions this agent calls and nothing else.
+/// polkit decides by user id — a capability does not help here — so there is
+/// no way to do this from inside the process.
+pub fn polkit_recipe(user: &str) -> String {
+    format!(
+        "sudo tee /etc/polkit-1/rules.d/50-tsunagi-resolved.rules > /dev/null <<'RULE'\n\
+         polkit.addRule(function(action, subject) {{\n\
+         \x20   var allowed = [\n\
+         \x20       \"org.freedesktop.resolve1.set-dns-servers\",\n\
+         \x20       \"org.freedesktop.resolve1.set-domains\",\n\
+         \x20       \"org.freedesktop.resolve1.set-default-route\",\n\
+         \x20       \"org.freedesktop.resolve1.revert\"\n\
+         \x20   ];\n\
+         \x20   if (allowed.indexOf(action.id) >= 0 && subject.user == \"{user}\") {{\n\
+         \x20       return polkit.Result.YES;\n\
+         \x20   }}\n\
+         }});\n\
+         RULE"
+    )
+}
+
 /// Arranges for the operating system to ask this server.
 pub trait DnsPublisher: Send + Sync + std::fmt::Debug + 'static {
     /// A short name used in diagnostics.
@@ -103,6 +153,10 @@ pub fn interface_index(name: &str) -> Option<u32> {
         .ok()
 }
 
+/// The kernel's index for an interface.
+///
+/// Only the systemd-resolved publisher needs one, so off Linux there is
+/// nothing to look up and this is always `None`.
 #[cfg(not(target_os = "linux"))]
 pub fn interface_index(_name: &str) -> Option<u32> {
     None
@@ -122,6 +176,36 @@ mod tests {
         assert!(PublishError::Failed("bang".into()).remedy().is_none());
     }
 
+    #[test]
+    fn only_a_refusal_waits_for_a_person() {
+        // The rest may come right on their own, so they are worth retrying
+        // at the ordinary pace; a refusal is not.
+        assert!(PublishError::Refused("no".into()).needs_a_human());
+        assert!(!PublishError::Unavailable("none".into()).needs_a_human());
+        assert!(!PublishError::Failed("bang".into()).needs_a_human());
+    }
+
+    #[test]
+    fn the_polkit_recipe_grants_what_is_called_and_no_more() {
+        let recipe = polkit_recipe("ab");
+        for action in [
+            "set-dns-servers",
+            "set-domains",
+            "set-default-route",
+            "revert",
+        ] {
+            assert!(recipe.contains(action), "{action} missing from:\n{recipe}");
+        }
+        // Nothing beyond what the agent calls: a rule that granted the lot
+        // would be handing out more than this feature needs.
+        for other in ["set-dnssec", "set-mdns", "register-service", "set-llmnr"] {
+            assert!(!recipe.contains(other), "{other} should not be granted");
+        }
+        assert!(recipe.contains("subject.user == \"ab\""));
+        // ES5: the rules engine is duktape and has no `startsWith`.
+        assert!(!recipe.contains("startsWith"));
+    }
+
     #[cfg(target_os = "linux")]
     #[test]
     fn an_interface_index_is_read_from_the_running_kernel() {

crates/tsunagi-core/src/dns/publish/resolved.rs

@@ -135,36 +135,53 @@ impl DnsPublisher for ResolvedPublisher {
                     published.interface
                 ))
             })?;
+            if published.servers.is_empty() {
+                return Err(PublishError::Unavailable(
+                    "there is no listening address to send questions to".to_string(),
+                ));
+            }
             let proxy = Self::proxy().await?;
             let index = ifindex as i32;
-            let (family, octets) = wire_address(published.server.ip());
+            // Every family in one call, because this replaces the link's
-            let port = published.server.port();
+            // whole list: sending them one at a time would leave only the
+            // last. A resolver then picks whichever it can reach.
+            let servers: Vec<(i32, Vec<u8>, u16, String)> = published
+                .servers
+                .iter()
+                .map(|server| {
+                    let (family, octets) = wire_address(server.ip());
+                    (family, octets, server.port(), String::new())
+                })
+                .collect();
 
-            match proxy
+            match proxy.set_link_dns_ex(index, &servers).await {
-                .set_link_dns_ex(index, &[(family, octets.clone(), port, String::new())])
-                .await
-            {
                 Ok(()) => {}
                 Err(err) => {
-                    let classified = classify(err, "setting the link's DNS server");
+                    let classified = classify(err, "setting the link's DNS servers");
                     // Older systemd has no `Ex` form, and the plain one is
                     // always port 53. Falling back to it when the server is
                     // somewhere else would point resolved at nothing.
                     if !matches!(classified, PublishError::Unavailable(_)) {
                         return Err(classified);
                     }
-                    if port != 53 {
+                    if let Some(elsewhere) =
+                        published.servers.iter().find(|server| server.port() != 53)
+                    {
                         return Err(PublishError::Unavailable(format!(
                             "this systemd-resolved cannot be given a port, and the server is on \
-                             {port}. Run the server on port 53, or point your resolver at \
+                             {}. Run the server on port 53, or point your resolver at \
-                             {} yourself.",
+                             {elsewhere} yourself.",
-                            published.server
+                            elsewhere.port()
                         )));
                     }
+                    let plain: Vec<(i32, Vec<u8>)> = servers
+                        .into_iter()
+                        .map(|(family, octets, _, _)| (family, octets))
+                        .collect();
                     proxy
-                        .set_link_dns(index, &[(family, octets)])
+                        .set_link_dns(index, &plain)
                         .await
-                        .map_err(|err| classify(err, "setting the link's DNS server"))?;
+                        .map_err(|err| classify(err, "setting the link's DNS servers"))?;
                 }
             }
 
@@ -243,7 +260,7 @@ mod tests {
         let publisher = ResolvedPublisher::new();
         let published = Published {
             interface: "tsunagi-no-such-interface".into(),
-            server: SocketAddr::from(([10, 13, 37, 69], 5354)),
+            servers: vec![SocketAddr::from(([10, 13, 37, 69], 5354))],
             domains: vec!["lab".into()],
         };
         let err = publisher.apply(&published).await.unwrap_err();

crates/tsunagi-core/src/dns/publish/unsupported.rs

@@ -62,7 +62,7 @@ mod tests {
         let publisher = UnsupportedPublisher::new();
         let published = Published {
             interface: "tsundemo".into(),
-            server: "10.0.0.1:5354".parse().unwrap(),
+            servers: vec!["10.0.0.1:5354".parse().unwrap()],
             domains: vec!["lab".into()],
         };
         let err = publisher.apply(&published).await.unwrap_err();

crates/tsunagi-core/src/dns/publish/windows.rs

@@ -0,0 +1,301 @@
+//! Telling the Windows DNS client to send some questions here, through the
+//! Name Resolution Policy Table.
+//!
+//! The NRPT is how Windows does split DNS: a rule says "names under this
+//! suffix are resolved by these servers", and names outside every rule are
+//! resolved the ordinary way. That is exactly the contract of [`Published`] —
+//! route these suffixes here and claim nothing else — so one NRPT rule per
+//! suffix is the whole of it. The rules are keyed by namespace, not by
+//! interface, so the interface name in [`Published`] is not needed here.
+//!
+//! # Why PowerShell
+//!
+//! The NRPT lives in the registry, but a rule only takes effect once the DNS
+//! client is told to reload its policy, and that notification is an RPC with
+//! no safe wrapper this crate may call — it forbids `unsafe`. The
+//! `DnsClient` PowerShell module does the write *and* the reload, so it is
+//! used by absolute path from `%SystemRoot%`. Every value passed to it is one
+//! this agent produced: the suffixes are validated zone names (letters,
+//! digits, `-`, `_`, `.` only) and the servers are addresses it is listening
+//! on. They are still single-quoted and escaped when built into the script,
+//! so nothing could be read as PowerShell rather than as data.
+//!
+//! # The port
+//!
+//! The Windows DNS client always asks on port 53; it cannot be told another.
+//! So a server on a different port cannot be reached this way, and rather than
+//! configure a rule that points at nothing, this says so — the same honest
+//! failure the systemd-resolved path gives for an old resolver with no port.
+//!
+//! # Privilege
+//!
+//! Writing an NRPT rule needs an elevated process. An ordinary user is
+//! refused, which is reported as its own kind of error because the answer to
+//! it — run elevated, or as a service — differs from "this is not Windows".
+
+use std::collections::BTreeSet;
+use std::net::IpAddr;
+use std::sync::Mutex;
+
+use crate::BoxFuture;
+
+use super::{DnsPublisher, PublishError, Published};
+
+/// The port the Windows DNS client is fixed to.
+const DNS_PORT: u16 = 53;
+
+/// Configures the Name Resolution Policy Table through PowerShell.
+#[derive(Debug, Default)]
+pub struct NrptPublisher {
+    /// The namespaces last configured, so shutdown knows what to undo.
+    applied: Mutex<Vec<String>>,
+}
+
+impl NrptPublisher {
+    /// Creates the publisher. Nothing is contacted until [`Self::apply`].
+    pub fn new() -> Self {
+        Self::default()
+    }
+}
+
+/// A suffix as the NRPT wants it: a leading dot means "this and everything
+/// under it".
+fn namespace(domain: &str) -> String {
+    let trimmed = domain.trim_matches('.');
+    format!(".{trimmed}")
+}
+
+/// A PowerShell single-quoted literal, with any embedded quote doubled.
+fn ps_literal(value: &str) -> String {
+    format!("'{}'", value.replace('\'', "''"))
+}
+
+/// A PowerShell array literal, `@('a','b')`, or `@()` when empty.
+fn ps_array(values: &[String]) -> String {
+    let items: Vec<String> = values.iter().map(|value| ps_literal(value)).collect();
+    format!("@({})", items.join(","))
+}
+
+/// The script that clears our rules for these namespaces and then adds them.
+fn apply_script(namespaces: &[String], servers: &[String]) -> String {
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $ns = {ns}\n\
+         $srv = {srv}\n\
+         foreach ($n in $ns) {{ Get-DnsClientNrptRule | Where-Object {{ $_.Namespace -eq $n }} | \
+         ForEach-Object {{ Remove-DnsClientNrptRule -Name $_.Name -Force }} }}\n\
+         foreach ($n in $ns) {{ Add-DnsClientNrptRule -Namespace $n -NameServers $srv }}\n",
+        ns = ps_array(namespaces),
+        srv = ps_array(servers),
+    )
+}
+
+/// The script that clears our rules for these namespaces.
+fn revert_script(namespaces: &[String]) -> String {
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $ns = {ns}\n\
+         foreach ($n in $ns) {{ Get-DnsClientNrptRule | Where-Object {{ $_.Namespace -eq $n }} | \
+         ForEach-Object {{ Remove-DnsClientNrptRule -Name $_.Name -Force }} }}\n",
+        ns = ps_array(namespaces),
+    )
+}
+
+/// Turns what PowerShell said on failure into the kind of failure it is.
+fn classify(text: &str) -> PublishError {
+    let lower = text.to_ascii_lowercase();
+    if lower.contains("access is denied")
+        || lower.contains("requires elevation")
+        || lower.contains("run as administrator")
+        || lower.contains("administrator privilege")
+        || lower.contains("permissiondenied")
+    {
+        PublishError::Refused(format!("the Windows DNS client refused the change: {text}"))
+    } else if lower.contains("is not recognized")
+        || lower.contains("commandnotfoundexception")
+        || lower.contains("not recognized as the name of a cmdlet")
+    {
+        PublishError::Unavailable(format!(
+            "the DnsClient PowerShell module is not available: {text}"
+        ))
+    } else {
+        PublishError::Failed(format!("the Windows DNS client failed the change: {text}"))
+    }
+}
+
+/// Runs a PowerShell script from `%SystemRoot%` and maps a failure to a
+/// [`PublishError`].
+async fn powershell(script: String) -> Result<(), PublishError> {
+    let program = powershell_path();
+    let output = tokio::task::spawn_blocking(move || {
+        std::process::Command::new(program)
+            .args([
+                "-NoProfile",
+                "-NonInteractive",
+                "-ExecutionPolicy",
+                "Bypass",
+                "-Command",
+                &script,
+            ])
+            .output()
+    })
+    .await
+    .map_err(|err| PublishError::Failed(format!("could not run PowerShell: {err}")))?
+    .map_err(|err| PublishError::Unavailable(format!("could not run PowerShell: {err}")))?;
+
+    if output.status.success() {
+        return Ok(());
+    }
+    let stderr = String::from_utf8_lossy(&output.stderr);
+    let text = if stderr.trim().is_empty() {
+        String::from_utf8_lossy(&output.stdout).trim().to_string()
+    } else {
+        stderr.trim().to_string()
+    };
+    Err(classify(if text.is_empty() { "no output" } else { &text }))
+}
+
+/// The absolute path to Windows PowerShell, so nothing on `PATH` can stand in
+/// for it.
+fn powershell_path() -> std::path::PathBuf {
+    let root = std::env::var_os("SystemRoot").unwrap_or_else(|| r"C:\Windows".into());
+    std::path::Path::new(&root)
+        .join("System32")
+        .join("WindowsPowerShell")
+        .join("v1.0")
+        .join("powershell.exe")
+}
+
+impl DnsPublisher for NrptPublisher {
+    fn name(&self) -> &str {
+        "windows-nrpt"
+    }
+
+    fn apply<'a>(&'a self, published: &'a Published) -> BoxFuture<'a, Result<(), PublishError>> {
+        Box::pin(async move {
+            if published.domains.is_empty() {
+                return Err(PublishError::Unavailable(
+                    "there are no suffixes to route, so there is no rule to write".to_string(),
+                ));
+            }
+
+            // Windows can only ask on port 53, so a server anywhere else is
+            // unreachable this way. Keep the ones it can use.
+            let servers: BTreeSet<IpAddr> = published
+                .servers
+                .iter()
+                .filter(|server| server.port() == DNS_PORT)
+                .map(|server| server.ip())
+                .collect();
+            if servers.is_empty() {
+                let elsewhere = published
+                    .servers
+                    .iter()
+                    .map(|server| server.to_string())
+                    .collect::<Vec<_>>()
+                    .join(", ");
+                return Err(PublishError::Unavailable(format!(
+                    "the Windows DNS client only asks on port {DNS_PORT}, and the server is on \
+                     {elsewhere}. Run the server on port {DNS_PORT}, or point your resolver at it \
+                     yourself."
+                )));
+            }
+
+            let namespaces: Vec<String> = published.domains.iter().map(|d| namespace(d)).collect();
+            let servers: Vec<String> = servers.iter().map(|ip| ip.to_string()).collect();
+
+            powershell(apply_script(&namespaces, &servers)).await?;
+
+            match self.applied.lock() {
+                Ok(mut guard) => *guard = namespaces,
+                Err(poisoned) => *poisoned.into_inner() = namespaces,
+            }
+            Ok(())
+        })
+    }
+
+    fn revert(&self) -> BoxFuture<'_, Result<(), PublishError>> {
+        Box::pin(async move {
+            let namespaces = match self.applied.lock() {
+                Ok(mut guard) => std::mem::take(&mut *guard),
+                Err(poisoned) => std::mem::take(&mut *poisoned.into_inner()),
+            };
+            if namespaces.is_empty() {
+                return Ok(());
+            }
+            powershell(revert_script(&namespaces)).await
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::net::SocketAddr;
+
+    #[test]
+    fn a_suffix_gets_a_leading_dot() {
+        assert_eq!(namespace("lab"), ".lab");
+        assert_eq!(namespace(".lab"), ".lab");
+        assert_eq!(namespace("a.b"), ".a.b");
+    }
+
+    #[test]
+    fn a_quote_in_a_value_cannot_escape_the_literal() {
+        // Zone names never contain a quote, but the escaping is what makes
+        // that a guarantee rather than a hope.
+        assert_eq!(ps_literal("a'b"), "'a''b'");
+        assert_eq!(ps_array(&["x".into(), "y".into()]), "@('x','y')");
+        assert_eq!(ps_array(&[]), "@()");
+    }
+
+    #[test]
+    fn the_apply_script_removes_before_it_adds() {
+        let script = apply_script(&[".lab".into()], &["10.13.37.69".into()]);
+        let remove = script.find("Remove-DnsClientNrptRule").unwrap();
+        let add = script.find("Add-DnsClientNrptRule").unwrap();
+        assert!(
+            remove < add,
+            "a stale rule must go before the new one:\n{script}"
+        );
+        assert!(script.contains("@('.lab')"));
+        assert!(script.contains("@('10.13.37.69')"));
+    }
+
+    #[test]
+    fn an_access_denied_is_a_refusal_a_missing_cmdlet_is_unavailable() {
+        assert!(matches!(
+            classify("Access is denied"),
+            PublishError::Refused(_)
+        ));
+        assert!(matches!(
+            classify("The term 'Add-DnsClientNrptRule' is not recognized"),
+            PublishError::Unavailable(_)
+        ));
+        assert!(matches!(
+            classify("something else"),
+            PublishError::Failed(_)
+        ));
+    }
+
+    #[tokio::test]
+    async fn a_server_only_on_a_nonstandard_port_is_unavailable_not_a_failure() {
+        // Nothing is run: Windows cannot ask there, and saying so is the
+        // honest answer without touching the DNS client.
+        let publisher = NrptPublisher::new();
+        let published = Published {
+            interface: "tsun0".into(),
+            servers: vec![SocketAddr::from(([10, 13, 37, 69], 5354))],
+            domains: vec!["lab".into()],
+        };
+        let err = publisher.apply(&published).await.unwrap_err();
+        assert!(matches!(err, PublishError::Unavailable(_)), "{err}");
+    }
+
+    #[tokio::test]
+    async fn reverting_without_having_applied_does_nothing_and_succeeds() {
+        NrptPublisher::new().revert().await.unwrap();
+    }
+}

crates/tsunagi-core/src/dns/server.rs

@@ -16,7 +16,7 @@ use std::sync::{Arc, RwLock};
 use simple_dns::rdata::{A, PTR, RData, SOA};
 use simple_dns::{Name, PacketFlag, QCLASS, QTYPE, RCODE, ResourceRecord, TYPE};
 
-use super::zone::{Answer, Query, Zone};
+use super::zone::{Answer, Query, Zone, Zones};
 
 /// How long an answer may be cached.
 ///
@@ -39,30 +39,35 @@ const TCP_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(5);
 /// How many TCP questions may be in flight at once.
 const MAX_TCP_CONNECTIONS: usize = 32;
 
-/// The zone the server answers from, swapped as the roster changes.
+/// The zones the server answers from, swapped as the rosters change.
 ///
-/// Shared rather than copied into the server so that a roster change is one
+/// Shared rather than copied into the server so that a roster change — or a
-/// write, not a restart: rebinding the socket would drop questions in flight
+/// network joining or leaving — is one write, not a restart: rebinding the
-/// for no reason.
+/// socket would drop questions in flight for no reason.
-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, Default)]
-pub struct SharedZone(Arc<RwLock<Arc<Zone>>>);
+pub struct SharedZone(Arc<RwLock<Arc<Zones>>>);
 
 impl SharedZone {
-    /// Wraps a zone.
+    /// Wraps a set of zones.
-    pub fn new(zone: Zone) -> Self {
+    pub fn new(zones: Zones) -> Self {
-        Self(Arc::new(RwLock::new(Arc::new(zone))))
+        Self(Arc::new(RwLock::new(Arc::new(zones))))
     }
 
-    /// Replaces it.
+    /// Wraps a single zone, which is the common case in a test.
-    pub fn set(&self, zone: Zone) {
+    pub fn one(zone: Zone) -> Self {
+        Self::new(Zones::new([zone]))
+    }
+
+    /// Replaces them.
+    pub fn set(&self, zones: Zones) {
         match self.0.write() {
-            Ok(mut guard) => *guard = Arc::new(zone),
+            Ok(mut guard) => *guard = Arc::new(zones),
-            Err(poisoned) => *poisoned.into_inner() = Arc::new(zone),
+            Err(poisoned) => *poisoned.into_inner() = Arc::new(zones),
         }
     }
 
-    /// The zone as it is now.
+    /// The zones as they are now.
-    pub fn get(&self) -> Arc<Zone> {
+    pub fn get(&self) -> Arc<Zones> {
         match self.0.read() {
             Ok(guard) => Arc::clone(&guard),
             Err(poisoned) => Arc::clone(&poisoned.into_inner()),
@@ -75,7 +80,7 @@ impl SharedZone {
 /// `None` means say nothing at all: the message was not a question this
 /// server should reply to, and replying anyway would make this a useful
 /// amplifier for somebody spoofing a source address.
-pub fn respond(zone: &Zone, query: &[u8]) -> Option<Vec<u8>> {
+pub fn respond(zones: &Zones, query: &[u8]) -> Option<Vec<u8>> {
     let packet = simple_dns::Packet::parse(query).ok()?;
     if packet.has_flags(PacketFlag::RESPONSE) {
         return None;
@@ -105,7 +110,16 @@ pub fn respond(zone: &Zone, query: &[u8]) -> Option<Vec<u8>> {
     }
 
     let qname = question.qname.to_string();
-    let answer = zone.lookup(&qname, query_kind(question.qtype));
+    // Which zone answers is part of the answer: a denial is bounded by the
+    // SOA of the zone that denied it, and this server may hold several.
+    let (zone, answer) = match zones.lookup(&qname, query_kind(question.qtype)) {
+        Some(answered) => answered,
+        None => {
+            reply.questions.push(question.clone());
+            *reply.rcode_mut() = RCODE::Refused;
+            return reply.build_bytes_vec_compressed().ok();
+        }
+    };
     reply.questions.push(question.clone());
 
     let name = Name::new(&qname).ok()?;
@@ -338,7 +352,11 @@ mod tests {
     use crate::dns::zone::ZoneName;
     use std::net::Ipv4Addr;
 
-    fn zone() -> Zone {
+    fn zone() -> Zones {
+        Zones::new([one_zone()])
+    }
+
+    fn one_zone() -> Zone {
         Zone::new(
             ZoneName::new("lab").unwrap(),
             [
@@ -468,7 +486,7 @@ mod tests {
         let many: Vec<(String, Ipv4Addr)> = (0..200)
             .map(|i| ("host".to_string(), Ipv4Addr::new(10, 13, 37, i as u8)))
             .collect();
-        let wide = Zone::new(ZoneName::new("lab").unwrap(), many);
+        let wide = Zones::new([Zone::new(ZoneName::new("lab").unwrap(), many)]);
         let query = ask("host.lab", TYPE::A);
         let full = respond(&wide, &query).unwrap();
         assert!(
@@ -518,9 +536,64 @@ mod tests {
         assert_eq!(reply.answers.len(), 1);
     }
 
+    #[tokio::test]
+    async fn the_same_zone_is_answered_over_ipv6_as_over_ipv4() {
+        // A question arrives over whichever family the resolver chooses, so
+        // a listener is opened from each list in the plan and both have to
+        // answer the same thing. Answering one family only leaves the other
+        // timing out, which looks like a broken overlay.
+        let shared = SharedZone::new(zone());
+        let plan = crate::dns::listen_plan(0);
+        let mut answered = 0;
+
+        for family in plan.families() {
+            let candidate = family[0];
+            let server = match DnsServer::bind(candidate, shared.clone()).await {
+                Ok(server) => server,
+                // A host with IPv6 switched off in the kernel cannot bind
+                // `::1`, and the agent copes with that by design; so does
+                // this. The other family still has to work.
+                Err(err) if candidate.is_ipv6() => {
+                    eprintln!("no IPv6 loopback on this host ({err}); skipping that family");
+                    continue;
+                }
+                Err(err) => panic!("cannot listen on {candidate}: {err}"),
+            };
+            let addr = server.local_addr();
+            assert_eq!(addr.is_ipv6(), candidate.is_ipv6());
+
+            let client = tokio::net::UdpSocket::bind(if addr.is_ipv6() {
+                "[::1]:0"
+            } else {
+                "127.0.0.1:0"
+            })
+            .await
+            .unwrap();
+            client
+                .send_to(&ask("music.lab", TYPE::A), addr)
+                .await
+                .unwrap();
+            let mut buffer = vec![0u8; MAX_MESSAGE_LEN];
+            let read = client.recv(&mut buffer).await.unwrap();
+            let reply = simple_dns::Packet::parse(&buffer[..read]).unwrap();
+            assert_eq!(reply.rcode(), RCODE::NoError, "over {addr}");
+            match &reply.answers[0].rdata {
+                // The same IPv4 answer either way: the family a question
+                // travelled over says nothing about what the answer is.
+                RData::A(a) => {
+                    assert_eq!(Ipv4Addr::from(a.address), Ipv4Addr::new(10, 13, 37, 237))
+                }
+                other => panic!("expected an A record over {addr}, got {other:?}"),
+            }
+            answered += 1;
+        }
+
+        assert!(answered > 0, "at least one family must have answered");
+    }
+
     #[tokio::test]
     async fn replacing_the_zone_changes_what_the_running_server_answers() {
-        let shared = SharedZone::new(Zone::new(ZoneName::new("lab").unwrap(), []));
+        let shared = SharedZone::one(Zone::new(ZoneName::new("lab").unwrap(), []));
         let server = DnsServer::bind("127.0.0.1:0".parse().unwrap(), shared.clone())
             .await
             .unwrap();

crates/tsunagi-core/src/dns/zone.rs

@@ -128,17 +128,14 @@ impl ZoneName {
                     .to_string(),
             );
         }
-        if tld::exist_case_insensitive(top) {
+        // A real top-level domain is the whole of it. A name that is not
-            return Some(format!(
+        // delegated today might be one day — and saying so fires on every
+        // private name anybody ever picks, which is a warning about
+        // nothing that teaches people to ignore warnings.
+        tld::exist_case_insensitive(top).then(|| {
+            format!(
                 "`.{top}` is a real top-level domain, so every public name under it \
                  becomes unreachable from this host while the overlay is up"
-            ));
-        }
-        // Not delegated today is not a promise about tomorrow.
-        (!self.0.contains('.')).then(|| {
-            format!(
-                "`.{top}` is not a delegated top-level domain today, but it could \
-                 become one; `.internal` is reserved for private use and never will"
             )
         })
     }
@@ -336,6 +333,86 @@ impl Zone {
     }
 }
 
+/// Every zone one agent serves.
+///
+/// One agent has one identity and as many networks as it likes, and each of
+/// them is a zone of its own named after the network. They are served
+/// together because they arrive at the same socket: a question is not
+/// labelled with the network it belongs to, only with a name, and the
+/// suffix is what decides. Nothing is shared between them — a member of one
+/// network is not a name in another.
+#[derive(Debug, Clone, Default)]
+pub struct Zones {
+    zones: Vec<Zone>,
+}
+
+impl Zones {
+    /// Collects zones, dropping any that repeats an origin already taken.
+    ///
+    /// Two networks can be given the same name, and then only one of them
+    /// can own the suffix. First wins, deterministically, rather than one
+    /// shadowing the other depending on the order they happened to start.
+    pub fn new(zones: impl IntoIterator<Item = Zone>) -> Self {
+        let mut taken: Vec<Zone> = Vec::new();
+        for zone in zones {
+            if taken.iter().any(|other| other.origin() == zone.origin()) {
+                continue;
+            }
+            taken.push(zone);
+        }
+        Self { zones: taken }
+    }
+
+    /// The zones, in the order they will be consulted.
+    pub fn iter(&self) -> impl Iterator<Item = &Zone> {
+        self.zones.iter()
+    }
+
+    /// How many zones are served.
+    pub fn len(&self) -> usize {
+        self.zones.len()
+    }
+
+    /// Whether nothing is served.
+    pub fn is_empty(&self) -> bool {
+        self.zones.is_empty()
+    }
+
+    /// How many names are answered for, across every zone.
+    pub fn names(&self) -> usize {
+        self.zones.iter().map(Zone::len).sum()
+    }
+
+    /// Answers one question, and says which zone answered it.
+    ///
+    /// The answer has to travel with its zone because the SOA that bounds a
+    /// denial belongs to the zone that denied it, and serving several means
+    /// that is no longer a foregone conclusion.
+    ///
+    /// A reverse question carries no suffix to match on, so every zone is
+    /// asked and the one that holds the address answers; if none does, the
+    /// denial comes from whichever is authoritative for that reverse zone,
+    /// and from nowhere at all when none is.
+    pub fn lookup(&self, qname: &str, query: Query) -> Option<(&Zone, Answer)> {
+        let mut denial: Option<(&Zone, Answer)> = None;
+        for zone in &self.zones {
+            match zone.lookup(qname, query) {
+                // Not this zone's question. Another may still own it.
+                Answer::NotOurs => continue,
+                // A zone that owns the suffix but not the name can still be
+                // overruled by one that has an answer: only reverse
+                // questions reach more than one zone, and there the zone
+                // holding the address is the one that knows.
+                answer @ (Answer::NoSuchName | Answer::NoData) => {
+                    denial.get_or_insert((zone, answer));
+                }
+                answer => return Some((zone, answer)),
+            }
+        }
+        denial
+    }
+}
+
 /// The address a reverse name asks about, if it is one.
 fn reverse_address(qname: &str) -> Option<Ipv4Addr> {
     let name = qname.trim_end_matches('.').to_ascii_lowercase();
@@ -436,11 +513,13 @@ mod tests {
         let local = ZoneName::new("local").unwrap().collision().unwrap();
         assert!(local.contains("multicast DNS"), "{local}");
 
-        // An undelegated single label is a maybe, not a yes.
+        // A name that collides with nothing is said nothing about. It
-        let lab = ZoneName::new("lab").unwrap().collision().unwrap();
+        // might be delegated one day, and warning about that fires on
-        assert!(lab.contains("could"), "{lab}");
+        // every private name anybody picks — a warning about nothing,
-        // A multi-label name under something undelegated is not worth a word.
+        // which is how people learn to ignore warnings.
-        assert_eq!(ZoneName::new("a.lab").unwrap().collision(), None);
+        for quiet in ["lab", "a.lab", "pidar", "twar", "kitchen-table"] {
+            assert_eq!(ZoneName::new(quiet).unwrap().collision(), None, "{quiet}");
+        }
     }
 
     #[test]
@@ -577,4 +656,94 @@ mod tests {
         );
         assert_ne!(a.serial(), changed.serial());
     }
+
+    #[test]
+    fn each_network_answers_only_for_its_own_zone() {
+        // One agent, several networks, one socket. A question carries a
+        // name and not the network it belongs to, so the suffix is what
+        // decides — and a member of one network is not a name in another.
+        let zones = Zones::new([
+            Zone::new(
+                ZoneName::new("lab").unwrap(),
+                [("music".to_string(), Ipv4Addr::new(10, 13, 37, 2))],
+            ),
+            Zone::new(
+                ZoneName::new("home").unwrap(),
+                [("music".to_string(), Ipv4Addr::new(10, 20, 0, 2))],
+            ),
+        ]);
+        assert_eq!(zones.len(), 2);
+        assert_eq!(zones.names(), 2);
+
+        let (zone, answer) = zones.lookup("music.lab", Query::A).unwrap();
+        assert_eq!(zone.origin().as_str(), "lab");
+        assert_eq!(
+            answer,
+            Answer::Addresses(vec![Ipv4Addr::new(10, 13, 37, 2)])
+        );
+
+        let (zone, answer) = zones.lookup("music.home", Query::A).unwrap();
+        assert_eq!(zone.origin().as_str(), "home");
+        assert_eq!(answer, Answer::Addresses(vec![Ipv4Addr::new(10, 20, 0, 2)]));
+
+        // A name under neither is nobody's business here.
+        assert!(zones.lookup("music.example.com", Query::A).is_none());
+        assert!(Zones::default().lookup("music.lab", Query::A).is_none());
+    }
+
+    #[test]
+    fn a_denial_comes_from_the_zone_that_owns_the_name() {
+        // The SOA that bounds a negative answer belongs to the zone that
+        // denied it; with several served that is no longer a given.
+        let zones = Zones::new([
+            Zone::new(ZoneName::new("lab").unwrap(), []),
+            Zone::new(
+                ZoneName::new("home").unwrap(),
+                [("music".to_string(), Ipv4Addr::new(10, 20, 0, 2))],
+            ),
+        ]);
+        let (zone, answer) = zones.lookup("nobody.lab", Query::A).unwrap();
+        assert_eq!(zone.origin().as_str(), "lab");
+        assert_eq!(answer, Answer::NoSuchName);
+    }
+
+    #[test]
+    fn a_reverse_question_is_answered_by_whichever_zone_holds_the_address() {
+        // It carries no suffix to match on, so every zone is asked.
+        let zones = Zones::new([
+            Zone::new(
+                ZoneName::new("lab").unwrap(),
+                [("music".to_string(), Ipv4Addr::new(10, 13, 37, 2))],
+            ),
+            Zone::new(
+                ZoneName::new("home").unwrap(),
+                [("kitchen".to_string(), Ipv4Addr::new(10, 20, 0, 2))],
+            ),
+        ]);
+        let (_, answer) = zones.lookup("2.0.20.10.in-addr.arpa", Query::Ptr).unwrap();
+        assert_eq!(answer, Answer::Name("kitchen.home".to_string()));
+        let (_, answer) = zones.lookup("2.37.13.10.in-addr.arpa", Query::Ptr).unwrap();
+        assert_eq!(answer, Answer::Name("music.lab".to_string()));
+    }
+
+    #[test]
+    fn two_networks_of_one_name_do_not_shadow_each_other_at_random() {
+        // Exactly the two-`LAB` case. Only one can own the suffix; which
+        // one must not depend on the order they happened to start in.
+        let first = Zone::new(
+            ZoneName::new("lab").unwrap(),
+            [("music".to_string(), Ipv4Addr::new(10, 13, 37, 2))],
+        );
+        let second = Zone::new(
+            ZoneName::new("lab").unwrap(),
+            [("music".to_string(), Ipv4Addr::new(10, 99, 0, 2))],
+        );
+        let zones = Zones::new([first, second]);
+        assert_eq!(zones.len(), 1, "one suffix, one zone");
+        let (_, answer) = zones.lookup("music.lab", Query::A).unwrap();
+        assert_eq!(
+            answer,
+            Answer::Addresses(vec![Ipv4Addr::new(10, 13, 37, 2)])
+        );
+    }
 }

crates/tsunagi-core/src/error.rs

@@ -110,6 +110,23 @@ pub enum Error {
     #[error("agent is stopped")]
     Stopped,
 
+    /// Something that had to answer in bounded time did not.
+    ///
+    /// Reported instead of waited out: an agent that has wedged must not
+    /// leave a command hanging with nothing on screen and no way out.
+    #[error("{what} did not answer in time")]
+    Timeout {
+        /// What was asked, phrased so the message stands on its own.
+        what: String,
+    },
+
+    /// The request is understood and cannot be done as things stand.
+    #[error("{reason}")]
+    Rejected {
+        /// Why, phrased so the message stands on its own.
+        reason: String,
+    },
+
     /// Discovery backend failure. Never fatal for the agent.
     #[error("discovery error: {0}")]
     Discovery(String),

crates/tsunagi-core/src/identity/mod.rs

@@ -49,11 +49,16 @@ impl DeviceIdentity {
         self.secret.public()
     }
 
-    /// The raw secret key bytes, for persistence only.
+    /// The raw secret key bytes, for persistence and transport certificate generation.
-    pub(crate) fn secret_bytes(&self) -> [u8; 32] {
+    pub fn secret_bytes(&self) -> [u8; 32] {
         self.secret.to_bytes()
     }
 
+    /// Signs a message with the device's persistent secret key.
+    pub fn sign(&self, message: &[u8]) -> iroh::Signature {
+        self.secret.sign(message)
+    }
+
     /// A clone of the iroh secret key, for endpoint construction only.
     pub(crate) fn secret_key(&self) -> SecretKey {
         self.secret.clone()

crates/tsunagi-core/src/identity/network.rs

@@ -340,6 +340,8 @@ pub struct NetworkKeys {
     network_id: NetworkId,
     discovery_key: DiscoveryKey,
     auth_key: Zeroizing<[u8; 32]>,
+    // Separate from the semi-public DiscoveryKey. Existing derivations stay frozen.
+    dht_write_key: Zeroizing<[u8; 32]>,
     name: NetworkName,
 }
 
@@ -373,6 +375,7 @@ impl NetworkKeys {
             network_id: NetworkId(expand("network-id")),
             discovery_key: DiscoveryKey(expand("discovery-key")),
             auth_key: Zeroizing::new(expand("handshake-auth")),
+            dht_write_key: Zeroizing::new(expand("mainline-rendezvous-write-v1")),
             name: name.clone(),
         }
     }
@@ -405,6 +408,11 @@ impl NetworkKeys {
     pub(crate) fn auth_key(&self) -> &[u8; 32] {
         &self.auth_key
     }
+
+    /// Secret signing material for network rendezvous; never exported or logged.
+    pub(crate) fn dht_write_key(&self) -> &[u8; 32] {
+        &self.dht_write_key
+    }
 }
 
 impl std::fmt::Debug for NetworkKeys {

crates/tsunagi-core/src/ipc/unix.rs

@@ -0,0 +1,227 @@
+//! A Unix socket adapter for the local control interface.
+//!
+//! One of the per-platform adapters; see [`super`]. It provides only what is
+//! particular to a Unix socket — binding a listener, connecting a client, and
+//! the owner-only permissions — and hands every accepted connection to the
+//! shared [`serve_connection`](super::serve_connection). The request framing,
+//! the dispatch and the client wrappers are the same on every platform and
+//! live in [`super`], re-exported here so `ipc::unix::request_status` and its
+//! siblings keep resolving.
+//!
+//! It is reachable only by a process that can open a file inside the agent's
+//! owner-only state directory, and the socket itself is created mode `0600`.
+
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+use tokio::net::{UnixListener, UnixStream};
+use tokio::task::JoinHandle;
+
+use crate::error::{Error, Result};
+
+// The transport-agnostic surface, re-exported so this module is a complete
+// view of the local control interface on its own.
+pub use super::{
+    CONTROL_PROTOCOL, ControlSocketAccess, EXCHANGE_TIMEOUT, ReportSource, is_serving,
+    join_network, leave_network, request_status, set_active, set_dns, set_hostname,
+};
+
+/// Serves the local control interface on a Unix socket.
+#[derive(Debug)]
+pub struct ControlSocket {
+    path: PathBuf,
+    task: Option<JoinHandle<()>>,
+}
+
+impl ControlSocket {
+    /// Binds the socket and starts serving.
+    ///
+    /// A socket file left behind by a crashed agent is replaced, but only
+    /// after checking that nothing is listening on it, so two live agents
+    /// never fight over one path.
+    pub async fn bind(
+        path: impl AsRef<Path>,
+        source: Arc<dyn ReportSource>,
+        access: ControlSocketAccess,
+    ) -> Result<Self> {
+        let path = path.as_ref().to_path_buf();
+
+        if let Some(parent) = path.parent() {
+            create_parent(parent, &access)?;
+        }
+
+        if path.exists() {
+            if is_serving(&path).await {
+                return Err(Error::StateLocked { path: path.clone() });
+            }
+            // Nothing is listening, so the file is a leftover.
+            std::fs::remove_file(&path).map_err(|source| Error::Io {
+                path: path.clone(),
+                source,
+            })?;
+        }
+
+        let listener = UnixListener::bind(&path).map_err(|source| Error::Io {
+            path: path.clone(),
+            source,
+        })?;
+        apply_access(&path, &access)?;
+
+        let task = tokio::spawn(serve(listener, source));
+        Ok(Self {
+            path,
+            task: Some(task),
+        })
+    }
+
+    /// The path being served.
+    pub fn path(&self) -> &Path {
+        &self.path
+    }
+
+    /// Stops serving and removes the socket file.
+    pub async fn shutdown(mut self) {
+        if let Some(task) = self.task.take() {
+            task.abort();
+            let _ = task.await;
+        }
+        let _ = std::fs::remove_file(&self.path);
+    }
+}
+
+impl Drop for ControlSocket {
+    fn drop(&mut self) {
+        if let Some(task) = self.task.take() {
+            task.abort();
+        }
+        let _ = std::fs::remove_file(&self.path);
+    }
+}
+
+/// Creates the socket's parent directory.
+///
+/// A group-shared socket must sit in a directory the group can traverse, so it
+/// is `0755`; a private socket keeps the owner-only state-directory default.
+fn create_parent(parent: &Path, access: &ControlSocketAccess) -> Result<()> {
+    match access {
+        ControlSocketAccess::Private => crate::storage::create_dir(parent),
+        ControlSocketAccess::Group(_) => {
+            std::fs::create_dir_all(parent).map_err(|source| Error::Io {
+                path: parent.to_path_buf(),
+                source,
+            })?;
+            set_mode(parent, 0o755)
+        }
+    }
+}
+
+/// Applies the access policy to the bound socket.
+///
+/// Private is owner-only `0600`. Group sets the socket's group to the gid and
+/// the mode to `0660`; if the group cannot be set — the agent is not a member,
+/// say — it falls back to owner-only rather than leaving the group wrong, so
+/// the socket is never more open than intended.
+fn apply_access(path: &Path, access: &ControlSocketAccess) -> Result<()> {
+    match access {
+        ControlSocketAccess::Private => set_mode(path, 0o600),
+        ControlSocketAccess::Group(gid) => {
+            if let Err(source) = std::os::unix::fs::chown(path, None, Some(*gid)) {
+                tracing::warn!(
+                    path = %path.display(), gid, %source,
+                    "cannot set the control socket group; keeping it owner-only"
+                );
+                return set_mode(path, 0o600);
+            }
+            set_mode(path, 0o660)
+        }
+    }
+}
+
+fn set_mode(path: &Path, mode: u32) -> Result<()> {
+    use std::os::unix::fs::PermissionsExt;
+    std::fs::set_permissions(path, std::fs::Permissions::from_mode(mode)).map_err(|source| {
+        Error::Io {
+            path: path.to_path_buf(),
+            source,
+        }
+    })
+}
+
+async fn serve(listener: UnixListener, source: Arc<dyn ReportSource>) {
+    loop {
+        let Ok((stream, _)) = listener.accept().await else {
+            continue;
+        };
+        let source = Arc::clone(&source);
+        tokio::spawn(async move {
+            if let Err(err) = super::serve_connection(stream, source).await {
+                tracing::debug!(%err, "local control request failed");
+            }
+        });
+    }
+}
+
+/// Connects a client to the socket at `path`.
+pub(crate) async fn connect(path: &Path) -> Result<UnixStream> {
+    UnixStream::connect(path).await.map_err(|source| Error::Io {
+        path: path.to_path_buf(),
+        source,
+    })
+}
+
+/// Whether an agent is listening on the socket at `path`.
+pub(crate) async fn probe(path: &Path) -> bool {
+    UnixStream::connect(path).await.is_ok()
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::time::Duration;
+
+    use super::super::{Request, Response, StatusReport, exchange};
+    use super::*;
+    use crate::BoxFuture;
+
+    #[tokio::test]
+    async fn a_silent_agent_is_reported_rather_than_waited_out() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("control.sock");
+        // A socket that accepts and then says nothing, like a wedged agent:
+        // the connection succeeds, the answer never comes. Unbounded, this
+        // call would never return.
+        let _listener = UnixListener::bind(&path).unwrap();
+
+        let error = exchange(&path, &Request::Status, Duration::from_millis(100))
+            .await
+            .expect_err("a silent agent cannot be reported as healthy");
+
+        match error {
+            Error::Timeout { what } => assert!(
+                what.contains("control.sock"),
+                "the message names what did not answer: {what}"
+            ),
+            other => panic!("expected a timeout, got {other:?}"),
+        }
+    }
+
+    #[tokio::test]
+    async fn an_answered_request_is_not_affected_by_the_bound() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("control.sock");
+        let source: Arc<dyn ReportSource> = Arc::new(|| -> BoxFuture<'static, StatusReport> {
+            Box::pin(async { StatusReport::default() })
+        });
+        let control = ControlSocket::bind(&path, source, ControlSocketAccess::Private)
+            .await
+            .unwrap();
+
+        let answer = exchange(&path, &Request::Status, EXCHANGE_TIMEOUT)
+            .await
+            .unwrap();
+        assert!(matches!(answer, Response::Status(_)));
+
+        control.shutdown().await;
+    }
+}

crates/tsunagi-core/src/ipc/windows.rs

@@ -0,0 +1,453 @@
+//! A named-pipe adapter for the local control interface.
+//!
+//! The Windows counterpart of [`super::unix`]. It provides only what is
+//! particular to a named pipe — creating the server instance, opening a client
+//! and turning the state directory path into a pipe name — and hands every
+//! accepted connection to the shared [`serve_connection`](super::serve_connection).
+//! The framing, the dispatch and the client wrappers are the same on every
+//! platform and live in [`super`], re-exported here so `ipc::windows::request_status`
+//! and its siblings resolve just as their Unix equivalents do.
+//!
+//! # From a path to a pipe
+//!
+//! The rest of the agent addresses the control interface by a filesystem path,
+//! the same one a Unix socket would live at. A named pipe has no filesystem
+//! path, so the path is hashed into a stable name under `\\.\pipe\`. The agent
+//! and a client derive it the same way from the same path, so neither has to
+//! be told where the other put it.
+//!
+//! # Who may connect
+//!
+//! The pipe rejects clients from other machines (`reject_remote_clients(true)`).
+//! An explicit SDDL security descriptor grants Full Control to Administrators
+//! and LocalSystem, while granting Read and Write (`GRGW`) to all local users
+//! with a Medium Integrity SACL label (`S:(ML;;NW;;;ME)`). This allows unprivileged
+//! user-space tools (such as tray applications, status monitors, or CLI commands
+//! launched by regular users) to connect to the agent even when the agent was
+//! started by an elevated Administrator or as a system service.
+//! A permission denial is reported as such, never as an absent agent.
+
+#![allow(unsafe_code)]
+
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::time::Duration;
+
+use sha2::{Digest, Sha256};
+use tokio::net::windows::named_pipe::{ClientOptions, NamedPipeClient, ServerOptions};
+use tokio::task::JoinHandle;
+
+use crate::error::{Error, Result};
+
+pub use super::{
+    CONTROL_PROTOCOL, ControlSocketAccess, EXCHANGE_TIMEOUT, ReportSource, is_serving,
+    join_network, leave_network, request_status, set_active, set_dns, set_hostname,
+};
+
+/// `ERROR_ACCESS_DENIED`: creating a first instance finds an existing pipe,
+/// or a client cannot access an existing pipe with its current privileges.
+const ERROR_ACCESS_DENIED: i32 = 5;
+/// `ERROR_PIPE_BUSY`: every instance is serving a client right now. A server
+/// is there; the client only has to wait for a free instance.
+const ERROR_PIPE_BUSY: i32 = 231;
+
+/// RAII wrapper for Windows named pipe security attributes with permissive local ACL.
+///
+/// SDDL specification:
+/// - `D:(A;;GRGW;;;WD)`: DACL allows Generic Read and Generic Write to Everyone (`WD`).
+///   Full control (`GA`) remains reserved to Administrators and LocalSystem.
+/// - `(A;;GA;;;BA)`: DACL allows Generic All to Builtin Administrators (`BA`).
+/// - `(A;;GA;;;SY)`: DACL allows Generic All to LocalSystem (`SY`).
+/// - `S:(ML;;NW;;;ME)`: SACL Mandatory Label with Medium Integrity (`ME`) and No Write Up (`NW`).
+///   This allows unprivileged user-space processes (Medium Integrity) to write to the
+///   named pipe even when the agent runs as an elevated Administrator (High Integrity).
+struct PipeSecurityAttributes {
+    attrs: windows_sys::Win32::Security::SECURITY_ATTRIBUTES,
+}
+
+impl std::fmt::Debug for PipeSecurityAttributes {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("PipeSecurityAttributes")
+            .field("nLength", &self.attrs.nLength)
+            .field("lpSecurityDescriptor", &self.attrs.lpSecurityDescriptor)
+            .field("bInheritHandle", &self.attrs.bInheritHandle)
+            .finish()
+    }
+}
+
+impl PipeSecurityAttributes {
+    fn new() -> Result<Self> {
+        use std::ffi::OsStr;
+        use std::os::windows::ffi::OsStrExt;
+        use windows_sys::Win32::Security::Authorization::{
+            ConvertStringSecurityDescriptorToSecurityDescriptorW, SDDL_REVISION_1,
+        };
+
+        const SDDL: &str = "D:(A;;GRGW;;;WD)(A;;GA;;;BA)(A;;GA;;;SY)S:(ML;;NW;;;ME)";
+        let wide: Vec<u16> = OsStr::new(SDDL)
+            .encode_wide()
+            .chain(std::iter::once(0))
+            .collect();
+
+        let mut p_sd = std::ptr::null_mut();
+        // SAFETY: We pass a valid null-terminated wide string, standard revision 1,
+        // and a valid pointer to receive the descriptor.
+        let success = unsafe {
+            ConvertStringSecurityDescriptorToSecurityDescriptorW(
+                wide.as_ptr(),
+                SDDL_REVISION_1,
+                &mut p_sd,
+                std::ptr::null_mut(),
+            )
+        };
+
+        if success == 0 || p_sd.is_null() {
+            let err = std::io::Error::last_os_error();
+            return Err(Error::Storage(format!(
+                "cannot create pipe security descriptor: {err}"
+            )));
+        }
+
+        let attrs = windows_sys::Win32::Security::SECURITY_ATTRIBUTES {
+            nLength: std::mem::size_of::<windows_sys::Win32::Security::SECURITY_ATTRIBUTES>()
+                as u32,
+            lpSecurityDescriptor: p_sd,
+            bInheritHandle: 0,
+        };
+
+        Ok(Self { attrs })
+    }
+
+    fn as_raw_mut(&mut self) -> *mut std::ffi::c_void {
+        &mut self.attrs as *mut _ as *mut std::ffi::c_void
+    }
+}
+
+impl Drop for PipeSecurityAttributes {
+    fn drop(&mut self) {
+        if !self.attrs.lpSecurityDescriptor.is_null() {
+            // SAFETY: The descriptor was allocated by ConvertStringSecurityDescriptorToSecurityDescriptorW
+            // and must be freed with LocalFree.
+            unsafe {
+                windows_sys::Win32::Foundation::LocalFree(self.attrs.lpSecurityDescriptor);
+            }
+            self.attrs.lpSecurityDescriptor = std::ptr::null_mut();
+        }
+    }
+}
+
+/// Serves the local control interface on a named pipe.
+#[derive(Debug)]
+pub struct ControlSocket {
+    path: PathBuf,
+    task: Option<JoinHandle<()>>,
+}
+
+impl ControlSocket {
+    /// Binds the pipe and starts serving.
+    ///
+    /// Claiming the first instance of the name is what detects a second agent:
+    /// if the name already exists, the create is refused and that is reported
+    /// as the state being locked, so two live agents never share one pipe.
+    ///
+    /// `_access` is accepted for parity with the Unix socket; granting a group
+    /// here means a pipe security descriptor and is a separate follow-up, so for
+    /// now the pipe keeps its default (creator plus the built-in access the
+    /// `PipeSecurityAttributes` set).
+    pub async fn bind(
+        path: impl AsRef<Path>,
+        source: Arc<dyn ReportSource>,
+        _access: ControlSocketAccess,
+    ) -> Result<Self> {
+        let path = path.as_ref().to_path_buf();
+        let name = pipe_name(&path);
+
+        let mut sec_attrs = PipeSecurityAttributes::new()?;
+        // SAFETY: sec_attrs points to valid initialized SECURITY_ATTRIBUTES.
+        let server = match unsafe {
+            ServerOptions::new()
+                .first_pipe_instance(true)
+                .reject_remote_clients(true)
+                .create_with_security_attributes_raw(&name, sec_attrs.as_raw_mut())
+        } {
+            Ok(server) => server,
+            Err(err) if err.raw_os_error() == Some(ERROR_ACCESS_DENIED) => {
+                return Err(Error::StateLocked { path });
+            }
+            Err(source) => return Err(Error::Io { path, source }),
+        };
+
+        let task = tokio::spawn(serve(name, server, source));
+        Ok(Self {
+            path,
+            task: Some(task),
+        })
+    }
+
+    /// The path the pipe name was derived from.
+    pub fn path(&self) -> &Path {
+        &self.path
+    }
+
+    /// Stops serving. The pipe goes with the server, so there is nothing to
+    /// remove.
+    pub async fn shutdown(mut self) {
+        if let Some(task) = self.task.take() {
+            task.abort();
+            let _ = task.await;
+        }
+    }
+}
+
+impl Drop for ControlSocket {
+    fn drop(&mut self) {
+        if let Some(task) = self.task.take() {
+            task.abort();
+        }
+    }
+}
+
+/// Accepts one client at a time, keeping a fresh instance ready for the next.
+///
+/// A named pipe server instance serves a single client, so a new instance is
+/// created as soon as one is taken — otherwise a second `tsng status` while
+/// the first is mid-flight would find nothing listening.
+async fn serve(
+    name: String,
+    first: tokio::net::windows::named_pipe::NamedPipeServer,
+    source: Arc<dyn ReportSource>,
+) {
+    let mut server = first;
+    loop {
+        if server.connect().await.is_err() {
+            match next_instance(&name) {
+                Some(next) => {
+                    server = next;
+                    continue;
+                }
+                None => return,
+            }
+        }
+        let connected = server;
+        match next_instance(&name) {
+            Some(next) => server = next,
+            None => {
+                // Nothing left to accept the next client on, but the one in
+                // hand is still answered before the loop ends.
+                let source = Arc::clone(&source);
+                tokio::spawn(async move {
+                    let _ = super::serve_connection(connected, source).await;
+                });
+                return;
+            }
+        }
+        let source = Arc::clone(&source);
+        tokio::spawn(async move {
+            if let Err(err) = super::serve_connection(connected, source).await {
+                tracing::debug!(%err, "local control request failed");
+            }
+        });
+    }
+}
+
+/// Creates the next pipe instance, or `None` if the name can no longer be
+/// served.
+fn next_instance(name: &str) -> Option<tokio::net::windows::named_pipe::NamedPipeServer> {
+    let mut sec_attrs = match PipeSecurityAttributes::new() {
+        Ok(attrs) => attrs,
+        Err(err) => {
+            tracing::debug!(%err, "cannot build pipe security attributes for next instance");
+            return None;
+        }
+    };
+    // SAFETY: sec_attrs points to valid initialized SECURITY_ATTRIBUTES.
+    match unsafe {
+        ServerOptions::new()
+            .reject_remote_clients(true)
+            .create_with_security_attributes_raw(name, sec_attrs.as_raw_mut())
+    } {
+        Ok(server) => Some(server),
+        Err(err) => {
+            tracing::debug!(%err, "cannot create the next control pipe instance");
+            None
+        }
+    }
+}
+
+/// Connects a client to the pipe for `path`, waiting out a busy pipe.
+///
+/// The wait is bounded by the caller's exchange timeout, so a pipe that is
+/// busy forever is given up on rather than spun on.
+pub(crate) async fn connect(path: &Path) -> Result<NamedPipeClient> {
+    let name = pipe_name(path);
+    loop {
+        match ClientOptions::new().open(&name) {
+            Ok(client) => return Ok(client),
+            Err(err) if err.raw_os_error() == Some(ERROR_PIPE_BUSY) => {
+                tokio::time::sleep(Duration::from_millis(50)).await;
+            }
+            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => {
+                return Err(Error::Io {
+                    path: path.to_path_buf(),
+                    source: std::io::Error::new(
+                        std::io::ErrorKind::PermissionDenied,
+                        format!(
+                            "the agent's control pipe exists, but Windows denied access. \
+                             Run this command as the same Windows user and with the same \
+                             elevation as `tsng up` (use an administrator terminal if \
+                             the agent is elevated). Windows error: {err}"
+                        ),
+                    ),
+                });
+            }
+            Err(source) => {
+                return Err(Error::Io {
+                    path: path.to_path_buf(),
+                    source,
+                });
+            }
+        }
+    }
+}
+
+/// Whether an agent is serving the pipe for `path`.
+///
+/// A busy or inaccessible pipe still means a server is there. In particular,
+/// access denied must not send a mutating command down the offline path,
+/// where it would only produce a misleading state-directory lock error.
+pub(crate) async fn probe(path: &Path) -> bool {
+    let name = pipe_name(path);
+    match ClientOptions::new().open(&name) {
+        Ok(_) => true,
+        Err(err) => {
+            err.raw_os_error() == Some(ERROR_PIPE_BUSY)
+                || err.kind() == std::io::ErrorKind::PermissionDenied
+        }
+    }
+}
+
+/// The pipe name for a control-socket path.
+///
+/// A hash rather than the path itself, because a pipe name may not contain the
+/// separators and drive letters a path does, and because two agents with
+/// different state directories must never collide.
+fn pipe_name(path: &Path) -> String {
+    let digest = Sha256::digest(path.as_os_str().as_encoded_bytes());
+    format!(r"\\.\pipe\tsunagi-{}", hex::encode(&digest[..16]))
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::time::Duration;
+
+    use super::super::{Request, Response, StatusReport, exchange};
+    use super::*;
+    use crate::BoxFuture;
+
+    #[test]
+    fn one_path_maps_to_one_name_and_two_paths_do_not_collide() {
+        let a = pipe_name(Path::new(r"C:\a\agent.sock"));
+        let b = pipe_name(Path::new(r"C:\b\agent.sock"));
+        assert!(a.starts_with(r"\\.\pipe\tsunagi-"), "{a}");
+        assert_eq!(a, pipe_name(Path::new(r"C:\a\agent.sock")));
+        assert_ne!(a, b);
+    }
+
+    #[tokio::test]
+    async fn a_request_makes_the_round_trip_over_a_real_pipe() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("agent.sock");
+        let source: Arc<dyn ReportSource> = Arc::new(|| -> BoxFuture<'static, StatusReport> {
+            Box::pin(async { StatusReport::default() })
+        });
+
+        assert!(!is_serving(&path).await, "nothing serves the pipe yet");
+        let control = ControlSocket::bind(&path, source).await.unwrap();
+        assert!(is_serving(&path).await, "the agent serves it now");
+
+        let answer = exchange(&path, &Request::Status, EXCHANGE_TIMEOUT)
+            .await
+            .unwrap();
+        assert!(matches!(answer, Response::Status(_)));
+
+        control.shutdown().await;
+    }
+
+    #[tokio::test]
+    async fn an_inaccessible_agent_is_not_mistaken_for_an_absent_one() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("agent.sock");
+        // An outbound-only pipe denies the client's duplex open even when
+        // both processes have the same privileges. This exercises the same
+        // Windows error as a client unable to access an elevated agent,
+        // without requiring elevation or changing any ACLs.
+        let _server = ServerOptions::new()
+            .first_pipe_instance(true)
+            .access_inbound(false)
+            .create(pipe_name(&path))
+            .unwrap();
+
+        assert!(is_serving(&path).await, "access denied is not absence");
+        for error in [
+            request_status(&path).await.unwrap_err(),
+            join_network(&path, "test", "local-ipc-test-secret")
+                .await
+                .unwrap_err(),
+            set_dns(&path, false, None).await.unwrap_err(),
+        ] {
+            let Error::Io { source, .. } = error else {
+                panic!("expected an access error, got {error}");
+            };
+            assert_eq!(source.kind(), std::io::ErrorKind::PermissionDenied);
+            let message = source.to_string();
+            assert!(message.contains("same Windows user"), "{message}");
+            assert!(message.contains("administrator"), "{message}");
+            assert!(!message.contains("local-ipc-test-secret"), "{message}");
+        }
+    }
+
+    #[tokio::test]
+    async fn a_silent_agent_is_reported_rather_than_waited_out() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("agent.sock");
+        // A server instance that accepts and then never answers, like a
+        // wedged agent. Unbounded, the exchange below would never return.
+        let name = pipe_name(&path);
+        let server = ServerOptions::new()
+            .first_pipe_instance(true)
+            .create(&name)
+            .unwrap();
+        let _accept = tokio::spawn(async move {
+            let _ = server.connect().await;
+            // Hold the connection open, answering nothing.
+            tokio::time::sleep(Duration::from_secs(30)).await;
+        });
+
+        let error = exchange(&path, &Request::Status, Duration::from_millis(200))
+            .await
+            .expect_err("a silent agent cannot be reported as healthy");
+        assert!(matches!(error, Error::Timeout { .. }), "{error:?}");
+    }
+
+    #[tokio::test]
+    async fn a_second_agent_on_the_same_pipe_is_refused() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("agent.sock");
+        let source: Arc<dyn ReportSource> = Arc::new(|| -> BoxFuture<'static, StatusReport> {
+            Box::pin(async { StatusReport::default() })
+        });
+
+        let first = ControlSocket::bind(&path, Arc::clone(&source))
+            .await
+            .unwrap();
+        let second = ControlSocket::bind(&path, source).await;
+        assert!(
+            matches!(second, Err(Error::StateLocked { .. })),
+            "{second:?}"
+        );
+
+        first.shutdown().await;
+    }
+}

crates/tsunagi-core/src/lib.rs

@@ -44,9 +44,13 @@ pub mod error;
 pub mod identity;
 pub mod ipc;
 pub mod net;
+pub mod overlay;
 pub mod proto;
 pub mod state;
 pub mod storage;
+pub mod task;
+#[cfg(feature = "testing")]
+pub mod testing;
 
 pub use agent::{Agent, AgentStatus, Event, NetworkStatus, PeerStatus};
 pub use config::{AgentConfig, Limits, ReconnectPolicy, StoragePaths, TransportPolicy};

crates/tsunagi-core/src/net.rs

@@ -240,6 +240,62 @@ impl EndpointAdapter {
             TransportPolicy::N0Defaults => builder.preset(presets::N0),
         };
 
+        if matches!(
+            config.transport,
+            TransportPolicy::N0Defaults | TransportPolicy::DirectOnly
+        ) {
+            let filter = iroh::address_lookup::AddrFilter::new(|addrs| {
+                let mut overlay_ips = std::collections::HashSet::new();
+                for iface in netdev::get_interfaces() {
+                    let name = iface.name.to_lowercase();
+                    let friendly = iface
+                        .friendly_name
+                        .as_deref()
+                        .map(str::to_lowercase)
+                        .unwrap_or_default();
+                    let desc = iface
+                        .description
+                        .as_deref()
+                        .map(str::to_lowercase)
+                        .unwrap_or_default();
+                    if name.starts_with("tsun")
+                        || name.contains("tsunagi")
+                        || name.contains("wintun")
+                        || friendly.starts_with("tsun")
+                        || friendly.contains("tsunagi")
+                        || friendly.contains("wintun")
+                        || desc.contains("wintun")
+                        || desc.contains("tsunagi")
+                    {
+                        for ip in iface.ipv4 {
+                            overlay_ips.insert(std::net::IpAddr::V4(ip.addr()));
+                        }
+                        for ip in iface.ipv6 {
+                            overlay_ips.insert(std::net::IpAddr::V6(ip.addr()));
+                        }
+                    }
+                }
+                if overlay_ips.is_empty() {
+                    return std::borrow::Cow::Borrowed(addrs);
+                }
+                let filtered: Vec<iroh::TransportAddr> = addrs
+                    .iter()
+                    .filter(|addr| match addr {
+                        iroh::TransportAddr::Ip(sock) => !overlay_ips.contains(&sock.ip()),
+                        _ => true,
+                    })
+                    .cloned()
+                    .collect();
+                std::borrow::Cow::Owned(filtered)
+            });
+            builder = builder.addr_filter(filter);
+        }
+
+        #[cfg(feature = "testing")]
+        if let Some(transport) = &config.test_quic_transport {
+            builder = builder.transport_config(transport.clone());
+        }
+
         if !config.bind_addrs.is_empty() {
             builder = builder.clear_ip_transports();
             for addr in &config.bind_addrs {
@@ -254,6 +310,91 @@ impl EndpointAdapter {
             .await
             .map_err(|err| Error::Endpoint(format!("cannot bind endpoint: {err}")))?;
 
+        tracing::debug!(
+            endpoint = %endpoint.id().fmt_short(),
+            bound_sockets = ?endpoint.bound_sockets(),
+            "iroh endpoint bound locally"
+        );
+
+        let ep_id = endpoint.id();
+        let mut addr_watcher = endpoint.watch_addr();
+        let mut relay_watcher = endpoint.home_relay_status();
+        let close_token = endpoint.clone();
+        tokio::spawn(async move {
+            use iroh::Watcher;
+            let mut last_addrs: Vec<SocketAddr> = Vec::new();
+            let mut last_relays: Vec<String> = Vec::new();
+
+            let warn_timer = tokio::time::sleep(Duration::from_secs(12));
+            tokio::pin!(warn_timer);
+            let mut warned = false;
+
+            loop {
+                tokio::select! {
+                    biased;
+                    _ = close_token.closed() => {
+                        break;
+                    }
+                    () = &mut warn_timer, if !warned => {
+                        warned = true;
+                        if last_addrs.is_empty() && !last_relays.is_empty() {
+                            tracing::warn!(
+                                endpoint = %ep_id.fmt_short(),
+                                relays = ?last_relays,
+                                "No direct reflexive or UPnP addresses discovered after probe; communications will rely on relay servers"
+                            );
+                        } else if last_addrs.is_empty() && last_relays.is_empty() {
+                            tracing::warn!(
+                                endpoint = %ep_id.fmt_short(),
+                                "No reachability discovered (neither direct reflexive/UPnP addresses nor relay servers available)"
+                            );
+                        }
+                    }
+                    res = addr_watcher.updated() => {
+                        let addr = match res {
+                            Ok(a) => a,
+                            Err(_) => break,
+                        };
+                        let addrs: Vec<SocketAddr> = addr.ip_addrs().copied().collect();
+                        let relays: Vec<String> = addr.relay_urls().map(|u| u.to_string()).collect();
+
+                        if addrs != last_addrs || relays != last_relays {
+                            tracing::debug!(
+                                endpoint = %ep_id.fmt_short(),
+                                direct_addrs = ?addrs,
+                                relays = ?relays,
+                                "Discovered endpoint reachability (STUN/UPnP/local/relay)"
+                            );
+                            last_addrs = addrs;
+                            last_relays = relays;
+                        }
+                    }
+                    res = relay_watcher.updated() => {
+                        let relays = match res {
+                            Ok(r) => r,
+                            Err(_) => break,
+                        };
+                        for relay in &relays {
+                            if relay.is_connected() {
+                                tracing::debug!(
+                                    endpoint = %ep_id.fmt_short(),
+                                    relay = %relay.url(),
+                                    "Connected to home relay server"
+                                );
+                            } else if let Some(err) = relay.last_error() {
+                                tracing::debug!(
+                                    endpoint = %ep_id.fmt_short(),
+                                    relay = %relay.url(),
+                                    error = %err,
+                                    "Relay connection attempt failed"
+                                );
+                            }
+                        }
+                    }
+                }
+            }
+        });
+
         Ok(Self { endpoint })
     }
 

crates/tsunagi-core/src/overlay/broadcast.rs

@@ -0,0 +1,231 @@
+//! Scoped LAN discovery fanout at IP ingress. A domain is one logical network;
+//! routed copies use ordinary encrypted peer links and are never reflooded.
+//! Future subnet exporters must supply explicitly authorized ingress domains
+//! here, not teach the encrypted transit router to inspect application bytes.
+
+use super::router::{NetworkRoutes, Route};
+use crate::NetworkId;
+use iroh::EndpointId;
+use std::collections::{HashMap, HashSet};
+use std::net::Ipv4Addr;
+use std::sync::Arc;
+
+/// Local participation and the currently authenticated participants in a domain.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct BroadcastPolicy {
+    /// Whether this agent originates and accepts LAN broadcasts in this network.
+    pub enabled: bool,
+    /// Live members explicitly advertising that they accept broadcasts.
+    pub peers: HashSet<EndpointId>,
+}
+impl Default for BroadcastPolicy {
+    fn default() -> Self {
+        Self {
+            enabled: true,
+            peers: HashSet::new(),
+        }
+    }
+}
+
+#[derive(Debug)]
+struct Domain {
+    network: NetworkId,
+    directed: Ipv4Addr,
+    enabled: bool,
+    recipients: Arc<[Route]>,
+    members: HashSet<EndpointId>,
+}
+
+/// Immutable ingress-domain index rebuilt with address/policy changes.
+#[derive(Debug, Default)]
+pub(crate) struct BroadcastTable {
+    sources: HashMap<Ipv4Addr, Arc<Domain>>,
+    networks: HashMap<NetworkId, Arc<Domain>>,
+    destinations: HashSet<Ipv4Addr>,
+}
+impl BroadcastTable {
+    pub fn build(networks: &HashMap<NetworkId, NetworkRoutes>) -> Self {
+        let mut table = Self::default();
+        for (&network, routes) in networks {
+            let (Some(range), Some(local)) = (routes.range, routes.local) else {
+                continue;
+            };
+            if range.prefix_len > 30 {
+                continue;
+            }
+            let directed = Ipv4Addr::from(u32::from(range.base) | (u32::MAX >> range.prefix_len));
+            let mut recipients: Vec<_> = routes
+                .peers
+                .iter()
+                .filter(|(_, peer)| routes.broadcast.peers.contains(peer))
+                .map(|(_, peer)| Route {
+                    network,
+                    peer: *peer,
+                })
+                .collect();
+            recipients.sort_by_key(|route| route.peer);
+            recipients.dedup();
+            let domain = Arc::new(Domain {
+                network,
+                directed,
+                enabled: routes.broadcast.enabled,
+                members: recipients.iter().map(|route| route.peer).collect(),
+                recipients: recipients.into(),
+            });
+            table.sources.insert(local, domain.clone());
+            table.networks.insert(network, domain);
+            table.destinations.insert(directed);
+        }
+        table
+    }
+    pub fn is_destination(&self, destination: Ipv4Addr) -> bool {
+        destination.is_broadcast() || self.destinations.contains(&destination)
+    }
+    pub fn outgoing(&self, source: Ipv4Addr, destination: Ipv4Addr) -> Option<Arc<[Route]>> {
+        let domain = self.sources.get(&source)?;
+        (domain.enabled && (destination.is_broadcast() || destination == domain.directed))
+            .then(|| domain.recipients.clone())
+    }
+    pub fn accepts(&self, network: NetworkId, peer: EndpointId, destination: Ipv4Addr) -> bool {
+        self.networks.get(&network).is_some_and(|domain| {
+            domain.network == network
+                && domain.enabled
+                && domain.members.contains(&peer)
+                && (destination.is_broadcast() || destination == domain.directed)
+        })
+    }
+}
+
+/// Validates IPv4/UDP lengths before broadcast fanout. No game-specific ports,
+/// payload rewriting or checksum changes. IP fragments retain their bytes and
+/// are reassembled by the destination OS; even later fragments have protocol 17.
+pub(crate) fn valid_udp(packet: &[u8]) -> bool {
+    if packet.len() < 20 || packet[0] >> 4 != 4 || packet[9] != 17 {
+        return false;
+    }
+    let header = usize::from(packet[0] & 15) * 4;
+    let total = usize::from(u16::from_be_bytes([packet[2], packet[3]]));
+    if header < 20 || total <= header || total > packet.len() {
+        return false;
+    }
+    let fragment = u16::from_be_bytes([packet[6], packet[7]]);
+    if fragment & 0x1fff != 0 {
+        return true;
+    }
+    if total < header + 8 {
+        return false;
+    }
+    let udp = usize::from(u16::from_be_bytes([packet[header + 4], packet[header + 5]]));
+    udp >= 8
+        && if fragment & 0x2000 != 0 {
+            udp >= total - header
+        } else {
+            udp == total - header
+        }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use crate::identity::{NetworkKeys, NetworkName, NetworkSecret};
+    fn id(name: &str) -> NetworkId {
+        NetworkKeys::derive(
+            &NetworkName::new(name).unwrap(),
+            &NetworkSecret::from_bytes([4; 32]).unwrap(),
+        )
+        .network_id()
+    }
+    #[test]
+    fn limited_and_directed_broadcasts_never_cross_domains_or_target_disabled_peers() {
+        let peer = iroh::SecretKey::from_bytes(&[8; 32]).public();
+        let a = NetworkRoutes {
+            exit: Default::default(),
+            range: Some("10.0.0.0/24".parse().unwrap()),
+            local: Some("10.0.0.1".parse().unwrap()),
+            peers: vec![("10.0.0.2".parse().unwrap(), peer)],
+            broadcast: BroadcastPolicy {
+                enabled: true,
+                peers: HashSet::from([peer]),
+            },
+        };
+        let b = NetworkRoutes {
+            exit: Default::default(),
+            range: Some("10.1.0.0/24".parse().unwrap()),
+            local: Some("10.1.0.1".parse().unwrap()),
+            peers: vec![("10.1.0.2".parse().unwrap(), peer)],
+            broadcast: BroadcastPolicy::default(),
+        };
+        let table = BroadcastTable::build(&HashMap::from([(id("a"), a.clone()), (id("b"), b)]));
+        let routes = table
+            .outgoing(a.local.unwrap(), Ipv4Addr::BROADCAST)
+            .unwrap();
+        assert_eq!(
+            &*routes,
+            &[Route {
+                network: id("a"),
+                peer
+            }]
+        );
+        assert!(
+            table
+                .outgoing(a.local.unwrap(), "10.1.0.255".parse().unwrap())
+                .is_none()
+        );
+        assert!(
+            table
+                .outgoing("192.168.1.2".parse().unwrap(), Ipv4Addr::BROADCAST)
+                .is_none()
+        );
+        assert!(
+            table
+                .outgoing("10.1.0.1".parse().unwrap(), Ipv4Addr::BROADCAST)
+                .unwrap()
+                .is_empty()
+        );
+        assert!(!table.accepts(id("a"), peer, "10.1.0.255".parse().unwrap()));
+        assert!(!table.accepts(id("b"), peer, Ipv4Addr::BROADCAST));
+        assert!(table.accepts(id("a"), peer, Ipv4Addr::BROADCAST));
+        let disabled = NetworkRoutes {
+            exit: Default::default(),
+            broadcast: BroadcastPolicy {
+                enabled: false,
+                ..a.broadcast.clone()
+            },
+            ..a.clone()
+        };
+        let table = BroadcastTable::build(&HashMap::from([(id("a"), disabled)]));
+        assert!(
+            table
+                .outgoing(a.local.unwrap(), Ipv4Addr::BROADCAST)
+                .is_none()
+        );
+        assert!(!table.accepts(id("a"), peer, Ipv4Addr::BROADCAST));
+    }
+    #[test]
+    fn malformed_udp_is_not_amplified_but_ipv4_fragments_are_supported() {
+        let mut packet = vec![0u8; 36];
+        packet[0] = 0x45;
+        packet[3] = 36;
+        packet[9] = 17;
+        packet[25] = 16;
+        assert!(valid_udp(&packet));
+        for length in 0..36 {
+            assert!(!valid_udp(&packet[..length]));
+        }
+        packet[9] = 6;
+        assert!(!valid_udp(&packet));
+        packet[9] = 17;
+        packet[0] = 0x44;
+        assert!(!valid_udp(&packet));
+        packet[0] = 0x45;
+        packet[25] = 8;
+        assert!(!valid_udp(&packet));
+        packet[6] = 0x20;
+        packet[25] = 64;
+        assert!(valid_udp(&packet));
+        packet[6] = 0;
+        packet[7] = 2;
+        assert!(valid_udp(&packet));
+    }
+}

crates/tsunagi-core/src/overlay/config.rs

@@ -11,10 +11,11 @@
 
 use std::net::{IpAddr, Ipv6Addr};
 
-use crate::dataplane::PluginError;
 use crate::identity::NetworkId;
+use crate::overlay::OverlayError;
 
-use super::overlay::OVERLAY_HOST_PREFIX_LEN;
+/// Prefix length of a single IPv6 host address.
+const HOST_PREFIX_LEN: u8 = 128;
 
 /// Longest interface name Linux accepts, excluding the terminating NUL.
 pub const MAX_INTERFACE_NAME_LEN: usize = 15;
@@ -33,13 +34,13 @@ pub struct Cidr {
 
 impl Cidr {
     /// Builds a CIDR, rejecting an impossible prefix length.
-    pub fn new(addr: IpAddr, prefix_len: u8) -> Result<Self, PluginError> {
+    pub fn new(addr: IpAddr, prefix_len: u8) -> Result<Self, OverlayError> {
         let max = match addr {
             IpAddr::V4(_) => 32,
             IpAddr::V6(_) => 128,
         };
         if prefix_len > max {
-            return Err(PluginError::Other(format!(
+            return Err(OverlayError::Other(format!(
                 "prefix length /{prefix_len} is impossible for {addr}"
             )));
         }
@@ -50,7 +51,7 @@ impl Cidr {
     pub fn host(addr: Ipv6Addr) -> Self {
         Self {
             addr: IpAddr::V6(addr),
-            prefix_len: OVERLAY_HOST_PREFIX_LEN,
+            prefix_len: HOST_PREFIX_LEN,
         }
     }
 }
@@ -66,9 +67,9 @@ impl std::fmt::Display for Cidr {
 /// The name is stable across restarts and short enough for the platform. Two
 /// agents on the same host in the same network must be given different
 /// prefixes, or they would derive the same name.
-pub fn interface_name(prefix: &str, network: NetworkId) -> Result<String, PluginError> {
+pub fn interface_name(prefix: &str, network: NetworkId) -> Result<String, OverlayError> {
     if prefix.is_empty() {
-        return Err(PluginError::Other(
+        return Err(OverlayError::Other(
             "interface prefix must not be empty".into(),
         ));
     }
@@ -76,12 +77,12 @@ pub fn interface_name(prefix: &str, network: NetworkId) -> Result<String, Plugin
         .chars()
         .all(|c| c.is_ascii_lowercase() || c.is_ascii_digit())
     {
-        return Err(PluginError::Other(
+        return Err(OverlayError::Other(
             "interface prefix must be lowercase ASCII letters and digits".into(),
         ));
     }
     if prefix.len() >= MAX_INTERFACE_NAME_LEN {
-        return Err(PluginError::Other(format!(
+        return Err(OverlayError::Other(format!(
             "interface prefix must be shorter than {MAX_INTERFACE_NAME_LEN} characters"
         )));
     }

crates/tsunagi-core/src/overlay/exit.rs

@@ -0,0 +1,32 @@
+//! Exit-node host rules: what makes this device an exit node for others, and
+//! what sends this device's own traffic through one.
+//!
+//! The packet side lives in [`super::router`]; this is the side the operating
+//! system sees. It is kept apart from the broadcast rules, which have a
+//! different lifetime and different tags.
+
+/// How one set of host rules went.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct RuleSetReport {
+    /// Whether every step took.
+    pub ok: bool,
+    /// One line: what is in place, or what is missing and why.
+    pub detail: String,
+}
+
+/// What the host rules of one network's exit-node settings are doing.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct ExitRulesReport {
+    /// Masquerading and forwarding for members' traffic, when this agent
+    /// offers itself as an exit node here. `None` means none are wanted.
+    pub offer: Option<RuleSetReport>,
+    /// Whether the kernel forwards packets for the overlay interface.
+    ///
+    /// `Some(false)` means the rules are in place but nothing passes until
+    /// forwarding is turned on, which the agent never does itself. `None`
+    /// when it could not be read.
+    pub forwarding: Option<bool>,
+    /// The routes that send this device's traffic through its exit node.
+    /// `None` means it uses none through this network.
+    pub client: Option<RuleSetReport>,
+}

crates/tsunagi-core/src/overlay/hostrules/exit.rs

@@ -0,0 +1,263 @@
+//! Host rules for exit nodes: what the operating system needs for this agent
+//! to carry other members' internet traffic, and for this device's own
+//! traffic to go through a member.
+//!
+//! Two independent halves, planned together because one interface carries
+//! both:
+//!
+//! * **Offering** — one set of rules per network that offers this agent as an
+//!   exit node, chosen by that network's range: masquerade the range out of
+//!   every other interface and let it through the forward chain. A network
+//!   that does not offer has no rules, so its members' packets are dropped by
+//!   the firewall as well as by the agent.
+//! * **Using** — policy routing that sends everything the host does not have
+//!   a more specific route for into the overlay interface, while the agent's
+//!   own sockets keep the ordinary route (see `linux_exit`).
+//!
+//! Nothing a peer announces becomes a host setting. Offering is the owner's
+//! switch; using is the user's choice of a member. Every object is tagged,
+//! reversible, and removed when it is no longer wanted, when the network
+//! leaves or stops, and when the agent stops.
+
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+use crate::state::Ipv4Range;
+
+use super::RuleOutcome;
+
+/// What the host should hold for one interface's exit settings.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct ExitHostPlan {
+    /// The overlay interface.
+    pub interface: String,
+    /// The ranges of every network that offers this agent as an exit node.
+    pub offer: Vec<Ipv4Range>,
+    /// Whether this device sends its internet traffic through a member.
+    pub client: bool,
+    /// The range of every network on the interface. Only the using side
+    /// reads it, to keep this agent's own traffic *to* the overlay inside it
+    /// where the host has no per-user routing (macOS).
+    pub overlay: Vec<Ipv4Range>,
+    /// Underlay addresses this agent talks to directly and must keep reaching
+    /// outside the overlay's default route. Only filled while `client` is set
+    /// and only for hosts that report [`ExitHostRules::needs_bypass`].
+    pub bypass: Vec<std::net::Ipv4Addr>,
+    /// Host names (relay servers) that must stay reachable the same way; the
+    /// host rules resolve them.
+    pub bypass_hosts: Vec<String>,
+}
+
+impl ExitHostPlan {
+    /// Whether there is nothing to hold.
+    pub fn is_empty(&self) -> bool {
+        self.offer.is_empty() && !self.client
+    }
+}
+
+/// What applying a plan achieved.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct ExitHostReport {
+    /// One outcome per offered range: its masquerade and forward rules.
+    pub offer: Vec<(Ipv4Range, RuleOutcome)>,
+    /// Whether the kernel forwards packets for the interface. `None` when it
+    /// could not be read. The agent only reads this: turning forwarding on is
+    /// the owner's decision.
+    pub forwarding: Option<bool>,
+    /// The routes and rules that send this device's traffic through the
+    /// interface, when the plan asked for them.
+    pub client: Option<RuleOutcome>,
+}
+
+/// Installs and removes exit-node rules for an overlay interface.
+///
+/// Like the broadcast rules, no call fails as a whole: a step that cannot be
+/// done is in the report, because the caller's only useful response is to
+/// say so.
+pub trait ExitHostRules: Send + Sync + std::fmt::Debug + 'static {
+    /// Makes the host match the plan, replacing what this agent put there
+    /// for the interface before.
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport>;
+
+    /// Removes everything this agent put there for the interface. Succeeds
+    /// when nothing is there: it runs on the shutdown path, and once at
+    /// startup to take away what a crashed run left.
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()>;
+
+    /// Whether using an exit node on this host needs the underlay addresses
+    /// the agent talks to, because the host cannot exempt the agent's own
+    /// traffic from the default route by user id (Windows). The default is no.
+    fn needs_bypass(&self) -> bool {
+        false
+    }
+}
+
+/// The tag every firewall object of one offered range carries.
+pub fn exit_tag(interface: &str, range: &Ipv4Range) -> String {
+    format!("{}{range}", exit_tag_prefix(interface))
+}
+
+/// The start of every tag for an interface; what cleanup selects by.
+pub fn exit_tag_prefix(interface: &str) -> String {
+    format!("tsunagi-exit:{interface}:")
+}
+
+/// The one-line detail for a set of offer outcomes.
+pub fn offer_detail(outcome: &RuleOutcome, range: &Ipv4Range) -> String {
+    match outcome {
+        RuleOutcome::Applied => format!("masquerade + forward for {range} in place"),
+        RuleOutcome::Failed(reason) => reason.clone(),
+    }
+}
+
+/// An in-memory host for exit rules, for tests: nothing is touched.
+#[derive(Debug, Clone, Default)]
+pub struct MockExitRules {
+    state: Arc<Mutex<Option<ExitHostPlan>>>,
+    calls: Arc<Mutex<Vec<String>>>,
+    forwarding: Arc<Mutex<Option<bool>>>,
+    failure: Arc<Mutex<Option<String>>>,
+    needs_bypass: Arc<Mutex<bool>>,
+}
+
+impl MockExitRules {
+    /// An empty host with forwarding turned on.
+    pub fn new() -> Self {
+        let rules = Self::default();
+        *lock(&rules.forwarding) = Some(true);
+        rules
+    }
+
+    /// What the host reports for kernel forwarding from now on.
+    pub fn set_forwarding(&self, forwarding: Option<bool>) {
+        *lock(&self.forwarding) = forwarding;
+    }
+
+    /// Whether this host asks for the underlay addresses, like Windows.
+    pub fn set_needs_bypass(&self, needs: bool) {
+        *lock(&self.needs_bypass) = needs;
+    }
+
+    /// Makes every step fail with this reason from now on.
+    pub fn fail_with(&self, reason: Option<String>) {
+        *lock(&self.failure) = reason;
+    }
+
+    /// What is installed now.
+    pub fn installed(&self) -> Option<ExitHostPlan> {
+        lock(&self.state).clone()
+    }
+
+    /// Every `apply:<interface>` / `clear:<interface>` call, in order.
+    pub fn calls(&self) -> Vec<String> {
+        lock(&self.calls).clone()
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+impl ExitHostRules for MockExitRules {
+    fn needs_bypass(&self) -> bool {
+        *lock(&self.needs_bypass)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport> {
+        Box::pin(async move {
+            lock(&self.calls).push(format!("apply:{}", plan.interface));
+            let outcome = match lock(&self.failure).clone() {
+                Some(reason) => RuleOutcome::Failed(reason),
+                None => {
+                    *lock(&self.state) = Some(plan.clone());
+                    RuleOutcome::Applied
+                }
+            };
+            ExitHostReport {
+                offer: plan
+                    .offer
+                    .iter()
+                    .map(|range| (*range, outcome.clone()))
+                    .collect(),
+                forwarding: *lock(&self.forwarding),
+                client: plan.client.then_some(outcome),
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            lock(&self.calls).push(format!("clear:{interface}"));
+            *lock(&self.state) = None;
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    fn range() -> Ipv4Range {
+        "10.13.37.0/24".parse().unwrap()
+    }
+
+    #[test]
+    fn tags_name_the_interface_and_the_range() {
+        assert_eq!(
+            exit_tag("tsun0", &range()),
+            "tsunagi-exit:tsun0:10.13.37.0/24"
+        );
+        assert!(exit_tag("tsun0", &range()).starts_with(&exit_tag_prefix("tsun0")));
+        // Another interface's tag never matches this one's prefix.
+        assert!(!exit_tag("tsun01", &range()).starts_with(&exit_tag_prefix("tsun0")));
+    }
+
+    #[test]
+    fn an_empty_plan_is_nothing_to_hold() {
+        let mut plan = ExitHostPlan::default();
+        assert!(plan.is_empty());
+        plan.client = true;
+        assert!(!plan.is_empty());
+    }
+
+    #[tokio::test]
+    async fn the_mock_applies_reports_and_clears() {
+        let rules = MockExitRules::new();
+        let plan = ExitHostPlan {
+            interface: "tsun0".into(),
+            offer: vec![range()],
+            client: true,
+            overlay: vec![range()],
+            ..Default::default()
+        };
+        let report = rules.apply(&plan).await;
+        assert_eq!(report.offer[0].1, RuleOutcome::Applied);
+        assert_eq!(report.forwarding, Some(true));
+        assert_eq!(report.client, Some(RuleOutcome::Applied));
+        assert_eq!(rules.installed(), Some(plan));
+
+        rules.set_forwarding(Some(false));
+        rules.fail_with(Some("denied".into()));
+        let report = rules
+            .apply(&ExitHostPlan {
+                interface: "tsun0".into(),
+                offer: vec![range()],
+                client: false,
+                overlay: Vec::new(),
+                ..Default::default()
+            })
+            .await;
+        assert_eq!(report.forwarding, Some(false));
+        assert!(matches!(report.offer[0].1, RuleOutcome::Failed(_)));
+        assert_eq!(report.client, None);
+
+        rules.clear("tsun0").await;
+        assert_eq!(rules.installed(), None);
+        assert_eq!(rules.calls(), ["apply:tsun0", "apply:tsun0", "clear:tsun0"]);
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/linux.rs

@@ -0,0 +1,687 @@
+//! Broadcast host rules on Linux: a netlink route and an `iptables` rule.
+//!
+//! The route is added by the netlink thread the interface provisioner already
+//! owns (see [`super::super::provision`]), because that thread is the only one
+//! that ever holds `CAP_NET_ADMIN`. It belongs to the interface and goes away
+//! with it.
+//!
+//! The firewall allowance is a rule in the `INPUT` chain, inserted at the top
+//! so a default-deny chain further down cannot drop discovery replies first:
+//!
+//! ```text
+//! iptables -I INPUT 1 -i <if> -p udp -s <range> -m comment --comment tsunagi:<if> -j ACCEPT
+//! ```
+//!
+//! `iptables` is a program, not a netlink call, so it is looked up in fixed
+//! system directories rather than on `PATH`, and its arguments are built from
+//! values this agent computed. It works with both the legacy backend and
+//! `iptables-nft`. A host with only `nft` or `firewalld` has no `iptables`;
+//! that is reported on the firewall half and the route is still installed.
+//!
+//! The rule carries a comment tag. That is how a rule left behind by a crashed
+//! run is found and replaced, and why nothing else in the chain is ever
+//! deleted.
+
+use std::net::Ipv4Addr;
+
+use crate::BoxFuture;
+use crate::overlay::provision::RouteHandle;
+
+use super::{BroadcastHostRules, BroadcastRulesPlan, BroadcastRulesReport, RuleOutcome, tag_for};
+
+const CHAIN: &str = "INPUT";
+
+/// Fixed places `iptables` is installed. Never `PATH`: the agent is
+/// privileged and a directory a user controls must not decide what it runs.
+const IPTABLES_DIRECTORIES: [&str; 5] = [
+    "/usr/sbin/iptables",
+    "/sbin/iptables",
+    "/usr/bin/iptables",
+    "/bin/iptables",
+    "/usr/local/sbin/iptables",
+];
+
+/// The first candidate that exists.
+fn pick_iptables(exists: impl Fn(&str) -> bool) -> Option<&'static str> {
+    IPTABLES_DIRECTORIES
+        .into_iter()
+        .find(|candidate| exists(candidate))
+}
+
+/// What running `iptables` produced.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub(super) enum Ran {
+    /// There is no `iptables` to run.
+    Missing,
+    /// It ran, and exited like this.
+    Exited {
+        success: bool,
+        stdout: String,
+        stderr: String,
+    },
+}
+
+/// Runs `iptables` with these arguments. A function value so the flow below
+/// can be tested without touching the host.
+pub(super) type Run<'a> = &'a dyn Fn(&[String]) -> Ran;
+
+/// The match and target of the rule, without the chain or position.
+///
+/// `comment` false leaves the tag off, for a kernel that has no `comment`
+/// match; the rule is then recognised by its shape, see [`tagged_deletions`].
+fn rule_spec(plan: &BroadcastRulesPlan, comment: bool) -> Vec<String> {
+    let range = plan.range.to_string();
+    let tag = plan.tag();
+    let mut words = vec!["-i", plan.interface.as_str(), "-p", "udp", "-s", &range];
+    if comment {
+        words.extend(["-m", "comment", "--comment", &tag]);
+    }
+    words.extend(["-j", "ACCEPT"]);
+    words.into_iter().map(str::to_string).collect()
+}
+
+/// Whether the kernel has the `comment` match, which a rule needs to carry
+/// its tag. Asked of iptables itself: checking a rule that does not exist
+/// fails either way, but only a missing match says so.
+pub(super) fn comment_supported(run: Run<'_>) -> bool {
+    let probe = locked(
+        [
+            "-t",
+            "filter",
+            "-C",
+            "FORWARD",
+            "-m",
+            "comment",
+            "--comment",
+            "tsunagi-probe",
+            "-j",
+            "ACCEPT",
+        ]
+        .map(str::to_string),
+    );
+    match run(&probe) {
+        Ran::Exited {
+            success: false,
+            stderr,
+            ..
+        } => !stderr.contains("No chain/target/match by that name"),
+        _ => true,
+    }
+}
+
+/// `iptables -S` words as option clauses: an option with its value, a negated
+/// one with its `!`. It prints the options in its own order, not the one a
+/// rule was written in, so rules are compared by their clauses, sorted.
+fn clauses(words: &[String]) -> Vec<String> {
+    let mut out = Vec::new();
+    let mut at = 0;
+    while at < words.len() {
+        let take = if words[at] == "!" { 3 } else { 2 };
+        let end = (at + take).min(words.len());
+        out.push(words[at..end].join(" "));
+        at = end;
+    }
+    out.sort();
+    out
+}
+
+/// Whether two rules say the same thing, whatever order they list it in.
+pub(super) fn same_rule(a: &[String], b: &[String]) -> bool {
+    clauses(a) == clauses(b)
+}
+
+/// Waits briefly for the xtables lock instead of failing when another tool
+/// holds it.
+pub(super) fn locked(args: impl IntoIterator<Item = String>) -> Vec<String> {
+    ["-w", "5"]
+        .into_iter()
+        .map(str::to_string)
+        .chain(args)
+        .collect()
+}
+
+fn check_args(plan: &BroadcastRulesPlan, comment: bool) -> Vec<String> {
+    locked(
+        ["-C".to_string(), CHAIN.to_string()]
+            .into_iter()
+            .chain(rule_spec(plan, comment)),
+    )
+}
+
+fn insert_args(plan: &BroadcastRulesPlan, comment: bool) -> Vec<String> {
+    locked(
+        ["-I".to_string(), CHAIN.to_string(), "1".to_string()]
+            .into_iter()
+            .chain(rule_spec(plan, comment)),
+    )
+}
+
+fn list_args() -> Vec<String> {
+    locked(["-S".to_string(), CHAIN.to_string()])
+}
+
+/// Splits one `iptables -S` line into words, honouring double quotes, which
+/// `iptables` puts around comments that need them.
+pub(super) fn split_rule_line(line: &str) -> Vec<String> {
+    let mut words = Vec::new();
+    let mut current = String::new();
+    let mut quoted = false;
+    let mut started = false;
+    for c in line.chars() {
+        match c {
+            '"' => {
+                quoted = !quoted;
+                started = true;
+            }
+            c if c.is_whitespace() && !quoted => {
+                if started {
+                    words.push(std::mem::take(&mut current));
+                    started = false;
+                }
+            }
+            c => {
+                current.push(c);
+                started = true;
+            }
+        }
+    }
+    if started {
+        words.push(current);
+    }
+    words
+}
+
+/// Whether a rule is the one this module writes for some range of the
+/// interface the tag belongs to, with no comment on it.
+fn is_uncommented_rule(tag: &str, rest: &[String]) -> bool {
+    let Some(interface) = tag.strip_prefix("tsunagi:") else {
+        return false;
+    };
+    let Some(range) = rest
+        .windows(2)
+        .find(|pair| pair[0] == "-s")
+        .map(|pair| pair[1].as_str())
+    else {
+        return false;
+    };
+    let ours: Vec<String> = ["-i", interface, "-p", "udp", "-s", range, "-j", "ACCEPT"]
+        .map(str::to_string)
+        .to_vec();
+    same_rule(&ours, rest)
+}
+
+/// The delete commands for every rule in an `iptables -S INPUT` listing that
+/// carries exactly this tag, or, with no `comment` match in the kernel, is
+/// exactly the rule written for some range. Anything else is left alone.
+fn tagged_deletions(listing: &str, tag: &str) -> Vec<Vec<String>> {
+    listing
+        .lines()
+        .filter_map(|line| {
+            let words = split_rule_line(line);
+            let [append, chain, rest @ ..] = words.as_slice() else {
+                return None;
+            };
+            if append != "-A" || chain != CHAIN {
+                return None;
+            }
+            let tagged = rest
+                .windows(2)
+                .any(|pair| pair[0] == "--comment" && pair[1] == tag)
+                || is_uncommented_rule(tag, rest);
+            tagged.then(|| {
+                locked(
+                    ["-D".to_string(), CHAIN.to_string()]
+                        .into_iter()
+                        .chain(rest.iter().cloned()),
+                )
+            })
+        })
+        .collect()
+}
+
+/// A failure message with the likely fix for the likely cause.
+pub(super) fn failure(stderr: &str) -> String {
+    let text = stderr.trim();
+    let lower = text.to_ascii_lowercase();
+    if lower.contains("permission denied") || lower.contains("must be root") {
+        format!("iptables needs root: {text}")
+    } else if text.is_empty() {
+        "iptables failed without saying why".to_string()
+    } else {
+        format!("iptables failed: {text}")
+    }
+}
+
+fn missing(plan: &BroadcastRulesPlan) -> String {
+    format!(
+        "iptables not found; install it (the iptables-nft package is enough) or allow inbound \
+         UDP from {} on `{}` in your firewall",
+        plan.range, plan.interface
+    )
+}
+
+/// Makes the firewall hold exactly one rule for this interface, the planned one.
+fn apply_firewall(run: Run<'_>, plan: &BroadcastRulesPlan) -> RuleOutcome {
+    let comment = comment_supported(run);
+    let present = match run(&check_args(plan, comment)) {
+        Ran::Missing => return RuleOutcome::Failed(missing(plan)),
+        Ran::Exited { success, .. } => success,
+    };
+    let listing = match run(&list_args()) {
+        Ran::Missing => return RuleOutcome::Failed(missing(plan)),
+        Ran::Exited {
+            success: true,
+            stdout,
+            ..
+        } => stdout,
+        Ran::Exited { stderr, .. } => return RuleOutcome::Failed(failure(&stderr)),
+    };
+    let stale = tagged_deletions(&listing, &plan.tag());
+    // `-C` matched an identical rule and nothing else carries our tag, so
+    // there is nothing to do. Otherwise start from a clean slate for this
+    // interface: a rule from an earlier run may name another range.
+    if present && stale.len() == 1 {
+        return RuleOutcome::Applied;
+    }
+    for delete in &stale {
+        if let Ran::Exited {
+            success: false,
+            stderr,
+            ..
+        } = run(delete)
+        {
+            return RuleOutcome::Failed(failure(&stderr));
+        }
+    }
+    match run(&insert_args(plan, comment)) {
+        Ran::Missing => RuleOutcome::Failed(missing(plan)),
+        Ran::Exited { success: true, .. } => RuleOutcome::Applied,
+        Ran::Exited { stderr, .. } => RuleOutcome::Failed(failure(&stderr)),
+    }
+}
+
+/// Removes every rule tagged for this interface. Best effort.
+fn clear_firewall(run: Run<'_>, interface: &str) {
+    let Ran::Exited {
+        success: true,
+        stdout,
+        ..
+    } = run(&list_args())
+    else {
+        return;
+    };
+    for delete in tagged_deletions(&stdout, &tag_for(interface)) {
+        if let Ran::Exited {
+            success: false,
+            stderr,
+            ..
+        } = run(&delete)
+        {
+            tracing::warn!(
+                interface,
+                "cannot remove the broadcast firewall rule: {}",
+                failure(&stderr)
+            );
+        }
+    }
+}
+
+/// Runs the real `iptables`.
+pub(super) fn run_iptables(args: &[String]) -> Ran {
+    let Some(program) = pick_iptables(|path| std::path::Path::new(path).is_file()) else {
+        return Ran::Missing;
+    };
+    match std::process::Command::new(program).args(args).output() {
+        Ok(output) => Ran::Exited {
+            success: output.status.success(),
+            stdout: String::from_utf8_lossy(&output.stdout).into_owned(),
+            stderr: String::from_utf8_lossy(&output.stderr).into_owned(),
+        },
+        Err(err) if err.kind() == std::io::ErrorKind::NotFound => Ran::Missing,
+        Err(err) => Ran::Exited {
+            success: false,
+            stdout: String::new(),
+            stderr: err.to_string(),
+        },
+    }
+}
+
+/// The Linux implementation: netlink for the route, `iptables` for the rule.
+#[derive(Debug)]
+pub struct LinuxHostRules {
+    pub(super) route: RouteHandle,
+}
+
+impl LinuxHostRules {
+    pub(crate) fn new(route: RouteHandle) -> Self {
+        Self { route }
+    }
+
+    async fn set_route(&self, interface: String, source: Ipv4Addr) -> RuleOutcome {
+        let route = self.route.clone();
+        match tokio::task::spawn_blocking(move || route.set(&interface, source)).await {
+            Ok(Ok(())) => RuleOutcome::Applied,
+            Ok(Err(err)) => RuleOutcome::Failed(err.to_string()),
+            Err(err) => RuleOutcome::Failed(format!("the route task failed: {err}")),
+        }
+    }
+}
+
+impl BroadcastHostRules for LinuxHostRules {
+    fn name(&self) -> &str {
+        "netlink+iptables"
+    }
+
+    fn exit_rules(&self) -> Option<&dyn super::ExitHostRules> {
+        Some(self)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            let route = self.set_route(plan.interface.clone(), plan.source).await;
+            let owned = plan.clone();
+            let firewall =
+                tokio::task::spawn_blocking(move || apply_firewall(&run_iptables, &owned))
+                    .await
+                    .unwrap_or_else(|err| {
+                        RuleOutcome::Failed(format!("the firewall task failed: {err}"))
+                    });
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route,
+                firewall,
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let route = self.route.clone();
+            let name = interface.to_string();
+            let route_name = name.clone();
+            if let Ok(Err(err)) =
+                tokio::task::spawn_blocking(move || route.clear(&route_name)).await
+            {
+                tracing::warn!(interface = %name, %err, "cannot remove the broadcast route");
+            }
+            let name = interface.to_string();
+            let _ = tokio::task::spawn_blocking(move || clear_firewall(&run_iptables, &name)).await;
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::cell::RefCell;
+
+    fn plan() -> BroadcastRulesPlan {
+        BroadcastRulesPlan {
+            interface: "tsun0".into(),
+            source: Ipv4Addr::new(10, 13, 37, 142),
+            range: "10.13.37.0/24".parse().unwrap(),
+        }
+    }
+
+    fn words(args: &[String]) -> String {
+        args.join(" ")
+    }
+
+    #[test]
+    fn the_rule_is_udp_only_from_the_overlay_range_on_this_interface() {
+        let spec = words(&rule_spec(&plan(), true));
+        assert_eq!(
+            spec,
+            "-i tsun0 -p udp -s 10.13.37.0/24 -m comment --comment tsunagi:tsun0 -j ACCEPT"
+        );
+        // Nothing here may widen to other protocols or every source.
+        assert!(spec.contains("-p udp"));
+        assert!(!spec.contains("0.0.0.0/0"));
+        assert!(words(&insert_args(&plan(), true)).starts_with("-w 5 -I INPUT 1 -i tsun0"));
+        assert!(words(&check_args(&plan(), true)).starts_with("-w 5 -C INPUT -i tsun0"));
+    }
+
+    #[test]
+    fn iptables_is_found_in_fixed_directories_never_on_path() {
+        assert_eq!(pick_iptables(|_| false), None);
+        assert_eq!(
+            pick_iptables(|p| p == "/sbin/iptables"),
+            Some("/sbin/iptables")
+        );
+        assert_eq!(pick_iptables(|_| true), Some("/usr/sbin/iptables"));
+    }
+
+    #[test]
+    fn only_rules_with_our_exact_tag_are_selected_for_deletion() {
+        let listing = "\
+-P INPUT DROP
+-A INPUT -s 10.13.37.0/24 -i tsun0 -p udp -m comment --comment \"tsunagi:tsun0\" -j ACCEPT
+-A INPUT -i tsun0 -m comment --comment tsunagi:tsun0 -j ACCEPT
+-A INPUT -i tsun1 -p udp -m comment --comment tsunagi:tsun1 -j ACCEPT
+-A INPUT -p tcp --dport 22 -m comment --comment \"allow ssh tsunagi:tsun0\" -j ACCEPT
+-A FORWARD -i tsun0 -m comment --comment tsunagi:tsun0 -j ACCEPT
+-A INPUT -p tcp --dport 22 -j ACCEPT
+";
+        let deletions = tagged_deletions(listing, "tsunagi:tsun0");
+        assert_eq!(deletions.len(), 2, "{deletions:?}");
+        assert_eq!(
+            words(&deletions[0]),
+            "-w 5 -D INPUT -s 10.13.37.0/24 -i tsun0 -p udp -m comment --comment tsunagi:tsun0 -j ACCEPT"
+        );
+        assert!(words(&deletions[1]).contains("-D INPUT -i tsun0 -m comment"));
+        assert!(tagged_deletions("", "tsunagi:tsun0").is_empty());
+    }
+
+    #[test]
+    fn a_kernel_without_the_comment_match_says_so_in_its_error() {
+        let answer = |stderr: &'static str| {
+            move |_: &[String]| Ran::Exited {
+                success: false,
+                stdout: String::new(),
+                stderr: stderr.to_string(),
+            }
+        };
+        assert!(!comment_supported(&answer(
+            "iptables: No chain/target/match by that name."
+        )));
+        assert!(comment_supported(&answer(
+            "iptables: Bad rule (does a matching rule exist in that chain?)."
+        )));
+        assert!(comment_supported(&|_: &[String]| Ran::Missing));
+    }
+
+    #[test]
+    fn a_rule_written_without_a_comment_is_found_by_its_shape() {
+        // Listed the way iptables prints it: its own option order.
+        let listing = "\
+-P INPUT ACCEPT
+-A INPUT -s 10.13.37.0/24 -i tsun0 -p udp -j ACCEPT
+-A INPUT -s 10.99.0.0/16 -i tsun0 -p udp -j ACCEPT
+-A INPUT -s 10.13.37.0/24 -i tsun1 -p udp -j ACCEPT
+-A INPUT -s 10.13.37.0/24 -i tsun0 -p udp -j DROP
+-A INPUT -s 10.13.37.0/24 -i tsun0 -p tcp -j ACCEPT
+";
+        let deletions = tagged_deletions(listing, "tsunagi:tsun0");
+        assert_eq!(deletions.len(), 2, "{deletions:?}");
+        assert!(words(&deletions[0]).ends_with("-s 10.13.37.0/24 -i tsun0 -p udp -j ACCEPT"));
+        assert!(words(&deletions[1]).contains("10.99.0.0/16"));
+    }
+
+    #[test]
+    fn without_a_comment_the_rule_has_none() {
+        let with = words(&rule_spec(&plan(), true));
+        let without = words(&rule_spec(&plan(), false));
+        assert!(with.contains("-m comment --comment tsunagi:tsun0"));
+        assert_eq!(without, "-i tsun0 -p udp -s 10.13.37.0/24 -j ACCEPT");
+    }
+
+    /// A scripted `iptables`: answers by the leading verb, records every call.
+    struct Fake {
+        calls: RefCell<Vec<String>>,
+        check_ok: bool,
+        listing: String,
+        insert_stderr: Option<&'static str>,
+        missing: bool,
+    }
+
+    impl Fake {
+        fn new(check_ok: bool, listing: &str) -> Self {
+            Self {
+                calls: RefCell::new(Vec::new()),
+                check_ok,
+                listing: listing.to_string(),
+                insert_stderr: None,
+                missing: false,
+            }
+        }
+        fn run(&self, args: &[String]) -> Ran {
+            // The probe for the `comment` match: not part of what is tested.
+            if args.iter().any(|word| word == "tsunagi-probe") {
+                return Ran::Exited {
+                    success: false,
+                    stdout: String::new(),
+                    stderr: "Bad rule".into(),
+                };
+            }
+            self.calls.borrow_mut().push(words(args));
+            if self.missing {
+                return Ran::Missing;
+            }
+            let ok = |stdout: &str| Ran::Exited {
+                success: true,
+                stdout: stdout.to_string(),
+                stderr: String::new(),
+            };
+            match args.get(2).map(String::as_str) {
+                Some("-C") if self.check_ok => ok(""),
+                Some("-C") => Ran::Exited {
+                    success: false,
+                    stdout: String::new(),
+                    stderr: "Bad rule".into(),
+                },
+                Some("-S") => ok(&self.listing),
+                Some("-I") => match self.insert_stderr {
+                    Some(stderr) => Ran::Exited {
+                        success: false,
+                        stdout: String::new(),
+                        stderr: stderr.into(),
+                    },
+                    None => ok(""),
+                },
+                _ => ok(""),
+            }
+        }
+        fn verbs(&self) -> Vec<String> {
+            self.calls
+                .borrow()
+                .iter()
+                .map(|call| call.split(' ').nth(2).unwrap().to_string())
+                .collect()
+        }
+    }
+
+    const ONE_OURS: &str =
+        "-A INPUT -s 10.13.37.0/24 -i tsun0 -p udp -m comment --comment tsunagi:tsun0 -j ACCEPT\n";
+
+    #[test]
+    fn an_identical_rule_already_in_place_is_left_alone() {
+        let fake = Fake::new(true, ONE_OURS);
+        assert_eq!(
+            apply_firewall(&|a| fake.run(a), &plan()),
+            RuleOutcome::Applied
+        );
+        assert_eq!(fake.verbs(), ["-C", "-S"], "no insert, no delete");
+    }
+
+    #[test]
+    fn a_missing_rule_is_inserted_once() {
+        let fake = Fake::new(false, "-P INPUT ACCEPT\n");
+        assert_eq!(
+            apply_firewall(&|a| fake.run(a), &plan()),
+            RuleOutcome::Applied
+        );
+        assert_eq!(fake.verbs(), ["-C", "-S", "-I"]);
+    }
+
+    #[test]
+    fn a_stale_rule_for_another_range_is_replaced_not_stacked() {
+        // A crashed run left a rule for an old range under our tag.
+        let stale = "-A INPUT -s 10.99.0.0/16 -i tsun0 -p udp -m comment --comment tsunagi:tsun0 -j ACCEPT\n";
+        let fake = Fake::new(false, stale);
+        assert_eq!(
+            apply_firewall(&|a| fake.run(a), &plan()),
+            RuleOutcome::Applied
+        );
+        assert_eq!(fake.verbs(), ["-C", "-S", "-D", "-I"]);
+    }
+
+    #[test]
+    fn duplicates_of_the_rule_collapse_to_one() {
+        let doubled = format!("{ONE_OURS}{ONE_OURS}");
+        let fake = Fake::new(true, &doubled);
+        assert_eq!(
+            apply_firewall(&|a| fake.run(a), &plan()),
+            RuleOutcome::Applied
+        );
+        assert_eq!(fake.verbs(), ["-C", "-S", "-D", "-D", "-I"]);
+    }
+
+    #[test]
+    fn a_missing_iptables_says_what_to_install_and_what_to_allow() {
+        let mut fake = Fake::new(false, "");
+        fake.missing = true;
+        let RuleOutcome::Failed(reason) = apply_firewall(&|a| fake.run(a), &plan()) else {
+            panic!("expected a failure");
+        };
+        assert!(reason.contains("iptables not found"), "{reason}");
+        assert!(reason.contains("10.13.37.0/24"), "{reason}");
+        assert!(reason.contains("tsun0"), "{reason}");
+    }
+
+    #[test]
+    fn a_refused_insert_is_reported_with_the_cause() {
+        let mut fake = Fake::new(false, "");
+        fake.insert_stderr = Some(
+            "iptables v1.8.10: can't initialize iptables table `filter': Permission denied (you must be root)",
+        );
+        let RuleOutcome::Failed(reason) = apply_firewall(&|a| fake.run(a), &plan()) else {
+            panic!("expected a failure");
+        };
+        assert!(reason.starts_with("iptables needs root"), "{reason}");
+    }
+
+    #[test]
+    fn clearing_removes_only_tagged_rules_and_ignores_a_missing_iptables() {
+        let listing = format!(
+            "-A INPUT -p tcp --dport 22 -j ACCEPT\n{ONE_OURS}-A INPUT -i tsun1 -m comment --comment tsunagi:tsun1 -j ACCEPT\n"
+        );
+        let fake = Fake::new(true, &listing);
+        clear_firewall(&|a| fake.run(a), "tsun0");
+        assert_eq!(fake.verbs(), ["-S", "-D"]);
+
+        let mut gone = Fake::new(true, "");
+        gone.missing = true;
+        clear_firewall(&|a| gone.run(a), "tsun0");
+        assert_eq!(gone.verbs(), ["-S"]);
+    }
+
+    #[test]
+    fn rule_listing_words_honour_quotes() {
+        assert_eq!(
+            split_rule_line("-A INPUT -m comment --comment \"a b\" -j ACCEPT"),
+            [
+                "-A",
+                "INPUT",
+                "-m",
+                "comment",
+                "--comment",
+                "a b",
+                "-j",
+                "ACCEPT"
+            ]
+        );
+        assert!(split_rule_line("   ").is_empty());
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/linux_exit.rs

@@ -0,0 +1,813 @@
+//! Exit-node host rules on Linux: `iptables` for the offering side, policy
+//! routing over netlink for the using side.
+//!
+//! # Offering
+//!
+//! For every network that offers this agent as an exit node, three rules,
+//! selected by that network's range and carrying the tag
+//! `tsunagi-exit:<if>:<range>`:
+//!
+//! ```text
+//! -t nat    POSTROUTING  -s <range> ! -o <if> -j MASQUERADE
+//! -t filter FORWARD      -i <if> -s <range> ! -o <if> -j ACCEPT
+//! -t filter FORWARD      -o <if> -d <range> -m conntrack --ctstate RELATED,ESTABLISHED -j ACCEPT
+//! ```
+//!
+//! `! -o <if>` is the egress without naming it: whatever interface the
+//! default route uses, now or after it changes. Kernel forwarding is only
+//! *read*. Turning it on changes how the whole host behaves, which is the
+//! owner's decision, so when it is off the rules are installed anyway and the
+//! report says so.
+//!
+//! # Using
+//!
+//! Policy routing, with no `/1` routes and nothing in the main table:
+//!
+//! ```text
+//! table 0x7473:  default dev <if>
+//! 5280: from all uidrange <euid>-<euid> lookup main
+//! 5290: from all lookup main suppress_prefixlength 0
+//! 5300: from all lookup 0x7473
+//! ```
+//!
+//! The first keeps the agent's *own* sockets off the tunnel: its QUIC, relay,
+//! DHT and DNS traffic would otherwise be routed into the interface it is
+//! carrying, a loop. The second lets every specific route — the LAN, another
+//! VPN, a tunnel to somewhere else — keep working, because only the default
+//! route is suppressed. The priorities sit after Tailscale's 5270.
+//!
+//! The first rule exempts a *user id*, so it is only right when the agent has
+//! one of its own, as the packaged service does. An agent started from a login
+//! session would take that user's own applications out of the tunnel with it,
+//! and the exit node would silently do nothing for them. That is refused, not
+//! guessed at: see [`client_refusal`].
+
+use crate::BoxFuture;
+use crate::state::Ipv4Range;
+
+use super::RuleOutcome;
+use super::exit::{ExitHostPlan, ExitHostReport, ExitHostRules, exit_tag, exit_tag_prefix};
+use super::linux::{
+    LinuxHostRules, Ran, Run, comment_supported, failure, locked, run_iptables, same_rule,
+    split_rule_line,
+};
+
+/// The chains the offering rules live in: `(table, chain)`.
+const CHAINS: [(&str, &str); 2] = [("nat", "POSTROUTING"), ("filter", "FORWARD")];
+
+/// One firewall rule: where it goes, and its match and target.
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct Rule {
+    table: &'static str,
+    chain: &'static str,
+    spec: Vec<String>,
+}
+
+fn owned(words: &[&str]) -> Vec<String> {
+    words.iter().map(|word| word.to_string()).collect()
+}
+
+/// The three rules of one offered range.
+///
+/// Each carries a comment naming its interface and range, which is how it is
+/// found again. A kernel without the `comment` match (small router builds
+/// leave it out) cannot take that, so `comment` false leaves it off and the
+/// rules are recognised by their shape instead; see [`tagged_rules`].
+fn offer_rules(interface: &str, range: &Ipv4Range, comment: bool) -> Vec<Rule> {
+    let range_text = range.to_string();
+    let tag = exit_tag(interface, range);
+    let comment: Vec<String> = if comment {
+        owned(&["-m", "comment", "--comment", tag.as_str()])
+    } else {
+        Vec::new()
+    };
+    let spec = |head: &[&str], tail: &[&str]| -> Vec<String> {
+        owned(head)
+            .into_iter()
+            .chain(comment.iter().cloned())
+            .chain(owned(tail))
+            .collect()
+    };
+    vec![
+        Rule {
+            table: "nat",
+            chain: "POSTROUTING",
+            spec: spec(
+                &["-s", &range_text, "!", "-o", interface],
+                &["-j", "MASQUERADE"],
+            ),
+        },
+        Rule {
+            table: "filter",
+            chain: "FORWARD",
+            spec: spec(
+                &["-i", interface, "-s", &range_text, "!", "-o", interface],
+                &["-j", "ACCEPT"],
+            ),
+        },
+        Rule {
+            table: "filter",
+            chain: "FORWARD",
+            spec: spec(
+                &[
+                    "-o",
+                    interface,
+                    "-d",
+                    &range_text,
+                    "-m",
+                    "conntrack",
+                    "--ctstate",
+                    "RELATED,ESTABLISHED",
+                ],
+                &["-j", "ACCEPT"],
+            ),
+        },
+    ]
+}
+
+fn check_args(rule: &Rule) -> Vec<String> {
+    locked(
+        owned(&["-t", rule.table, "-C", rule.chain])
+            .into_iter()
+            .chain(rule.spec.iter().cloned()),
+    )
+}
+
+fn insert_args(rule: &Rule) -> Vec<String> {
+    locked(
+        owned(&["-t", rule.table, "-I", rule.chain, "1"])
+            .into_iter()
+            .chain(rule.spec.iter().cloned()),
+    )
+}
+
+fn list_args(table: &str, chain: &str) -> Vec<String> {
+    locked(owned(&["-t", table, "-S", chain]))
+}
+
+/// A rule found in a listing whose comment starts with our prefix.
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct Found {
+    /// The whole tag.
+    tag: String,
+    /// The words after `-A <chain>`, which is what `-D` needs.
+    rest: Vec<String>,
+}
+
+/// Every rule in an `iptables -S <chain>` listing that is ours: carrying a
+/// tag that starts with `prefix`, or, on a kernel with no `comment` match,
+/// being exactly one of the three rules this module writes for some range.
+/// Anything else in the chain is left alone.
+fn tagged_rules(listing: &str, chain: &str, prefix: &str) -> Vec<Found> {
+    // `tsunagi-exit:<interface>:`
+    let interface = prefix
+        .strip_prefix("tsunagi-exit:")
+        .and_then(|rest| rest.strip_suffix(':'))
+        .unwrap_or_default();
+    listing
+        .lines()
+        .filter_map(|line| {
+            let words = split_rule_line(line);
+            let [append, found_chain, rest @ ..] = words.as_slice() else {
+                return None;
+            };
+            if append != "-A" || found_chain != chain {
+                return None;
+            }
+            let tag = rest
+                .windows(2)
+                .find(|pair| pair[0] == "--comment" && pair[1].starts_with(prefix))
+                .map(|pair| pair[1].clone())
+                .or_else(|| uncommented_tag(interface, chain, rest))?;
+            Some(Found {
+                tag,
+                rest: rest.to_vec(),
+            })
+        })
+        .collect()
+}
+
+/// The tag a rule would have carried, when it is one of ours written without
+/// a comment: its words are those of the rule for the range it names.
+fn uncommented_tag(interface: &str, chain: &str, rest: &[String]) -> Option<String> {
+    if interface.is_empty() {
+        return None;
+    }
+    let range: Ipv4Range = rest
+        .windows(2)
+        .find(|pair| pair[0] == "-s" || pair[0] == "-d")
+        .and_then(|pair| pair[1].parse().ok())?;
+    offer_rules(interface, &range, false)
+        .iter()
+        .any(|rule| rule.chain == chain && same_rule(&rule.spec, rest))
+        .then(|| exit_tag(interface, &range))
+}
+
+fn delete_args(table: &str, chain: &str, found: &Found) -> Vec<String> {
+    locked(
+        owned(&["-t", table, "-D", chain])
+            .into_iter()
+            .chain(found.rest.iter().cloned()),
+    )
+}
+
+fn missing_iptables(range: &Ipv4Range, interface: &str) -> String {
+    format!(
+        "iptables not found; install it (the iptables-nft package is enough), or masquerade \
+         {range} out of every interface but `{interface}` and forward it yourself"
+    )
+}
+
+/// Reads one chain, or says why not.
+fn read_chain(run: Run<'_>, table: &str, chain: &str) -> Result<String, Failure> {
+    match run(&list_args(table, chain)) {
+        Ran::Missing => Err(Failure::Missing),
+        Ran::Exited {
+            success: true,
+            stdout,
+            ..
+        } => Ok(stdout),
+        Ran::Exited { stderr, .. } => Err(Failure::Failed(failure(&stderr))),
+    }
+}
+
+enum Failure {
+    Missing,
+    Failed(String),
+}
+
+/// Deletes found rules; the first failure ends it.
+fn delete_all(
+    run: Run<'_>,
+    table: &str,
+    chain: &str,
+    found: &[Found],
+    keep: impl Fn(&Found) -> bool,
+) -> Result<(), String> {
+    for rule in found.iter().filter(|rule| !keep(rule)) {
+        if let Ran::Exited {
+            success: false,
+            stderr,
+            ..
+        } = run(&delete_args(table, chain, rule))
+        {
+            return Err(failure(&stderr));
+        }
+    }
+    Ok(())
+}
+
+/// Makes the firewall hold exactly the offering rules of these ranges for the
+/// interface, and no others of ours.
+fn apply_offer(
+    run: Run<'_>,
+    interface: &str,
+    ranges: &[Ipv4Range],
+) -> Vec<(Ipv4Range, RuleOutcome)> {
+    let fail_all = |reason: String| -> Vec<(Ipv4Range, RuleOutcome)> {
+        ranges
+            .iter()
+            .map(|range| (*range, RuleOutcome::Failed(reason.clone())))
+            .collect()
+    };
+    let prefix = exit_tag_prefix(interface);
+    let wanted: Vec<String> = ranges.iter().map(|r| exit_tag(interface, r)).collect();
+    let comment = comment_supported(run);
+
+    // What is there now, per chain.
+    let mut present: Vec<(&str, &str, Vec<Found>)> = Vec::new();
+    for (table, chain) in CHAINS {
+        match read_chain(run, table, chain) {
+            Ok(listing) => present.push((table, chain, tagged_rules(&listing, chain, &prefix))),
+            Err(Failure::Missing) => {
+                return match ranges.first() {
+                    Some(range) => fail_all(missing_iptables(range, interface)),
+                    None => Vec::new(),
+                };
+            }
+            Err(Failure::Failed(reason)) => return fail_all(reason),
+        }
+    }
+
+    // Rules for a range no longer offered, or left by an earlier run.
+    for (table, chain, found) in &present {
+        if let Err(reason) = delete_all(run, table, chain, found, |rule| wanted.contains(&rule.tag))
+        {
+            return fail_all(reason);
+        }
+    }
+
+    ranges
+        .iter()
+        .map(|range| {
+            let rules = offer_rules(interface, range, comment);
+            let tag = exit_tag(interface, range);
+            let in_place = rules
+                .iter()
+                .all(|rule| matches!(run(&check_args(rule)), Ran::Exited { success: true, .. }))
+                && present.iter().all(|(_, chain, found)| {
+                    let have = found.iter().filter(|rule| rule.tag == tag).count();
+                    have == rules.iter().filter(|rule| rule.chain == *chain).count()
+                });
+            if in_place {
+                return (*range, RuleOutcome::Applied);
+            }
+            // Start this range from a clean slate: a half-installed set, or
+            // one that differs from the plan, is replaced as a whole.
+            for (table, chain, found) in &present {
+                if let Err(reason) = delete_all(run, table, chain, found, |rule| rule.tag != tag) {
+                    return (*range, RuleOutcome::Failed(reason));
+                }
+            }
+            for rule in &rules {
+                match run(&insert_args(rule)) {
+                    Ran::Missing => {
+                        return (
+                            *range,
+                            RuleOutcome::Failed(missing_iptables(range, interface)),
+                        );
+                    }
+                    Ran::Exited { success: true, .. } => {}
+                    Ran::Exited { stderr, .. } => {
+                        return (*range, RuleOutcome::Failed(failure(&stderr)));
+                    }
+                }
+            }
+            (*range, RuleOutcome::Applied)
+        })
+        .collect()
+}
+
+/// Removes every offering rule of ours for the interface. Best effort.
+fn clear_offer(run: Run<'_>, interface: &str) {
+    let prefix = exit_tag_prefix(interface);
+    for (table, chain) in CHAINS {
+        let Ok(listing) = read_chain(run, table, chain) else {
+            continue;
+        };
+        let found = tagged_rules(&listing, chain, &prefix);
+        if let Err(reason) = delete_all(run, table, chain, &found, |_| false) {
+            tracing::warn!(
+                interface,
+                "cannot remove an exit node firewall rule: {reason}"
+            );
+        }
+    }
+}
+
+/// The kernel's forwarding switch for packets arriving on the interface,
+/// read from the text of the sysctl files. Never written.
+fn forwarding_state(read: impl Fn(&str) -> Option<String>, interface: &str) -> Option<bool> {
+    let flag = |text: String| match text.trim() {
+        "1" => Some(true),
+        "0" => Some(false),
+        _ => None,
+    };
+    // The name is derived here, but it ends up in a path.
+    if !interface.is_empty() && !interface.contains(['/', '\0']) && interface != ".." {
+        let per_interface = format!("/proc/sys/net/ipv4/conf/{interface}/forwarding");
+        if let Some(state) = read(&per_interface).and_then(flag) {
+            return Some(state);
+        }
+    }
+    read("/proc/sys/net/ipv4/ip_forward").and_then(flag)
+}
+
+fn read_sysctl(path: &str) -> Option<String> {
+    std::fs::read_to_string(path).ok()
+}
+
+/// The effective user id from the text of `/proc/self/status`.
+fn parse_euid(status: &str) -> Option<u32> {
+    status
+        .lines()
+        .find_map(|line| line.strip_prefix("Uid:"))
+        .and_then(|rest| rest.split_whitespace().nth(1))
+        .and_then(|effective| effective.parse().ok())
+}
+
+/// Why the using side must not be installed for this process, if it must not.
+///
+/// The bypass rule exempts this process's *user id* from the tunnel. That is
+/// right for a service with an id of its own and wrong for a process started
+/// from somebody's login session, which shares its id with every application
+/// that person runs. A login session is recognised by `XDG_RUNTIME_DIR`,
+/// which `systemd --user` and the login managers set and a system service
+/// does not. Crude on purpose: refusing a service that was started oddly is
+/// a message, and exempting a user's applications is a silent failure.
+fn client_refusal(xdg_runtime_dir: Option<&str>) -> Option<String> {
+    xdg_runtime_dir.filter(|dir| !dir.is_empty()).map(|_| {
+        "the exit node needs the agent to run as a system service with a user of its own: \
+         this agent was started from a login session, so routing everything through the \
+         overlay would take that user's own applications out of the tunnel with it"
+            .to_string()
+    })
+}
+
+impl LinuxHostRules {
+    async fn blocking<T: Send + 'static>(
+        work: impl FnOnce() -> T + Send + 'static,
+    ) -> Result<T, String> {
+        tokio::task::spawn_blocking(work)
+            .await
+            .map_err(|err| format!("the task failed: {err}"))
+    }
+
+    async fn apply_client(&self, interface: &str) -> RuleOutcome {
+        let xdg = std::env::var("XDG_RUNTIME_DIR").ok();
+        if let Some(reason) = client_refusal(xdg.as_deref()) {
+            // Whatever an earlier run left must not stay in force.
+            self.clear_client().await;
+            return RuleOutcome::Failed(reason);
+        }
+        let Some(euid) = std::fs::read_to_string("/proc/self/status")
+            .ok()
+            .and_then(|status| parse_euid(&status))
+        else {
+            return RuleOutcome::Failed("cannot tell which user this agent runs as".to_string());
+        };
+        let route = self.route.clone();
+        let interface = interface.to_string();
+        match Self::blocking(move || route.set_exit_client(&interface, euid)).await {
+            Ok(Ok(())) => RuleOutcome::Applied,
+            Ok(Err(err)) => RuleOutcome::Failed(err.to_string()),
+            Err(reason) => RuleOutcome::Failed(reason),
+        }
+    }
+
+    async fn clear_client(&self) {
+        let route = self.route.clone();
+        if let Ok(Err(err)) = Self::blocking(move || route.clear_exit_client()).await {
+            tracing::warn!(%err, "cannot remove the exit node routing rules");
+        }
+    }
+}
+
+impl ExitHostRules for LinuxHostRules {
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport> {
+        Box::pin(async move {
+            let interface = plan.interface.clone();
+            let ranges = plan.offer.clone();
+            let offer = {
+                let (job_interface, job_ranges) = (interface.clone(), ranges.clone());
+                Self::blocking(move || apply_offer(&run_iptables, &job_interface, &job_ranges))
+                    .await
+                    .unwrap_or_else(|reason| {
+                        ranges
+                            .iter()
+                            .map(|range| (*range, RuleOutcome::Failed(reason.clone())))
+                            .collect()
+                    })
+            };
+            let forwarding = if ranges.is_empty() {
+                None
+            } else {
+                forwarding_state(read_sysctl, &interface)
+            };
+            let client = if plan.client {
+                Some(self.apply_client(&interface).await)
+            } else {
+                self.clear_client().await;
+                None
+            };
+            ExitHostReport {
+                offer,
+                forwarding,
+                client,
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let name = interface.to_string();
+            let _ = Self::blocking(move || clear_offer(&run_iptables, &name)).await;
+            self.clear_client().await;
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::cell::RefCell;
+
+    use super::*;
+
+    fn range() -> Ipv4Range {
+        "10.13.37.0/24".parse().unwrap()
+    }
+
+    fn words(args: &[String]) -> String {
+        args.join(" ")
+    }
+
+    #[test]
+    fn the_rules_masquerade_the_range_out_of_every_other_interface_and_forward_it() {
+        let rules = offer_rules("tsun0", &range(), true);
+        assert_eq!(rules.len(), 3);
+        assert_eq!(
+            words(&rules[0].spec),
+            "-s 10.13.37.0/24 ! -o tsun0 -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j MASQUERADE"
+        );
+        assert_eq!((rules[0].table, rules[0].chain), ("nat", "POSTROUTING"));
+        assert_eq!(
+            words(&rules[1].spec),
+            "-i tsun0 -s 10.13.37.0/24 ! -o tsun0 -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT"
+        );
+        assert_eq!((rules[1].table, rules[1].chain), ("filter", "FORWARD"));
+        assert_eq!(
+            words(&rules[2].spec),
+            "-o tsun0 -d 10.13.37.0/24 -m conntrack --ctstate RELATED,ESTABLISHED -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT"
+        );
+        // Nothing may widen to every source or to the other direction.
+        for rule in &rules {
+            assert!(!words(&rule.spec).contains("0.0.0.0/0"));
+        }
+        assert!(words(&insert_args(&rules[1])).starts_with("-w 5 -t filter -I FORWARD 1 -i tsun0"));
+        assert!(words(&check_args(&rules[0])).starts_with("-w 5 -t nat -C POSTROUTING -s"));
+    }
+
+    #[test]
+    fn only_rules_with_our_tag_for_this_interface_are_selected() {
+        let listing = "\
+-P FORWARD DROP
+-A FORWARD -i tsun0 -s 10.13.37.0/24 ! -o tsun0 -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT
+-A FORWARD -i tsun1 -m comment --comment tsunagi-exit:tsun1:10.99.0.0/24 -j ACCEPT
+-A FORWARD -i tsun01 -m comment --comment tsunagi-exit:tsun01:10.5.0.0/24 -j ACCEPT
+-A FORWARD -m comment --comment \"allow tsunagi-exit:tsun0:x\" -j ACCEPT
+-A FORWARD -m comment --comment tsunagi:tsun0 -j ACCEPT
+-A INPUT -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT
+-A FORWARD -j ACCEPT
+";
+        let found = tagged_rules(listing, "FORWARD", &exit_tag_prefix("tsun0"));
+        assert_eq!(found.len(), 1, "{found:?}");
+        assert_eq!(found[0].tag, "tsunagi-exit:tsun0:10.13.37.0/24");
+        assert_eq!(
+            words(&delete_args("filter", "FORWARD", &found[0])),
+            "-w 5 -t filter -D FORWARD -i tsun0 -s 10.13.37.0/24 ! -o tsun0 -m comment --comment tsunagi-exit:tsun0:10.13.37.0/24 -j ACCEPT"
+        );
+        assert!(tagged_rules("", "FORWARD", "tsunagi-exit:tsun0:").is_empty());
+    }
+
+    /// A scripted `iptables` that remembers which rules it was given.
+    struct Fake {
+        calls: RefCell<Vec<String>>,
+        /// Rules present, as `-S` lines per `(table, chain)`.
+        listing: RefCell<Vec<(String, String)>>,
+        missing: bool,
+        check_ok: bool,
+        /// A kernel with no `comment` match, as small router builds have.
+        no_comment: bool,
+    }
+
+    impl Fake {
+        fn new(check_ok: bool) -> Self {
+            Self {
+                calls: RefCell::new(Vec::new()),
+                listing: RefCell::new(Vec::new()),
+                missing: false,
+                check_ok,
+                no_comment: false,
+            }
+        }
+
+        fn with_rule(self, table: &str, line: &str) -> Self {
+            self.listing
+                .borrow_mut()
+                .push((table.to_string(), line.to_string()));
+            self
+        }
+
+        fn run(&self, args: &[String]) -> Ran {
+            if self.missing {
+                return Ran::Missing;
+            }
+            let line = args.join(" ");
+            self.calls.borrow_mut().push(line.clone());
+            let table = args
+                .iter()
+                .position(|word| word == "-t")
+                .map(|at| args[at + 1].clone())
+                .unwrap_or_default();
+            let ok = |stdout: String| Ran::Exited {
+                success: true,
+                stdout,
+                stderr: String::new(),
+            };
+            if self.no_comment && line.contains("-m comment") {
+                return Ran::Exited {
+                    success: false,
+                    stdout: String::new(),
+                    stderr: "iptables: No chain/target/match by that name.".to_string(),
+                };
+            }
+            if line.contains(" -S ") {
+                let listing: Vec<String> = self
+                    .listing
+                    .borrow()
+                    .iter()
+                    .filter(|(t, _)| *t == table)
+                    .map(|(_, l)| l.clone())
+                    .collect();
+                return ok(listing.join("\n"));
+            }
+            if line.contains(" -C ") {
+                return Ran::Exited {
+                    success: self.check_ok,
+                    stdout: String::new(),
+                    stderr: String::new(),
+                };
+            }
+            ok(String::new())
+        }
+
+        fn verbs(&self, verb: &str) -> Vec<String> {
+            self.calls
+                .borrow()
+                .iter()
+                .filter(|call| call.contains(&format!(" {verb} ")))
+                .cloned()
+                .collect()
+        }
+    }
+
+    #[test]
+    fn a_kernel_without_the_comment_match_gets_the_rules_without_comments() {
+        let mut fake = Fake::new(false);
+        fake.no_comment = true;
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        assert_eq!(outcome, vec![(range(), RuleOutcome::Applied)]);
+        let inserted = fake.verbs("-I");
+        assert_eq!(inserted.len(), 3, "{inserted:?}");
+        assert!(
+            inserted.iter().all(|call| !call.contains("comment")),
+            "{inserted:?}"
+        );
+    }
+
+    #[test]
+    fn rules_written_without_a_comment_are_ours_and_are_removed_with_them() {
+        // The way iptables prints them: its own option order.
+        let listing = "\
+-P FORWARD ACCEPT
+-A FORWARD -s 10.13.37.0/24 -i tsun0 ! -o tsun0 -j ACCEPT
+-A FORWARD -d 10.13.37.0/24 -o tsun0 -m conntrack --ctstate RELATED,ESTABLISHED -j ACCEPT
+-A FORWARD -i tsun0 -j ACCEPT
+-A FORWARD -s 10.13.37.0/24 -i tsun1 ! -o tsun1 -j ACCEPT
+";
+        let found = tagged_rules(listing, "FORWARD", &exit_tag_prefix("tsun0"));
+        assert_eq!(found.len(), 2, "{found:?}");
+        assert!(
+            found
+                .iter()
+                .all(|rule| rule.tag == exit_tag("tsun0", &range()))
+        );
+        let nat = "-A POSTROUTING -s 10.13.37.0/24 ! -o tsun0 -j MASQUERADE\n";
+        assert_eq!(
+            tagged_rules(nat, "POSTROUTING", &exit_tag_prefix("tsun0")).len(),
+            1
+        );
+    }
+
+    #[test]
+    fn a_fresh_host_gets_the_three_rules() {
+        let fake = Fake::new(false);
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        assert_eq!(outcome, vec![(range(), RuleOutcome::Applied)]);
+        let inserted = fake.verbs("-I");
+        assert_eq!(inserted.len(), 3, "{inserted:?}");
+        assert!(inserted[0].contains("-t nat -I POSTROUTING 1"));
+        assert!(fake.verbs("-D").is_empty());
+    }
+
+    #[test]
+    fn rules_already_in_place_are_left_alone() {
+        let tag = exit_tag("tsun0", &range());
+        let fake = Fake::new(true)
+            .with_rule(
+                "nat",
+                &format!("-A POSTROUTING -s 10.13.37.0/24 ! -o tsun0 -m comment --comment {tag} -j MASQUERADE"),
+            )
+            .with_rule("filter", &format!("-A FORWARD -i tsun0 -m comment --comment {tag} -j ACCEPT"))
+            .with_rule("filter", &format!("-A FORWARD -o tsun0 -m comment --comment {tag} -j ACCEPT"));
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        assert_eq!(outcome, vec![(range(), RuleOutcome::Applied)]);
+        assert!(fake.verbs("-I").is_empty());
+        assert!(fake.verbs("-D").is_empty());
+    }
+
+    #[test]
+    fn a_stale_range_is_removed_and_a_half_installed_set_is_replaced() {
+        let old: Ipv4Range = "10.99.0.0/24".parse().unwrap();
+        let old_tag = exit_tag("tsun0", &old);
+        let tag = exit_tag("tsun0", &range());
+        let fake = Fake::new(true)
+            .with_rule(
+                "nat",
+                &format!(
+                    "-A POSTROUTING -s 10.99.0.0/24 -m comment --comment {old_tag} -j MASQUERADE"
+                ),
+            )
+            // Only one of the three for the current range.
+            .with_rule(
+                "filter",
+                &format!("-A FORWARD -i tsun0 -m comment --comment {tag} -j ACCEPT"),
+            )
+            // Somebody else's rule, which must survive.
+            .with_rule("filter", "-A FORWARD -i eth0 -j ACCEPT");
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        assert_eq!(outcome, vec![(range(), RuleOutcome::Applied)]);
+        let deleted = fake.verbs("-D");
+        assert!(
+            deleted.iter().any(|call| call.contains(&old_tag)),
+            "{deleted:?}"
+        );
+        assert!(
+            deleted.iter().any(|call| call.contains(&tag)),
+            "{deleted:?}"
+        );
+        assert!(
+            deleted.iter().all(|call| !call.contains("eth0")),
+            "{deleted:?}"
+        );
+        assert_eq!(fake.verbs("-I").len(), 3);
+    }
+
+    #[test]
+    fn clearing_removes_only_our_rules_for_the_interface() {
+        let tag = exit_tag("tsun0", &range());
+        let other = exit_tag("tsun1", &range());
+        let fake = Fake::new(true)
+            .with_rule(
+                "nat",
+                &format!("-A POSTROUTING -m comment --comment {tag} -j MASQUERADE"),
+            )
+            .with_rule(
+                "nat",
+                &format!("-A POSTROUTING -m comment --comment {other} -j MASQUERADE"),
+            )
+            .with_rule("filter", "-A FORWARD -j ACCEPT");
+        clear_offer(&|args| fake.run(args), "tsun0");
+        let deleted = fake.verbs("-D");
+        assert_eq!(deleted.len(), 1, "{deleted:?}");
+        assert!(deleted[0].contains(&tag));
+    }
+
+    #[test]
+    fn a_missing_iptables_is_a_failure_that_says_what_to_do() {
+        let mut fake = Fake::new(false);
+        fake.missing = true;
+        let outcome = apply_offer(&|args| fake.run(args), "tsun0", &[range()]);
+        let RuleOutcome::Failed(reason) = &outcome[0].1 else {
+            panic!("{outcome:?}");
+        };
+        assert!(reason.contains("iptables not found"), "{reason}");
+        assert!(reason.contains("10.13.37.0/24"), "{reason}");
+        // Nothing wanted and nothing to run is not a failure of anything.
+        assert!(apply_offer(&|args| fake.run(args), "tsun0", &[]).is_empty());
+    }
+
+    #[test]
+    fn forwarding_is_read_per_interface_then_globally_and_never_written() {
+        let files = |per: Option<&'static str>, global: Option<&'static str>| {
+            move |path: &str| match path {
+                "/proc/sys/net/ipv4/conf/tsun0/forwarding" => per.map(str::to_string),
+                "/proc/sys/net/ipv4/ip_forward" => global.map(str::to_string),
+                _ => None,
+            }
+        };
+        assert_eq!(
+            forwarding_state(files(Some("1\n"), Some("0\n")), "tsun0"),
+            Some(true)
+        );
+        assert_eq!(
+            forwarding_state(files(Some("0\n"), Some("1\n")), "tsun0"),
+            Some(false)
+        );
+        assert_eq!(
+            forwarding_state(files(None, Some("1\n")), "tsun0"),
+            Some(true)
+        );
+        assert_eq!(forwarding_state(files(None, None), "tsun0"), None);
+        assert_eq!(forwarding_state(files(Some("junk"), None), "tsun0"), None);
+        // A name that would leave the directory is not looked up.
+        assert_eq!(forwarding_state(files(Some("1"), None), "../../x"), None);
+    }
+
+    #[test]
+    fn the_effective_uid_is_the_second_field() {
+        let status = "Name:\ttsunagi\nUid:\t1001\t1002\t1001\t1001\nGid:\t5\t5\t5\t5\n";
+        assert_eq!(parse_euid(status), Some(1002));
+        assert_eq!(parse_euid("Name:\tx\n"), None);
+    }
+
+    #[test]
+    fn a_login_session_is_refused_and_a_service_is_not() {
+        let reason = client_refusal(Some("/run/user/1000")).unwrap();
+        assert!(reason.contains("system service"), "{reason}");
+        assert!(client_refusal(None).is_none());
+        assert!(client_refusal(Some("")).is_none());
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/macos.rs

@@ -0,0 +1,233 @@
+//! Broadcast host rules on macOS: a `route` entry; the firewall is deferred.
+//!
+//! The limited-broadcast route is installed with `/sbin/route`, so a game's
+//! LAN discovery packet leaves through the overlay interface. It belongs to
+//! the interface and is removed when broadcast is turned off, the network
+//! leaves, or the agent stops — and, because a utun is torn down with the
+//! process, it also goes away on its own if the agent is killed.
+//!
+//! The inbound-UDP firewall allowance is **not** implemented here. On macOS
+//! that means a `pf` anchor, which is a good deal more machinery; until it is
+//! written the firewall half reports itself as missing with what to allow by
+//! hand, and the route is still installed. Failure is contained: neither half
+//! can stop the agent, and `status` shows exactly which is in place.
+//!
+//! `route` is a program, not a system call, so it is run by absolute path from
+//! `/sbin` rather than off `PATH`, and its arguments are built from values
+//! this agent computed — the interface name the kernel assigned, nothing a
+//! peer said.
+
+use crate::BoxFuture;
+
+use std::sync::{Arc, Mutex};
+
+use super::exit::{ExitHostPlan, ExitHostReport, ExitHostRules};
+use super::macos_exit::{apply_plan, clear_all, run_tool};
+use super::{BroadcastHostRules, BroadcastRulesPlan, BroadcastRulesReport, RuleOutcome};
+
+/// The limited-broadcast destination routed through the overlay interface.
+const LIMITED_BROADCAST: &str = "255.255.255.255";
+
+/// The `route` arguments that direct the limited broadcast out an interface.
+fn route_add_args(interface: &str) -> Vec<String> {
+    vec![
+        "-q".into(),
+        "-n".into(),
+        "add".into(),
+        "-host".into(),
+        LIMITED_BROADCAST.into(),
+        "-interface".into(),
+        interface.into(),
+    ]
+}
+
+/// The `route` arguments that remove it again.
+fn route_del_args(interface: &str) -> Vec<String> {
+    vec![
+        "-q".into(),
+        "-n".into(),
+        "delete".into(),
+        "-host".into(),
+        LIMITED_BROADCAST.into(),
+        "-interface".into(),
+        interface.into(),
+    ]
+}
+
+/// What to do by hand while the `pf` allowance is unimplemented.
+fn firewall_todo(plan: &BroadcastRulesPlan) -> String {
+    format!(
+        "the inbound-UDP firewall allowance is not implemented on macOS yet; if the host \
+         firewall drops discovery replies, allow inbound UDP from {} on `{}` with a pf rule \
+         yourself",
+        plan.range, plan.interface
+    )
+}
+
+/// The macOS implementation: `route` for the route, no firewall yet, and
+/// `pf` plus `route` for exit nodes (see [`super::macos_exit`]).
+#[derive(Debug, Default)]
+pub struct MacosHostRules {
+    /// The reference `pfctl -E` handed out, held while exit-node rules need
+    /// `pf` on and released when they are removed.
+    pub(super) pf_token: Arc<Mutex<Option<String>>>,
+}
+
+impl MacosHostRules {
+    /// Creates the host rules.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Installs the route, replacing one this agent added earlier so the call
+    /// can be repeated. The delete is best effort: a route that is not there
+    /// is the state the add then reaches anyway.
+    async fn set_route(&self, interface: String) -> RuleOutcome {
+        let _ = run(route_del_args(&interface)).await;
+        match run(route_add_args(&interface)).await {
+            Ok(()) => RuleOutcome::Applied,
+            Err(reason) => RuleOutcome::Failed(reason),
+        }
+    }
+}
+
+impl ExitHostRules for MacosHostRules {
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport> {
+        Box::pin(async move {
+            let job_plan = plan.clone();
+            let token = Arc::clone(&self.pf_token);
+            tokio::task::spawn_blocking(move || apply_plan(&run_tool, &token, &job_plan))
+                .await
+                .unwrap_or_else(|err| {
+                    let failed = RuleOutcome::Failed(format!("the task failed: {err}"));
+                    ExitHostReport {
+                        offer: plan
+                            .offer
+                            .iter()
+                            .map(|range| (*range, failed.clone()))
+                            .collect(),
+                        forwarding: None,
+                        client: plan.client.then_some(failed),
+                    }
+                })
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let name = interface.to_string();
+            let token = Arc::clone(&self.pf_token);
+            let _ = tokio::task::spawn_blocking(move || clear_all(&run_tool, &token, &name)).await;
+        })
+    }
+}
+
+impl BroadcastHostRules for MacosHostRules {
+    fn name(&self) -> &str {
+        "route"
+    }
+
+    fn exit_rules(&self) -> Option<&dyn ExitHostRules> {
+        Some(self)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            let route = self.set_route(plan.interface.clone()).await;
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route,
+                firewall: RuleOutcome::Failed(firewall_todo(plan)),
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            if let Err(reason) = run(route_del_args(interface)).await {
+                tracing::debug!(interface = %interface, %reason, "cannot remove the broadcast route");
+            }
+        })
+    }
+}
+
+/// Runs `/sbin/route`, turning a non-zero exit into the message it printed.
+async fn run(args: Vec<String>) -> Result<(), String> {
+    let display = format!("route {}", args.join(" "));
+    let output = tokio::task::spawn_blocking(move || {
+        std::process::Command::new("/sbin/route")
+            .args(&args)
+            .output()
+    })
+    .await
+    .map_err(|err| format!("could not run {display}: {err}"))?
+    .map_err(|err| format!("could not run {display}: {err}"))?;
+
+    if output.status.success() {
+        return Ok(());
+    }
+    let stderr = String::from_utf8_lossy(&output.stderr);
+    let text = stderr.trim();
+    if text.is_empty() {
+        Err(format!("{display} failed with {}", output.status))
+    } else {
+        Err(format!("{display} failed: {text}"))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::net::Ipv4Addr;
+
+    use super::*;
+
+    fn plan() -> BroadcastRulesPlan {
+        BroadcastRulesPlan {
+            interface: "utun5".into(),
+            source: Ipv4Addr::new(10, 13, 37, 142),
+            range: "10.13.37.0/24".parse().unwrap(),
+        }
+    }
+
+    #[test]
+    fn the_route_is_the_limited_broadcast_out_the_interface() {
+        assert_eq!(
+            route_add_args("utun5"),
+            vec![
+                "-q",
+                "-n",
+                "add",
+                "-host",
+                "255.255.255.255",
+                "-interface",
+                "utun5"
+            ]
+        );
+        assert_eq!(
+            route_del_args("utun5"),
+            vec![
+                "-q",
+                "-n",
+                "delete",
+                "-host",
+                "255.255.255.255",
+                "-interface",
+                "utun5"
+            ]
+        );
+    }
+
+    #[test]
+    fn the_firewall_todo_names_the_range_and_interface() {
+        let todo = firewall_todo(&plan());
+        assert!(todo.contains("10.13.37.0/24"), "{todo}");
+        assert!(todo.contains("utun5"), "{todo}");
+        assert!(todo.contains("pf"), "{todo}");
+        // Deferred, so the firewall half is always reported as not in place;
+        // the route runs for real and so is left to an on-host check. Running
+        // `route` here would need root and would change the host routing
+        // table, which the default test suite must never do.
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/macos_exit.rs

@@ -0,0 +1,915 @@
+//! Exit-node host rules on macOS: a `pf` anchor and `route` entries.
+//!
+//! Compiled on every platform so the rule text and the parsers are tested
+//! everywhere; only `MacosHostRules` runs them, and only on macOS.
+//!
+//! # Offering
+//!
+//! Per offered range, in the anchor `com.apple/tsunagi-exit-<if>`:
+//!
+//! ```text
+//! nat on <egress> inet from <range> to any -> (<egress>)
+//! pass in quick on <if> inet from <range> to ! <range>
+//! ```
+//!
+//! The anchor lives under `com.apple/` on purpose: the stock `/etc/pf.conf`
+//! evaluates `nat-anchor "com.apple/*"` and `anchor "com.apple/*"`, so the
+//! rules take effect without editing the user's ruleset. `pf` cannot say
+//! "every interface but one" in a translation address, so the egress is the
+//! interface of the default route **when the rules are applied**; after the
+//! default route moves (Wi-Fi to Ethernet) the rules are re-applied by
+//! switching the exit node off and on, or restarting the agent.
+//!
+//! The kernel's `net.inet.ip.forwarding` is only read, never written: it
+//! changes how the whole host behaves, which is the owner's decision.
+//!
+//! # Using
+//!
+//! macOS has no `uidrange` policy routing, and the agent is root, so the
+//! Linux scheme does not carry over. Instead:
+//!
+//! ```text
+//! route add -net 0.0.0.0/1   -interface <if>
+//! route add -net 128.0.0.0/1 -interface <if>
+//! pass out quick on <if> route-to (<egress> <gateway>) inet proto { tcp udp }
+//!     from any to ! <overlay ranges> user <agent uid>
+//! nat on <egress> inet from (<if>) to any -> (<egress>)
+//! ```
+//!
+//! The two halves outrank the default route without replacing it, and every
+//! more specific route (LAN, another VPN) keeps working. The `route-to` rule
+//! sends the *agent's own* traffic — QUIC, relay, DHT, which would otherwise
+//! loop into the tunnel it carries — back out the physical interface, and the
+//! translation gives it that interface's address, since an unbound socket
+//! picked the overlay address when the route pointed at the overlay. Like the
+//! Linux exemption this is by user id, so anything else running as root also
+//! bypasses the tunnel.
+//!
+//! IPv6 is blocked while an exit node is in use, as on Linux: `::/1` and
+//! `8000::/1` are rejecting routes, so applications fall back to IPv4 at once
+//! instead of leaking.
+//!
+//! Every object is derived from local state and the interface the agent owns;
+//! none of it comes from a peer.
+
+use std::net::Ipv4Addr;
+use std::sync::Mutex;
+
+use crate::state::Ipv4Range;
+
+use super::RuleOutcome;
+use super::exit::{ExitHostPlan, ExitHostReport};
+
+/// The programs this module runs, always by absolute path.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(super) enum Tool {
+    Route,
+    Pfctl,
+    Sysctl,
+    Id,
+}
+
+impl Tool {
+    fn path(self) -> &'static str {
+        match self {
+            Self::Route => "/sbin/route",
+            Self::Pfctl => "/sbin/pfctl",
+            Self::Sysctl => "/usr/sbin/sysctl",
+            Self::Id => "/usr/bin/id",
+        }
+    }
+}
+
+/// What running a tool came to.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub(super) enum Ran {
+    /// The program is not there.
+    Missing,
+    /// It ran.
+    Exited {
+        success: bool,
+        stdout: String,
+        stderr: String,
+    },
+}
+
+/// Runs a tool with optional standard input; injectable so the logic is
+/// tested without touching the host.
+pub(super) type Run<'a> = &'a dyn Fn(Tool, &[String], Option<&str>) -> Ran;
+
+/// Runs a real tool. Blocking: callers use `spawn_blocking`.
+pub(super) fn run_tool(tool: Tool, args: &[String], stdin: Option<&str>) -> Ran {
+    use std::io::Write;
+    use std::process::{Command, Stdio};
+
+    let mut command = Command::new(tool.path());
+    command
+        .args(args)
+        .stdin(if stdin.is_some() {
+            Stdio::piped()
+        } else {
+            Stdio::null()
+        })
+        .stdout(Stdio::piped())
+        .stderr(Stdio::piped());
+    let mut child = match command.spawn() {
+        Ok(child) => child,
+        Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ran::Missing,
+        Err(err) => {
+            return Ran::Exited {
+                success: false,
+                stdout: String::new(),
+                stderr: format!("could not run {}: {err}", tool.path()),
+            };
+        }
+    };
+    if let (Some(text), Some(mut pipe)) = (stdin, child.stdin.take()) {
+        // A write error shows up as the tool's own failure below.
+        let _ = pipe.write_all(text.as_bytes());
+    }
+    match child.wait_with_output() {
+        Ok(output) => Ran::Exited {
+            success: output.status.success(),
+            stdout: String::from_utf8_lossy(&output.stdout).into_owned(),
+            stderr: String::from_utf8_lossy(&output.stderr).into_owned(),
+        },
+        Err(err) => Ran::Exited {
+            success: false,
+            stdout: String::new(),
+            stderr: format!("could not run {}: {err}", tool.path()),
+        },
+    }
+}
+
+fn owned(words: &[&str]) -> Vec<String> {
+    words.iter().map(|word| word.to_string()).collect()
+}
+
+/// A name that is safe in a `pf` rule, an anchor path and a table name.
+fn safe_name(name: &str) -> bool {
+    !name.is_empty() && name.len() <= 32 && name.chars().all(|c| c.is_ascii_alphanumeric())
+}
+
+/// The anchor holding every rule of one interface.
+fn anchor_name(interface: &str) -> Result<String, String> {
+    if safe_name(interface) {
+        Ok(format!("com.apple/tsunagi-exit-{interface}"))
+    } else {
+        Err(format!(
+            "`{interface}` is not an interface name this agent can write a pf rule for"
+        ))
+    }
+}
+
+/// The default route as `route -n get default` reports it.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub(super) struct DefaultRoute {
+    pub interface: String,
+    pub gateway: Option<Ipv4Addr>,
+}
+
+/// Reads `interface:` and `gateway:` out of `route -n get default`.
+fn parse_default_route(text: &str) -> Option<DefaultRoute> {
+    let field = |name: &str| {
+        text.lines().find_map(|line| {
+            let (key, value) = line.split_once(':')?;
+            (key.trim() == name).then(|| value.trim().to_string())
+        })
+    };
+    let interface = field("interface").filter(|name| safe_name(name))?;
+    let gateway = field("gateway").and_then(|value| value.parse().ok());
+    Some(DefaultRoute { interface, gateway })
+}
+
+fn failure(stderr: &str, stdout: &str) -> String {
+    let text = if stderr.trim().is_empty() {
+        stdout.trim()
+    } else {
+        stderr.trim()
+    };
+    if text.is_empty() {
+        "the command failed without saying why".to_string()
+    } else {
+        text.to_string()
+    }
+}
+
+/// Runs a tool and turns a failure into a sentence a person can act on.
+fn run_ok(
+    run: Run<'_>,
+    tool: Tool,
+    args: &[String],
+    stdin: Option<&str>,
+) -> Result<String, String> {
+    match run(tool, args, stdin) {
+        Ran::Missing => Err(format!("{} not found", tool.path())),
+        Ran::Exited {
+            success: true,
+            stdout,
+            ..
+        } => Ok(stdout),
+        Ran::Exited { stderr, stdout, .. } => Err(format!(
+            "{} {} failed: {}",
+            tool.path(),
+            args.join(" "),
+            failure(&stderr, &stdout)
+        )),
+    }
+}
+
+/// The route traffic leaves by now, which is the egress for translation and
+/// for the agent's own sockets. Read *before* the overlay routes are added.
+fn default_route(run: Run<'_>, overlay: &str) -> Result<DefaultRoute, String> {
+    let out = run_ok(run, Tool::Route, &owned(&["-n", "get", "default"]), None)
+        .map_err(|why| format!("cannot find the default route: {why}"))?;
+    let route = parse_default_route(&out)
+        .ok_or_else(|| "the host has no default route to send traffic out of".to_string())?;
+    if route.interface == overlay {
+        return Err(format!(
+            "the default route already goes through `{overlay}`"
+        ));
+    }
+    Ok(route)
+}
+
+/// Whether the kernel forwards IPv4. Read only.
+fn forwarding_state(run: Run<'_>) -> Option<bool> {
+    let out = run_ok(
+        run,
+        Tool::Sysctl,
+        &owned(&["-n", "net.inet.ip.forwarding"]),
+        None,
+    )
+    .ok()?;
+    match out.trim() {
+        "1" => Some(true),
+        "0" => Some(false),
+        _ => None,
+    }
+}
+
+/// The effective user id of this process, from `id -u`.
+fn agent_uid(run: Run<'_>) -> Result<u32, String> {
+    run_ok(run, Tool::Id, &owned(&["-u"]), None)
+        .ok()
+        .and_then(|out| out.trim().parse().ok())
+        .ok_or_else(|| "cannot tell which user this agent runs as".to_string())
+}
+
+/// What the using side adds to the rules.
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct ClientRules<'a> {
+    uid: u32,
+    /// Every overlay range: the agent's own traffic *to* those still goes
+    /// through the interface.
+    overlay: &'a [Ipv4Range],
+}
+
+/// The text of the anchor: translation first, filtering after, which is the
+/// order `pf` insists on.
+fn pf_rules(
+    interface: &str,
+    egress: &DefaultRoute,
+    offer: &[Ipv4Range],
+    client: Option<&ClientRules<'_>>,
+) -> String {
+    let out = &egress.interface;
+    let mut nat = Vec::new();
+    let mut filter = Vec::new();
+    for range in offer {
+        nat.push(format!("nat on {out} inet from {range} to any -> ({out})"));
+        filter.push(format!(
+            "pass in quick on {interface} inet from {range} to ! {range}"
+        ));
+    }
+    let mut tables = Vec::new();
+    if let Some(client) = client {
+        nat.push(format!(
+            "nat on {out} inet from ({interface}) to any -> ({out})"
+        ));
+        let target = if client.overlay.is_empty() {
+            "any".to_string()
+        } else {
+            let ranges: Vec<String> = client.overlay.iter().map(|r| r.to_string()).collect();
+            // A table, because `! { a, b }` expands to one rule per element
+            // in older `pf`, which matches everything.
+            tables.push(format!(
+                "table <tsunagi_overlay_{interface}> const {{ {} }}",
+                ranges.join(", ")
+            ));
+            format!("! <tsunagi_overlay_{interface}>")
+        };
+        let via = match egress.gateway {
+            Some(gateway) => format!("({out} {gateway})"),
+            None => out.clone(),
+        };
+        filter.push(format!(
+            "pass out quick on {interface} route-to {via} inet proto {{ tcp udp }} \
+             from any to {target} user {}",
+            client.uid
+        ));
+    }
+    tables
+        .into_iter()
+        .chain(nat)
+        .chain(filter)
+        .collect::<Vec<_>>()
+        .join("\n")
+        + "\n"
+}
+
+/// Turns `pf` on if nobody has, holding the reference `pfctl` hands out so it
+/// can be released when the agent is done. One reference at most.
+fn acquire_pf(run: Run<'_>, slot: &Mutex<Option<String>>) -> Result<(), String> {
+    if lock(slot).is_some() {
+        return Ok(());
+    }
+    match run(Tool::Pfctl, &owned(&["-E"]), None) {
+        Ran::Missing => Err(format!("{} not found", Tool::Pfctl.path())),
+        Ran::Exited {
+            success,
+            stdout,
+            stderr,
+        } => {
+            // `pfctl -E` reports `Token : <n>` on stderr, and also says so
+            // when pf is already enabled.
+            let token = stdout.lines().chain(stderr.lines()).find_map(parse_token);
+            if !success && token.is_none() {
+                return Err(format!("cannot enable pf: {}", failure(&stderr, &stdout)));
+            }
+            *lock(slot) = token;
+            Ok(())
+        }
+    }
+}
+
+fn parse_token(line: &str) -> Option<String> {
+    let (key, value) = line.split_once(':')?;
+    let value = value.trim();
+    (key.trim().eq_ignore_ascii_case("token")
+        && !value.is_empty()
+        && value.chars().all(|c| c.is_ascii_digit()))
+    .then(|| value.to_string())
+}
+
+fn release_pf(run: Run<'_>, slot: &Mutex<Option<String>>) {
+    if let Some(token) = lock(slot).take()
+        && let Err(why) = run_ok(run, Tool::Pfctl, &["-X".to_string(), token], None)
+    {
+        tracing::debug!(%why, "cannot release the pf reference");
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+fn flush_args(anchor: &str) -> Vec<String> {
+    owned(&["-q", "-a", anchor, "-F", "all"])
+}
+
+fn load_args(anchor: &str) -> Vec<String> {
+    owned(&["-q", "-a", anchor, "-f", "-"])
+}
+
+/// The two halves that out-rank the default route.
+const HALVES: [&str; 2] = ["0.0.0.0/1", "128.0.0.0/1"];
+/// The IPv6 halves, made unreachable.
+const HALVES_V6: [&str; 2] = ["::/1", "8000::/1"];
+
+fn route_add_args(half: &str, interface: &str) -> Vec<String> {
+    owned(&["-q", "-n", "add", "-net", half, "-interface", interface])
+}
+
+fn route_del_args(half: &str, interface: &str) -> Vec<String> {
+    owned(&["-q", "-n", "delete", "-net", half, "-interface", interface])
+}
+
+fn route_add_v6_args(half: &str) -> Vec<String> {
+    owned(&["-q", "-n", "add", "-inet6", "-net", half, "::1", "-reject"])
+}
+
+fn route_del_v6_args(half: &str) -> Vec<String> {
+    owned(&["-q", "-n", "delete", "-inet6", "-net", half, "::1"])
+}
+
+/// Takes away the overlay default routes. Best effort: what is not there is
+/// the state wanted.
+fn clear_routes(run: Run<'_>, interface: &str) {
+    for half in HALVES {
+        let _ = run(Tool::Route, &route_del_args(half, interface), None);
+    }
+    for half in HALVES_V6 {
+        let _ = run(Tool::Route, &route_del_v6_args(half), None);
+    }
+}
+
+fn set_routes(run: Run<'_>, interface: &str) -> Result<(), String> {
+    clear_routes(run, interface);
+    for half in HALVES {
+        if let Err(why) = run_ok(run, Tool::Route, &route_add_args(half, interface), None) {
+            clear_routes(run, interface);
+            return Err(why);
+        }
+    }
+    for half in HALVES_V6 {
+        // A host without IPv6 has nothing to block; not a failure.
+        if let Err(why) = run_ok(run, Tool::Route, &route_add_v6_args(half), None) {
+            tracing::debug!(%why, "cannot block IPv6 for the exit node");
+        }
+    }
+    Ok(())
+}
+
+/// Makes the host match the plan.
+pub(super) fn apply_plan(
+    run: Run<'_>,
+    token: &Mutex<Option<String>>,
+    plan: &ExitHostPlan,
+) -> ExitHostReport {
+    let forwarding = if plan.offer.is_empty() {
+        None
+    } else {
+        forwarding_state(run)
+    };
+    let outcome = apply_rules(run, token, plan);
+    let verdict = |ok: bool, why: &str| {
+        if ok {
+            RuleOutcome::Applied
+        } else {
+            RuleOutcome::Failed(why.to_string())
+        }
+    };
+    match outcome {
+        Ok(()) => ExitHostReport {
+            offer: plan
+                .offer
+                .iter()
+                .map(|range| (*range, RuleOutcome::Applied))
+                .collect(),
+            forwarding,
+            client: plan.client.then_some(RuleOutcome::Applied),
+        },
+        Err(why) => ExitHostReport {
+            offer: plan
+                .offer
+                .iter()
+                .map(|range| (*range, verdict(false, &why)))
+                .collect(),
+            forwarding,
+            client: plan.client.then(|| verdict(false, &why)),
+        },
+    }
+}
+
+fn apply_rules(
+    run: Run<'_>,
+    token: &Mutex<Option<String>>,
+    plan: &ExitHostPlan,
+) -> Result<(), String> {
+    let interface = plan.interface.as_str();
+    let anchor = anchor_name(interface)?;
+    let egress = default_route(run, interface)?;
+    let uid = if plan.client {
+        Some(agent_uid(run)?)
+    } else {
+        None
+    };
+    let client = uid.map(|uid| ClientRules {
+        uid,
+        overlay: &plan.overlay,
+    });
+    let rules = pf_rules(interface, &egress, &plan.offer, client.as_ref());
+
+    acquire_pf(run, token)?;
+    run_ok(run, Tool::Pfctl, &load_args(&anchor), Some(&rules))
+        .map_err(|why| format!("cannot load the pf rules: {why}"))?;
+    // The routes come after the rules that keep the agent's own traffic off
+    // them: the other order would loop for as long as it takes.
+    if plan.client {
+        set_routes(run, interface)?;
+    } else {
+        clear_routes(run, interface);
+    }
+    Ok(())
+}
+
+/// Removes everything this agent put there for the interface.
+pub(super) fn clear_all(run: Run<'_>, token: &Mutex<Option<String>>, interface: &str) {
+    clear_routes(run, interface);
+    if let Ok(anchor) = anchor_name(interface) {
+        // Flushing an anchor nothing was ever loaded into is not an error
+        // worth reporting.
+        if let Err(why) = run_ok(run, Tool::Pfctl, &flush_args(&anchor), None) {
+            tracing::debug!(%why, "cannot flush the exit node pf anchor");
+        }
+    }
+    release_pf(run, token);
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::cell::RefCell;
+
+    use super::*;
+
+    fn range() -> Ipv4Range {
+        "10.13.37.0/24".parse().unwrap()
+    }
+
+    const DEFAULT_ROUTE: &str = "   route to: default\n\
+destination: default\n\
+       mask: default\n\
+    gateway: 192.168.1.1\n\
+  interface: en0\n\
+      flags: <UP,GATEWAY,DONE,STATIC,PRCLONING,GLOBAL>\n";
+
+    fn egress() -> DefaultRoute {
+        DefaultRoute {
+            interface: "en0".into(),
+            gateway: Some(Ipv4Addr::new(192, 168, 1, 1)),
+        }
+    }
+
+    /// A scripted host that records every call.
+    struct Fake {
+        calls: RefCell<Vec<String>>,
+        stdins: RefCell<Vec<String>>,
+        default_route: Option<String>,
+        forwarding: &'static str,
+        pf_missing: bool,
+        pf_load_fails: bool,
+        route_add_fails: bool,
+        pf_enable_output: &'static str,
+    }
+
+    impl Fake {
+        fn new() -> Self {
+            Self {
+                calls: RefCell::new(Vec::new()),
+                stdins: RefCell::new(Vec::new()),
+                default_route: Some(DEFAULT_ROUTE.to_string()),
+                forwarding: "1\n",
+                pf_missing: false,
+                pf_load_fails: false,
+                route_add_fails: false,
+                pf_enable_output: "pf enabled\nToken : 4242\n",
+            }
+        }
+
+        fn run(&self, tool: Tool, args: &[String], stdin: Option<&str>) -> Ran {
+            let line = format!("{:?} {}", tool, args.join(" "));
+            self.calls.borrow_mut().push(line);
+            if let Some(text) = stdin {
+                self.stdins.borrow_mut().push(text.to_string());
+            }
+            let ok = |stdout: &str, stderr: &str| Ran::Exited {
+                success: true,
+                stdout: stdout.to_string(),
+                stderr: stderr.to_string(),
+            };
+            let fail = |why: &str| Ran::Exited {
+                success: false,
+                stdout: String::new(),
+                stderr: why.to_string(),
+            };
+            match tool {
+                Tool::Id => ok("0\n", ""),
+                Tool::Sysctl => ok(self.forwarding, ""),
+                Tool::Route if args.first().map(String::as_str) == Some("-n") => {
+                    match &self.default_route {
+                        Some(text) => ok(text, ""),
+                        None => fail("route: writing to routing socket: not in table"),
+                    }
+                }
+                Tool::Route if args.contains(&"add".to_string()) && self.route_add_fails => {
+                    fail("route: writing to routing socket: File exists")
+                }
+                Tool::Route => ok("", ""),
+                Tool::Pfctl if self.pf_missing => Ran::Missing,
+                Tool::Pfctl if args == ["-E"] => ok("", self.pf_enable_output),
+                Tool::Pfctl if args.contains(&"-f".to_string()) && self.pf_load_fails => {
+                    fail("pfctl: Syntax error")
+                }
+                Tool::Pfctl => ok("", ""),
+            }
+        }
+
+        fn calls(&self) -> Vec<String> {
+            self.calls.borrow().clone()
+        }
+
+        fn position(&self, needle: &str) -> Option<usize> {
+            self.calls.borrow().iter().position(|c| c.contains(needle))
+        }
+
+        fn count(&self, needle: &str) -> usize {
+            self.calls
+                .borrow()
+                .iter()
+                .filter(|c| c.contains(needle))
+                .count()
+        }
+    }
+
+    fn plan(offer: bool, client: bool) -> ExitHostPlan {
+        ExitHostPlan {
+            interface: "utun5".into(),
+            offer: if offer { vec![range()] } else { Vec::new() },
+            client,
+            overlay: vec![range()],
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn the_default_route_is_read_from_route_output() {
+        assert_eq!(parse_default_route(DEFAULT_ROUTE), Some(egress()));
+        // A point-to-point default has no gateway line to use.
+        let ppp = "destination: default\n  interface: ppp0\n";
+        assert_eq!(
+            parse_default_route(ppp),
+            Some(DefaultRoute {
+                interface: "ppp0".into(),
+                gateway: None
+            })
+        );
+        assert_eq!(parse_default_route("route: not in table"), None);
+        // An interface name that is not plain letters and digits never ends
+        // up in a rule.
+        assert_eq!(parse_default_route("interface: en0; pass all\n"), None);
+    }
+
+    #[test]
+    fn the_anchor_is_under_com_apple_and_only_for_plain_names() {
+        assert_eq!(
+            anchor_name("utun5").unwrap(),
+            "com.apple/tsunagi-exit-utun5"
+        );
+        assert!(anchor_name("").is_err());
+        assert!(anchor_name("utun5 pass all").is_err());
+        assert!(anchor_name("../x").is_err());
+    }
+
+    #[test]
+    fn offering_translates_the_range_and_admits_it_in() {
+        let text = pf_rules("utun5", &egress(), &[range()], None);
+        assert_eq!(
+            text,
+            "nat on en0 inet from 10.13.37.0/24 to any -> (en0)\n\
+             pass in quick on utun5 inet from 10.13.37.0/24 to ! 10.13.37.0/24\n"
+        );
+        // Translation precedes filtering, as pf requires.
+        assert!(text.find("nat on").unwrap() < text.find("pass in").unwrap());
+    }
+
+    #[test]
+    fn using_sends_the_agents_own_traffic_back_out_the_physical_interface() {
+        let other: Ipv4Range = "10.99.0.0/16".parse().unwrap();
+        let overlay = [range(), other];
+        let client = ClientRules {
+            uid: 0,
+            overlay: &overlay,
+        };
+        let text = pf_rules("utun5", &egress(), &[], Some(&client));
+        assert!(
+            text.starts_with(
+                "table <tsunagi_overlay_utun5> const { 10.13.37.0/24, 10.99.0.0/16 }\n"
+            ),
+            "{text}"
+        );
+        assert!(
+            text.contains("nat on en0 inet from (utun5) to any -> (en0)\n"),
+            "{text}"
+        );
+        assert!(
+            text.contains(
+                "pass out quick on utun5 route-to (en0 192.168.1.1) inet proto { tcp udp } \
+                 from any to ! <tsunagi_overlay_utun5> user 0\n"
+            ),
+            "{text}"
+        );
+        // Only the agent's user, and only what is not the overlay.
+        assert!(!text.contains("from any to any"));
+
+        let direct = DefaultRoute {
+            interface: "ppp0".into(),
+            gateway: None,
+        };
+        let text = pf_rules(
+            "utun5",
+            &direct,
+            &[],
+            Some(&ClientRules {
+                uid: 501,
+                overlay: &[],
+            }),
+        );
+        assert!(text.contains("route-to ppp0 inet"), "{text}");
+        assert!(text.contains("to any user 501"), "{text}");
+        assert!(!text.contains("table"), "{text}");
+    }
+
+    #[test]
+    fn the_overlay_default_routes_are_halves_and_ipv6_is_rejected() {
+        assert_eq!(
+            route_add_args("0.0.0.0/1", "utun5").join(" "),
+            "-q -n add -net 0.0.0.0/1 -interface utun5"
+        );
+        assert_eq!(
+            route_del_args("128.0.0.0/1", "utun5").join(" "),
+            "-q -n delete -net 128.0.0.0/1 -interface utun5"
+        );
+        assert_eq!(
+            route_add_v6_args("8000::/1").join(" "),
+            "-q -n add -inet6 -net 8000::/1 ::1 -reject"
+        );
+        assert_eq!(
+            route_del_v6_args("::/1").join(" "),
+            "-q -n delete -inet6 -net ::/1 ::1"
+        );
+    }
+
+    #[test]
+    fn the_pf_token_is_found_in_either_stream() {
+        assert_eq!(parse_token("Token : 1234"), Some("1234".into()));
+        assert_eq!(parse_token("token: 9"), Some("9".into()));
+        assert_eq!(parse_token("pf enabled"), None);
+        assert_eq!(parse_token("Token : x; rm"), None);
+    }
+
+    #[test]
+    fn offering_loads_the_anchor_and_reads_but_never_writes_forwarding() {
+        let fake = Fake::new();
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        assert_eq!(report.offer, vec![(range(), RuleOutcome::Applied)]);
+        assert_eq!(report.forwarding, Some(true));
+        assert_eq!(report.client, None);
+        let calls = fake.calls();
+        assert!(
+            calls
+                .iter()
+                .any(|c| c == "Pfctl -q -a com.apple/tsunagi-exit-utun5 -f -"),
+            "{calls:?}"
+        );
+        assert_eq!(
+            fake.stdins.borrow()[0],
+            "nat on en0 inet from 10.13.37.0/24 to any -> (en0)\n\
+             pass in quick on utun5 inet from 10.13.37.0/24 to ! 10.13.37.0/24\n"
+        );
+        // sysctl is only ever asked, never told.
+        assert!(
+            calls
+                .iter()
+                .filter(|c| c.starts_with("Sysctl"))
+                .all(|c| c == "Sysctl -n net.inet.ip.forwarding"),
+            "{calls:?}"
+        );
+        // No client routes were added for an offer alone.
+        assert_eq!(fake.count(" add "), 0);
+        assert_eq!(token.lock().unwrap().as_deref(), Some("4242"));
+    }
+
+    #[test]
+    fn forwarding_off_is_reported_and_unreadable_is_none() {
+        let mut fake = Fake::new();
+        fake.forwarding = "0\n";
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        assert_eq!(report.forwarding, Some(false));
+        fake.forwarding = "junk";
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        assert_eq!(report.forwarding, None);
+    }
+
+    #[test]
+    fn the_client_rules_are_loaded_before_the_routes_that_need_them() {
+        let fake = Fake::new();
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(false, true));
+        assert_eq!(report.client, Some(RuleOutcome::Applied));
+        assert!(report.offer.is_empty());
+        let load = fake.position("-f -").unwrap();
+        let add = fake.position(" add -net 0.0.0.0/1").unwrap();
+        assert!(load < add, "{:?}", fake.calls());
+        assert!(fake.position("add -net 128.0.0.0/1").is_some());
+        assert!(fake.position("-inet6 -net ::/1 ::1 -reject").is_some());
+        assert!(fake.stdins.borrow()[0].contains("user 0"));
+        // The default route was read before any overlay route existed.
+        assert!(fake.position("-n get default").unwrap() < add);
+    }
+
+    #[test]
+    fn no_routes_are_added_when_the_rules_did_not_load() {
+        let mut fake = Fake::new();
+        fake.pf_load_fails = true;
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, true));
+        let Some(RuleOutcome::Failed(why)) = report.client else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("cannot load the pf rules"), "{why}");
+        assert!(why.contains("Syntax error"), "{why}");
+        assert!(matches!(report.offer[0].1, RuleOutcome::Failed(_)));
+        // Nothing sent the host's traffic into a tunnel it could not exempt
+        // the agent from.
+        assert_eq!(fake.count(" add "), 0);
+    }
+
+    #[test]
+    fn a_failing_route_is_a_failure_and_leaves_nothing_behind() {
+        let mut fake = Fake::new();
+        fake.route_add_fails = true;
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(false, true));
+        let Some(RuleOutcome::Failed(why)) = report.client else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("File exists"), "{why}");
+        // The cleanup after the failed add removed whatever did land.
+        let last_add = fake
+            .calls()
+            .iter()
+            .rposition(|c| c.contains(" add "))
+            .unwrap();
+        assert!(
+            fake.calls()[last_add..]
+                .iter()
+                .any(|c| c.contains("delete"))
+        );
+    }
+
+    #[test]
+    fn a_missing_pfctl_or_default_route_says_what_is_wrong() {
+        let mut fake = Fake::new();
+        fake.pf_missing = true;
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        let RuleOutcome::Failed(why) = &report.offer[0].1 else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("/sbin/pfctl not found"), "{why}");
+
+        let mut fake = Fake::new();
+        fake.default_route = None;
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        let RuleOutcome::Failed(why) = &report.offer[0].1 else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("default route"), "{why}");
+        assert_eq!(fake.count("-f -"), 0);
+    }
+
+    #[test]
+    fn a_default_route_through_the_overlay_is_refused() {
+        let mut fake = Fake::new();
+        fake.default_route = Some("gateway: 10.13.37.1\ninterface: utun5\n".into());
+        let token = Mutex::new(None);
+        let report = apply_plan(&|t, a, s| fake.run(t, a, s), &token, &plan(true, false));
+        let RuleOutcome::Failed(why) = &report.offer[0].1 else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("already goes through"), "{why}");
+    }
+
+    #[test]
+    fn pf_is_enabled_once_and_released_once_on_clear() {
+        let fake = Fake::new();
+        let token = Mutex::new(None);
+        let run = |t, a: &[String], s: Option<&str>| fake.run(t, a, s);
+        apply_plan(&run, &token, &plan(true, false));
+        apply_plan(&run, &token, &plan(true, true));
+        assert_eq!(fake.count("Pfctl -E"), 1, "{:?}", fake.calls());
+
+        clear_all(&run, &token, "utun5");
+        assert_eq!(fake.count("Pfctl -X 4242"), 1);
+        assert!(
+            fake.position("-a com.apple/tsunagi-exit-utun5 -F all")
+                .is_some()
+        );
+        assert!(
+            fake.position("delete -net 0.0.0.0/1 -interface utun5")
+                .is_some()
+        );
+        assert_eq!(*token.lock().unwrap(), None);
+        // A second clear has nothing to release and still succeeds.
+        clear_all(&run, &token, "utun5");
+        assert_eq!(fake.count("Pfctl -X"), 1);
+    }
+
+    #[test]
+    fn dropping_the_client_removes_its_routes() {
+        let fake = Fake::new();
+        let token = Mutex::new(None);
+        let run = |t, a: &[String], s: Option<&str>| fake.run(t, a, s);
+        apply_plan(&run, &token, &plan(true, false));
+        assert!(fake.position("delete -net 0.0.0.0/1").is_some());
+        assert_eq!(fake.count(" add "), 0);
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/mod.rs

@@ -0,0 +1,365 @@
+//! Host rules that let a game's LAN discovery reach the overlay interface.
+//!
+//! Broadcast fanout happens at local TUN ingress (see [`super::broadcast`]),
+//! so a limited broadcast only reaches Tsunagi if the operating system routes
+//! it through the overlay interface, and a discovery *reply* only reaches the
+//! game if the host firewall lets it in. Left to the user that is two manual
+//! steps per host that nobody remembers, so when a network's broadcast is on
+//! the agent installs exactly these two things for the interface it owns:
+//!
+//! * a `255.255.255.255/32` route through the overlay interface, with the
+//!   overlay address as its preferred source, so the game's discovery packet
+//!   leaves with a source address the router recognises;
+//! * an inbound firewall allowance for **UDP only**, from the overlay range
+//!   only, on that interface only. Replies to a broadcast come from a
+//!   different address than the one the request was sent to, so a stateful
+//!   firewall never matches them to the request. Opening every protocol
+//!   would expose services such as SSH to every member of the network, which
+//!   is not what a game needs.
+//!
+//! # Why this is allowed here
+//!
+//! The architecture says plugins never touch routing or firewall settings,
+//! and nothing a peer announces becomes an OS setting. This is neither: it is
+//! the system level, which owns the interface, configuring that interface from
+//! values it computed itself (the interface name is derived locally, the
+//! address and range come from signed state it already verified). Every object
+//! is tagged `tsunagi:<interface>`, reversible, and removed when broadcast is
+//! turned off, the network leaves, or the agent stops. A crash can leave a
+//! firewall rule behind; the next apply for the same interface removes it.
+//!
+//! # Failure is contained
+//!
+//! Neither step can stop the agent. Each reports its own outcome, so a missing
+//! firewall tool still leaves the route in place and says what is missing.
+//!
+//! # Why a trait
+//!
+//! Linux, Windows and the tests do this very differently, while *what* to
+//! install is the same everywhere and lives in [`BroadcastRulesPlan`] and
+//! [`rules_summary`], tested on every platform.
+
+use std::net::Ipv4Addr;
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+use crate::state::Ipv4Range;
+
+mod exit;
+pub use exit::{
+    ExitHostPlan, ExitHostReport, ExitHostRules, MockExitRules, exit_tag, exit_tag_prefix,
+    offer_detail,
+};
+
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+mod linux;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+mod linux_exit;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub use linux::LinuxHostRules;
+
+// Compiled everywhere so the scripts it builds are tested everywhere; only
+// the type that runs them is exported on Windows.
+mod windows;
+#[cfg_attr(not(target_os = "windows"), allow(dead_code))]
+mod windows_exit;
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use windows::WindowsHostRules;
+
+// Compiled everywhere so its rule text and parsers are tested everywhere.
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+mod macos;
+#[cfg_attr(not(target_os = "macos"), allow(dead_code))]
+mod macos_exit;
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use macos::MacosHostRules;
+
+/// What the host should be told for one interface.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct BroadcastRulesPlan {
+    /// The overlay interface. Derived locally from the network id, never from
+    /// anything a peer said.
+    pub interface: String,
+    /// This agent's overlay address, the route's preferred source.
+    pub source: Ipv4Addr,
+    /// The overlay range inbound UDP is admitted from.
+    pub range: Ipv4Range,
+}
+
+impl BroadcastRulesPlan {
+    /// The tag every object carries, so it can be found again and nothing
+    /// else is ever touched.
+    pub fn tag(&self) -> String {
+        tag_for(&self.interface)
+    }
+}
+
+/// The tag for an interface's objects.
+pub fn tag_for(interface: &str) -> String {
+    format!("tsunagi:{interface}")
+}
+
+/// What happened to one of the two host objects.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum RuleOutcome {
+    /// It is in place.
+    Applied,
+    /// It is not, with a reason a person can act on.
+    Failed(String),
+}
+
+impl RuleOutcome {
+    /// Whether the object is in place.
+    pub fn is_applied(&self) -> bool {
+        matches!(self, Self::Applied)
+    }
+}
+
+/// What applying a plan achieved.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct BroadcastRulesReport {
+    /// What was asked for.
+    pub plan: BroadcastRulesPlan,
+    /// The `255.255.255.255` route.
+    pub route: RuleOutcome,
+    /// The inbound firewall allowance.
+    pub firewall: RuleOutcome,
+}
+
+impl BroadcastRulesReport {
+    /// Whether both objects are in place.
+    pub fn is_ok(&self) -> bool {
+        self.route.is_applied() && self.firewall.is_applied()
+    }
+
+    /// One line for `status`.
+    pub fn summary(&self) -> String {
+        rules_summary(&self.route, &self.firewall)
+    }
+}
+
+/// The status line for a pair of outcomes.
+pub fn rules_summary(route: &RuleOutcome, firewall: &RuleOutcome) -> String {
+    match (route, firewall) {
+        (RuleOutcome::Applied, RuleOutcome::Applied) => "ok (route + firewall)".to_string(),
+        (route, firewall) => {
+            let part = |what: &str, outcome: &RuleOutcome| match outcome {
+                RuleOutcome::Applied => format!("{what} ok"),
+                RuleOutcome::Failed(reason) => format!("{what} failed: {reason}"),
+            };
+            format!(
+                "incomplete; {}; {}",
+                part("route", route),
+                part("firewall", firewall)
+            )
+        }
+    }
+}
+
+/// Installs and removes the host rules for an overlay interface.
+///
+/// Both calls are idempotent. Neither can fail as a whole: a step that cannot
+/// be done is reported in the [`BroadcastRulesReport`], because the caller's
+/// only useful response is to say so.
+pub trait BroadcastHostRules: Send + Sync + std::fmt::Debug + 'static {
+    /// A short name used in diagnostics.
+    fn name(&self) -> &str;
+
+    /// Makes the host match the plan, replacing whatever this agent put there
+    /// for the same interface before.
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport>;
+
+    /// Removes what this agent put there for the interface. Succeeds when
+    /// nothing is there: it runs on the shutdown path.
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()>;
+
+    /// The exit-node rules that go with these, when the platform has them.
+    ///
+    /// One object per host, because both halves need the same privileged
+    /// helpers; the default is that this platform has none, which is said
+    /// where it matters rather than pretended.
+    fn exit_rules(&self) -> Option<&dyn ExitHostRules> {
+        None
+    }
+}
+
+/// The platform has no implementation, which is said rather than pretended.
+#[derive(Debug, Default)]
+pub struct UnsupportedHostRules;
+
+impl BroadcastHostRules for UnsupportedHostRules {
+    fn name(&self) -> &str {
+        "unsupported"
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            let reason = "host rules are not implemented on this platform; route \
+                          255.255.255.255 through the overlay interface and allow inbound \
+                          UDP from the overlay range yourself"
+                .to_string();
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route: RuleOutcome::Failed(reason.clone()),
+                firewall: RuleOutcome::Failed(reason),
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, _interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {})
+    }
+}
+
+/// What a [`MockHostRules`] currently holds, per interface.
+pub type MockRulesState = std::collections::BTreeMap<String, BroadcastRulesPlan>;
+
+/// An in-memory host, for tests: nothing is touched.
+#[derive(Debug, Clone, Default)]
+pub struct MockHostRules {
+    exit: MockExitRules,
+    state: Arc<Mutex<MockRulesState>>,
+    calls: Arc<Mutex<Vec<String>>>,
+    firewall_failure: Arc<Mutex<Option<String>>>,
+}
+
+impl MockHostRules {
+    /// An empty host.
+    pub fn new() -> Self {
+        Self {
+            exit: MockExitRules::new(),
+            ..Self::default()
+        }
+    }
+
+    /// The exit-node half of this host, for inspecting and steering tests.
+    pub fn exit(&self) -> &MockExitRules {
+        &self.exit
+    }
+
+    /// Makes the firewall step fail with this reason from now on.
+    pub fn fail_firewall(&self, reason: impl Into<String>) {
+        *lock(&self.firewall_failure) = Some(reason.into());
+    }
+
+    /// What is installed now.
+    pub fn installed(&self) -> MockRulesState {
+        lock(&self.state).clone()
+    }
+
+    /// Every `apply:<interface>` / `clear:<interface>` call, in order.
+    pub fn calls(&self) -> Vec<String> {
+        lock(&self.calls).clone()
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+impl BroadcastHostRules for MockHostRules {
+    fn name(&self) -> &str {
+        "mock"
+    }
+
+    fn exit_rules(&self) -> Option<&dyn ExitHostRules> {
+        Some(&self.exit)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            lock(&self.calls).push(format!("apply:{}", plan.interface));
+            lock(&self.state).insert(plan.interface.clone(), plan.clone());
+            let firewall = match lock(&self.firewall_failure).clone() {
+                Some(reason) => RuleOutcome::Failed(reason),
+                None => RuleOutcome::Applied,
+            };
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route: RuleOutcome::Applied,
+                firewall,
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            lock(&self.calls).push(format!("clear:{interface}"));
+            lock(&self.state).remove(interface);
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    fn plan() -> BroadcastRulesPlan {
+        BroadcastRulesPlan {
+            interface: "tsun0".into(),
+            source: Ipv4Addr::new(10, 13, 37, 142),
+            range: "10.13.37.0/24".parse().unwrap(),
+        }
+    }
+
+    #[test]
+    fn every_object_is_tagged_with_its_interface() {
+        assert_eq!(plan().tag(), "tsunagi:tsun0");
+        assert_eq!(tag_for("tsun7"), "tsunagi:tsun7");
+    }
+
+    #[test]
+    fn the_summary_says_which_half_is_missing() {
+        assert_eq!(
+            rules_summary(&RuleOutcome::Applied, &RuleOutcome::Applied),
+            "ok (route + firewall)"
+        );
+        let line = rules_summary(
+            &RuleOutcome::Applied,
+            &RuleOutcome::Failed("iptables not found".into()),
+        );
+        assert!(line.contains("route ok"), "{line}");
+        assert!(
+            line.contains("firewall failed: iptables not found"),
+            "{line}"
+        );
+        let report = BroadcastRulesReport {
+            plan: plan(),
+            route: RuleOutcome::Failed("denied".into()),
+            firewall: RuleOutcome::Applied,
+        };
+        assert!(!report.is_ok());
+        assert!(report.summary().contains("route failed: denied"));
+    }
+
+    #[tokio::test]
+    async fn the_mock_applies_replaces_and_clears_idempotently() {
+        let rules = MockHostRules::new();
+        let first = rules.apply(&plan()).await;
+        assert!(first.is_ok());
+        assert_eq!(rules.installed().len(), 1);
+
+        let mut moved = plan();
+        moved.source = Ipv4Addr::new(10, 13, 37, 7);
+        rules.apply(&moved).await;
+        assert_eq!(rules.installed()["tsun0"].source, moved.source);
+
+        rules.clear("tsun0").await;
+        rules.clear("tsun0").await;
+        assert!(rules.installed().is_empty());
+        assert_eq!(rules.calls().len(), 4);
+    }
+
+    #[tokio::test]
+    async fn an_unsupported_platform_reports_both_steps_failed() {
+        let report = UnsupportedHostRules.apply(&plan()).await;
+        assert!(!report.route.is_applied());
+        assert!(!report.firewall.is_applied());
+        UnsupportedHostRules.clear("tsun0").await;
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/windows.rs

@@ -0,0 +1,341 @@
+//! Broadcast host rules on Windows: a route and a Defender Firewall rule.
+//!
+//! Both are written with the `NetTCPIP` and `NetSecurity` PowerShell modules,
+//! run by absolute path from `%SystemRoot%` like the DNS publisher does: the
+//! in-process alternatives are `unsafe` IP Helper calls this crate forbids.
+//!
+//! * The route is `255.255.255.255/32` on the overlay interface, on-link, with
+//!   metric 1, written to the active store only so it cannot outlive a reboot.
+//!   Windows picks the interface's own address as the source of a packet sent
+//!   along it; with several networks on one interface that is whichever
+//!   address Windows treats as primary, which this cannot choose.
+//! * The firewall rule allows inbound UDP on that interface alias from the
+//!   overlay range only, in every profile. Wintun adapters are usually
+//!   classified as a public network, where inbound UDP is otherwise dropped.
+//!   Its display name is the tag `tsunagi:<interface>`, which is how it is
+//!   found again, replaced and removed.
+//!
+//! Every value in a script is one this agent produced: the interface name is
+//! derived from the network id and the range comes from signed state. They are
+//! still single-quoted and escaped, so nothing could be read as PowerShell
+//! rather than as data. Changing routes and firewall rules needs an elevated
+//! process, which is reported as its own cause.
+//!
+//! This module is compiled everywhere so its script building is tested
+//! everywhere. It could not be run against a real Windows host while it was
+//! written.
+
+#![cfg_attr(not(target_os = "windows"), allow(dead_code))]
+
+use crate::BoxFuture;
+
+use super::exit::{ExitHostPlan, ExitHostReport, ExitHostRules};
+use super::windows_exit;
+use super::{BroadcastHostRules, BroadcastRulesPlan, BroadcastRulesReport, RuleOutcome, tag_for};
+
+/// A PowerShell single-quoted literal, with any embedded quote doubled.
+fn ps_literal(value: &str) -> String {
+    format!("'{}'", value.replace('\'', "''"))
+}
+
+/// The step that makes the route match the plan.
+///
+/// A leftover route (from a crashed run, or for an old interface index) is
+/// removed first so the result is exactly one route.
+fn route_script(plan: &BroadcastRulesPlan) -> String {
+    let alias = ps_literal(&plan.interface);
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $alias = {alias}\n\
+         Get-NetRoute -DestinationPrefix '255.255.255.255/32' -InterfaceAlias $alias \
+         -PolicyStore ActiveStore -ErrorAction SilentlyContinue | \
+         Remove-NetRoute -Confirm:$false -ErrorAction SilentlyContinue\n\
+         New-NetRoute -DestinationPrefix '255.255.255.255/32' -InterfaceAlias $alias \
+         -NextHop '0.0.0.0' -RouteMetric 1 -PolicyStore ActiveStore | Out-Null\n"
+    )
+}
+
+/// The step that makes the firewall hold exactly one rule for the interface.
+fn firewall_script(plan: &BroadcastRulesPlan) -> String {
+    let alias = ps_literal(&plan.interface);
+    let tag = ps_literal(&plan.tag());
+    let range = ps_literal(&plan.range.to_string());
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $alias = {alias}\n\
+         $tag = {tag}\n\
+         Get-NetFirewallRule -DisplayName $tag -ErrorAction SilentlyContinue | \
+         Remove-NetFirewallRule -ErrorAction SilentlyContinue\n\
+         New-NetFirewallRule -DisplayName $tag -Direction Inbound -Action Allow -Protocol UDP \
+         -InterfaceAlias $alias -RemoteAddress {range} -Profile Any | Out-Null\n"
+    )
+}
+
+/// The script that removes both, ignoring what is not there.
+fn clear_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    let tag = ps_literal(&tag_for(interface));
+    format!(
+        "$ErrorActionPreference = 'SilentlyContinue'\n\
+         Get-NetRoute -DestinationPrefix '255.255.255.255/32' -InterfaceAlias {alias} \
+         -PolicyStore ActiveStore | Remove-NetRoute -Confirm:$false\n\
+         Get-NetFirewallRule -DisplayName {tag} | Remove-NetFirewallRule\n\
+         exit 0\n"
+    )
+}
+
+/// Turns what PowerShell said into a cause a person can act on.
+fn classify(what: &str, text: &str) -> String {
+    let lower = text.to_ascii_lowercase();
+    if lower.contains("access is denied")
+        || lower.contains("requires elevation")
+        || lower.contains("run as administrator")
+        || lower.contains("permissiondenied")
+    {
+        format!("{what}: Windows refused; run tsng from an elevated (administrator) process")
+    } else if lower.contains("is not recognized") || lower.contains("commandnotfound") {
+        format!("{what}: the NetTCPIP/NetSecurity PowerShell modules are not available: {text}")
+    } else if text.is_empty() {
+        format!("{what}: failed without saying why")
+    } else {
+        format!("{what}: {text}")
+    }
+}
+
+/// The absolute path to Windows PowerShell, so nothing on `PATH` can stand in
+/// for it.
+fn powershell_path() -> std::path::PathBuf {
+    let root = std::env::var_os("SystemRoot").unwrap_or_else(|| r"C:\Windows".into());
+    std::path::Path::new(&root)
+        .join("System32")
+        .join("WindowsPowerShell")
+        .join("v1.0")
+        .join("powershell.exe")
+}
+
+/// Runs a script and returns what it said on failure.
+fn run_powershell(script: &str) -> Result<(), String> {
+    run_powershell_output(script).map(|_| ())
+}
+
+/// Runs a script and returns what it printed, or what it said on failure.
+fn run_powershell_output(script: &str) -> Result<String, String> {
+    let output = std::process::Command::new(powershell_path())
+        .args([
+            "-NoProfile",
+            "-NonInteractive",
+            "-ExecutionPolicy",
+            "Bypass",
+            "-Command",
+            script,
+        ])
+        .output()
+        .map_err(|err| format!("could not run PowerShell: {err}"))?;
+    if output.status.success() {
+        return Ok(String::from_utf8_lossy(&output.stdout).into_owned());
+    }
+    let stderr = String::from_utf8_lossy(&output.stderr);
+    let text = if stderr.trim().is_empty() {
+        String::from_utf8_lossy(&output.stdout).trim().to_string()
+    } else {
+        stderr.trim().to_string()
+    };
+    Err(text)
+}
+
+async fn step(what: &'static str, script: String) -> RuleOutcome {
+    match tokio::task::spawn_blocking(move || run_powershell(&script)).await {
+        Ok(Ok(())) => RuleOutcome::Applied,
+        Ok(Err(text)) => RuleOutcome::Failed(classify(what, &text)),
+        Err(err) => RuleOutcome::Failed(format!("{what}: the task failed: {err}")),
+    }
+}
+
+/// The Windows implementation, through PowerShell.
+#[derive(Debug, Default)]
+pub struct WindowsHostRules;
+
+impl WindowsHostRules {
+    /// Creates it. Nothing is run until [`BroadcastHostRules::apply`].
+    pub fn new() -> Self {
+        Self
+    }
+}
+
+/// The IPv4 addresses a host name resolves to; empty when it does not.
+fn resolve_host(host: &str) -> Vec<std::net::Ipv4Addr> {
+    use std::net::{IpAddr, ToSocketAddrs};
+    (host, 443)
+        .to_socket_addrs()
+        .map(|addrs| {
+            addrs
+                .filter_map(|addr| match addr.ip() {
+                    IpAddr::V4(ip) => Some(ip),
+                    IpAddr::V6(_) => None,
+                })
+                .collect()
+        })
+        .unwrap_or_default()
+}
+
+impl ExitHostRules for WindowsHostRules {
+    fn needs_bypass(&self) -> bool {
+        true
+    }
+
+    fn apply<'a>(&'a self, plan: &'a ExitHostPlan) -> BoxFuture<'a, ExitHostReport> {
+        Box::pin(async move {
+            let job = plan.clone();
+            tokio::task::spawn_blocking(move || {
+                windows_exit::apply_plan(
+                    &|script| run_powershell_output(script),
+                    &resolve_host,
+                    &job,
+                )
+            })
+            .await
+            .unwrap_or_else(|err| {
+                let failed = RuleOutcome::Failed(format!("the task failed: {err}"));
+                ExitHostReport {
+                    offer: plan
+                        .offer
+                        .iter()
+                        .map(|range| (*range, failed.clone()))
+                        .collect(),
+                    forwarding: None,
+                    client: plan.client.then_some(failed),
+                }
+            })
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            let name = interface.to_string();
+            let _ = tokio::task::spawn_blocking(move || {
+                windows_exit::clear_all(&|script| run_powershell_output(script), &name)
+            })
+            .await;
+        })
+    }
+}
+
+impl BroadcastHostRules for WindowsHostRules {
+    fn name(&self) -> &str {
+        "windows-netsecurity"
+    }
+
+    fn exit_rules(&self) -> Option<&dyn ExitHostRules> {
+        Some(self)
+    }
+
+    fn apply<'a>(&'a self, plan: &'a BroadcastRulesPlan) -> BoxFuture<'a, BroadcastRulesReport> {
+        Box::pin(async move {
+            let route = step("route", route_script(plan)).await;
+            let firewall = step("firewall", firewall_script(plan)).await;
+            BroadcastRulesReport {
+                plan: plan.clone(),
+                route,
+                firewall,
+            }
+        })
+    }
+
+    fn clear<'a>(&'a self, interface: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            if let RuleOutcome::Failed(reason) = step("clear", clear_script(interface)).await {
+                tracing::warn!(
+                    interface,
+                    "cannot remove the broadcast host rules: {reason}"
+                );
+            }
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::net::Ipv4Addr;
+
+    fn plan() -> BroadcastRulesPlan {
+        BroadcastRulesPlan {
+            interface: "tsun0".into(),
+            source: Ipv4Addr::new(10, 13, 37, 142),
+            range: "10.13.37.0/24".parse().unwrap(),
+        }
+    }
+
+    #[test]
+    fn a_quote_in_a_value_cannot_escape_the_literal() {
+        assert_eq!(ps_literal("a'b"), "'a''b'");
+        let mut hostile = plan();
+        hostile.interface = "x'; Remove-Item C:\\ #".into();
+        let script = route_script(&hostile);
+        assert!(
+            script.contains("$alias = 'x''; Remove-Item C:\\ #'"),
+            "{script}"
+        );
+        assert!(firewall_script(&hostile).contains("'x''; Remove-Item C:\\ #'"));
+    }
+
+    #[test]
+    fn the_route_is_on_link_metric_one_in_the_active_store_and_replaces_its_own() {
+        let script = route_script(&plan());
+        assert!(script.contains("'255.255.255.255/32'"));
+        assert!(script.contains("-NextHop '0.0.0.0'"));
+        assert!(script.contains("-RouteMetric 1"));
+        assert!(script.contains("-PolicyStore ActiveStore"));
+        assert!(
+            script.find("Remove-NetRoute").unwrap() < script.find("New-NetRoute").unwrap(),
+            "a stale route goes before the new one:\n{script}"
+        );
+    }
+
+    #[test]
+    fn the_firewall_rule_is_udp_only_from_the_overlay_range_on_this_interface() {
+        let script = firewall_script(&plan());
+        assert!(script.contains("-Protocol UDP"));
+        assert!(script.contains("-Direction Inbound -Action Allow"));
+        assert!(script.contains("-InterfaceAlias $alias"));
+        assert!(script.contains("-RemoteAddress '10.13.37.0/24'"));
+        assert!(script.contains("$tag = 'tsunagi:tsun0'"));
+        assert!(!script.contains("-Protocol Any") && !script.contains("-Protocol TCP"));
+        assert!(
+            script.find("Remove-NetFirewallRule").unwrap()
+                < script.find("New-NetFirewallRule").unwrap(),
+            "{script}"
+        );
+    }
+
+    #[test]
+    fn clearing_touches_only_our_route_and_our_tagged_rule() {
+        let script = clear_script("tsun0");
+        assert!(script.contains("-InterfaceAlias 'tsun0'"));
+        assert!(script.contains("Get-NetFirewallRule -DisplayName 'tsunagi:tsun0'"));
+        assert!(script.contains("exit 0"));
+    }
+
+    #[test]
+    fn a_refusal_names_elevation_and_a_missing_module_names_the_module() {
+        assert!(classify("route", "Access is denied").contains("elevated"));
+        assert!(
+            classify(
+                "firewall",
+                "The term 'New-NetFirewallRule' is not recognized"
+            )
+            .contains("modules are not available")
+        );
+        assert_eq!(classify("route", "boom"), "route: boom");
+        assert!(classify("route", "").contains("without saying why"));
+    }
+
+    #[test]
+    fn powershell_is_found_by_absolute_path() {
+        let path = powershell_path();
+        assert!(path.ends_with("powershell.exe"));
+        assert!(path.to_string_lossy().contains("System32"), "{path:?}");
+    }
+}

crates/tsunagi-core/src/overlay/hostrules/windows_exit.rs

@@ -0,0 +1,603 @@
+//! Exit-node host rules on Windows: WinNAT through PowerShell.
+//!
+//! Compiled on every platform so the scripts and the plan logic are tested
+//! everywhere; only `WindowsHostRules` runs them, and only on Windows. It
+//! could not be run against a real Windows host while it was written.
+//!
+//! # Offering
+//!
+//! One NAT object per offered range, named with the usual tag
+//! `tsunagi-exit:<if>:<range>`:
+//!
+//! ```text
+//! New-NetNat -Name <tag> -InternalIPInterfaceAddressPrefix <range>
+//! ```
+//!
+//! WinNAT translates that range out of whichever interface the traffic leaves
+//! by, which is the Windows counterpart of the Linux masquerade. Windows
+//! forwards per interface, and the egress is not known here, so forwarding is
+//! only *read* for the overlay interface; turning it on for the interfaces
+//! involved is the owner's decision (`Set-NetIPInterface -Forwarding Enabled`).
+//! Some Windows versions allow a single NAT per host: a second offered range
+//! then reports its own failure.
+//!
+//! # Using
+//!
+//! Windows has no per-user or per-process routing, so the Linux scheme (an
+//! exemption for the agent's user) does not carry over. Instead the agent's
+//! own underlay traffic is exempted by *destination*:
+//!
+//! ```text
+//! New-NetRoute 0.0.0.0/1   on the overlay interface, metric 1
+//! New-NetRoute 128.0.0.0/1 on the overlay interface, metric 1
+//! New-NetRoute <ip>/32 via the physical default gateway, metric 7373
+//! ```
+//!
+//! The halves out-rank the default route without replacing it, so every more
+//! specific route (LAN, another VPN) keeps working. A `/32` per address the
+//! agent talks to directly (the verified direct paths of its connections, and
+//! the relays it uses, resolved when the rules are applied) keeps that traffic
+//! on the physical network instead of looping into the tunnel it carries. The
+//! metric 7373 marks those routes as ours, which is how stale ones are found
+//! and removed. Everything is written to the active store, so a reboot clears
+//! it. IPv6 is blackholed with `::/1` and `8000::/1` on the overlay interface,
+//! which has no IPv6: applications fall back to IPv4 after a delay, where
+//! Linux and macOS answer at once. The bypass follows the agent's paths as
+//! they change; a path that appears after the last apply is reached through
+//! the tunnel until the next one.
+//!
+//! The bypass routes are written before the halves, and removed after them.
+
+use super::RuleOutcome;
+use std::net::Ipv4Addr;
+
+use super::exit::{ExitHostPlan, ExitHostReport, exit_tag, exit_tag_prefix};
+use crate::state::Ipv4Range;
+
+/// The route metric that marks a bypass route as this agent's.
+const BYPASS_METRIC: u32 = 7373;
+
+/// Runs a PowerShell script and returns what it printed.
+pub(super) type RunScript<'a> = &'a dyn Fn(&str) -> Result<String, String>;
+
+/// Resolves a host name to its IPv4 addresses.
+pub(super) type Resolve<'a> = &'a dyn Fn(&str) -> Vec<Ipv4Addr>;
+
+/// A PowerShell single-quoted literal, with any embedded quote doubled.
+fn ps_literal(value: &str) -> String {
+    format!("'{}'", value.replace('\'', "''"))
+}
+
+/// Creates the NAT of one range, replacing one this agent made before.
+fn nat_script(interface: &str, range: &Ipv4Range) -> String {
+    let name = ps_literal(&exit_tag(interface, range));
+    let prefix = ps_literal(&range.to_string());
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $name = {name}\n\
+         Get-NetNat -Name $name -ErrorAction SilentlyContinue | \
+         Remove-NetNat -Confirm:$false -ErrorAction SilentlyContinue\n\
+         New-NetNat -Name $name -InternalIPInterfaceAddressPrefix {prefix} | Out-Null\n"
+    )
+}
+
+/// Removes this interface's NAT objects except those named in `keep`.
+fn prune_script(interface: &str, keep: &[Ipv4Range]) -> String {
+    let prefix = ps_literal(&exit_tag_prefix(interface));
+    let keep: Vec<String> = keep
+        .iter()
+        .map(|range| ps_literal(&exit_tag(interface, range)))
+        .collect();
+    format!(
+        "$ErrorActionPreference = 'SilentlyContinue'\n\
+         $prefix = {prefix}\n\
+         $keep = @({})\n\
+         Get-NetNat | Where-Object {{ $_.Name.StartsWith($prefix) -and \
+         ($keep -notcontains $_.Name) }} | Remove-NetNat -Confirm:$false\n\
+         exit 0\n",
+        keep.join(", ")
+    )
+}
+
+/// Prints `Enabled` or `Disabled` for the interface's IPv4 forwarding.
+fn forwarding_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         (Get-NetIPInterface -InterfaceAlias {alias} -AddressFamily IPv4).Forwarding\n"
+    )
+}
+
+/// Prints `<interface index> <next hop>` of the best default route that is
+/// not through the overlay interface.
+fn gateway_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $alias = {alias}\n\
+         $best = Get-NetRoute -DestinationPrefix '0.0.0.0/0' -AddressFamily IPv4 \
+         -PolicyStore ActiveStore | Where-Object {{ $_.InterfaceAlias -ne $alias -and \
+         $_.NextHop -ne '0.0.0.0' }} | Sort-Object {{ $_.RouteMetric + \
+         (Get-NetIPInterface -InterfaceIndex $_.InterfaceIndex -AddressFamily IPv4).InterfaceMetric }} | \
+         Select-Object -First 1\n\
+         if (-not $best) {{ throw 'the host has no default gateway outside the overlay' }}\n\
+         \"$($best.InterfaceIndex) $($best.NextHop)\"\n"
+    )
+}
+
+/// The default gateway as the script above printed it.
+fn parse_gateway(text: &str) -> Option<(u32, Ipv4Addr)> {
+    let mut words = text.split_whitespace();
+    let index = words.next()?.parse().ok()?;
+    let hop = words.next()?.parse().ok()?;
+    words.next().is_none().then_some((index, hop))
+}
+
+/// Replaces the bypass routes with one per address, through the gateway.
+fn bypass_script(index: u32, gateway: Ipv4Addr, addresses: &[Ipv4Addr]) -> String {
+    let list: Vec<String> = addresses
+        .iter()
+        .map(|address| ps_literal(&address.to_string()))
+        .collect();
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         Get-NetRoute -AddressFamily IPv4 -PolicyStore ActiveStore -ErrorAction SilentlyContinue | \
+         Where-Object {{ $_.RouteMetric -eq {BYPASS_METRIC} -and $_.DestinationPrefix.EndsWith('/32') }} | \
+         Remove-NetRoute -Confirm:$false -ErrorAction SilentlyContinue\n\
+         foreach ($ip in @({})) {{\n\
+         New-NetRoute -DestinationPrefix \"$ip/32\" -InterfaceIndex {index} -NextHop '{gateway}' \
+         -RouteMetric {BYPASS_METRIC} -PolicyStore ActiveStore -ErrorAction SilentlyContinue | Out-Null\n\
+         }}\n",
+        list.join(", ")
+    )
+}
+
+/// Sends everything else through the overlay interface.
+fn halves_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    format!(
+        "$ErrorActionPreference = 'Stop'\n\
+         $alias = {alias}\n\
+         foreach ($p in '0.0.0.0/1', '128.0.0.0/1') {{\n\
+         Get-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -PolicyStore ActiveStore \
+         -ErrorAction SilentlyContinue | Remove-NetRoute -Confirm:$false -ErrorAction SilentlyContinue\n\
+         New-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -NextHop '0.0.0.0' \
+         -RouteMetric 1 -PolicyStore ActiveStore | Out-Null\n\
+         }}\n\
+         foreach ($p in '::/1', '8000::/1') {{\n\
+         Get-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -PolicyStore ActiveStore \
+         -ErrorAction SilentlyContinue | Remove-NetRoute -Confirm:$false -ErrorAction SilentlyContinue\n\
+         New-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -NextHop '::' \
+         -RouteMetric 1 -PolicyStore ActiveStore -ErrorAction SilentlyContinue | Out-Null\n\
+         }}\n"
+    )
+}
+
+/// Removes the halves and every bypass route. What is not there is the state
+/// wanted.
+fn remove_client_script(interface: &str) -> String {
+    let alias = ps_literal(interface);
+    format!(
+        "$ErrorActionPreference = 'SilentlyContinue'\n\
+         $alias = {alias}\n\
+         foreach ($p in '0.0.0.0/1', '128.0.0.0/1', '::/1', '8000::/1') {{\n\
+         Get-NetRoute -DestinationPrefix $p -InterfaceAlias $alias -PolicyStore ActiveStore | \
+         Remove-NetRoute -Confirm:$false\n\
+         }}\n\
+         Get-NetRoute -AddressFamily IPv4 -PolicyStore ActiveStore | \
+         Where-Object {{ $_.RouteMetric -eq {BYPASS_METRIC} -and $_.DestinationPrefix.EndsWith('/32') }} | \
+         Remove-NetRoute -Confirm:$false\n\
+         exit 0\n"
+    )
+}
+
+/// Every address to keep direct: the ones given and what the names resolve
+/// to, sorted and without duplicates.
+fn bypass_addresses(
+    resolve: Resolve<'_>,
+    addresses: &[Ipv4Addr],
+    hosts: &[String],
+) -> Vec<Ipv4Addr> {
+    let mut all: std::collections::BTreeSet<Ipv4Addr> = addresses.iter().copied().collect();
+    for host in hosts {
+        all.extend(resolve(host));
+    }
+    all.into_iter()
+        .filter(|a| !a.is_loopback() && !a.is_unspecified() && !a.is_multicast())
+        .collect()
+}
+
+fn apply_client(
+    run: RunScript<'_>,
+    resolve: Resolve<'_>,
+    plan: &ExitHostPlan,
+) -> Result<(), String> {
+    let interface = plan.interface.as_str();
+    let gateway = run(&gateway_script(interface))?;
+    let (index, hop) = parse_gateway(&gateway)
+        .ok_or_else(|| format!("cannot read the default gateway from `{}`", gateway.trim()))?;
+    let addresses = bypass_addresses(resolve, &plan.bypass, &plan.bypass_hosts);
+    run(&bypass_script(index, hop, &addresses))?;
+    if let Err(why) = run(&halves_script(interface)) {
+        // Do not leave a half-installed set: bypass routes alone are harmless
+        // but the next apply starts clean.
+        let _ = run(&remove_client_script(interface));
+        return Err(why);
+    }
+    Ok(())
+}
+
+fn parse_forwarding(text: &str) -> Option<bool> {
+    match text.trim().to_ascii_lowercase().as_str() {
+        "enabled" | "1" => Some(true),
+        "disabled" | "0" => Some(false),
+        _ => None,
+    }
+}
+
+/// Makes the host match the plan.
+pub(super) fn apply_plan(
+    run: RunScript<'_>,
+    resolve: Resolve<'_>,
+    plan: &ExitHostPlan,
+) -> ExitHostReport {
+    let interface = plan.interface.as_str();
+    // What an earlier plan left and this one no longer wants.
+    let pruned = run(&prune_script(interface, &plan.offer));
+    let offer = plan
+        .offer
+        .iter()
+        .map(|range| {
+            let outcome = match &pruned {
+                Err(why) => RuleOutcome::Failed(why.clone()),
+                Ok(_) => match run(&nat_script(interface, range)) {
+                    Ok(_) => RuleOutcome::Applied,
+                    Err(why) => RuleOutcome::Failed(why),
+                },
+            };
+            (*range, outcome)
+        })
+        .collect();
+    let client = if plan.client {
+        Some(match apply_client(run, resolve, plan) {
+            Ok(()) => RuleOutcome::Applied,
+            Err(why) => RuleOutcome::Failed(why),
+        })
+    } else {
+        // Whatever an earlier plan left must not stay in force.
+        let _ = run(&remove_client_script(interface));
+        None
+    };
+    let forwarding = if plan.offer.is_empty() {
+        None
+    } else {
+        run(&forwarding_script(interface))
+            .ok()
+            .and_then(|text| parse_forwarding(&text))
+    };
+    ExitHostReport {
+        offer,
+        forwarding,
+        client,
+    }
+}
+
+/// Removes every NAT object of this agent's for the interface.
+pub(super) fn clear_all(run: RunScript<'_>, interface: &str) {
+    if let Err(why) = run(&prune_script(interface, &[])) {
+        tracing::warn!(interface, "cannot remove the exit node NAT: {why}");
+    }
+    if let Err(why) = run(&remove_client_script(interface)) {
+        tracing::warn!(interface, "cannot remove the exit node routes: {why}");
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use std::cell::RefCell;
+
+    use super::*;
+
+    fn range() -> Ipv4Range {
+        "10.13.37.0/24".parse().unwrap()
+    }
+
+    fn plan(client: bool) -> ExitHostPlan {
+        ExitHostPlan {
+            interface: "tsun0".into(),
+            offer: vec![range()],
+            client,
+            overlay: vec![range()],
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn the_nat_is_named_with_the_tag_and_replaces_its_own() {
+        let script = nat_script("tsun0", &range());
+        assert!(
+            script.contains("$name = 'tsunagi-exit:tsun0:10.13.37.0/24'"),
+            "{script}"
+        );
+        assert!(script.contains("-InternalIPInterfaceAddressPrefix '10.13.37.0/24'"));
+        assert!(
+            script.find("Remove-NetNat").unwrap() < script.find("New-NetNat").unwrap(),
+            "{script}"
+        );
+    }
+
+    #[test]
+    fn a_quote_in_a_value_cannot_escape_the_literal() {
+        let script = nat_script("x'; Remove-Item C:\\ #", &range());
+        assert!(script.contains("'tsunagi-exit:x''; Remove-Item C:\\ #:10.13.37.0/24'"));
+        assert!(prune_script("x'y", &[]).contains("'tsunagi-exit:x''y:'"));
+        assert!(forwarding_script("x'y").contains("'x''y'"));
+    }
+
+    #[test]
+    fn pruning_keeps_wanted_ranges_and_only_selects_this_interfaces_nats() {
+        let script = prune_script("tsun0", &[range()]);
+        assert!(script.contains("$prefix = 'tsunagi-exit:tsun0:'"));
+        assert!(script.contains("$keep = @('tsunagi-exit:tsun0:10.13.37.0/24')"));
+        assert!(script.contains("StartsWith($prefix)"));
+        assert!(script.contains("exit 0"));
+        assert!(prune_script("tsun0", &[]).contains("$keep = @()"));
+    }
+
+    #[test]
+    fn forwarding_text_is_read_never_written() {
+        assert_eq!(parse_forwarding("Enabled\r\n"), Some(true));
+        assert_eq!(parse_forwarding("Disabled"), Some(false));
+        assert_eq!(parse_forwarding("what"), None);
+        let script = forwarding_script("tsun0");
+        assert!(script.contains("Get-NetIPInterface"));
+        assert!(!script.contains("Set-NetIPInterface"));
+    }
+
+    #[test]
+    fn offering_prunes_then_creates_and_reads_forwarding() {
+        let calls = RefCell::new(Vec::new());
+        let run = |script: &str| -> Result<String, String> {
+            calls.borrow_mut().push(script.to_string());
+            Ok(if script.contains("Get-NetIPInterface") {
+                "Disabled\n".to_string()
+            } else {
+                String::new()
+            })
+        };
+        let report = apply_plan(&run, &no_names, &plan(false));
+        assert_eq!(report.offer, vec![(range(), RuleOutcome::Applied)]);
+        assert_eq!(report.forwarding, Some(false));
+        assert_eq!(report.client, None);
+        let calls = calls.borrow();
+        assert!(calls[0].contains("StartsWith"), "{calls:?}");
+        assert!(calls[1].contains("New-NetNat"), "{calls:?}");
+        // Without a client, the routes of an earlier plan are taken away.
+        assert!(calls[2].contains("RouteMetric -eq 7373"), "{calls:?}");
+        assert!(calls[3].contains("Get-NetIPInterface"), "{calls:?}");
+    }
+
+    #[test]
+    fn a_failing_nat_is_that_ranges_failure() {
+        let run = |script: &str| -> Result<String, String> {
+            if script.contains("New-NetNat") {
+                Err("A NAT with this prefix already exists".into())
+            } else {
+                Ok(String::new())
+            }
+        };
+        let report = apply_plan(&run, &no_names, &plan(false));
+        let RuleOutcome::Failed(why) = &report.offer[0].1 else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("already exists"), "{why}");
+    }
+
+    fn no_names(_: &str) -> Vec<Ipv4Addr> {
+        Vec::new()
+    }
+
+    fn client_plan() -> ExitHostPlan {
+        ExitHostPlan {
+            interface: "tsun0".into(),
+            offer: Vec::new(),
+            client: true,
+            overlay: vec![range()],
+            bypass: vec![Ipv4Addr::new(138, 201, 61, 182)],
+            bypass_hosts: vec!["relay.example".into()],
+        }
+    }
+
+    #[test]
+    fn the_gateway_is_read_from_the_script_output() {
+        assert_eq!(
+            parse_gateway("12 192.168.1.1\r\n"),
+            Some((12, Ipv4Addr::new(192, 168, 1, 1)))
+        );
+        assert_eq!(parse_gateway(""), None);
+        assert_eq!(parse_gateway("12"), None);
+        assert_eq!(parse_gateway("12 gw"), None);
+        assert_eq!(parse_gateway("12 1.1.1.1 extra"), None);
+        let script = gateway_script("tsun0");
+        assert!(script.contains("-ne $alias"));
+        assert!(script.contains("-ne '0.0.0.0'"));
+    }
+
+    #[test]
+    fn bypass_routes_are_marked_host_routes_through_the_physical_gateway() {
+        let script = bypass_script(
+            12,
+            Ipv4Addr::new(192, 168, 1, 1),
+            &[Ipv4Addr::new(1, 2, 3, 4), Ipv4Addr::new(5, 6, 7, 8)],
+        );
+        assert!(script.contains("@('1.2.3.4', '5.6.7.8')"), "{script}");
+        assert!(script.contains("-InterfaceIndex 12 -NextHop '192.168.1.1'"));
+        assert!(script.contains("-RouteMetric 7373"));
+        assert!(script.contains("-PolicyStore ActiveStore"));
+        // Stale marked routes go first, and only /32 ones with our metric.
+        assert!(script.find("Remove-NetRoute").unwrap() < script.find("New-NetRoute").unwrap());
+        assert!(script.contains("RouteMetric -eq 7373"));
+        assert!(script.contains("EndsWith('/32')"));
+    }
+
+    #[test]
+    fn the_halves_go_through_the_overlay_and_ipv6_is_blackholed() {
+        let script = halves_script("tsun0");
+        assert!(script.contains("'0.0.0.0/1', '128.0.0.0/1'"));
+        assert!(script.contains("-InterfaceAlias $alias -NextHop '0.0.0.0'"));
+        assert!(script.contains("'::/1', '8000::/1'"));
+        assert!(script.contains("-PolicyStore ActiveStore"));
+        assert!(
+            !script.contains("0.0.0.0/0"),
+            "the default route is left alone"
+        );
+        assert!(halves_script("x'y").contains("'x''y'"));
+    }
+
+    #[test]
+    fn addresses_are_resolved_merged_and_cleaned() {
+        let resolve = |host: &str| -> Vec<Ipv4Addr> {
+            match host {
+                "relay.example" => vec![Ipv4Addr::new(9, 9, 9, 9), Ipv4Addr::new(1, 2, 3, 4)],
+                _ => Vec::new(),
+            }
+        };
+        let all = bypass_addresses(
+            &resolve,
+            &[
+                Ipv4Addr::new(1, 2, 3, 4),
+                Ipv4Addr::LOCALHOST,
+                Ipv4Addr::UNSPECIFIED,
+            ],
+            &["relay.example".into(), "gone.example".into()],
+        );
+        assert_eq!(
+            all,
+            vec![Ipv4Addr::new(1, 2, 3, 4), Ipv4Addr::new(9, 9, 9, 9)]
+        );
+    }
+
+    #[test]
+    fn the_bypass_is_written_before_the_halves_that_need_it() {
+        let calls = RefCell::new(Vec::new());
+        let run = |script: &str| -> Result<String, String> {
+            calls.borrow_mut().push(script.to_string());
+            Ok(if script.contains("Select-Object -First 1") {
+                "12 192.168.1.1\n".to_string()
+            } else {
+                String::new()
+            })
+        };
+        let resolve = |_: &str| vec![Ipv4Addr::new(9, 9, 9, 9)];
+        let report = apply_plan(&run, &resolve, &client_plan());
+        assert_eq!(report.client, Some(RuleOutcome::Applied));
+        let calls = calls.borrow();
+        let at = |needle: &str| calls.iter().position(|c| c.contains(needle)).unwrap();
+        assert!(at("Select-Object -First 1") < at("-RouteMetric 7373"));
+        assert!(at("-RouteMetric 7373") < at("'0.0.0.0/1', '128.0.0.0/1'"));
+        let bypass = &calls[at("-RouteMetric 7373")];
+        assert!(
+            bypass.contains("'9.9.9.9'") && bypass.contains("'138.201.61.182'"),
+            "{bypass}"
+        );
+    }
+
+    #[test]
+    fn nothing_goes_through_the_overlay_when_the_bypass_cannot_be_written() {
+        let run = |script: &str| -> Result<String, String> {
+            if script.contains("Select-Object -First 1") {
+                Ok("12 192.168.1.1".to_string())
+            } else if script.contains("-RouteMetric 7373") && script.contains("foreach ($ip") {
+                Err("Access is denied".to_string())
+            } else {
+                assert!(!script.contains("-NextHop '0.0.0.0'"), "{script}");
+                Ok(String::new())
+            }
+        };
+        let report = apply_plan(&run, &no_names, &client_plan());
+        assert_eq!(
+            report.client,
+            Some(RuleOutcome::Failed("Access is denied".into()))
+        );
+    }
+
+    #[test]
+    fn a_missing_gateway_is_a_failure_that_installs_no_route() {
+        let run = |script: &str| -> Result<String, String> {
+            assert!(!script.contains("New-NetRoute"), "{script}");
+            if script.contains("Select-Object -First 1") {
+                Err("the host has no default gateway outside the overlay".into())
+            } else {
+                Ok(String::new())
+            }
+        };
+        let report = apply_plan(&run, &no_names, &client_plan());
+        let Some(RuleOutcome::Failed(why)) = report.client else {
+            panic!("{report:?}");
+        };
+        assert!(why.contains("no default gateway"), "{why}");
+    }
+
+    #[test]
+    fn a_failed_halves_step_removes_what_was_written() {
+        let calls = RefCell::new(Vec::new());
+        let run = |script: &str| -> Result<String, String> {
+            calls.borrow_mut().push(script.to_string());
+            if script.contains("Select-Object -First 1") {
+                Ok("12 192.168.1.1".to_string())
+            } else if script.contains("'128.0.0.0/1'") && script.contains("New-NetRoute") {
+                Err("boom".to_string())
+            } else {
+                Ok(String::new())
+            }
+        };
+        let report = apply_plan(&run, &no_names, &client_plan());
+        assert!(matches!(report.client, Some(RuleOutcome::Failed(_))));
+        assert!(calls.borrow().last().unwrap().contains("Remove-NetRoute"));
+    }
+
+    #[test]
+    fn dropping_the_client_removes_its_routes_and_clearing_does_too() {
+        let calls = RefCell::new(Vec::new());
+        let run = |script: &str| -> Result<String, String> {
+            calls.borrow_mut().push(script.to_string());
+            Ok(String::new())
+        };
+        let report = apply_plan(&run, &no_names, &plan(false));
+        assert_eq!(report.client, None);
+        assert!(
+            calls
+                .borrow()
+                .iter()
+                .any(|c| c.contains("RouteMetric -eq 7373"))
+        );
+        let script = remove_client_script("tsun0");
+        assert!(script.contains("'0.0.0.0/1', '128.0.0.0/1', '::/1', '8000::/1'"));
+        assert!(script.contains("exit 0"));
+        // Only routes on our interface, or marked ours, are touched.
+        assert!(script.contains("-InterfaceAlias $alias"));
+        calls.borrow_mut().clear();
+        clear_all(&run, "tsun0");
+        assert!(
+            calls
+                .borrow()
+                .iter()
+                .any(|c| c.contains("RouteMetric -eq 7373"))
+        );
+    }
+
+    #[test]
+    fn clearing_prunes_everything_for_the_interface() {
+        let seen = RefCell::new(Vec::<String>::new());
+        let run = |script: &str| -> Result<String, String> {
+            seen.borrow_mut().push(script.to_string());
+            Ok(String::new())
+        };
+        clear_all(&run, "tsun0");
+        let seen = seen.borrow();
+        assert!(seen[0].contains("$keep = @()"));
+        assert!(seen[0].contains("'tsunagi-exit:tsun0:'"));
+    }
+}

crates/tsunagi-core/src/overlay/mod.rs

@@ -0,0 +1,80 @@
+//! The overlay interface: the one network interface an agent owns.
+//!
+//! One agent, one interface. It belongs to the system level rather than to
+//! any protocol, and that is the whole point: a packet leaving it is routed
+//! to whichever peer owns its destination address, over whichever protocol
+//! currently has a link to that peer. Several protocols can be live at once
+//! without arguing over who holds the address, because neither of them
+//! holds it — the agent does.
+//!
+//! What lives here:
+//!
+//! * [`provision`] creates the interface and configures it, and removes it
+//!   again. Platform mechanics behind one decision function.
+//! * [`hostrules`] installs the route and firewall allowance that let a
+//!   game's LAN discovery reach the interface, while broadcast is on.
+//! * [`tun`] is the packet interface itself, real or in memory.
+//! * [`packet`] reads just enough of an IP header to route by it.
+//! * [`config`] names interfaces and describes addresses.
+//!
+//! What does not live here: encryption, peer discovery, and how a packet
+//! actually reaches another machine. Those belong to a protocol plugin,
+//! which this level hands packets to and takes packets from.
+
+/// Why the overlay interface cannot be brought into the state asked for.
+///
+/// Its own error rather than the plugin one it used to borrow: this level
+/// owns the interface now, and a plugin failing is a different event from
+/// the interface failing.
+#[derive(Debug, thiserror::Error)]
+#[non_exhaustive]
+pub enum OverlayError {
+    /// The interface cannot be created or configured right now, with the
+    /// reason and, where there is one, what to do about it.
+    #[error("overlay unavailable: {0}")]
+    Unavailable(String),
+    /// Anything else.
+    #[error("{0}")]
+    Other(String),
+}
+
+pub mod broadcast;
+pub mod config;
+pub mod exit;
+pub mod hostrules;
+pub mod interface;
+pub mod packet;
+pub mod provision;
+pub mod router;
+pub mod tun;
+
+pub use config::{Cidr, DEFAULT_INTERFACE_PREFIX, MAX_INTERFACE_NAME_LEN, interface_name};
+pub use exit::{ExitRulesReport, RuleSetReport};
+pub use hostrules::{
+    BroadcastHostRules, BroadcastRulesPlan, BroadcastRulesReport, ExitHostPlan, ExitHostReport,
+    ExitHostRules, MockExitRules, MockHostRules, RuleOutcome, UnsupportedHostRules,
+};
+pub use interface::{Counters, Interface, PacketCarrier, Rejected};
+pub use packet::IpHeader;
+pub use provision::{
+    Changes, InterfacePlan, InterfaceProvisioner, InterfaceState, LinkKind, ManagedTunFactory,
+    MockHost, MockProvisioner, Privilege, Provisioned, UnsupportedProvisioner, plan_changes,
+    probe_net_admin,
+};
+pub use router::{ExitPolicy, NetworkRoutes, Route, RouteError, RoutingTable};
+pub use tun::{MemoryTun, MemoryTunFactory, TunDevice, TunFactory, TunRequest, address_is_local};
+
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub use hostrules::LinuxHostRules;
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use hostrules::MacosHostRules;
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use hostrules::WindowsHostRules;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub use provision::NetlinkProvisioner;
+
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use provision::UtunProvisioner;
+
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use provision::WintunProvisioner;

crates/tsunagi-core/src/overlay/provision/factory.rs

@@ -9,22 +9,21 @@
 use std::sync::Arc;
 
 use crate::BoxFuture;
-use crate::dataplane::PluginError;
+use crate::overlay::OverlayError;
 
-use super::super::config::Cidr;
+use super::super::hostrules::BroadcastHostRules;
 use super::super::tun::{TunDevice, TunFactory, TunRequest};
 use super::{InterfacePlan, InterfaceProvisioner};
 
 /// Turns a [`TunRequest`] into the plan for a host interface.
-fn plan_for(request: &TunRequest) -> Result<InterfacePlan, PluginError> {
+fn plan_for(request: &TunRequest) -> Result<InterfacePlan, OverlayError> {
-    let mut addresses = vec![Cidr::new(request.address.into(), request.prefix_len)?];
+    // Every address, because the plan is exhaustive: the provisioner adds
-    if let Some(address) = request.address_v4 {
+    // what is missing and removes what is not in it. Passing one network's
-        addresses.push(Cidr::new(address.into(), request.prefix_len_v4)?);
+    // address would take every other network's off the host.
-    }
     Ok(InterfacePlan::new(
         request.name.clone(),
         request.mtu,
-        addresses,
+        request.addresses.clone(),
     ))
 }
 
@@ -32,12 +31,23 @@ fn plan_for(request: &TunRequest) -> Result<InterfacePlan, PluginError> {
 #[derive(Debug)]
 pub struct ManagedTunFactory {
     provisioner: Arc<dyn InterfaceProvisioner>,
+    host_rules: Option<Arc<dyn BroadcastHostRules>>,
 }
 
 impl ManagedTunFactory {
     /// Wraps a provisioner.
     pub fn new(provisioner: Arc<dyn InterfaceProvisioner>) -> Self {
-        Self { provisioner }
+        Self {
+            provisioner,
+            host_rules: None,
+        }
+    }
+
+    /// Lets the agent install the broadcast route and firewall allowance for
+    /// the interface this factory manages.
+    pub fn with_host_rules(mut self, rules: Arc<dyn BroadcastHostRules>) -> Self {
+        self.host_rules = Some(rules);
+        self
     }
 
     /// The provisioner underneath.
@@ -51,10 +61,19 @@ impl TunFactory for ManagedTunFactory {
         self.provisioner.name()
     }
 
+    /// It creates the interface on the host and holds it open.
+    fn on_host(&self) -> bool {
+        true
+    }
+
+    fn host_rules(&self) -> Option<Arc<dyn BroadcastHostRules>> {
+        self.host_rules.clone()
+    }
+
     fn create<'a>(
         &'a self,
         request: TunRequest,
-    ) -> BoxFuture<'a, Result<Arc<dyn TunDevice>, PluginError>> {
+    ) -> BoxFuture<'a, Result<Arc<dyn TunDevice>, OverlayError>> {
         Box::pin(async move {
             let plan = plan_for(&request)?;
             let provisioned = self.provisioner.reconcile(&plan).await?;
@@ -67,7 +86,7 @@ impl TunFactory for ManagedTunFactory {
                 // Reaching here would mean the interface already existed and
                 // was held open by us, which cannot be true on the path that
                 // creates a device.
-                PluginError::Other(format!(
+                OverlayError::Other(format!(
                     "interface `{}` was reconciled but no device came back",
                     plan.name
                 ))
@@ -75,7 +94,7 @@ impl TunFactory for ManagedTunFactory {
         })
     }
 
-    fn reconfigure<'a>(&'a self, request: TunRequest) -> BoxFuture<'a, Result<(), PluginError>> {
+    fn reconfigure<'a>(&'a self, request: TunRequest) -> BoxFuture<'a, Result<(), OverlayError>> {
         Box::pin(async move {
             let plan = plan_for(&request)?;
             let provisioned = self.provisioner.reconcile(&plan).await?;
@@ -106,18 +125,23 @@ impl TunFactory for ManagedTunFactory {
 mod tests {
     #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-    use std::net::{Ipv4Addr, Ipv6Addr};
+    use std::net::Ipv4Addr;
 
+    use super::super::super::config::Cidr;
     use super::super::{LinkKind, MockHost, MockProvisioner};
     use super::*;
 
     fn request(v4: Option<Ipv4Addr>) -> TunRequest {
+        addressed(v4.into_iter().collect())
+    }
+
+    fn addressed(v4: Vec<Ipv4Addr>) -> TunRequest {
         TunRequest {
             name: "tsunfactory".into(),
-            address: Ipv6Addr::new(0xfd00, 0, 0, 0, 0, 0, 0, 1),
+            addresses: v4
-            prefix_len: 64,
+                .into_iter()
-            address_v4: v4,
+                .map(|address| Cidr::new(address.into(), 24).unwrap())
-            prefix_len_v4: 24,
+                .collect(),
             mtu: 1280,
         }
     }
@@ -137,7 +161,7 @@ mod tests {
         let state = provisioner.host().get("tsunfactory").unwrap();
         assert_eq!(state.kind, LinkKind::Tun);
         assert_eq!(state.mtu, 1280);
-        assert_eq!(state.addresses.len(), 2, "both families are assigned");
+        assert_eq!(state.addresses.len(), 1, "the overlay address is assigned");
     }
 
     #[tokio::test]
@@ -198,4 +222,50 @@ mod tests {
         let err = factory.create(request(None)).await.unwrap_err();
         assert!(err.to_string().contains("bridge"), "{err}");
     }
+
+    #[tokio::test]
+    async fn every_network_s_address_reaches_the_host_not_just_the_first() {
+        // One agent has one interface and a network apiece on it. Carrying
+        // only the first address is carrying only the first network: the
+        // rest have addresses the operating system has never heard of, and
+        // their traffic goes nowhere while the status says all is well.
+        let provisioner = Arc::new(MockProvisioner::default());
+        let factory = ManagedTunFactory::new(provisioner.clone());
+        factory
+            .create(addressed(vec![
+                Ipv4Addr::new(10, 13, 37, 69),
+                Ipv4Addr::new(10, 156, 200, 116),
+            ]))
+            .await
+            .unwrap();
+
+        let state = provisioner.host().get("tsunfactory").unwrap();
+        let addresses: Vec<String> = state
+            .addresses
+            .iter()
+            .map(|entry| entry.to_string())
+            .collect();
+        assert!(
+            addresses.contains(&"10.13.37.69/24".to_string()),
+            "{addresses:?}"
+        );
+        assert!(
+            addresses.contains(&"10.156.200.116/24".to_string()),
+            "{addresses:?}"
+        );
+
+        // And leaving one network takes its address off, because the plan
+        // is what the interface should carry and nothing else.
+        factory
+            .reconfigure(addressed(vec![Ipv4Addr::new(10, 13, 37, 69)]))
+            .await
+            .unwrap();
+        let state = provisioner.host().get("tsunfactory").unwrap();
+        let addresses: Vec<String> = state
+            .addresses
+            .iter()
+            .map(|entry| entry.to_string())
+            .collect();
+        assert_eq!(addresses, vec!["10.13.37.69/24".to_string()]);
+    }
 }

crates/tsunagi-core/src/overlay/provision/macos.rs

@@ -0,0 +1,531 @@
+//! Managing the overlay interface on macOS, through utun and `ifconfig`/`route`.
+//!
+//! The shape is the same as the Linux and Windows provisioners: observe what
+//! is there, [`plan_changes`](super::plan_changes) to decide what to do, apply
+//! the difference. Only the mechanisms differ.
+//!
+//! # Creating the interface
+//!
+//! The interface is a `utun` device, created through the `tun` dependency's
+//! safe wrapper — [`open_tun`](super::super::tun) — over the in-kernel
+//! `com.apple.net.utun_control` control socket. No kernel extension and no
+//! Network Extension entitlement is involved; what it needs is root. The
+//! device is not made persistent, so closing the last file descriptor removes
+//! the interface — it goes away when the agent does, however the agent ends.
+//!
+//! # The kernel names the interface, not us
+//!
+//! Unlike Linux and Windows, a utun interface cannot be given an arbitrary
+//! name: it is always `utunN`, and the unit number is assigned by the kernel.
+//! So the name derived from the network id (`tsun…`) is only ever a *request*;
+//! the real name is read back off the device after it is opened, and
+//! everything downstream — the host rules, the resolver publication — uses the
+//! real one. The agent adopts it as soon as the device is created (see
+//! [`Interface`](crate::overlay::Interface)).
+//!
+//! # Configuring it
+//!
+//! Addresses, MTU and routes are applied with `ifconfig` and `route`. On Linux
+//! the same work is done in process over a netlink socket; here there is no
+//! equivalent this crate may call, because the in-process route is raw `ioctl`
+//! and `PF_ROUTE` and that is `unsafe`. The tools are invoked by absolute path
+//! from `/sbin`, so nothing on `PATH` can stand in for them, and every value
+//! handed to them is one this agent derived — an interface name chosen by the
+//! kernel, an address it allocated — never anything a peer said.
+//!
+//! # Observing it
+//!
+//! The interface table is read through `netdev`, a safe wrapper over
+//! `getifaddrs`. An interface is matched by its real `utunN` name, which only
+//! matters on reconfiguration: the first create always makes a fresh device.
+
+use std::net::{IpAddr, Ipv4Addr, Ipv6Addr};
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+use super::super::config::Cidr;
+use super::super::tun::{TunDevice, TunRequest, open_tun};
+use super::privilege::{Privilege, probe_net_admin};
+use super::{
+    InterfacePlan, InterfaceProvisioner, InterfaceState, LinkKind, Provisioned, plan_changes,
+};
+
+/// Manages the overlay interface with utun and `ifconfig`/`route`.
+#[derive(Debug)]
+pub struct UtunProvisioner {
+    /// Interfaces this process created, by their real `utunN` name, so a
+    /// reconfiguration adjusts them rather than trying to create them again.
+    ours: Mutex<Vec<String>>,
+    /// Devices kept alive for as long as the interface should exist. Dropping
+    /// one closes the last descriptor, which removes the utin interface.
+    held: Mutex<Vec<(String, Arc<dyn TunDevice>)>>,
+}
+
+impl UtunProvisioner {
+    /// Creates the provisioner, after checking the platform supports one.
+    ///
+    /// Whether the process is actually root is not checked here — reading the
+    /// effective uid needs a libc call this crate forbids — so a missing
+    /// privilege surfaces at utun creation with a precise message, the same
+    /// way the Linux and Windows paths treat the open itself as the honest
+    /// answer.
+    pub fn new() -> Result<Self, OverlayError> {
+        match probe_net_admin() {
+            Privilege::Available => {}
+            Privilege::Missing(reason) => return Err(OverlayError::Unavailable(reason)),
+            Privilege::Unsupported => {
+                return Err(OverlayError::Unavailable(
+                    "interface management is not compiled in".to_string(),
+                ));
+            }
+        }
+        Ok(Self {
+            ours: Mutex::new(Vec::new()),
+            held: Mutex::new(Vec::new()),
+        })
+    }
+
+    fn is_ours(&self, name: &str) -> bool {
+        lock(&self.ours).iter().any(|owned| owned == name)
+    }
+
+    /// Opens the utun device, which is what creates it. The kernel assigns the
+    /// name, so the returned device reports a `utunN` that is not `plan.name`.
+    fn create_device(&self, plan: &InterfacePlan) -> Result<Arc<dyn TunDevice>, OverlayError> {
+        let request = TunRequest::bare(plan.name.clone(), plan.mtu);
+        open_tun(&request)
+    }
+}
+
+impl InterfaceProvisioner for UtunProvisioner {
+    fn name(&self) -> &str {
+        "utun"
+    }
+
+    fn reconcile<'a>(
+        &'a self,
+        plan: &'a InterfacePlan,
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>> {
+        Box::pin(async move {
+            // A reconfiguration: the plan carries the real `utunN` name the
+            // agent adopted after the first create, so the interface is there
+            // and only the addresses and MTU need bringing into line.
+            if self.is_ours(&plan.name) {
+                let current = observe(&plan.name).await;
+                let changes = plan_changes(&current, plan, true)?;
+                configure(&plan.name, &changes).await?;
+                return Ok(Provisioned {
+                    changes,
+                    device: None,
+                });
+            }
+
+            // The first create. The kernel assigns the unit, so the device is
+            // opened first and its real name used for everything after.
+            let device = self.create_device(plan)?;
+            let real = device.name().to_string();
+
+            // A fresh interface starts with nothing on it, so the whole plan is
+            // applied. `plan_changes` against an absent interface is exactly
+            // that decision; its `create_link` flag is already satisfied by the
+            // open above, so `configure` ignores it and only applies the rest.
+            let changes = plan_changes(&InterfaceState::absent(), plan, false)?;
+
+            // A failure here leaves an interface that exists but cannot carry
+            // traffic. Drop the device so the utun closes, and report, rather
+            // than leave a trap behind.
+            if let Err(err) = configure(&real, &changes).await {
+                drop(device);
+                return Err(err);
+            }
+
+            lock(&self.ours).push(real.clone());
+            lock(&self.held).push((real, Arc::clone(&device)));
+            Ok(Provisioned {
+                changes,
+                device: Some(device),
+            })
+        })
+    }
+
+    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move {
+            // Dropping the device closes the last descriptor, which removes the
+            // utun interface and the routes that were bound to it. There is
+            // nothing to flush: a utun never outlives the process that holds it.
+            lock(&self.held).retain(|(held, _)| held != name);
+            lock(&self.ours).retain(|owned| owned != name);
+            Ok(())
+        })
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+/// `fe80::/10` and `169.254.0.0/16`, which macOS assigns on its own.
+fn is_link_local(addr: IpAddr) -> bool {
+    match addr {
+        IpAddr::V4(addr) => addr.is_link_local(),
+        IpAddr::V6(addr) => (addr.segments()[0] & 0xffc0) == 0xfe80,
+    }
+}
+
+/// What the interface named `name` currently looks like.
+///
+/// Never fails: a name nothing is using reads as [`InterfaceState::absent`],
+/// which is what the plan then acts on.
+async fn observe(name: &str) -> InterfaceState {
+    let name = name.to_string();
+    let found = tokio::task::spawn_blocking(move || {
+        netdev::get_interfaces()
+            .into_iter()
+            .find(|iface| iface.name == name)
+    })
+    .await
+    .ok()
+    .flatten();
+
+    let Some(iface) = found else {
+        return InterfaceState::absent();
+    };
+
+    let mut addresses: Vec<Cidr> = Vec::new();
+    for net in &iface.ipv4 {
+        if let Ok(cidr) = Cidr::new(IpAddr::V4(net.addr()), net.prefix_len())
+            && !is_link_local(cidr.addr)
+        {
+            addresses.push(cidr);
+        }
+    }
+    for net in &iface.ipv6 {
+        if let Ok(cidr) = Cidr::new(IpAddr::V6(net.addr()), net.prefix_len())
+            && !is_link_local(cidr.addr)
+        {
+            addresses.push(cidr);
+        }
+    }
+    addresses.sort();
+
+    InterfaceState {
+        // Only ever reached for an interface this agent created (see
+        // `reconcile`), so it is ours and a TUN by construction.
+        kind: LinkKind::Tun,
+        attached: iface.is_up(),
+        up: iface.is_up(),
+        mtu: iface.mtu.unwrap_or(0),
+        addresses,
+    }
+}
+
+/// Applies MTU, link state, addresses and the on-link routes they imply.
+async fn configure(name: &str, changes: &super::Changes) -> Result<(), OverlayError> {
+    if let Some(mtu) = changes.set_mtu {
+        set_mtu(name, mtu).await?;
+    }
+    for cidr in &changes.remove {
+        // Best effort: an address that is already gone is the outcome wanted.
+        if let Err(err) = address_del(name, *cidr).await {
+            tracing::debug!(interface = %name, %cidr, %err, "could not remove address");
+        }
+        if let Err(err) = route_del(name, *cidr).await {
+            tracing::debug!(interface = %name, %cidr, %err, "could not remove route");
+        }
+    }
+    for cidr in &changes.add {
+        address_add(name, *cidr).await?;
+        // The subnet route makes the whole overlay range reach the interface;
+        // a point-to-point utun does not install one from the address alone. A
+        // host address (/32 or /128) needs none. Best effort: a route that is
+        // already there is the outcome wanted.
+        if let Err(err) = route_add(name, *cidr).await {
+            tracing::debug!(interface = %name, %cidr, %err, "could not add route");
+        }
+    }
+    if changes.bring_up {
+        // Best effort: a utun with a descriptor open is up already, and a
+        // failure to nudge it is not a reason to fail the whole reconcile.
+        if let Err(err) = set_up(name).await {
+            tracing::debug!(interface = %name, %err, "could not enable the interface");
+        }
+    }
+    Ok(())
+}
+
+/// The network address of a CIDR, with the host bits cleared.
+fn network_of(cidr: Cidr) -> IpAddr {
+    match cidr.addr {
+        IpAddr::V4(addr) => {
+            let bits = u32::from(addr);
+            let mask = if cidr.prefix_len == 0 {
+                0
+            } else {
+                u32::MAX << (32 - u32::from(cidr.prefix_len))
+            };
+            IpAddr::V4(Ipv4Addr::from(bits & mask))
+        }
+        IpAddr::V6(addr) => {
+            let bits = u128::from(addr);
+            let mask = if cidr.prefix_len == 0 {
+                0
+            } else {
+                u128::MAX << (128 - u32::from(cidr.prefix_len))
+            };
+            IpAddr::V6(Ipv6Addr::from(bits & mask))
+        }
+    }
+}
+
+/// Whether a CIDR is a single host address, which needs no subnet route.
+fn is_host_route(cidr: Cidr) -> bool {
+    match cidr.addr {
+        IpAddr::V4(_) => cidr.prefix_len >= 32,
+        IpAddr::V6(_) => cidr.prefix_len >= 128,
+    }
+}
+
+/// The `ifconfig` arguments that add an address to a utun interface.
+///
+/// A utun is point-to-point, so the IPv4 form names the address as its own
+/// peer; `alias` lets several coexist on the one interface.
+fn address_add_args(name: &str, cidr: Cidr) -> Vec<String> {
+    match cidr.addr {
+        IpAddr::V4(addr) => vec![
+            name.into(),
+            "inet".into(),
+            format!("{addr}/{}", cidr.prefix_len),
+            addr.to_string(),
+            "alias".into(),
+        ],
+        IpAddr::V6(addr) => vec![
+            name.into(),
+            "inet6".into(),
+            format!("{addr}/{}", cidr.prefix_len),
+            "alias".into(),
+        ],
+    }
+}
+
+/// The `ifconfig` arguments that remove an address from a utun interface.
+fn address_del_args(name: &str, cidr: Cidr) -> Vec<String> {
+    let family = match cidr.addr {
+        IpAddr::V4(_) => "inet",
+        IpAddr::V6(_) => "inet6",
+    };
+    vec![
+        name.into(),
+        family.into(),
+        cidr.addr.to_string(),
+        "-alias".into(),
+    ]
+}
+
+/// The `route` arguments that point a CIDR's subnet at the interface.
+fn route_add_args(name: &str, cidr: Cidr) -> Vec<String> {
+    let family = match cidr.addr {
+        IpAddr::V4(_) => "-inet",
+        IpAddr::V6(_) => "-inet6",
+    };
+    vec![
+        "-q".into(),
+        "-n".into(),
+        "add".into(),
+        family.into(),
+        "-net".into(),
+        format!("{}/{}", network_of(cidr), cidr.prefix_len),
+        "-interface".into(),
+        name.into(),
+    ]
+}
+
+/// The `route` arguments that remove that subnet route.
+fn route_del_args(name: &str, cidr: Cidr) -> Vec<String> {
+    let family = match cidr.addr {
+        IpAddr::V4(_) => "-inet",
+        IpAddr::V6(_) => "-inet6",
+    };
+    vec![
+        "-q".into(),
+        "-n".into(),
+        "delete".into(),
+        family.into(),
+        "-net".into(),
+        format!("{}/{}", network_of(cidr), cidr.prefix_len),
+        "-interface".into(),
+        name.into(),
+    ]
+}
+
+async fn address_add(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    ifconfig(address_add_args(name, cidr)).await
+}
+
+async fn address_del(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    ifconfig(address_del_args(name, cidr)).await
+}
+
+async fn route_add(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    if is_host_route(cidr) {
+        return Ok(());
+    }
+    route(route_add_args(name, cidr)).await
+}
+
+async fn route_del(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    if is_host_route(cidr) {
+        return Ok(());
+    }
+    route(route_del_args(name, cidr)).await
+}
+
+async fn set_mtu(name: &str, mtu: u32) -> Result<(), OverlayError> {
+    ifconfig(vec![name.into(), "mtu".into(), mtu.to_string()]).await
+}
+
+async fn set_up(name: &str) -> Result<(), OverlayError> {
+    ifconfig(vec![name.into(), "up".into()]).await
+}
+
+/// Runs `/sbin/ifconfig`, turning a non-zero exit into an error with its text.
+async fn ifconfig(args: Vec<String>) -> Result<(), OverlayError> {
+    run("/sbin/ifconfig", args).await
+}
+
+/// Runs `/sbin/route`, turning a non-zero exit into an error with its text.
+async fn route(args: Vec<String>) -> Result<(), OverlayError> {
+    run("/sbin/route", args).await
+}
+
+/// Runs a system tool by absolute path, so nothing on `PATH` can stand in for
+/// it, and turns a non-zero exit into an error carrying what it said.
+async fn run(program: &'static str, args: Vec<String>) -> Result<(), OverlayError> {
+    let display = format!("{program} {}", args.join(" "));
+    let output = tokio::task::spawn_blocking(move || {
+        std::process::Command::new(program).args(&args).output()
+    })
+    .await
+    .map_err(|err| OverlayError::Other(format!("could not run {display}: {err}")))?
+    .map_err(|err| OverlayError::Unavailable(format!("could not run {display}: {err}")))?;
+
+    if output.status.success() {
+        return Ok(());
+    }
+    let message = {
+        let stderr = String::from_utf8_lossy(&output.stderr);
+        let text = if stderr.trim().is_empty() {
+            String::from_utf8_lossy(&output.stdout).trim().to_string()
+        } else {
+            stderr.trim().to_string()
+        };
+        if text.is_empty() {
+            format!("exit code {}", output.status)
+        } else {
+            text
+        }
+    };
+    Err(OverlayError::Unavailable(format!(
+        "{display} failed: {message}"
+    )))
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    fn v4(last: u8, prefix: u8) -> Cidr {
+        Cidr::new(IpAddr::V4(Ipv4Addr::new(10, 13, 37, last)), prefix).unwrap()
+    }
+
+    fn v6(last: u16, prefix: u8) -> Cidr {
+        Cidr::new(
+            IpAddr::V6(Ipv6Addr::new(0xfd00, 0, 0, 0, 0, 0, 0, last)),
+            prefix,
+        )
+        .unwrap()
+    }
+
+    #[test]
+    fn a_v4_address_is_added_point_to_point_with_an_alias() {
+        let args = address_add_args("utun5", v4(69, 24));
+        assert_eq!(
+            args,
+            vec!["utun5", "inet", "10.13.37.69/24", "10.13.37.69", "alias"]
+        );
+    }
+
+    #[test]
+    fn a_v6_address_is_added_with_a_prefix_and_an_alias() {
+        let args = address_add_args("utun5", v6(1, 128));
+        assert_eq!(args, vec!["utun5", "inet6", "fd00::1/128", "alias"]);
+    }
+
+    #[test]
+    fn deleting_an_address_names_the_family_and_removes_the_alias() {
+        let args = address_del_args("utun5", v4(69, 24));
+        assert_eq!(args, vec!["utun5", "inet", "10.13.37.69", "-alias"]);
+        let args = address_del_args("utun5", v6(1, 128));
+        assert_eq!(args, vec!["utun5", "inet6", "fd00::1", "-alias"]);
+    }
+
+    #[test]
+    fn a_subnet_route_points_the_masked_network_at_the_interface() {
+        let args = route_add_args("utun5", v4(69, 24));
+        assert_eq!(
+            args,
+            vec![
+                "-q",
+                "-n",
+                "add",
+                "-inet",
+                "-net",
+                "10.13.37.0/24",
+                "-interface",
+                "utun5"
+            ]
+        );
+    }
+
+    #[test]
+    fn a_host_address_installs_no_subnet_route() {
+        assert!(is_host_route(v4(69, 32)));
+        assert!(is_host_route(v6(1, 128)));
+        assert!(!is_host_route(v4(69, 24)));
+        assert!(!is_host_route(v6(1, 64)));
+    }
+
+    #[test]
+    fn the_network_address_clears_the_host_bits() {
+        assert_eq!(
+            network_of(v4(69, 24)),
+            IpAddr::V4(Ipv4Addr::new(10, 13, 37, 0))
+        );
+        assert_eq!(
+            network_of(Cidr::new(IpAddr::V4(Ipv4Addr::new(10, 13, 37, 69)), 16).unwrap()),
+            IpAddr::V4(Ipv4Addr::new(10, 13, 0, 0))
+        );
+        assert_eq!(
+            network_of(v6(0x1234, 64)),
+            IpAddr::V6(Ipv6Addr::new(0xfd00, 0, 0, 0, 0, 0, 0, 0))
+        );
+    }
+
+    #[test]
+    fn link_local_addresses_are_not_ours_to_manage() {
+        assert!(is_link_local(IpAddr::V6(Ipv6Addr::new(
+            0xfe80, 0, 0, 0, 0, 0, 0, 1
+        ))));
+        assert!(is_link_local(IpAddr::V4(Ipv4Addr::new(169, 254, 1, 1))));
+        assert!(!is_link_local(IpAddr::V4(Ipv4Addr::new(10, 13, 37, 1))));
+        assert!(!is_link_local(IpAddr::V6(Ipv6Addr::new(
+            0xfd00, 0, 0, 0, 0, 0, 0, 1
+        ))));
+    }
+}

crates/tsunagi-core/src/overlay/provision/mock.rs

@@ -9,7 +9,7 @@ use std::collections::HashMap;
 use std::sync::{Arc, Mutex};
 
 use crate::BoxFuture;
-use crate::dataplane::PluginError;
+use crate::overlay::OverlayError;
 
 use super::super::config::Cidr;
 use super::super::tun::{MemoryTun, MemoryTunFactory, TunFactory, TunRequest};
@@ -125,9 +125,9 @@ impl MockProvisioner {
         }
     }
 
-    fn apply(&self, plan: &InterfacePlan) -> Result<Changes, PluginError> {
+    fn apply(&self, plan: &InterfacePlan) -> Result<Changes, OverlayError> {
         if let Some(reason) = &self.failure {
-            return Err(PluginError::Unavailable(reason.clone()));
+            return Err(OverlayError::Unavailable(reason.clone()));
         }
 
         let ours = self.owned().iter().any(|name| name == &plan.name);
@@ -179,7 +179,7 @@ impl InterfaceProvisioner for MockProvisioner {
     fn reconcile<'a>(
         &'a self,
         plan: &'a InterfacePlan,
-    ) -> BoxFuture<'a, Result<Provisioned, PluginError>> {
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>> {
         Box::pin(async move {
             let changes = self.apply(plan)?;
             let device = if changes.create_link {
@@ -195,7 +195,7 @@ impl InterfaceProvisioner for MockProvisioner {
         })
     }
 
-    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), PluginError>> {
+    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>> {
         Box::pin(async move {
             self.host.lock().remove(name);
             self.owned().retain(|owned| owned != name);

crates/tsunagi-core/src/overlay/provision/mod.rs

@@ -20,9 +20,13 @@
 //! pure and tested on every platform, and only the execution is behind
 //! [`InterfaceProvisioner`].
 //!
-//! Three implementations:
+//! Five implementations:
 //!
 //! * `NetlinkProvisioner` on Linux, which needs `CAP_NET_ADMIN`.
+//! * `WintunProvisioner` on Windows, a Wintun adapter configured with `netsh`,
+//!   which needs an elevated process.
+//! * `UtunProvisioner` on macOS, a utun interface configured with `ifconfig`
+//!   and `route`, which needs root.
 //! * [`MockProvisioner`], an in-memory host used by the tests.
 //! * [`UnsupportedProvisioner`] elsewhere, which fails with an explanation
 //!   and a pointer at the manual route rather than pretending to work.
@@ -38,7 +42,7 @@ use std::collections::BTreeSet;
 use std::sync::Arc;
 
 use crate::BoxFuture;
-use crate::dataplane::PluginError;
+use crate::overlay::OverlayError;
 
 use super::config::Cidr;
 use super::tun::TunDevice;
@@ -50,6 +54,18 @@ pub use mock::{MockHost, MockProvisioner};
 mod linux;
 #[cfg(all(feature = "tun-device", target_os = "linux"))]
 pub use linux::NetlinkProvisioner;
+#[cfg(all(feature = "tun-device", target_os = "linux"))]
+pub(crate) use linux::RouteHandle;
+
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+mod windows;
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use windows::WintunProvisioner;
+
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+mod macos;
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use macos::UtunProvisioner;
 
 mod privilege;
 pub use privilege::{Privilege, probe_net_admin};
@@ -203,19 +219,19 @@ pub fn plan_changes(
     current: &InterfaceState,
     plan: &InterfacePlan,
     ours: bool,
-) -> Result<Changes, PluginError> {
+) -> Result<Changes, OverlayError> {
     let mut changes = Changes::default();
 
     match &current.kind {
         LinkKind::Foreign(kind) => {
-            return Err(PluginError::Unavailable(format!(
+            return Err(OverlayError::Unavailable(format!(
                 "`{}` already exists and is a {kind} interface, not one of ours. \
                  Refusing to touch it. Run with a different interface prefix.",
                 plan.name
             )));
         }
         LinkKind::Tun if !ours && current.attached => {
-            return Err(PluginError::Unavailable(format!(
+            return Err(OverlayError::Unavailable(format!(
                 "`{}` already exists and another process is attached to it. \
                  That is most likely a second agent on this host in the same \
                  network; give one of them a different interface prefix.",
@@ -293,13 +309,13 @@ pub trait InterfaceProvisioner: Send + Sync + std::fmt::Debug + 'static {
     fn reconcile<'a>(
         &'a self,
         plan: &'a InterfacePlan,
-    ) -> BoxFuture<'a, Result<Provisioned, PluginError>>;
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>>;
 
     /// Removes an interface this provisioner created.
     ///
     /// Removing one that is already gone succeeds: this runs on the shutdown
     /// path, where the interface having vanished is the desired outcome.
-    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), PluginError>>;
+    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>>;
 }
 
 #[cfg(test)]

crates/tsunagi-core/src/overlay/provision/privilege.rs

@@ -15,7 +15,7 @@
 //! Grant it with:
 //!
 //! ```text
-//! sudo setcap cap_net_admin+p /usr/local/bin/tsunagi
+//! sudo setcap cap_net_admin,cap_net_bind_service+p /usr/local/bin/tsng
 //! ```
 //!
 //! `+p` rather than `+ep`: with `+p` the capability is permitted but not
@@ -49,7 +49,7 @@ impl Privilege {
     /// How to obtain it, for a diagnostic.
     pub fn how_to_grant(program: &str) -> String {
         format!(
-            "Grant it once with `sudo setcap cap_net_admin+p {program}` \
+            "Grant it once with `sudo setcap cap_net_admin,cap_net_bind_service+p {program}` \
              and the agent manages its own interface. Without it, run with \
              `--no-tun`: the tunnels still form, they just do not reach the \
              operating system."
@@ -60,10 +60,24 @@ impl Privilege {
 #[cfg(all(feature = "tun-device", target_os = "linux"))]
 pub use linux_impl::{NetAdmin, probe_net_admin};
 
-#[cfg(not(all(feature = "tun-device", target_os = "linux")))]
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+pub use windows_impl::probe_net_admin;
+
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+pub use macos_impl::probe_net_admin;
+
+#[cfg(not(any(
+    all(feature = "tun-device", target_os = "linux"),
+    all(feature = "tun-device", target_os = "windows"),
+    all(feature = "tun-device", target_os = "macos")
+)))]
 pub use other_impl::probe_net_admin;
 
-#[cfg(not(all(feature = "tun-device", target_os = "linux")))]
+#[cfg(not(any(
+    all(feature = "tun-device", target_os = "linux"),
+    all(feature = "tun-device", target_os = "windows"),
+    all(feature = "tun-device", target_os = "macos")
+)))]
 mod other_impl {
     use super::Privilege;
 
@@ -73,12 +87,50 @@ mod other_impl {
     }
 }
 
+#[cfg(all(feature = "tun-device", target_os = "windows"))]
+mod windows_impl {
+    use super::Privilege;
+
+    /// Whether this process can configure interfaces.
+    ///
+    /// Windows has no capability to hold and lower the way Linux does:
+    /// creating an adapter simply needs the process to be elevated. Reading
+    /// whether it *is* elevated means inspecting the process token through a
+    /// raw call this crate forbids, so the probe is deliberately optimistic —
+    /// it reports that the platform can manage interfaces — and the real
+    /// check is left to Wintun's adapter creation, which fails with a precise
+    /// message when the process is not elevated. This matches how the Linux
+    /// path treats the open itself as the honest answer.
+    pub fn probe_net_admin() -> Privilege {
+        Privilege::Available
+    }
+}
+
+#[cfg(all(feature = "tun-device", target_os = "macos"))]
+mod macos_impl {
+    use super::Privilege;
+
+    /// Whether this process can configure interfaces.
+    ///
+    /// macOS has no capability to hold and lower the way Linux does: creating
+    /// a utun interface simply needs the process to be root. Reading the
+    /// effective uid means a libc call this crate forbids, so the probe is
+    /// deliberately optimistic — it reports that the platform can manage
+    /// interfaces — and the real check is left to the utun open, which fails
+    /// with a precise "run with sudo" message when the process is not root.
+    /// This matches how the Linux path treats the open itself as the honest
+    /// answer.
+    pub fn probe_net_admin() -> Privilege {
+        Privilege::Available
+    }
+}
+
 #[cfg(all(feature = "tun-device", target_os = "linux"))]
 mod linux_impl {
     use caps::{CapSet, Capability};
 
     use super::Privilege;
-    use crate::dataplane::PluginError;
+    use crate::overlay::OverlayError;
 
     /// Whether this thread holds `CAP_NET_ADMIN` in its permitted set.
     pub fn probe_net_admin() -> Privilege {
@@ -108,18 +160,18 @@ mod linux_impl {
 
     impl NetAdmin {
         /// Raises `CAP_NET_ADMIN` into the effective set.
-        pub fn acquire() -> Result<Self, PluginError> {
+        pub fn acquire() -> Result<Self, OverlayError> {
             let already = caps::has_cap(None, CapSet::Effective, Capability::CAP_NET_ADMIN)
                 .map_err(|err| {
-                    PluginError::Unavailable(format!("cannot read capabilities: {err}"))
+                    OverlayError::Unavailable(format!("cannot read capabilities: {err}"))
                 })?;
             if already {
                 return Ok(Self { raised: false });
             }
             caps::raise(None, CapSet::Effective, Capability::CAP_NET_ADMIN).map_err(|err| {
-                PluginError::Unavailable(format!(
+                OverlayError::Unavailable(format!(
                     "cannot raise CAP_NET_ADMIN: {err}. {}",
-                    Privilege::how_to_grant("tsunagi")
+                    Privilege::how_to_grant("tsng")
                 ))
             })?;
             Ok(Self { raised: true })
@@ -152,8 +204,8 @@ mod tests {
 
     #[test]
     fn the_grant_instructions_name_the_program() {
-        let text = Privilege::how_to_grant("/usr/local/bin/tsunagi");
+        let text = Privilege::how_to_grant("/usr/local/bin/tsng");
-        assert!(text.contains("setcap cap_net_admin+p /usr/local/bin/tsunagi"));
+        assert!(text.contains("setcap cap_net_admin,cap_net_bind_service+p /usr/local/bin/tsng"));
         assert!(text.contains("--no-tun"), "the fallback is offered too");
     }
 

crates/tsunagi-core/src/overlay/provision/unsupported.rs

@@ -7,7 +7,7 @@
 //! of provisioning and says what to do instead.
 
 use crate::BoxFuture;
-use crate::dataplane::PluginError;
+use crate::overlay::OverlayError;
 
 use super::{InterfacePlan, InterfaceProvisioner, Provisioned};
 
@@ -31,8 +31,8 @@ impl UnsupportedProvisioner {
         }
     }
 
-    fn refusal(&self) -> PluginError {
+    fn refusal(&self) -> OverlayError {
-        PluginError::Unavailable(format!(
+        OverlayError::Unavailable(format!(
             "managing the overlay interface is not implemented on {} yet. \
              Run with `--no-tun` until it is: the tunnels still form, they just \
              do not reach the operating system.",
@@ -49,11 +49,11 @@ impl InterfaceProvisioner for UnsupportedProvisioner {
     fn reconcile<'a>(
         &'a self,
         _plan: &'a InterfacePlan,
-    ) -> BoxFuture<'a, Result<Provisioned, PluginError>> {
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>> {
         Box::pin(async move { Err(self.refusal()) })
     }
 
-    fn remove<'a>(&'a self, _name: &'a str) -> BoxFuture<'a, Result<(), PluginError>> {
+    fn remove<'a>(&'a self, _name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>> {
         // Nothing was ever created, so there is nothing to clean up and no
         // reason to fail a shutdown path.
         Box::pin(async move { Ok(()) })

crates/tsunagi-core/src/overlay/provision/windows.rs

@@ -0,0 +1,513 @@
+//! Managing the overlay interface on Windows, through Wintun and `netsh`.
+//!
+//! The shape is the same as the Linux provisioner: observe what is there,
+//! [`plan_changes`](super::plan_changes) to decide what to do, apply the
+//! difference. Only the two mechanisms differ.
+//!
+//! # Creating the interface
+//!
+//! The interface is a Wintun adapter, created through the `tun` dependency's
+//! safe wrapper — [`open_tun`](super::super::tun) — because loading
+//! `wintun.dll` is an `unsafe` call and this crate forbids `unsafe`. The
+//! adapter is created, not made persistent: closing the last handle to it
+//! removes it, so it goes away when the agent does, however the agent ends.
+//! The reader and writer inside the device hold that handle, so the adapter
+//! lives exactly as long as the device does.
+//!
+//! # Configuring it
+//!
+//! Addresses and MTU are applied with `netsh`. On Linux the same work is done
+//! in process over a netlink socket; here there is no equivalent that this
+//! crate may call, because the in-process route is the IP Helper API and that
+//! is `unsafe`. `netsh` is the supported tool for the job, it is invoked by
+//! absolute path from `%SystemRoot%` so nothing on `PATH` can stand in for it,
+//! and every value handed to it is one this agent derived — an interface name
+//! from the network id, an address it allocated — never anything a peer said.
+//!
+//! # Observing it
+//!
+//! The interface table is read through `netdev`, whose enumeration is a safe
+//! wrapper over the same IP Helper API. An adapter is matched by the friendly
+//! name it was created with. A match that carries a default gateway is treated
+//! as foreign and left alone: a tsunagi overlay interface never has one, so a
+//! gateway is the mark of a real device that happens to share the name.
+//!
+//! # What Windows cannot do that Linux can
+//!
+//! Wintun exposes no way to delete an adapter this process did not create, so
+//! a leftover from a run that was killed is *reused* rather than replaced: its
+//! addresses are flushed and the planned ones applied, which reaches the same
+//! end state. A reused adapter is not removed on exit, only closed — that is a
+//! Wintun limitation, and the common case of a clean start and a clean stop is
+//! unaffected.
+
+use std::net::{IpAddr, Ipv4Addr};
+use std::sync::{Arc, Mutex};
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+use super::super::config::Cidr;
+use super::super::tun::{TunDevice, TunRequest, open_tun};
+use super::privilege::{Privilege, probe_net_admin};
+use super::{
+    InterfacePlan, InterfaceProvisioner, InterfaceState, LinkKind, Provisioned, plan_changes,
+};
+
+/// Manages the overlay interface with Wintun and `netsh`.
+#[derive(Debug)]
+pub struct WintunProvisioner {
+    /// Interfaces this process created, so they are adjusted rather than
+    /// replaced. See [`plan_changes`].
+    ours: Mutex<Vec<String>>,
+    /// Devices kept alive for as long as the interface should exist. Dropping
+    /// one closes the adapter, which removes it if this process created it.
+    held: Mutex<Vec<(String, Arc<dyn TunDevice>)>>,
+}
+
+impl WintunProvisioner {
+    /// Creates the provisioner, after checking the platform supports one.
+    ///
+    /// Whether the process is actually elevated is not checked here — that
+    /// cannot be read without an `unsafe` call — so a missing privilege
+    /// surfaces at adapter creation with a precise message, the same way the
+    /// Linux path treats the open itself as the honest answer.
+    pub fn new() -> Result<Self, OverlayError> {
+        match probe_net_admin() {
+            Privilege::Available => {}
+            Privilege::Missing(reason) => return Err(OverlayError::Unavailable(reason)),
+            Privilege::Unsupported => {
+                return Err(OverlayError::Unavailable(
+                    "interface management is not compiled in".to_string(),
+                ));
+            }
+        }
+        Ok(Self {
+            ours: Mutex::new(Vec::new()),
+            held: Mutex::new(Vec::new()),
+        })
+    }
+
+    fn is_ours(&self, name: &str) -> bool {
+        lock(&self.ours)
+            .iter()
+            .any(|owned| owned.eq_ignore_ascii_case(name))
+    }
+
+    /// Opens the Wintun adapter, which is what creates it.
+    fn create_device(&self, plan: &InterfacePlan) -> Result<Arc<dyn TunDevice>, OverlayError> {
+        let request = TunRequest::bare(plan.name.clone(), plan.mtu);
+        open_tun(&request)
+    }
+}
+
+impl InterfaceProvisioner for WintunProvisioner {
+    fn name(&self) -> &str {
+        "wintun"
+    }
+
+    fn reconcile<'a>(
+        &'a self,
+        plan: &'a InterfacePlan,
+    ) -> BoxFuture<'a, Result<Provisioned, OverlayError>> {
+        Box::pin(async move {
+            let current = observe(&plan.name).await;
+            let changes = plan_changes(&current, plan, self.is_ours(&plan.name))?;
+
+            // Wintun cannot delete an adapter this process did not create, so a
+            // "delete then create" from `plan_changes` becomes "flush then
+            // reuse": clear the stale addresses off the interface that is
+            // there, and let the create step below reopen the same adapter.
+            // The end state — an adapter carrying exactly the plan's addresses
+            // — is identical.
+            let flush: Vec<Cidr> = if changes.delete_link {
+                tracing::info!(
+                    interface = %plan.name,
+                    "reusing an abandoned interface left by an earlier run"
+                );
+                current.addresses.clone()
+            } else {
+                Vec::new()
+            };
+
+            for cidr in flush.iter().chain(changes.remove.iter()) {
+                // Best effort: an address that is already gone is the outcome
+                // wanted, and refusing to start over one is not worth it.
+                if let Err(err) = address_del(&plan.name, *cidr).await {
+                    tracing::debug!(interface = %plan.name, %cidr, %err, "could not remove address");
+                }
+            }
+
+            let device = if changes.create_link {
+                let device = self.create_device(plan)?;
+                lock(&self.ours).push(plan.name.clone());
+                lock(&self.held).push((plan.name.clone(), Arc::clone(&device)));
+                Some(device)
+            } else {
+                None
+            };
+
+            // A failure past this point leaves an interface that exists but
+            // cannot carry traffic. Undo the bookkeeping for one we just
+            // created so it is not mistaken for a working interface, and drop
+            // the device so the adapter closes.
+            let configured = configure(plan, &changes).await;
+            if let Err(err) = configured {
+                if changes.create_link {
+                    lock(&self.held).retain(|(held, _)| held != &plan.name);
+                    lock(&self.ours).retain(|owned| owned != &plan.name);
+                }
+                return Err(err);
+            }
+
+            Ok(Provisioned { changes, device })
+        })
+    }
+
+    fn remove<'a>(&'a self, name: &'a str) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move {
+            // Dropping the device closes the adapter, which removes it if this
+            // process created it. A reused adapter is only closed, not removed;
+            // flushing its addresses first keeps no stale configuration behind.
+            let current = observe(name).await;
+            for cidr in &current.addresses {
+                if let Err(err) = address_del(name, *cidr).await {
+                    tracing::debug!(interface = %name, %cidr, %err, "could not remove address");
+                }
+            }
+            lock(&self.held).retain(|(held, _)| !held.eq_ignore_ascii_case(name));
+            lock(&self.ours).retain(|owned| !owned.eq_ignore_ascii_case(name));
+            Ok(())
+        })
+    }
+}
+
+fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+    match mutex.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => poisoned.into_inner(),
+    }
+}
+
+/// `fe80::/10` and `169.254.0.0/16`, which Windows assigns on its own.
+fn is_link_local(addr: IpAddr) -> bool {
+    match addr {
+        IpAddr::V4(addr) => addr.is_link_local(),
+        IpAddr::V6(addr) => (addr.segments()[0] & 0xffc0) == 0xfe80,
+    }
+}
+
+/// What the interface named `name` currently looks like.
+///
+/// Never fails: a name nothing is using reads as [`InterfaceState::absent`],
+/// which is what the plan then acts on.
+async fn observe(name: &str) -> InterfaceState {
+    let name = name.to_string();
+    let found = tokio::task::spawn_blocking(move || {
+        netdev::get_interfaces()
+            .into_iter()
+            .find(|iface| match &iface.friendly_name {
+                Some(friendly) => friendly.eq_ignore_ascii_case(&name),
+                None => false,
+            })
+    })
+    .await
+    .ok()
+    .flatten();
+
+    let Some(iface) = found else {
+        return InterfaceState::absent();
+    };
+
+    let mut addresses: Vec<Cidr> = Vec::new();
+    for net in &iface.ipv4 {
+        if let Ok(cidr) = Cidr::new(IpAddr::V4(net.addr()), net.prefix_len())
+            && !is_link_local(cidr.addr)
+        {
+            addresses.push(cidr);
+        }
+    }
+    for net in &iface.ipv6 {
+        if let Ok(cidr) = Cidr::new(IpAddr::V6(net.addr()), net.prefix_len())
+            && !is_link_local(cidr.addr)
+        {
+            addresses.push(cidr);
+        }
+    }
+    addresses.sort();
+
+    // A tsunagi overlay interface only ever carries the addresses it was given
+    // and no gateway. One that has a gateway is a real device sharing the name,
+    // so it is foreign and never touched.
+    let kind = if iface.gateway.is_some() {
+        LinkKind::Foreign("gatewayed".to_string())
+    } else {
+        LinkKind::Tun
+    };
+
+    InterfaceState {
+        kind,
+        // Wintun reports an adapter with a live session as up and an abandoned
+        // one as down, so "up" doubles as "some process is holding it".
+        attached: iface.is_up(),
+        up: iface.is_up(),
+        mtu: iface.mtu.unwrap_or(0),
+        addresses,
+    }
+}
+
+/// Applies MTU, link state and addresses.
+async fn configure(plan: &InterfacePlan, changes: &super::Changes) -> Result<(), OverlayError> {
+    if let Some(mtu) = changes.set_mtu {
+        set_mtu(&plan.name, mtu).await?;
+    }
+    for cidr in &changes.add {
+        if let Err(err) = address_add(&plan.name, *cidr).await {
+            let err_str = err.to_string();
+            if !err_str.contains("already exists") && !err_str.contains("0x80071392") {
+                return Err(err);
+            }
+        }
+    }
+    if changes.bring_up {
+        // Best effort: a Wintun adapter with a session is up already, and a
+        // failure to nudge it is not a reason to fail the whole reconcile.
+        if let Err(err) = set_up(&plan.name).await {
+            tracing::debug!(interface = %plan.name, %err, "could not enable the interface");
+        }
+    }
+    Ok(())
+}
+
+/// The dotted netmask for an IPv4 prefix length.
+fn ipv4_mask(prefix_len: u8) -> Ipv4Addr {
+    if prefix_len == 0 {
+        Ipv4Addr::UNSPECIFIED
+    } else {
+        Ipv4Addr::from(u32::MAX << (32 - u32::from(prefix_len)))
+    }
+}
+
+/// The `netsh` arguments that add an address to an interface.
+fn address_add_args(name: &str, cidr: Cidr) -> Vec<String> {
+    match cidr.addr {
+        IpAddr::V4(addr) => vec![
+            "interface".into(),
+            "ipv4".into(),
+            "add".into(),
+            "address".into(),
+            format!("name={name}"),
+            format!("address={addr}"),
+            format!("mask={}", ipv4_mask(cidr.prefix_len)),
+            "store=active".into(),
+        ],
+        IpAddr::V6(addr) => vec![
+            "interface".into(),
+            "ipv6".into(),
+            "add".into(),
+            "address".into(),
+            format!("interface={name}"),
+            format!("address={addr}/{}", cidr.prefix_len),
+            "store=active".into(),
+        ],
+    }
+}
+
+/// The `netsh` arguments that remove an address from an interface.
+fn address_del_args(name: &str, cidr: Cidr) -> Vec<String> {
+    match cidr.addr {
+        IpAddr::V4(addr) => vec![
+            "interface".into(),
+            "ipv4".into(),
+            "delete".into(),
+            "address".into(),
+            format!("name={name}"),
+            format!("address={addr}"),
+            "store=active".into(),
+        ],
+        IpAddr::V6(addr) => vec![
+            "interface".into(),
+            "ipv6".into(),
+            "delete".into(),
+            "address".into(),
+            format!("interface={name}"),
+            format!("address={addr}"),
+            "store=active".into(),
+        ],
+    }
+}
+
+/// The `netsh` arguments that set an interface's MTU, one call per family.
+fn mtu_args(name: &str, family: &str, mtu: u32) -> Vec<String> {
+    vec![
+        "interface".into(),
+        family.into(),
+        "set".into(),
+        "subinterface".into(),
+        name.into(),
+        format!("mtu={mtu}"),
+        "store=active".into(),
+    ]
+}
+
+async fn address_add(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    netsh(address_add_args(name, cidr)).await
+}
+
+async fn address_del(name: &str, cidr: Cidr) -> Result<(), OverlayError> {
+    netsh(address_del_args(name, cidr)).await
+}
+
+/// Sets the MTU on both families.
+///
+/// A family that is turned off on the adapter cannot take an MTU, so this
+/// fails only when *neither* would — one family carrying the plan's MTU is
+/// enough, and refusing the whole interface because the other is disabled
+/// would be wrong.
+async fn set_mtu(name: &str, mtu: u32) -> Result<(), OverlayError> {
+    let v4 = netsh(mtu_args(name, "ipv4", mtu)).await;
+    let v6 = netsh(mtu_args(name, "ipv6", mtu)).await;
+    match (&v4, &v6) {
+        (Err(_), Err(_)) => v4,
+        _ => {
+            if let Err(err) = &v4 {
+                tracing::debug!(interface = %name, %err, "could not set the IPv4 MTU");
+            }
+            if let Err(err) = &v6 {
+                tracing::debug!(interface = %name, %err, "could not set the IPv6 MTU");
+            }
+            Ok(())
+        }
+    }
+}
+
+async fn set_up(name: &str) -> Result<(), OverlayError> {
+    netsh(vec![
+        "interface".into(),
+        "set".into(),
+        "interface".into(),
+        format!("name={name}"),
+        "admin=enabled".into(),
+    ])
+    .await
+}
+
+/// Runs `netsh` from `%SystemRoot%\System32`, so nothing on `PATH` can stand
+/// in for it, and turns a non-zero exit into an error carrying what it said.
+async fn netsh(args: Vec<String>) -> Result<(), OverlayError> {
+    let program = system32("netsh.exe");
+    let display = format!("netsh {}", args.join(" "));
+    let output = tokio::task::spawn_blocking(move || {
+        std::process::Command::new(program).args(&args).output()
+    })
+    .await
+    .map_err(|err| OverlayError::Other(format!("could not run {display}: {err}")))?
+    .map_err(|err| OverlayError::Unavailable(format!("could not run {display}: {err}")))?;
+
+    if output.status.success() {
+        return Ok(());
+    }
+    let message = {
+        let stderr = String::from_utf8_lossy(&output.stderr);
+        let text = if stderr.trim().is_empty() {
+            String::from_utf8_lossy(&output.stdout).trim().to_string()
+        } else {
+            stderr.trim().to_string()
+        };
+        if text.is_empty() {
+            format!("exit code {}", output.status)
+        } else {
+            text
+        }
+    };
+    Err(OverlayError::Unavailable(format!(
+        "{display} failed: {message}"
+    )))
+}
+
+/// The absolute path to a program in `System32`.
+fn system32(exe: &str) -> std::path::PathBuf {
+    let root = std::env::var_os("SystemRoot").unwrap_or_else(|| r"C:\Windows".into());
+    std::path::Path::new(&root).join("System32").join(exe)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use std::net::Ipv6Addr;
+
+    fn v4(last: u8, prefix: u8) -> Cidr {
+        Cidr::new(IpAddr::V4(Ipv4Addr::new(10, 13, 37, last)), prefix).unwrap()
+    }
+
+    fn v6(last: u16) -> Cidr {
+        Cidr::new(
+            IpAddr::V6(Ipv6Addr::new(0xfd00, 0, 0, 0, 0, 0, 0, last)),
+            64,
+        )
+        .unwrap()
+    }
+
+    #[test]
+    fn a_prefix_length_becomes_the_right_dotted_mask() {
+        assert_eq!(ipv4_mask(24), Ipv4Addr::new(255, 255, 255, 0));
+        assert_eq!(ipv4_mask(16), Ipv4Addr::new(255, 255, 0, 0));
+        assert_eq!(ipv4_mask(32), Ipv4Addr::new(255, 255, 255, 255));
+        assert_eq!(ipv4_mask(0), Ipv4Addr::UNSPECIFIED);
+    }
+
+    #[test]
+    fn a_v4_address_is_added_by_name_with_a_dotted_mask() {
+        let args = address_add_args("tsun0", v4(69, 24));
+        assert!(args.contains(&"ipv4".to_string()));
+        assert!(args.contains(&"add".to_string()));
+        assert!(args.contains(&"name=tsun0".to_string()));
+        assert!(args.contains(&"address=10.13.37.69".to_string()));
+        assert!(args.contains(&"mask=255.255.255.0".to_string()));
+    }
+
+    #[test]
+    fn a_v6_address_is_added_by_interface_with_a_prefix() {
+        let args = address_add_args("tsun0", v6(1));
+        assert!(args.contains(&"ipv6".to_string()));
+        assert!(args.contains(&"interface=tsun0".to_string()));
+        assert!(args.contains(&"address=fd00::1/64".to_string()));
+    }
+
+    #[test]
+    fn deleting_an_address_names_no_mask() {
+        let args = address_del_args("tsun0", v4(69, 24));
+        assert!(args.contains(&"delete".to_string()));
+        assert!(args.contains(&"address=10.13.37.69".to_string()));
+        assert!(!args.iter().any(|arg| arg.starts_with("mask=")));
+    }
+
+    #[test]
+    fn the_mtu_is_set_per_family_on_the_subinterface() {
+        let args = mtu_args("tsun0", "ipv4", 1280);
+        assert!(args.contains(&"subinterface".to_string()));
+        assert!(args.contains(&"tsun0".to_string()));
+        assert!(args.contains(&"mtu=1280".to_string()));
+    }
+
+    #[test]
+    fn link_local_addresses_are_not_ours_to_manage() {
+        assert!(is_link_local(IpAddr::V6(Ipv6Addr::new(
+            0xfe80, 0, 0, 0, 0, 0, 0, 1
+        ))));
+        assert!(is_link_local(IpAddr::V4(Ipv4Addr::new(169, 254, 1, 1))));
+        assert!(!is_link_local(IpAddr::V4(Ipv4Addr::new(10, 13, 37, 1))));
+        assert!(!is_link_local(IpAddr::V6(Ipv6Addr::new(
+            0xfd00, 0, 0, 0, 0, 0, 0, 1
+        ))));
+    }
+
+    #[test]
+    fn system32_is_absolute_and_ends_with_the_program() {
+        let path = system32("netsh.exe");
+        assert!(path.is_absolute(), "{path:?}");
+        assert!(path.ends_with("netsh.exe"), "{path:?}");
+    }
+}

crates/tsunagi-core/src/overlay/router.rs

@@ -0,0 +1,937 @@
+//! Which peer a packet is for, and whether a peer may have sent it.
+//!
+//! One agent has one interface, so a packet coming off it could belong to any
+//! network the agent is in and to any protocol currently carrying traffic.
+//! Deciding which is this table's job, and it is the reason the interface
+//! belongs to the system level: the answer comes from signed state, which no
+//! protocol owns.
+//!
+//! Two questions, and they are not the same one:
+//!
+//! * **Outbound**: this destination address — whose is it? A packet for
+//!   nobody is dropped rather than broadcast.
+//! * **Inbound**: this peer decrypted a packet with this source address —
+//!   is that address actually its? A peer may not speak for anybody else,
+//!   and the check consults the signed claim rather than anything the peer
+//!   said.
+//!
+//! Nothing here is derived from a protocol's key. An address is allocated at
+//! the system level and signed by the member that holds it, so every protocol
+//! carries traffic for the same addresses and the table is the same whichever
+//! one is in use.
+
+use super::broadcast::{BroadcastPolicy, BroadcastTable};
+use arc_swap::ArcSwap;
+use std::collections::HashMap;
+use std::net::Ipv4Addr;
+use std::sync::{Arc, RwLock};
+
+use iroh::EndpointId;
+
+use crate::identity::NetworkId;
+use crate::state::Ipv4Range;
+
+/// Where an outbound packet is going.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct Route {
+    /// The network the destination belongs to.
+    pub network: NetworkId,
+    /// The member that holds the destination address.
+    pub peer: EndpointId,
+}
+
+/// A network's exit-node settings, as far as packets are concerned.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct ExitPolicy {
+    /// This agent offers itself as an exit node here: packets from members
+    /// for addresses outside every overlay range are handed to the host.
+    pub offer: bool,
+    /// The member all other traffic is sent to, if this device uses one.
+    pub via: Option<EndpointId>,
+}
+
+/// What one network contributes to the table.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct NetworkRoutes {
+    /// Exit-node settings of this network.
+    pub exit: ExitPolicy,
+    /// Local broadcast policy and authenticated willing recipients.
+    pub broadcast: BroadcastPolicy,
+    /// The range this network allocates from, once it has agreed one.
+    pub range: Option<Ipv4Range>,
+    /// This agent's own address in the network.
+    pub local: Option<Ipv4Addr>,
+    /// Every other member's address, as the signed state has it.
+    pub peers: Vec<(Ipv4Addr, EndpointId)>,
+}
+
+/// Why a network cannot be added to the table.
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+pub enum RouteError {
+    /// Two networks would answer for the same addresses.
+    #[error(
+        "network {network} uses {range}, which overlaps {other_range} already in use by \
+         {other}. One agent has one interface, so an address can belong to only one of \
+         them: give one network a different range."
+    )]
+    Overlap {
+        /// The network being added.
+        network: NetworkId,
+        /// The range it wants.
+        range: Ipv4Range,
+        /// The network already using an overlapping range.
+        other: NetworkId,
+        /// That network's range.
+        other_range: Ipv4Range,
+    },
+}
+
+/// Address ownership across every network this agent is in.
+#[derive(Debug, Default)]
+pub struct RoutingTable {
+    broadcast: ArcSwap<BroadcastTable>,
+    networks: RwLock<HashMap<NetworkId, NetworkRoutes>>,
+}
+
+impl RoutingTable {
+    /// An empty table.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    fn read(&self) -> std::sync::RwLockReadGuard<'_, HashMap<NetworkId, NetworkRoutes>> {
+        match self.networks.read() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        }
+    }
+
+    fn write(&self) -> std::sync::RwLockWriteGuard<'_, HashMap<NetworkId, NetworkRoutes>> {
+        match self.networks.write() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        }
+    }
+
+    /// Replaces what one network contributes.
+    ///
+    /// Rejected when its range overlaps another network's. With one
+    /// interface an address can belong to only one network, and guessing
+    /// which would send somebody's traffic to a stranger.
+    pub fn set_network(&self, network: NetworkId, routes: NetworkRoutes) -> Result<(), RouteError> {
+        let mut networks = self.write();
+        if let Some(range) = routes.range {
+            for (other, existing) in networks.iter() {
+                if *other == network {
+                    continue;
+                }
+                let Some(other_range) = existing.range else {
+                    continue;
+                };
+                if ranges_overlap(range, other_range) {
+                    return Err(RouteError::Overlap {
+                        network,
+                        range,
+                        other: *other,
+                        other_range,
+                    });
+                }
+            }
+        }
+        networks.insert(network, routes);
+        self.broadcast
+            .store(Arc::new(BroadcastTable::build(&networks)));
+        Ok(())
+    }
+
+    /// Forgets a network.
+    pub fn remove_network(&self, network: NetworkId) {
+        let mut networks = self.write();
+        networks.remove(&network);
+        self.broadcast
+            .store(Arc::new(BroadcastTable::build(&networks)));
+    }
+
+    /// Replaces only a network's broadcast participation, without changing addresses.
+    pub fn set_broadcast(&self, network: NetworkId, policy: BroadcastPolicy) {
+        let mut networks = self.write();
+        if let Some(routes) = networks.get_mut(&network) {
+            if routes.broadcast == policy {
+                return;
+            }
+            routes.broadcast = policy;
+            self.broadcast
+                .store(Arc::new(BroadcastTable::build(&networks)));
+        }
+    }
+
+    /// Replaces only a network's exit-node settings, without changing addresses.
+    ///
+    /// A network that has no entry yet (its range is not agreed) takes them
+    /// from [`NetworkRoutes::exit`] when it is added.
+    pub fn set_exit(&self, network: NetworkId, policy: ExitPolicy) {
+        if let Some(routes) = self.write().get_mut(&network) {
+            routes.exit = policy;
+        }
+    }
+
+    /// Recognizes limited and configured subnet broadcasts before unicast lookup.
+    pub fn is_broadcast(&self, destination: Ipv4Addr) -> bool {
+        self.broadcast.load().is_destination(destination)
+    }
+
+    /// A precomputed domain-scoped recipient list; no per-packet graph search.
+    pub fn broadcast_recipients(
+        &self,
+        source: Ipv4Addr,
+        destination: Ipv4Addr,
+    ) -> Option<Arc<[Route]>> {
+        self.broadcast.load().outgoing(source, destination)
+    }
+
+    /// Local admission after the plugin authenticated and decrypted the sender.
+    pub fn accepts_broadcast(
+        &self,
+        network: NetworkId,
+        peer: EndpointId,
+        destination: Ipv4Addr,
+    ) -> bool {
+        self.broadcast.load().accepts(network, peer, destination)
+    }
+
+    /// The peer that holds a destination address, if anybody does.
+    ///
+    /// This agent's own address returns `None`: a packet for ourselves does
+    /// not go over a tunnel, and answering with a peer would send it to one.
+    ///
+    /// An address nobody holds, outside every overlay range and an ordinary
+    /// unicast destination, goes to this device's exit node when it has one:
+    /// that is what sending *all* traffic through a member means.
+    pub fn route(&self, destination: Ipv4Addr) -> Option<Route> {
+        let networks = self.read();
+        for (network, routes) in networks.iter() {
+            if routes.local == Some(destination) {
+                return None;
+            }
+            if let Some((_, peer)) = routes
+                .peers
+                .iter()
+                .find(|(address, _)| *address == destination)
+            {
+                return Some(Route {
+                    network: *network,
+                    peer: *peer,
+                });
+            }
+        }
+        // Inside an overlay range, an unowned address is nobody's, not the
+        // internet's.
+        if !is_internet_destination(destination) || in_any_range(&networks, destination) {
+            return None;
+        }
+        exit_via(&networks).map(|(network, peer)| Route { network, peer })
+    }
+
+    /// Whether a peer may send from a source address.
+    ///
+    /// A peer speaks only for the address the network agreed it holds. The
+    /// one exception is this device's exit node: what comes back from the
+    /// internet carries the internet's addresses, so from *that* peer any
+    /// source outside the overlay ranges is accepted.
+    pub fn may_send_from(&self, network: NetworkId, peer: EndpointId, source: Ipv4Addr) -> bool {
+        let networks = self.read();
+        let Some(routes) = networks.get(&network) else {
+            return false;
+        };
+        if routes
+            .peers
+            .iter()
+            .any(|(address, holder)| *address == source && *holder == peer)
+        {
+            return true;
+        }
+        routes.exit.via == Some(peer)
+            && is_internet_destination(source)
+            && !in_any_range(&networks, source)
+    }
+
+    /// Whether this agent, as an exit node of `network`, takes a member's
+    /// packet for `destination` to the host.
+    ///
+    /// Only an ordinary unicast address outside every overlay range, so a
+    /// packet for another member, for this agent or for the local network
+    /// segment never leaves by this door. The sender's own source address is
+    /// checked separately, by [`Self::may_send_from`].
+    pub fn accepts_exit_traffic(&self, network: NetworkId, destination: Ipv4Addr) -> bool {
+        let networks = self.read();
+        networks
+            .get(&network)
+            .is_some_and(|routes| routes.exit.offer)
+            && is_internet_destination(destination)
+            && !in_any_range(&networks, destination)
+    }
+
+    /// The network and member this device sends its internet traffic through.
+    ///
+    /// One default route, so one answer; if several networks claim one (they
+    /// should not) the lowest network id decides, so two runs agree.
+    pub fn exit_via(&self) -> Option<(NetworkId, EndpointId)> {
+        exit_via(&self.read())
+    }
+
+    /// Networks that offer this agent as an exit node, with the range each
+    /// agreed. A network that has not agreed one yet has nothing to
+    /// masquerade and is left out.
+    pub fn exit_offers(&self) -> Vec<(NetworkId, Ipv4Range)> {
+        let mut offers: Vec<(NetworkId, Ipv4Range)> = self
+            .read()
+            .iter()
+            .filter(|(_, routes)| routes.exit.offer)
+            .filter_map(|(network, routes)| Some((*network, routes.range?)))
+            .collect();
+        offers.sort_by_key(|(network, _)| *network);
+        offers
+    }
+
+    /// The range of every network in the table that has agreed one, sorted.
+    pub fn overlay_ranges(&self) -> Vec<Ipv4Range> {
+        let mut ranges: Vec<Ipv4Range> = self
+            .read()
+            .values()
+            .filter_map(|routes| routes.range)
+            .collect();
+        ranges.sort_by_key(|range| (range.base, range.prefix_len));
+        ranges.dedup();
+        ranges
+    }
+
+    /// One network's exit-node settings and its range, if it is in the table.
+    pub fn exit_settings(&self, network: NetworkId) -> Option<(ExitPolicy, Option<Ipv4Range>)> {
+        self.read()
+            .get(&network)
+            .map(|routes| (routes.exit, routes.range))
+    }
+
+    /// Whether an ordinary packet terminates on this host in this network.
+    /// Exported LAN subnets must extend this admission policy explicitly later.
+    pub fn is_local_destination(&self, network: NetworkId, destination: Ipv4Addr) -> bool {
+        self.read()
+            .get(&network)
+            .is_some_and(|routes| routes.local == Some(destination))
+    }
+
+    /// Every address this agent should answer to, with its prefix length.
+    pub fn local_addresses(&self) -> Vec<(Ipv4Addr, u8)> {
+        let mut addresses: Vec<(Ipv4Addr, u8)> = self
+            .read()
+            .values()
+            .filter_map(|routes| Some((routes.local?, routes.range?.prefix_len)))
+            .collect();
+        addresses.sort();
+        addresses
+    }
+
+    /// The one network whose address the limited broadcast route is bound to.
+    ///
+    /// The host has a single `255.255.255.255` route per interface, and one
+    /// interface carries every network, so only one overlay address can be its
+    /// source. The choice is the lowest [`NetworkId`] among networks that have
+    /// broadcast enabled, a range and a local address: deterministic, so two
+    /// runs with the same networks pick the same one.
+    pub fn broadcast_host_target(&self) -> Option<(Ipv4Addr, Ipv4Range)> {
+        let networks = self.read();
+        networks
+            .iter()
+            .filter(|(_, routes)| routes.broadcast.enabled)
+            .filter_map(|(network, routes)| Some((*network, routes.local?, routes.range?)))
+            .min_by_key(|(network, _, _)| *network)
+            .map(|(_, local, range)| (local, range))
+    }
+
+    /// Whether a range would collide with one another network already uses.
+    ///
+    /// Asked before proposing a range rather than after: with one interface
+    /// an agent that claimed an address it could not route would also be
+    /// telling everybody else to use that range, and "the range of the
+    /// lowest author wins" would spread the collision instead of containing
+    /// it.
+    pub fn would_overlap(&self, network: NetworkId, range: Ipv4Range) -> Option<Ipv4Range> {
+        self.read().iter().find_map(|(other, existing)| {
+            if *other == network {
+                return None;
+            }
+            let other_range = existing.range?;
+            ranges_overlap(range, other_range).then_some(other_range)
+        })
+    }
+
+    /// How many networks the table covers.
+    pub fn len(&self) -> usize {
+        self.read().len()
+    }
+
+    /// Whether it covers none.
+    pub fn is_empty(&self) -> bool {
+        self.read().is_empty()
+    }
+}
+
+/// Whether an address is an ordinary unicast one: not this host, not a
+/// multicast or broadcast group, not link-local, not unspecified.
+fn is_internet_destination(address: Ipv4Addr) -> bool {
+    !(address.is_unspecified()
+        || address.is_loopback()
+        || address.is_multicast()
+        || address.is_broadcast()
+        || address.is_link_local())
+}
+
+/// Whether an address lies in any network's range.
+fn in_any_range(networks: &HashMap<NetworkId, NetworkRoutes>, address: Ipv4Addr) -> bool {
+    networks
+        .values()
+        .any(|routes| routes.range.is_some_and(|range| range.contains(address)))
+}
+
+/// The exit node in use, if any, from the lowest network that names one.
+fn exit_via(networks: &HashMap<NetworkId, NetworkRoutes>) -> Option<(NetworkId, EndpointId)> {
+    networks
+        .iter()
+        .filter_map(|(network, routes)| Some((*network, routes.exit.via?)))
+        .min_by_key(|(network, _)| *network)
+}
+
+/// Whether two ranges share any address.
+fn ranges_overlap(one: Ipv4Range, other: Ipv4Range) -> bool {
+    // A range contains the other's base, or the other way round. With
+    // prefix-aligned ranges that is the whole of it: two blocks either nest
+    // or are disjoint.
+    one.contains(other.base) || other.contains(one.base)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+    use crate::identity::{NetworkKeys, NetworkName, NetworkSecret};
+
+    fn network(name: &str) -> NetworkId {
+        NetworkKeys::derive(
+            &NetworkName::new(name).unwrap(),
+            &NetworkSecret::from_bytes(vec![4u8; 32]).unwrap(),
+        )
+        .network_id()
+    }
+
+    fn peer(seed: u8) -> EndpointId {
+        iroh::SecretKey::from_bytes(&[seed; 32]).public()
+    }
+
+    fn addr(last: u8) -> Ipv4Addr {
+        Ipv4Addr::new(10, 13, 37, last)
+    }
+
+    fn routes() -> NetworkRoutes {
+        NetworkRoutes {
+            exit: Default::default(),
+            broadcast: Default::default(),
+            range: Some("10.13.37.0/24".parse().unwrap()),
+            local: Some(addr(1)),
+            peers: vec![(addr(2), peer(2)), (addr(3), peer(3))],
+        }
+    }
+
+    #[test]
+    fn a_packet_goes_to_the_member_that_holds_its_destination() {
+        let table = RoutingTable::new();
+        let id = network("routing");
+        table.set_network(id, routes()).unwrap();
+
+        assert_eq!(
+            table.route(addr(2)),
+            Some(Route {
+                network: id,
+                peer: peer(2)
+            })
+        );
+        assert_eq!(
+            table.route(addr(3)),
+            Some(Route {
+                network: id,
+                peer: peer(3)
+            })
+        );
+    }
+
+    #[test]
+    fn a_packet_for_nobody_has_nowhere_to_go() {
+        // Dropped rather than sent to everybody: a tunnel is not a broadcast
+        // domain, and flooding would hand one member another's traffic.
+        let table = RoutingTable::new();
+        table.set_network(network("routing"), routes()).unwrap();
+        assert_eq!(table.route(addr(200)), None);
+        assert_eq!(table.route(Ipv4Addr::new(192, 0, 2, 1)), None);
+    }
+
+    #[test]
+    fn a_packet_for_ourselves_is_not_sent_to_a_peer() {
+        let table = RoutingTable::new();
+        table.set_network(network("routing"), routes()).unwrap();
+        assert_eq!(table.route(addr(1)), None, "that address is ours");
+    }
+
+    #[test]
+    fn a_peer_may_only_send_from_the_address_it_holds() {
+        let table = RoutingTable::new();
+        let id = network("ownership");
+        table.set_network(id, routes()).unwrap();
+
+        assert!(table.may_send_from(id, peer(2), addr(2)));
+        // Somebody else's address, and its own network's.
+        assert!(!table.may_send_from(id, peer(2), addr(3)));
+        assert!(!table.may_send_from(id, peer(2), addr(1)));
+        assert!(!table.may_send_from(id, peer(2), addr(99)));
+        // A peer that holds nothing here.
+        assert!(!table.may_send_from(id, peer(9), addr(2)));
+        // And the question is per network, not global.
+        assert!(!table.may_send_from(network("elsewhere"), peer(2), addr(2)));
+    }
+
+    #[test]
+    fn two_networks_with_overlapping_ranges_are_refused() {
+        // One interface means an address belongs to one network. Accepting
+        // both and guessing would send somebody's traffic to a stranger.
+        let table = RoutingTable::new();
+        let first = network("first");
+        let second = network("second");
+        table.set_network(first, routes()).unwrap();
+
+        let err = table.set_network(second, routes()).unwrap_err();
+        let RouteError::Overlap { other, .. } = err;
+        assert_eq!(other, first);
+        assert!(err.to_string().contains("different range"), "{err}");
+        assert_eq!(table.len(), 1, "the overlapping network is not added");
+
+        // A range of its own is fine.
+        table
+            .set_network(
+                second,
+                NetworkRoutes {
+                    exit: Default::default(),
+                    broadcast: Default::default(),
+                    range: Some("10.99.0.0/16".parse().unwrap()),
+                    local: Some(Ipv4Addr::new(10, 99, 0, 1)),
+                    peers: vec![(Ipv4Addr::new(10, 99, 0, 2), peer(4))],
+                },
+            )
+            .unwrap();
+        assert_eq!(table.len(), 2);
+        assert_eq!(
+            table.route(Ipv4Addr::new(10, 99, 0, 2)),
+            Some(Route {
+                network: second,
+                peer: peer(4)
+            })
+        );
+        // And the first network still answers for its own.
+        assert_eq!(table.route(addr(2)).map(|route| route.network), Some(first));
+    }
+
+    #[test]
+    fn the_host_broadcast_target_is_the_lowest_enabled_network_with_an_address() {
+        let table = RoutingTable::new();
+        assert_eq!(table.broadcast_host_target(), None);
+
+        let (low, high) = {
+            let (a, b) = (network("first"), network("second"));
+            if a < b { (a, b) } else { (b, a) }
+        };
+        let mk = |enabled: bool, base: [u8; 4], local: Option<Ipv4Addr>| NetworkRoutes {
+            exit: Default::default(),
+            broadcast: BroadcastPolicy {
+                enabled,
+                ..Default::default()
+            },
+            range: Some(
+                format!("{}.{}.{}.0/24", base[0], base[1], base[2])
+                    .parse()
+                    .unwrap(),
+            ),
+            local,
+            peers: Vec::new(),
+        };
+        let a = Ipv4Addr::new(10, 1, 0, 1);
+        let b = Ipv4Addr::new(10, 2, 0, 1);
+
+        // The lower id has broadcast off, so the higher one is chosen.
+        table
+            .set_network(low, mk(false, [10, 1, 0, 0], Some(a)))
+            .unwrap();
+        table
+            .set_network(high, mk(true, [10, 2, 0, 0], Some(b)))
+            .unwrap();
+        assert_eq!(
+            table.broadcast_host_target(),
+            Some((b, "10.2.0.0/24".parse().unwrap()))
+        );
+
+        // Both enabled: the lowest id wins, whatever the insertion order.
+        table
+            .set_network(low, mk(true, [10, 1, 0, 0], Some(a)))
+            .unwrap();
+        assert_eq!(
+            table.broadcast_host_target(),
+            Some((a, "10.1.0.0/24".parse().unwrap()))
+        );
+
+        // No local address yet: it cannot be the route's source.
+        table
+            .set_network(low, mk(true, [10, 1, 0, 0], None))
+            .unwrap();
+        assert_eq!(
+            table.broadcast_host_target(),
+            Some((b, "10.2.0.0/24".parse().unwrap()))
+        );
+
+        table.remove_network(high);
+        assert_eq!(table.broadcast_host_target(), None);
+    }
+
+    #[test]
+    fn a_range_can_be_asked_about_before_it_is_proposed() {
+        // The point of asking first: an agent that claims an address it
+        // cannot route also tells everybody else to use that range.
+        let table = RoutingTable::new();
+        let first = network("first");
+        let second = network("second");
+        table.set_network(first, routes()).unwrap();
+
+        let mine: Ipv4Range = "10.13.37.0/24".parse().unwrap();
+        assert_eq!(table.would_overlap(second, mine), Some(mine));
+        // Its own range is not a collision with itself.
+        assert_eq!(table.would_overlap(first, mine), None);
+        assert_eq!(
+            table.would_overlap(second, "10.99.0.0/16".parse().unwrap()),
+            None
+        );
+    }
+
+    #[test]
+    fn a_nested_range_counts_as_overlapping() {
+        let table = RoutingTable::new();
+        table
+            .set_network(
+                network("wide"),
+                NetworkRoutes {
+                    exit: Default::default(),
+                    broadcast: Default::default(),
+                    range: Some("10.0.0.0/8".parse().unwrap()),
+                    ..Default::default()
+                },
+            )
+            .unwrap();
+        // Inside the /8, in either direction.
+        assert!(
+            table
+                .set_network(
+                    network("narrow"),
+                    NetworkRoutes {
+                        exit: Default::default(),
+                        broadcast: Default::default(),
+                        range: Some("10.13.37.0/24".parse().unwrap()),
+                        ..Default::default()
+                    }
+                )
+                .is_err()
+        );
+    }
+
+    #[test]
+    fn replacing_a_network_does_not_count_as_overlapping_itself() {
+        let table = RoutingTable::new();
+        let id = network("refresh");
+        table.set_network(id, routes()).unwrap();
+
+        let mut updated = routes();
+        updated.peers.push((addr(4), peer(4)));
+        table.set_network(id, updated).unwrap();
+        assert_eq!(table.len(), 1);
+        assert!(table.route(addr(4)).is_some());
+    }
+
+    #[test]
+    fn forgetting_a_network_takes_its_addresses_with_it() {
+        let table = RoutingTable::new();
+        let id = network("leaving");
+        table.set_network(id, routes()).unwrap();
+        table.remove_network(id);
+
+        assert!(table.is_empty());
+        assert_eq!(table.route(addr(2)), None);
+        assert!(!table.may_send_from(id, peer(2), addr(2)));
+    }
+
+    #[test]
+    fn the_local_addresses_are_what_the_interface_should_carry() {
+        let table = RoutingTable::new();
+        table.set_network(network("one"), routes()).unwrap();
+        table
+            .set_network(
+                network("two"),
+                NetworkRoutes {
+                    exit: Default::default(),
+                    broadcast: Default::default(),
+                    range: Some("10.99.0.0/16".parse().unwrap()),
+                    local: Some(Ipv4Addr::new(10, 99, 0, 1)),
+                    peers: Vec::new(),
+                },
+            )
+            .unwrap();
+
+        assert_eq!(
+            table.local_addresses(),
+            vec![(addr(1), 24), (Ipv4Addr::new(10, 99, 0, 1), 16)]
+        );
+    }
+
+    #[test]
+    fn a_network_with_no_address_yet_contributes_nothing() {
+        let table = RoutingTable::new();
+        let id = network("waiting");
+        table.set_network(id, NetworkRoutes::default()).unwrap();
+
+        assert!(table.local_addresses().is_empty());
+        assert_eq!(table.route(addr(2)), None);
+        assert_eq!(table.len(), 1, "the network is known, it just has nothing");
+    }
+
+    fn with_exit(offer: bool, via: Option<EndpointId>) -> NetworkRoutes {
+        NetworkRoutes {
+            exit: ExitPolicy { offer, via },
+            ..routes()
+        }
+    }
+
+    const INTERNET: Ipv4Addr = Ipv4Addr::new(93, 184, 216, 34);
+
+    #[test]
+    fn without_an_exit_node_the_internet_has_nowhere_to_go() {
+        let table = RoutingTable::new();
+        table.set_network(network("plain"), routes()).unwrap();
+        assert_eq!(table.route(INTERNET), None);
+        assert_eq!(table.exit_via(), None);
+    }
+
+    #[test]
+    fn with_an_exit_node_every_ordinary_destination_goes_to_it() {
+        let table = RoutingTable::new();
+        let id = network("client");
+        table
+            .set_network(id, with_exit(false, Some(peer(2))))
+            .unwrap();
+        let exit = Route {
+            network: id,
+            peer: peer(2),
+        };
+        assert_eq!(table.route(INTERNET), Some(exit));
+        assert_eq!(table.route(Ipv4Addr::new(192, 168, 1, 10)), Some(exit));
+        assert_eq!(table.exit_via(), Some((id, peer(2))));
+        // A member's own address still goes to that member.
+        assert_eq!(
+            table.route(addr(3)),
+            Some(Route {
+                network: id,
+                peer: peer(3)
+            })
+        );
+    }
+
+    #[test]
+    fn the_exit_node_is_never_the_answer_for_what_is_not_the_internet() {
+        let table = RoutingTable::new();
+        table
+            .set_network(network("client"), with_exit(false, Some(peer(2))))
+            .unwrap();
+        // Ourselves, a free address inside the overlay range, and everything
+        // that is not an ordinary unicast destination.
+        for destination in [
+            addr(1),
+            addr(200),
+            Ipv4Addr::new(127, 0, 0, 1),
+            Ipv4Addr::new(224, 0, 0, 251),
+            Ipv4Addr::new(239, 255, 255, 250),
+            Ipv4Addr::BROADCAST,
+            Ipv4Addr::UNSPECIFIED,
+            Ipv4Addr::new(169, 254, 1, 1),
+        ] {
+            assert_eq!(table.route(destination), None, "{destination}");
+        }
+    }
+
+    #[test]
+    fn another_networks_range_is_not_the_internet_either() {
+        let table = RoutingTable::new();
+        table
+            .set_network(network("client"), with_exit(false, Some(peer(2))))
+            .unwrap();
+        table
+            .set_network(
+                network("other"),
+                NetworkRoutes {
+                    range: Some("10.99.0.0/16".parse().unwrap()),
+                    local: Some(Ipv4Addr::new(10, 99, 0, 1)),
+                    ..Default::default()
+                },
+            )
+            .unwrap();
+        assert_eq!(table.route(Ipv4Addr::new(10, 99, 7, 7)), None);
+    }
+
+    #[test]
+    fn two_exit_choices_resolve_to_the_lowest_network_id() {
+        let table = RoutingTable::new();
+        let (a, b) = (network("one"), network("two"));
+        table
+            .set_network(a, with_exit(false, Some(peer(2))))
+            .unwrap();
+        table
+            .set_network(
+                b,
+                NetworkRoutes {
+                    exit: ExitPolicy {
+                        offer: false,
+                        via: Some(peer(9)),
+                    },
+                    range: Some("10.99.0.0/16".parse().unwrap()),
+                    ..Default::default()
+                },
+            )
+            .unwrap();
+        let (chosen, _) = table.exit_via().unwrap();
+        assert_eq!(chosen, a.min(b));
+        assert_eq!(table.route(INTERNET).unwrap().network, a.min(b));
+    }
+
+    #[test]
+    fn the_exit_peer_may_send_from_any_internet_address_but_nobody_else_may() {
+        let table = RoutingTable::new();
+        let id = network("client");
+        table
+            .set_network(id, with_exit(false, Some(peer(2))))
+            .unwrap();
+
+        // What comes back from the internet carries the internet's addresses.
+        assert!(table.may_send_from(id, peer(2), INTERNET));
+        // Another member still speaks only for its own address.
+        assert!(!table.may_send_from(id, peer(3), INTERNET));
+        assert!(table.may_send_from(id, peer(3), addr(3)));
+        // The exit peer cannot pass itself off as somebody else in the
+        // overlay, nor as this agent, nor as something that is not a source.
+        assert!(!table.may_send_from(id, peer(2), addr(3)));
+        assert!(!table.may_send_from(id, peer(2), addr(1)));
+        assert!(!table.may_send_from(id, peer(2), addr(200)));
+        for source in [
+            Ipv4Addr::LOCALHOST,
+            Ipv4Addr::UNSPECIFIED,
+            Ipv4Addr::new(224, 0, 0, 1),
+            Ipv4Addr::BROADCAST,
+        ] {
+            assert!(!table.may_send_from(id, peer(2), source), "{source}");
+        }
+        // Without choosing an exit node nobody gets the exception.
+        let plain = RoutingTable::new();
+        plain.set_network(id, routes()).unwrap();
+        assert!(!plain.may_send_from(id, peer(2), INTERNET));
+    }
+
+    #[test]
+    fn an_exit_node_takes_a_members_internet_packets_only_while_it_offers() {
+        let table = RoutingTable::new();
+        let id = network("server");
+        table.set_network(id, with_exit(false, None)).unwrap();
+        assert!(!table.accepts_exit_traffic(id, INTERNET));
+
+        table.set_exit(
+            id,
+            ExitPolicy {
+                offer: true,
+                via: None,
+            },
+        );
+        assert!(table.accepts_exit_traffic(id, INTERNET));
+        assert!(table.accepts_exit_traffic(id, Ipv4Addr::new(192, 168, 1, 10)));
+        // Never for another member, for this agent, a free overlay address,
+        // or anything that is not an ordinary unicast destination.
+        for destination in [
+            addr(1),
+            addr(2),
+            addr(200),
+            Ipv4Addr::LOCALHOST,
+            Ipv4Addr::new(224, 0, 0, 251),
+            Ipv4Addr::BROADCAST,
+            Ipv4Addr::UNSPECIFIED,
+        ] {
+            assert!(
+                !table.accepts_exit_traffic(id, destination),
+                "{destination}"
+            );
+        }
+        // An unknown network offers nothing.
+        assert!(!table.accepts_exit_traffic(network("elsewhere"), INTERNET));
+
+        table.set_exit(id, ExitPolicy::default());
+        assert!(!table.accepts_exit_traffic(id, INTERNET));
+    }
+
+    #[test]
+    fn offers_are_listed_by_network_with_the_range_each_agreed() {
+        let table = RoutingTable::new();
+        let (a, b) = (network("a"), network("b"));
+        table.set_network(a, with_exit(true, None)).unwrap();
+        table
+            .set_network(
+                b,
+                NetworkRoutes {
+                    exit: ExitPolicy {
+                        offer: true,
+                        via: None,
+                    },
+                    // No range agreed yet: nothing to masquerade.
+                    ..Default::default()
+                },
+            )
+            .unwrap();
+        assert_eq!(
+            table.exit_offers(),
+            vec![(a, "10.13.37.0/24".parse().unwrap())]
+        );
+        assert_eq!(
+            table.exit_settings(a),
+            Some((
+                ExitPolicy {
+                    offer: true,
+                    via: None
+                },
+                Some("10.13.37.0/24".parse().unwrap())
+            ))
+        );
+        assert_eq!(table.exit_settings(network("none")), None);
+    }
+
+    #[test]
+    fn setting_exit_for_a_network_with_no_entry_is_remembered_by_the_caller() {
+        // The runtime keeps the policy and passes it with the network when
+        // its range is agreed; a setter for a network the table does not know
+        // yet changes nothing.
+        let table = RoutingTable::new();
+        table.set_exit(
+            network("later"),
+            ExitPolicy {
+                offer: true,
+                via: None,
+            },
+        );
+        assert!(table.is_empty());
+    }
+}

crates/tsunagi-core/src/overlay/tun.rs

@@ -0,0 +1,569 @@
+//! The boundary to the operating system's packet interface.
+//!
+//! One agent has one of these, and it belongs to the agent rather than to
+//! any protocol: a packet arriving here is routed to whichever peer owns its
+//! destination, over whichever protocol has a link to that peer.
+//!
+//! Two implementations:
+//!
+//! * [`MemoryTun`] keeps packets in memory. It needs no privileges at all and
+//!   is what the test suite uses, so the entire data plane — handshake,
+//!   encryption, routing — is exercised without touching the host.
+//! * `SystemTun`, behind the `tun-device` feature, is a real TUN interface.
+//!   Creating one needs `CAP_NET_ADMIN`, and it is
+//!   [`provision`](super::provision) that holds that and creates it.
+
+use std::sync::Arc;
+
+use bytes::Bytes;
+
+use crate::BoxFuture;
+use crate::overlay::OverlayError;
+
+/// What a device should look like once created.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct TunRequest {
+    /// Interface name to ask for.
+    pub name: String,
+    /// Every overlay address this host answers to, and no others.
+    ///
+    /// One per network the agent is in — one agent has one interface, and
+    /// it carries them all. A list rather than one address because
+    /// carrying only the first is carrying only the first network: the
+    /// others have addresses the operating system has never heard of, and
+    /// their traffic goes nowhere.
+    ///
+    /// Allocated and signed at the system level, never derived from a
+    /// protocol's key: every protocol carries traffic for the same
+    /// addresses.
+    pub addresses: Vec<super::Cidr>,
+    /// Interface MTU.
+    pub mtu: u32,
+}
+
+impl TunRequest {
+    /// A request carrying nothing but a name and an MTU.
+    ///
+    /// Used where the addresses have already been applied to the host, so the
+    /// device itself only needs opening.
+    pub fn bare(name: impl Into<String>, mtu: u32) -> Self {
+        Self {
+            name: name.into(),
+            addresses: Vec::new(),
+            mtu,
+        }
+    }
+}
+
+/// A packet interface.
+///
+/// `recv` yields packets the operating system wants sent; `send` delivers
+/// packets that arrived from a peer.
+pub trait TunDevice: Send + Sync + std::fmt::Debug + 'static {
+    /// The interface name the operating system actually gave us.
+    fn name(&self) -> &str;
+
+    /// The interface MTU.
+    fn mtu(&self) -> u32;
+
+    /// The next packet the operating system wants to send, or `None` once the
+    /// device is gone.
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>>;
+
+    /// Delivers a packet to the operating system.
+    fn send<'a>(&'a self, packet: Bytes) -> BoxFuture<'a, Result<(), OverlayError>>;
+}
+
+/// Creates packet interfaces.
+pub trait TunFactory: Send + Sync + std::fmt::Debug + 'static {
+    /// A short name used in diagnostics.
+    fn name(&self) -> &str;
+
+    /// Whether what this factory creates exists on the host.
+    ///
+    /// `false` for an in-memory device: it has a name and an MTU and nothing
+    /// else — no link, no addresses, nothing for the operating system to
+    /// route through or to attach a resolver setting to. Stated rather than
+    /// guessed from the name, because a name is not evidence: an in-memory
+    /// `tsun0` and a real `tsun0` belonging to another agent look identical
+    /// from here, and configuring the operating system for the first would
+    /// land on the second.
+    fn on_host(&self) -> bool;
+
+    /// Creates a device.
+    fn create<'a>(
+        &'a self,
+        request: TunRequest,
+    ) -> BoxFuture<'a, Result<Arc<dyn TunDevice>, OverlayError>>;
+
+    /// Applies a changed request to an interface that already exists.
+    ///
+    /// The overlay IPv4 address is allocated at run time, so it can change
+    /// while the agent runs. A factory that manages the host applies that to
+    /// the live interface, without recreating it: recreating would drop every
+    /// tunnel riding on it.
+    ///
+    /// The default does nothing, which is right for a factory that only
+    /// attaches to an interface somebody else prepared.
+    fn reconfigure<'a>(&'a self, _request: TunRequest) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move { Ok(()) })
+    }
+
+    /// The host rules for this factory's interface, when it manages the host.
+    ///
+    /// `None` for an in-memory device or an interface somebody else prepared:
+    /// there is nothing of ours to route or to open in a firewall.
+    fn host_rules(&self) -> Option<Arc<dyn super::hostrules::BroadcastHostRules>> {
+        None
+    }
+
+    /// Removes an interface this factory created.
+    ///
+    /// Runs on the teardown path, so it reports rather than fails: there is
+    /// nothing useful to do about a failure at that point, and an interface
+    /// that is already gone is the desired outcome anyway.
+    fn destroy<'a>(&'a self, _name: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {})
+    }
+}
+
+/// An in-memory packet interface.
+///
+/// Nothing reaches the operating system. Packets the device "sends" can be
+/// read back with [`MemoryTun::pop_to_os`], and packets can be injected as if
+/// the operating system produced them with [`MemoryTun::push_from_os`].
+#[derive(Debug)]
+pub struct MemoryTun {
+    name: String,
+    mtu: u32,
+    from_os_tx: tokio::sync::mpsc::UnboundedSender<Bytes>,
+    from_os_rx: tokio::sync::Mutex<tokio::sync::mpsc::UnboundedReceiver<Bytes>>,
+    to_os_tx: tokio::sync::mpsc::UnboundedSender<Bytes>,
+    to_os_rx: tokio::sync::Mutex<tokio::sync::mpsc::UnboundedReceiver<Bytes>>,
+    /// Set once the device is gone, so a reader stops rather than waiting
+    /// for a packet that will never come.
+    ///
+    /// A `watch` rather than a `Notify`: closing must be seen by a reader
+    /// that has not started waiting yet as well as by one already waiting —
+    /// a device can be removed before the loop reading it has been polled
+    /// even once. The receiver is kept for the same reason: a `watch` send
+    /// with nobody subscribed does nothing at all.
+    closed: tokio::sync::watch::Sender<bool>,
+    closed_rx: tokio::sync::watch::Receiver<bool>,
+}
+
+impl MemoryTun {
+    /// Creates a device with the given name and MTU.
+    pub fn new(name: impl Into<String>, mtu: u32) -> Arc<Self> {
+        let (from_os_tx, from_os_rx) = tokio::sync::mpsc::unbounded_channel();
+        let (to_os_tx, to_os_rx) = tokio::sync::mpsc::unbounded_channel();
+        let (closed, closed_rx) = tokio::sync::watch::channel(false);
+        Arc::new(Self {
+            name: name.into(),
+            mtu,
+            from_os_tx,
+            from_os_rx: tokio::sync::Mutex::new(from_os_rx),
+            to_os_tx,
+            to_os_rx: tokio::sync::Mutex::new(to_os_rx),
+            closed,
+            closed_rx,
+        })
+    }
+
+    /// Marks the device gone, so the packet loop reading it ends.
+    ///
+    /// A real interface reports end of stream when it is removed; this is
+    /// how that happens here. Without it teardown waits out the whole grace
+    /// period for a reader that had no way of knowing.
+    pub fn close(&self) {
+        let _ = self.closed.send(true);
+    }
+
+    /// Injects a packet as if the operating system had produced it.
+    pub fn push_from_os(&self, packet: Bytes) {
+        let _ = self.from_os_tx.send(packet);
+    }
+
+    /// Takes the next packet the device delivered to the operating system.
+    pub async fn pop_to_os(&self) -> Option<Bytes> {
+        self.to_os_rx.lock().await.recv().await
+    }
+}
+
+impl TunDevice for MemoryTun {
+    fn name(&self) -> &str {
+        &self.name
+    }
+
+    fn mtu(&self) -> u32 {
+        self.mtu
+    }
+
+    fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+        Box::pin(async move {
+            let mut closed = self.closed_rx.clone();
+            if *closed.borrow_and_update() {
+                return None;
+            }
+            let mut queue = self.from_os_rx.lock().await;
+            tokio::select! {
+                packet = queue.recv() => packet,
+                _ = closed.changed() => None,
+            }
+        })
+    }
+
+    fn send<'a>(&'a self, packet: Bytes) -> BoxFuture<'a, Result<(), OverlayError>> {
+        Box::pin(async move {
+            let _ = self.to_os_tx.send(packet);
+            Ok(())
+        })
+    }
+}
+
+/// Whether an address is assigned to some interface on this host.
+///
+/// Binding a UDP socket to a specific address only succeeds when the address
+/// is local, which makes this a cheap check that needs no privileges and no
+/// platform-specific code. It does not say *which* interface has it, which is
+/// enough here: the agent chose the address, so anything else holding it is a
+/// problem in its own right.
+pub fn address_is_local(address: std::net::IpAddr) -> bool {
+    if std::net::UdpSocket::bind((address, 0)).is_ok() {
+        return true;
+    }
+    #[cfg(all(feature = "tun-device", target_os = "windows"))]
+    {
+        if let Ok(interfaces) = std::panic::catch_unwind(netdev::get_interfaces) {
+            for iface in interfaces {
+                if iface.is_up() {
+                    match address {
+                        std::net::IpAddr::V4(v4) => {
+                            if iface.ipv4.iter().any(|ip| ip.addr() == v4) {
+                                return true;
+                            }
+                        }
+                        std::net::IpAddr::V6(v6) => {
+                            if iface.ipv6.iter().any(|ip| ip.addr() == v6) {
+                                return true;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+    false
+}
+
+/// Creates [`MemoryTun`] devices.
+#[derive(Debug, Clone, Default)]
+pub struct MemoryTunFactory {
+    created: Arc<std::sync::Mutex<Vec<Arc<MemoryTun>>>>,
+}
+
+impl MemoryTunFactory {
+    /// Creates a factory.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// The device created for an interface name, if any.
+    pub fn device(&self, name: &str) -> Option<Arc<MemoryTun>> {
+        let guard = match self.created.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        };
+        guard
+            .iter()
+            .find(|device| device.name() == name)
+            .map(Arc::clone)
+    }
+
+    /// Every device created so far.
+    pub fn devices(&self) -> Vec<Arc<MemoryTun>> {
+        let guard = match self.created.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        };
+        guard.clone()
+    }
+}
+
+impl TunFactory for MemoryTunFactory {
+    fn name(&self) -> &str {
+        "memory"
+    }
+
+    /// Nothing here is on the host, which is the whole point of it.
+    fn on_host(&self) -> bool {
+        false
+    }
+
+    /// Closes the device so whatever is reading it stops.
+    ///
+    /// The device itself is kept, because a test asks what went through it
+    /// after the agent that owned it has gone.
+    fn destroy<'a>(&'a self, name: &'a str) -> BoxFuture<'a, ()> {
+        Box::pin(async move {
+            if let Some(device) = self.device(name) {
+                device.close();
+            }
+        })
+    }
+
+    fn create<'a>(
+        &'a self,
+        request: TunRequest,
+    ) -> BoxFuture<'a, Result<Arc<dyn TunDevice>, OverlayError>> {
+        Box::pin(async move {
+            let device = MemoryTun::new(request.name, request.mtu);
+            let mut guard = match self.created.lock() {
+                Ok(guard) => guard,
+                Err(poisoned) => poisoned.into_inner(),
+            };
+            guard.push(Arc::clone(&device));
+            Ok(device as Arc<dyn TunDevice>)
+        })
+    }
+}
+
+#[cfg(feature = "tun-device")]
+pub(crate) use system::open_tun;
+
+#[cfg(feature = "tun-device")]
+mod system {
+    use std::sync::Arc;
+
+    use bytes::Bytes;
+    use tokio::sync::Mutex;
+
+    use super::{TunDevice, TunRequest};
+    use crate::BoxFuture;
+    use crate::overlay::OverlayError;
+
+    /// A real TUN interface.
+    ///
+    /// On Linux it is created by opening `/dev/net/tun`, which needs
+    /// `CAP_NET_ADMIN`; on Windows it is a Wintun adapter, which needs the
+    /// process to be elevated and `wintun.dll` to be reachable. Either way
+    /// [`open_tun`] is only ever called from
+    /// [`provision`](super::super::provision), which holds whatever privilege
+    /// the platform requires for the length of the call and no longer.
+    ///
+    /// It is deliberately **not** made persistent, so the operating system
+    /// removes the interface when this value is dropped — however the process
+    /// ends. (A Wintun adapter created this way is likewise torn down when the
+    /// last handle to it closes, which the reader and writer below hold.)
+    pub struct SystemTun {
+        name: String,
+        mtu: u32,
+        reader: Mutex<tokio::io::ReadHalf<tun::AsyncDevice>>,
+        writer: Mutex<tokio::io::WriteHalf<tun::AsyncDevice>>,
+    }
+
+    impl std::fmt::Debug for SystemTun {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            f.debug_struct("SystemTun")
+                .field("name", &self.name)
+                .field("mtu", &self.mtu)
+                .finish()
+        }
+    }
+
+    impl TunDevice for SystemTun {
+        fn name(&self) -> &str {
+            &self.name
+        }
+
+        fn mtu(&self) -> u32 {
+            self.mtu
+        }
+
+        fn recv(&self) -> BoxFuture<'_, Option<Bytes>> {
+            Box::pin(async move {
+                use tokio::io::AsyncReadExt;
+                let mut buffer = vec![0u8; self.mtu as usize + 64];
+                let mut reader = self.reader.lock().await;
+                match reader.read(&mut buffer).await {
+                    Ok(0) => None,
+                    Ok(read) => {
+                        buffer.truncate(read);
+                        Some(Bytes::from(buffer))
+                    }
+                    Err(err) => {
+                        tracing::debug!(%err, "tun read failed");
+                        None
+                    }
+                }
+            })
+        }
+
+        fn send<'a>(&'a self, packet: Bytes) -> BoxFuture<'a, Result<(), OverlayError>> {
+            Box::pin(async move {
+                use tokio::io::AsyncWriteExt;
+                let mut writer = self.writer.lock().await;
+                writer
+                    .write_all(&packet)
+                    .await
+                    .map_err(|err| OverlayError::Other(format!("tun write failed: {err}")))
+            })
+        }
+    }
+
+    /// Creates the TUN interface by opening it.
+    ///
+    /// Synchronous, and deliberately so: the caller holds a capability guard
+    /// across this call, and such a guard must not span an `await` because
+    /// Linux capabilities are per thread.
+    pub(crate) fn open_tun(request: &TunRequest) -> Result<Arc<dyn TunDevice>, OverlayError> {
+        // macOS mutates neither field below (it neither names the interface nor
+        // sets a platform config), so `mut` is unused only there.
+        #[cfg_attr(target_os = "macos", allow(unused_mut))]
+        let mut config = tun::Configuration::default();
+        // macOS names its tunnels `utunN` and assigns the unit number itself.
+        // A name derived from the network id (`tsun…`) is not a legal utun
+        // name, so there the kernel is left to choose and the real name is read
+        // back below. Linux and Windows take the requested name verbatim.
+        #[cfg(not(target_os = "macos"))]
+        config.tun_name(&request.name);
+        #[cfg(target_os = "linux")]
+        config.platform_config(|platform| {
+            // The crate's own root check is not the check we want: this holds
+            // CAP_NET_ADMIN without being root. Whether the open succeeds is
+            // the honest answer.
+            platform.ensure_root_privileges(false);
+        });
+        // Packet information stays off, so reads and writes are raw IP
+        // packets. `ip tuntap add ... mode tun` also defaults to no packet
+        // information, so the flags match when attaching to one. The address
+        // and MTU are left off the configuration on purpose: the provisioner
+        // applies those to the interface separately, the same way on every
+        // platform, so the plan stays the single source of what it carries.
+
+        let device = tun::create_as_async(&config).map_err(|err| open_error(&request.name, err))?;
+
+        // The interface name the operating system actually gave us: the
+        // requested one everywhere except macOS, where the kernel chose the
+        // utun unit and so the real name has to be read back off the device.
+        #[cfg(target_os = "macos")]
+        let name = {
+            use tun::AbstractDevice;
+            device.tun_name().unwrap_or_else(|_| request.name.clone())
+        };
+        #[cfg(not(target_os = "macos"))]
+        let name = request.name.clone();
+
+        let (reader, writer) = tokio::io::split(device);
+        Ok(Arc::new(SystemTun {
+            name,
+            mtu: request.mtu,
+            reader: Mutex::new(reader),
+            writer: Mutex::new(writer),
+        }) as Arc<dyn TunDevice>)
+    }
+
+    fn open_error(name: &str, error: tun::Error) -> OverlayError {
+        // Wintun preserves the Windows code in its I/O variant. Normalize it
+        // before testing permissions; matching localized error text would
+        // miss the same failure on another Windows language.
+        #[cfg(target_os = "windows")]
+        let error = match error {
+            tun::Error::WintunError(error) => tun::Error::Io(error.into()),
+            error => error,
+        };
+        let denied = match &error {
+            tun::Error::Io(error) => {
+                error.kind() == std::io::ErrorKind::PermissionDenied
+                    || (cfg!(target_os = "windows")
+                        && matches!(error.raw_os_error(), Some(5 | 740 | 1314)))
+            }
+            _ => false,
+        };
+        if denied {
+            let hint = if cfg!(target_os = "windows") {
+                "Open PowerShell or Command Prompt with 'Run as administrator' and \
+                 start `tsng up` there."
+            } else if cfg!(target_os = "linux") {
+                "Start `tsng up` as root or grant CAP_NET_ADMIN with \
+                 `sudo setcap cap_net_admin,cap_net_bind_service+p /path/to/tsunagi`."
+            } else if cfg!(target_os = "macos") {
+                "Creating a utun interface needs root on macOS; start `tsng up` \
+                 with `sudo`, or install it as a root LaunchDaemon."
+            } else {
+                "Start `tsng up` with the privileges required to create a TUN interface."
+            };
+            return OverlayError::Unavailable(format!(
+                "permission denied while creating the TUN interface `{name}`. \
+                 {hint} OS error: {error}"
+            ));
+        }
+        OverlayError::Unavailable(format!(
+            "cannot create the TUN interface `{name}`: {error}. {}",
+            open_hint()
+        ))
+    }
+
+    /// The platform-appropriate tail of a "cannot create the interface" error:
+    /// what privilege it needs and the fallback that always works.
+    fn open_hint() -> &'static str {
+        #[cfg(target_os = "linux")]
+        {
+            "Creating one needs CAP_NET_ADMIN; grant it with `setcap cap_net_admin,cap_net_bind_service+p`, \
+             or run with `--no-tun` to keep the tunnels off the operating system."
+        }
+        #[cfg(target_os = "windows")]
+        {
+            "Creating one needs an elevated process and `wintun.dll` on the search path; \
+             run as Administrator with the DLL beside the executable, or run with `--no-tun` \
+             to keep the tunnels off the operating system."
+        }
+        #[cfg(target_os = "macos")]
+        {
+            "Creating a utun interface needs root; start the agent with `sudo` or as a root \
+             LaunchDaemon, or run with `--no-tun` to keep the tunnels off the operating system."
+        }
+        #[cfg(not(any(target_os = "linux", target_os = "windows", target_os = "macos")))]
+        {
+            "Run with `--no-tun` to keep the tunnels off the operating system."
+        }
+    }
+
+    #[cfg(test)]
+    mod tests {
+        use super::*;
+
+        #[test]
+        fn a_tun_permission_error_explains_how_to_start_the_agent() {
+            let error = tun::Error::Io(std::io::Error::from(std::io::ErrorKind::PermissionDenied));
+            let message = open_error("tsun0", error).to_string();
+            assert!(message.contains("permission denied"), "{message}");
+            assert!(message.contains("tsun0"), "{message}");
+            assert!(message.contains("tsng up"), "{message}");
+            #[cfg(target_os = "windows")]
+            assert!(message.contains("Run as administrator"), "{message}");
+            #[cfg(target_os = "linux")]
+            assert!(message.contains("CAP_NET_ADMIN"), "{message}");
+        }
+
+        #[cfg(target_os = "windows")]
+        #[test]
+        fn wintun_access_and_elevation_errors_get_the_permission_hint() {
+            for code in [5, 740, 1314] {
+                let error = tun::Error::WintunError(std::io::Error::from_raw_os_error(code).into());
+                let message = open_error("tsun0", error).to_string();
+                assert!(message.contains("permission denied"), "{message}");
+                assert!(message.contains("Run as administrator"), "{message}");
+                assert!(message.contains(&format!("os error {code}")), "{message}");
+                assert!(!message.contains("wintun.dll"), "{message}");
+            }
+        }
+
+        #[test]
+        fn other_tun_errors_keep_their_cause_without_claiming_access_was_denied() {
+            let message = open_error("tsun0", tun::Error::InvalidName).to_string();
+            assert!(message.contains("invalid device tun name"), "{message}");
+            assert!(!message.contains("permission denied"), "{message}");
+        }
+    }
+}

crates/tsunagi-core/src/proto/message.rs

@@ -20,7 +20,7 @@ use crate::error::ProtocolError;
 /// The version in the ALPN is the wire-compatibility version of the control
 /// protocol. It is independent of the network identity scheme version, so
 /// bumping it must not change any existing [`crate::NetworkId`].
-pub const ALPN: &[u8] = b"tsunagi/ctrl/1";
+pub const ALPN: &[u8] = b"tsunagi/ctrl/5";
 
 /// ALPN of the tsunagi data plane.
 ///
@@ -28,7 +28,13 @@ pub const ALPN: &[u8] = b"tsunagi/ctrl/1";
 /// They carry one IP plugin's packets for one network and nothing else, so a
 /// saturated or broken data plane cannot disturb control traffic, and the
 /// transport underneath can be replaced without touching the control protocol.
-pub const DATA_ALPN: &[u8] = b"tsunagi/data/1";
+///
+/// Version 4 adds source, destination, hop limit and stable flow id above
+/// transport fragmentation. Version 5 adds the end-to-end protocol to that
+/// envelope, so a tunnel no longer depends on the kind of link under it, and
+/// a first-claimed time to a signed hostname claim. All members must upgrade
+/// together. Persistent identities and network configuration do not change.
+pub const DATA_ALPN: &[u8] = b"tsunagi/data/5";
 
 /// Largest plugin protocol identifier accepted when opening a data channel.
 pub const MAX_DATA_PROTOCOL_LEN: usize = 32;
@@ -37,7 +43,7 @@ pub const MAX_DATA_PROTOCOL_LEN: usize = 32;
 pub const MAX_SIGNATURE_LEN: usize = 64;
 
 /// Control protocol version carried inside the handshake.
-pub const PROTOCOL_VERSION: u16 = 1;
+pub const PROTOCOL_VERSION: u16 = 5;
 
 /// First message of the handshake, sent by the initiator.
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
@@ -69,6 +75,13 @@ pub struct AuthProof {
 /// What this agent tells a peer about itself.
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct Announcement {
+    /// Local opt-in to receive LAN broadcasts in this network.
+    pub broadcast: bool,
+    /// This agent offers itself as an exit node in this network: members may
+    /// send all their internet traffic through it. Off unless its owner
+    /// turned it on, and only said once the rules that make it work are in
+    /// place.
+    pub exit_node: bool,
     /// Human-readable hostname. A mutable binding, not an identity.
     pub hostname: String,
     /// Announced IP plugin capabilities. Opaque to the core.
@@ -81,6 +94,8 @@ pub struct Announcement {
 pub struct DataOpen {
     /// Which IP plugin's packets this channel will carry.
     pub protocol: String,
+    /// Maximum reassembled payload the initiator is willing to receive.
+    pub max_datagram: u32,
 }
 
 /// The responder's answer to [`DataOpen`].
@@ -124,6 +139,32 @@ pub enum ControlMessage {
         /// The records. Bounded by [`crate::config::Limits::max_state_records`].
         records: Vec<crate::state::SignedRecord>,
     },
+    /// Members this sender knows of, and where it has seen them.
+    ///
+    /// An introduction, not a vouching: everything here is an unverified
+    /// candidate, and membership is still decided by the handshake. It is
+    /// what turns a star around whoever was named on the command line into
+    /// a mesh — a device that was told about one member ends up talking to
+    /// all of them — and it is the same shape a lookup in a distributed
+    /// hash table would return.
+    Peers {
+        /// The members, with whatever addresses the sender has for them.
+        peers: Vec<PeerHint>,
+    },
+    /// Which peers this sender has a live data link with, right now.
+    ///
+    /// First-hand and nothing else: a sender speaks only for itself, never
+    /// about what somebody else can reach. That is what makes it usable
+    /// without weighing hearsay — the claim is proved or disproved by
+    /// sending through it.
+    ///
+    /// A snapshot rather than a change, because a snapshot is idempotent
+    /// and repairs itself; and volatile, so it lives here and not in
+    /// signed state, which is for what has to survive a member being away.
+    Reachable {
+        /// Direct links, scoped to the plugin protocol they actually carry.
+        links: Vec<ReachableLink>,
+    },
     /// Graceful goodbye.
     ///
     /// A peer going away is not a revocation of anything.
@@ -133,6 +174,38 @@ pub enum ControlMessage {
     },
 }
 
+/// Somewhere a member has been seen.
+///
+/// Addresses are in iroh's own text form, which is what the endpoint takes
+/// back: this is a hint to try, never a fact about where anybody is.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct PeerHint {
+    /// The member's endpoint key.
+    pub endpoint: [u8; 32],
+    /// Where the sender has seen it, as `ip:<socket>` or `relay:<url>`.
+    ///
+    /// The same spelling the disposable cache uses, so one decoder serves
+    /// both and neither can drift from the other. An empty list is still
+    /// worth sending: an id alone is enough for a receiver whose endpoint
+    /// can resolve it.
+    pub addrs: Vec<String>,
+}
+
+/// One directly observed edge, advertised only by its authenticated source.
+#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize)]
+pub struct ReachableLink {
+    /// The other end of this transport link.
+    pub peer: [u8; 32],
+    /// Protocol carried by the link.
+    pub protocol: String,
+}
+
+/// Longest address text accepted in a hint.
+pub const MAX_HINT_ADDR_LEN: usize = 128;
+
+/// Most addresses accepted for one member in a hint.
+pub const MAX_HINT_ADDRS: usize = 8;
+
 /// A control message together with the network it belongs to.
 ///
 /// Every session is bound to exactly one network at handshake time. The
@@ -191,6 +264,28 @@ pub fn validate_capability(
 /// network, the other networks or the agent.
 pub fn validate(message: &ControlMessage, limits: &Limits) -> Result<(), ProtocolError> {
     match message {
+        ControlMessage::Peers { peers } => {
+            check_len("peers", peers.len(), limits.max_state_records)?;
+            for hint in peers {
+                check_len("peers.addrs", hint.addrs.len(), MAX_HINT_ADDRS)?;
+                for addr in &hint.addrs {
+                    check_len("peers.addr", addr.len(), MAX_HINT_ADDR_LEN)?;
+                }
+            }
+        }
+        ControlMessage::Reachable { links } => {
+            check_len("reachable.links", links.len(), limits.max_state_records)?;
+            for link in links {
+                if link.protocol.is_empty() {
+                    return Err(ProtocolError::Malformed("empty routing protocol"));
+                }
+                check_len(
+                    "reachable.protocol",
+                    link.protocol.len(),
+                    MAX_PROTOCOL_ID_LEN,
+                )?;
+            }
+        }
         ControlMessage::Announce(announcement) => {
             check_len(
                 "announce.hostname",
@@ -229,6 +324,8 @@ pub fn kind(message: &ControlMessage) -> &'static str {
         ControlMessage::Ping { .. } => "ping",
         ControlMessage::Pong { .. } => "pong",
         ControlMessage::State { .. } => "state",
+        ControlMessage::Reachable { .. } => "reachable",
+        ControlMessage::Peers { .. } => "peers",
         ControlMessage::Bye { .. } => "bye",
     }
 }

crates/tsunagi-core/src/state/mod.rs

@@ -120,8 +120,37 @@ pub const DEFAULT_IPV4_RANGE: Ipv4Range = Ipv4Range {
     prefix_len: 24,
 };
 
+/// A range derived from a network's own identity.
+///
+/// One agent has one interface, so two of its networks cannot both use the
+/// configured range. The second needs another one — and it cannot be picked
+/// locally, because every member has to arrive at the same answer without
+/// being told. So it comes from the network id, which every member already
+/// has and nobody can choose: the same network derives the same range on
+/// every device.
+///
+/// Inside 10/8 like the default, and never equal to it, so the first
+/// network keeps what it has always had.
+pub fn derived_ipv4_range(network: NetworkId) -> Ipv4Range {
+    let bytes = network.as_bytes();
+    let second = bytes[0];
+    let third = bytes[1];
+    let candidate = Ipv4Addr::new(10, second, third, 0);
+    // The default is somebody's already. One step along is still derived
+    // from the id and still the same everywhere.
+    let base = if candidate == DEFAULT_IPV4_RANGE.base {
+        Ipv4Addr::new(10, second, third.wrapping_add(1), 0)
+    } else {
+        candidate
+    };
+    Ipv4Range {
+        base,
+        prefix_len: 24,
+    }
+}
+
 /// Frozen domain separator for the bytes a record signature covers.
-pub const RECORD_DOMAIN: &str = "tsunagi-signed-record-v2";
+pub const RECORD_DOMAIN: &str = "tsunagi-signed-record-v3";
 
 /// Longest hostname a record may carry.
 ///
@@ -178,6 +207,18 @@ pub enum RecordBody {
         range: Option<Ipv4Range>,
         /// The name this author answers to.
         hostname: Option<String>,
+        /// When this author first claimed that name, in milliseconds since
+        /// the Unix epoch, kept across every later version that keeps the
+        /// name. It is what makes a name belong to whoever claimed it first
+        /// rather than to whoever happens to sort lowest, so a member that
+        /// arrives later with the same name cannot take it over.
+        ///
+        /// The author's own word: a member with a wrong clock can claim an
+        /// earlier time. That costs it nothing it could not do by other
+        /// means, since anybody who knows the network secret is a member,
+        /// and it cannot take an address or reach anything it was not
+        /// already entitled to.
+        hostname_since: Option<u64>,
     },
     /// This author gives up everything it claimed.
     ///
@@ -214,6 +255,14 @@ impl RecordBody {
         }
     }
 
+    /// When the author first claimed the hostname, if it has one.
+    pub fn hostname_since(&self) -> Option<u64> {
+        match self {
+            RecordBody::Claim { hostname_since, .. } => *hostname_since,
+            RecordBody::Release => None,
+        }
+    }
+
     fn canonical(&self) -> Vec<u8> {
         let mut out = Vec::with_capacity(96);
         match self {
@@ -221,6 +270,7 @@ impl RecordBody {
                 address,
                 range,
                 hostname,
+                hostname_since,
             } => {
                 push_lp(&mut out, b"claim");
                 push_opt(
@@ -241,6 +291,13 @@ impl RecordBody {
                         .as_deref(),
                 );
                 push_opt(&mut out, hostname.as_deref().map(str::as_bytes));
+                push_opt(
+                    &mut out,
+                    hostname_since
+                        .map(|since| since.to_be_bytes())
+                        .as_ref()
+                        .map(|bytes| &bytes[..]),
+                );
             }
             RecordBody::Release => {
                 push_lp(&mut out, b"release");
@@ -362,8 +419,12 @@ impl SignedRecord {
             address,
             range,
             hostname,
+            hostname_since,
         } = &self.body
         {
+            if hostname.is_none() && hostname_since.is_some() {
+                return Err(StateError::Malformed("a time was claimed for no hostname"));
+            }
             // Bounds before anything is believed, because all of this came
             // off the network.
             if let Some(range) = range {
@@ -556,25 +617,33 @@ impl StateSet {
 
     /// Who currently holds each claimed hostname.
     ///
-    /// The same rule as addresses, for the same reason: a name is owned, two
+    /// A name belongs to the member that claimed it first, and the others
-    /// members may claim one, and every replica has to reach the same answer
+    /// that claim it keep working by address but are not found by that name.
-    /// about who has it without asking anybody.
+    /// "First" is the time each author signed into its own claim, with the
+    /// lower id deciding a tie and a claim with no time counting as last, so
+    /// every replica reaches the same answer about who has it without
+    /// asking anybody. It is a pure function of the records: nothing depends
+    /// on the order they arrived in, so there is no race to lose.
     pub fn hostname_holders(&self) -> HashMap<String, EndpointId> {
-        let mut holders: HashMap<String, EndpointId> = HashMap::new();
+        let mut holders: HashMap<String, (u64, EndpointId)> = HashMap::new();
         for (author, record) in &self.records {
             let Some(hostname) = record.body.hostname() else {
                 continue;
             };
+            let claim = (record.body.hostname_since().unwrap_or(u64::MAX), *author);
             holders
                 .entry(hostname.to_string())
                 .and_modify(|held| {
-                    if author.as_bytes() < held.as_bytes() {
+                    if (claim.0, claim.1.as_bytes()) < (held.0, held.1.as_bytes()) {
-                        *held = *author;
+                        *held = claim;
                     }
                 })
-                .or_insert(*author);
+                .or_insert(claim);
         }
         holders
+            .into_iter()
+            .map(|(hostname, (_, author))| (hostname, author))
+            .collect()
     }
 
     /// The hostname an author holds, if it holds one uncontested.
@@ -583,6 +652,19 @@ impl StateSet {
         (self.hostname_holders().get(hostname) == Some(author)).then_some(hostname)
     }
 
+    /// The name an author claimed and the member that holds it instead, when
+    /// somebody who claimed it earlier does.
+    pub fn hostname_taken(&self, author: &EndpointId) -> Option<(&str, EndpointId)> {
+        let hostname = self.records.get(author)?.body.hostname()?;
+        let holder = *self.hostname_holders().get(hostname)?;
+        (holder != *author).then_some((hostname, holder))
+    }
+
+    /// When an author first claimed the name it holds now, if it has one.
+    pub fn hostname_since_of(&self, author: &EndpointId) -> Option<u64> {
+        self.records.get(author)?.body.hostname_since()
+    }
+
     /// The address an author holds, if it holds one uncontested.
     pub fn address_of(&self, author: &EndpointId) -> Option<Ipv4Addr> {
         let address = self.records.get(author)?.body.claimed_address()?;
@@ -632,6 +714,7 @@ mod tests {
             address: Some(address.parse().unwrap()),
             range: Some(range()),
             hostname: None,
+            hostname_since: None,
         }
     }
 
@@ -691,6 +774,7 @@ mod tests {
                 address: Some("10.99.0.1".parse().unwrap()),
                 range: Some(range()),
                 hostname: None,
+                hostname_since: None,
             },
         );
         assert!(matches!(outside.verify(id), Err(StateError::Malformed(_))));
@@ -706,6 +790,7 @@ mod tests {
                     prefix_len: 31,
                 }),
                 hostname: None,
+                hostname_since: None,
             },
         );
         assert!(matches!(no_hosts.verify(id), Err(StateError::Malformed(_))));
@@ -747,6 +832,7 @@ mod tests {
                     address: None,
                     range: None,
                     hostname: Some(bad.to_string()),
+                    hostname_since: None,
                 },
             );
             assert!(
@@ -768,6 +854,7 @@ mod tests {
                 address: Some("10.13.37.5".parse().unwrap()),
                 range: None,
                 hostname: None,
+                hostname_since: None,
             },
         );
         assert!(matches!(record.verify(id), Err(StateError::Malformed(_))));
@@ -789,6 +876,7 @@ mod tests {
                     address: None,
                     range: None,
                     hostname: Some("music".into()),
+                    hostname_since: None,
                 },
             )
         };
@@ -831,6 +919,7 @@ mod tests {
                     address: None,
                     range: None,
                     hostname: Some(name.to_string()),
+                    hostname_since: None,
                 },
             )
         };
@@ -850,6 +939,109 @@ mod tests {
         assert_eq!(set.hostname_of(&secret.public()), Some("new"));
     }
 
+    fn named_at(
+        secret: &SecretKey,
+        id: NetworkId,
+        version: u64,
+        name: &str,
+        since: u64,
+    ) -> SignedRecord {
+        SignedRecord::sign(
+            secret,
+            id,
+            version,
+            RecordBody::Claim {
+                address: None,
+                range: None,
+                hostname: Some(name.into()),
+                hostname_since: Some(since),
+            },
+        )
+    }
+
+    #[test]
+    fn a_name_belongs_to_whoever_claimed_it_first_whatever_their_ids_or_arrival_order() {
+        let id = network("naming");
+        let mut keys = [SecretKey::generate(), SecretKey::generate()];
+        keys.sort_by_key(|key| *key.public().as_bytes());
+        let [low, high] = keys;
+
+        // The member with the higher id claimed first, so it holds the name,
+        // and the lower id does not take it back by sorting first.
+        let early = named_at(&high, id, 1, "music", 1_000);
+        let late = named_at(&low, id, 1, "music", 2_000);
+        for order in [[&early, &late], [&late, &early]] {
+            let mut set = StateSet::new();
+            for record in order {
+                set.merge(id, (*record).clone()).unwrap();
+            }
+            assert_eq!(set.hostname_of(&high.public()), Some("music"));
+            assert_eq!(set.hostname_of(&low.public()), None);
+            assert_eq!(
+                set.hostname_taken(&low.public()),
+                Some(("music", high.public()))
+            );
+            assert_eq!(set.hostname_taken(&high.public()), None);
+        }
+
+        // At the same instant the lower id decides, the same way everywhere.
+        let mut set = StateSet::new();
+        set.merge(id, named_at(&high, id, 1, "music", 5)).unwrap();
+        set.merge(id, named_at(&low, id, 1, "music", 5)).unwrap();
+        assert_eq!(set.hostname_of(&low.public()), Some("music"));
+        assert_eq!(set.hostname_of(&high.public()), None);
+    }
+
+    #[test]
+    fn a_name_that_loses_is_found_again_once_the_holder_lets_it_go() {
+        let id = network("naming");
+        let first = SecretKey::generate();
+        let second = SecretKey::generate();
+        let mut set = StateSet::new();
+        set.merge(id, named_at(&first, id, 1, "music", 1)).unwrap();
+        set.merge(id, named_at(&second, id, 1, "music", 2)).unwrap();
+        assert_eq!(set.hostname_of(&second.public()), None);
+
+        // Renaming starts the clock afresh, so it cannot be used to jump the
+        // queue for a name somebody else already holds.
+        set.merge(id, named_at(&first, id, 2, "other", 3)).unwrap();
+        assert_eq!(set.hostname_of(&second.public()), Some("music"));
+        set.merge(id, named_at(&first, id, 3, "music", 4)).unwrap();
+        assert_eq!(set.hostname_of(&second.public()), Some("music"));
+        assert_eq!(set.hostname_of(&first.public()), None);
+
+        // And a release hands it on.
+        set.merge(id, SignedRecord::sign(&second, id, 2, RecordBody::Release))
+            .unwrap();
+        assert_eq!(set.hostname_of(&first.public()), Some("music"));
+    }
+
+    #[test]
+    fn a_time_with_no_name_is_not_a_valid_claim() {
+        let id = network("naming");
+        let secret = SecretKey::generate();
+        let record = SignedRecord::sign(
+            &secret,
+            id,
+            1,
+            RecordBody::Claim {
+                address: None,
+                range: None,
+                hostname: None,
+                hostname_since: Some(1),
+            },
+        );
+        assert!(matches!(record.verify(id), Err(StateError::Malformed(_))));
+
+        // The time is signed: moving it breaks the signature, so nobody can
+        // improve another member's place in the queue.
+        let mut forged = named_at(&secret, id, 1, "music", 9_000);
+        if let RecordBody::Claim { hostname_since, .. } = &mut forged.body {
+            *hostname_since = Some(1);
+        }
+        assert_eq!(forged.verify(id).unwrap_err(), StateError::BadSignature);
+    }
+
     #[test]
     fn a_release_gives_up_the_name_as_well_as_the_address() {
         let secret = SecretKey::generate();
@@ -865,6 +1057,7 @@ mod tests {
                     address: Some("10.13.37.5".parse().unwrap()),
                     range: Some(range()),
                     hostname: Some("music".into()),
+                    hostname_since: None,
                 },
             ),
         )
@@ -1028,6 +1221,7 @@ mod tests {
                     address: Some("10.99.0.7".parse().unwrap()),
                     range: Some(custom),
                     hostname: None,
+                    hostname_since: None,
                 },
             ),
         )
@@ -1053,4 +1247,32 @@ mod tests {
             SignedRecord::canonical_bytes(network("other"), author, 1, &claim("10.13.37.5"))
         );
     }
+
+    #[test]
+    fn a_derived_range_is_the_same_wherever_it_is_derived() {
+        // Every member has to arrive at it without being told, so it comes
+        // from the one thing they all already agree on.
+        let network = NetworkKeys::derive(
+            &NetworkName::new("second").unwrap(),
+            &NetworkSecret::from_bytes([7u8; 32]).unwrap(),
+        )
+        .network_id();
+        let once = derived_ipv4_range(network);
+        assert_eq!(once, derived_ipv4_range(network));
+        assert_eq!(once.prefix_len, 24);
+        assert_eq!(once.base.octets()[0], 10, "inside 10/8 like the default");
+        assert_ne!(
+            once, DEFAULT_IPV4_RANGE,
+            "the default belongs to whichever network asked first"
+        );
+
+        // A different network derives a different range, which is the
+        // whole point of deriving it.
+        let other = NetworkKeys::derive(
+            &NetworkName::new("third").unwrap(),
+            &NetworkSecret::from_bytes([9u8; 32]).unwrap(),
+        )
+        .network_id();
+        assert_ne!(once, derived_ipv4_range(other));
+    }
 }

crates/tsunagi-core/src/storage/mod.rs

@@ -56,8 +56,13 @@ fn restrict_permissions(_file: &std::fs::File, _path: &Path) -> Result<()> {
     Ok(())
 }
 
+/// Restricts a file to the user that owns it.
+///
+/// Part of what the system level offers a protocol: a plugin keeping keys of
+/// its own on disk has the same obligation as the agent, and should not have
+/// to work out the platform details again to meet it.
 #[cfg(unix)]
-pub(crate) fn restrict_path_permissions(path: &Path) -> Result<()> {
+pub fn restrict_path_permissions(path: &Path) -> Result<()> {
     use std::os::unix::fs::PermissionsExt;
     std::fs::set_permissions(path, std::fs::Permissions::from_mode(0o600)).map_err(|source| {
         Error::Io {
@@ -67,13 +72,15 @@ pub(crate) fn restrict_path_permissions(path: &Path) -> Result<()> {
     })
 }
 
+/// Restricts a file to the user that owns it. A no-op off Unix.
 #[cfg(not(unix))]
-pub(crate) fn restrict_path_permissions(_path: &Path) -> Result<()> {
+pub fn restrict_path_permissions(_path: &Path) -> Result<()> {
     Ok(())
 }
 
+/// Restricts a directory to the user that owns it.
 #[cfg(unix)]
-pub(crate) fn restrict_dir_permissions(path: &Path) -> Result<()> {
+pub fn restrict_dir_permissions(path: &Path) -> Result<()> {
     use std::os::unix::fs::PermissionsExt;
     std::fs::set_permissions(path, std::fs::Permissions::from_mode(0o700)).map_err(|source| {
         Error::Io {
@@ -83,12 +90,14 @@ pub(crate) fn restrict_dir_permissions(path: &Path) -> Result<()> {
     })
 }
 
+/// Restricts a directory to the user that owns it. A no-op off Unix.
 #[cfg(not(unix))]
-pub(crate) fn restrict_dir_permissions(_path: &Path) -> Result<()> {
+pub fn restrict_dir_permissions(_path: &Path) -> Result<()> {
     Ok(())
 }
 
-pub(crate) fn create_dir(path: &Path) -> Result<()> {
+/// Creates a directory and every parent, restricted to this user.
+pub fn create_dir(path: &Path) -> Result<()> {
     std::fs::create_dir_all(path).map_err(|source| Error::Io {
         path: path.to_path_buf(),
         source,
@@ -239,6 +248,24 @@ impl Storage {
             .await
     }
 
+    /// Persists local broadcast participation for one configured network.
+    pub async fn set_broadcast(&self, network_id: NetworkId, enabled: bool) -> Result<()> {
+        self.with_state(move |state| state.set_broadcast(network_id, enabled))
+            .await
+    }
+
+    /// Persists whether this agent offers itself as an exit node in a network.
+    pub async fn set_exit_node(&self, network_id: NetworkId, enabled: bool) -> Result<()> {
+        self.with_state(move |state| state.set_exit_node(network_id, enabled))
+            .await
+    }
+
+    /// Persists the member this device sends its internet traffic through.
+    pub async fn set_exit_via(&self, network_id: NetworkId, via: Option<[u8; 32]>) -> Result<()> {
+        self.with_state(move |state| state.set_exit_via(network_id, via.as_ref()))
+            .await
+    }
+
     /// Updates the auto-start flag of a network.
     pub async fn set_auto_start(&self, network_id: NetworkId, auto_start: bool) -> Result<()> {
         self.with_state(move |state| state.set_auto_start(network_id, auto_start))
@@ -249,7 +276,8 @@ impl Storage {
     pub async fn remove_network(&self, network_id: NetworkId) -> Result<()> {
         self.with_state(move |state| {
             state.remove_network(network_id)?;
-            state.forget_signed_records(network_id)
+            state.forget_signed_records(network_id)?;
+            state.forget_peer_hostnames(network_id)
         })
         .await?;
         self.with_cache((), move |cache| cache.forget_network(network_id))
@@ -268,6 +296,26 @@ impl Storage {
             .await
     }
 
+    /// Remembers the hostname a member announced, so it can still be named
+    /// once it is away.
+    pub async fn remember_peer_hostname(
+        &self,
+        network_id: NetworkId,
+        endpoint_id: [u8; 32],
+        hostname: String,
+    ) -> Result<()> {
+        self.with_state(move |state| {
+            state.remember_peer_hostname(network_id, &endpoint_id, &hostname)
+        })
+        .await
+    }
+
+    /// The hostnames members of a network last announced.
+    pub async fn peer_hostnames(&self, network_id: NetworkId) -> Result<Vec<([u8; 32], String)>> {
+        self.with_state(move |state| state.peer_hostnames(network_id))
+            .await
+    }
+
     /// Records an address hint. Failures are non-fatal.
     pub async fn record_hint(
         &self,
@@ -341,3 +389,137 @@ impl Storage {
         guard.take();
     }
 }
+
+/// What a wipe covers: every entry in the two directories.
+///
+/// Reported before anything is removed, and again afterwards, because a
+/// command whose whole job is to destroy state has to be able to say exactly
+/// what it is about to destroy.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct WipePlan {
+    /// Entries in the state directory, which is the mandatory state and
+    /// anything a plugin keeps beside it.
+    pub state: Vec<std::path::PathBuf>,
+    /// Entries in the cache directory.
+    pub cache: Vec<std::path::PathBuf>,
+}
+
+impl WipePlan {
+    /// Whether there is anything to remove at all.
+    pub fn is_empty(&self) -> bool {
+        self.state.is_empty() && self.cache.is_empty()
+    }
+
+    /// Every entry, state first.
+    pub fn entries(&self) -> impl Iterator<Item = &std::path::Path> {
+        self.state
+            .iter()
+            .chain(self.cache.iter())
+            .map(std::path::PathBuf::as_path)
+    }
+}
+
+/// Lists what [`wipe`] would remove, without removing anything.
+pub fn wipe_plan(paths: &StoragePaths) -> Result<WipePlan> {
+    Ok(WipePlan {
+        state: removable(&paths.state_dir, &paths.state_db())?,
+        cache: removable(&paths.cache_dir, &paths.cache_db())?,
+    })
+}
+
+/// Removes everything in both directories, leaving this device a stranger.
+///
+/// The device identity, every network it belongs to, every signed record and
+/// everything a plugin kept beside them. The next start generates a new
+/// identity and knows nothing, which is the point: after this there is no
+/// membership left to be surprised by.
+///
+/// What it is **not** is a goodbye. Other members keep the claims this device
+/// signed, because signed state has no expiry and there is nobody left here
+/// to sign a release. Leaving each network first with
+/// [`crate::Agent::leave_network`] is what frees the address and the name.
+///
+/// Refuses while an agent owns the state directory, and refuses a directory
+/// with no `state.sqlite` in it — a wipe pointed at the wrong place by a
+/// mistyped `--state-dir` would otherwise remove somebody's documents.
+pub fn wipe(paths: &StoragePaths) -> Result<WipePlan> {
+    let plan = wipe_plan(paths)?;
+    // Held across the removal, so an agent cannot start into a directory
+    // that is half gone. The lock file is itself removed at the end: on
+    // Unix the lock lives on the open handle, not on the name.
+    let lock = DirectoryLock::acquire(paths.lock_file())?;
+
+    for entry in plan.entries() {
+        remove_entry(entry)?;
+    }
+    drop(lock);
+    // Created by acquiring the lock a moment ago, so it is ours to remove
+    // and nothing is left behind claiming the directory.
+    let _ = std::fs::remove_file(paths.lock_file());
+    Ok(plan)
+}
+
+/// Every entry of `dir`, once `marker` proves it is one of ours.
+///
+/// An empty or missing directory is nothing to remove rather than an error:
+/// wiping twice must be as ordinary as wiping once.
+fn removable(dir: &Path, marker: &Path) -> Result<Vec<std::path::PathBuf>> {
+    if !dir.exists() {
+        return Ok(Vec::new());
+    }
+    let mut entries = Vec::new();
+    let read = std::fs::read_dir(dir).map_err(|source| Error::Io {
+        path: dir.to_path_buf(),
+        source,
+    })?;
+    for entry in read {
+        let entry = entry.map_err(|source| Error::Io {
+            path: dir.to_path_buf(),
+            source,
+        })?;
+        entries.push(entry.path());
+    }
+    if entries.is_empty() {
+        return Ok(entries);
+    }
+
+    // The marker is what tells a state directory from a directory that
+    // merely got named as one. The lock file counts too: an agent that
+    // crashed before writing anything still leaves that.
+    let ours = marker.exists() || dir.join("state.lock").exists();
+    if !ours {
+        return Err(Error::Storage(format!(
+            "{} does not look like a tsunagi directory: no {} in it. Nothing was removed.",
+            dir.display(),
+            marker.file_name().map_or_else(
+                || marker.display().to_string(),
+                |name| name.to_string_lossy().into_owned()
+            )
+        )));
+    }
+    entries.sort();
+    Ok(entries)
+}
+
+fn remove_entry(path: &Path) -> Result<()> {
+    let metadata = match std::fs::symlink_metadata(path) {
+        Ok(metadata) => metadata,
+        // Gone between the plan and the removal is the outcome asked for.
+        Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ok(()),
+        Err(source) => {
+            return Err(Error::Io {
+                path: path.to_path_buf(),
+                source,
+            });
+        }
+    };
+    let removed = if metadata.is_dir() {
+        std::fs::remove_dir_all(path)
+    } else {
+        std::fs::remove_file(path)
+    };
+    removed.map_err(|source| Error::Io {
+        path: path.to_path_buf(),
+        source,
+    })
+}

crates/tsunagi-core/src/storage/state.rs

@@ -21,7 +21,7 @@ use crate::identity::{DeviceIdentity, NetworkId, NetworkName, NetworkSecret};
 use crate::state::{RecordBody, SignedRecord};
 
 /// Schema version written by this build.
-pub const SCHEMA_VERSION: i64 = 3;
+pub const SCHEMA_VERSION: i64 = 7;
 
 /// Key of the stored hostname setting.
 const SETTING_HOSTNAME: &str = "hostname";
@@ -40,6 +40,12 @@ pub struct StoredNetwork {
     pub secret: NetworkSecret,
     /// Whether the network is activated automatically at agent startup.
     pub auto_start: bool,
+    /// Local broadcast participation, enabled unless explicitly disabled.
+    pub broadcast: bool,
+    /// This agent offers itself as an exit node here. Off unless turned on.
+    pub exit_node: bool,
+    /// The member this device sends all its internet traffic through, if any.
+    pub exit_via: Option<[u8; 32]>,
 }
 
 /// The mandatory state store.
@@ -183,13 +189,68 @@ impl StateStore {
                 .execute_batch(SIGNED_RECORDS_SCHEMA)
                 .map_err(|err| self.corrupt(format!("cannot create schema: {err}")))?;
         }
+        if found < 4 {
+            self.conn.execute_batch("BEGIN;
+                ALTER TABLE networks ADD COLUMN broadcast INTEGER NOT NULL DEFAULT 1 CHECK (broadcast IN (0,1));
+                PRAGMA user_version = 4;
+                COMMIT;").map_err(|err| self.corrupt(format!("cannot migrate schema to 4: {err}")))?;
+        }
+        if found < 5 {
+            self.conn
+                .execute_batch(
+                    "BEGIN;
+                     CREATE TABLE IF NOT EXISTS peer_hostnames (
+                         network_id  BLOB NOT NULL,
+                         endpoint_id BLOB NOT NULL,
+                         hostname    TEXT NOT NULL,
+                         PRIMARY KEY (network_id, endpoint_id)
+                     );
+                     PRAGMA user_version = 5;
+                     COMMIT;",
+                )
+                .map_err(|err| self.corrupt(format!("cannot migrate schema to 5: {err}")))?;
+        }
+        if found < 6 {
+            self.conn
+                .execute_batch(
+                    "BEGIN;
+                     ALTER TABLE networks ADD COLUMN exit_node INTEGER NOT NULL DEFAULT 0 CHECK (exit_node IN (0,1));
+                     ALTER TABLE networks ADD COLUMN exit_via BLOB;
+                     PRAGMA user_version = 6;
+                     COMMIT;",
+                )
+                .map_err(|err| self.corrupt(format!("cannot migrate schema to 6: {err}")))?;
+        }
+        // Migration 6 -> 7: a claim gained the time its hostname was first
+        // claimed, which changed the signing domain. Records signed before
+        // no longer verify, so they are discarded rather than kept as rows
+        // every read has to reject; each member re-publishes its own claim on
+        // its next run and the others arrive again from their authors. The
+        // version counter is kept, so nothing is ever published twice at one
+        // version.
+        if found < 7 {
+            self.conn
+                .execute_batch(
+                    "BEGIN;
+                     DELETE FROM signed_records;
+                     PRAGMA user_version = 7;
+                     COMMIT;",
+                )
+                .map_err(|err| self.corrupt(format!("cannot migrate schema to 7: {err}")))?;
+        }
         Ok(())
     }
 
     /// Confirms the expected tables exist, so that a truncated or foreign
     /// database is reported rather than used.
     fn verify_shape(&self) -> Result<()> {
-        for table in ["device_identity", "networks", "settings", "signed_records"] {
+        for table in [
+            "device_identity",
+            "networks",
+            "settings",
+            "signed_records",
+            "peer_hostnames",
+        ] {
             let present: Option<String> = self
                 .conn
                 .query_row(
@@ -396,6 +457,21 @@ impl StateStore {
         Ok(())
     }
 
+    /// Saves local participation without changing identity, membership or auto-start.
+    pub fn set_broadcast(&self, network_id: NetworkId, enabled: bool) -> Result<()> {
+        let changed = self
+            .conn
+            .execute(
+                "UPDATE networks SET broadcast = ?2 WHERE network_id = ?1",
+                params![network_id.as_bytes().as_slice(), enabled as i64],
+            )
+            .map_err(|err| Error::Storage(format!("cannot update broadcast policy: {err}")))?;
+        if changed == 0 {
+            return Err(Error::NetworkUnknown(network_id));
+        }
+        Ok(())
+    }
+
     /// Removes a network configuration entirely.
     pub fn remove_network(&self, network_id: NetworkId) -> Result<()> {
         self.conn
@@ -407,11 +483,56 @@ impl StateStore {
         Ok(())
     }
 
+    /// Saves whether this agent offers itself as an exit node in a network.
+    pub fn set_exit_node(&self, network_id: NetworkId, enabled: bool) -> Result<()> {
+        let changed = self
+            .conn
+            .execute(
+                "UPDATE networks SET exit_node = ?2 WHERE network_id = ?1",
+                params![network_id.as_bytes().as_slice(), enabled as i64],
+            )
+            .map_err(|err| Error::Storage(format!("cannot update exit node offer: {err}")))?;
+        if changed == 0 {
+            return Err(Error::NetworkUnknown(network_id));
+        }
+        Ok(())
+    }
+
+    /// Saves which member this device sends its internet traffic through.
+    ///
+    /// One device has one default route, so choosing an exit node in one
+    /// network clears the choice in every other.
+    pub fn set_exit_via(&self, network_id: NetworkId, via: Option<&[u8; 32]>) -> Result<()> {
+        let tx = self
+            .conn
+            .unchecked_transaction()
+            .map_err(|err| Error::Storage(format!("cannot update exit node: {err}")))?;
+        let changed = tx
+            .execute(
+                "UPDATE networks SET exit_via = ?2 WHERE network_id = ?1",
+                params![network_id.as_bytes().as_slice(), via.map(|v| v.as_slice())],
+            )
+            .map_err(|err| Error::Storage(format!("cannot update exit node: {err}")))?;
+        if changed == 0 {
+            return Err(Error::NetworkUnknown(network_id));
+        }
+        if via.is_some() {
+            tx.execute(
+                "UPDATE networks SET exit_via = NULL WHERE network_id != ?1",
+                params![network_id.as_bytes().as_slice()],
+            )
+            .map_err(|err| Error::Storage(format!("cannot update exit node: {err}")))?;
+        }
+        tx.commit()
+            .map_err(|err| Error::Storage(format!("cannot update exit node: {err}")))?;
+        Ok(())
+    }
+
     /// Lists every configured network.
     pub fn list_networks(&self) -> Result<Vec<StoredNetwork>> {
         let mut stmt = self
             .conn
-            .prepare("SELECT network_id, name, secret, auto_start FROM networks ORDER BY name")
+            .prepare("SELECT network_id, name, secret, auto_start, broadcast, exit_node, exit_via FROM networks ORDER BY name")
             .map_err(|err| Error::Storage(format!("cannot list networks: {err}")))?;
         let rows = stmt
             .query_map([], |row| {
@@ -419,13 +540,21 @@ impl StateStore {
                 let name: String = row.get(1)?;
                 let secret: Vec<u8> = row.get(2)?;
                 let auto_start: i64 = row.get(3)?;
-                Ok((id, name, secret, auto_start != 0))
+                Ok((
+                    id,
+                    name,
+                    secret,
+                    auto_start != 0,
+                    row.get::<_, bool>(4)?,
+                    row.get::<_, bool>(5)?,
+                    row.get::<_, Option<Vec<u8>>>(6)?,
+                ))
             })
             .map_err(|err| Error::Storage(format!("cannot list networks: {err}")))?;
 
         let mut out = Vec::new();
         for row in rows {
-            let (id, name, secret, auto_start) =
+            let (id, name, secret, auto_start, broadcast, exit_node, exit_via) =
                 row.map_err(|err| Error::Storage(format!("cannot read network row: {err}")))?;
             let id: [u8; 32] = id
                 .as_slice()
@@ -436,6 +565,9 @@ impl StateStore {
                 name: NetworkName::new(name)?,
                 secret: NetworkSecret::from_bytes(secret)?,
                 auto_start,
+                broadcast,
+                exit_node,
+                exit_via: exit_via.and_then(|bytes| <[u8; 32]>::try_from(bytes.as_slice()).ok()),
             });
         }
         Ok(out)
@@ -611,6 +743,61 @@ impl StateStore {
         Ok(())
     }
 
+    /// Remembers the hostname a member last announced.
+    pub fn remember_peer_hostname(
+        &self,
+        network_id: NetworkId,
+        endpoint_id: &[u8; 32],
+        hostname: &str,
+    ) -> Result<()> {
+        self.conn
+            .execute(
+                "INSERT INTO peer_hostnames (network_id, endpoint_id, hostname)
+                 VALUES (?1, ?2, ?3)
+                 ON CONFLICT(network_id, endpoint_id) DO UPDATE SET hostname = excluded.hostname",
+                params![
+                    network_id.as_bytes().as_slice(),
+                    endpoint_id.as_slice(),
+                    hostname
+                ],
+            )
+            .map_err(|err| Error::Storage(format!("cannot remember a hostname: {err}")))?;
+        Ok(())
+    }
+
+    /// The last hostname every member of a network announced.
+    pub fn peer_hostnames(&self, network_id: NetworkId) -> Result<Vec<([u8; 32], String)>> {
+        let mut stmt = self
+            .conn
+            .prepare("SELECT endpoint_id, hostname FROM peer_hostnames WHERE network_id = ?1")
+            .map_err(|err| Error::Storage(format!("cannot read hostnames: {err}")))?;
+        let rows = stmt
+            .query_map(params![network_id.as_bytes().as_slice()], |row| {
+                Ok((row.get::<_, Vec<u8>>(0)?, row.get::<_, String>(1)?))
+            })
+            .map_err(|err| Error::Storage(format!("cannot read hostnames: {err}")))?;
+        let mut out = Vec::new();
+        for row in rows {
+            let (id, hostname) =
+                row.map_err(|err| Error::Storage(format!("cannot read hostnames: {err}")))?;
+            if let Ok(id) = <[u8; 32]>::try_from(id.as_slice()) {
+                out.push((id, hostname));
+            }
+        }
+        Ok(out)
+    }
+
+    /// Forgets every remembered hostname of a network.
+    pub fn forget_peer_hostnames(&self, network_id: NetworkId) -> Result<()> {
+        self.conn
+            .execute(
+                "DELETE FROM peer_hostnames WHERE network_id = ?1",
+                params![network_id.as_bytes().as_slice()],
+            )
+            .map_err(|err| Error::Storage(format!("cannot forget hostnames: {err}")))?;
+        Ok(())
+    }
+
     /// Reads an arbitrary setting.
     pub fn get_setting(&self, key: &str) -> Result<Option<String>> {
         self.conn
@@ -668,3 +855,117 @@ pub(crate) fn now_unix() -> i64 {
         .map(|d| d.as_secs() as i64)
         .unwrap_or(0)
 }
+#[cfg(test)]
+mod broadcast_storage_tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use crate::identity::NetworkKeys;
+
+    #[test]
+    fn exit_node_choices_are_off_by_default_and_one_choice_clears_the_others() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = StateStore::open(dir.path().join("state.sqlite")).unwrap();
+        let secret = NetworkSecret::from_bytes([5; 32]).unwrap();
+        let make = |name: &str| {
+            let name = NetworkName::new(name).unwrap();
+            let id = NetworkKeys::derive(&name, &secret).network_id();
+            store.upsert_network(id, &name, &secret, true).unwrap();
+            id
+        };
+        let (a, b) = (make("exit-a"), make("exit-b"));
+        let listed = || store.list_networks().unwrap();
+        assert!(
+            listed()
+                .iter()
+                .all(|n| !n.exit_node && n.exit_via.is_none())
+        );
+
+        store.set_exit_node(a, true).unwrap();
+        store.set_exit_via(a, Some(&[7; 32])).unwrap();
+        store.set_exit_via(b, Some(&[8; 32])).unwrap();
+        let by_id = |id: NetworkId| listed().into_iter().find(|n| n.network_id == id).unwrap();
+        assert!(by_id(a).exit_node, "offering is per network and kept");
+        assert_eq!(by_id(a).exit_via, None, "one default route: a was cleared");
+        assert_eq!(by_id(b).exit_via, Some([8; 32]));
+
+        store.set_exit_via(b, None).unwrap();
+        assert_eq!(by_id(b).exit_via, None);
+    }
+
+    #[test]
+    fn a_remembered_hostname_survives_a_restart_and_follows_the_latest_name() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("state.sqlite");
+        let name = NetworkName::new("hostnames").unwrap();
+        let secret = NetworkSecret::from_bytes([9; 32]).unwrap();
+        let id = NetworkKeys::derive(&name, &secret).network_id();
+        let peer = [3u8; 32];
+
+        let store = StateStore::open(&path).unwrap();
+        store.remember_peer_hostname(id, &peer, "laptop").unwrap();
+        store.remember_peer_hostname(id, &peer, "laptop-2").unwrap();
+        drop(store);
+
+        let store = StateStore::open(&path).unwrap();
+        assert_eq!(
+            store.peer_hostnames(id).unwrap(),
+            vec![(peer, "laptop-2".to_string())]
+        );
+        store.forget_peer_hostnames(id).unwrap();
+        assert!(store.peer_hostnames(id).unwrap().is_empty());
+    }
+
+    #[test]
+    fn v3_migration_preserves_identity_and_networks_and_defaults_broadcast_on() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("state.sqlite");
+        let name = NetworkName::new("migration-broadcast").unwrap();
+        let secret = NetworkSecret::from_bytes([7; 32]).unwrap();
+        let id = NetworkKeys::derive(&name, &secret).network_id();
+        let store = StateStore::open(&path).unwrap();
+        let identity = store
+            .load_or_create_device_identity()
+            .unwrap()
+            .endpoint_id();
+        store.upsert_network(id, &name, &secret, false).unwrap();
+        store.set_hostname("old-host").unwrap();
+        store
+            .conn
+            .execute_batch(
+                "ALTER TABLE networks DROP COLUMN broadcast; \
+             ALTER TABLE networks DROP COLUMN exit_node; \
+             ALTER TABLE networks DROP COLUMN exit_via; \
+             PRAGMA user_version=3;",
+            )
+            .unwrap();
+        drop(store);
+        let store = StateStore::open(&path).unwrap();
+        let networks = store.list_networks().unwrap();
+        assert_eq!(networks.len(), 1);
+        assert_eq!(networks[0].network_id, id);
+        assert!(!networks[0].auto_start);
+        assert!(networks[0].broadcast);
+        assert_eq!(store.hostname().unwrap().as_deref(), Some("old-host"));
+        assert_eq!(
+            store
+                .load_or_create_device_identity()
+                .unwrap()
+                .endpoint_id(),
+            identity
+        );
+        store.set_broadcast(id, false).unwrap();
+        store.upsert_network(id, &name, &secret, true).unwrap();
+        drop(store);
+        let store = StateStore::open(&path).unwrap();
+        assert!(
+            !store.list_networks().unwrap()[0].broadcast,
+            "rejoin and restart preserve opt-out"
+        );
+        store.remove_network(id).unwrap();
+        store.upsert_network(id, &name, &secret, true).unwrap();
+        assert!(
+            store.list_networks().unwrap()[0].broadcast,
+            "forgotten network gets defaults"
+        );
+    }
+}

crates/tsunagi-core/src/task.rs

@@ -0,0 +1,74 @@
+//! Stopping the tasks an agent started.
+//!
+//! Shutdown is cooperative first and bounded always: every loop watches a
+//! cancellation token, and anything still running when its grace period ends
+//! is aborted. Nothing an agent owns may outlive the agent, and no peer may
+//! hold shutdown open by refusing to read.
+//!
+//! Public because a protocol crate spawns tasks of its own and is held to
+//! the same rule: the agent gives each plugin a grace period, but abandoning
+//! the future it is awaiting does not stop the task behind it.
+
+use std::time::Duration;
+
+use tokio::task::JoinHandle;
+
+/// How long a task gets to notice cancellation before it is cut off.
+///
+/// Winding down is cooperative first: a task in the middle of a write to a
+/// peer gets a moment to finish it. After that it is aborted, because a peer
+/// that stops reading must not be able to hold shutdown open.
+pub const TASK_GRACE: Duration = Duration::from_secs(5);
+
+/// Waits for `task` to finish, aborting it if it outlasts `grace`.
+///
+/// Returns whether it finished on its own. `what` names it in the warning,
+/// which is the only signal that something was cut off rather than stopped.
+pub async fn wind_down(task: JoinHandle<()>, grace: Duration, what: &str) -> bool {
+    let abort = task.abort_handle();
+    if tokio::time::timeout(grace, task).await.is_err() {
+        tracing::warn!(
+            task = what,
+            grace_secs = grace.as_secs_f32(),
+            "task did not wind down in time; aborting it"
+        );
+        abort.abort();
+        return false;
+    }
+    true
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+    use super::*;
+
+    #[tokio::test]
+    async fn a_task_that_stops_on_its_own_is_not_aborted() {
+        let (tx, mut rx) = tokio::sync::mpsc::channel::<()>(1);
+        let task = tokio::spawn(async move {
+            tx.send(()).await.ok();
+        });
+        assert!(wind_down(task, Duration::from_secs(5), "cooperative").await);
+        assert_eq!(rx.recv().await, Some(()), "it ran to completion");
+    }
+
+    #[tokio::test]
+    async fn a_task_that_ignores_the_grace_is_aborted() {
+        // A stand-in for a task blocked writing to a peer that stopped
+        // reading: it will never finish by itself.
+        let (tx, mut rx) = tokio::sync::mpsc::channel::<()>(1);
+        let task = tokio::spawn(async move {
+            let _held = tx;
+            std::future::pending::<()>().await;
+        });
+
+        let finished = wind_down(task, Duration::from_millis(50), "wedged").await;
+        assert!(!finished, "it had to be cut off");
+
+        // Aborted for real: the task is gone, so what it held is dropped.
+        let closed = tokio::time::timeout(Duration::from_secs(5), rx.recv()).await;
+        assert_eq!(closed, Ok(None), "the aborted task dropped its sender");
+    }
+}

crates/tsunagi-core/src/testing.rs

@@ -1,26 +1,31 @@
-//! Shared helpers for the integration tests.
+//! A harness for testing against a real agent.
 //!
-//! Every test uses real iroh endpoints on loopback, its own temporary SQLite
+//! Behind the `testing` feature, because it is only of use to a test and
-//! files and independent agent instances. Only discovery is substituted; iroh,
+//! pulls in `tempfile` and a tracing subscriber. It lives here rather than
-//! the handshake, message passing and persistent storage are not.
+//! being copied into each crate's `tests/` directory: a protocol plugin in
+//! its own crate needs exactly the same harness, and two copies of it would
+//! drift.
 //!
-//! Synchronisation is always "wait for a specific event or condition, under one
+//! Everything here builds agents that touch nothing outside a temporary
-//! overall deadline", never a fixed multi-second sleep.
+//! directory: loopback only, no relay, no address publication.
 
-#![allow(dead_code, clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+// A harness fails a test loudly and at once; that is the whole job. The
+// crate forbids these elsewhere for the opposite reason — nothing off the
+// network may panic — and this module never sees anything off the network.
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
 use std::future::Future;
 use std::path::Path;
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
+use crate::agent::Event;
+use crate::config::{AgentConfig, StoragePaths, TransportPolicy};
+use crate::discovery::SharedMemoryDiscovery;
+use crate::identity::{NetworkName, NetworkSecret};
+use crate::{Agent, Result};
 use tempfile::TempDir;
 use tokio::sync::broadcast::error::RecvError;
-use tsunagi::agent::Event;
-use tsunagi::config::{AgentConfig, StoragePaths, TransportPolicy};
-use tsunagi::discovery::SharedMemoryDiscovery;
-use tsunagi::identity::{NetworkName, NetworkSecret};
-use tsunagi::{Agent, Result};
 
 /// Installs a tracing subscriber when `TSUNAGI_TEST_LOG` is set.
 ///
@@ -50,7 +55,7 @@ const POLL_INTERVAL: Duration = Duration::from_millis(25);
 /// port mapping, so the suite needs neither the internet nor privileges.
 /// Timeouts are shortened so that failure paths finish quickly.
 pub fn local_config(dir: &Path) -> AgentConfig {
-    let limits = tsunagi::Limits {
+    let limits = crate::Limits {
         dial_timeout: Duration::from_millis(1500),
         handshake_timeout: Duration::from_secs(5),
         ..Default::default()
@@ -69,7 +74,9 @@ pub fn config_with(dir: &Path, discovery: &SharedMemoryDiscovery) -> AgentConfig
 
 /// A temporary directory plus the agent running on it.
 pub struct TestAgent {
+    /// The directory it keeps its state in, removed when this is dropped.
     pub dir: TempDir,
+    /// The agent itself.
     pub agent: Agent,
 }
 
@@ -148,14 +155,22 @@ where
 {
     let deadline = Instant::now() + DEADLINE;
     loop {
-        if let Some(value) = probe().await {
+        let remaining = deadline
-            return value;
+            .checked_duration_since(Instant::now())
-        }
+            .unwrap_or_default();
         assert!(
-            Instant::now() < deadline,
+            !remaining.is_zero(),
             "timed out waiting for condition: {what}"
         );
-        tokio::time::sleep(POLL_INTERVAL).await;
+
+        // The probe is bounded too. A deadline checked only between probes is
+        // no deadline at all: one call into a wedged agent that never answers
+        // would hang the test process for ever instead of failing it.
+        match tokio::time::timeout(remaining, probe()).await {
+            Ok(Some(value)) => return value,
+            Ok(None) => tokio::time::sleep(POLL_INTERVAL).await,
+            Err(_) => panic!("timed out waiting for condition: {what}"),
+        }
     }
 }
 
@@ -170,7 +185,7 @@ pub async fn settle() {
 /// Waits until `agent` has `count` authenticated peers in `network`.
 pub async fn wait_for_peers(
     agent: &Agent,
-    network: tsunagi::NetworkId,
+    network: crate::NetworkId,
     count: usize,
 ) -> Vec<iroh::EndpointId> {
     wait_until(&format!("{count} peers in {network}"), || async move {
@@ -183,3 +198,20 @@ pub async fn wait_for_peers(
     })
     .await
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test(start_paused = true)]
+    #[should_panic(expected = "timed out waiting for condition")]
+    async fn a_probe_that_never_answers_fails_the_deadline_instead_of_hanging() {
+        // Virtual time, so the deadline arrives at once rather than in thirty
+        // real seconds — and the deadline does arrive, which is the point: a
+        // probe that never returns used to hang the process instead.
+        wait_until::<(), _, _>("an answer that never comes", || async {
+            std::future::pending::<Option<()>>().await
+        })
+        .await;
+    }
+}

crates/tsunagi-core/tests/authentication.rs

@@ -8,9 +8,6 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
-use common::{TestAgent, network, settle, wait_event};
 use iroh::endpoint::{Connection, PortmapperConfig, RecvStream, SendStream, presets};
 use iroh::{Endpoint, EndpointAddr, RelayMode};
 use tsunagi::agent::Event;
@@ -22,6 +19,7 @@ use tsunagi::proto::message::{
 };
 use tsunagi::proto::{read_frame, write_frame};
 use tsunagi::test_support;
+use tsunagi::testing::{TestAgent, network, settle, wait_event};
 
 /// A bare iroh endpoint with no tsunagi agent behind it.
 async fn raw_endpoint() -> Endpoint {

crates/tsunagi-core/tests/cache_and_state.rs

@@ -2,14 +2,12 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
 use std::io::Write;
 
-use common::{TestAgent, config_with, local_config, network, wait_for_peers};
 use tsunagi::config::StoragePaths;
 use tsunagi::discovery::SharedMemoryDiscovery;
 use tsunagi::storage::CacheOutcome;
+use tsunagi::testing::{TestAgent, config_with, local_config, network, wait_for_peers};
 use tsunagi::{Agent, Error};
 
 /// Overwrites a file with bytes that are definitely not a SQLite database.
@@ -284,3 +282,85 @@ async fn secrets_never_appear_in_status_or_debug_output() {
 
     agent.agent.shutdown().await;
 }
+
+#[tokio::test]
+async fn a_wipe_removes_everything_and_the_next_start_is_a_stranger() {
+    let dir = tempfile::tempdir().unwrap();
+    let paths = StoragePaths::under(dir.path());
+    let discovery = SharedMemoryDiscovery::new();
+    let (name, secret) = network("wiped");
+
+    let agent = Agent::spawn(config_with(dir.path(), &discovery))
+        .await
+        .unwrap();
+    let before = agent.endpoint_id();
+    agent.join_network(&name, &secret).await.unwrap();
+    // Something a plugin keeps beside the state, which a wipe must take too.
+    std::fs::create_dir_all(paths.state_dir.join("wireguard")).unwrap();
+    std::fs::write(paths.state_dir.join("wireguard/keys.sqlite"), b"key").unwrap();
+
+    // Not while an agent owns the directory: a half-wiped state under a
+    // running agent is worse than no wipe at all.
+    let refused = tsunagi::storage::wipe(&paths).unwrap_err();
+    assert!(matches!(refused, Error::StateLocked { .. }), "{refused}");
+    agent.shutdown().await;
+
+    let plan = tsunagi::storage::wipe_plan(&paths).unwrap();
+    assert!(
+        plan.entries().any(|path| path.ends_with("state.sqlite")),
+        "{plan:?}"
+    );
+    assert!(
+        plan.entries().any(|path| path.ends_with("wireguard")),
+        "what a plugin kept is state too: {plan:?}"
+    );
+
+    let wiped = tsunagi::storage::wipe(&paths).unwrap();
+    assert_eq!(wiped, plan);
+    assert!(!paths.state_db().exists());
+    assert!(!paths.state_dir.join("wireguard").exists());
+    assert!(!paths.lock_file().exists(), "no lock is left claiming it");
+
+    // A stranger: new identity, no networks, nothing to be surprised by.
+    let fresh = Agent::spawn(config_with(dir.path(), &discovery))
+        .await
+        .unwrap();
+    assert_ne!(fresh.endpoint_id(), before);
+    assert!(fresh.list_networks().await.unwrap().is_empty());
+    fresh.shutdown().await;
+}
+
+#[tokio::test]
+async fn wiping_twice_is_as_ordinary_as_wiping_once() {
+    let dir = tempfile::tempdir().unwrap();
+    let paths = StoragePaths::under(dir.path());
+    let discovery = SharedMemoryDiscovery::new();
+
+    let agent = Agent::spawn(config_with(dir.path(), &discovery))
+        .await
+        .unwrap();
+    agent.shutdown().await;
+
+    assert!(!tsunagi::storage::wipe(&paths).unwrap().is_empty());
+    let again = tsunagi::storage::wipe(&paths).unwrap();
+    assert!(again.is_empty(), "nothing left to remove: {again:?}");
+}
+
+#[tokio::test]
+async fn a_directory_that_is_not_ours_is_refused_rather_than_emptied() {
+    // The mistyped `--state-dir` that would otherwise remove somebody's
+    // documents. A marker decides, not the name of the directory.
+    let dir = tempfile::tempdir().unwrap();
+    let paths = StoragePaths::new(dir.path(), dir.path().join("cache"));
+    std::fs::write(dir.path().join("thesis.txt"), b"years of work").unwrap();
+
+    let err = tsunagi::storage::wipe(&paths).unwrap_err();
+    assert!(
+        err.to_string().contains("does not look like a tsunagi"),
+        "{err}"
+    );
+    assert!(
+        dir.path().join("thesis.txt").exists(),
+        "nothing was removed"
+    );
+}

crates/tsunagi-core/tests/device_identity.rs

@@ -8,9 +8,6 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
-use common::{config_with, network, wait_until};
 use tempfile::TempDir;
 use tsunagi::Agent;
 use tsunagi::config::StoragePaths;
@@ -18,6 +15,7 @@ use tsunagi::discovery::SharedMemoryDiscovery;
 use tsunagi::identity::{NetworkKeys, NetworkName, NetworkSecret};
 use tsunagi::state::{RecordBody, StateSet};
 use tsunagi::storage::StateStore;
+use tsunagi::testing::{config_with, network, wait_until};
 
 #[tokio::test]
 async fn a_changed_name_reaches_peers_and_replaces_the_old_claim() {

crates/tsunagi-core/tests/discovery.rs

@@ -3,16 +3,14 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
 use std::sync::Arc;
 
-use common::{TestAgent, local_config, network, settle, wait_for_peers};
 use tsunagi::Agent;
 use tsunagi::discovery::{
     CandidateSource, CompositeDiscovery, NetworkDiscovery, SharedMemoryDiscovery, StaticBootstrap,
 };
 use tsunagi::identity::NetworkKeys;
+use tsunagi::testing::{TestAgent, local_config, network, settle, wait_for_peers};
 
 #[tokio::test]
 async fn a_static_bootstrap_candidate_is_enough_to_join() {
@@ -139,3 +137,63 @@ async fn forgetting_a_network_removes_it_from_the_state_store() {
     drop(restarted);
     drop(dir);
 }
+
+#[tokio::test]
+async fn being_told_about_one_member_is_enough_to_meet_them_all() {
+    // A network is a mesh, and which member happened to be named on a
+    // command line must not decide who a device ends up talking to. The
+    // newcomer is given one address and nothing else; everybody else is
+    // learned from the members it meets.
+    let (name, secret) = network("introductions");
+
+    // Three that already know each other, and nothing that resolves by
+    // network: no shared discovery here, only what members pass on.
+    let first_dir = tempfile::TempDir::new().unwrap();
+    let first = Agent::spawn(local_config(first_dir.path())).await.unwrap();
+    let network_id = first.join_network(&name, &secret).await.unwrap();
+
+    let second_dir = tempfile::TempDir::new().unwrap();
+    let second = Agent::spawn(
+        local_config(second_dir.path())
+            .with_discovery(Arc::new(StaticBootstrap::new([first.local_addr()]))),
+    )
+    .await
+    .unwrap();
+    second.join_network(&name, &secret).await.unwrap();
+    wait_for_peers(&first, network_id, 1).await;
+
+    let third_dir = tempfile::TempDir::new().unwrap();
+    let third = Agent::spawn(
+        local_config(third_dir.path())
+            .with_discovery(Arc::new(StaticBootstrap::new([first.local_addr()]))),
+    )
+    .await
+    .unwrap();
+    third.join_network(&name, &secret).await.unwrap();
+
+    // Each of the two newcomers was given only the first one's address,
+    // and each ends up with both of the others.
+    wait_for_peers(&second, network_id, 2).await;
+    wait_for_peers(&third, network_id, 2).await;
+    wait_for_peers(&first, network_id, 2).await;
+
+    // And the one they were not told about is there as an introduction:
+    // a candidate like any other, still authenticated by the handshake.
+    let status = second.network_status(network_id).await.unwrap();
+    let introduced = status
+        .candidates
+        .iter()
+        .find(|candidate| candidate.endpoint_id == third.endpoint_id())
+        .map(|candidate| candidate.source);
+    assert!(
+        matches!(
+            introduced,
+            Some(CandidateSource::Introduced) | Some(CandidateSource::Member)
+        ),
+        "the third should have arrived from the others: {introduced:?}"
+    );
+
+    second.shutdown().await;
+    third.shutdown().await;
+    first.shutdown().await;
+}

crates/tsunagi-core/tests/discovery_lifecycle.rs

@@ -0,0 +1,117 @@
+//! Discovery is a cancellable bootstrap job, never the network's event loop.
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+use iroh::{EndpointAddr, EndpointId};
+use std::sync::{
+    Arc,
+    atomic::{AtomicUsize, Ordering},
+};
+use std::time::Duration;
+use tsunagi::config::DiscoveryPolicy;
+use tsunagi::discovery::{BoxFuture, Candidate, NetworkDiscovery, SharedMemoryDiscovery};
+use tsunagi::identity::DiscoveryKey;
+use tsunagi::testing::{local_config, network, wait_for_peers, wait_until};
+use tsunagi::{Agent, Result};
+
+#[derive(Debug, Default)]
+struct Counted {
+    table: SharedMemoryDiscovery,
+    reads: AtomicUsize,
+    writes: AtomicUsize,
+}
+
+impl NetworkDiscovery for Counted {
+    fn name(&self) -> &str {
+        "counted"
+    }
+    fn publish<'a>(&'a self, key: DiscoveryKey, addr: EndpointAddr) -> BoxFuture<'a, Result<()>> {
+        self.writes.fetch_add(1, Ordering::SeqCst);
+        self.table.publish(key, addr)
+    }
+    fn unpublish<'a>(&'a self, key: DiscoveryKey, id: EndpointId) -> BoxFuture<'a, Result<()>> {
+        self.table.unpublish(key, id)
+    }
+    fn resolve<'a>(&'a self, key: DiscoveryKey) -> BoxFuture<'a, Result<Vec<Candidate>>> {
+        self.reads.fetch_add(1, Ordering::SeqCst);
+        self.table.resolve(key)
+    }
+}
+
+fn policy() -> DiscoveryPolicy {
+    DiscoveryPolicy {
+        publish_interval: Duration::from_millis(100),
+        lookup_interval: Duration::from_millis(50),
+        max_lookup_interval: Duration::from_millis(100),
+        reconnect_delay: Duration::from_millis(300),
+        request_timeout: Duration::from_secs(2),
+    }
+}
+
+#[tokio::test]
+async fn connected_peers_only_publish_and_isolated_peers_resume_search() {
+    let discovery = Arc::new(Counted::default());
+    let first_dir = tempfile::tempdir().unwrap();
+    let second_dir = tempfile::tempdir().unwrap();
+    let config = |path: &std::path::Path| {
+        local_config(path)
+            .with_discovery(discovery.clone())
+            .with_discovery_policy(policy())
+    };
+    let first = Agent::spawn(config(first_dir.path())).await.unwrap();
+    let second = Agent::spawn(config(second_dir.path())).await.unwrap();
+    let (name, secret) = network("discovery-lifecycle");
+    let id = first.join_network(&name, &secret).await.unwrap();
+    second.join_network(&name, &secret).await.unwrap();
+    wait_for_peers(&first, id, 1).await;
+    wait_for_peers(&second, id, 1).await;
+    let reads = discovery.reads.load(Ordering::SeqCst);
+    let writes = discovery.writes.load(Ordering::SeqCst);
+    wait_until("publications continue while connected", || async {
+        (discovery.writes.load(Ordering::SeqCst) >= writes + 6).then_some(())
+    })
+    .await;
+    assert_eq!(discovery.reads.load(Ordering::SeqCst), reads);
+    second.shutdown().await;
+    wait_for_peers(&first, id, 0).await;
+    wait_until("isolation resumes discovery", || async {
+        (discovery.reads.load(Ordering::SeqCst) > reads).then_some(())
+    })
+    .await;
+    first.shutdown().await;
+}
+
+#[derive(Debug)]
+struct Stalled;
+impl NetworkDiscovery for Stalled {
+    fn name(&self) -> &str {
+        "stalled"
+    }
+    fn publish<'a>(&'a self, _: DiscoveryKey, _: EndpointAddr) -> BoxFuture<'a, Result<()>> {
+        Box::pin(std::future::pending())
+    }
+    fn unpublish<'a>(&'a self, _: DiscoveryKey, _: EndpointId) -> BoxFuture<'a, Result<()>> {
+        Box::pin(std::future::pending())
+    }
+    fn resolve<'a>(&'a self, _: DiscoveryKey) -> BoxFuture<'a, Result<Vec<Candidate>>> {
+        Box::pin(std::future::pending())
+    }
+}
+
+#[tokio::test]
+async fn stalled_discovery_does_not_block_status_or_shutdown() {
+    let dir = tempfile::tempdir().unwrap();
+    let agent = Agent::spawn(local_config(dir.path()).with_discovery(Arc::new(Stalled)))
+        .await
+        .unwrap();
+    let (name, secret) = network("stalled-discovery");
+    let id = agent.join_network(&name, &secret).await.unwrap();
+    // A recheck schedules work; it must not await the unreachable backend.
+    agent.recheck_network(id).await.unwrap();
+    tokio::time::timeout(Duration::from_secs(2), agent.network_status(id))
+        .await
+        .unwrap()
+        .unwrap();
+    tokio::time::timeout(Duration::from_secs(2), agent.shutdown())
+        .await
+        .unwrap();
+}

crates/tsunagi-core/tests/end_to_end.rs

@@ -3,12 +3,10 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
-use common::{TestAgent, network, wait_event, wait_for_peers};
 use tsunagi::agent::Event;
 use tsunagi::discovery::SharedMemoryDiscovery;
 use tsunagi::proto::ControlMessage;
+use tsunagi::testing::{TestAgent, network, wait_event, wait_for_peers};
 
 #[tokio::test]
 async fn two_agents_authenticate_and_exchange_messages() {

crates/tsunagi-core/tests/identity.rs

@@ -5,12 +5,10 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
-use common::{TestAgent, config_with, network};
 use tsunagi::Agent;
 use tsunagi::discovery::SharedMemoryDiscovery;
 use tsunagi::identity::{NetworkKeys, NetworkName, NetworkSecret};
+use tsunagi::testing::{TestAgent, config_with, network};
 
 #[test]
 fn derivation_is_a_pure_function_of_name_and_secret() {

crates/tsunagi-core/tests/leaving.rs

@@ -0,0 +1,171 @@
+//! Scenario 11: leaving a network, and what the others are left holding.
+//!
+//! Leaving is not deactivating and not forgetting. It is a signed statement
+//! that this member gives up what it claimed, published while there is still
+//! somebody to hear it, because signed state has no expiry and nothing else
+//! will ever free the address.
+
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+use tsunagi::discovery::SharedMemoryDiscovery;
+use tsunagi::testing::{TestAgent, network, wait_for_peers, wait_until};
+
+#[tokio::test]
+async fn leaving_frees_the_address_for_everyone_still_there() {
+    let discovery = SharedMemoryDiscovery::new();
+    let (name, secret) = network("leaving-frees");
+
+    let leaver = TestAgent::spawn(&discovery).await.unwrap();
+    let stayer = TestAgent::spawn(&discovery).await.unwrap();
+    let network_id = leaver.agent.join_network(&name, &secret).await.unwrap();
+    stayer.agent.join_network(&name, &secret).await.unwrap();
+    wait_for_peers(&stayer.agent, network_id, 1).await;
+
+    let leaver_id = leaver.agent.endpoint_id();
+    let address = wait_until("the one leaving holds an address", || async {
+        let status = stayer.agent.network_status(network_id).await.ok()?;
+        status
+            .members
+            .iter()
+            .find(|member| member.endpoint_id == leaver_id)?
+            .overlay_address_v4
+    })
+    .await;
+
+    let outcome = leaver.agent.leave_network(network_id).await.unwrap();
+    assert!(outcome.announced, "there was a session to announce it on");
+    assert_eq!(outcome.peers_told, 1);
+
+    // The one still there drops it from the roster. The tombstone stays in
+    // the record set — it has to, or a replica that never heard of it would
+    // reinstate the old claim — but a member that gave everything up is not
+    // a member, and listing it as one makes leaving look like a fault.
+    wait_until("the member is gone from the roster", || async {
+        let status = stayer.agent.network_status(network_id).await.ok()?;
+        status
+            .members
+            .iter()
+            .all(|member| member.endpoint_id != leaver_id)
+            .then_some(())
+    })
+    .await;
+    let taken = stayer.agent.network_status(network_id).await.unwrap();
+    assert!(
+        !taken
+            .members
+            .iter()
+            .any(|member| member.overlay_address_v4 == Some(address)),
+        "the address is free for somebody else"
+    );
+
+    // And locally there is no membership left to be surprised by.
+    assert!(
+        leaver.agent.list_networks().await.unwrap().is_empty(),
+        "the network is gone from the store"
+    );
+    assert!(!leaver.agent.is_active(network_id).await);
+
+    leaver.agent.shutdown().await;
+    stayer.agent.shutdown().await;
+}
+
+#[tokio::test]
+async fn leaving_with_nobody_connected_says_so_rather_than_pretending() {
+    let discovery = SharedMemoryDiscovery::new();
+    let (name, secret) = network("leaving-alone");
+
+    let agent = TestAgent::spawn(&discovery).await.unwrap();
+    let network_id = agent.agent.join_network(&name, &secret).await.unwrap();
+
+    // Nobody is here, so the tombstone reaches nobody. The network is still
+    // left — the caller asked — but the outcome does not claim an audience
+    // there was not one for.
+    let outcome = agent.agent.leave_network(network_id).await.unwrap();
+    assert!(outcome.announced, "it was published locally");
+    assert_eq!(outcome.peers_told, 0);
+    assert!(agent.agent.list_networks().await.unwrap().is_empty());
+
+    agent.agent.shutdown().await;
+}
+
+#[tokio::test]
+async fn leaving_a_network_that_is_not_running_tells_nobody() {
+    let discovery = SharedMemoryDiscovery::new();
+    let (name, secret) = network("leaving-inactive");
+
+    let agent = TestAgent::spawn(&discovery).await.unwrap();
+    let network_id = agent.agent.join_network(&name, &secret).await.unwrap();
+    agent.agent.deactivate_network(network_id).await.unwrap();
+
+    // There is no runtime to sign and send from, so nothing was announced
+    // and the outcome says exactly that instead of a quiet success.
+    let outcome = agent.agent.leave_network(network_id).await.unwrap();
+    assert!(!outcome.announced);
+    assert_eq!(outcome.peers_told, 0);
+    assert!(agent.agent.list_networks().await.unwrap().is_empty());
+
+    agent.agent.shutdown().await;
+}
+
+#[tokio::test]
+async fn rejoining_after_leaving_is_not_taken_for_a_stale_record() {
+    let discovery = SharedMemoryDiscovery::new();
+    let (name, secret) = network("leaving-rejoin");
+
+    let returner = TestAgent::spawn(&discovery).await.unwrap();
+    let stayer = TestAgent::spawn(&discovery).await.unwrap();
+    let network_id = returner.agent.join_network(&name, &secret).await.unwrap();
+    stayer.agent.join_network(&name, &secret).await.unwrap();
+    wait_for_peers(&stayer.agent, network_id, 1).await;
+
+    let returner_id = returner.agent.endpoint_id();
+    wait_until("the one leaving holds an address", || async {
+        stayer
+            .agent
+            .network_status(network_id)
+            .await
+            .ok()?
+            .members
+            .iter()
+            .find(|member| member.endpoint_id == returner_id)?
+            .overlay_address_v4
+    })
+    .await;
+
+    returner.agent.leave_network(network_id).await.unwrap();
+    wait_until("the release reached the other one", || async {
+        let status = stayer.agent.network_status(network_id).await.ok()?;
+        status
+            .members
+            .iter()
+            .all(|member| member.endpoint_id != returner_id)
+            .then_some(())
+    })
+    .await;
+
+    // The same device, the same key, the same network. The version counter
+    // survived leaving on purpose: a claim numbered below the release would
+    // be ignored by every replica that already has the release, and this
+    // member would be invisible for good.
+    returner.agent.join_network(&name, &secret).await.unwrap();
+    wait_for_peers(&stayer.agent, network_id, 1).await;
+    let again = wait_until("the returning member is seen again", || async {
+        stayer
+            .agent
+            .network_status(network_id)
+            .await
+            .ok()?
+            .members
+            .iter()
+            .find(|member| member.endpoint_id == returner_id)?
+            .overlay_address_v4
+    })
+    .await;
+    assert!(
+        tsunagi::state::DEFAULT_IPV4_RANGE.contains(again),
+        "an address in the network's range: {again}"
+    );
+
+    returner.agent.shutdown().await;
+    stayer.agent.shutdown().await;
+}

crates/tsunagi-core/tests/mainline.rs

@@ -0,0 +1,92 @@
+//! Local Mainline Testnet, real iroh authentication, and durable agent state.
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+use std::time::Duration;
+use tsunagi::testing::{local_config, network, wait_for_peers};
+use tsunagi::{Agent, config::DiscoveryPolicy, discovery::MainlineDiscovery};
+
+#[tokio::test]
+#[ignore = "uses public Mainline DHT and iroh relay services"]
+async fn public_dht_finds_and_authenticates_two_agents() {
+    tsunagi::testing::init_tracing();
+    let a_dir = tempfile::tempdir().unwrap();
+    let b_dir = tempfile::tempdir().unwrap();
+    let config = |path: &std::path::Path| {
+        tsunagi::config::AgentConfig::new(tsunagi::config::StoragePaths::under(path))
+            .with_transport(tsunagi::config::TransportPolicy::N0Defaults)
+            .with_dht(MainlineDiscovery::default())
+    };
+    let a = Agent::spawn(config(a_dir.path())).await.unwrap();
+    let b = Agent::spawn(config(b_dir.path())).await.unwrap();
+    let (name, secret) = network("mainline-public-smoke");
+    let id = a.join_network(&name, &secret).await.unwrap();
+    b.join_network(&name, &secret).await.unwrap();
+    let found = tokio::time::timeout(Duration::from_secs(180), async {
+        loop {
+            if !a.network_status(id).await.unwrap().peers.is_empty()
+                && !b.network_status(id).await.unwrap().peers.is_empty()
+            {
+                break;
+            }
+            tokio::time::sleep(Duration::from_millis(250)).await;
+        }
+    })
+    .await;
+    a.shutdown().await;
+    b.shutdown().await;
+    assert!(
+        found.is_ok(),
+        "public DHT/relay discovery did not connect within three minutes"
+    );
+}
+#[tokio::test]
+async fn two_agents_restore_after_all_dht_records_and_addresses_are_gone() {
+    let first_dir = tempfile::tempdir().unwrap();
+    let second_dir = tempfile::tempdir().unwrap();
+    let (name, secret) = network("dht-restart");
+    let mut identities = None;
+    for _ in 0..2 {
+        if identities.is_some() {
+            for path in [first_dir.path(), second_dir.path()] {
+                let cache = tsunagi::config::StoragePaths::under(path).cache_dir;
+                if cache.exists() {
+                    std::fs::remove_dir_all(cache).unwrap();
+                }
+            }
+        }
+        // A fresh Testnet means no old rendezvous record survives. The second
+        // pass restores the networks from SQLite without another join command.
+        let net = mainline::Testnet::builder(5).build().unwrap();
+        let config = |path: &std::path::Path| {
+            local_config(path)
+                .with_dht(MainlineDiscovery::local_testnet(&net.bootstrap).unwrap())
+                .with_discovery_policy(DiscoveryPolicy {
+                    lookup_interval: Duration::from_millis(100),
+                    max_lookup_interval: Duration::from_millis(300),
+                    ..Default::default()
+                })
+        };
+        let (a, b) = tokio::join!(
+            Agent::spawn(config(first_dir.path())),
+            Agent::spawn(config(second_dir.path()))
+        );
+        let a = a.unwrap();
+        let b = b.unwrap();
+        let id = tsunagi::identity::NetworkKeys::derive(&name, &secret).network_id();
+        if let Some(old) = identities {
+            assert_eq!((a.endpoint_id(), b.endpoint_id()), old);
+        } else {
+            identities = Some((a.endpoint_id(), b.endpoint_id()));
+            let (ra, rb) = tokio::join!(
+                a.join_network(&name, &secret),
+                b.join_network(&name, &secret)
+            );
+            assert_eq!(ra.unwrap(), id);
+            assert_eq!(rb.unwrap(), id);
+        }
+        wait_for_peers(&a, id, 1).await;
+        wait_for_peers(&b, id, 1).await;
+        a.shutdown().await;
+        b.shutdown().await;
+    }
+}

crates/tsunagi-core/tests/mainline_socket.rs

@@ -0,0 +1,68 @@
+//! The real Mainline UDP receive loop must be quiet while idle on Windows too.
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+use std::net::Ipv4Addr;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::time::Duration;
+
+use tracing_subscriber::layer::SubscriberExt;
+
+struct SocketWarnings(Arc<AtomicUsize>);
+
+impl<S: tracing::Subscriber> tracing_subscriber::Layer<S> for SocketWarnings {
+    fn on_event(&self, event: &tracing::Event<'_>, _: tracing_subscriber::layer::Context<'_, S>) {
+        if event.metadata().target() == "mainline::rpc::socket"
+            && *event.metadata().level() <= tracing::Level::WARN
+        {
+            self.0.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+}
+
+#[tokio::test]
+async fn an_idle_dht_socket_stays_quiet_and_still_answers_packets() {
+    // Mainline owns an OS thread, so a thread-local subscriber cannot see its
+    // events. This integration-test binary has just this test and installs
+    // its own subscriber; it cannot affect the application or other binaries.
+    let warnings = Arc::new(AtomicUsize::new(0));
+    tracing::subscriber::set_global_default(
+        tracing_subscriber::registry().with(SocketWarnings(Arc::clone(&warnings))),
+    )
+    .unwrap();
+
+    tokio::time::timeout(Duration::from_secs(5), async {
+        let dht = mainline::Dht::builder()
+            .no_bootstrap()
+            .server_mode()
+            .bind_address(Ipv4Addr::LOCALHOST)
+            .port(0)
+            .build()
+            .unwrap()
+            .as_async();
+
+        let address = dht.info().await.local_addr();
+        // Each info response comes from the actor after another receive-loop
+        // iteration. With no peers or packets, these exercise real socket
+        // read timeouts, without an arbitrary sleep or public bootstrap.
+        for _ in 0..8 {
+            assert_eq!(dht.info().await.local_addr(), address);
+        }
+        assert_eq!(warnings.load(Ordering::Relaxed), 0, "idle receive warnings");
+
+        let client = tokio::net::UdpSocket::bind((Ipv4Addr::LOCALHOST, 0))
+            .await
+            .unwrap();
+        let ping = b"d1:ad2:id20:abcdefghijklmnopqrste1:q4:ping1:t4:ping1:y1:qe";
+        client.send_to(ping, address).await.unwrap();
+        let mut response = [0u8; 2048];
+        let (len, from) = client.recv_from(&mut response).await.unwrap();
+        assert_eq!(from, std::net::SocketAddr::V4(address));
+        let response = &response[..len];
+        assert!(response.windows(9).any(|field| field == b"1:t4:ping"));
+        assert!(response.windows(6).any(|field| field == b"1:y1:r"));
+        assert_eq!(warnings.load(Ordering::Relaxed), 0);
+    })
+    .await
+    .expect("the DHT receive loop keeps making progress");
+}

crates/tsunagi-core/tests/mixed_transports.rs

@@ -0,0 +1,308 @@
+//! Connectivity does not depend on what kind of link each hop happens to be,
+//! and a member is online however far away it is.
+
+#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+
+use std::collections::{HashMap, HashSet};
+use std::sync::{Arc, Mutex};
+
+use bytes::Bytes;
+use iroh::EndpointId;
+use tsunagi::NetworkId;
+use tsunagi::agent::DataPath;
+use tsunagi::dataplane::transport::SharedLink;
+use tsunagi::dataplane::{IpPlugin, PluginCapability, PluginError};
+use tsunagi::discovery::SharedMemoryDiscovery;
+use tsunagi::testing::{TestAgent, network, wait_for_peers, wait_until};
+
+/// A protocol that moves bytes between two members unchanged and keeps the
+/// link it was handed, so a test can use the link directly.
+#[derive(Debug)]
+struct Carrier {
+    protocol: String,
+    links: Mutex<HashMap<EndpointId, SharedLink>>,
+}
+
+impl Carrier {
+    fn new(protocol: &str) -> Arc<Self> {
+        Arc::new(Self {
+            protocol: protocol.into(),
+            links: Mutex::default(),
+        })
+    }
+
+    fn link(&self, peer: EndpointId) -> Option<SharedLink> {
+        self.links.lock().unwrap().get(&peer).cloned()
+    }
+}
+
+impl IpPlugin for Carrier {
+    fn protocol_id(&self) -> &str {
+        &self.protocol
+    }
+
+    fn protocol_version(&self) -> u16 {
+        1
+    }
+
+    fn local_capability(
+        &self,
+        _network: NetworkId,
+    ) -> Result<Option<PluginCapability>, PluginError> {
+        Ok(Some(PluginCapability {
+            protocol: self.protocol.clone(),
+            version: 1,
+            enabled: true,
+            data: Vec::new(),
+        }))
+    }
+
+    fn on_peer_capability(
+        &self,
+        _network: NetworkId,
+        _peer: EndpointId,
+        _capability: &PluginCapability,
+    ) -> Result<(), PluginError> {
+        Ok(())
+    }
+
+    fn on_peer_link(&self, _network: NetworkId, peer: EndpointId, link: SharedLink) {
+        self.links.lock().unwrap().insert(peer, link);
+    }
+
+    fn on_peer_gone(&self, _network: NetworkId, peer: EndpointId) {
+        self.links.lock().unwrap().remove(&peer);
+    }
+
+    fn on_network_deactivated(&self, _network: NetworkId) {}
+}
+
+type Blocked = Arc<Mutex<HashSet<EndpointId>>>;
+
+async fn member(
+    discovery: &SharedMemoryDiscovery,
+    plugins: &[Arc<Carrier>],
+    blocked: &Blocked,
+) -> TestAgent {
+    let plugins: Vec<_> = plugins.to_vec();
+    let blocked = Arc::clone(blocked);
+    TestAgent::spawn_with(
+        move |config| {
+            let config = config.with_unreachable_data_peers(blocked);
+            plugins
+                .into_iter()
+                .fold(config, |config, plugin| config.with_plugin(plugin))
+        },
+        discovery,
+    )
+    .await
+    .unwrap()
+}
+
+async fn exchange(from: &SharedLink, to: &SharedLink, payload: &'static [u8]) {
+    from.send(Bytes::from_static(payload)).unwrap();
+    let got = tokio::time::timeout(tsunagi::testing::DEADLINE, to.recv())
+        .await
+        .expect("the datagram should arrive")
+        .expect("the link stays open");
+    assert_eq!(got, Bytes::from_static(payload));
+}
+
+#[tokio::test]
+async fn a_tunnel_crosses_a_member_that_shares_no_protocol_with_either_end_and_moves_to_a_direct_link_by_itself()
+ {
+    // A and B share one protocol and nothing else. C shares a different
+    // one with each, so every link here is of a different kind and the
+    // middle has never heard of what A and B speak. A and B cannot reach
+    // each other directly, which is the case a blocked UDP path produces.
+    let discovery = SharedMemoryDiscovery::new();
+    let (name, secret) = network("mixed-transports");
+
+    let blocked_a: Blocked = Arc::default();
+    let blocked_b: Blocked = Arc::default();
+    let a_shared = Carrier::new("shared");
+    let b_shared = Carrier::new("shared");
+    let a = member(
+        &discovery,
+        &[Carrier::new("a-and-c"), Arc::clone(&a_shared)],
+        &blocked_a,
+    )
+    .await;
+    let b = member(
+        &discovery,
+        &[Carrier::new("c-and-b"), Arc::clone(&b_shared)],
+        &blocked_b,
+    )
+    .await;
+    let c = member(
+        &discovery,
+        &[Carrier::new("a-and-c"), Carrier::new("c-and-b")],
+        &Blocked::default(),
+    )
+    .await;
+    let (a_id, b_id, c_id) = (
+        a.agent.endpoint_id(),
+        b.agent.endpoint_id(),
+        c.agent.endpoint_id(),
+    );
+    blocked_a.lock().unwrap().insert(b_id);
+    blocked_b.lock().unwrap().insert(a_id);
+
+    let network_id = a.agent.join_network(&name, &secret).await.unwrap();
+    b.agent.join_network(&name, &secret).await.unwrap();
+    c.agent.join_network(&name, &secret).await.unwrap();
+    for agent in [&a, &b, &c] {
+        wait_for_peers(&agent.agent, network_id, 2).await;
+    }
+
+    // The tunnel comes up through C, and data goes both ways.
+    let to_b = wait_until("a has a link to b", || async { a_shared.link(b_id) }).await;
+    let to_a = wait_until("b has a link to a", || async { b_shared.link(a_id) }).await;
+    exchange(&to_b, &to_a, b"a to b, sealed").await;
+    exchange(&to_a, &to_b, b"b to a, sealed").await;
+
+    // The status says what is true: online, and through whom.
+    let status = a.agent.network_status(network_id).await.unwrap();
+    let b_peer = status.peers.iter().find(|p| p.endpoint_id == b_id).unwrap();
+    assert_eq!(b_peer.data_path, DataPath::Relayed { hops: 2, via: c_id });
+    let c_peer = status.peers.iter().find(|p| p.endpoint_id == c_id).unwrap();
+    assert!(matches!(&c_peer.data_path, DataPath::Direct { transport } if transport == "a-and-c"));
+    let b_member = wait_until("a lists b as a member", || async {
+        let status = a.agent.network_status(network_id).await.ok()?;
+        status.members.into_iter().find(|m| m.endpoint_id == b_id)
+    })
+    .await;
+    assert!(b_member.online);
+    let carried = c.agent.network_status(network_id).await.unwrap().relay;
+    assert!(carried.forwarded >= 2, "{carried:?}");
+
+    // The direct path becomes possible: the same link carries on, now
+    // without C, and nobody has to do anything about it.
+    blocked_a.lock().unwrap().clear();
+    blocked_b.lock().unwrap().clear();
+    wait_until("a reaches b directly", || async {
+        let status = a.agent.network_status(network_id).await.ok()?;
+        let peer = status.peers.iter().find(|p| p.endpoint_id == b_id)?;
+        matches!(&peer.data_path, DataPath::Direct { transport } if transport == "shared")
+            .then_some(())
+    })
+    .await;
+    let forwarded = c
+        .agent
+        .network_status(network_id)
+        .await
+        .unwrap()
+        .relay
+        .forwarded;
+    exchange(&to_b, &to_a, b"direct now").await;
+    assert!(
+        Arc::ptr_eq(&to_b, &a_shared.link(b_id).unwrap()),
+        "the tunnel's link is the one it always had"
+    );
+    assert_eq!(
+        c.agent
+            .network_status(network_id)
+            .await
+            .unwrap()
+            .relay
+            .forwarded,
+        forwarded,
+        "nothing goes through C any more"
+    );
+
+    a.stop().await;
+    b.stop().await;
+    c.stop().await;
+}
+
+#[tokio::test]
+async fn a_name_belongs_to_the_member_that_claimed_it_first_on_every_member() {
+    let discovery = SharedMemoryDiscovery::new();
+    let (name, secret) = network("first-claim");
+    let spawn = |hostname: &'static str| {
+        let discovery = discovery.clone();
+        async move {
+            TestAgent::spawn_with(move |config| config.with_hostname(hostname), &discovery)
+                .await
+                .unwrap()
+        }
+    };
+
+    let first = spawn("music").await;
+    let network_id = first.agent.join_network(&name, &secret).await.unwrap();
+    let first_id = first.agent.endpoint_id();
+    wait_until("the first claim is signed", || async {
+        let status = first.agent.network_status(network_id).await.ok()?;
+        status
+            .members
+            .iter()
+            .any(|m| m.endpoint_id == first_id && m.hostname.as_deref() == Some("music"))
+            .then_some(())
+    })
+    .await;
+
+    // Whatever the ids turn out to be, the later one does not take it.
+    let second = spawn("music").await;
+    let third = spawn("studio").await;
+    second.agent.join_network(&name, &secret).await.unwrap();
+    third.agent.join_network(&name, &secret).await.unwrap();
+    let second_id = second.agent.endpoint_id();
+    for agent in [&first, &second, &third] {
+        wait_for_peers(&agent.agent, network_id, 2).await;
+    }
+
+    for (viewer, label) in [(&first, "first"), (&second, "second"), (&third, "third")] {
+        let status = wait_until(&format!("{label} sees both claims"), || async {
+            let status = viewer.agent.network_status(network_id).await.ok()?;
+            // The roster is the authors of signed records, so a member
+            // being in it means its claim has arrived.
+            let both = [first_id, second_id]
+                .iter()
+                .all(|id| status.members.iter().any(|m| m.endpoint_id == *id));
+            both.then_some(status)
+        })
+        .await;
+        let holder = status
+            .members
+            .iter()
+            .find(|m| m.endpoint_id == first_id)
+            .unwrap();
+        let loser = status
+            .members
+            .iter()
+            .find(|m| m.endpoint_id == second_id)
+            .unwrap();
+        assert_eq!(holder.hostname.as_deref(), Some("music"), "{label}");
+        assert_eq!(loser.hostname, None, "{label}: no name for the later one");
+        let conflict = loser.hostname_conflict.as_ref().expect(label);
+        assert_eq!(conflict.name, "music");
+        assert_eq!(conflict.holder, first_id);
+        assert!(holder.hostname_conflict.is_none(), "{label}");
+        // Nothing about being second makes it less of a member.
+        assert!(
+            loser.online || viewer.agent.endpoint_id() == second_id,
+            "{label}"
+        );
+    }
+
+    // The later one is told so itself, and keeps working.
+    let own = second.agent.network_status(network_id).await.unwrap();
+    let conflict = own.hostname_conflict.expect("told its name is taken");
+    assert_eq!(
+        (conflict.name.as_str(), conflict.holder),
+        ("music", first_id)
+    );
+    assert!(
+        first
+            .agent
+            .network_status(network_id)
+            .await
+            .unwrap()
+            .hostname_conflict
+            .is_none()
+    );
+
+    first.stop().await;
+    second.stop().await;
+    third.stop().await;
+}

crates/tsunagi-core/tests/multi_peer.rs

@@ -3,16 +3,14 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
 use std::collections::HashSet;
 use std::sync::Arc;
 
-use common::{TestAgent, network, wait_event, wait_for_peers, wait_until};
 use tsunagi::agent::Event;
 use tsunagi::dataplane::TestCapabilityPlugin;
 use tsunagi::discovery::SharedMemoryDiscovery;
 use tsunagi::proto::ControlMessage;
+use tsunagi::testing::{TestAgent, network, wait_event, wait_for_peers, wait_until};
 
 #[tokio::test]
 async fn four_agents_form_a_mesh_and_exchange_distinguishable_messages() {

crates/tsunagi-core/tests/network_isolation.rs

@@ -2,9 +2,6 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
-use common::{TestAgent, network, settle, wait_event, wait_for_peers};
 use iroh::endpoint::{PortmapperConfig, presets};
 use iroh::{Endpoint, RelayMode};
 use tsunagi::agent::Event;
@@ -16,6 +13,7 @@ use tsunagi::proto::message::{
 };
 use tsunagi::proto::{read_frame, write_frame};
 use tsunagi::test_support;
+use tsunagi::testing::{TestAgent, network, settle, wait_event, wait_for_peers, wait_until};
 
 const LIMIT: usize = 64 * 1024;
 
@@ -264,3 +262,73 @@ async fn deactivating_one_network_leaves_the_agent_and_others_running() {
 
     agent.agent.shutdown().await;
 }
+
+#[tokio::test]
+async fn two_networks_on_one_agent_each_get_a_range_of_their_own() {
+    // One agent, one interface: the configured range can only belong to one
+    // of them. The second does not go without — it takes the range derived
+    // from its own id, which every one of its members derives identically,
+    // so it is an agreement and not a local invention.
+    let discovery = SharedMemoryDiscovery::new();
+    let (first, first_secret) = network("two-ranges-one");
+    let (second, second_secret) = network("two-ranges-two");
+
+    let agent = TestAgent::spawn(&discovery).await.unwrap();
+    let one = agent
+        .agent
+        .join_network(&first, &first_secret)
+        .await
+        .unwrap();
+    let two = agent
+        .agent
+        .join_network(&second, &second_secret)
+        .await
+        .unwrap();
+
+    let own = agent.agent.endpoint_id();
+    let addresses = wait_until("both networks allocate an address", || {
+        let handle = agent.agent.clone();
+        async move {
+            let mine = async |network| {
+                handle
+                    .network_status(network)
+                    .await
+                    .ok()?
+                    .members
+                    .into_iter()
+                    .find(|member| member.endpoint_id == own)?
+                    .overlay_address_v4
+            };
+            Some((mine(one).await?, mine(two).await?))
+        }
+    })
+    .await;
+
+    assert_ne!(
+        addresses.0, addresses.1,
+        "different networks, different addresses"
+    );
+    assert!(
+        tsunagi::state::DEFAULT_IPV4_RANGE.contains(addresses.0),
+        "the first keeps the configured range: {}",
+        addresses.0
+    );
+    assert!(
+        tsunagi::state::derived_ipv4_range(two).contains(addresses.1),
+        "the second is in the range derived from its id: {}",
+        addresses.1
+    );
+    assert!(
+        !tsunagi::state::DEFAULT_IPV4_RANGE.contains(addresses.1),
+        "and not in the one the first network holds"
+    );
+
+    // And the ranges do not overlap, which is what lets one interface
+    // carry both without a packet being ambiguous.
+    assert!(
+        !tsunagi::state::derived_ipv4_range(two).contains(addresses.0),
+        "the two ranges must not overlap"
+    );
+
+    agent.agent.shutdown().await;
+}

crates/tsunagi-core/tests/resilience.rs

@@ -3,18 +3,18 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
 use std::net::SocketAddr;
 use std::time::Duration;
 
-use common::{TestAgent, config_with, local_config, network, settle, wait_event, wait_for_peers};
 use iroh::{EndpointAddr, SecretKey};
 use tsunagi::agent::Event;
 use tsunagi::config::ReconnectPolicy;
 use tsunagi::discovery::SharedMemoryDiscovery;
 use tsunagi::identity::NetworkKeys;
 use tsunagi::proto::ControlMessage;
+use tsunagi::testing::{
+    TestAgent, config_with, local_config, network, settle, wait_event, wait_for_peers,
+};
 use tsunagi::{Agent, Error};
 
 /// An endpoint id nobody is listening for, at an address nothing answers on.

crates/tsunagi-core/tests/restart.rs

@@ -2,13 +2,11 @@
 
 #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 
-mod common;
-
-use common::{TestAgent, config_with, network, settle, wait_event, wait_for_peers};
 use tsunagi::agent::Event;
 use tsunagi::discovery::SharedMemoryDiscovery;
 use tsunagi::identity::{NetworkName, NetworkSecret};
 use tsunagi::proto::ControlMessage;
+use tsunagi::testing::{TestAgent, config_with, network, settle, wait_event, wait_for_peers};
 use tsunagi::{Agent, Error};
 
 #[tokio::test]

crates/tsunagi-tcp-tls/Cargo.toml

@@ -0,0 +1,41 @@
+[package]
+name = "tsunagi-tcp-tls"
+version = "0.1.0"
+description = "TCP TLS 443 transport and E2EE tunnel protocol for tsunagi mesh networks."
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[dependencies]
+tsunagi-core = { path = "../tsunagi-core", version = "0.1.0" }
+iroh.workspace = true
+tokio = { workspace = true, features = ["net", "io-util", "time", "sync", "macros"] }
+rustls = "0.23"
+tokio-rustls = "0.26"
+rcgen = { version = "0.14", default-features = false, features = ["ring"] }
+x25519-dalek = { version = "2.0", features = ["static_secrets"] }
+chacha20poly1305 = "0.10"
+sha2 = "0.11"
+hkdf = "0.13"
+rand = "0.10"
+subtle = "2.6"
+zeroize = { version = "1.9", features = ["derive"] }
+bytes.workspace = true
+serde.workspace = true
+postcard.workspace = true
+tracing.workspace = true
+thiserror.workspace = true
+hex.workspace = true
+netdev = "0.46"
+
+[target.'cfg(target_os = "linux")'.dependencies]
+caps = "0.5"
+
+[dev-dependencies]
+tsunagi-core = { path = "../tsunagi-core", features = ["testing"] }
+tempfile.workspace = true
+tracing-subscriber.workspace = true
+
+[lints]
+workspace = true

crates/tsunagi-tcp-tls/src/cert.rs

@@ -0,0 +1,147 @@
+//! TLS certificate generation and verification.
+//!
+//! Generates a self-signed TLS certificate derived from the device's persistent
+//! identity, and verifies remote certificates against the expected peer's
+//! [`iroh::EndpointId`].
+
+use std::sync::Arc;
+
+use iroh::EndpointId;
+use rustls::client::danger::{ServerCertVerified, ServerCertVerifier};
+use rustls::pki_types::{CertificateDer, ServerName, UnixTime};
+use rustls::{ClientConfig, Error as RustlsError, ServerConfig};
+
+/// Verifies that the server certificate's public key matches the expected peer.
+#[derive(Debug)]
+pub struct PinnedEndpointVerifier {
+    expected: EndpointId,
+}
+
+impl PinnedEndpointVerifier {
+    /// Creates a verifier expecting `peer`.
+    pub fn new(peer: EndpointId) -> Self {
+        Self { expected: peer }
+    }
+}
+
+impl ServerCertVerifier for PinnedEndpointVerifier {
+    fn verify_server_cert(
+        &self,
+        end_entity: &CertificateDer<'_>,
+        _intermediates: &[CertificateDer<'_>],
+        _server_name: &ServerName<'_>,
+        _ocsp_response: &[u8],
+        _now: UnixTime,
+    ) -> Result<ServerCertVerified, RustlsError> {
+        // Parse the certificate's SPKI / public key
+        let cert_bytes = end_entity.as_ref();
+        let expected_bytes = self.expected.as_bytes();
+
+        // Check if the expected 32-byte Ed25519 public key is in the certificate
+        if cert_bytes
+            .windows(32)
+            .any(|window| window == expected_bytes)
+        {
+            Ok(ServerCertVerified::assertion())
+        } else {
+            Err(RustlsError::InvalidCertificate(
+                rustls::CertificateError::NotValidForName,
+            ))
+        }
+    }
+
+    fn verify_tls12_signature(
+        &self,
+        _message: &[u8],
+        _cert: &CertificateDer<'_>,
+        _dss: &rustls::DigitallySignedStruct,
+    ) -> Result<rustls::client::danger::HandshakeSignatureValid, RustlsError> {
+        // We require TLS 1.3
+        Err(RustlsError::PeerIncompatible(
+            rustls::PeerIncompatible::NoSignatureSchemesInCommon,
+        ))
+    }
+
+    fn verify_tls13_signature(
+        &self,
+        message: &[u8],
+        cert: &CertificateDer<'_>,
+        dss: &rustls::DigitallySignedStruct,
+    ) -> Result<rustls::client::danger::HandshakeSignatureValid, RustlsError> {
+        rustls::crypto::verify_tls13_signature(
+            message,
+            cert,
+            dss,
+            &rustls::crypto::ring::default_provider().signature_verification_algorithms,
+        )
+    }
+
+    fn supported_verify_schemes(&self) -> Vec<rustls::SignatureScheme> {
+        vec![rustls::SignatureScheme::ED25519]
+    }
+}
+
+/// Default SNI domain used to disguise TLS connections.
+pub const DEFAULT_SNI: &str = "cloudflare.com";
+
+/// Generates a self-signed certificate and private key for TLS 1.3 using
+/// a given 32-byte Ed25519 secret key and specified SNI domain.
+pub fn generate_self_signed_cert(
+    secret_key_bytes: &[u8; 32],
+    sni: &str,
+) -> Result<
+    (
+        CertificateDer<'static>,
+        rustls::pki_types::PrivateKeyDer<'static>,
+    ),
+    Box<dyn std::error::Error + Send + Sync>,
+> {
+    // PKCS#8 prefix for v1 Ed25519 private keys:
+    // 30 2e 02 01 00 30 05 06 03 2b 65 70 04 22 04 20 <32 bytes>
+    let mut pkcs8 = Vec::with_capacity(48);
+    pkcs8.extend_from_slice(&[
+        0x30, 0x2e, 0x02, 0x01, 0x00, 0x30, 0x05, 0x06, 0x03, 0x2b, 0x65, 0x70, 0x04, 0x22, 0x04,
+        0x20,
+    ]);
+    pkcs8.extend_from_slice(secret_key_bytes);
+
+    let key_der = rustls::pki_types::PrivatePkcs8KeyDer::from(pkcs8);
+    let key_pair = rcgen::KeyPair::from_pkcs8_der_and_sign_algo(&key_der, &rcgen::PKCS_ED25519)?;
+    let mut params = rcgen::CertificateParams::new(vec![sni.to_string()])?;
+    params
+        .distinguished_name
+        .push(rcgen::DnType::CommonName, sni);
+
+    let cert = params.self_signed(&key_pair)?;
+    let cert_der = CertificateDer::from(cert.der().to_vec());
+    let priv_key = rustls::pki_types::PrivateKeyDer::Pkcs8(key_der);
+
+    Ok((cert_der, priv_key))
+}
+
+/// Creates a `rustls::ServerConfig` for the node.
+pub fn make_server_config(
+    cert: CertificateDer<'static>,
+    key: rustls::pki_types::PrivateKeyDer<'static>,
+) -> Result<Arc<ServerConfig>, rustls::Error> {
+    let provider = Arc::new(rustls::crypto::ring::default_provider());
+    let mut config = ServerConfig::builder_with_provider(provider)
+        .with_safe_default_protocol_versions()?
+        .with_no_client_auth()
+        .with_single_cert(vec![cert], key)?;
+    config.alpn_protocols = vec![b"http/1.1".to_vec()];
+    Ok(Arc::new(config))
+}
+
+/// Creates a `rustls::ClientConfig` expecting a specific peer's `EndpointId`.
+pub fn make_client_config(expected_peer: EndpointId) -> Result<Arc<ClientConfig>, rustls::Error> {
+    let provider = Arc::new(rustls::crypto::ring::default_provider());
+    let verifier = Arc::new(PinnedEndpointVerifier::new(expected_peer));
+    let mut config = ClientConfig::builder_with_provider(provider)
+        .with_safe_default_protocol_versions()?
+        .dangerous()
+        .with_custom_certificate_verifier(verifier)
+        .with_no_client_auth();
+    config.alpn_protocols = vec![b"http/1.1".to_vec()];
+    Ok(Arc::new(config))
+}

crates/tsunagi-tcp-tls/src/crypto.rs

@@ -0,0 +1,104 @@
+//! End-to-end packet encryption for `tcp-tls`.
+//!
+//! Uses X25519 Diffie-Hellman key exchange and ChaCha20-Poly1305 AEAD.
+//! This ensures that even when traffic passes through intermediate relay nodes,
+//! the intermediate nodes cannot inspect or tamper with the payload.
+
+use std::sync::atomic::{AtomicU64, Ordering};
+
+use chacha20poly1305::aead::{Aead, KeyInit};
+use chacha20poly1305::{ChaCha20Poly1305, Nonce};
+use hkdf::Hkdf;
+use iroh::EndpointId;
+use sha2::Sha256;
+use tsunagi::dataplane::PluginError;
+use x25519_dalek::{PublicKey, StaticSecret};
+
+/// Length of the X25519 public key.
+pub const KEY_LEN: usize = 32;
+
+/// Domain separator for session key derivation.
+const HKDF_SALT: &[u8] = b"tsunagi/e2ee/noise/v1";
+
+/// A session cipher for communicating with one peer.
+pub struct PeerCipher {
+    tx_cipher: ChaCha20Poly1305,
+    rx_cipher: ChaCha20Poly1305,
+    tx_nonce: AtomicU64,
+}
+
+impl PeerCipher {
+    /// Establishes a new cipher between `local_id` and `remote_id` given
+    /// our local secret and their public key.
+    pub fn new(
+        local_id: EndpointId,
+        remote_id: EndpointId,
+        local_secret: &StaticSecret,
+        remote_public: &PublicKey,
+    ) -> Self {
+        let shared = local_secret.diffie_hellman(remote_public);
+        let hk = Hkdf::<Sha256>::new(Some(HKDF_SALT), shared.as_bytes());
+
+        let mut key_a_to_b = [0u8; 32];
+        let mut key_b_to_a = [0u8; 32];
+        let _ = hk.expand(b"tsunagi-key-a-to-b", &mut key_a_to_b);
+        let _ = hk.expand(b"tsunagi-key-b-to-a", &mut key_b_to_a);
+
+        // Deterministic role: smaller EndpointId is party A.
+        let is_party_a = local_id.as_bytes() < remote_id.as_bytes();
+        let (tx_key, rx_key) = if is_party_a {
+            (key_a_to_b, key_b_to_a)
+        } else {
+            (key_b_to_a, key_a_to_b)
+        };
+
+        let tx_cipher = ChaCha20Poly1305::new(&tx_key.into());
+        let rx_cipher = ChaCha20Poly1305::new(&rx_key.into());
+
+        Self {
+            tx_cipher,
+            rx_cipher,
+            tx_nonce: AtomicU64::new(0),
+        }
+    }
+
+    /// Encrypts an outbound packet. Prefixes the ciphertext with the 8-byte nonce.
+    pub fn encrypt(&self, packet: &[u8]) -> Result<bytes::Bytes, PluginError> {
+        let counter = self.tx_nonce.fetch_add(1, Ordering::Relaxed);
+        let mut nonce_bytes = [0u8; 12];
+        nonce_bytes[4..12].copy_from_slice(&counter.to_be_bytes());
+        let nonce = Nonce::from_slice(&nonce_bytes);
+
+        let ciphertext = self
+            .tx_cipher
+            .encrypt(nonce, packet)
+            .map_err(|err| PluginError::Other(format!("encryption failed: {err}")))?;
+
+        let mut out = Vec::with_capacity(8 + ciphertext.len());
+        out.extend_from_slice(&counter.to_be_bytes());
+        out.extend_from_slice(&ciphertext);
+        Ok(bytes::Bytes::from(out))
+    }
+
+    /// Decrypts an inbound packet prefixed with the 8-byte nonce.
+    pub fn decrypt(&self, payload: &[u8]) -> Result<bytes::Bytes, PluginError> {
+        if payload.len() < 8 + 16 {
+            return Err(PluginError::Other("payload too short".into()));
+        }
+
+        let counter_bytes: [u8; 8] = match payload[..8].try_into() {
+            Ok(b) => b,
+            Err(_) => return Err(PluginError::Other("invalid nonce".into())),
+        };
+        let mut nonce_bytes = [0u8; 12];
+        nonce_bytes[4..12].copy_from_slice(&counter_bytes);
+        let nonce = Nonce::from_slice(&nonce_bytes);
+
+        let plaintext = self
+            .rx_cipher
+            .decrypt(nonce, &payload[8..])
+            .map_err(|err| PluginError::Other(format!("decryption failed: {err}")))?;
+
+        Ok(bytes::Bytes::from(plaintext))
+    }
+}