A reload or a Play after a pause continues from the register instead of
this page's saved position, a seek during preparation is kept, a player
that keeps listening after being superseded writes again as a new
session, a player that loses the playback role records its listening,
marking the loaded episode closes its session, and Prev no longer
resets an episode to 0.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The enclosure type of a feed or of a client-sent episode reference no
longer makes an application/octet-stream response acceptable.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Feed downloads and proxy streams have per-account and server-wide
limits, new subscriptions stop at 1000 per account, and scheduled
refreshes take accounts in turns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Users see the kind of failure only; resolved addresses and transport
details go to the logs, without URLs. Feed and audio requests are
limited to the web ports.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clippy flagged the helper that returned the 401 response as its error
variant (result_large_err); it now returns the account id as an Option.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Playing, completing or restarting an episode from the player updated
its row but left the unplayed and in-progress counters of the podcast
page as they were loaded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
README and the administration guide describe the Podcasts section,
resume and completion, handoffs, unsubscribing, what stays private and
what the publisher sees, and the security model of the optional audio
proxy. The development guide lists the endpoints, how the player
writes progress and the node tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The sidebar gets Podcasts: add an RSS feed (fetched by the server; a
refused or broken feed shows why), the subscriptions with unplayed
counts and the last fetch error, and a podcast page with seasons, an
episode search, played/sort filters, paging, manual refresh and
unsubscribe (the confirmation says progress and history are kept), or
subscribe again. Episodes show the saved position, the time left and a
played mark, with Play/Resume, Start over, Mark played, Mark unplayed
and Add to queue; Jam guests cannot add episodes. The podcast page also
lists the recent listening history.
The queue labels episodes, the podcast name of the current episode
opens its page, like and share are hidden for episodes, and the player
bar shows when an episode is loading or cannot be played or resumed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Episodes in the queue are no longer skipped. The player asks the
podcast API for the audio source and the resume position, plays the
publisher URL (or the same-origin proxy when the administrator enabled
it and the direct URL fails or is http:// on an https:// page) and
seeks before it starts, so it never reports a resumed episode that
plays from elsewhere: a server that cannot seek to the saved position
shows that and keeps the position until the user plays from the start.
Progress follows the contract: Play starts a listening session as an
explicit action above the handed-off generation, checkpoints go out
every 20 s and on pause, a seek is an explicit action, the end of the
stream completes the episode, and a stop, a page hide (beacon) or a
handoff saves the final position. Writes of a session run one after
another in the background; a superseded session stops writing, and a
device that loses the playback role ends its session without a report.
Reports always carry the position of their own episode.
Jam placeholders stay unplayable and episodes stay private for Jam
rules. A podcasts store holds the list, the podcast page and the
explicit episode actions for the UI of the next commit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A browser queue item keeps the handoff hint it arrived with, which is
older than the checkpoints this account has written since. Every
podcast episode in a playback state the web sends to a trusted device
now carries max(hint, register generation), so the device taking over
writes its explicit action above everything the web has seen. Which
state reads the resume position from the register is decided on the
converted queue, so the bridge itself stays testable without a
database.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
/api/podcasts now lists subscriptions with counters and fetch state,
subscribes by feed URL (fetched first; nothing is stored for a URL that
is not a public feed), shows a podcast with its seasons and history,
pages and searches episodes, and runs the explicit progress actions
(start over, played, unplayed) and unsubscribe/subscribe again.
The web player gets two endpoints: playback returns the audio source
and the resume position and, when the browser starts playing, writes a
new listening session as an explicit action above the carried handoff
generation; progress stores checkpoints, pause, stop and handoff at the
session generation, a seek as an explicit action, completes the episode
once at the end or within the completion tail, and records the
listening session. Handlers delegate to podcasts::api, which the new
Postgres test drives directly.
Podcast queue items are no longer marked unavailable in browser JSON
and carry the episode artwork. The temporary dead-code allowance of the
podcast module is gone; helpers used only by tests are cfg(test) and
complete_session gave way to complete_if_current.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Browsers stream episode audio from the publisher by default. For
publishers they cannot play directly (CORS, mixed content), the
administrator can enable podcast_audio_proxy_enabled (admin Settings,
FURU_PODCAST_AUDIO_PROXY_ENABLED). The endpoints live in a new app at
/api/podcasts:
- GET /episodes/{episode_id}/audio streams an episode of the signed-in
account. The URL comes from the account's catalog, never from the
request. Requests go through SafeHttp with one sanitized byte Range and
no client headers; 200/206/416 pass through, only audio content types
are served, with nosniff and a sandbox CSP, and responses stop at 1 GiB
or 4 hours.
- GET /episodes/{episode_id}/source returns the publisher URL and, when
enabled, the proxy URL.
- POST /{podcast_id}/refresh refreshes one subscribed feed now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The podcast_refresh job runs every five minutes and refreshes the podcasts
whose next check is due: conditional requests (ETag, Last-Modified), 32 MiB
and 60 s limits, the shared music-dht parser and PodcastStore::apply_feed.
A chain of permanent redirects moves the feed and becomes an alias.
Healthy feeds are checked hourly with a stable spread; failures back off
from 15 minutes up to a day, honouring Retry-After. Accounts subscribed to
the same feed URL share one request, at most four run at once, and one
run at a time holds an advisory lock. runner::refresh_now is the manual
refresh for one subscribed podcast, at most every 30 seconds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Feed and enclosure URLs come from users and third-party feeds. SafeHttp
accepts only http(s) URLs without credentials, resolves names with a
resolver that refuses the whole answer unless every address is public
(loopback, private, CGNAT, link-local, cloud metadata, unique local,
multicast, reserved; IPv4-mapped, NAT64 and 6to4 judged by their IPv4
part), and connects to exactly the checked addresses, so DNS rebinding
cannot swap them. IP literals and every redirect hop are checked; system
proxies, cookies and Referer are off. Loopback is reachable only through a
#[cfg(test)] policy used by the local test server.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Received podcast operations and snapshot sections are applied to the
podcast store, this device's snapshots are filled from it, and an
unknown episode position resumes from its progress register. The
relay module is removed; operations, gating and snapshot bounds on the
wire are unchanged. The device-sync Postgres tests now migrate their
database themselves and cover locally emitted podcast operations.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PodcastStore implements the music_dht::podcast contract per account:
catalog, feed aliases, episodes with a derived season view, the
subscription, share and progress registers, listening sessions and the
local-only fetch state. Received operations, snapshot records and local
actions share one apply path (materializers plus merge functions) under
a per-account advisory lock. Local actions append their operation to the
device-sync log with the next stamp of the web device and apply exactly
that operation in the same transaction.
Migration m_0056 creates the tables; m_0057 imports the registers that
0.13 kept in furumusic__fed_podcast_relay and the catalog rows of the
episode references in the operation log. The relay table is kept.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The podcast store resolves episode ids from parsed feeds with the
identity hints of music_dht::podcast::feed, which are part of the id
contract.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The ignored device-sync PostgreSQL test now also covers a handoff and a
remote Next onto an episode with an unknown position (the relayed
progress register supplies 1200 s and its generation), and checks that
0.12's operation types fail on a relayed podcast operation and read the
log again after the documented cleanup statement, which the m_0055
backwards operation now shares.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Release 0.12 fails a whole sync batch on a podcast operation kind, so
rolling back is only safe after the relayed podcast operations are
removed: document the step in docs/administration.md and run it in the
m_0055 backwards operation.
The typed queue now records the current track id and queue position
written with it, next to the legacy ids. A 0.12 release that only moved
to another track in the same queue no longer leaves a stale typed
position that the next upgrade would restore. Jam placeholders are never
stored in a user's own typed queue.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The server checked a participant's queue_remove/queue_clear against the
host state of the last report, but the host browser runs its pending
commands in one batch: after a move, a checked index can name a private
item. Participant commands are now stamped `jam_participant`, and the
host browser refuses to remove, clear, add or bring back its private
items for them, judged on the queue it is about to change.
A guest that stops controlling a Jam drops the host's private
placeholders from its own queue, so they are not saved or relayed as
the guest's queue.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The web player only knows a podcast episode's position from a handoff or
a seek. When Next, auto-advance, a click or a play command makes an
episode current, it now reports `position_unknown` instead of 0, and the
server takes the resume position (and the register generation) from the
relayed progress register before it hands the state to another device.
Remote Next/Prev/play commands to a device do the same.
Next with repeat all or shuffle wraps past a trailing unplayable item
instead of stopping on it. Prev leaves an unplayable item instead of
"restarting" it, and seeking such an item moves its kept position.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Use the shared music_dht::device_sync types instead of private copies
(byte-identical for v3 content, checked against the 0.12.x types and a
golden fixture) and announce protocol 4, because every v4 element is
kept:
- podcast operations stay in the operation log and are relayed only to
peers that reported v4; their registers and the podcast snapshot
sections are merged with the shared rules into an opaque relay table
and returned in this server's snapshots (m_0055);
- the browser bridge carries PlaybackTrack.source: items that are not
library tracks keep the verbatim wire item, so podcast episodes and
Jam placeholders survive the round trip; the player labels and skips
them instead of dropping them;
- the typed queue is persisted in queue_items_json next to queue_json,
which is still written for rollback (m_0054).
Peer versions come from self-reports (profile_protocol_version); relayed
profiles only raise them and DeviceProfileSet no longer stores this
server's constant. Invites carry the inviter's version and a
DeviceProfileSet announces v4 once. Handoffs of a podcast queue to a
device below v4 are refused with an update message, and its playback
commands are ignored while the queue has podcast items. Jam guests see
podcast items as private placeholders, which are mapped back on their
commands.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The podcast contract and device-sync protocol v4 live in music-dht 0.6.
Cargo.lock currently resolves it from a local patch of the frid branch
podcasts/contract and must be regenerated against the published crate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>