feat(agent): switch LLM client from Ollama to OpenAI-compatible API (LM Studio support)

- Replace /api/chat with /v1/chat/completions endpoint
- Use json_schema response_format (LM Studio does not support json_object)
- Make schema parameter optional in call_ollama to support different schemas per use case
- Add dedicated normalize schema (normalized_metadata) with release_kind field
  instead of release_type to avoid model repetition loops
- Add dedicated merge schema (artist_merge) so model no longer confuses
  normalize and merge response structures
- Add retry with frequency_penalty=1.5 on parse failure to suppress repetition
- Add id3 crate as fallback metadata reader for MP3 files with large embedded
  cover art that exceed Symphonia probe limit of 1MB

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

Cargo.lock

@@ -1191,6 +1191,7 @@ dependencies = [
  "tokio",
  "tokio-util",
  "tower 0.4.13",
+ "tower-http",
  "tracing",
  "tracing-subscriber",
  "urlencoding",

furumi-agent/src/ingest/normalize.rs

@@ -25,16 +25,35 @@ pub async fn normalize(
 ) -> anyhow::Result<NormalizedFields> {
     let user_message = build_user_message(raw, hints, similar_artists, similar_albums, folder_ctx);
 
+    let schema = normalize_schema();
     let response = call_ollama(
         &state.config.ollama_url,
         &state.config.ollama_model,
         &state.system_prompt,
         &user_message,
         state.config.ollama_auth.as_deref(),
+        0.5,
+        Some(("normalized_metadata", schema.clone())),
     )
     .await?;
 
-    parse_response(&response)
+    match parse_response(&response) {
+        Ok(fields) => Ok(fields),
+        Err(e) => {
+            tracing::warn!(error = %e, "LLM parse failed, retrying with higher frequency_penalty");
+            let response2 = call_ollama(
+                &state.config.ollama_url,
+                &state.config.ollama_model,
+                &state.system_prompt,
+                &user_message,
+                state.config.ollama_auth.as_deref(),
+                1.5,
+                Some(("normalized_metadata", schema)),
+            )
+            .await?;
+            parse_response(&response2)
+        }
+    }
 }
 
 fn build_user_message(
@@ -116,9 +135,12 @@ fn build_user_message(
 struct ChatRequest {
     model: String,
     messages: Vec<ChatMessage>,
-    response_format: ChatResponseFormat,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    response_format: Option<ChatResponseFormat>,
     stream: bool,
     temperature: f64,
+    max_tokens: u32,
+    frequency_penalty: f64,
 }
 
 #[derive(Serialize)]
@@ -162,11 +184,18 @@ pub async fn call_ollama(
     system_prompt: &str,
     user_message: &str,
     auth: Option<&str>,
+    frequency_penalty: f64,
+    schema: Option<(&str, serde_json::Value)>,
 ) -> anyhow::Result<String> {
     let client = reqwest::Client::builder()
         .timeout(std::time::Duration::from_secs(120))
         .build()?;
 
+    let response_format = schema.map(|(name, schema)| ChatResponseFormat {
+        kind: "json_schema".to_owned(),
+        json_schema: JsonSchemaWrapper { name: name.to_owned(), strict: true, schema },
+    });
+
     let request = ChatRequest {
         model: model.to_owned(),
         messages: vec![
@@ -179,32 +208,11 @@ pub async fn call_ollama(
                 content: user_message.to_owned(),
             },
         ],
-        response_format: ChatResponseFormat {
-            kind: "json_schema".to_owned(),
-            json_schema: JsonSchemaWrapper {
-                name: "normalized_metadata".to_owned(),
-                strict: true,
-                schema: serde_json::json!({
-                    "type": "object",
-                    "properties": {
-                        "artist":           { "type": ["string", "null"] },
-                        "album":            { "type": ["string", "null"] },
-                        "title":            { "type": ["string", "null"] },
-                        "year":             { "type": ["integer", "null"] },
-                        "track_number":     { "type": ["integer", "null"] },
-                        "genre":            { "type": ["string", "null"] },
-                        "featured_artists": { "type": "array", "items": { "type": "string" } },
-                        "release_type":     { "type": ["string", "null"] },
-                        "confidence":       { "type": ["number", "null"] },
-                        "notes":            { "type": ["string", "null"] }
-                    },
-                    "required": ["artist", "album", "title", "year", "track_number", "genre", "featured_artists", "release_type", "confidence", "notes"],
-                    "additionalProperties": false
-                }),
-            },
-        },
+        response_format,
         stream: false,
         temperature: 0.1,
+        max_tokens: 512,
+        frequency_penalty,
     };
 
     let url = format!("{}/v1/chat/completions", base_url.trim_end_matches('/'));
@@ -242,6 +250,26 @@ pub async fn call_ollama(
     Ok(content)
 }
 
+fn normalize_schema() -> serde_json::Value {
+    serde_json::json!({
+        "type": "object",
+        "properties": {
+            "artist":           { "type": ["string", "null"] },
+            "album":            { "type": ["string", "null"] },
+            "title":            { "type": ["string", "null"] },
+            "year":             { "type": ["integer", "null"] },
+            "track_number":     { "type": ["integer", "null"] },
+            "genre":            { "type": ["string", "null"] },
+            "featured_artists": { "type": "array", "items": { "type": "string" } },
+            "release_kind":     { "type": ["string", "null"] },
+            "confidence":       { "type": ["number", "null"] },
+            "notes":            { "type": ["string", "null"] }
+        },
+        "required": ["artist", "album", "title", "year", "track_number", "genre", "featured_artists", "release_kind", "confidence", "notes"],
+        "additionalProperties": false
+    })
+}
+
 /// Parse the LLM JSON response into NormalizedFields.
 /// Handles both clean JSON and JSON wrapped in markdown code fences.
 fn parse_response(response: &str) -> anyhow::Result<NormalizedFields> {
@@ -266,6 +294,7 @@ fn parse_response(response: &str) -> anyhow::Result<NormalizedFields> {
         genre: Option<String>,
         #[serde(default)]
         featured_artists: Vec<String>,
+        #[serde(rename = "release_kind")]
         release_type: Option<String>,
         confidence: Option<f64>,
         notes: Option<String>,

furumi-agent/src/merge.rs

@@ -35,12 +35,38 @@ pub async fn propose_merge(state: &Arc<AppState>, merge_id: Uuid) -> anyhow::Res
 
     let user_message = build_merge_message(&artists_data);
 
+    let schema = serde_json::json!({
+        "type": "object",
+        "properties": {
+            "canonical_artist_name": { "type": "string" },
+            "winner_artist_id":      { "type": "integer" },
+            "album_mappings": {
+                "type": "array",
+                "items": {
+                    "type": "object",
+                    "properties": {
+                        "source_album_id":     { "type": "integer" },
+                        "canonical_name":      { "type": "string" },
+                        "merge_into_album_id": { "type": ["integer", "null"] }
+                    },
+                    "required": ["source_album_id", "canonical_name", "merge_into_album_id"],
+                    "additionalProperties": false
+                }
+            },
+            "notes": { "type": "string" }
+        },
+        "required": ["canonical_artist_name", "winner_artist_id", "album_mappings", "notes"],
+        "additionalProperties": false
+    });
+
     let response = call_ollama(
         &state.config.ollama_url,
         &state.config.ollama_model,
         &state.merge_prompt,
         &user_message,
         state.config.ollama_auth.as_deref(),
+        0.5,
+        Some(("artist_merge", schema)),
     ).await?;
 
     let proposal = parse_merge_response(&response)?;