feat(agent): add decade/genre filters, tune model params, update backlog

Add genre, decade, year_from, year_to filters to SearchLibrary tool and
LibraryQuery. Parse decade strings ("90s" -> 1990-1999) programmatically.
Tune build_agent() params: temperature 0.3, max_tokens 2048, repeat_penalty
1.1. Add playlist naming guidance and decade strategy to system prompt.
Update task-277 ACs and notes with migration details.

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

Author: pythoninthegrass

AGENTS.md

@@ -25,13 +25,15 @@ Always use Context7 MCP when I need library/API documentation, code generation,
 - microsoft/playwright
 - mrlesk/backlog.md
 - nextest-rs/nextest
+- proptest-rs/proptest
 - serial-ata/lofty-rs
 - sharkdp/hyperfine
 - taiko2k/tauon
 - tailwindlabs/tailwindcss
 - tranxuanthang/lrclib
 - websites/deno
 - websites/last_fm_api
+- websites/rs_rig-core
 - websites/rs_tauri_2_9_5
 - websites/taskfile_dev
 

backlog/tasks/task-277 - Genius-playlist-creator.md

@@ -4,7 +4,7 @@ title: Genius playlist creator
 status: In Progress
 assignee: []
 created_date: '2026-02-18 05:58'
-updated_date: '2026-04-02 22:34'
+updated_date: '2026-04-04 03:49'
 labels:
   - feature
   - playlists
@@ -14,6 +14,7 @@ labels:
   - lastfm
 dependencies:
   - TASK-308
+  - TASK-277.1
 references:
   - docs/genius.md
 priority: high
@@ -32,13 +33,13 @@ Design document: `~/.claude/plans/steady-juggling-scroll.md`
 
 ## Acceptance Criteria
 <!-- AC:BEGIN -->
-- [ ] #1 User can enter a natural language prompt to generate a playlist
-- [ ] #2 System uses local LLM (Ollama + llama3.2:1b) to interpret prompts and call appropriate tools
-- [ ] #3 8 tools available: GetRecentlyPlayed, GetTopArtists, SearchLibrary, GetSimilarTracks, GetSimilarArtists, GetTrackTags, GetTopArtistsByTag, GetTopTracksByCountry
-- [ ] #4 Generated playlist only contains tracks that exist in user's local library
-- [ ] #5 Graceful degradation: works without Last.fm (local-only tools), works without Ollama (returns setup instructions)
+- [x] #1 User can enter a natural language prompt to generate a playlist
+- [x] #2 System uses local LLM (Ollama + qwen3.5:9b) to interpret prompts and call appropriate tools
+- [x] #3 8 tools available: GetRecentlyPlayed, GetTopArtists, SearchLibrary, GetSimilarTracks, GetSimilarArtists, GetTrackTags, GetTopArtistsByTag, GetTopTracksByCountry
+- [x] #4 Generated playlist only contains tracks that exist in user's local library
+- [x] #5 Graceful degradation: works without Last.fm (local-only tools), works without Ollama (returns setup instructions)
 - [x] #6 Feature-flagged behind `agent` — zero overhead when disabled
-- [ ] #7 Onboarding wizard: Ollama check → model download → ready
+- [x] #7 Onboarding wizard: Ollama check → model download → ready
 - [x] #8 Agent evals pass (tool selection, output format, degradation)
 <!-- AC:END -->
 
@@ -214,6 +215,34 @@ Total: 764 tests pass (762 existing + 2 new)
 - Mixed-history request (`make me a chill playlist like what I listened to last Friday`): default prompt used 4 turns; override prompt used 2 turns and treated weak recent-history results as a weighting signal instead of spending extra turns matching them exactly.
 
 2026-04-02: Conclusion from prompt experiments: tighter stop rules materially reduce turn count, but prompt-only business-rule enforcement remains unreliable for cases like seed-artist caps. The most promising direction is to keep LLM-driven discovery/tool routing while moving playlist compilation and policy enforcement into deterministic business logic that scores and filters candidates using empirical evidence (tool source overlap, local genre, Last.fm tags/similarity, last played date, play history, and explicit duplicate-artist caps).
+
+## 2026-04-03: Python-to-Rust migration finalized
+
+Swapped AC#2 model from llama3.2:1b to qwen3.5:9b (matches Python agent). Ported remaining Python logic to Rust via rig:
+
+### prompt.rs
+- Added PLAYLIST NAMING section (creative synonyms, not parroting user's words)
+- Added decade/era strategy (search_library with decade+genre, parallel get_top_artists_by_tag)
+- Documented all search_library filters (keyword, artist, album, genre, decade, year range)
+- Clarified search_library(query=...) vs search_library(genre=...) in CRITICAL section
+
+### tools.rs — SearchLibrary expanded
+- Added genre, decade, year_from, year_to args + tool definition
+- Programmatic `parse_decade()` handles any century ("90s" -> 1990-1999, "1780s" -> 1780-1789)
+- 6 new tests: parse_decade (4) + genre/decade filter integration (2)
+
+### db/library.rs — LibraryQuery expanded
+- Added genre (LIKE), year_from, year_to fields to LibraryQuery struct
+- Added filtering logic in get_all_tracks() SQL builder
+
+### mod.rs — build_agent() tuned
+- Temperature: 0.2 -> 0.3 (matches Python)
+- max_tokens: 1024 -> 2048 (prevents response truncation)
+- Added repeat_penalty: 1.1 (prevents token repetition/gibberish)
+- Preamble set via rig's `.preamble()` builder method
+
+### Test coverage
+770 tests pass (764 existing + 6 new). No regressions with or without agent feature.
 <!-- SECTION:NOTES:END -->
 
 ## Definition of Done

backlog/tasks/task-277.1 - Refresh-Genius-chat-style-UI.md

@@ -1,9 +1,10 @@
 ---
 id: TASK-277.1
 title: Refresh Genius chat-style UI
-status: To Do
+status: In Progress
 assignee: []
 created_date: '2026-04-04 03:38'
+updated_date: '2026-04-04 03:43'
 labels:
   - feature
   - frontend
@@ -14,6 +15,7 @@ documentation:
   - docs/genius.md
 parent_task_id: TASK-277
 priority: medium
+ordinal: 2500
 ---
 
 ## Description

backlog/tasks/task-301 - Global-playback-control-shortcuts-stop-after-current-etc.-from-all-views.md

@@ -1,17 +1,17 @@
 ---
 id: TASK-301
 title: 'Global playback control shortcuts (stop after current, etc.) from all views'
-status: In Progress
+status: To Do
 assignee: []
 created_date: '2026-03-31 03:36'
-updated_date: '2026-03-31 03:38'
+updated_date: '2026-04-04 03:43'
 labels:
   - enhancement
   - ux
   - playback
 dependencies: []
 priority: medium
-ordinal: 1500
+ordinal: 2859.375
 ---
 
 ## Description

crates/mt-tauri/src/agent/mod.rs

@@ -31,18 +31,21 @@ use crate::lastfm::LastFmClient;
 /// Generate a unique playlist name by appending a number if needed.
 ///
 /// If the base name exists, tries "Name (2)", "Name (3)", etc.
-fn generate_unique_playlist_name(conn: &rusqlite::Connection, base_name: &str) -> crate::db::DbResult<String> {
+fn generate_unique_playlist_name(
+    conn: &rusqlite::Connection,
+    base_name: &str,
+) -> crate::db::DbResult<String> {
     // Check if base name is available
     let exists: bool = conn.query_row(
         "SELECT EXISTS(SELECT 1 FROM playlists WHERE name = ? LIMIT 1)",
         [base_name],
         |row| row.get(0),
     )?;
-    
+
     if !exists {
         return Ok(base_name.to_string());
     }
-    
+
     // Find an available suffix
     for i in 2..=100 {
         let candidate = format!("{} ({})", base_name, i);
@@ -55,7 +58,7 @@ fn generate_unique_playlist_name(conn: &rusqlite::Connection, base_name: &str) -
             return Ok(candidate);
         }
     }
-    
+
     // Fallback: append timestamp if all numbers are taken
     use std::time::{SystemTime, UNIX_EPOCH};
     let timestamp = SystemTime::now()
@@ -79,10 +82,7 @@ fn shuffle_spread_artists(tracks: &[(i64, String)]) -> Vec<i64> {
     let mut artist_keys: HashMap<i64, &str> = HashMap::new();
 
     for (id, artist) in tracks {
-        by_artist
-            .entry(artist.as_str())
-            .or_default()
-            .push(*id);
+        by_artist.entry(artist.as_str()).or_default().push(*id);
         artist_keys.insert(*id, artist.as_str());
     }
 
@@ -189,10 +189,11 @@ pub(crate) fn build_agent(
     client
         .agent(DEFAULT_MODEL)
         .preamble(&system_prompt)
-        .temperature(0.2)
-        .max_tokens(1024)
+        .temperature(0.3)
+        .max_tokens(2048)
         .additional_params(serde_json::json!({
             "top_p": 0.9,
+            "repeat_penalty": 1.1,
         }))
         .tool(GetRecentlyPlayed {
             ctx: Arc::clone(&ctx),
@@ -323,7 +324,9 @@ pub async fn agent_generate_playlist(
         .with_conn(|conn| {
             let mut results = Vec::new();
             for id in &parsed.track_ids {
-                let mut stmt = conn.prepare("SELECT id, artist FROM library WHERE id = ? AND missing = 0 LIMIT 1")?;
+                let mut stmt = conn.prepare(
+                    "SELECT id, artist FROM library WHERE id = ? AND missing = 0 LIMIT 1",
+                )?;
                 let mut rows = stmt.query([id])?;
                 if let Some(row) = rows.next()? {
                     let track_id: i64 = row.get(0)?;
@@ -337,7 +340,7 @@ pub async fn agent_generate_playlist(
 
     if track_details.is_empty() {
         return Ok(AgentResponse::error(
-            "None of the selected tracks exist in your library".to_string()
+            "None of the selected tracks exist in your library".to_string(),
         ));
     }
 
@@ -355,7 +358,7 @@ pub async fn agent_generate_playlist(
     let playlist_name: String = db
         .with_conn(|conn| generate_unique_playlist_name(conn, &parsed.name))
         .map_err(|e: crate::db::DbError| format!("Failed to generate playlist name: {e}"))?;
-    
+
     let playlist = db
         .with_conn(|conn| playlists::create_playlist(conn, &playlist_name))
         .map_err(|e| format!("Failed to create playlist: {e}"))?;
@@ -371,9 +374,7 @@ pub async fn agent_generate_playlist(
     };
 
     let added = db
-        .with_conn(|conn| {
-            playlists::add_tracks_to_playlist(conn, playlist.id, &shuffled_ids, None)
-        })
+        .with_conn(|conn| playlists::add_tracks_to_playlist(conn, playlist.id, &shuffled_ids, None))
         .map_err(|e| format!("Failed to add tracks to playlist: {e}"))?;
 
     info!(
@@ -694,23 +695,23 @@ mod tests {
     #[test]
     fn unique_name_appends_number_when_exists() {
         let db = Database::new_in_memory().expect("in-memory db");
-        
+
         // Create first playlist
         db.with_conn(|conn| playlists::create_playlist(conn, "Chill Vibes"))
             .expect("create first")
             .expect("first playlist created");
-        
+
         // Second should get (2)
         let name2 = db
             .with_conn(|conn| generate_unique_playlist_name(conn, "Chill Vibes"))
             .expect("generate name2");
         assert_eq!(name2, "Chill Vibes (2)");
-        
+
         // Create second playlist
         db.with_conn(|conn| playlists::create_playlist(conn, &name2))
             .expect("create second")
             .expect("second playlist created");
-        
+
         // Third should get (3)
         let name3 = db
             .with_conn(|conn| generate_unique_playlist_name(conn, "Chill Vibes"))

crates/mt-tauri/src/agent/prompt.rs

@@ -35,23 +35,35 @@ STRATEGY — pick the approach that fits the request:
 - Artist-based requests ("similar to Radiohead", "like Bjork"):
   Call get_similar_artists AND search_library(artist=...) in parallel on the first turn.
   Then use get_similar_tracks on seed tracks to expand.
+- Decade/era requests ("90s rock", "80s pop", "2000s indie"):
+  Call search_library with decade="90s" AND genre="rock" (or relevant genre) with limit=50.
+  Also call get_top_artists_by_tag with era-appropriate genre tags IN PARALLEL.
+  The library has year metadata on most tracks — USE IT instead of guessing artist names.
 - General/mixed requests:
   Use get_recently_played or get_top_artists to understand listening habits, then combine
   with get_similar_tracks, get_similar_artists, or get_top_artists_by_tag.
 - Regional requests ("Japanese music", "Brazilian"):
   Use get_top_tracks_by_country with limit=50.
-- search_library is ONLY for: exact artist names, exact album names, or specific song titles. NEVER for mood keywords.
+- search_library supports keyword, artist, album, genre, decade, and year range filters.
+- search_library is ONLY for: exact artist names, exact album names, specific song titles, genre/decade filtering. NEVER for mood keywords.
 - Use get_track_tags to understand a track's mood/genre before expanding with get_top_artists_by_tag.
 
 CRITICAL: Avoid these common mistakes:
 - NEVER call search_library with mood words like "chill", "relax", "calm", "soft", "dream", "slow" — this matches titles containing those words, not actual chill music
-- NEVER call search_library with genre words like "ambient", "electronic", "indie" — use get_top_artists_by_tag instead
+- NEVER call search_library(query=...) with genre words like "ambient", "electronic", "indie" — use get_top_artists_by_tag or search_library(genre=...) instead
 - If get_top_artists_by_tag returns 0 matches, try related tags (e.g., "ambient" -> "chillout", "electronic" -> "electronica") rather than falling back to search_library
 
 RESPONSE FORMAT (final answer only):
 Playlist: [descriptive name]
 Tracks: [comma-separated track IDs]
 
+PLAYLIST NAMING:
+- Use a creative synonym or evocative phrase, not the user's exact words
+- "chill" -> "Midnight Drift", "Velvet Haze", "Slow Burn Frequencies"
+- "upbeat" -> "Solar Flare", "Electric Momentum", "Daybreak Drive"
+- "sad" -> "Rain on Glass", "Quiet Ache", "Blue Hour Confessions"
+- Capture the FEELING, don't parrot the request
+
 Only include track IDs you received from tool results. Never invent IDs."#
     )
 }