release: v1.5.7

BullMQ enrichment migration
- Rewrote enrichment pipeline on BullMQ v5: artist, track, podcast workers with
  pause/resume/stop support and Bull Board visibility
- Essentia publishes audio:analysis:complete events; CLAP subscribes reactively
  instead of polling (eliminates scan delay between phases)
- Thread-safe DB pool in CLAP worker, all DB calls in run_in_executor
- Fixed psycopg2.pool submodule import crash on BullMQ vibe worker startup
- Silenced EnrichmentStateService disconnect error on already-closed Redis conn

Enrichment fixes
- lastfmTags NULL caused mood-tags phase to silently skip all tracks; migration
  backfills NULL -> '{}' and sets column default
- Cover art fetch errors for temp-MBID albums (temp-* passed to Cover Art Archive)
- VIBE-VOCAB vocabulary JSON not copied to Docker image (TypeScript omits .json)
- Wired cleanupOldResolved() to run daily; added missing enrichment status indexes
- Fixed circuit breaker reset, orphan cleanup, podcast entityType, hang detection

Soulseek
- Track search-page downloads in activity tracker
- Use async fs.promises.access instead of synchronous existsSync
- Verify file exists on disk before emitting download:complete (#110)

Docker image size: 28.4 GB -> 12.2 GB
- Removed all CUDA/NVIDIA dependencies; switched to CPU-only PyTorch/TensorFlow
- tensorflow-cpu + essentia-tensorflow --no-deps (avoids GPU TF transitive dep)
- Fixed .dockerignore: **/node_modules and **/.next now excluded from build context

PWA / mobile
- Background audio session loss on iOS/Android: SilenceKeepalive singleton,
  tryResume() in MediaSession play handler, direct track load on 'ended' event,
  visibilitychange/pageshow foreground recovery
- Lock orientation to portrait for Android device lock (#117)

Discovery
- Retry All re-importing albums already in library: apply same three-level filter
  as GET /current before creating download jobs; delete stale UnavailableAlbum
  records for already-present albums. Closes #34

CI / release
- linux/arm64 added to release and nightly Docker builds (#87)
- Isolated release CI from nightly GHA cache (cache-from/cache-to removed from
  docker-publish.yml to guarantee clean release builds)
- Redis vm.overcommit_memory=1 sysctl added to prod and server compose files

Other fixes
- Cross-artist album fallback by title+year prevents library splitting (#50)
- Retry temp-MBID artists after 24h not 7 days; hide temp MBIDs from API (#112)
- 3-attempt ECONNRESET retry on all Deezer getPlaylist call sites (#119)
- check response.ok on health probe — fetch does not throw on 5xx (#104)
- Z-index stacking hierarchy established (MiniPlayer through OverlayPlayer)
- API token display overflow on iPhone (min-w-0/overflow-hidden on flex container)

Author: Your Name

.dockerignore

@@ -0,0 +1,23 @@
+.git
+.worktrees
+.claude
+.serena
+.aider-desk
+.aider.tags.cache.v4
+.roo
+.ruff_cache
+.vscode
+context_portal
+logs
+docs
+scripts
+*.md
+*.log
+**/node_modules
+**/.next
+.env*
+coverage
+*.test.ts
+__tests__
+android
+ios

.github/workflows/docker-nightly.yml

@@ -25,6 +25,9 @@ jobs:
                   sudo rm -rf /usr/local/share/boost
                   sudo rm -rf "$AGENT_TOOLSDIRECTORY"
 
+            - name: Set up QEMU
+              uses: docker/setup-qemu-action@v3
+
             - name: Set up Docker Buildx
               uses: docker/setup-buildx-action@v3
 
@@ -52,5 +55,4 @@ jobs:
                       org.opencontainers.image.version=nightly-${{ steps.sha.outputs.short }}
                   cache-from: type=gha
                   cache-to: type=gha,mode=max
-                  # ARM64 disabled due to QEMU emulation issues with npm packages
-                  platforms: linux/amd64
+                  platforms: linux/amd64,linux/arm64

.github/workflows/docker-publish.yml

@@ -60,9 +60,7 @@ jobs:
                       ${{ env.IMAGE_NAME }}:latest
                   cache-from: type=gha
                   cache-to: type=gha,mode=max
-                  # Note: ARM64 removed due to QEMU emulation issues with npm packages
-                  # Can be re-added when using native ARM64 runners
-                  platforms: linux/amd64
+                  platforms: linux/amd64,linux/arm64
 
     create-release:
         needs: [build]

CHANGELOG.md

@@ -5,12 +5,28 @@ All notable changes to Kima will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [1.5.6] - 2026-02-22
+## [1.5.7] - 2026-02-23
+
+### Added
+
+- **BullMQ enrichment infrastructure**: Rewrote the entire enrichment pipeline on top of BullMQ v5, replacing the custom BLPOP/Redis queue loops. Artist, track, and podcast enrichment all run as proper BullMQ Worker instances with job-level pause, resume, and stop support. All queues are visible in the Bull Board admin dashboard. The orchestrator pushes jobs into BullMQ and uses a sentinel pattern to track when all jobs in a phase have completed before advancing.
+- **Reactive vibe queuing**: The Essentia audio analyzer now publishes an `audio:analysis:complete` event to Redis when each track finishes. The CLAP service subscribes and immediately queues a vibe embedding job for that track — eliminating the previous polling-based approach where CLAP scanned the database on a fixed interval looking for newly-completed Essentia tracks.
 
 ### Fixed
 
+- **PWA background audio session lost on iOS and Android**: Pausing from lock-screen / notification controls while the app was backgrounded caused iOS to reclaim the audio session, blocking any subsequent `audio.play()` call until the app was foregrounded. Fixes two related symptoms: (1) resuming from lock-screen controls appeared to do nothing until the app was opened, (2) music stopped after extended background playback during track transitions. Fixed by: calling `audioEngine.tryResume()` synchronously inside the MediaSession `play` handler (within the user-activation window iOS grants to MediaSession callbacks); adding a silent looping audio keepalive (`silence-keepalive.ts`) that holds the OS audio session while user audio is paused and the app is backgrounded; loading the next track directly from the `ended` event handler to eliminate the inter-track silence gap that triggered session reclaim; and adding `visibilitychange` / `pageshow` foreground recovery to retry playback if the engine is paused when the app returns to the foreground.
+- **Discovery "Retry All" importing entire albums already in library**: The `POST /discover/retry-unavailable` endpoint fetched all raw `UnavailableAlbum` records for the week without applying the same three-level filter the `GET /current` endpoint uses before displaying them. As a result, clicking "Retry All" triggered full re-downloads of albums that were already present in the library (matched by discovery MBID, library MBID, or fuzzy title+artist). The retry handler now applies all three filters before creating download jobs, and deletes stale `UnavailableAlbum` records for albums already in the library so they do not reappear. Closes #34.
+- **Mood-tags phase silently skipping all tracks**: `lastfmTags` was `NULL` for tracks that had been enriched before the column was added. The mood-tags enrichment phase queries `WHERE lastfmTags != '{}'`, which never matches `NULL` — so every track was silently skipped every cycle. Migration backfills all `NULL` values to `'{}'` and sets the column default, so newly enriched tracks are never NULL.
+- **Docker image size (28.4 GB → 12.2 GB)**: Removed all CUDA and NVIDIA dependencies from the Docker image. The `audio-analyzer` and `audio-analyzer-clap` services now run on CPU-only PyTorch and TensorFlow. Changed pip installs to use the CPU-only PyTorch wheel index (`--index-url https://download.pytorch.org/whl/cpu`), replaced `tensorflow` with `tensorflow-cpu`, and installed `essentia-tensorflow --no-deps` to prevent pip from pulling the GPU TensorFlow variant as a transitive dependency. Removed `nvidia-cudnn-cu12`, `torchvision` (not imported), the `/opt/cudnn8` CUDA layer, and all NVIDIA library paths from the supervisor `LD_LIBRARY_PATH`. No regressions: TensorFlow confirmed running on CPU, all 9 MusiCNN classification heads load normally.
+- **Docker build context bloat**: `frontend/node_modules/` (598 MB) and `frontend/.next/` (313 MB) were not excluded from the Docker build context. The `.dockerignore` `node_modules` pattern only matched root-level; changed to `**/node_modules`. Added `**/.next`. Combined these reduced the `COPY frontend/ ./` layer from 946 MB to ~50 MB.
+- **Cover art fetch errors for temp-MBID albums**: Albums with temporary MBIDs (temp-*) were being passed to the Cover Art Archive API, causing 400 errors. Added validation to skip temp-MBIDs in artist enrichment and data cache.
+- **VIBE-VOCAB vocabulary file missing**: The vocabulary JSON file wasn't being copied to the Docker image because TypeScript doesn't copy .json files automatically. Added explicit import to force tsc to copy it.
+- **Redis memory overcommit warning**: Added `vm.overcommit_memory=1` sysctl to docker-compose.prod.yml and docker-compose.server.yml.
 - **Z-index stacking order**: MiniPlayer was z-50 (same tier as modals), causing it to appear above open dialogs due to DOM ordering. Established a consistent stacking hierarchy: MiniPlayer z-[45] → TopBar z-50 → VibeOverlay/toasts z-[55] → MobileSidebar backdrop z-[60] / drawer z-[70] → all modals z-[80] → nested confirm z-[85] → toast z-[100] → OverlayPlayer z-[9999]. MobileSidebar was also using non-standard `z-100` which is not a valid Tailwind class.
 - **API token display overflowing viewport on iPhone**: The newly-generated token `<code>` block extended beyond the screen on narrow viewports due to missing `min-w-0` / `overflow-hidden` on its flex container; added both.
+- **CLAP BullMQ worker crash on startup**: `import psycopg2` does not implicitly import `psycopg2.pool`; the BullMQ vibe worker was crashing immediately because `psycopg2.pool.ThreadedConnectionPool` was referenced without the submodule being imported. Added explicit `import psycopg2.pool`.
+- **EnrichmentStateService Redis disconnect error**: Calling `disconnect()` on an already-closed Redis connection raised an unhandled error. The disconnect is now silenced when the connection is already in a closed state.
+- **CLAP worker thread-safety**: All PostgreSQL calls in the CLAP BullMQ worker are now wrapped in `run_in_executor` so they execute on a thread-pool thread rather than blocking the asyncio event loop. Connection pool is initialized once per process and shared safely across concurrent jobs.
 
 ## [1.5.5] - 2026-02-21
 

Dockerfile

@@ -42,25 +42,38 @@ RUN mkdir -p /app/backend /app/frontend /app/audio-analyzer /app/models \
 # ============================================
 WORKDIR /app/audio-analyzer
 
-# Install Python dependencies for audio analysis
-# Note: TensorFlow must be installed explicitly for Python 3.11+ compatibility
+# Install all Python dependencies in a single layer to minimize image size
+# CPU-only torch/torchaudio: install first via the CPU index so downstream
+# packages (laion-clap, transformers) reuse the already-installed CPU wheels.
+# tensorflow-cpu replaces tensorflow to avoid pulling in CUDA runtime libs.
+# essentia-tensorflow declares a dependency on `tensorflow` (not tensorflow-cpu)
+# so we install it with --no-deps after tensorflow-cpu is already present.
 RUN pip3 install --no-cache-dir --break-system-packages \
-    'tensorflow>=2.13.0,<2.16.0' \
+    torch torchaudio torchvision \
+    --index-url https://download.pytorch.org/whl/cpu \
+    && pip3 install --no-cache-dir --break-system-packages \
+    'tensorflow-cpu>=2.13.0,<2.14.0' \
+    && pip3 install --no-cache-dir --break-system-packages --no-deps \
     essentia-tensorflow \
+    && pip3 install --no-cache-dir --break-system-packages \
     redis \
-    psycopg2-binary
-
-# Install cuDNN 8 for TensorFlow GPU (separate from PyTorch's cuDNN 9)
-# TF 2.15 needs cuDNN 8, PyTorch needs cuDNN 9 -- installed to isolated path to avoid conflicts
-RUN pip3 install --no-cache-dir --break-system-packages --target=/opt/cudnn8 'nvidia-cudnn-cu12==8.9.7.29'
+    psycopg2-binary \
+    'laion-clap>=1.1.4' \
+    'librosa>=0.10.0' \
+    'transformers>=4.30.0' \
+    'pgvector>=0.2.0' \
+    'python-dotenv>=1.0.0' \
+    'requests>=2.31.0' \
+    'bullmq==2.19.5' \
+    && pip cache purge \
+    && find /usr -name "*.pyc" -delete \
+    && find /usr -name "__pycache__" -type d -exec rm -rf {} + 2>/dev/null || true
 
-# Download Essentia ML models (~200MB total) - these enable Enhanced vibe matching
+# Download all ML models in a single layer (~800MB total)
 # IMPORTANT: Using MusiCNN models to match analyzer.py expectations
-RUN echo "Downloading Essentia ML models for Enhanced vibe matching..." && \
-    # Base MusiCNN embedding model (required for all predictions)
+RUN echo "Downloading ML models..." && \
     curl -L --retry 3 --retry-delay 5 --connect-timeout 30 --max-time 300 -o /app/models/msd-musicnn-1.pb \
         "https://essentia.upf.edu/models/autotagging/msd/msd-musicnn-1.pb" && \
-    # Mood classification heads (using MusiCNN architecture)
     curl -L --retry 3 --retry-delay 5 --connect-timeout 30 --max-time 300 -o /app/models/mood_happy-msd-musicnn-1.pb \
         "https://essentia.upf.edu/models/classification-heads/mood_happy/mood_happy-msd-musicnn-1.pb" && \
     curl -L --retry 3 --retry-delay 5 --connect-timeout 30 --max-time 300 -o /app/models/mood_sad-msd-musicnn-1.pb \
@@ -75,46 +88,26 @@ RUN echo "Downloading Essentia ML models for Enhanced vibe matching..." && \
         "https://essentia.upf.edu/models/classification-heads/mood_acoustic/mood_acoustic-msd-musicnn-1.pb" && \
     curl -L --retry 3 --retry-delay 5 --connect-timeout 30 --max-time 300 -o /app/models/mood_electronic-msd-musicnn-1.pb \
         "https://essentia.upf.edu/models/classification-heads/mood_electronic/mood_electronic-msd-musicnn-1.pb" && \
-    # Other classification heads
     curl -L --retry 3 --retry-delay 5 --connect-timeout 30 --max-time 300 -o /app/models/danceability-msd-musicnn-1.pb \
         "https://essentia.upf.edu/models/classification-heads/danceability/danceability-msd-musicnn-1.pb" && \
     curl -L --retry 3 --retry-delay 5 --connect-timeout 30 --max-time 300 -o /app/models/voice_instrumental-msd-musicnn-1.pb \
         "https://essentia.upf.edu/models/classification-heads/voice_instrumental/voice_instrumental-msd-musicnn-1.pb" && \
-    echo "ML models downloaded successfully" && \
+    curl -L --retry 3 --retry-delay 5 --connect-timeout 30 --max-time 300 -o /app/models/music_audioset_epoch_15_esc_90.14.pt \
+        "https://huggingface.co/lukewys/laion_clap/resolve/main/music_audioset_epoch_15_esc_90.14.pt" && \
+    echo "All ML models downloaded" && \
     ls -lh /app/models/
 
-# Copy audio analyzer script
+# Copy audio analyzer scripts
 COPY services/audio-analyzer/analyzer.py /app/audio-analyzer/
 
 # ============================================
 # CLAP ANALYZER SETUP (Vibe Similarity)
 # ============================================
 WORKDIR /app/audio-analyzer-clap
 
-# Install CLAP Python dependencies
-# Note: torch is large (~2GB) but required for CLAP embeddings
-RUN pip3 install --no-cache-dir --break-system-packages \
-    'laion-clap>=1.1.4' \
-    'torch>=2.0.0' \
-    'torchaudio>=2.0.0' \
-    'torchvision>=0.15.0' \
-    'librosa>=0.10.0' \
-    'transformers>=4.30.0' \
-    'pgvector>=0.2.0' \
-    'python-dotenv>=1.0.0' \
-    'requests>=2.31.0'
-
 # Copy CLAP analyzer script
 COPY services/audio-analyzer-clap/analyzer.py /app/audio-analyzer-clap/
 
-# Pre-download CLAP model (~600MB) during build to avoid runtime download
-# The analyzer expects the model at /app/models/music_audioset_epoch_15_esc_90.14.pt
-RUN echo "Downloading CLAP model for vibe similarity..." && \
-    curl -L --retry 3 --retry-delay 5 --connect-timeout 30 --max-time 300 -o /app/models/music_audioset_epoch_15_esc_90.14.pt \
-        "https://huggingface.co/lukewys/laion_clap/resolve/main/music_audioset_epoch_15_esc_90.14.pt" && \
-    echo "CLAP model downloaded successfully" && \
-    ls -lh /app/models/music_audioset_epoch_15_esc_90.14.pt
-
 # Create database readiness check script
 RUN cat > /app/wait-for-db.sh << 'EOF'
 #!/bin/bash
@@ -166,7 +159,9 @@ RUN npx prisma generate
 # Copy backend source and build
 COPY backend/src ./src
 COPY backend/tsconfig.json ./
-RUN npm run build
+RUN npm run build && \
+    npm prune --production && \
+    rm -rf src tests __tests__ tsconfig*.json
 
 COPY backend/docker-entrypoint.sh ./
 COPY backend/healthcheck.js ./healthcheck-backend.js
@@ -275,7 +270,7 @@ stdout_logfile=/dev/stdout
 stdout_logfile_maxbytes=0
 stderr_logfile=/dev/stderr
 stderr_logfile_maxbytes=0
-environment=DATABASE_URL="postgresql://kima:kima@localhost:5432/kima",REDIS_URL="redis://localhost:6379",MUSIC_PATH="/music",BATCH_SIZE="10",SLEEP_INTERVAL="5",MAX_ANALYZE_SECONDS="90",BRPOP_TIMEOUT="30",MODEL_IDLE_TIMEOUT="300",NUM_WORKERS="2",THREADS_PER_WORKER="1",LD_LIBRARY_PATH="/opt/cudnn8/nvidia/cudnn/lib:/usr/local/lib/python3.11/dist-packages/nvidia/cublas/lib:/usr/local/lib/python3.11/dist-packages/nvidia/cufft/lib:/usr/local/lib/python3.11/dist-packages/nvidia/cuda_runtime/lib:/usr/local/lib/python3.11/dist-packages/nvidia/cuda_nvrtc/lib:/usr/local/lib/python3.11/dist-packages/nvidia/cusolver/lib:/usr/local/lib/python3.11/dist-packages/nvidia/cusparse/lib:/usr/local/lib/python3.11/dist-packages/nvidia/nccl/lib"
+environment=DATABASE_URL="postgresql://kima:kima@localhost:5432/kima",REDIS_URL="redis://localhost:6379",MUSIC_PATH="/music",BATCH_SIZE="10",SLEEP_INTERVAL="5",MAX_ANALYZE_SECONDS="90",BRPOP_TIMEOUT="30",MODEL_IDLE_TIMEOUT="300",NUM_WORKERS="2",THREADS_PER_WORKER="1"
 priority=50
 
 [program:audio-analyzer-clap]