Replaces per-member Nextcloud credentials with a BandStorage model that
supports multiple providers. Credentials are Fernet-encrypted at rest;
worker receives audio via an internal streaming endpoint instead of
direct storage access.
- Add BandStorage DB model with partial unique index (one active per band)
- Add migrations 0007 (create band_storage) and 0008 (drop old nc columns)
- Add StorageFactory that builds the correct StorageClient from BandStorage
- Add storage router: connect/nextcloud, OAuth2 authorize/callback, list, disconnect
- Add Fernet encryption helpers in security/encryption.py
- Rewrite watcher for per-band polling via internal API config endpoint
- Update worker to stream audio from API instead of accessing storage directly
- Update frontend: new storage API in bands.ts, rewritten StorageSection,
simplified band creation modal (no storage step)
- Add STORAGE_ENCRYPTION_KEY to all docker-compose files
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two bugs fixed:
1. handle_transcode was writing cdn_hls_base = "hls/{version_id}" to the DB
even though HLS files were only in a temp dir (never uploaded to Nextcloud).
The stream endpoint then tried to serve this non-existent path, returning 404
and breaking audio playback for every transcoded version. Removed the
cdn_hls_base write — stream endpoint falls back to nc_file_path (raw file),
which works correctly.
2. Added extract_peaks worker job type: lightweight job that downloads audio
and computes waveform_peaks + waveform_peaks_mini only. No transcode, no HLS,
no full analysis.
3. Added POST /internal/reindex-peaks endpoint (protected by internal secret):
finds all audio_versions with null waveform_peaks and enqueues extract_peaks
jobs. Safe to call multiple times. Use after a fresh DB scan or peak algorithm
changes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Store waveform peaks inline in audio_versions (JSONB columns) so WaveSurfer
can render the waveform immediately on page load without waiting for audio
decode. Adds a 100-point mini-waveform for version selector thumbnails.
Backend:
- Migration 0006: adds waveform_peaks and waveform_peaks_mini JSONB columns
- Worker generates both resolutions (500-pt full, 100-pt mini) during transcode
and stores them directly in DB — replaces file-based waveform_url approach
- AudioVersionRead schema exposes both fields inline (no extra HTTP round-trip)
- GET /versions/{id}/waveform reads from DB; adds ?resolution=mini support
Frontend:
- audioService.initialize() accepts peaks and calls ws.load(url, Float32Array)
so waveform renders instantly without audio decode
- useWaveform hook threads peaks option through to audioService
- PlayerPanel passes waveform_peaks from the active version to the hook
- New MiniWaveform SVG component (no WaveSurfer) renders mini peaks in the
version selector buttons
Fix: docker-compose.dev.yml now runs alembic upgrade head before starting
the API server, so a fresh volume gets the full schema automatically.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- New split layout: waveform/transport/queue left, comment panel right
- Avatar pins above waveform positioned by timestamp with hover tooltips
- Transport bar: speed selector, ±30s skip, 46px amber play/pause, volume
- Comment compose: live timestamp pill, suggestion/issue/keeper tag buttons
- Comment list: per-author colour avatars, amber timestamp seek chips,
playhead-proximity highlight, delete only shown on own comments
- Queue panel showing other songs in the same session
- Waveform colours updated to amber/dim palette (104px height)
- Add GET /songs/{song_id} endpoint for song metadata
- Add tag field to SongComment (model, schema, router, migration 0005)
- Fix migration 0005 down_revision to use short ID "0004"
- Fix ESLint no-unused-expressions in keyboard shortcut handler
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Race condition (worker "Job not found in DB"):
- RedisJobQueue.enqueue() was pushing job IDs to Redis immediately after
flush() but before the API transaction committed, so the worker would
read an ID that didn't exist yet in the DB from its own session.
- Fix: defer the Redis rpush until after session.commit() via a pending-
push list drained by get_session() after each successful commit.
- Worker: drain stale Redis queue entries on startup to clear any IDs
left over from previously uncommitted transactions.
- Worker: add 3-attempt retry with 200ms sleep when a job is not found,
as a safety net for any remaining propagation edge cases.
NC scan folder structure (YYMMDD rehearsal subfolders):
- Previously used dir_name as song title for all files in a subdirectory,
meaning every file got the folder name (e.g. "231015") as its title.
- Fix: derive song title from Path(sub_rel).stem so each audio file gets
its own name; use the file's parent path as nc_folder for version grouping.
- Rehearsal folder name stored in song.notes as "Rehearsal: YYMMDD".
- Added structured logging throughout the scan: entries found, per-folder
file counts, skip/create/import decisions, and final summary count.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
