Tidal, Deezer, and Beatport discovery workers found metadata matches
but only stored them in memory and the discovery cache — never wrote
them back to the mirrored playlist tracks. This meant playlists had
to be re-discovered every time. Now writes {discovered, provider,
confidence, matched_data} to each mirrored track's extra_data after
discovery completes, matching the pattern YouTube and automation
pipeline discovery already used. Matches tracks by source_track_id
first, falls back to position index. Purely additive — discovery
logic is untouched.
The metadata-only optimization broke two things:
1. Cards showed 0 tracks because tracks were no longer in the listing
2. Auto-mirror skipped all playlists because tracks array was empty
Fix: cards render instantly from metadata, then tracks are fetched
per-playlist in the background via /api/tidal/playlist/<id>. As each
playlist's tracks arrive, the card count updates and the playlist
is auto-mirrored. Also tried multiple V2 attribute names for track
count (numberOfTracks, numberOfItems, etc.) and fixed the card DOM
selector for count updates.
get_user_playlists_metadata_only() was fetching full track lists for
every playlist sequentially (1+ API calls per playlist with 1s sleep
between pagination pages). For 20+ playlists this took 30-60 seconds.
Now returns only metadata (name, ID, track count, image) from a
single V2 API call. Track count comes from the numberOfTracks
attribute. Tracks are fetched on-demand when the user selects a
specific playlist to sync/mirror via the existing get_playlist()
endpoint.
Fresh installs now default to hybrid download mode (HiFi → YouTube →
Soulseek) instead of Soulseek-only, and Deezer as the metadata
fallback source instead of iTunes. Existing users with saved settings
are unaffected — defaults only apply when config keys don't exist.
The tooltip on failed/not_found tracks was offset by 3-6 items because
the backend cleanup step removed owned tracks from the wishlist between
when the frontend rendered the table and when the backend assigned
track indices. Surviving tracks got new enumeration indices (0,1,2...)
that didn't match their original table row positions (0,1,3,4...).
Fix: stamp each track with its position in the frontend's track_ids
array as _original_index, so the track_index always matches the modal
table row regardless of how many tracks were cleaned during processing.
The original #221 fix only covered Genius and AudioDB. All other
workers (Spotify, iTunes, Last.fm, MusicBrainz, Deezer, Tidal,
Qobuz) had the same bug: enrichment overwrites manual match status
to not_found when name search fails. Each worker now checks for an
existing service ID before searching by name and returns early if
one exists, preserving the manual match.
Worker threads were started before paused flag was set, allowing one
loop iteration (and API call) before pause took effect. Now sets
paused=True BEFORE start() so the thread sees it immediately. Fixes
the Spotify rate limit re-trigger on every container restart for users
with paused enrichment. Also increased max-retries rate limit ban
from 1 hour to 4 hours to prevent endless retry cycles.
When spotipy exhausts all retries on 429 errors, the actual
Retry-After value (often 10+ hours) is consumed internally by
spotipy and not passed in the exception. The default ban was
only 1 hour, causing an endless retry cycle. Increased to 4
hours to match the escalation max and give Spotify's ban time
to expire.
YouTube and ListenBrainz discovery toasts had source_label hardcoded
to 'iTunes' when not using Spotify. Now uses discovery_source.upper()
like the other discovery functions, so it correctly shows DEEZER when
Deezer is the active metadata source.
cover.jpg was always using the Spotify/iTunes thumbnail (640x640).
Now tries Cover Art Archive first (1200x1200+) when MusicBrainz
release ID is available. One line stores the MBID on album_info
during _enhance_file_metadata so _download_cover_art can use it.
Falls back to the source thumbnail if CAA has no art. Works for
all metadata sources since MusicBrainz enrichment is source-agnostic.
Status text and indicators now use fixed Material Design colors
instead of accent-dependent values — green for running/idle, amber
for paused, red for stopped, dim white for not configured. Readable
regardless of the user's chosen accent color.
normalize_string() was running unidecode on all text, converting
Japanese kanji to Chinese pinyin gibberish (命の灯火 → "tvanimedei").
Now detects CJK characters (kanji, hiragana, katakana, hangul,
fullwidth forms) and skips unidecode for text containing them —
just lowercases instead. Non-CJK text (Latin accents, Cyrillic)
still goes through unidecode normally.
"Believe" was falsely matching "Believe In Me" because SequenceMatcher
gives high scores when the search string is fully contained in the
match. Added a length ratio penalty: when cleaned titles differ in
length by more than 30%, the similarity score is multiplied by the
ratio (min/max length). This crushes prefix/suffix false positives
while leaving exact matches and cleaned variants (remastered, deluxe)
unaffected.
Playlist auto-sync was dropping tracks that failed iTunes/Apple Music
discovery — they never reached the wishlist or download pipeline. Now
undiscovered tracks continue through using available metadata: first
from the spotify_hint (embed scraper data with real Spotify track ID,
name, artists), then from raw playlist fields if a source track ID
exists. Album cover art from the mirrored playlist is included. Only
tracks with no usable ID or name are skipped.
Always-visible button in Spotify API section that clears the OAuth
token cache, pauses enrichment, and switches to the configured
fallback metadata source. Also fixed the dashboard service card to
show the actual active source name (Spotify/iTunes/Deezer) instead
of always showing "Spotify" with an amber fallback indicator.
New dedicated Explorer page with interactive node graph visualization.
Users select a mirrored playlist, choose Albums or Discographies mode,
and the app builds a branching tree: playlist root → artist nodes →
album nodes → track nodes. Supports all metadata sources (Spotify,
iTunes, Deezer) with source-aware discovery cache integration.
Features:
- Streaming NDJSON builds tree progressively as artist data arrives
- Circular artist nodes with photos, rounded album nodes with art
- SVG bezier connections that draw in on completion, fade on hover
- Click artist to expand albums, double-click album for track listing
- Single-click albums to select, Select All/Deselect for bulk ops
- Wishlist confirmation modal with per-album progress (NDJSON streaming)
- Artist nodes glow when any of their albums are selected
- Playlist picker with source tabs, discovery % gate (50% minimum)
- Zoom (scroll/pinch/buttons), pan (right/middle-drag), fit-to-view
- Metadata cache for discographies and album track listings
- Owned album detection from library database
- Fallback track-name matching when album names are missing
Plain (unsynced) lyrics were being saved with .lrc extension despite
having no timestamps, making them invalid for Plex and other players
that expect LRC format. Synced lyrics now write as .lrc, plain lyrics
write as .txt. Both types still get embedded in audio file tags.
Updated all file move/rename operations to handle .txt sidecars
alongside .lrc.
Enrichment chips now show live activity: 24h call count for all
services and daily budget usage (used/3,000) with gradient progress
bar for Spotify. Tracking is centralized in _get_enrichment_status
using cumulative stat diffs over a rolling deque — no worker files
modified. Added section header, "Configure →" label for unconfigured
services, and full 1h/24h breakdown in tooltips.
Single path template was missing _artists_list and _itunes_artist_id
context keys, so the collab mode first-artist extraction in
_apply_path_template had nothing to work with — $albumartist resolved
to the full multi-artist string. Added both keys matching the exact
pattern used by album and playlist modes, including the iTunes
spotify_album.external_urls fallback. Updated settings UI hints to
show $albumartist as available for single and playlist templates.
When a user manually matched an artist to a service ID then triggered
enrichment, the worker re-searched by name, failed to find a match,
and overwrote the status back to not_found — despite the ID being
valid. Now both Genius and AudioDB workers check for existing service
IDs before searching by name. If an ID exists (from manual match),
the worker uses it for a direct API lookup to enrich metadata while
preserving the matched status. Added AudioDB lookup-by-ID client
methods for artist, album, and track.
Top-level try/except in do_GET ensures an HTTP response is always sent
— previously, unhandled exceptions caused BaseHTTPRequestHandler to
silently close the connection (ERR_EMPTY_RESPONSE). All callback
logging now uses the app logger instead of print() so output appears
in app.log rather than only Docker stdout. Added health check at / to
verify the callback server is running, and startup now logs the actual
bind address to help diagnose port conflicts.
Dashboard now displays all enrichment services as live-status chips
below the core service cards. Each chip shows Running, Idle, Paused,
Stopped, or Not Configured state with color-coded left border accents.
Unconfigured services appear dimmed with dashed borders — clicking any
configurable chip navigates to Settings → Connections and scrolls to
the relevant service section.
Also fixes the Spotify card always being labeled "Apple Music" when
using iTunes fallback — card now always says "Spotify" with an amber
"Using iTunes/Deezer" indicator when fallback is active.
Qobuz login was only available on the Downloads tab when Qobuz was
selected as download source. But Qobuz credentials are also needed
for the enrichment worker which runs independently. Users saw
"Connect Qobuz in Settings" on the dashboard but couldn't find it.
Adds a Qobuz section to Settings → Connections (same pattern as
Tidal's existing Connections section). checkQobuzAuthStatus() now
syncs both the Connections and Downloads tab sections. Login from
either tab updates both. No backend changes — same API endpoints.
The tooltip only checked paused/authenticated/idle/running states.
When Spotify was rate limited or daily budget exhausted, the worker
thread was still alive (sleeping in guards) so it showed "Running"
with no current item and stale 0% progress.
Now checks rate_limited and daily_budget.exhausted before running:
- Rate limited: "Rate Limited — Waiting Xm for rate limit to clear"
- Budget exhausted: "Daily Limit Reached — Resets in Xh Xm"
- No current item: "Waiting for next item..." instead of blank
Also adds rate_limit info object to get_stats() response for the
countdown display.
Cache maintenance:
- Input validation rejects junk entities (Unknown Artist, empty names)
from being cached, with exemptions for synthetic entries (_features,
_tracks suffixes)
- CacheEvictorJob expanded to 4 phases: TTL eviction, junk cleanup,
orphaned search cleanup, MusicBrainz failed lookup cleanup
- MusicBrainz null results now expire after 30 days (was 90) so failed
lookups get retried sooner
Cache health UI:
- Polished modal accessible from Dashboard "Cache Health" button and
repair dashboard health bar
- Shows health status banner (healthy/fair/poor), stat cards, source
breakdown with colored progress bars, type pills, and metrics table
- Repair dashboard shows compact bar with health dot indicator
downloadSelectedCategory() was passing only the category name to the
download function, which fetched ALL tracks in that category. Now
collects checked track IDs from checkboxes BEFORE closing the modal
(DOM is destroyed on close), then filters the fetched tracks to only
the selected ones.
If nothing is checked, downloads the full category (same as before).
Other callers of openDownloadMissingWishlistModal are unaffected —
the new selectedTrackIds parameter defaults to null.
The auth_tidal() endpoint was overriding the user's configured
redirect_uri with one built from request.host. In Docker, request.host
is the container hostname (e.g. "soulsync-webui"), not the external
URL the user configured in settings.
Now checks config_manager for the user's configured redirect_uri first.
Only falls back to request.host dynamic detection if no redirect URI
is configured.
Album art embedding now tries Cover Art Archive first (using the
MusicBrainz release ID from source ID embedding) before falling back
to Spotify/iTunes/Deezer URLs. CAA provides original-quality artwork,
often 1200x1200 or higher vs Spotify's 640x640 max.
Reordered _embed_source_ids to run before _embed_album_art_metadata
so the MusicBrainz release ID is available for the CAA lookup. Also
fixed hardcoded 640x640 FLAC picture dimensions — now detects actual
size from image bytes. Falls back to existing behavior if CAA fails
or no release ID exists.
When the fingerprint score is >=0.95 but title/artist don't match
(e.g. English expected vs Japanese returned), SKIP instead of FAIL.
A 95%+ fingerprint means the audio IS the correct recording — the
metadata mismatch is just a language/script difference, not a wrong
file. Prevents Japanese, Chinese, Korean, and other non-Latin tracks
from being falsely quarantined.
Happy path unchanged — matching title/artist still returns PASS at
the earlier check before this code is reached.
Previously, tags.clear() only cleared in memory — if any later step
threw (metadata extraction, API calls, album art download), the file
was moved with its original Soulseek source tags intact. This caused
album fragmentation in media servers when some tracks had MusicBrainz
IDs and others didn't.
Now the cleared tags are saved to disk immediately after wiping. If
enhancement succeeds, the file is saved again with full metadata
(identical to before). If it fails, the file has clean empty tags
instead of inconsistent junk — media servers group by folder structure
which is always correct.
When artist or title contains non-ASCII characters (Japanese, Chinese,
Korean, etc.), prepend the original un-romanized text as the first
search query. unidecode converts Japanese kanji to Chinese pinyin
(e.g. "藤澤慶昌" → "wu zhi zhuan sheng") which never matches on
Soulseek. The original characters match filenames directly.
Romanized fallback queries are still generated after for coverage.
Zero impact on ASCII-only tracks (isascii check skips them).
Replaces the fire-and-forget button with a premium modal that shows
exactly which artists will be added before confirming. Features:
- Glassmorphic modal with stat cards, two-column artist grid, search
filter, collapsible ineligible section, and loading spinner
- Source-aware filtering: only shows artists with the active source's
ID (Spotify/iTunes/Deezer) as eligible
- Frontend and backend both paginate at 400 to avoid SQLite variable
limit (SQLITE_MAX_VARIABLE_NUMBER=999) that silently broke queries
above ~500 artists
- Backend source detection aligned with frontend — uses only the
active source's ID, falls back to configured metadata source
The bulk watchlist add had no Deezer ID support — artists with only a
deezer_id were silently skipped (artist_id stayed None). Also fixed the
source detection to use the actual ID field picked instead of a numeric
heuristic that could assign Deezer IDs to the wrong service column.
Fallback chain is now: active source first, then Spotify → iTunes → Deezer.
fix_finding() was using a potentially stale transfer_folder that was
only refreshed during scheduled job runs, not on manual fix attempts.
Now re-reads the transfer path from config before each fix, matching
the same logic used by _run_next_job().
Also surfaces fix failure reasons to the user — bulk fix now logs each
failure with finding ID and error, and the frontend toast shows the
actual error message instead of just "X failed".
Artist/album/track matching previously took the first Spotify search
result above the 0.80 similarity threshold. If Spotify returned a
near-match before the exact match (e.g. "Brother's Keeper" before
"Brothers Keepers"), the wrong entity would be selected.
Now scores all candidates and picks the highest, so an exact match
(1.0) always wins over a near-match (0.94). No change to threshold
or batch matching logic — strictly better or equal results.
When Spotify falls back to iTunes/Deezer (auth failure, rate limit, API
error), the manual match modals were storing numeric fallback IDs in
spotify_*_id columns, breaking Spotify links and metadata fetching.
Fix detects the actual provider by inspecting result IDs (alphanumeric =
Spotify, numeric = fallback) rather than checking auth status, which can
be misleading during rate limit bans. The frontend now passes the real
provider to the storage endpoint so IDs land in the correct column.
The background enrichment worker now caps itself at 3,000 processed items
per calendar day. Counter resets at midnight automatically. When exhausted,
the worker sleeps and checks every 5 minutes for a new day.
This is scoped entirely to the enrichment worker — user-initiated Spotify
API calls (searches, playlist ops, album lookups, etc.) are completely
unaffected. Budget status is exposed in the worker's get_stats() response
for the dashboard widget.
The image update code only checked Spotify's images array format,
missing iTunes direct image_url. Also used spotify_artist_id for
DB lookup which missed Deezer-only artists. Now handles both image
formats and tries all provider IDs for the DB update.
Old Deezer watchlist artists missing images will get backfilled
on the next watchlist scan automatically.
OggOpus (Mutagen) is a separate class from OggVorbis but uses the
same VorbisComment tag API. All isinstance checks for (FLAC, OggVorbis)
missed OggOpus, so tags were cleared but never rewritten — resulting
in empty metadata and "Unknown Artist" on the media server.
Added _is_ogg_opus() helper and applied to all 15 format checks in
web_server.py and 2 in core/tag_writer.py. No change to FLAC/OGG/
MP3/M4A handling (short-circuit evaluation skips the new check).
The fix was failing silently — returning error results without
writing to any log. Added info/warning logs at each failure point
(missing paths, source not found, path escapes transfer, destination
conflict) so Docker path resolution issues can be diagnosed.
New "Download Discography" button in artist hero section opens a modal
showing the full catalog — albums, EPs, and singles — with filter
toggles, select/deselect all, and per-album owned/missing indicators.
Modal features:
- Glassmorphic design with artist image blurred background header
- Filter pills for Albums/EPs/Singles with instant grid filtering
- Album cards with cover art, year, track count, and checkbox
- Owned albums dimmed and unchecked by default, missing pre-selected
- Live NDJSON streaming: each album updates in real-time as processed
- "Process Wishlist Now" button after completion
- Albums sorted by track count (Deluxe first) to prevent duplicate
folder contexts from standard/deluxe edition ordering
Backend: NDJSON streaming endpoint POST /api/artist/<id>/download-discography
- Fetches tracks per album via active metadata client
- Adds to wishlist with dedup (no slow fuzzy matching)
- Streams one JSON line per album as it completes
- Works on both Artists search page and Library artist detail page
Added 7 new sections to version modal: Stream Source, YouTube Fix,
Completion Badges, Collab Album Handling, Per-Artist Sync, Stability
fixes. Updated helper What's New with 5 new entries.
New "Sync" button in the enhanced view header validates an artist's
library entries against files on disk. Removes stale tracks (missing
files), cleans empty albums, and updates track counts.
- POST /api/library/artist/<id>/sync endpoint
- Checks each track's file_path via _resolve_library_file_path
- Empty album cleanup checks ALL tracks (not just this artist's)
to avoid deleting albums shared with other artists
- Toast shows results: stale removed, albums cleaned, or "All files
verified" if everything checks out
- Auto-refreshes enhanced view when changes are made
The DELETE /api/library/album/<id> route used <int:album_id> which
rejected non-integer IDs (e.g., Navidrome string IDs). Flask returned
its default 404 HTML page, causing "unexpected token <" JSON parse
error on the frontend. Changed to <album_id> to match all other
album routes which already accept any ID type.
Apple Music returns censored titles like "B*****t Faucet" for
"Bullshit Faucet". The string similarity function now detects
asterisk patterns and matches by comparing non-censored words
exactly and censored words by prefix/suffix characters.
Only fires when * is present in one string — zero impact on
normal comparisons. Prevents daily re-downloads of explicit
tracks that exist in the library under their uncensored names.
New setting in Settings → Library → File Organization: "Collaborative
Album Artist" — choose between first listed artist (default) or all
artists combined for $albumartist in folder paths and album_artist tag.
Per-source resolution:
- Spotify: artists array has separate objects — picks first directly
- Deezer: API already returns first artist only — no change needed
- iTunes: combined string ("Larry June, Curren$y & The Alchemist") —
resolves via artistId API lookup to get primary name ("Larry June").
Safe for "Tyler, the Creator" and "Simon & Garfunkel" because their
IDs resolve to the same combined name (no change).
Applied to both folder path ($albumartist template) and album_artist
metadata tag for consistency. Track artist tag always keeps all artists.
iTunes lookup only fires when source is iTunes (numeric ID + not Deezer).
Allows external apps to trigger wishlist processing without needing
to look up automation IDs. Returns 409 if already running. Uses the
same _process_wishlist_automatically function as the automation engine.
