Sokhi (again): downloading the base 'Mushoku Tensei S2 Original Soundtrack' embedded
the cour-2 '…サウンドトラック2' cover. numeric_tokens_differ stripped titles to
[a-z0-9], turning CJK into spaces — so the trailing '2' collapsed to a bare '2'
that '第2期' (season 2) already supplied on BOTH sides, leaving the digit sets equal
and the guard blind. Tokenise on \W (Unicode word-aware) instead, so a digit stays
attached to its word ('サウンドトラック2' is its own digit-bearing token). Latin
behaviour is byte-identical (Vol.4 vs Vol.4.5 etc.). Shared guard, so the art picker
AND the MusicBrainz->CAA path are both fixed. Regression tests added.
Single tracks (esp. Deezer-sourced) imported as "01 - Title" regardless
of their real album position — e.g. Fly Away (track 2 of Greatest Hits)
landed as 01, littering album folders with duplicate "01" files.
Root cause: a Deezer single track is matched via /search/track, which
omits track_position, so the context never carried the real number; then
service.py + context.py fabricated a confident track_number=1 from that
gap. Because the resolver puts that first, the fake 1 beat the source.
It is source-agnostic (slskd-with-Deezer-metadata hits it too) — albums
work because /album/<id>/tracks DOES include positions.
Fix (at the shared import funnel, strictly additive):
- track_number.py: new read_embedded_track_number() (mutagen, local, no
network) + an optional embedded_track_number arg on resolve_track_number.
The downloaded file already carries the source-written position (deemix
wrote it); consult it LAST — only when metadata AND the "NN - Title"
filename both come up empty — so it can only fill the gap that would
otherwise hit the default-1 floor. Never overrides a value the pre-fix
resolver produced (no regression for correctly-named/mistagged files).
- pipeline.py: read the file tag at the resolve step and pass it in.
- De-poison: service.py:217 + context.py default to 0 (the existing
"unknown" sentinel, like total_tracks), NOT 1 — so the fake 1 no longer
blocks recovery. Frontend already treats falsy track_number as unknown
(omits it), so this also drops the bogus "1." in the UI.
13 new resolver tests incl. the no-regression precedence guards; full
imports + wishlist suites green (583), no behavior change for albums.
A user who removes a wishlist track, or cancels an in-flight wishlist
download, would have it re-added on the next auto cycle (watchlist scan,
failed-track capture, or the cancel handler's own re-add), so the same
release downloaded -> failed/cancelled -> re-queued forever.
Adds a TTL'd skip-gate (30 days), softer than the blocklist: it expires
so the track is reconsidered later, and never blocks a manual
force-download — only the automatic re-queue.
- core/wishlist/ignore.py: pure TTL/normalization/display logic + a
best-effort orchestrator (no DB handle, caller passes now).
- database/music_database.py: migration-safe wishlist_ignore table +
add/check/remove/list(+purge)/clear methods, and the gate in
add_to_wishlist beside the blocklist guard. Fail-open throughout — an
ignore error can never block a legitimate add; a manual add bypasses
the gate AND clears the ignore.
- routes.py: user remove (single/album/batch) records an ignore. Hooked
at the route layer, NOT the DB remove, so success-cleanup never
ignores (regression-tested).
- web_server.py: cancel now ignores + removes from the wishlist instead
of re-adding for endless retry; three /api/wishlist/ignore-list*
endpoints.
- downloads.js: 'Ignored' modal (view / un-ignore / clear all).
- 13 tests: pure logic, DB seam, gate (block/bypass/fail-open),
route wiring, and the success-cleanup-does-not-ignore regression.
Multiple failed source attempts at one song each land in quarantine as
separate entries. Group them by the *intended* target (sidecar context
track_info isrc -> id -> uri, falling back to normalized artist|title for
legacy thin sidecars) — an exact relationship across siblings, since the
bad files' own tags differ but the target track is constant.
- core: quarantine_group_key() + find_quarantine_siblings() seams; list
entries now carry group_key.
- approve endpoint: remove_siblings flag auto-deletes the other attempts
once one is accepted (captured BEFORE approve restores the file out of
quarantine, or the id lookup would resolve nothing). Scoped to the
quarantine manager; download-modal chooser + version-mismatch fallback
pass no flag and are unaffected.
- UI: multi-member groups render as a collapsible parent row (album art +
'N alternatives'); singletons unchanged. Toast reports removed count.
- 11 tests incl. ordering regression for capture-before-approve.
The Favorites collection walker (_iter_collection_resource_ids) broke on ANY
non-200 — including a transient 429. So a rate-limit mid-pagination silently
truncated the collection: the log shows `status=429` then `Retrieved 98/100`,
and the mirror saved 98 of a 524-track favorites list. The auto-sync cycle only
"worked" because it dodged the 429. The regular-playlist paginator already
retries 429; the collection walker didn't.
Fix: retry the same cursor page with backoff (5/10/15/20s, 4 attempts) on 429,
mirroring the playlist paginator; 401/403 still bail (+ reconnect flag), other
non-200 still break. Regression tests: 429 mid-walk completes the full chain;
exhausted retries return partial without hanging; 429 doesn't set reconnect.
Adds a dedicated `get_library_history_unverified()` DB query that fetches
every library_history row with verification_status IN ('unverified',
'force_imported') with no recency cap. This is loaded unconditionally in
`build_unified_downloads_response` — not gated on `len(items) < limit` —
so historical unverified entries are never buried by a busy batch filling
the 200-row general limit, and entries from weeks/months ago aren't lost
in the 50-row recency-ordered history tail. Adds idx_lh_verification_status
for query performance and two regression tests.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The scan was only doing the header-based quality gate (mutagen) — fast but
shallow. The download/import pipeline ALSO runs detect_broken_audio first,
which uses ffmpeg to actually DECODE the file (astats truncation check +
silencedetect) to verify the REAL audio, not just the metadata. That's the
whole point of unifying onto the download quality pipeline.
- Each file now runs both stages: (1) ffmpeg AudioGuard (detect_broken_audio),
(2) header quality gate (probe_audio_quality + quality_meets_profile).
A finding is created for broken/incomplete audio OR below-profile quality,
with quality_issue + broken_audio_reason in details and a 'warning' severity
for broken audio vs 'info' for below-profile.
- New setting deep_audio_verify (default True) toggles the ffmpeg decode pass;
off = fast header-only. Slower full scan is expected — it decodes every file.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
library_tracks_only defaulted ON, which skipped every file when the DB (reset
by the user) no longer matched the files on disk → scanned=0, nothing tested.
Default it OFF: check every audio file in the Music Library output folder, which
is what users expect. DB matching is still used opportunistically for better
finding metadata, just no longer required. Power users can re-enable the filter.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The scan was walking soulseek.download_path (/app/downloads) too, which is the
raw download/staging area full of pre-import leftovers — not the library. Walk
only the "Output Folder (Music Library)" (soulseek.transfer_path) plus any
custom library.music_paths. A user's custom output-folder path is respected
since it's read live from config.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The folder walk found 403 files in transfer/downloads when the user's library
is only ~18 tracks — the rest are pre-import leftovers (residue after a DB
reset). Those are orphans, not library tracks, and belong to the Orphan File
Detector, not a quality scan.
- New setting library_tracks_only (default True): match each walked file to a
DB track via the suffix index BEFORE probing; skip anything with no DB row.
So the scan reflects the real library, not download junk, and avoids probing
hundreds of orphan files.
- Split _lookup_meta into _match_db (cheap DB suffix match) + _read_file_tags
(only used when library_tracks_only is off, for loose files).
- Log how many files were skipped as not-in-library.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The quality job kept resolving 0/N because DB-path resolution failed in the
deployed environment for reasons the logs wouldn't surface. Switch to the
mechanism the WORKING file tools use: os.walk the real music folders
(transfer + download + configured library paths, abspath'd) exactly like
orphan_file_detector and fake_lossless_detector — those reliably see files
because they never touch the DB's stored relative paths.
- Walk all existing music dirs, collect audio files (dedup by realpath),
probe each with the same probe_audio_quality the import guard uses, check
quality_meets_profile (strict). Below-profile files become findings.
- Match each walked file back to its DB track via a path-suffix index (last
1-3 components) for real title/artist/album + track id; fall back to the
file's own tags when no DB row matches (finding filed as 'file').
- Loud diagnostics: logs the folders walked and the audio-file count, and
warns clearly when no music folder exists to walk.
The fix handler already works with the now-absolute file_path and an optional
entity_id (deletes the file by real path; DB row only when known).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Answers "does import respect quality?": yes — the pipeline already runs the
quality gate (check_quality_target) BEFORE AcoustID and quarantines files that
don't meet the profile (unless fallback/downsample is on). This adds an explicit
user switch over that behaviour.
- New config import.quality_filter_enabled (default True). When False,
check_quality_target returns None early so EVERY file imports regardless of
quality; the file is still probed and the library Quality Upgrade Scanner
still flags below-profile tracks. Default preserves current behaviour.
- Settings → Library: the Import Settings group is now a collapsible tile
(same pattern as Post-Processing) and gains the "Only import tracks that meet
your quality profile" toggle at the top, alongside replace-lower-quality and
folder-artist-override.
- settings.js populate/collect the new key; config schema default added.
- Tests: key-aware config stub (a blanket-False mock would wrongly disable the
filter) + a new test pinning toggle-OFF = accept below-target file.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The quality job still resolved 0/18 because the shared resolver kept relative
config paths ("./Transfer") as-is and gated them behind os.path.isdir("./Transfer"),
which only holds when the calling thread's CWD is the app root. The repair
worker thread's CWD isn't guaranteed to be /app, so base_dirs came back empty
and every track was "unresolved".
- _collect_base_dirs now also adds os.path.abspath() of every relative
candidate, so "./Transfer" → "/app/Transfer" regardless of CWD.
- quality_upgrade_scanner logs a one-shot [QualityResolve] diagnostic on the
first unresolved track (cwd, transfer_folder + abspath + isdir, base dirs
tried, abs-join existence) so any remaining mount mismatch is pinpointable
instead of a silent "all skipped".
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaces the wishlist-only Quality Scanner with a proper Library Maintenance
job that produces actionable findings — same model as the AcoustID/orphan
tools, per user request ("mach ein finding wie jedes anderes Tool").
- New core/repair_jobs/quality_upgrade_scanner.py: iterates DB library
tracks, resolves each path via the shared resolver (now index-0 correct for
relative library paths), probes REAL audio quality with the same
probe_audio_quality the download import guard uses, and checks it against the
user's v3 ranked targets via quality_meets_profile (strict — no extension
guessing, no fallback). Below-profile tracks become 'quality_upgrade'
findings with current vs target quality in details.
- repair_worker._fix_quality_upgrade: redownload (wishlist + delete file/row),
delete (file + row), or ignore (dismiss in UI). Registered in _execute_fix
dispatch + bulk fixable_types.
- Frontend (enrichment.js): 'Low Quality' type label, 'Upgrade' fix button, a
3-way _promptQualityUpgradeAction modal (Re-download / Delete / Ignore),
wired into both single-finding fix and bulk-fix (Ignore → dismiss inline).
- Tools "Quality Scanner" button now triggers Run Now of this job and points
the user to Library Maintenance → Findings.
The old standalone /api/quality-scanner endpoints are left intact (unused by the
button) to avoid churn. Verified: job registers, fix handler dispatches.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ROOT CAUSE of the quality scanner's "18/18 could not be probed". The shared
resolver (core/library/path_resolver.py) suffix-walked starting at index 1,
which is correct for absolute media-server paths (/music/Artist/... — index 0
is the empty leading segment) but WRONG for SoulSync's own library, which
stores RELATIVE paths like "Asketa/Another Side/track.flac". Index 0 there is
the artist folder; dropping it meant the resolver joined base/Another Side/...
(no artist) and nothing ever matched — so every library track came back
unresolved and the probe opened a relative path that didn't exist from CWD.
Start the suffix walk at index 0 so the FULL relative path is tried first.
Safe for absolute paths (i=0 yields base//Artist/... which harmlessly misses
and falls through to i=1) and Windows drive parts (E: fails on POSIX, falls
through). Other tools (orphan/fake-lossless detectors) were unaffected because
they os.walk the transfer folder directly and never used this resolver.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- list_quarantine_entries now surfaces the probed quality (context._audio_quality,
recorded before the gates) so each quarantine row shows what the file actually
is when deciding to approve/delete. Rendered as a quality chip in the review UI.
- _resolve_library_file_path logs the searched base dirs once when it can't
resolve a path, so a remaining mount/path mismatch is diagnosable.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1) _resolve_library_file_path now tries the FULL relative path (index 0)
against each base dir, not just suffixes. The library scanner stores clean
"Artist/Album/Track.flac" paths; skipping index 0 dropped the artist folder
so the file never resolved — every quality-scanner probe failed ("20/20
could not be probed"). Now they resolve under the transfer/library dir.
2) Quality gate moved BEFORE the AcoustID check in post_process_matched_download.
- A wrong-quality file is rejected without paying for an AcoustID fingerprint.
- context['_audio_quality'] is set before either gate quarantines, so the
real quality is recorded on the sidecar for EVERY quarantine trigger —
it's known when reviewing/approving any quarantined file.
- force_import still never fires on a quality mismatch (only AcoustID).
normalize_import_context mutates in place, so the moved block keeps its
context fields. New test pins the order + that AcoustID isn't run on a
quality reject.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diagnostics revealed the real cause: the tracks table stores file_path
RELATIVE to the library root (e.g. "Asketa/Another Side/01-01 - Another
Side.flac"), so probing the raw path failed for the entire library — every
track came back unprobeable and was left unflagged ("20/20 could not be
probed").
The scanner now resolves each path via _resolve_library_file_path (checks
transfer/download/library dirs, same helper the rest of the app uses) before
probing, falling back to docker_resolve_path. Injected via deps for testability.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A task stuck in 'post_processing' past the cutoff was force-marked 'completed'
("assume it worked"). In a large batch, post-processing (AcoustID + quality +
import) is serialized and backs up, so tasks sit in post_processing while merely
QUEUED — then got falsely completed, showing as downloaded with no file on disk
(/Transfer empty).
Now: the cutoff is 30 min (was 5) so legit backlog isn't cut off, and when it
does fire the task is only completed if it actually produced a file
(final_file_path exists on disk) — otherwise marked failed (honest + retryable).
Applied at both stuck-detection sites (check_batch_completion + _v2).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The worker used logging.getLogger(__name__) → "core.discovery.quality_scanner",
which the app log view (soulsync.*) doesn't surface — so the scan looked like it
did nothing ("API Starting scan" straight to "quality_scan_completed" with no
worker output). Switched to get_logger("discovery.quality_scanner") so "Found N
tracks", "Profile targets", and the unprobeable-file diagnostics show up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The unified scanner must READ each file to judge real bit depth/sample rate
(extension alone can't tell 16-bit from 24-bit FLAC). If the stored library
path doesn't resolve to a readable file in this container, every probe returns
None and — since an unprobeable file can't be judged — the whole library passes
silently ("scans nothing").
Now: resolve the path via docker_resolve_path before probing, and count +
log unprobeable files (first 5 paths at WARNING, plus an end-of-scan summary
"N/M tracks could not be probed"). This makes a systematic path/mount mismatch
visible instead of an empty result.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The library quality scanner judged quality by FILE EXTENSION only
(get_quality_tier_from_extension) and read the legacy v2 `qualities` dict —
so every FLAC was "lossless tier 1" regardless of bit depth / sample rate. It
could never flag a 16-bit FLAC as upgradeable under a 24-bit profile, and it
ignored the v3 ranked_targets entirely. Completely inconsistent with the
download guard.
Now both share one core:
- selection.targets_from_profile(profile) — single profile→targets conversion
(v2→v3 migration), reused by load_profile_targets.
- selection.quality_meets_profile(aq, targets) — strict: meets iff the real
measured quality satisfies a ranked target (fallback ignored — it's a
download concession, not a definition of "good enough").
- guards.check_quality_target refactored to use both.
- quality_scanner probes real quality (probe_audio_quality) and checks against
the v3 targets via quality_meets_profile. Extension tier kept only as a
fallback label when a file can't be probed.
Result: the scan flags exactly what the download gate would reject — 16-bit
when you want 24-bit, wrong sample rate, MP3 when you want FLAC.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A file quarantined for QUALITY (e.g. MP3-VBR rejected by a FLAC-only profile)
showed up in BOTH the Completed and Quarantine tabs. Cause: the verification
wrapper (post_process_matched_download_with_verification) handled the
_acoustid_quarantined / _integrity_failure_msg / _race_guard_failed markers but
NOT the quality marker _bitdepth_rejected (nor _silence_rejected). A quality
quarantine leaves no _final_processed_path, so the wrapper hit the
"no final path — assuming success" branch and marked the task Completed.
Unlike acoustid/integrity (retry driven by the wrapper), the inner pipeline
already owns the quality/audio-guard outcome — it quarantines then re-queues the
next-best candidate or marks the task failed. So the wrapper now just returns
when it sees _bitdepth_rejected/_silence_rejected, without marking completed
(which clobbered both the quarantine state AND any successful retry).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause of "completed/failed/unverified don't show during a running
batch, only after it ends" (F5 didn't help): build_unified_downloads_response
sorted live tasks active-first (downloading/searching/queued = priority 0-3,
completed/failed = 4-7) then truncated the whole array at items[:limit] (300).
During a busy batch the active+queued tasks filled the limit and pushed every
terminal task off the end, so /api/downloads/all never returned them — the
Completed/Failed/Unverified tabs filter client-side and had nothing to show.
Fix: `limit` now bounds only the persistent-history tail. Live in-memory
tasks are always returned in full — they're already bounded by the 5-min
cleanup automation, and array order is presentation-only since the page
filters per tab client-side.
Verified with a repro (320 queued + 1 completed + 1 failed → terminal rows
were absent at limit=300; now present).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three fixes from on-device testing of best-quality mode:
1. clear_completed_local no longer prunes terminal tasks that belong to a
STILL-ACTIVE batch (one with non-terminal work remaining). The 5-min
"Clean Completed Downloads" automation was yanking completed/failed/
unverified rows out of download_tasks mid-run — and failed/cancelled
aren't in library_history — so they only reappeared after the batch
ended. Now the whole active batch stays intact until it finishes.
2. search_all_sources runs every source CONCURRENTLY (asyncio.gather)
instead of sequentially, so the pool waits only for the slowest source
(e.g. usenet/Prowlarr) in parallel rather than summing all latencies.
3. The pool log now reports per-source contribution counts
(e.g. "usenet=0, hifi=11, soulseek=1") instead of just echoing the
chain, so a release-level source that returns nothing for a track-title
query is visible rather than appearing to have been searched.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds an opt-in search strategy toggle in the Quality Profile:
- priority (default): unchanged — first source in the hybrid chain that
meets a quality target wins.
- best_quality: pool candidates from EVERY source per query and download
them best→worst by actual audio quality; source order only breaks ties.
Implementation reuses existing plumbing so the retry system is untouched:
- engine.search_all_sources pools raw tracks across all configured,
non-exhausted sources (no first-source short-circuit).
- candidates.order_candidates: new quality_first sort path — profile
quality rank dominates, confidence/peer signals break ties. Priority
path is byte-for-byte unchanged (regression-locked by tests).
- task_worker passes quality_first + targets through; skips the redundant
hybrid-fallback block in best-quality mode (pool already covered it).
- Per-source retry budgets unchanged: a source that spends its budget is
added to exhausted_download_sources and thus dropped from the whole
pool. Independent of post_processing.retry_exhaustive.
- Query generator NOT touched.
Also clarifies the "Allow fallback" setting wording: it accepts OFF-LIST
quality as a last resort (not "walk down my list"), and notes that
lossy_copy.downsample_hires also bypasses the quality gate — the cause of
16-bit/MP3 files slipping through a 24-bit-only profile.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
get_audio_quality_string now appends the FLAC sample rate so the Downloads
quality chip and library history read e.g. 'FLAC 24bit/96kHz' instead of just
'FLAC 24bit' — surfaces hi-res frequency (44.1/48/96/192kHz) at a glance.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Unverified review actions (play/audit/approve/delete) only rendered for
persistent-history rows, so a freshly-completed unverified download — still a
live task without a 'history-<id>' task_id — showed no buttons until it aged
into history (Quarantine always worked because it uses the quarantine entry
id). Thread the library_history row id from import through to the live task
(add_library_history_entry now returns lastrowid -> context._history_id ->
task.history_id -> /api/downloads/all), and resolve verifHistoryId from it.
Also surface the real probed audio quality (mutagen-read from the file, e.g.
'FLAC 24bit') on completed rows as a chip, so you can see what was actually
downloaded.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Some Monochrome instances only have 30-second Tidal DOWNLOAD access: the HLS
variant playlist for a 220s track comes back as ~30s of segments + ENDLIST
(verified live on us-west.monochrome.tf — lossless=30s, hires=403). The client
downloaded that 30s file, which then got quarantined by the new audio guard.
Detect it at manifest time: sum the playlist's EXTINF runtime and compare to
the track's real duration (get_track_info). When the playlist is < 85% of the
track, decline the manifest and rotate the instance, so the download falls
through to a real source (Soulseek/Qobuz/Tidal/Deezer) instead of fetching a
preview. Best-effort — unknown duration disables the check (the post-download
audio guard remains the safety net).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The actual HiFi/Monochrome bug isn't silence padding — it's a TRUNCATED file:
the container claims the full length (e.g. 3:08) but only ~30s of audio
decodes. silencedetect finds nothing (there's no silent audio, just missing
audio) and ffmpeg's time= even reports 0 with no error, so the duration and
quality guards all pass.
Detect it by decoding and comparing the real audio length (astats sample
count / sample rate) against the container duration: reject when the real
audio covers < 85% of the claimed length. detect_broken_audio() runs this
truncation check first, then the silence-ratio check. Wire it into the guard
that runs at the integrity/length verification point.
Verified on the real file: 'only ~30s actually decodes of a 188s file (16%)';
a normal 180s file is not flagged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The silence guard sat after the quality guard, so a strict quality profile
quarantined every file before silence detection ever ran. Move it to right
after check_audio_integrity (where the length is verified) and before the
AcoustID/quality gates, so a mostly-silent file is caught regardless of its
quality verdict and reported with the correct reason. Same quarantine +
next-candidate retry pattern (trigger='silence').
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
HiFi/Monochrome HLS assembly can produce a file with the correct container
duration but only ~30s of real audio + silence padding — the duration and
quality guards both pass, so nothing caught it until you listened. Add
core/imports/silence.py: ffmpeg silencedetect over the audio, reject when the
silent fraction exceeds 50%. Wire it into the post-download pipeline with the
same quarantine + next-candidate retry pattern as the quality guard
(trigger='silence'), and surface it via import_rejection_reason. Fails open
when ffmpeg/mutagen are unavailable so tooling problems never quarantine a
legit file.
Also mark 'quality filter' and 'silence guard' failures as recoverable
quarantine rows in the downloads UI (were shown as plain failures).
Verified end-to-end: a 30s-tone + 180s-silence FLAC is flagged '86% silence
(only ~30s audible of 210s)'; a 210s tone passes. 7 parser unit tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Remove the per-source download-quality dropdowns (Tidal/HiFi/Qobuz/Deezer/
Amazon) — with the global ranked-targets system they were redundant and
conflicting. Add quality_tier_for_source(): picks the LOWEST source tier
that satisfies the user's top target (respects the quality ceiling, saves
bandwidth) or the source's max as best effort. Every source's search +
download + retry path now derives its tier from the global profile instead
of config_manager.get('<source>_download.quality').
Settings keep the per-source allow_fallback toggles; the quality selects are
replaced with a note pointing at Quality Profile.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
search_and_download_best applied confidence scoring to streaming results
but never quality-ranked them — only the Soulseek path did. Apply
rank_for_profile to the confidence-passing survivors so the best version
wins (match first, then quality). Stable ranking keeps confidence order
within an equal tier; an "or scored" fail-safe keeps a candidate to try.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add core/quality/selection.py: rank_with_targets() returns (ranked,
satisfied) where satisfied = a candidate meets a real target (strict).
load_profile_targets()/rank_for_profile() are the DB-backed wrappers.
search_with_fallback now skips a source that can deliver no target-meeting
quality and escalates to the next (source priority still wins among
satisfying sources; first source's results kept as fallback unless the
profile disables it). Returns RAW tracks — the satisfied check is a coarse
source gate; match-filtering + final ranking stay in the orchestrator so
the correct track is never pruned. Ranking is fail-open: a ranking error
never drops a source's real results.
Tested: rank_with_targets satisfied/fallback matrix + engine escalation,
stop-on-first, raw-not-pruned, fallback on/off. Amazon field test updated
for the corrected format token.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add core/quality/source_map.py centralising each source's tier->AudioQuality
mapping (Tidal/HiFi tiers, Qobuz real kHz/bit-depth, Deezer codes, Amazon
codec/tier). Add TrackResult.set_quality() to merge a mapped AudioQuality
onto a result. Wire HiFi, Qobuz, Deezer, Tidal, Amazon search results to
stamp real sample_rate/bit_depth so the global ranker no longer relies on
crude kbps heuristics for streaming sources. Fixes Qobuz/Amazon display
labels ('FLAC 24-bit/192kHz', 'Lossless') breaking format derivation.
Tested: 22 passing (mappers + set_quality merge semantics).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
wolf's report: Spotify shows 'Calma - Remix', Find & Add searches that literal
string, but the library stores the track as just 'Calma' (only the 3:58 duration
marks it the remix). The literal LIKE '%calma - remix%' misses, so it fell to the
OR-fuzzy fallback which floods on the common word 'remix' (20 unrelated '... remix'
hits). Dropping '- Remix' (searching 'Calma') finds it instantly.
Fix: search_tracks (and api_search_tracks) now retry on the BASE title — the part
before Spotify's ' - ' version separator — BEFORE the OR-fuzzy flood. So
'Calma - Remix' resolves to 'Calma' (or 'Calma (Remix)') and the noise fallback is
never reached when the base matches. New core.text.title_match.base_title_before_dash
(splits the first spaced ' - '; leaves bare hyphens like 'Up-Tight' alone).
Tests: pure helper (3) + real-DB integration reproducing the Calma case, the
parenthesized-remix variant, plain-title-unaffected, and no-flood (4). 64
search/match tests green.
Enrichment matched artists by NAME ONLY (0.85 gate), so for a common name
('Rone' has ~5 artists) it stored whichever the source ranked first — often the
wrong one, which then drove a wrong/sparse library 'Standard' discography while
'Enhanced' (the real owned albums) showed the full set.
Fix — use the decisive signal the library already has (the albums you OWN):
- worker_utils: pick_artist_by_catalog() + catalog_overlap_score() +
owned_album_titles()/release_titles(). When 2+ candidates clear the name gate,
fetch each one's catalog and choose the one overlapping the owned albums; falls
back to the current best-by-name pick when there's nothing to disambiguate or
no overlap (so the common single-candidate path makes no extra API calls).
- Wired into Spotify (covers Spotify-Free, same client), iTunes, Deezer (now
multi-candidate search_artists + get_artist_info store), and MusicBrainz
(match_artist gains owned_titles; release-groups as the catalog).
Re-match path (#868):
- build_reset_query now also clears the stored source-ID column for artist/album
item resets — previously a 're-match' only nulled match_status, so the worker's
existing-id short-circuit re-confirmed the WRONG id and never re-resolved. Tracks
excluded (ids live in tags, not a column).
- MusicBrainz also self-corrects its 90-day name->mbid cache: match_artist bypasses
a cached mbid whose catalog has ZERO overlap with the owned albums, so a re-match
isn't blocked by a stale wrong cache entry.
Tests: shared selector (9), per-worker disambiguation for all 4 sources + MB
backward-compat + MB cache-revalidation (8), reset-clears-id (2). 99 worker/
enrichment tests green.
Four refinements on top of the tiered matcher:
1. Direct source track-ID tier (new top tier): enrichment writes each source's own
track ID into the file tags (spotify_track_id/deezer_track_id/itunes_track_id/...).
If we have the active source's track ID, fetch that exact track by ID via
get_track_details — zero search. Tiers are now: track-ID -> ISRC -> album->track
-> artist+title. _read_file_ids reads ISRC + all per-source IDs in one tag read.
2. Skip already-proposed tracks: a re-run loads existing finding entity_ids for the
job and skips those tracks before any API call (pending stays deduped, dismissed
stays dismissed) — re-runs are cheap.
3. Wrong-version guard: the fuzzy tiers (album-search + track search) reject a
candidate whose length differs from ours by >5s (live/edit/remix with same title).
_load_tracks now selects t.duration; exact tiers (track-ID/ISRC/stored-album-ID)
skip the guard.
4. Tighter album matching: same-title cuts in an album are disambiguated by closest
duration when track_number doesn't decide it.
Findings record matched_via = track_id | isrc | album | search. 30 repair tests pass
(added track-ID tier, duration guard, dedup-skip, and unit coverage).
Replaces the Soulseek-only bit-depth heuristic with a source-agnostic
quality system that works across all download sources.
## core/quality/model.py (new)
- AudioQuality dataclass: format, bitrate, sample_rate, bit_depth
- QualityTarget: one ranked entry in the user's priority list
- filter_and_rank(): source-neutral candidate ranking
- rank_candidate(): scores any AudioQuality against ranked_targets
- v2_qualities_to_ranked_targets(): migration helper
## core/download_plugins/types.py
- SearchResult gains sample_rate + bit_depth fields
- audio_quality property returns unified AudioQuality
- AlbumResult gets audio_quality aggregated from tracks
## core/soulseek_client.py
- Parses slskd attributes array (type 4=sample_rate, type 5=bit_depth)
- Real values instead of kbps heuristic
- filter_results_by_quality_preference() replaced by filter_and_rank()
## database/music_database.py
- Quality profile v3 with ranked_targets list
- Auto-migration v2 → v3 on load
- Presets (audiophile/balanced/space_saver) updated to v3
## core/imports/file_ops.py
- probe_audio_quality(): reads actual downloaded file via mutagen
returns AudioQuality with ground-truth values
## core/imports/guards.py
- check_quality_target(): replaces check_flac_bit_depth
checks all formats/sources against ranked_targets
- check_flac_bit_depth() kept as backwards-compat wrapper
## core/imports/pipeline.py
- Uses check_quality_target() instead of check_flac_bit_depth()
- Quality mismatch triggers _requeue_quarantined_task_for_retry('quality')
so next-best candidate is tried before failing (same as AcoustID)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaces the blind fuzzy search with a smart hierarchy that uses the data we
already have, best identity first:
1. ISRC embedded in the file tags (enriched track) -> exact track.
2. Album -> track: use the album's stored source ID (albums.spotify_album_id /
itunes_album_id / deezer_id / musicbrainz_release_id / audiodb_id) when the
ALBUM is enriched (even if the track isn't); else find the album by searching
'artist album', then locate our track in that album's tracklist by normalized
title (track_number breaks ties). Pins the exact album context. (artist->album->track)
3. Plain artist+title search with similarity scoring. (artist->track) — loosest.
_load_tracks now returns dict rows (adds track_number + the album source-id
columns). Findings record matched_via = isrc | album | search. All clients
(spotify/deezer/itunes/discogs) expose search_albums + get_album_tracks with a
uniform {'items': [...]} shape, so the album tier is source-agnostic.
26 repair tests pass (added album-tier + _find_track_in_album coverage).
The job was doing a blind fuzzy search for every low-quality track, ignoring that
enrichment writes each track's ISRC + per-source IDs into the file tags. Now it
reads the file's embedded ISRC and resolves the EXACT track via each source's
'isrc:' search (universal cross-source key), guarded by an ISRC-equality check so
a source that ignores the syntax can't produce a false match — exact track, exact
album context, one call. Falls back to the name/artist fuzzy search only for
un-enriched tracks with no usable ISRC. Findings record matched_via=isrc|search.
4 new seam tests (guard accept/reject, ISRC-preferred-over-fuzzy, fuzzy fallback).
Phase 2 of the redesign. The tool that judged quality by extension and auto-dumped
matches into the wishlist is gone; quality scanning is now the reviewed
quality_upgrade repair job.
Removed:
- Frontend: Tools-page Quality Scanner card, its JS handlers/poller/socket listener,
help tooltip + tour entry (webui index.html, core.js, helper.js, wishlist-tools.js).
- Backend: /api/quality-scanner/{start,status,stop} endpoints, the in-memory state +
executor + 1s socket broadcast, the QualityScannerDeps/run_quality_scanner shim.
- core/discovery/quality_scanner.py: the auto-acting worker + deps class (the shared
match/normalize helpers stay — the new job imports them).
Rewired:
- Automation 'start_quality_scan' action now triggers the quality_upgrade repair job
via repair_worker.run_job_now() (AutomationDeps gains run_repair_job_now, drops the
4 scanner fields). Action block's vestigial scope field removed (scope lives in the
job's settings now). NOTE: the 'quality_scan_completed' trigger no longer fires (the
repair job doesn't emit it).
- Updated all automation test _build_deps helpers + conftest tool-progress harness;
deleted the obsolete worker test. 528 affected tests pass; 6123 collect cleanly.
QUALITY_TIERS / _get_quality_tier_from_extension kept (used elsewhere).
The old Quality Scanner tool judged quality by file EXTENSION only (a 128k and a
320k MP3 looked identical), ignored the bitrate-based quality profile, used min()
of enabled tiers so the default profile flagged the ENTIRE non-lossless library,
and auto-dumped every match into the wishlist with no review.
This new repair job does it properly:
- meets_preferred_quality(): pure, bitrate-AWARE decision honoring every enabled
quality bucket (320 MP3 passes a FLAC+320+256 profile; 128 MP3 doesn't). Floor
is the worst enabled bucket, not the best.
- scans watchlist artists or whole library, finds below-quality tracks, matches a
better version at scan time (reusing the existing tested match helpers), emits a
FINDING showing the match + confidence. Off by default; nothing auto-queued.
- _fix_quality_upgrade apply handler adds the matched track WITH album context to
the wishlist — the user-approved version of what the old tool did silently.
- Transcode/fake-lossless detection intentionally left to the existing Fake
Lossless Detector job.
12 seam tests incl. a regression pinning the default-profile flooding bug. The old
tool is still in place; removing it + rewiring its automation action is the next step.
Two issues in the same path:
1. The shared discovery modal pre-renders one row per track from a
separately-fetched frontend track list, then the poll dropped any backend
result without a pre-rendered row (if (!row) return). When the frontend's
track fetch came back rate-limited/partial (~21) while discovery's own fetch
got all 59, the surplus results vanished. Now the modal CREATES a row for any
result lacking one, so authoritative backend results drive the list (fixes
all sources sharing the modal).
2. get_playlist hydrated a whole relationships page in one _get_tracks_batch
call, but Tidal caps filter[id] at 20/request, silently truncating larger
pages. Chunk to the cap like get_album_tracks already does.
Seam + regression tests (tests/test_tidal_playlist_batch_chunking.py).
Status checks asked is_spotify_authenticated() (official OAuth only) instead of
is_spotify_metadata_available(), so a Spotify-Free primary read as disconnected.
get_primary_source_status had spotify_free awareness but it was dead code:
get_client_for_source('spotify') returns None unless officially authed, so the
free-availability probe never had a client. Fetch the client directly for that
check; add the missing free branch to the dashboard test message. Seam + regression tests.
The reconcile read each completed task's final_file_path to find paths — but not
every import path sets it (the verification worker marks the task completed
without it), so tracks that imported via that path were silently dropped (user
saw 3 of 5 symlinks). Root cause: leaning on a fragile per-task field.
Now reconcile_batch_playlists identifies the organize playlists the batch touched
(its own + any reached via a completed track's source_info provenance) and
rebuilds each from CURRENT library ownership via _rebuild_one_from_db
(check_track_exists over membership). It just asks the library what's owned, so
it's robust to HOW a track imported (modal worker / slskd monitor / verification
worker) and still prunes tracks that left. Takes a db handle; all three callers
pass MusicDatabase().
Reconcile tests rewritten for the DB-rebuild form (organize batch, wishlist
provenance, non-organize skip, plain no-op). 973 downloads/imports/playlist
tests pass.
on_download_completed and check_batch_completion_v2 are duplicate completion
paths. Monitor-detected downloads (Deezer / slskd-monitor / verification-worker
imports) finish the batch via the V2 path, but the materialize reconcile was
only added to on_download_completed — so those batches never built playlist
folders (no '[Playlist Folder] Rebuilt' line at all). Add the same non-fatal
reconcile to the V2 path. Now all three completion points (both lifecycle paths
+ the master.py all-owned path) materialize. 550 tests pass.
Symmetric to the post-download reconcile (which handles ADDITIONS): when a
playlist's membership is re-synced (the mirror step — scheduled refresh or the
manual mirror endpoint), rebuild its folder from current membership WITH prune
IF it's organize-by-playlist. So a track that just LEFT the playlist has its
symlink cleaned up the instant membership changes, not only on the next download.
Factored a shared _rebuild_one_from_db (used by the manual 'Rebuild' button and
the mirror hook) + rebuild_mirrored_playlist_if_organized. Gated to organized
playlists, non-fatal at both mirror call sites.
Now the invariant 'folder = the playlist's current owned members' holds on every
change: additions caught at download, removals caught at mirror. 2 new tests
(removed track pruned; non-organized skipped). 985 + 277 tests pass.
Replaces the two organize-only triggers with a single reconcile_batch_playlists
called at both batch-completion points. It groups the batch's newly-resolved
tracks by their per-track playlist provenance:
- the batch's OWN organize playlist → full (re)build with prune, and
- a track that completed for a DIFFERENT playlist (e.g. a WISHLIST fulfilling a
track that belongs to an organize playlist) → ADDED to that folder, no prune.
So a late wishlist arrival now lands in its playlist folder immediately, instead
of only on the next sync/manual rebuild — the folder = the playlist's owned
members, kept true on every ownership change regardless of download path. Uses
the paths the batch already captured (no DB re-match, no waiting on the server
scan/sync). Non-fatal.
3 new reconcile tests (organize full-rebuild, wishlist add-without-prune, plain
batch no-op). 983 downloads/imports/playlist tests pass.
- Settings: 'Playlists Folder' path field (Unlock pattern, separate-root help
text), a Symlinks/Copies selector, and a 'Rebuild playlist folders now' button
(standard test-button style). Wired through PATH_INPUT_IDS / load / save, plus
'playlists' added to the settings save allowlist so it persists.
- POST /api/playlists/materialize/rebuild → rebuild_organized_playlists_from_db:
rebuilds every organize-by-playlist folder from CURRENT ownership, re-matching
each track with check_track_exists (name, not IDs) so it self-heals after a
reorganize / membership change. +1 test.
70 materialize tests + JS integrity pass; settings round-trip wiring verified.
- Routing (step 5): organize-by-playlist tracks no longer set the per-track
_playlist_folder_mode flag, so they import NORMALLY into Artist/Album — exactly
what a normal download does. _playlist_name provenance is kept (origin.py).
- Triggers (step 4): build the playlist folder from the batch's own payload at
both end-of-flow points — the all-owned path in master.py (no downloads, so the
lifecycle never runs) and the batch-complete hook in lifecycle.py (after
downloads). Both gated on playlist_folder_mode, both non-fatal.
Works for the all-owned case (the smack test that did nothing before) and for
mixed owned/downloaded, with no source-ID or mirrored-playlist dependency. The
materialized folder uses the default ./Playlists root + symlink mode until the
Settings UI is added.
Updated the master test to assert the new contract (provenance kept, routing
flag gone). 979 tests pass.
materialize_playlist_from_batch(batch, download_tasks, config) collects the real
on-disk path of every resolved track from the batch's OWN payload — owned via
analysis_results.matched_file_path, downloaded via tasks.final_file_path — runs
each through the playback path resolver (Docker-correct), de-dupes, and hands the
list to rebuild_playlist_folder. Gated on playlist_folder_mode.
No re-matching, no source IDs, no mirrored-playlist lookup — works for any
organize-by-playlist download including the all-owned case. 5 tests. Still
isolated; the triggers wire it in next.
- settings: playlists.materialize_path (separate root, mapped apart from the
music library so the media server never double-scans it) + materialize_mode
(symlink|copy).
- core/playlists/materialize.py: pure filesystem engine that (re)builds a
playlist folder of relative symlinks (or copies) into the real library —
idempotent, prunes stale entries, disambiguates filename collisions, never
escapes the root, and auto-falls-back to copy when the FS can't symlink.
No DB, no app state; ops injectable. 13 unit tests.
Isolated + additive — nothing live calls this yet (stitcher/trigger/routing
come next).
The download analysis already matches every track to a library row via
check_track_exists / manual match, then discarded the result. Keep it: each
analysis_results entry now carries matched_file_path + matched_track_id (the
owned file's real location, or None). Symmetrically, a completed download task
now records final_file_path (where the import landed).
Purely additive, no behavior change, no new matching, zero perf cost — just
stops throwing away what the pipeline already computed. This is the foundation
for playlist materialization: owned + downloaded tracks both report where their
real file is, so the folder can be built by name match, not source IDs.
A DB<->filesystem path mismatch (Docker volume change, remount, Music
Paths unset for the container) makes EVERY library file fail to resolve
to a DB track, so the orphan detector flags the whole library as
orphaned. The mass-orphan check only logged a warning and then created
the findings anyway — so a user batch-applying 'move to staging' or
'delete' would relocate or wipe their entire library.
Make it a hard skip (create zero findings) like the dead-file cleaner
and stale-removal paths already do (#828). Centralise the predicate as
is_implausible_orphan_flood() alongside is_implausible_stale_removal()
so the rule lives in one tested place. Small genuine orphan sets still
surface unchanged — only an implausibly large flood (>50% and >20) is
suppressed.
Tests: seam cases for the new predicate + scan-level regressions (mass
mismatch -> 0 findings; small genuine set -> still reported).
Extends the watchlist export to the full library. The exporter is now general
(core/exports/artist_export.py, renamed from watchlist_export) — adds tidal/qobuz
links and an extra_fields passthrough, so the library export also carries
lastfm/genius URLs + soul_id, and an optional "library counts" toggle adds owned
album/track counts per artist.
- GET /api/library/artists/export?format=&links=&contents= — pulls every artists
row, normalizes onto the canonical *_artist_id keys, optionally GROUP-BY counts
for album/track totals.
- The export modal is now openArtistExportModal(scope): "Export Library" button in
the library header + the existing "Export" on the watchlist bar (a thin wrapper).
Library mode shows the extra "library counts" toggle.
Tests (11): builder across formats + the new tidal/qobuz links + extra_fields
columns; watchlist + library endpoint wiring. 64 integrity green; ruff clean.
An "Export" button on the watchlist filter bar opens a modal (same aesthetic as the
artist DB-record inspector) to export your whole watchlist roster — each artist's
name + source IDs (spotify / musicbrainz / deezer / discogs / itunes / amazon),
with an optional "external links" toggle that adds the discography URLs built from
those IDs. Live preview, copy, and download in the chosen format.
- core/exports/watchlist_export.py: pure builder (json/csv/txt + links, present-IDs
only, deterministic columns) — the single source of truth, fully unit-tested.
- GET /api/watchlist/export?format=&links= shapes the roster + returns it (with
X-Export-Count / X-Export-Ext headers for the modal).
- Frontend reuses the DB-record helpers (_jsonSyntaxHighlight / _arecCopy).
Tests (8): builder across json/csv/txt, links on/off, present-ids-only, empty +
bad-format fallback, mime/ext, and endpoint wiring. ruff clean; 64 integrity green.
Scoped to the watchlist for v1; library-wide export + a "library contents"
(owned albums/tracks) option are natural follow-ups.
New Aria2 JSON-RPC adapter, alongside qBittorrent / Transmission / Deluge. Aria2's
RPC (default :6800/jsonrpc) maps cleanly onto the uniform adapter contract:
- the --rpc-secret token leads every call as "token:<secret>" (no username — the
secret uses the existing password field),
- addUri returns a GID (our torrent id); tellStatus → TorrentStatus with state
mapping (active→downloading, or seeding once the payload is complete; waiting→
queued; etc.),
- remove picks forceRemove vs removeDownloadResult by status, and (since aria2
doesn't delete files on remove) unlinks the file paths itself for delete_files,
- bare-host URLs get /jsonrpc appended.
Wired into adapter_for_type + the Settings dropdown (with a help note: port 6800,
secret in the Password field). All adapter methods go through the same interface,
so the stall/orphan handling and downloads pipeline work unchanged.
Tests (9): registry wiring, state mapping (incl. active→seeding), token-prefixed
params, /jsonrpc fixup, status parse (+ name fallback, no div-by-zero). 126 torrent
tests green; ruff clean.
A maintenance job to keep the music library tidy — finds empty folders left behind
by imports/relocations/deletions (empty artist/album dirs, or dirs holding only OS
junk like .DS_Store/Thumbs.db) and removes them.
Safety is the focus (deleting directories is destructive):
- only TRULY empty folders are flagged — a folder with a cover image or any audio
is never touched; only OS-junk files count as "no real content" (a setting),
- the library root + symlinked dirs are never removed,
- walks bottom-up so a parent left empty by its removable children cascades,
- the apply handler RE-CHECKS emptiness at delete time, so a folder that gained a
file between scan and apply is left alone.
dir_is_removable + remove_empty_folder are pure/injectable seams. Wired through the
job registry, repair_worker apply handler (_fix_empty_folder), fixable-types, and
the findings UI. Opt-in (default off), weekly interval.
Tests (10): removable decision (empty / real-file / surviving-subdir / junk-only /
strict mode) + apply re-check (removes empty + junk, refuses content/root/symlink).
Repair + integrity suites green; ruff clean.
A transient ping failure (network blip, Navidrome busy mid-scan) makes
_setup_client null out the configured creds, and _connection_attempted then
latches the client "disconnected" — so is_connected() returned False forever until
the user hit the manual Test button to re-read config. That's the reported
"disconnects every 5-10 min, reconnects instantly on Test."
Fix: ensure_connection no longer latches on a failed attempt — once a short
throttle (_RECONNECT_THROTTLE_S = 20s) elapses it re-attempts, and is_connected()
triggers that retry whenever it's currently disconnected. So a blip recovers on its
own within the next status check, no manual reconnect. The throttle prevents ping
storms when Navidrome is genuinely down.
Tests: transient failure self-heals after the throttle (and doesn't re-ping within
it); a connected client never re-pings; first connect attempts once. 115 navidrome/
media-server tests green.
Invariant: while security.require_login is on, every profile must have a login
password or it's locked out. Previously only the admin's own anti-lockout existed,
so members could be stranded (created without a password, or login flipped on while
passwordless members existed). Closed all the write-points:
core/security/login_provisioning.py (pure policy, single source of truth):
- members_without_password(profiles) — non-admin profiles that can't sign in
- create_needs_password(require_login) / removing_password_strands(require_login)
Wired into web_server:
- create_profile: while login is on, a new member must be given a password (400
otherwise) and it's set on creation.
- enable-login (settings save): refuses to turn login on while any member lacks a
password — lists them — same shape as the existing admin anti-lockout.
- set-password: refuses to CLEAR a password while login is on (would strand them).
UI: Create Profile form gains a login-password field (alongside the optional PIN);
the Manage Profiles per-member password button (prior commit) covers existing
members + changes.
Tests: pure policy seam + endpoint enforcement (create blocked w/o password when
on, allowed w/ password, no friction when off, clear blocked when on). 442
profile/settings/auth tests green; ruff clean.
Post-processing applies ReplayGain only to slskd/WebUI downloads — content added
via Lidarr, the REST API, or by hand never got it, and there was no way to (re)apply
RG to existing tracks or fix ones where analysis failed (raised in #437 + comments).
New ReplayGain Filler repair job (sibling of Lyrics/Cover Art fillers): scans for
tracks with no ReplayGain track-gain tag and creates a finding per track; the scan
only READS tags (cheap) and no-ops when ffmpeg is absent. Applying a finding runs
the same ffmpeg ebur128 analysis the import pipeline uses (gain = ref - LUFS) and
writes the RG tags in place — no moves, no re-matching. Opt-in (default off),
schedulable like the other maintenance jobs.
Wired: job registry, repair_worker apply handler (_fix_missing_replaygain) +
fixable-types, and the findings UI (label / fix-button / detail rows).
Tests: pure needs_replaygain decision (missing/blank/present/+0.00-is-tagged) +
the apply handler's analyze→compute→write seam with the pipeline gain formula,
ffmpeg-absent + missing-file guards, and registration. 93 repair tests green.
The dev-nightly build runs `ruff check .` before "Build and push to GHCR" in the
same job, so the three S110 (try/except/pass) errors introduced since the last
green build (ce6ce4d) failed the lint step and SKIPPED the image push entirely —
every dev-nightly since #704 went red, so the dev image was never rebuilt and none
of the recent fixes (incl. the #852 WebSocket login-bypass fix) ever shipped to
the image users pull.
All three are deliberate best-effort swallows; annotate them with the repo's
existing `# noqa: S110 — <reason>` convention rather than adding dead logging:
- relocate.py: tag write is best-effort (re-import re-derives tags)
- acoustid_scanner.py: verification-status tag is optional context
- web_server.py: audio-duration probe falls through to 0
ruff check . + compileall now clean; pytest already passed in CI at ce6ce4d.
PR #853 added artist album-list caching to Deezer, but unlike the Spotify path it
had no equivalent of the truncated-fetch guard: Deezer paginates, and a transient/
malformed response on page 2+ (artist with >100 albums) broke the loop and cached
the PARTIAL list as the full discography — serving an incomplete album list from
cache until TTL.
Fix: track whether pagination finished cleanly. A malformed/empty-of-data response
mid-walk now clears a `complete` flag and the artist→album-LIST is cached only when
complete. Individual album entities still cache regardless (each is complete; we
just have fewer). A clean end (short page / empty page / reached limit) still caches
as before.
Tests: a page-1-ok / page-2-errors walk no longer caches (second call refetches
instead of serving a permanently-incomplete list); a clean two-page walk still
caches (happy path intact). 181 deezer/metadata tests green.
noldevin: a magnet stuck "downloading metadata" ran 11h despite a 15-min stall
timeout, got cleared from SoulSync but left active in qbit, then re-grabbed as a
duplicate. Two bugs:
1. Stall never fired on metaDL. StallTracker reset its clock on any `downloaded`
byte increase, but a metaDL torrent's byte counter still ticks up from DHT/peer
protocol overhead while making no real progress — so the clock reset forever.
Fix: the byte counter only counts once metadata is in (size>0). During the
metadata phase (size==0) the only thing that counts as progress is *obtaining*
metadata, so a magnet that can't even do that within the timeout is correctly
flagged stalled. size=None preserves the old byte-only behavior (back-compat).
2. Orphaned in qbit. The monitor's stall exit removed the torrent, but the `error`
exit and the 6h deadline exit only marked the download failed — leaving the
torrent active in qbit, untracked here, so SoulSync re-grabbed the same dead
torrent (qbit logs the duplicate-add). Fix: both terminal exits now run
_cleanup_torrent (shared with the stall path), which removes+deletes (abandon)
or pauses per the stall action — nothing is left orphaned.
Tests (10 new): metaDL byte-noise no longer resets the clock (stalls at timeout);
obtaining metadata resets it; real byte-progress still tracked after metadata;
_cleanup_torrent removes+delete_files on abandon / pauses on pause / no-ops on
empty hash or no adapter / swallows a client error. 151 torrent tests green.
The #832 fix enforces the launch PIN / login via a Flask before_request hook, but
that hook does NOT run for the socketio handshake — empirically a normal endpoint
401s while /socket.io/ returns 200 with the gate on. So removing the client overlay
(Safari "Hide Distracting Items", devtools) + opening a socket streams live data
(downloads, logs, dashboard, notifications) completely unauthenticated.
Fix: the socketio connect handler now enforces the same check and returns False
(rejects the connection) when a gate is active and the session isn't verified.
Rejecting connect blocks every downstream WS event (subscribe/join), so all live
data is covered. core/security/ws_gate.is_ws_connection_blocked is the pure seam:
login mode (when on) > launch PIN > open, mirroring the HTTP gate exactly. Fails
OPEN on a config-read error, same as the HTTP gate.
Audited every other surface empirically with the gate on + unauthenticated: SSE
streams, catch-all pages, library/dashboard data, admin endpoints, search,
image-proxy, audio-stream (incl. a /etc/passwd traversal probe) all 401; /api/v1
key-gated. The WebSocket was the only hole.
Tests (10): pure gate logic (login>pin precedence, all on/off combos) + real
socketio.test_client integration — connect rejected when gate on + unauthenticated,
allowed when gate off or PIN verified.
Root cause (from the reporter's app.log): a ListenBrainz weekly playlist syncs
through the in-memory youtube_playlist_states discovery machine. When that live
state is lost — a Docker restart, or the discovery process ending while the user
waits for the media-server scan — the DB discover-download snapshot survives but
the live state is gone. Every recovery action (Cancel/Reset/Delete) then hit
`key not in states` and returned 404 "YouTube playlist not found" (hence the
confusing "Youtube" on a ListenBrainz playlist), leaving the playlist permanently
wedged with no way to dismiss or re-sync. Works for the maintainer because a
single session with no restart keeps the live state alive.
Fix — these are cleanup ops, so "the thing is already gone" is SUCCESS, not 404:
- cancel_sync core (shared by YouTube + ListenBrainz + Tidal/Deezer/Qobuz/...) →
missing key returns idempotent success.
- reset_youtube_playlist / delete_youtube_playlist → same.
The playlist becomes recoverable: Cancel/Reset clears the dead state and the user
re-syncs fresh.
Tests: cancel_sync core (missing key = idempotent 200 not 404; present key still
cancels + clears the worker + reverts phase); endpoint-level idempotency for
cancel/reset/delete; updated the old test that locked the 404 wedge. 834 sync/
discovery tests green.
The 'retag' fix corrects a mismatched file's tags/DB but leaves it in the WRONG
artist/album folder, so the library shows the right title while the file sits under
the previous track. AcoustID yields only title+artist (no reliable album), so an
in-place move has no safe target.
New 'relocate' action: retag the file, move it into Staging, drop the stale tracks
row, and clean up the emptied folder. The auto-import worker (which watches Staging)
re-identifies it with full metadata and files it correctly — reusing the import
pipeline instead of guessing a destination.
- core/repair_jobs/relocate.py: pure, injectable orchestration (retag -> move ->
drop row) + collision-safe staging_destination. Row is dropped only AFTER a
successful move, so a failed move never orphans the library entry.
- _fix_acoustid_mismatch gains the 'relocate' branch (thin wrapper: resolve path,
staging dir, drop-row closure, empty-parent cleanup).
- UI: "Relocate" button on the AcoustID-mismatch fix modal.
Tests (8): staging-dest collision suffixing; relocate happy path; tag-write failure
still relocates; FAILED move does NOT drop the row; no-tags skips write; a real
file move through safe_move_file; and a handler integration test (file moved to
staging + tracks row deleted end-to-end). Repair + integrity suites green.
resolve_history_audio_path drives a DESTRUCTIVE delete (os.remove), but lived
endpoint-bound in web_server with zero tests. Lifted to core/matching/history_paths
with injected effects (exists / resolve_library_path / lookup_titled_paths) so the
fallback chain — and the collision-safety that stops delete() from removing the
wrong same-title file — is a clean importable seam. web_server now wraps it (DB
lookup + os.path.exists + prefix resolver injected); behavior preserved.
9 tests lock it: recorded-path hit, prefix-resolve fallback, single tracks-table
candidate, and the safety rules — multiple same-title candidates with NO artist ->
None (refuse to guess), artist filter picks only the matching path, artist named
but unmatched -> None, no-title/empty-lookup -> None. Full suite green (5906).
The candidate matcher rejected valid downloads of titles with a '/' or ':' (e.g.
Sawano's "You See Big Girl / T:T") because the unified normalize() removed those
chars ("t:t" -> "tt") while keeping the '_' that source filenames substitute for
them ("T_T" -> "t_t") — the asymmetry tanked the similarity score. Now '/ \ : _'
all map to spaces before the strip, so "/ T:T" and "_ T_T" both normalize to "t t".
Verified on the real library: similarity for the Sawano pair 0.927 -> 1.000; only
348/40786 strings (0.85%, all containing those separators) change; worst-case
joined-variant match (e.g. "12:05" vs "1205") stays 0.889, well above the 0.70
title threshold — no match regressions. Fixes the matching half of #851.
Closes the forgot-login-password gap. A per-profile recovery question + answer lets
a locked-out user reset their own password.
- DB: additive recovery_question + recovery_answer_hash columns (idempotent
migration). set/get-question/verify/has methods; answer is hashed (pbkdf2) and
matched forgivingly (trim + lowercase + collapse whitespace). No recovery set →
never verifies.
- Endpoints (allowlisted in the login gate so they work pre-auth):
GET /api/auth/recovery-question?username= (generic 404 when absent),
POST /api/auth/recovery-reset {username, answer, new_password} — brute-force
limited; a correct answer sets the new password + authenticates the session.
POST /api/profiles/<id>/set-recovery (admin or self) to configure it.
Tests: set/get/verify, forgiving match, hashed-not-plaintext, no-recovery-never-
verifies, full reset flow (wrong answer rejected + password intact; correct answer
resets), unknown-user 404. 25 tests pass. Next: the Settings + login-screen UI.
The backend auth for opt-in username/password mode (security.require_login, default
off → zero change; the launch PIN + picker behave exactly as today).
- core/security/login_gate.py: pure gate (mirrors launch_lock) — when login mode is
on, an unauthenticated session reaches only the page shell, /api/auth/login,
/api/auth/logout, /api/profiles/current, /api/setup/status, and the key-authed
/api/v1 API. Deliberately does NOT expose the profile list pre-auth (you type your
name, not pick from a roster).
- _enforce_login before_request enforces it; _enforce_launch_pin no-ops when login
mode is on (login replaces the shared PIN, per design).
- POST /api/auth/login (username = profile name, case-insensitive; brute-force
limited per IP; generic error so names don't leak) + POST /api/auth/logout.
- Anti-lockout: the settings save refuses to turn ON login mode until the admin
account has a password.
Tests: gate blocks→login→access→logout→blocked; case-insensitive username; wrong
password / passwordless profile / unknown user all 401 generically; login list not
exposed pre-auth; can't enable login without an admin password. 12 tests pass. Next:
the login screen + set-password UI + the toggle (increment 3).
Lets SoulSync sit behind Authelia/Authentik/oauth2-proxy as the gatekeeper: when
security.auth_proxy_header names a header (e.g. Remote-User), a request carrying it
is treated as already-authenticated and passes the launch lock — the proxy did the
login (with 2FA).
- core/security/auth_proxy.py: trusted_proxy_user(get_header, header_name) — returns
the user iff the configured header is present + non-empty; empty header name (the
default) → always None → feature off.
- _enforce_launch_pin ORs it into pin_verified. OFF by default, so a direct install
is unaffected AND a client-spoofed header does nothing unless the operator opted in.
- Doc'd in Support/REVERSE-PROXY.md with the must-strip-client-headers warning.
This is the lightweight Tier 3 (auth-proxy integration), not a full per-user login —
the proxy owns identity; SoulSync trusts it.
Tests: helper off/on/blank/exception-safe; integration — trusted header passes the
gate, no header is locked, and (the safety pin) a spoofed header is IGNORED when the
feature is off. 6 tests pass.
Fold a conservative security-header set into the SAME opt-in proxy mode, so it's
zero-impact when off. When security.trust_reverse_proxy is on, an after_request
adds X-Content-Type-Options: nosniff, X-Frame-Options: SAMEORIGIN, and HSTS
(safe — only honoured over the proxy's HTTPS), via setdefault so it never clobbers
a header the proxy already set. No CSP (needs per-deploy tuning; better at the
proxy). When OFF (default), the after_request isn't registered → no headers added.
Tests: off adds none of the headers; on adds all three. Doc updated. 6 tests pass.
A publicly-exposed instance gated only by the launch PIN was brute-forceable. Added
a lenient in-memory failed-attempt limiter (core/security/rate_limit.py): 10 wrong
PINs from one IP within 5 min → 429 with Retry-After, failures age out on their own
(self-heal, no persistent lockout), and a CORRECT entry clears that IP instantly.
Wired into /api/profiles/verify-launch-pin. By design it can only ever trigger on a
flood of WRONG PINs — correct entry, a couple of typos, or a no-PIN install are
never affected, so normal use sees no change. Keyed per-IP so an attacker can't
lock out a legit user.
Tests: limiter is lenient under threshold, trips on a flood, success clears it,
failures self-heal, per-IP isolation; endpoint returns 429 after 10 wrong PINs with
Retry-After. 6 tests pass.
Tier 1 of "secure behind a reverse proxy". STRICTLY opt-in so direct/LAN installs
are byte-for-byte unchanged.
- core/security/reverse_proxy.py: apply_reverse_proxy_mode(app, config_get) — a
no-op unless security.trust_reverse_proxy=true. When OFF (default), the app is
untouched: no ProxyFix, X-Forwarded-* stays UNtrusted (a direct client can't
spoof its IP/scheme), session cookie keeps Flask defaults. When ON (operator is
behind nginx/Caddy/Traefik with TLS): trust one proxy hop's X-Forwarded-*, and
mark the session cookie Secure + SameSite=Lax. Any config error → safe no-op,
never breaks startup.
- Wired once at app init.
- Support/REVERSE-PROXY.md: nginx (with the Socket.IO Upgrade headers people
always miss) / Caddy / Traefik configs, the setting, and the "put auth in front
(Authelia/Authentik/oauth2-proxy)" recommendation + the off-for-plain-HTTP note.
Tests: off (and missing-key, and a config exception) is a strict no-op — not
ProxyFix-wrapped, cookie defaults intact; on wraps ProxyFix + secures the cookie;
and the real web_server app is NOT in proxy mode by default. 5 tests pass.
The enhanced-tab "Sync" button's stale-removal phase deleted any track whose file
wasn't on disk, with NO guard — so if the music storage was momentarily
unavailable (sleeping NAS, dropped mount, unmounted Docker volume, WSL hiccup),
os.path.exists returned False for EVERY file and one click wiped the whole artist
(tracks + their now-"empty" albums) from the DB. The deep-scan path already had a
50%-stale safety net (#828); this endpoint never got one.
- New core/library/stale_guard.py: is_implausible_stale_removal(missing, total) —
a tested rule (skip removal when missing > 50% of a >=5-track set), centralised
so every stale-removal site can share it.
- sync_artist_library: if the guard trips, SKIP removal (delete nothing), return
removal_skipped + warn; the frontend shows "storage may be offline — skipped"
instead of silently deleting. Empty-album cleanup now also only runs on the
non-skipped path and uses `album_id IS NOT NULL` (fixes the NOT IN-with-NULL
no-op). Frontend also refreshes the view on additions, not just removals.
- @admin_only on the endpoint — it deletes tracks + albums but was ungated, while
the sibling delete_album endpoint is gated.
Deep scan was already safe (different mechanism: server-diff + its own 50% guard).
Tests: guard unit rules; endpoint skips removal when all files missing (keeps the
tracks), removes only the genuinely-gone few otherwise, and 403s for non-admins.
7 new tests pass.
