Root cause of albums re-downloading forever despite completing (Linkin Park "Hybrid
Theory"): the MB resolver read the tracklist from an arbitrary releases[0], often a
later reissue/deluxe. HT resolved to a 2023 13-track reissue (~52 min); a complete
2000 12-track download (~38 min) then failed _download_problem's duration gate
(2267s < 2877s×0.85) and fell through peer after peer, never importing.
- _pick_release: choose the ORIGINAL standard edition (Official, earliest date)
instead of releases[0], so completeness/duration are judged against the edition
sources actually deliver.
- _DURATION_MIN_RATIO 0.85 -> 0.65: a source legitimately delivering a shorter
edition than MB's release shouldn't be rejected; only genuinely truncated/preview
downloads (a fraction of expected runtime) should be.
Together with the est-time peer ranking, this stops the download→reject→re-download
loop. 334 tests pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Show a live "~Xm left" countdown on the download progress bar. The slskd download
loop already polls byte transfer every 3s; measure the actual throughput (EWMA-
smoothed bytes/sec), compute seconds-remaining from the album's total bytes, and
surface it. New nullable Job.downloadEtaSeconds (migration; web entrypoint runs
prisma migrate deploy), threaded through the on_progress callback (optional 2nd arg,
so other adapters/callers are unaffected — they report no ETA). API exposes it;
queue.tsx renders etaLabel() after the track count. Null when unknown (no sample yet
or a source that doesn't report bytes). Worker 330 tests, web 214 tests, both green;
verified live-rendered as "42% · 6/14 tracks · ~4m left".
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Draining the queue via Soulseek was stalling: slow peers (~100KB/s) never finished a
full FLAC album within the fixed 10-min timeout, got cancelled at ~89%, and the
fall-through restarted from scratch on the next peer — while single-song folders named
like the album were also being tried.
- Stall detection: abandon a peer with no byte progress for ~90s (queued/dead) instead
of waiting the full timeout; a peer that IS progressing keeps its slot up to a 60-min
backstop, so a slow-but-working transfer can finish.
- Rank search candidates by peer: free upload slot, then speed, then shortest queue.
- Drop candidates with far fewer tracks than the release (single-song folders) up front,
so no download attempt is wasted on them.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
With qualityCutoff=3 (want hi-res), a Soulseek-grabbed CD copy stays below cutoff and
is re-attempted for upgrade. Without a guard, when Qobuz is gated by pacing those
attempts would (a) re-search every owned album each monitor cycle and (b) re-download
a same-quality Soulseek copy that keep-if-better discards — pure churn.
Add an upgrade guard: a >= CD-lossless copy can only be improved by hi-res Qobuz, so
when Qobuz is gated, skip the job WITHOUT searching (pre-search skip); and as a safety
net, skip the download when the best-ranked candidate can't beat the existing copy.
Force jobs always proceed. Extracted the shared keep-existing path into a helper.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two compounding causes of 'many downloads of one album from different sources' in
slskd: (1) the slskd client raised on timeout/error WITHOUT cancelling the transfer
it enqueued, so as the pipeline fell through to other peers each left a transfer
running; (2) the incomplete-download fall-through was uncapped, grinding through
every candidate of a popular album (~200 Soulseek peers), amplified now that Qobuz
pacing pushes most jobs to Soulseek.
- SlskdClient.download now cancels its enqueued transfers (DELETE) on any abandon.
- The download loop caps at _MAX_DOWNLOAD_ATTEMPTS (6) candidates.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Avoid re-triggering a USER_BLOCKED account by capping Qobuz to a human-looking
volume and cadence. A per-album gate (worker/lyra_worker/qobuz_gate.py) allows
Qobuz only within active hours, under today's cap (a warm-up ramp of a steady
cap), and past a randomized spacing since the last Qobuz download; the gap spreads
the day's budget across active hours with jitter. When the gate is closed the
pipeline drops Qobuz candidates so the job falls through to another source (and
monitored albums upgrade to Qobuz later); a Qobuz-only setup is never gated.
Config keys read live each loop; counters reset at the day boundary via the DB
clock. Tunable in Settings -> Qobuz -> Pacing (new /api/qobuz/pacing route).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Completeness/duration verification was applied only to the first candidate that
downloaded successfully; if that copy was incomplete or too short the whole job
failed, even when a later-ranked source had a complete copy (e.g. John Mayer 'The
Search for Everything' — Qobuz reliably fails one hi-res track, so it never fell
back to the deluxe edition or Soulseek).
Extract the checks into a pure _download_problem() and run it INSIDE the download
loop: a candidate that downloads but fails verification now falls through to the
next-ranked source, and the job only fails (needs_attention) when every candidate
is incomplete — reporting the last shortfall reason.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The completeness + duration guards used MusicBrainz's target.track_count /
total_duration_s, but MB picks an arbitrary releases[0] that is often a
deluxe/expanded edition with more tracks (and runtime) than the standard album a
source legitimately delivers. So complete standard-edition downloads were falsely
rejected — The Script 'No Sound Without Silence' (11 vs MB 12) as 'incomplete
download', Skillet 'Unleashed' (12 vs MB's 20-track deluxe) as 'too short'.
Now completeness is judged against the chosen source's own track count
(winner.track_count): a truncated download (fewer files than the source promised)
still fails, and the duration expectation is scaled to the delivered edition's
size. A separate 'implausibly small vs MB' guard (source has <= half MB's tracks)
still rejects a lone-track 'full album' video. Genuine partials (e.g. John Mayer,
a Qobuz track failing mid-download) still correctly fail.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Soulseek's slskd search polls up to ~2 min; searching adapters sequentially
added that on top of Qobuz's instant search. Run the per-adapter searches
concurrently via a ThreadPoolExecutor so total match time is the slowest
single source, not their sum. Only Qobuz drives streamrip's shared asyncio
loop and there's one Qobuz adapter, so no two threads use that loop at once.
Results collected in adapter order (map preserves order) for deterministic
candidate ordering; per-adapter failures still swallowed (down source = no
candidates).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Own-state ("in library") relied on an exact artist-NAME match, so an owned
album whose name differed from the MB canonical (punctuation/locale/feat.)
showed as not-owned even when followed. Capture the MusicBrainz artist MBID
on LibraryItem and prefer it for matching.
- New LibraryItem.artistMbid (migration add_library_artist_mbid). Populated at
import (pipeline) + scan, both of which already hold the resolved target;
the pipeline also now stores rgMbid (it previously didn't). COALESCE on
conflict backfills MBIDs on re-scan without clobbering.
- /api/artists "in library, not followed" matches by artistMbid first, name
case-insensitively as fallback (rows predating the column). Library route
prefers the item's own artistMbid over the release join.
Last.fm/discover own-state still use name-match (they work) — a smaller
follow-up. worker 233, web 186 tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The verify step checked track COUNT but not playtime, so a truncated file or
a short preview substituted for a real track could import as "complete". Add
a total-duration check: sum the staged audio's playtime (mutagen) and fail
the job when it falls below 85% of MusicBrainz's expected total
(target.total_duration_s, already resolved). Generous tolerance so
edition/encoding differences don't false-positive; bonus tracks (over-long)
always pass.
The duration reader is injected (measure_duration param, default the real
mutagen-based reader) so the fail/pass paths are unit-tested; it returns None
for the offline fakes (no real audio), skipping the check exactly like the
staged-count guard. worker 233 tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add Request.force (migration add_request_force): the pipeline's intake now
skips the "already in library" dedupe for a forced job, so an owned album is
re-downloaded. The import step still keeps the new copy only when it's higher
quality, so a forced upgrade can never downgrade what's on disk.
- Worker: _is_force_job → dedupe bypassed when Request.force.
- POST /api/library/[id]/upgrade finds the matching MonitoredRelease and
enqueues a force request (guards against stacking on an in-flight job).
- Library route exposes monitoredReleaseId; the album modal shows a
"Replace / upgrade" button when a release exists.
Worker + web tests added (pipeline force re-acquire; upgrade route).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Completeness only compared the adapter's self-reported track_count to expected,
so a silently-skipped track (0-byte placeholder — see the Qobuz 24K Magic bug)
imported as a "complete" album. Now the verify also counts the NON-EMPTY audio
actually staged and fails "incomplete download" when fewer than expected are
present (guarded with `staged and ...` so fake adapters that stage nothing keep
using the count path). A real broken download now goes needs_attention (or falls
through to another source) instead of importing a dead track.
Full per-track duration matching against MB recording lengths (the rest of #16)
remains a larger follow-up. worker 224 tests / 7-skip.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Bruno Mars "24K Magic" track 1 repeatedly downloaded as a 0-byte file. The
worker log showed the real cause: streamrip's track download hit
HTTPException("got more than 100 headers") — Qobuz's CDN returns >100 response
headers for some tracks and Python's http.client rejects them at the default
_MAXHEADERS=100, so streamrip skips the track and leaves an empty placeholder.
- Raise http.client._MAXHEADERS to 1000 on import of the Qobuz adapter, so those
responses parse and the track downloads.
- _count_staged_audio now skips 0-byte files (a failed-track placeholder is not a
real track) — improves the download-progress estimate and stops counting dead
tracks. (A verify-level "reject 0-byte tracks" safety net is a follow-up under
#16 per-track duration matching — it needs the pipeline fakes to write real
files, deferred to avoid disproportionate churn here.)
worker 223 tests / 7-skip.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Library modal showed MusicBrainz's canonical tracklist, hiding incomplete
downloads / edition or track-order mismatches, and showed nothing for albums
without an rgMbid. Now it shows the REAL files on disk.
- LibraryItem gains trackNames String[] (migration add_library_track_names):
the album's on-disk "## Title.ext" audio filenames.
- Captured at import (pipeline._import lists the final album folder) and at scan
(scan._record_owned now also refreshes trackNames on re-scan via ON CONFLICT
DO UPDATE + xmax=0 to preserve the newly-imported flag). New shared
library.list_audio_files() helper.
- /api/library exposes trackNames; the AlbumModal prefers on-disk tracks when
present (parsed "## Title" → position/title, zero network, labeled "On disk ·
N tracks"), falling back to the MusicBrainz listing otherwise. Items imported
before this column show the MB fallback until re-scanned.
worker 221 tests / 7-skip, web 153, tsc + build clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Floor bar used a file-count poller that jumps 0→~99% for small/fast albums
(streamrip reports only 0/1). But yt-dlp and slskd already emit real byte-level
progress via the on_progress callback the pipeline was discarding
(lambda _pct: None).
- Thread a throttled on_progress → Job.downloadProgress (write on a >=1% move or
every 0.5s; clamped 0..1; write failures swallowed). Called synchronously from
adapter.download on the pipeline thread, so it reuses the main conn safely.
- The file-count poller stays as the FALLBACK: it skips writing while a
`reports` flag is set (an adapter is driving real progress), and the flag is
cleared per attempt so a fall-through to streamrip still gets the estimate.
- Smooth bars for YouTube + Soulseek now; Qobuz/streamrip keeps the file-count
estimate until its per-track callback is hooked later.
worker 216 tests / 7-skip (on_progress write/throttle/clamp covered; existing
pipeline tests now exercise the fake's on_progress end-to-end).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add Job.downloadProgress (0.0-1.0), derived from a worker background
poller that counts audio files landing in staging / expected track
count, since streamrip only reports 0/1. No UI yet (slice B).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
_mark_chosen ran only after a download succeeded, so throughout the 'downloading'
state no candidate was chosen — the Floor row's "source · format · tracks" line was
blank (only elapsed showed). Now mark the candidate exclusively at each download
attempt so `chosen` reflects the source in flight (and updates on fall-through).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Staging is no longer forced under the library. The worker reads STAGING_ROOT
(container path, default /music/.staging) and stages per-job downloads there;
docker-compose mounts ${STAGING_DIR:-${MUSIC_DIR}/.staging} at /staging with
STAGING_ROOT=/staging. Set STAGING_DIR to a fast local disk when the library
(MUSIC_DIR) is a network share so temp download I/O stays off the share.
Safe cross-volume: import_album already assembles into a temp dir on the
library volume and swaps atomically there, so partial downloads never touch
the library and the final swap stays atomic. clear_staging_root now clears the
root's contents (mount-safe) rather than removing the root. Documented in
.env.example; new tests cover the separate-root path.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rewrite run_pipeline's download->import section to download into a per-job
staging dir (cleared before each candidate), verify completeness there,
promote only a complete album into the library via import_album, tag the
final dir, then DB-import. The staging dir is always removed in a
try/finally, so a partial/failed download never touches or corrupts the
library, and a retry can no longer leave orphan files behind.
Also point test_upgrade_e2e.py's run_pipeline call at dest_root="/tmp/lib"
(matching every other pipeline test) since import_album now unconditionally
creates the final dir, which requires a writable dest_root.
The monitor's quality-upgrade path was inert: run_pipeline's intake
short-circuit treated any existing LibraryItem as "done", so a
monitor-driven upgrade job for a below-cutoff copy jumped straight to
imported without downloading. Intake now checks whether the job is
monitor-driven (Request.monitoredReleaseId IS NOT NULL) and, if so,
only short-circuits when the existing copy already meets the quality
cutoff; plain manual requests keep exact slice-1 behavior. Also drops
an unused `field` import in browser.py.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Wrap each adapter's search() call in the match stage in try/except so
a down source (dead slskd, Qobuz auth error, etc.) contributes no
candidates instead of crashing run_pipeline and killing the worker.
Jonathan