RemoveTrack took the admin's id and dropped it, so a merge left an
audit trail and the more destructive removal did not. It now reads the
track's title, artist, album and file path before the delete and writes
them, with whether the album and artist went too, under the new
track_remove action.
The audit action groups lose their plan-era U1/U2/U3 headings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
type=recent orders the caller's albums by their latest play (skips
included); type=frequent orders them by completed plays (skips
excluded). Plays of tracks whose file is gone do not count, and both
page by size/offset like the other types.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An admin asking for another username got their own name and roles back.
The handler now looks the user up by exact username and answers with
that user's payload; an unknown username is ErrDataNotFound. A
non-admin asking about someone else is still ErrNotAuthorized.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The marker list behind the earlier server and web passes did not catch
"for v1", "Task 5", "Phase 2" or "today: Flutter". Those comments now say
what the code does.
Corrected: coverart said "both v1 providers" ship on. There are four
providers (MBCAA, TheAudioDB, Deezer, Last.fm), and Last.fm, which needs a
key, ships off.
Non-comment tokens unchanged against HEAD: 10 Go files, 7 TS/Svelte files.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A wider pass for "previous", "before the fix", "we shipped" and the
like. The chi double-subtree, uppercase Vorbis key and stopPropagation
comments now describe the trap they guard; commit hashes, CI run
numbers and a deleted file name are gone; uuidString's comment says
what it returns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Applies rule 32's migration exception: each header still says what the
schema was, what changes and what it prevents, but task, milestone,
decision and rule ids are replaced by the practice they stood for, the
operator quotes by the decision, single-install counts are dropped, and
pointers to a deleted package and an external design doc are gone.
golang-migrate keeps no checksum, so editing applied migrations'
comments is safe.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rewrites the internal/db/queries comments to rule 32 and regenerates
dbq, whose doc comments and embedded SQL carry them. The bugs behind the
tiered mixes, the md5 order seed, the score bands and the round-robin
arms are restated as the constraint each one enforces; counts from one
library and stale cross-references are gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rewrites the internal/library comment blocks to rule 32. Task, milestone,
decision and rule ids are gone; the incidents behind them are restated as
the constraint they left (why a file is deleted before its row, why the
missing mark is a mark, why duplicates block on duration). Counts and
album names from one library are dropped from production comments; test
comments keep the names their fixtures use. The tagReadVersion log keeps
each version's before and after, without the measurements.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rewrites the internal/api comment blocks to rule 32: task, milestone and
rule ids gone, the stories behind them restated as the constraint they
left, stale references corrected (the stream secret is wired from config,
the timezone handler already refreshes the scheduler, the Android app is
the native client, the design docs cited are not in the repo). Also
rewraps comment lines over 80 columns left by the earlier batches.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The same rewrite as d191a282, for four packages whose comments carried the
most history: tuning stories, quotes from the operator, single-install
measurements in library_scale.go, and 'used to' accounts of bugs. The
constraint each story protected stays as a present-tense reason (why tier 3
is reduced not removed, why the randomised arms take an order seed, why the
seed artist is eligible in Songs-like). Comments naming system_cron.go and
config.RecommendationConfig, neither of which holds what they claimed, now
name the scheduler and internal/recsettings.
Checked: non-comment tokens identical to HEAD; gofmt clean.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Comment blocks in cmd/minstrel and 30 smaller internal packages, rewritten
to rule 32's standard with each citation read in context: task, milestone
and rule ids go, and so do the stories they were attached to ("used to",
"Before #N", measurements from one install, quotes, pointers to Fabled
Librarian's records and to docs/superpowers files that are not in the
repo). Where a story wrapped a real constraint, the constraint stays as a
present-tense fact.
Comments that had gone stale are corrected:
- taste.Config: the tuned values are overlaid by internal/recsettings;
it is no longer "a later follow-up".
- StorageConfig: the data dir holds art, caches and the stream secret,
not only playlist covers.
- the Subsonic album list and getUser comments describe what they do
today rather than a milestone that would change it.
Checked: the non-comment token stream of every changed file matches
HEAD, and gofmt -l is empty.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
http.Server.Shutdown waits for in-flight requests but never cancels them,
and an SSE stream only ends when its client leaves, so any open tab or
signed-in phone held every redeploy for the full 10s and the old task
exited 1. With the swarm's stop-first update that was ~10s of downtime per
deploy. The bus now closes on shutdown, ending each stream; clients
reconnect to the next instance.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
seedUser stores users as test-<name>, so the assertion compares against
alice.Username rather than the literal "alice".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
/admin/requests showed "by user f7e1ec86" — the first eight characters of
the requester's id, a placeholder from the first build. The admin list now
sends each requester's username, looked up for the whole page in one query,
and the row reads "by <username>". The id stays as the fallback for a list
that comes back without names.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The duplicates review offered only Merge for a wrong-slot group. A merge keeps
one copy's title whether or not it is the audio's, and neither sets the release
aside nor wants the missing song back.
For class mismatch the page now asks which song the recording is. "Get it
again" (the operator can't tell) removes every copy; "Keep <title>" keeps that
copy and removes the ones under other titles. Either way it runs the resolver's
own path: the release the files came from is set aside, the removed copies'
tracks stay marked missing and emptied, and the album is wanted again.
- library.SettleWrongSlot, POST /api/admin/library/duplicates/{id}/refetch
- ListDuplicateGroupsForResolve takes an optional group_id
- removeWrongSlots takes the acting operator (manual: not automatic, actor in
the audit row) and removes files through removeTrackFileOnDisk, so a
read-only library answers library_not_writable
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured on the operator's library (26,005 tracks, pool scale 2.28): the
arms the library scale grows all have far more rows than it asks for, and
two arms collapsed whatever their limits were.
- taste_overlap ordered by weight then track id, so the strongest taste
artist's catalogue filled it: 20 tracks of saib. (197 tracks) out of 174
taste artists, the same lowest-id tracks every time. It now takes one
track per taste artist, strongest first, md5-seeded, as similar_artists
does since #5297.
- tag_overlap broke ties in no order. A one-tag seed ties 7,804 tracks,
and the LIMIT took whichever Postgres reached first, which the order
seed did not control. Ties now follow md5(id || seed).
- library_scale.go records the per-arm measurement beside the reference
size it leaves as is.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Black Keys' Ohio Players was grabbed 40 times in a day. FabledExchange
already held the torrent under another category from before Minstrel took
the downloads over: the add answered "duplicate", which counted as added,
and Exchange's listing, filtered to Minstrel's category, never showed it.
Half an hour later the grab ended gone, and the next pass took it again.
- A duplicate add that the client does not list for Minstrel is not a
grab: the release is held back for a day with the client's own reason
(already downloading X (category)), and the search moves on. Forced,
it is refused with that reason.
- A release that ends gone a second time within a week is held back for
a day, whatever the cause.
- release_blocks gains the reason 'gone' (migration 0098), shown as
"Held back: its download client lost track of it".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The image's fpcalc 1.5.1 exits 3 with "Error decoding audio frame (End
of file)" whenever it reaches the end of a file before -length, at any
length, so every track under two minutes failed both the AcoustID lookup
print and the duplicate-detection print: 471 lookups and all 1,979
present short tracks' chromaprints.
- runFingerprintTool returns the tool's stdout beside a toolExitError.
- fpcalcReadToEnd keeps a partial exit whose stderr is the end of file
and whose declared DURATION is within the span asked for. A damaged
file that stops early declares more than it read, and still fails.
- Migration 0097 clears the stored failures so the lookup worker and
the fingerprint backfill take those tracks again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
LB's similarity scores are counts (16 to 1445 for recordings, to 10,875
for artists), and they reached a scorer that reads similarity as 0 to 1.
One edge added 64 to 5,800 points under Songs-like, so likes, skips and
recency stopped mattering for any candidate with an edge.
- ResolveListenBrainzTrackEdges divides each edge by the seed's best
score in its whole cached answer. The cache keeps the raw answer.
- The worker stores matched and unmatched artist rows over the seed's
top artist.
- Migration 0096 rescales the rows already stored.
- The candidate arms hold sim_score bands so tier 1 stays above tier 2:
lb_similar 0.5 + 0.5 x edge, tag_overlap halved, likes_overlap 0.3.
Otherwise a seed's weaker LB edges fell below a shared genre tag.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Under e45561e2 a file with no group tag took the oldest same-named album row,
which left three of Weezer's 1994 bonus tracks on the 2016 album. The lookup
now ranks same-named rows by how many tracks they hold from the file's
folder; tagReadVersion 8 re-places the existing ones.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The scanner found an album by artist and title alone, so Weezer's six
self-titled albums (1994–2026) were one 118-track row and their merges,
wants and upgrades counted them as one. A same-named row whose release group
is known and differs from the file's is now another album; tagReadVersion 7
re-reads the library so existing tracks move to their own rows.
WantHoldings compared discs with IS NOT DISTINCT FROM, so a missing track 6
with no disc was never covered by a present disc 1 track 6. Air's Love 2
stayed wanted and was downloaded again for nothing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first album-copy fold pass on 2026-10-20 removed five FLACs (Måneskin
Rush!, Weezer 2016) in favour of mp3s whose only edge was a recording MBID,
and an explicit "Suite-Pee" for its clean version. ProposeSurvivor now ranks
lossless before tag fit; tag fit still outranks file size, which is what the
Humanz rips needed. The fold keys copies by their version words (title and
file name), so a clean, live or instrumental copy holding the same title tag
is linked, not folded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On 2026-10-20 the emptied "Open the Gateway" slot of Auvic's Speed Demons
was folded by title into another file tagged "Open the Gateway" that holds
Onward (244 s against the song's 254 s), so the album was never wanted
back. A wrong-slot removal now stamps tracks.emptied_at (migration 0095)
when the album holds no copy of the song as long as its MusicBrainz slot,
and missing-pair folding pairs such a row by title only with a track added
since: a copy fetched again. A slot whose song the album does hold (VOID
DLC's Weeper beside Void's) folds into it as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A cross-release group was only ever linked, never merged, so an album that
held the song twice among the releases kept both: Humanz has Saturnz Barz,
Andromeda, We Got the Power and Garage Palace as its FLAC and a YouTube mp3,
each beside the single. The resolver now merges an album's extra copies
into its best (the one that is not a video rip, else ProposeSurvivor:
lossless first) with ReplaceTrack, history and all, then links what is
left as one song.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The importer wrote track=N/M for every format, which ffmpeg stores in a
Vorbis comment as TRACKNUMBER=3/20; dhowden/tag reads a Vorbis TRACKNUMBER
with a bare Atoi, so it came back 0. All 245 tracks imported since
2026-10-10 (Humanz, Plastic Beach, the Gorillaz super deluxe and twelve
more albums) have no track or disc number, and their album pages have no
order.
- The importer writes a Vorbis track and disc number alone, with
TRACKTOTAL and DISCTOTAL, as Picard does. ID3 and MP4 keep N/M.
- ReadTrackAndDisc takes the number before a slash when the Vorbis reader
returns 0, for the scanner and for the import's own check of a download.
- tagReadVersion 6 re-reads the files already held (tag reads only).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At boot the scan lists the library, the resolver removes wrong-slot copies
and marks their tracks missing, then the scan reconciles from its list and
marks them present: 7 of the 13 removed on 2026-10-20 15:00 read as
present with no file, and would not be wanted back until a later scan. A
pass now waits while a scan younger than StuckScanThreshold is in flight
and tries again a minute later; a stuck run does not hold it up.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first pass on the operator's library held 16 groups. In several, every
copy's own lookup names one recording (Wombats Record Shop/Tormentor,
Alchemist (2)/Prelude, Monday Chill 8/9, 404sessions), but one copy sits 3-10s
from its slot, so the every-copy-fits precondition kept AcoustID out. Now:
copies whose lookups all agree are that recording, whatever their
fingerprints read; copies of the same audio are asked about as before; and
the track list places the recording in one copy's slot, or none. Copies
filed as one song are not settled here.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Savant's Mortals held one recording as 16 of its 50 tracks, each trimmed to
its own slot's MusicBrainz length, so the lengths could not say which slot
was its own and every group waited for the operator. When a one-album
mismatch is the same audio (identical, or fingerprints within 0.02) and
every copy fits its own slot, the resolver asks what the recording is: the
copies' own AcoustID lookups, or a fresh lookup of one copy (five a pass).
The track list names each slot's recording; the copy in that slot is kept
and the others removed, or every copy when the recording is in none of
their slots or AcoustID knows none. AcoustID unreachable or answers in
several copies' slots: the group waits.
Whenever wrong-slot copies are removed, the download the files were
imported from is set aside, so the album fetched again is not the same
files. Albums imported before Minstrel downloaded anything have none.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each group now states its evidence: identical audio data, or the share of
audio fingerprints that agree. Copies titled as different songs in one
acoustic group say so and ask the reviewer to listen, and each copy has a
play button so they can.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The release id was only ever used to find the group. 238 albums matched to a
group carried none, so their losses, five of the resolver's wrong-slot
removals among them, were never fetched again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On 2026-10-20 twelve of the 24 slots the resolver emptied were folded by MBID
into the kept copy, so their albums were never wanted back.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Shnabubula's whole album, one 45-minute file, was filed as both Fading Light
and Looking for Answers. Neither slot fits, so wrongSlots had nothing to keep
and held the group. Every copy now goes when each is clearly another song's;
a copy near its slot still holds the group.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
20 of the 34 groups left for review had no cached track list for their album:
the catalogue only reads one when the wanted list or the importer needs it.
The resolver now asks catalogue.RefreshReleaseGroup for an album it must
decide and the catalogue has never read, through the shared MusicBrainz pacer,
at most five a pass (hourly). An unreachable MusicBrainz stops the pass's
fetching; a group read and found empty is not asked about again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- songKey drops a dashed version suffix (" - Instrumental Version", " - Radio
Edit", live, remix, mix, acoustic, demo, reprise, rework, VIP), so Andora's
"Goodbye Yesterday - Instrumental Version" is distinct, not a wrong slot.
- wrongSlots accepts several fitting copies when they share one title
(Killjoy twice beside Dreamland): the other slot's copy is removed, and the
doubled song is left for the next sweep to find as a pair the resolver merges.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The operator's library had 536 groups waiting for review, 2026-10-19. A sample
showed most were one recording under two titles on two releases.
- The headline and the Duplicates count leave out cross-release groups, which
the resolver keeps and links: the headline read 3,479 when 536 needed review.
- Any group spanning albums is cross_release and is linked as one song,
whatever its titles say. Nothing is deleted.
- One album, two versions named as such ("Lovesick (Instrumental)") and not
byte-identical: a new class, distinct, dismissed by the resolver.
- One album, one recording under different song titles (mismatch): the
album's cached MusicBrainz track list decides. When exactly one slot's song
is as long as the file and every other slot's song is clearly not, the other
copies are removed and their tracks marked missing, so the reacquisition
sweeper wants the album back. Otherwise the group stays for review.
- The duplicates_found notice moves from the sweep to the resolver: it names
only groups a pass first left for review.
- Resolved automatically lists dismissals and wrong-slot removals beside
merges.
Migration 0094 adds 'distinct' to the class CHECK (rule 36).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- An update to a grouped notice keeps its arrival time and email state, so the
phone and the email no longer go off again each time it grows.
- Notices carry the names of what changed (newest ten), shown as three and
"and N more": tracks still missing after the scan's move/replacement fold,
duplicate groups new to a sweep, songs merged automatically, and the track
and listener a playback error was for.
- Duplicates found counts only the groups new to that sweep; playback errors
count reports, not the unresolved total.
- A scan failure says what stopped and links to Admin > Library.
- Android: a pile of more than three alerts lists the newest five titles.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Migration 0091 drops lidarr_config and renames the quarantine tables, types,
indexes and the history column off the Lidarr name. The Integrations Lidarr
card, the Reports "Delete via Lidarr" action, the merge and remove-track
unmonitor options, the Lidarr audit actions and error codes go, on the
server, web and Android. The README describes acquisition as it is now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
