resizedItem.Reader closes what open() hands back, so an open() that reports "no image"
as (nil, nil) rather than an error takes the request down with a nil-pointer panic. Every
caller returns an error today, and no test covered it: the resolution e2e harness stubs
the resize reader out entirely, so no e2e path reaches this code at all.
Guard it and cover Reader directly.
resolveDisc duplicated the loop selectImageReader already ran: try each source in
priority order, take the first that yields an image. The serving path and the CLI
diverged on two details as a result — only selectImageReader checked ctx between
candidates and logged each attempt.
Both now call discArtworkReader.selectImage, which takes the chainState the CLI already
uses for the other kinds. The serving path passes an untraced one, whose nil trace makes
recording a no-op. selectImageReader had no other caller and is gone.
The disc tests move from fromDiscArtPriority to discCandidates, so they assert the skip
reason for an entry that maps to no source rather than that it silently vanished, and
cancellation mid-walk is now covered.
`artwork explain` rejected `dc` and `mf` because it validated against RecheckKinds,
the list of kinds the backfill revisits. Those are different questions: a kind with no
recheck path still has artwork someone can report as wrong.
Disc artwork now walks DiscArtPriority under a trace, so explain reports which entry won
and why the others lost, including entries that map to no source at all (external is
unsupported, a disc with no subtitle, an album folder with no images). Media file artwork
traces its single embedded candidate, separating "EnableMediaFileCoverArt is off" from
"the track has no embedded art" — stored state cannot tell those apart.
Each command now validates against the kinds it can actually serve: explain takes all six,
refresh takes artwork.RefreshableKinds (which nativeapi now shares instead of keeping its
own copy), reprocess still takes RecheckKinds. Disc artwork stays out of refresh: the
worker cannot resolve it, so the queue row would be rejected on every drain.
WalksPriorityChain becomes Explainable, and ResolveArtist/ResolveAlbum collapse into
Resolve(kind, id).
validateSources checks the requested sources against the ones item_artwork
actually uses, to catch a typo. The reserved empty source (spelled 'absent' on
the CLI) is a valid filter even when it matches nothing, so a scheduled
'artwork reprocess --source absent --yes' stopped working the moment the
library finished resolving. Treat it as intrinsically valid and let the
existing zero-match path report it.
explainResult qualified the verdict whenever an external OutcomeError
appeared before the winning hit. When a later external agent returns an
image, fetchArtistImage/fetchAlbumImage discard the earlier error, so
extError is false: the worker settles the item and schedules no retry.
Telling the operator it may resolve differently on a retry was wrong.
The warning is only correct when a lower-priority local source won while
an external error was recorded, which is the case that carries extError.
The Agents: line printed the raw config while the Chain only showed the agents the CLI could
construct, with nothing explaining the gap: plugin agents are never registered in a CLI that does
not start the plugin manager, and a built-in without credentials returns nil. Three of five agents
could vanish, including ones ranked above the one shown.
Also treat a live external error before the winning hit like the already-handled would-try case:
the resolver serves such a hit provisionally and retries later, so the verdict is indeterminate.
The Result line is still not qualified when an unavailable agent might have won; that needs agent
ranking, and is left to the follow-up that makes the CLI load plugin agents for real.
Neither bound is true. A ceiling is false because plugin agents are invisible to
a CLI that never starts the plugin manager, and a floor is false because a local
hit ends the walk before any agent is asked and a hit on the first agent skips
the rest. "at least N" traded one wrong claim for another.
The line now names its blind spots instead:
External lookups: ~340 estimated (plugin agents not counted; local hits may
need fewer).
The same line is reused in the confirmation prompt, and the zero case still
reads "External lookups: none." with the prompt dropping the clause entirely.
The count itself is unchanged.
The reprocess prompt billed one external lookup per externally-capable item.
fetchArtistImage/fetchAlbumImage try every enabled image agent and stop early
only on a hit, and resolvePlaylist can fetch the m3u image and then resolve up
to four sampled albums for the grid, each walking the album agents again. The
number the operator confirmed could understate real provider traffic several
fold, in the prompt whose whole job is to stop a provider flood.
ExternalLookupsPerItem now multiplies by the visible image-agent count and adds
the playlist grid factor. It stays a floor: the CLI never calls Manager.Start(),
so the plugin registry is empty and plugin-provided agents are dropped by
getEnabledAgentNames. On an install with 5 agents of which 3 are plugins the
count is well under the truth, so the wording is now "at least N" rather than
"up to N" — a zero visible count still bills one lookup for the same reason.
Fixing the plugin visibility is out of scope: Manager.Start() needs a Subsonic
router and writes to the DB via syncPlugins, breaking this command group's
read-only guarantee.
The estimate used WalksPriorityChain, which is true only for artist and album,
so a playlist-only reprocess reported "External lookups: none" and the
confirmation prompt dropped the external-cost warning. Playlists do reach the
network: through the m3u ExternalImageURL fetch when EnableM3UExternalAlbumArt
is on, and — verified by test — through the generated grid, whose tiles resolve
album art via the full album priority chain.
Adds artwork.MayFetchExternal, a config-aware predicate for "can this kind's
resolver reach the network", and uses it for the estimate. WalksPriorityChain
keeps its separate job of deciding whether explain prints a chain block.
One seeded hash and source together and then asserted the two counts agree,
so its setup guaranteed the result; the other repeated the count-does-not-
enqueue property already covered by the CountBySource spec.
Reuse artwork.RecheckKinds and artwork.WalksPriorityChain, extract
newTabWriter and externalEstimate, fold reprocessSelectsAll into
selectedKinds, and derive the queue total and the walks-chain flag instead of
carrying them in the report structs.
Export RecheckKinds and add WalksPriorityChain so the CLI stops keeping its
own copies of both, and unexport externalCandidate, which nothing outside the
package consumes.
`artwork explain` rendered the stored empty source as "(absent)", while
`artwork reprocess --source` only accepts "absent", so pasting what explain
printed straight back into reprocess was rejected as an unknown source.
ChainTrace and TraceStep were unexported types re-exported through aliases,
which existed only so the CLI had a name to refer to them by. The types are
public API — Resolver.Steps returns []TraceStep and the CLI constructs a
ChainTrace — so name them that way and drop the indirection.
Encapsulation is unchanged: add, mu and steps stay unexported, so only this
package can write a step.
A stale stored fingerprint with items already queued is the worst state the
system can be in: a second full re-enqueue is pending on top of the one running.
The line carried the weakest wording of the three, and was untested.
The CLI hardcoded the outcome literals and the "external:" prefix, so renaming a
constant's value in core/artwork left cmd compiling and the suite green while
`artwork explain` silently degraded its verdict.
Renaming a value now fails the golden vocabulary test in core/artwork and the
explainResult tests in cmd.
By the time anyone runs a diagnostic, backfill has usually already stored the new
fingerprint, so 'up to date' was printed while thousands of items churned through external
providers. The backlog is the finding; the fingerprint is context.
Also echoes the config inputs the fingerprint covers, so a change can be traced to the
setting that caused it, and pins the rendered rows: the Absent values, the queue TOTAL and
a queue-scoped kind/priority pair were all unasserted, so kindName and priorityName were
effectively untested. FingerprintInputs is now the single listing ConfigFingerprint hashes;
a pinned hash proves the value did not change.
Reports the queue, where artwork currently resolves from, absent counts against the 24h
recheck window, and the stored config fingerprint versus the current one — the line that
turns 'why is my server re-resolving everything?' into one command.
fingerprint() and staleAbsentAge are exported so the CLI reports the values backfill
itself compares, instead of a second copy of the formula that can silently drift.
Mutating the --yes check to `if true` left the suite green, so the one bypass
of the confirmation was unverified. The choice is now reprocessConfirm(yes, in),
covered in both directions.
The external estimate only reached the operator through the prompt, which
--dry-run skips — hiding the number in the one mode that exists to show it
before committing. The preview now carries it, and the prompt drops the clause
when no lookup will be made.
An empty selection says so again under --dry-run.
The reconciliation that makes --source alone target every kind was only
exercised through runReprocess, which no test calls: mutating it to
`all := reprocessAll` left the suite green. It is now reprocessSelectsAll,
covered for all three selectors.
Scoping source validation to the selected kinds made the same well-formed
filter valid or invalid depending on which other kinds were selected, and its
error read the same for a typo as for a source that simply does not apply to
the chosen kind. Validation is now table-wide: a typo still aborts, while a
valid-but-inapplicable source falls through to "Nothing matches".
Also: the prompt now counts only the kinds that reach an external agent as
external cost, and --dry-run on an empty selection reports a dry run.
Bulk re-enqueues artwork by kind and/or by the source an item currently
resolves from, previewing the matched count and confirming before queueing.
The preview counts with CountBySource (rows matched) and reports separately
what EnqueueBySource inserted: its DO NOTHING conflict policy leaves an
already-queued row untouched, so the two numbers differ and the output must
not claim the skipped rows were re-queued.
An unknown --source is rejected against the sources present in item_artwork,
rather than silently matching nothing and printing a reassuring 0.
A transient external failure traced as 'error' fell through to 'not resolved', which
is the most common state behind a missing-artwork report. It is now indeterminate, and
an offline win that a skipped higher-priority external candidate could have taken says
so instead of naming a winner the live chain might not pick.
A configured 'external' token vanished from the chain when no enabled agent provided
images for that entity type, and for synthetic artists, leaving the trace unable to
say whether the tier was even considered.
Prints why an item has the artwork it has: the stored state, its queue row, the
governing config, the resolver's priority-chain walk and the verdict. Offline by
default so a diagnostic run cannot add load to an external provider; --live asks
the agents for real. Playlists and radios do not walk a priority chain, so they
report that instead of an empty chain table.
The explain CLI must report whether an item is queued, at what priority and when it
retries; the queue repository could only be drained in eligibility batches, which
cannot see a row that is still backing off.
* feat(insights): report the app store or hosting platform via ND_PLATFORM
Insights had no way to tell where an instance is deployed. The existing `os.package` field is written only by our own packagers and holds just `deb`, `rpm` or `msi`, so it answers "which installer", not "which platform". Overloading it would mix two unrelated dimensions in the same field.
This adds a separate top-level `platform` field, self-declared by the deployer through the `ND_PLATFORM` environment variable. App stores and hosting providers (ZimaOS, PikaPods, TrueNAS, Unraid, and others) generally deploy our container image unmodified and can only inject environment variables, so an env var is the one marker they can all set. It is deliberately not a config option: it is a packager marker, not something users should tune, and it stays out of the config surface.
Both values are now whitespace-trimmed. The msi packager writes the file with `echo`, so `os.package` has been arriving as `"msi\n"` and sorting separately from `"msi"` in any aggregation.
* test(insights): isolate hostingPlatform specs from an inherited ND_PLATFORM
The spec asserting an empty result read the real environment, so it failed on any machine that already had ND_PLATFORM set. Unset it per-spec, using Setenv first so Ginkgo restores the original value on cleanup.
* fix(artwork): serve images whose format has no registered decoder
The new pipeline derives dimensions, mime and the placeholder hashes at
resolution time, so a decode became a precondition for recording artwork at
all. An image in a format Go has no decoder for therefore failed acquisition,
retried until the 12h budget ran out, and then settled as absent, serving a
placeholder from that point on. The old pipeline decoded only to resize and
fell back to the original bytes when that failed, so these covers used to work.
A local file is picked by matching an image extension, so bytes it cannot
decode are most likely a codec we lack: image.ErrFormat on a folder, upload or
embedded source now yields an Artwork row carrying just the hash and mime, and
the bytes stay servable. An external response carries no such guarantee, so it
still fails and retries rather than pinning a non-image body as a cover. A
corrupt image of a known format and an over-cap declared size still fail, so
the decompression bomb guard is unchanged. Absent rows recorded by earlier
builds are re-resolved by the existing stale-absent recheck within a day, so
no epoch bump is needed to repair them.
Reusing a stored image now re-decodes when it carries no dimensions, so a
row recorded while a decoder was missing can still be upgraded later.
Registers jxl, heic and heif in mime_types.yaml: image detection resolves the
extension through the host MIME table, and the Alpine release image ships no
/etc/mime.types, so those covers were never recorded in folder.ImageFiles
there and never reached the pipeline at all.
* fix(artwork): never record empty bytes as artwork
image.DecodeConfig returns image.ErrFormat for an empty payload just as it
does for a codec with no registered decoder, so a zero-byte cover file was
recorded as found artwork and served as an empty response instead of falling
back to the placeholder. A truncated image of a known format already fails
with unexpected EOF rather than ErrFormat, so only the empty case needed the
guard.
Finamp's connection test GETs /System/Endpoint and treats anything other
than a 200 carrying an IsInNetwork key as "not a Jellyfin server", so the
test failed against Navidrome and dual-connection setups could never
switch to the local address.
IsInNetwork mirrors Jellyfin's default LAN set (NetworkManager with no
LocalNetworkSubnets configured): loopback, the RFC 1918 ranges, fc00::/7
and fe80::/10. Notably that set omits 169.254.0.0/16, so Go's
IsLinkLocalUnicast is deliberately restricted to its IPv6 half.
IsLocal mirrors HttpContext.IsLocal(): the caller shares the connection's
local address, not merely "is loopback". It falls back to the loopback
check when the local address is unavailable.
* fix(plugins): stop reporting plugin call failures as not-found
MetadataAgent joined agents.ErrNotFound onto every failed plugin call, so a
transport fault was indistinguishable from a definitive miss. The artwork
circuit breaker treats a not-found as a successful, definitive answer and
resets its failure counter, so it never opened for a failing plugin and kept
calling it on every request. Observed with the apple-music plugin against
prod: ~900 iTunes 429s in 27 minutes with the breaker never tripping.
Return the underlying error instead. The genuine empty-result branches still
return agents.ErrNotFound, and agent fallback is unaffected because
callAgentMethod/callAgentSliceMethod continue on any error, not only on
ErrNotFound.
* test(plugins): fold duplicate metadata agent error specs into one table
The error-handling container drove all 11 MetadataAgent methods twice: once
to assert the message, once to assert the failure is not an ErrNotFound. The
argument lists were identical, so each method cost two WASM instantiations for
one method's worth of coverage, and a new capability had to be registered in
two places to stay guarded.
Fold both assertions into a single DescribeTable, document the ErrNotFound
contract at the sentinel where agent implementers will read it, and collapse
breaker.record's hand-inlined predicate onto isTransientExternal, which it
already duplicated by hand with a keep-in-sync comment.
* fix(plugins): keep an unimplemented plugin method a definitive miss
Returning the raw plugin error made errNotImplemented and errFunctionNotFound
look like provider faults. Every MetadataAgent satisfies ArtistImageRetriever
and AlbumImageRetriever regardless of what the plugin actually exports, so
artwork resolution calls those stubs on a partially-implemented plugin: each
call counted toward the artwork circuit breaker and kept the item in the retry
queue instead of settling it absent.
Map both sentinels back onto agents.ErrNotFound, joined so the underlying
reason survives for diagnostics, and leave real call failures untouched. This
matches what ScrobblerPlugin already does for the same two sentinels.
The partial-implementation specs asserted only MatchError(errNotImplemented),
which the previous errors.Join satisfied incidentally, so nothing caught the
lost not-found semantics. They now assert both and are folded into one table.
* test(plugins): cover the missing-export arm of agentErr
The partial-metadata-agent fixture registers through the Go PDK, which exports
every method and answers with the not-implemented code, so no fixture reaches
the errFunctionNotFound branch. Building one would mean hand-writing Extism
exports to deliberately omit a function, which tests the manager's function
lookup rather than the mapping this PR added.
Cover agentErr directly instead: both sentinels classify as a definitive miss,
a call failure and a non-zero exit stay faults, and the underlying reason
survives in every case.
* fix(artwork): stop counting a cancelled run against the circuit breaker
callPluginFunction returns ctx.Err() when a plugin call is cancelled, and that
reached breaker.record as an ordinary error, so cancellations counted toward
the five consecutive failures that open a gate. A cancellation says nothing
about the provider, so it now neither counts nor clears the failure run.
Deliberately scoped to breaker.record rather than isTransientExternal: the
latter also drives whether the queue item is rescheduled, and a cancelled item
must still be retried rather than settling absent. context.DeadlineExceeded is
left counting as a fault, since a provider that blows the budget is one worth
backing off from.
Reachable today only at shutdown, where the in-memory breaker state is
discarded anyway. It becomes live the moment Worker.gate is used on a
request-scoped context, which is why it is worth closing now.
* fix(instant-mix): top short mixes up instead of returning what the first source found
SimilarSongs returned the first non-empty source's tracks, however few. For a
thinly-represented artist that meant a 3-track mix no matter the requested count:
the artist agent found nothing, the similar-artists fallback matched 3 library
tracks, and `len(res) > 0` kept seed-track sampling from ever running.
Clients treat that as a failed mix and retry with a bigger limit forever. Finamp
cycles limit 34 through 472 and starts over, ~1 request every 2s indefinitely,
each one re-hitting Last.fm, Deezer and AudioMuse.
Sources are now chained rather than raced: each one tops the mix up until it
holds count tracks, so the agent's picks, the similar-artists fallback and
seed-track sampling all contribute instead of the first one winning outright.
* refactor(external): move similar-songs code to its own file
provider.go held two distinct concerns: artist/album external metadata and the
similar-songs mix pipeline. The mix code was already one contiguous block, and
maxSeeds, maxSimilarSongs and dedupByID were used by nothing else.
Moved SimilarSongs and its helpers to provider_similarsongs.go, matching the
existing provider_similarsongs_test.go. Pure code motion: the moved block is
byte-for-byte unchanged and provider.go has no additions, only deletions.
* fix(instant-mix): dedup before deciding a mix is full
topUp measured res before deduplicating it. Matcher.MatchSongs deliberately
re-emits a library track when the same input song repeats, and the similar-artists
fallback can reach one track through several artists, so len(res) could equal count
while holding fewer unique tracks. That returned a mix with duplicates in it and
stopped the top-up early; the caller then deduplicated and handed back a short mix,
which is the client retry loop this branch set out to fix.
Deduplicate first, so the length check counts what the client will actually receive.
* perf(instant-mix): skip a fallback once the mix is already full
The artist path nested one topUp inside another, so the inner one measured only
similarSongsFallback's own result against the full count. With 49 agent matches and
one fallback match for count=50 the mix was already full, yet seed-track sampling
still ran and fired up to five GetSimilarSongsByTrack calls whose results the outer
topUp then truncated away.
topUp now takes the sources as a variadic list and re-checks the accumulated mix
before each one, so a later, costlier source only runs while the mix is still short.
That also flattens the artist case: the agent, the similar-artists fallback and
seed-track sampling are now three peers in one chain instead of two nested calls.
* fix(instant-mix): count distinct tracks when picking the fallback mix
similarSongsFallback stopped after count picks from the weighted chooser, but a
track can sit in that chooser once per artist listing it in their top songs, and
Pick removes the entry it returns. Repeats therefore consumed pick slots and left
unique candidates stranded, so the batch could come back short of count. On the
track path this is the only source, so that short mix reached the client and kept
the retry loop alive.
Track the ids already picked and keep drawing until count distinct tracks are held
or the chooser is empty.
* fix(instant-mix): match the whole agent response before trimming
MatchSongs was capped at count, and it re-emits a track when the same song repeats,
so [A, A, B] with count=2 returned [A, A] and never reached B. topUp then shrank
that to [A] and, with an empty or overlapping fallback, the mix stayed short even
though B had been available all along. seedMix already matched its full merged set
for this reason; mixFromAgent now does the same and leaves the trim to topUp.
Also drop the capacity hint on the picked-ids map. It was sized from the caller's
count, which CodeQL flags as an allocation sized by user input (go/uncontrolled-
allocation-size). SimilarSongs clamps count to maxSimilarSongs long before this
point, so the hint bought nothing worth the alert.
* refactor(instant-mix): tidy the mix chain and its specs
Quality pass over the new code, no behaviour change:
- topUp: drop the first-vs-last error bookkeeping (the value is only read when the
mix is empty, so the distinction is unobservable) and the redundant nil guard
(dedupByID returns nil for an empty result, so both branches already agreed).
- mixFromAgent: assign through the if-scoped err instead of a second error name.
- Hoist the similar-artists fallback closure written verbatim in two switch arms.
- Use map[string]struct{} in the pick loop, matching dedupByID in the same file.
- Trim three comments back within budget; two restated the line below them and one
carried commit-message rationale.
- Tests: add an ids() helper for the ID assertion repeated seven times, and fold
the track-entity stub block copied into three specs into stubTrackEntity. The
block hard-coded .Twice() on GetEntityByID, which pinned an implementation
detail no spec asserts.
* revert(instant-mix): inline the similar-artists fallback closure again
Hoisting it to a shared artistFallback var moved the call away from the arm that
uses it and saved nothing: each arm reads better spelling out its own sources.
* feat(agents): local agent genre-hint similar songs fallback
* feat(external): playlist instant mix via seed-track sampling
* test(external): cover playlist mix never-empty fallback and maxSeeds cap
Adds coverage for the empty-match seed fallback and the maxSeeds
call cap on GetSimilarSongsByTrack, per code review finding.
* feat(external): genre instant mix via seed-track sampling
* feat(external): album instant mix falls back to AudioMuse track similarity
* feat(external): artist instant mix falls back to seed-track sampling
* fix(jellyfin): route genre seeds through instant mix instead of empty
* feat(jellyfin): add /Albums/{id}/Similar route for albumMix radio
* perf(external): bound playlist seed sampling to a random N
samplePlaylistTracks loaded an entire playlist's joined rows just to keep
5 random seeds; push the bound and randomization into the query instead,
matching the other samplers (GetRandom/GetAllByTags with Max).
Fixing this surfaced a real bug: resetSeededRandom's SEEDEDRAND rewrite
assumed every table's id is TEXT, but playlist_tracks.id is an INTEGER
position, so the random sort silently dropped every row. Cast the id to
TEXT before hashing (no-op for the other, TEXT-id tables).
Also trims a changelog-flavored comment and a duplicated rationale in
server/jellyfin/similar_test.go.
* refactor(external): parallelize seed mix and dedup mix helpers
Run the up-to-5 per-seed GetSimilarSongsByTrack calls concurrently (errgroup),
route the four container cases through a shared seedMix helper, flatten the
genre lookup, and sample playlist seeds without forcing a smart-playlist
rebuild. Share the media-file->Song mapping in the local agent.
* perf(agents): use the indexed genre filter for local similarity
Replace GetAllByTags (a json_tree scan of every media_file row) with the
media_file_tags semi-join from #5940, deriving the seed's genre tag ids
locally since they hash from (name, value).
Also carry the library id and the recording MBID on the returned songs:
the matcher resolves by id first and looks up mbz_recording_id, so the
release-track id it got before matched nothing and the local fallback
silently returned no songs.
* refactor: drop redundant MBID and fold mixFromSeeds into seedMix
The local agent returns library tracks, so the id alone resolves them in the
matcher's first phase; the MBID was never consulted. mixFromSeeds had no
caller other than seedMix.
* docs: trim redundant comments
* fix(jellyfin): adopt the GUID id codec in the merged similar routes
getSimilarAlbums still used resolveItemID/DecodeID, which #5942 replaced with
itemIDParam; its tests passed raw ids that the strict codec now rejects.
* fix(external): guard non-positive counts and blend every seed
A negative Subsonic count reached matched[:count] and panicked. The matcher
also keeps input order and stops at count, so seed-grouped results let the
first seed fill the whole mix; interleaving gives every seed a share.
Drops the duplicate playlist-track mock in favour of tests.MockPlaylistTrackRepo,
which pages like the real repository and records the query options.
* fix(external): refresh smart playlists before sampling seeds
A smart playlist materializes no playlist_tracks until it is evaluated, so
sampling without the refresh mixed an empty seed set. The refresh is a no-op
for regular playlists, inside the refresh delay, and for non-owners.
* fix(external): skip missing tracks and a nil playlist-track repo when sampling
Tracks() logs and returns a nil repository when its own lookup fails, so the
chained GetAll panicked. Seeds can also reach the mix verbatim when the agents
find nothing, so a missing file would surface as an unplayable entry.
* fix(jellyfin): never report the seed album as its own similar album
The sampled-seed fallback returns the album's own tracks, which similarAlbums
mapped straight back to the requested album, often as the only result.
* test(agents): assert the genre predicate instead of relying on the mock
MockMediaFileRepo ignores QueryOptions.Filters, so the spec passed even with
no genre filter at all. It now checks the generated predicate carries the
seed's own tag id, the indexed join and the missing exclusion.
* fix(external): clamp the requested count before it becomes a query limit
Subsonic passes the client's count through unbounded. At MaxInt64 the local
agent's count+1 overflows negative, and GetRandom omits the SQL limit unless
Max is positive, so one request would hydrate every matching track. 500 is
what the widest caller (similarAlbums, limit*5) legitimately asks for.
* fix(external): deduplicate playlist seeds by media file
A playlist can hold the same file at several positions, so sampling its rows
could seed the mix twice: a wasted agent call, and a duplicate track whenever
the seed fallback kicks in.
* fix(external): drop tracks two seeds both recommend
The matcher re-emits a track when two inputs are identical, so overlapping
recommendations took several slots in the mix. Match the whole merged set and
dedup before trimming. Playlist sampling now over-fetches before its own
dedup, so repeated positions cannot collapse the seed count.
* test(external): make the seed-blend assertion independent of the shuffle
It matched four tracks and kept two at random, so both could come from the
first seed once in six runs. Keeping three of the four makes a seed-two track
unavoidable.
* fix(external): seed artist mixes from every credited role
media_file.artist_id is the deprecated primary artist, so an artist credited
only on the album, as on compilations, sampled no seeds at all. Use the same
participant filter the artist listings use.
* refactor(external): drop the now-vestigial seed interleaving
Matching the whole merged set removed the early truncation the interleave
guarded against, and the shuffle before the trim makes input order irrelevant.
Its comment described the old behaviour.
* test(agents): give the id-mapping fixture a matching genre
The related track carried no genre, so the real query would never return it;
the spec only passed because the mock ignores QueryOptions.Filters.
* test(agents): drop the MBID from the id-mapping fixture
Local agent candidates are non-missing library rows, so the matcher always
resolves them in its id phase and never reads the MBID. The field guarded a
regression that could not change behaviour.
* test(agents): remove unnecessary comment about MBID in GetArtistTopSongs test
* fix(jellyfin): only let a not-found entity fall through in getInstantMix
Discarding the error conflated a genre id, which never resolves, with a real
lookup failure, which then made a provider call that fails the same way.
* test: pin the invariants the specs only appeared to cover
The missing filter was asserted by substring, so flipping it to true passed
everywhere, including the spec named for it. Matching the whole merged set,
the local agent's over-fetch, and its no-genres early return had no coverage
at all; each is now pinned by a spec that fails when the code is broken.
* test: make the remaining specs say what they actually guard
The playlist-track spec named a sort whitelist it does not exercise; it guards
the integer-id CAST, so it now asserts no rows are dropped. The maxSeeds cap
passed with either bound removed, and the over-fetch was pinned by its literal
value rather than the duplicate positions it exists for. Also drops setup the
count guard returns before reaching.
* fix(external): fall back when the agent's picks are not in this library
A non-empty answer whose songs are all absent locally matched nothing and was
returned as-is, so the mix came back empty with sampleable source tracks
sitting right there.
* refactor(external): name the agent-then-fallback flow once
Each entity case repeated the same error and emptiness plumbing around the
matcher. mixFromAgent states it once and each case supplies only what differs:
how to ask, and what to do when the answer is unusable.
* test(jellyfin): use canonical ids in dto fixtures
Fixtures used short placeholder strings, which are not valid Navidrome ids. Deriving them from
id.NewHash keeps the labels readable while exercising the real id shape.
* test(jellyfin): use canonical ids in handler fixtures
Fixtures used short placeholder strings, which are not valid Navidrome ids. Deriving them from
id.NewHash keeps the labels readable while exercising the real id shape. Playlist entry positions
stay decimal, matching the integer playlist_tracks.id column.
* test(jellyfin): use canonical ids in e2e fixtures
Fixtures used short placeholder strings, which are not valid Navidrome ids. Deriving them from
id.NewHash keeps the labels readable while exercising the real id shape.
* test(jellyfin): use canonical ids in audiomuse fixtures
audiomuse_test.go passes ids as bare function args (mf(id, ...), call(query, user)) rather than
via ID: struct-literal fields, so the original grep-built file list missed it. Same conversion as
the rest of the fixtures: fake labels through id.NewHash via testID.
* test(jellyfin): convert remaining nonexistent-id sentinels in e2e tests
Reviewer swept for enc("literal") sites the brief's dto.EncodeID grep missed. These "does not
exist" fixtures must stay well-formed GUIDs under the strict codec, or the test degrades from
"resolves to nothing" to "empty path segment".
* refactor(jellyfin): emit real 128-bit GUIDs as item ids
Navidrome ids are now a canonical 22-char base62 encoding of exactly 128 bits, so they map
losslessly onto Jellyfin GUIDs. Previously the API hex-encoded the id string itself, producing
44 hex chars where Jellyfin uses 32.
Integer library ids, the synthetic playlists folder, and playlist entry positions (a
playlist_tracks.id, an integer column) aren't 128-bit values, so they get a reserved GUID space
tagged by kind. DecodeID is now strict: malformed input returns an empty string instead of
passing through unchanged.
BREAKING: Jellyfin clients see entirely new item ids.
* fix(jellyfin): 404 malformed playlist ids instead of silently creating
updatePlaylist decoded a malformed playlistId to "", the same sentinel core/playlists.Create
uses to mean "make a new playlist" — the overload createPlaylist deliberately relies on. A
malformed id now 404s before reaching Create.
Also tightens id-codec test fixtures: several tests set chi params to a raw canonical id, which
now decodes to "" and only passed because the fakes ignore the id argument; and a batch of
not-found sentinels now use well-formed-but-nonexistent GUIDs so they exercise the intended path
instead of the malformed-id path. READMEs "lossless" claim softened to note the reserved space.
* refactor(jellyfin): drop the id truncation workaround
Finamp's saved-queue packing keeps the first 16 bytes of each item id. That was lossy only
because our ids were 44 hex chars; now they are 32, so the packing round-trips exactly and the
server-side prefix recovery is dead code.
Removes an indexed range scan per restored queue and the ambiguous-prefix path that could
resolve to the wrong item.
* fix(jellyfin): emit ServerId and PlaySessionId in Jellyfin's id format
Jellyfin serializes GUIDs without dashes; ServerId was emitting the dashed UUID form. A
ServerId persisted before this change is normalized on read rather than rewritten.
PlaySessionId was emitting a raw internal id instead of the encoded form.
BREAKING: the ServerId change makes clients treat the server as new, so users re-login once.
* fix(jellyfin): 404 on undecodable id filters instead of widening the query
DecodeID collapsed an absent param and an undecodable one into the empty string, and downstream
an empty id means no filter. A client sending a stale pre-upgrade id therefore had its filter
silently dropped: ParentId, ArtistIds and AlbumArtistIds each returned the whole library instead
of a scoped result. Every existing client hits this on first launch after the id format changes.
Scalar id params now distinguish the two cases and report not-found. List-valued params already
failed closed. EncodeID logs a diagnostic when a non-empty id is not canonical, which should not
happen post-migration and would otherwise ship an unaddressable item silently.
* test(jellyfin): drop comments that restate the spec names
* refactor(jellyfin): decode reserved GUIDs from bytes, not hex strings
DecodeID already had the 16 decoded bytes, then re-derived the kind tag and payload by slicing the
hex string and parsing it a second time. Reading them off the byte slice matches how the format is
specified and removes the duplicate parse.
Bounding the payload inside encodeReserved gives both encoders the 32-char guarantee, which only
EncodePlaylistEntryID enforced before.
Playlist entries now decode through DecodePlaylistEntryID, which rejects other kinds. The tag was
being encoded and then discarded, so a song id passed as an EntryId reached RemoveTracks as a
playlist_tracks position.
Drops the per-field log.Warn from EncodeID: it sat in a leaf codec without a ctx and would emit
once per item per request on exactly the bad-data population it was meant to surface.
* refactor(jellyfin): make DecodeID report whether the id was decodable
DecodeID returned the empty string for both an absent param and an undecodable one, and
downstream an empty id means no filter. That conflation is what let a stale id widen /Items to
the whole library; it had been patched at two call sites, leaving three different policies for an
undecodable id in one package and ~16 handlers correct only because a repo Get("") happens to fail.
Returning (string, bool) makes the ambiguity unrepresentable, and the compiler forces each of the
~22 sites to decide. URL params share one itemIDParam helper that 404s; id lists go through
DecodeIDs, which is all-or-nothing because dropping bad entries would empty a list and make its
len() > 0 filter gate vanish — the original bug by another route.
A well-formed but unknown id is still 200 with zero results; only malformed ids 404. Malformed
ids now also 404 on the image and similar/instant-mix routes, which previously answered with a
placeholder or an empty list.
* fix(scanner): accept absolute paths in selective scan --target
The scanner's fs.FS only accepts paths relative to the library root, so an
absolute --target path (e.g. 2:/jukebox/collection) failed with an opaque
"invalid argument" error. Rebase absolute targets onto the library root
before scanning; relative paths are unchanged.
Fixes#5943
* refactor(scanner): simplify libraryRelativePath with IsLocal and slice.ToMap
* fix(scanner): make libraryRelativePath cross-platform
Windows CI failed: the tests hardcoded Unix-style absolute paths, which are
not absolute on Windows, and filepath.Rel yields backslash-separated paths
that the io/fs-based scanner FS rejects. Build the test paths with
filepath.Abs so they are absolute on every OS, and normalize the rebased
result with filepath.ToSlash.
* fix(scanner): resolve relative library root before rebasing target
filepath.Rel cannot rebase an absolute target onto a relative library root
(e.g. the default MusicFolder=./music), so an absolute --target was left
unchanged and rejected by the rooted io/fs. Make the library root absolute
first; it resolves against the same cwd as the scanner's fs.
* feat(ui): add Share button to artist detail page
* feat(ui): add Download button to artist detail page
* fix(ui): scope artist share/download to album-artist content
Gate the artist Share/Download actions on album-artist stats and show the
album-artist size, since ZipArtist and the share query only cover
album_artist_id songs. Previously the total (role-inclusive) size was shown
and guest-only artists could produce an empty archive. Applies to the artist
toolbar, the shared context menu, and the download dialog title.
* fix: match artist download/share to album-artist participation
ZipArtist and the artist share query filtered the deprecated album_artist_id
column, which only stores the first album artist of a track. Secondary
album-artists (co-credited but not first) got an empty download/share even
though the UI offered it. Filter by the album-artist role participation
instead, matching the artist's album-artist stats used to gate the actions.
Also cover the artist-specific size branch of the download dialog.
* fix(share): scope artist shares to the owner's libraries
The artist share query broadened to album-artist participation, which could
pull a secondary album artist's tracks from libraries the (non-admin) share
owner cannot access into the public share. Load the artist share as the owner
so their library access is applied, mirroring how playlist shares already work.
Adds a repository test covering co-album-artist inclusion and library scoping.
* test(share): assert album participation branch of artist shares
Link the co-album-artist fixtures to albums and assert share.Albums (used by
Subsonic getShares) includes the accessible album and excludes the one in a
library the owner cannot access, so the album participation + scoping branch
is covered too.
* fix: exclude missing files from artist download/share actions
An artist's stats still count files that went missing, so the toolbar/context
menu could offer Download/Share for an artist whose files are all gone, while
the share query (missing=false) returns nothing and downloads open dead paths.
Hide the actions when the artist is missing and exclude missing files from
ZipArtist, matching the share semantics.
* refactor: dedupe artist download-size and share-owner lookups
Extract the 'album-artist download size (or none when missing)' rule into a
single artistDownloadSize() helper shared by the toolbar, context menu, and
download dialog, and factor the duplicated share-owner context lookup into a
shareRepository.ownerContext() method used by both the artist and playlist
share cases.
* refactor(ui): move artistDownloadSize helper to common
utils is for domain-agnostic, potentially portable code; this helper is
Navidrome-specific (artist stats shape), so it belongs in common. Consumers
import it directly from common/artist to avoid pulling in the common barrel.
* refactor(artwork): let callers pin the blurhash component counts
* feat(artwork): synthesize a unique placeholder blurhash from a seed
* fix(artwork): base the synthetic-hash no-collision guarantee on the prefix, not length
The prior comment and test claimed a synthetic value could never collide with a
real one because it's always shorter. That's false: components() targets ~16
tiles by scaling one axis down as the other hits the 9 cap, so an extreme
aspect ratio (e.g. 10x200) collapses to 1x9 = 9 components, which encodes to
the same 22 characters as a 3x3 synthetic hash. The old test only exercised a
square 64x64 gradient, so it never caught this.
The real guarantee is structural, not length-based: the first character encodes
shape as (xComp-1)+(yComp-1)*9, and components() derives xf*yf = 16 exactly
before flooring/capping (xf = sqrt(16w/h), yf = xf*h/w = sqrt(16h/w), so
xf*yf = sqrt(256) = 16). Two factors both in [2,3) can't multiply to 16, so
Encode can never derive 3x3 - the synthetic prefix 'K' is structurally
exclusive to Synthetic. Replaced the length-based test with one that sweeps
extreme aspect ratios and asserts no real encode ever produces prefix 'K',
alongside the assertion that Synthetic always does.
* feat(artwork): tint a synthetic blurhash from a base colour
Parses baseColor into HSL and clamps saturation/lightness so the tint
stays muted; per-cell hashing (unchanged) is what keeps a shared tint
across an album's tracks from colliding, as the new 1M-seed spec
proves under a single fixed colour.
* fix(artwork): trim over-long comment in synthetic blurhash test
Review flagged the collision spec's comment for exceeding the 2-line
budget; the Finamp rationale it restated already lives in the design
doc and commit history.
* feat(jellyfin): send a synthetic blurhash for unresolved artwork
Clients render nothing where a placeholder belongs when ImageBlurHashes
is omitted for pending artwork. primaryImage now synthesizes a value
(seeded on the tag, so a cover swap re-keys it) whenever a tag is
present but no blurhash has been computed yet. Known-absent artwork
(ImageAbsent) is unaffected: it still emits neither tag nor blurhash,
since GetOrPlaceholder would otherwise pin a shared placeholder under a
distinct cache key for a year.
* fix(jellyfin): avoid computing a synthetic blurhash when a real one exists
cmp.Or evaluates both arguments before choosing between them, so
blurhash.Synthetic ran (and was discarded) on every call even when
img.BlurHash was already set. That's the common case in production:
artwork_repository.go populates BlurHash from persisted values once a
scan resolves it, so a healthy library paid the synthesis cost (xxh3
hashing, HSL conversion, image alloc, DCT encode) on every mapped item
for a value it never used. Branch on emptiness first instead.
* feat(jellyfin): tint a pending track's placeholder with its album's colour
primaryImage never reaches the embeddedArtPending branch of
SongToBaseItem, since there is no resolved image yet to feed it. Seed
the synthetic blurhash on mf.ID (so the value stays unique and the
client still issues the read-through request) but tint it with the
album's DominantColor, which is already hydrated on MediaFile at no
extra cost.
* style(artwork): trim synthetic blurhash comments to the budget
* docs(jellyfin): fix stale blurhash README bullet + two review nits
The "Blurhashes are synthetic" bullet under Known limitations described
dto/blurhash.go, which was deleted when the real core/artwork-computed
blurhash + synthetic-fallback pipeline landed; every claim in it was
false. Replaced it with an accurate paragraph in the Images section,
since the described behaviour is now the finished design, not a gap.
Also: fix a doc/body comment mismatch in blurhash.go (component counts
are 2..9, not 1..9), and deduplicate the inline DC-extraction logic in
synthetic_test.go by reusing the existing dcOf helper.
* refactor(artwork): slice the synthetic cell jitter on byte boundaries
The three perturbations came off one hash with mismatched masks and shifts
(0xFF at 0, 0x3F at 8, 0x3F at 14), so a reader had to do the arithmetic to
confirm the fields did not overlap. Only 20 of the 64 bits were in use either
way, so the narrower fields bought nothing.
Uniform byte slices at 0/8/16 are non-overlapping by inspection and give each
field the full 8 bits. Both 1,000,000-seed collision specs still measure
1,000,000 distinct values. Also drops the local `n` alias, which was a second
name for synthComponents inside a 15-line function.
* docs(jellyfin): fix the blurhash paragraph's opening sentence
It opened with "follow the same principle", pointing back at the preceding
paragraph on admin-context artwork resolution — an unrelated subject, so the
reader looks for a connection that is not there.
* fix(artwork): render the synthetic grid larger than its component count
The 3x3 cell grid was handed straight to the encoder as a 3x3 image, so the
source had exactly as many samples as basis functions. Blurhash normalises its
coefficients by 1/(w*h) and 2/(w*h), which assumes many samples per component,
so the AC terms came out far too large. At the bottom-right corner the x and y
bases are both [1, -0.5, -0.5], everything lines up negative, and the result
clamped to black — a dark blob on every synthetic placeholder.
Rendering the same nine colours bilinearly at 8x8 first removes it: measured
over three seeds, the darkest corner goes from 13 to 74 and the darkest pixel
from 1 to 63. 8px is the smallest size that clears the artefact; 12 and 16 are
visually indistinguishable and cost 1.7x and 2.7x more.
The interpolation is hand-rolled rather than x/image's scaler, which allocated
528 times per call against 14 for this. Both 1,000,000-seed collision specs
still measure 1,000,000 distinct values.
* refactor(artwork): tidy the synthetic upscale helpers
cellWeight nudged its upper bound with a 1e-9 epsilon so int() could never
land on the last cell. Clamping the coordinate and then the index says the
same thing without a magic constant, and makes the clamp-don't-extrapolate
intent explicit — edge pixels map outside the cell centres, so the fraction
would otherwise run past 1.
The grid type is spelled once as colorGrid rather than repeated in the local
and the upscale signature, and encodeAt's doc now states the source-size
contract that Synthetic depends on, so the next caller sees it at the
function rather than only in synthetic.go.
Output is unchanged: all seven sample hashes match byte for byte.
* perf(artwork): make the synthetic upscale separable
Both axes are square and constant-sized, so the per-pixel cell index and
weight were the same 8 values recomputed 64 times per call. They are now
built once by sync.OnceValue, matching the srgbToLinearTable pattern.
The interpolation is also separable: stretching each of the 3 grid rows
horizontally once and then blending rows vertically does 264 lerps where
the per-pixel form did 576.
upscale drops from 347ns to 260ns, Synthetic from 1780ns to 1669ns. Output
is byte-identical across all seven sample hashes — same operations in the
same order, only hoisted.
* refactor(artwork): let a caller supply the cosine basis
encodeAt built its cosine tables inline, so Synthetic rebuilt bit-identical
ones on every call — its shape is always 3 components over 8 pixels. Splitting
the table construction into cosBasis and the encoder proper into encodePixels
lets Synthetic build the basis once via sync.OnceValue, and leaves encodeAt's
signature and behaviour untouched.
Synthetic drops from 14 allocations to 6 (1669ns to 1508ns). The time saving
is small and this path only runs while artwork is still unresolved; the
allocation cut is the point, alongside a shorter encodeAt.
Every hash is unchanged: nine real Encode outputs spanning square, 10x200,
200x10, 1x50 and a real JPEG, plus all seven synthetic samples, all byte for
byte identical before and after.
