Completes the server side of #392 Half B.
PUT /api/me/timezone now calls scheduler.Refresh(ctx, userID) after
the DB write so the rescheduled daily-at-03:00-local job takes
effect synchronously. Failure to refresh is logged but doesn't
undo the DB write — the hourly reconciliation pass would pick it
up within an hour regardless.
POST /api/auth/register calls Refresh after successful user
insert so brand-new users get scheduled immediately rather than
waiting for the hourly pass to discover them.
system_cron.go deleted: the new scheduler subsumes its
responsibilities. The StartSystemPlaylistCron call in main.go is
also removed. Server restart now runs the new scheduler's startup
recovery + catch-up instead.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Builds the per-user daily-at-03:00-local scheduler for #392 Half B.
Uses github.com/go-co-op/gocron/v2 with WithLocation(userTZ) for
each user's job. Hourly reconciliation pass keeps the in-memory
job set in sync with the active-users query.
Start sequence: clear stale in_flight rows; register a daily job
for every active user at their stored timezone; fire a one-shot
runOnce to catch up missed schedules during downtime; start gocron.
Stop drains and shuts down the gocron loop.
Refresh(ctx, userID) removes the user's existing job (if any) and
registers a fresh one at their current timezone — wired into the
PUT /api/me/timezone handler and new-user registration in the next
commit.
Server struct gains PlaylistScheduler; main.go constructs it after
the eventbus and threads it through to api.Mount. The existing
StartSystemPlaylistCron call stays for one more commit so we don't
have a window where no scheduler is running; Task 3 deletes it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Schema + endpoint scaffolding for #392 Half B (per-user timezone
scheduling). Adds two columns to the users table:
- timezone text NOT NULL DEFAULT 'UTC' (IANA name)
- timezone_updated_at timestamptz (nullable; populated on each PUT)
PUT /api/me/timezone validates the IANA value via time.LoadLocation
and writes the row. No scheduler integration yet — the scheduler
struct lands in the next commit and the handler-side Refresh call
in the commit after.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five diversity mechanics — applied to both For-You and Songs-Like-X.
1. For-You seed rotates daily across the user's top-5 most-played
tracks. pickForYouSeedForDay uses userIDHash(user, day) mod
len(seeds) so today's mix uses an entirely different similarity
pool than tomorrow's. Within-day determinism preserved.
2. JitterMagnitude bumped 0.0 → 0.1. The scoring RNG is now seeded
by userIDHash(user, day) rather than the no-op, so near-tied
candidates shuffle daily without breaking within-day stability.
3. Head/tail split moves from 20+5 to 12+13. Roughly half the
playlist comes from the tail now (daily-deterministic via
tieBreakHash), giving the user substantially different content
while a 12-track anchor of strong similarity matches keeps the
mix recognizable.
4. Songs-Like-X seed artists shuffle daily across the user's top-5
played artists. pickSeedArtistsForDay applies a userIDHash-seeded
Fisher-Yates and takes 3.
5. scoreAndSortCandidates / pickTopN / pickHeadAndTail gain a userID
parameter so the RNG can be seeded per-user; existing call sites
updated; noopRNG removed.
Test fixtures widened similarity gaps (e.g. float64(50-i) instead of
(50-i)/50) so the new jitter (±0.1) doesn't perturb head ordering in
assertions about the head/tail mechanism. New seed_selection_test
coverage for userIDHash + pickForYouSeedForDay + pickSeedArtistsForDay
spans deterministic-within-day, varies-across-days, and graceful
degradation with small candidate pools.
PickTopPlayedTrackForUser replaced by PickTopPlayedTracksForUser
:many in the prior commit (b4801c2). The For-You seed lookup now
goes through pickForYouSeedForDay over the returned slice.
PickSeedArtists's LIMIT widened to 5 in the same prior commit.
For #392 Half A — system playlist content diversity. Half B
(per-user timezone scheduling) is a separate spec.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
For-You + Songs-Like-X seed selection moves to Go-side daily rotation
(next commit). The SQL change just widens the candidate pool: top-5
played tracks instead of single top played track; top-5 artists
instead of top-3.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Slice 4 — completes the #392 hybrid live-refresh loop.
live_events_provider.dart subscribes to /api/events/stream via dio's
streaming response mode, parses SSE frames (kind + JSON data + UserID
scope), and exposes them as a Riverpod StreamProvider. Heartbeat
comments are silently dropped; malformed JSON frames are skipped. The
provider auto-rebuilds when auth state changes (token rotation,
sign-out → sign-in), so reconnect is implicit.
live_events_dispatcher.dart listens to the stream and invalidates the
small set of publicly-importable providers we know about:
- myQuarantineProvider + homeProvider on any quarantine.* event
- homeProvider on any playlist.* event (Home renders the Playlists row)
Screen-private providers (library_screen.dart's _liked* /
_libraryAlbums / _history, admin screens, etc.) opt in to live-refresh
by themselves listening to liveEventsProvider in follow-up commits;
the dispatcher stays small and avoids back-edge dependencies on every
feature folder.
The dispatcher also installs an AppLifecycleState observer for
resume-time defensive invalidation. SSE will catch up on its own when
the app returns from background, but the invalidate flushes any stale
data immediately so the first frame back is fresh.
app.dart wires the dispatcher into the post-first-frame callback
alongside the other startup activations.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The hand-unrolled byte-pair version in reconciler.go tripped gofmt -s
and goimports. Replaced with the same loop-based formatter that
internal/library/eventbus.go uses — same behaviour, fewer lines, lint
clean. Two copies of the helper still exist (one per package) to keep
the no-back-edge property for both internal/library and
internal/lidarrrequests, but they're now identical and the duplication
is tiny.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Slice 3c — completes the server-side producer set.
Publishes scan.run_started when InsertScanRun succeeds and
scan.run_finished after FinishScanRun completes (success or with an
error_message payload). Per-file progress events are intentionally NOT
emitted — they'd flood the stream during large library scans without
giving the admin scan card anything actionable. The two-beat signal
gives the UI enough to invalidate scanStatusProvider at the right
moments.
Bus access uses a package-level setter on internal/library because
threading bus through RunScan + TryStartScan + Scheduler + all their
callers would touch ~10 sites without changing behavior at the boundaries
that don't publish. Per-process singleton, matches the log.SetDefault
idiom. cmd/minstrel/main.go calls library.SetEventBus(bus) once at
startup; test contexts that never call it skip publishing safely
(publishScanEvent is a no-op when bus is nil).
This completes the server side of #392. Slice 4 wires the Flutter
consumer: live_events_provider.dart (StreamProvider) +
live_events_dispatcher.dart (event-kind → provider invalidation)
+ AppLifecycleState cold-start invalidation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Slice 3b — extends event publishing to background workers.
When the reconciler matches an approved Lidarr request against the
library and flips it to 'completed', it now publishes a
request.status_changed event scoped to the original requester so their
/requests page invalidates without polling. Three call sites — one per
kind (artist / album / track) — capture the completed row instead of
discarding it so the publish has the user_id.
Bus plumbing: the reconciler runs in cmd/minstrel/main.go which
constructs services before server.New is called. Moved bus
construction into main; the Server struct gained a Bus field that
Router() reads, falling back to a fresh local instance when nil
(test contexts). Result: one bus per process shared by every
publisher and the SSE subscriber endpoint.
reconciler.go inlines a uuid->string helper rather than reaching into
internal/api for one — avoids a back-edge dependency.
Test compat: 9 NewReconciler call sites in reconciler_integration_test.go
get `nil` for the new bus param. The reconciler's publishCompleted helper
is a no-op when the bus is nil so tests that don't care about events
keep passing.
Scanner producer (scan.progress) deferred to slice 3c — needs more
thought about which lifecycle transitions warrant events vs. flood the
stream.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Slice 3a — extends the producer set onto the SSE bus.
quarantine events (5 producer sites):
- quarantine.flagged (user-side flag): broadcast so the flagger's other
clients invalidate their Hidden tab and admins' clients invalidate
their queue.
- quarantine.unflagged (user-side unflag): scoped to the user.
- quarantine.resolved / .file_deleted / .deleted_via_lidarr (admin
actions): broadcast because a single admin action can affect every
user who'd flagged that track.
playlist events (6 producer sites):
- playlist.created / .updated / .deleted: owner-scoped.
- playlist.tracks_changed: emitted for AppendTracks / RemoveTrack /
Reorder. Owner-scoped. Single kind for all three mutations because
the client invalidation logic is identical (refetch the detail).
Public-playlist subscribers (other users viewing someone else's public
playlist) intentionally left out — needs a separate broadcast kind, not
exercised yet at single-household scale.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Slice 2 of #392 — wires the first producers onto the bus that slice 1
built. After this commit, an SSE subscriber sees real events fire:
- track.liked / track.unliked when the user toggles the heart on a track
(handleLikeTrack / handleUnlikeTrack). Album + artist like events
intentionally deferred — they're symmetric trivial follow-ups but the
operator's primary like surface is tracks.
- request.status_changed when a Lidarr request is created, cancelled,
approved, or rejected. Auto-approve will fire twice (pending then
approved) in rapid succession, which is semantically correct; client
invalidation handles that fine.
Events are user-scoped via row.UserID so admin approve/reject route to
the requester, not the admin acting. Helpers live in events_publish.go
so the wire shape (kind names, payload keys) stays in one place — future
producers in slice 3 reuse the same pattern.
events_publish.go is no-op when h.eventbus is nil so tests that
construct handlers without a bus continue to pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
go vet caught a missed test caller of api.Mount after slice 1's signature
change added the *eventbus.Bus parameter. Pass eventbus.New() in the test
— the TestRoutesRegisteredInMount test only walks the route table for
existence, never publishes or subscribes, so a throwaway bus is fine.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Slice 1 of the #392 hybrid live-refresh work. Ships the in-process pub/sub
bus and the SSE subscriber endpoint; no producers wired yet, so the stream
emits only heartbeats today. Verifiable in isolation by curl-ing the
endpoint with a valid Bearer token — the connection opens, ": heartbeat"
lines arrive every 15s, the connection closes cleanly on client disconnect.
eventbus.Bus is a small fan-out broadcaster: subscribers register through
Subscribe (returns a receive channel + an unsubscribe closure), writers
call Publish, and the bus drops events for any subscriber whose buffer is
full rather than blocking the writer. No persistence — clients are
expected to resync via normal /api/* fetches on (re)connect.
The SSE handler emits an initial ": connected" comment so the client sees
the connection open immediately, then forwards events whose UserID matches
the authenticated user (or is empty for broadcast). Heartbeat comments
keep proxy connections alive. Context cancellation cleanly tears down the
subscription on client disconnect.
Producers (likes, request status, quarantine, scanner, playlist mutations)
land in subsequent slices.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three small parity gaps in _HiddenTab vs the web /library/hidden page:
- No unhide affordance — once a track was flagged via the kebab, the only
way to reverse it was to find the track in another surface and toggle
via the kebab again. Added an Icons.restore IconButton on each tile that
calls myQuarantineProvider.notifier.unflag(trackId) (the optimistic
remove-with-rollback already lived on the notifier).
- No album cover art — added a 56px ServerImage thumb matching the web
page's 14×14 thumb, with the same fs.slate fallback compact_track_card
uses for missing covers.
- No relative timestamp — appended _relativeTime(row.createdAt) next to
the reason pill so the user can tell "I hid this 3d ago" at a glance.
Also collapsed the duplicate provider: _HiddenTab was watching a local
FutureProvider that didn't see flag/unflag mutations, while the kebab's
HideTrackSheet flow goes through the canonical myQuarantineProvider
(AsyncNotifier). Switched _HiddenTab to watch myQuarantineProvider so
flag-from-anywhere and unhide-from-the-tab stay in sync. The local
_quarantineProvider was deleted; one source of truth now.
Caught during the #375 DRY audit cross-check against the #356 inventory.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The DRY pass audit (#375) found two inline mock patterns repeated across the
vitest suite: a default empty-playlists mock for $lib/api/playlists (4 exact
copies in menu/card tests) and an api-client spy mock for $lib/api/client
(9 callers split between get-only and full-RESTy shapes — unified into one
helper that always returns all four verb spies).
Mirrors the existing test-utils/mocks/likes.ts and test-utils/mocks/quarantine.ts
convention. Tests with intentionally divergent shapes (AddToPlaylistMenu's
richer createPlaylist payload, PlaylistCard's getPlaylist, route-specific
mocks, events.svelte's specific resolved value) stay inline.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The cross-user bucket query combined SELECT DISTINCT with ORDER BY md5(...),
which Postgres rejects at plan time (SQLSTATE 42P10). buildDiscoverCandidates
returned that error on first bucket failure, so the whole Discover playlist
was skipped every nightly run — even though the random bucket pool was
healthy. Switched to GROUP BY so the md5 ordering expression no longer needs
to appear in the select list, and hardened the function so future single-
bucket failures degrade gracefully via slot redistribution instead of taking
out the whole playlist.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two unrelated issues, batched.
Full-player kebab → "Go to album/artist" still crashed with
_debugCheckDuplicatedPageKeys despite the prior onBeforeNavigate
fix. Cause: pop() and push() ran in the same frame, so go_router's
page-key reservation table briefly contained both /now-playing AND
the destination shell-child, tripping the duplicate-key assert.
Refactor: replace onBeforeNavigate with onNavigate(path), where the
host receives the path AFTER the sheet pops and owns the
navigation entirely. The full player wires:
onNavigate: (path) async {
await Navigator.of(context).maybePop();
if (context.mounted) GoRouter.of(context).push(path);
}
Awaiting the pop guarantees /now-playing is fully gone from the
navigator's page list before the push starts.
Mini player + track rows + everywhere else use the default (no
onNavigate) path that does context.push(path) inline — they're
already inside the ShellRoute, no race possible.
Restored alwaysRefresh: true on artistAlbumsProvider. Dropping it
in the cache-loop fix had a side effect: if drift's cachedAlbums
held only a subset of an artist's albums (user previously visited
just one album by them), the artist detail page rendered that
partial list forever — provider only fetched on empty drift, never
on partial drift. The metadata prefetcher only mass-warms
artistProvider (single row), so re-enabling SWR on artistAlbums
won't recreate the storm.
Same root cause as the /queue duplicate-page-key crash: /now-playing
is a top-level route (lives outside the ShellRoute), but
/artists/:id and /albums/:id are shell-children. Pushing a shell-
child from a top-level route makes go_router attempt to mount a
second ShellRoute on top of the active one, leaving navigation in a
broken state. The mini player works because it's already inside the
shell.
Add an optional onBeforeNavigate callback to TrackActionsSheet
(forwarded through TrackActionsButton). When set, fires after
sheet.pop() and before context.push() of the destination route.
Wire the full player's TrackActionsButton with onBeforeNavigate:
() => Navigator.of(context).maybePop() so /now-playing dismisses
itself before the detail route is pushed. Result: clean navigation
into the destination, mini player visible underneath as expected.
Mini player keeps the default (no callback) since it's already in
the shell.
Server has been generating system_variant='discover' playlists since
M6a (internal/playlists/system.go and POST
/api/playlists/system/discover/refresh) but neither client surfaced
it. The Flutter home filtered system playlists by exact match on
'for_you' and 'songs_like_artist', dropping Discover. The web home
did the same. Tiles were generated server-side and silently
discarded by both clients.
Add a Discover slot to the playlists row in both:
- Flutter: 5-slot row now (For You, Discover, 3× Songs-like) with
matching placeholder state machine.
- Web: same shape, same placeholderVariant() call.
When the engine has built it, the real playlist tile renders;
otherwise a placeholder with the same building/pending/failed
status semantics as the other system tiles.
The lazy source-build commit (1ddde12) introduced a race: when the
user taps play on a new track while a previous queue's
_fillRemainingSources is still working, the stale fill keeps
calling _player.addAudioSource() on the new player state — appending
old-playlist tracks into the new queue and confusing the player into
the "locked to one song" symptom.
Fix: queue-generation counter. setQueueFromTracks bumps
_queueGeneration first thing; the background fill captures its gen
at start and aborts before any further player mutation if a newer
queue has taken over. The previous play() never gets to mutate the
new player state.
Also resets _suppressIndexUpdates at the start of every
setQueueFromTracks (defensive — covers the case where a prior
backward-fill bailed on its gen check before reaching `finally`)
and only releases the flag in finally if we're still the active
gen.
Symptoms this should resolve:
- "locked to one song" after rapid play taps
- Late `play() returned 73189ms` lines indicating a previous
hung play() call finally resolving and stepping on current state
- Player stuck in odd processingState after queue swaps
Same pattern as album + artist detail. PlaylistDetailScreen accepts
optional Playlist seed via go_router extra. While
playlistDetailProvider is still resolving, render the header
(description if any + track count) plus a "loading tracks…" inline
spinner instead of an opaque full-screen CircularProgressIndicator.
The body fills in when tracks arrive via drift watch re-emit.
Routing reads s.extra as Playlist. PlaylistCard +
PlaylistsListScreen pass the playlist via extra. Surfaces with no
seed available (deep links etc.) fall back to the original full-
screen spinner.
Track row rendering already uses ListView.builder so visible row
count was never the bottleneck — the wait was for the detail fetch
itself. Header-on-tap is the win that makes the screen feel snappy.
Two unrelated wins as a single batch.
Flutter — lazy source building in setQueueFromTracks:
Today: Future.wait builds all N AudioSource objects (drift queries +
LockCaching ctor) before the player can call setAudioSources →
play(). Measured at 83ms for a 25-track playlist, on top of the
~285ms initial-source preload.
New flow: build only the initial source, hand it to setAudioSources
([initial], initialIndex: 0) so play() can start, then background-
fill the rest. Forward direction (skipNext targets) added via
addAudioSource. Backward direction (skipPrev) inserted at index 0..
initialIndex-1 with _suppressIndexUpdates true so the unavoidable
currentIndex shifts don't push the wrong MediaItem onto the stream.
Saves the up-front source-build wait — tap-to-audio for long queues
should drop by ~80-100ms even on cache hits.
Server — Cache-Control on the three byte-serving endpoints:
- /api/albums/{id}/cover: max-age=86400, must-revalidate. Covers
change rarely (re-scan, MBID enrichment); a day of cache is safe
and skips conditional GETs for the bulk of a session.
- /api/playlists/{id}/cover: max-age=300, must-revalidate. Collages
recompute when contents change; short enough for edits to feel
fresh, long enough to skip repeat fetches during a session.
- /api/tracks/{id}/stream: max-age=31536000, immutable. Track bytes
are immutable for a given id (scanner re-indexes by file_path; new
files get new ids). LockCachingAudioSource on the Flutter side
already disk-caches, but proper headers let it skip even the
conditional 304 on repeat plays.
Logs were spitting "RenderFlex overflowed by 1.00 pixels on the
bottom" from album_card.dart whenever the library Albums tab or
artist detail album grid rendered. Cell height was computed as
cover + gap + title + artist + 4px slack, which assumes pixel-perfect
14sp/12sp line heights — Flutter's actual rendering with ascent/
descent + line-height multipliers wants one more pixel.
Bump slack from 4 to 8 in both grids. AlbumCard layout unchanged;
the warnings stop.
Otherwise the log shape is healthy now: prefetcher fires once per
library page emit (expected, one batch per pagination), cache misses
fetch sequentially with clean drift-write → re-emit cycles, and
album taps are single-round-trip cold-fetches followed by drift hits.
No more cycles of duplicate fetches.
Same fix shape as albums: drift's CachedPlaylistAdapter.toRef was
returning coverUrl: '' because the cache table doesn't persist the
server-derived URL. Set it to /api/playlists/<id>/cover in the
adapter — handleGetPlaylistCover serves the cached collage from
disk, so the URL is deterministic and the round-trip through drift
no longer drops it.
PlaylistCard + PlaylistsListScreen pass the existing queue_music
icon as ServerImage's fallback, so when the server hasn't built a
collage yet (system playlists with no tracks at build time), the
endpoint 404s and the icon shows over the slate background instead
of an empty box.
Tap→audio measured at ~370ms — that path's fast. Real symptom: a few
seconds of audio, then silence with no event surfaced to Flutter.
ExoPlayer is failing/completing somewhere and we have no log to act
on.
Three changes, all log-only:
- Add onError to playbackEventStream so stream failures (404, range-
request bugs, decoder errors, network drops) print instead of
silently halting playback.
- Subscribe to playerStateStream and log playing + processingState
on every transition. Silent stops will now show as a state shift
to completed / idle / buffering with no resumption.
- Subscribe to processingStateStream separately to catch fine-grained
state transitions ExoPlayer reports between source advances.
After hot-restart, tap-then-go-quiet should produce a sequence we
can read — most likely either "processingState=completed" partway
through (server returning premature EOS or wrong Content-Length) or
a thrown error from ExoPlayer's source-loading path.
CI fixes:
- artist_detail_screen.dart: drop unnecessary foundation import (debugPrint
comes from material) and unused metadata_prefetcher import.
Playback timing visibility (so we can stop guessing where the lag
lives):
- playTracks now logs serverUrl / token / configure / setQueue /
play() returned, each stage in milliseconds. The next time you
tap play, we'll see exactly where the seconds go.
- setQueueFromTracks adds two more measurements: total source-build
time across all tracks, and setAudioSources duration.
Small concrete win:
- audio_handler caches the application cache dir path on first use
(already cached in _maybeRegisterStreamCache; now also used in
_buildAudioSource for the LockCachingAudioSource path). One less
platform channel hit per track on cache-miss queue builds.
Once we see real numbers we can decide whether the fix is to build
sources lazily (initial source first → play → background-add the
rest), pre-warm the audio handler at app start so playTracks skips
serverUrl + token reads entirely, or something else.
The prefetcher + alwaysRefresh combination was creating a feedback
loop visible in the logs as repeated `metadataPrefetcher: warming N
albums` cycles, each kicking N parallel getAlbum fetches that then
triggered drift writes that triggered re-emits that re-ran the
prefetcher. Tap-to-play was queueing behind 14+ in-flight cache
fetches.
Three structural fixes:
1. Prefetcher hard-dedupes per session via _warmedArtists Set.
Re-rendering a screen no longer re-fires fetches for ids we've
already seen.
2. Prefetcher only warms artistProvider, not albumProvider. Albums
carry track lists; pre-warming N albums fans out N parallel
"fetch tracks" round trips for content the user may never visit.
Artist rows are single-row lookups — cheap. Album detail loads
on tap (still fast: server-side perf work makes it ~one round
trip).
3. Drop alwaysRefresh from albumProvider, artistProvider,
artistAlbumsProvider, artistTracksProvider. Each was kicking one
silent background refresh per first cache hit. With the prefetcher
creating many subscriptions in parallel, that meant every
prewarmed id triggered an extra fetch even when drift was already
populated. playlistsListProvider keeps alwaysRefresh — system
playlists genuinely rotate UUIDs and need the catch-up. Pull-to-
refresh remains the explicit invalidation path everywhere else.
Removed the warmAlbums calls from the library Albums tab and artist
detail album grid (the storm sources).
Net effect: cold app boot warms ~12-15 artist rows once, period.
Tapping a tile still fetches its detail on demand (one round trip,
fast). User-initiated playback isn't queued behind cache work.
MetadataPrefetcher gains warmAlbums(ids) / warmArtists(ids) public
methods so callers can fan out drift-cache warm-ups for whatever
collection just landed.
Wired into:
- Library Artists tab — warms first 8 artists from the page on every
data emit (initial + paginate + refresh).
- Library Albums tab — same for first 8 albums.
- Artist detail album grid — warms first 8 albums from the artist's
album list as soon as it loads, so tapping into any of them is a
drift hit.
Hard-cap of 8 per call (same as the home prefetch). Set spans across
calls aren't deduped at this layer because providers themselves
short-circuit on cached values.
Also instrument the artist-detail play button: try/catch around the
artistTracksProvider read + playTracks call, snackbar on
empty-tracks or thrown-error so silent failures stop being silent.
The current behavior was an early return on tracks.isEmpty with no
visible feedback.
Across-the-board sluggishness was every cold-cache tap doing one
network round trip while the user waits. SWR helps on re-visits but
the first time you tap a tile from home you eat the latency.
MetadataPrefetcher listens to homeProvider. When /api/home returns,
it fires fire-and-forget reads on albumProvider + artistProvider for
the top-N items in each home section (recently added, rediscover,
most played, last played). Each provider read triggers the existing
cold-cache path, which writes to drift. By the time the user
actually taps a tile, it's already a drift hit and the detail
screen renders instantly.
Cap N=8 per section (covers what's visible on a typical phone
without scrolling). Set spans dedupe across sections so popular
artists don't get fetched five times. Errors are swallowed — a
failed prefetch is silent and the tile falls back to its on-tap
fetch behavior.
Wired alongside the existing audio prefetcher in app.dart's
postFrameCallback.
Detail screen showed empty rows for system playlists because the
fetch batch wrote cachedPlaylists, cachedPlaylistTracks (positions),
cachedArtists, and cachedAlbums — but never cachedTracks themselves.
The detail screen's LEFT OUTER JOIN cachedTracks then returned null
on every row, so trackId was null and titles came back empty.
PlaylistTrack on the wire carries enough to populate cachedTracks
(id, title, albumId, artistId, durationSec). Adds a third dedup map
in fetchAndPopulate, batched with the existing artist + album writes.
track_number / disc_number aren't on the wire so they default to 0;
the detail screen doesn't surface them.
Reusing cachedTracks across albumProvider + playlistDetailProvider
also means tapping a playlist's track to play it now finds the row
in drift instead of triggering another fetch.
The 404 on tapping the For-You tile traced to stale drift rows.
BuildSystemPlaylists rotates UUIDs on every rebuild, so old For-You
/ Songs-Like ids accumulate in cachedPlaylists. The list provider's
fetchAndPopulate was only doing insertOrReplace, which adds new rows
but never removes the obsolete ones — so the home tile renders 8
playlists when the server only knows 4, and 4 of those tiles 404 on
tap.
Two fixes:
playlistsListProvider.fetchAndPopulate now reconciles. After
fetching the fresh list, deleteWhere any user-owned drift row whose
id isn't in the fresh response, then upsert the fresh set in the
same batch. Public-from-others rows are left alone — they're not
keyed by ownership and we don't want to drop someone else's public
playlist just because the current user's response didn't enumerate
it. Operates inside the existing batch so it's atomic.
playlistDetailProvider.fetchAndPopulate now treats a DioException
404 as "this row is stale": delete the cachedPlaylists + any
cachedPlaylistTracks rows for this id, return false so the UI yields
emptyDetail. The next render of the home row sees the row gone and
the tile disappears, completing the cleanup.
Side note: every system-playlist rebuild discards drift rows for the
just-evicted UUIDs and writes the new ones. That cycle's been
silently churning since system playlists shipped — this is the
first time the cleanup actually runs.
handleGetHome itself is well-architected (5 sections in parallel via
goroutines, latency-bound by the slowest single query). The cold-
start lag is two of those queries doing wider scans than necessary.
ListLastPlayedArtistsForUser was iterating FROM artists a with a
LATERAL play_events join per row — O(total_artists in library) plan
even for users who've only played a handful. Inverted: aggregate the
user's plays by artist_id first via the play_events → tracks join
(uses play_events_user_track_idx + tracks pkey), then attach the
artist row and lateral cover/count subqueries only for the artists
that actually appear. Cost now bounded by play history, not library
size.
ListMostPlayedTracksForUser was joining tracks/albums/artists for
every play_event row before grouping — O(total plays) work for
joins. Pre-aggregated play_events into a CTE keyed by track_id +
count(*), then joined to tracks/albums/artists only for the
distinct-tracks survivors. Order-by uses the pre-computed count.
No handler or generated-Go signature changes — both queries return
the same rowset shape, just much faster on libraries where total
artists/plays >> distinct-played-artists/distinct-played-tracks.
Two new sqlc queries replace three sequential per-album round trips
that were dominating detail-screen latency.
GetAlbumWithArtist: handleGetAlbum was doing GetAlbumByID then
GetArtistByID — separate round trips for one logical lookup. The new
query joins albums + artists with sqlc.embed and returns both in one
SELECT. Detail-page DB cost: 3 trips → 2.
ListAlbumsByArtistWithTrackCount: handleGetArtist was loading the
artist's album list, then issuing one CountTracksByAlbum per album to
populate track_count. On a 30-album artist that's 32 sequential
queries — each ~5ms over a local DB, ~30ms over a remote one. The
new query embeds the album row + a correlated count(*) subquery, so
every album's track count comes back in one SELECT regardless of
album count. Detail-page DB cost: 1 + N → 1 + 1.
Together these account for the bulk of cold-cache navigation latency
on the Flutter client. Combined with the existing SWR + nav
hydration on the client side, detail screens should render their
header instantly and the body within one round trip instead of
N+constant.
Closes the gap where LockCachingAudioSource wrote files to disk but
never told AudioCacheManager about them — meaning evict() couldn't
reclaim stream-cached files when usage exceeded the cap, only
explicitly-pinned downloads.
Wire just_audio's bufferedPositionStream as the "download complete"
signal: when bufferedPosition reaches duration (with 200ms slack for
header bytes), look up the on-disk file at the LockCaching path,
read its size, and insert an audio_cache_index row via the new
AudioCacheManager.registerStreamCache(). Source defaults to
incidental so stream-cached tracks are first to be evicted under
pressure.
Dedupe via _streamCacheRegistered Set so we don't hit drift on every
~200ms buffered-position emit. Cache the application cache dir path
on first use for the same reason.
Eviction now sees the full set of files on disk; usageBytes() (which
already walks the dir) and evict() (which reads the index) are
finally consistent for stream-cached tracks. Pinned tracks keep
their existing manual-download flow unchanged.
The Storage card has been showing "0 B" for users who only stream
(never explicitly pin or download an album). usageBytes() summed
SUM(size_bytes) from audio_cache_index — but the streaming path
through audio_handler writes files via LockCachingAudioSource without
ever inserting an index row, so the index undercounts (often to
zero) for normal use.
Walk the cache directory instead. Catches everything on disk:
manually pinned tracks (registered in the index), stream-cached
tracks (LockCaching), partial downloads. Falls back gracefully when
a file is racing against concurrent writes / deletes.
Eviction still operates on the index (it needs the source/recency
metadata to pick eviction order). Stream-cached files aren't subject
to eviction today — separate problem; addressed when we wire a
download-complete hook from LockCaching back into the index.
Tapping the queue button from /now-playing crashed with
NavigatorState._debugCheckDuplicatedPageKeys. Root cause: when I
moved /now-playing out of the ShellRoute earlier, /queue was left
inside the shell. Pushing /queue while /now-playing was on top
made go_router try to mount a second ShellRoute instance under the
existing one — both shells got the same page key.
Move /queue out of the ShellRoute too, sibling to /now-playing.
Shell stays mounted (with whatever child it had) underneath both
top-level routes; pop from /queue returns to /now-playing or the
shell's previous child as appropriate.
Move shuffle / repeat / queue out from below the play controls and
combine with the like + kebab into a single row sitting just above
the seek bar. Title row drops back to title-only (truly centered now,
no Stack needed since nothing competes for the row's right edge).
Layout order is now: art → title → artist → album → [shuffle, repeat,
queue, like, kebab] → seek → prev/play/next.
The "queue" icon in the top-right of the AppBar stays for now —
redundant with the new row but matches what users have already
muscle-memoried for opening the queue from any player state.
Both providers were FutureProvider<Paged<T>> returning a single page
of 50, so once the user scrolled past 50 items the list just ended.
Replace with AsyncNotifier<Paged<T>> that exposes loadMore(): fetches
the next page using items.length as the offset and appends to the
existing list. Idempotent guard via _loadingMore flag — concurrent
scroll events near the bottom collapse to a single fetch.
Both tabs now wrap the GridView in NotificationListener<ScrollNotification>
that fires loadMore() when within 800px of maxScrollExtent. The
lookahead is enough that the next page lands before the user hits
the visible bottom on a typical phone scroll.
Pull-to-refresh updated to invalidate + re-await the provider so a
manual refresh always starts from the first page.
Update installs were failing at the system scanner step because every
release APK was signed with a different signature: build.gradle.kts
fell back to signingConfigs.getByName("debug"), and the debug
keystore is generated per-machine — so every CI runner had a fresh
random key. Android refuses to upgrade an installed app when the
signature doesn't match, hence the "App not installed" failure
after the scan.
Two-part fix:
1. build.gradle.kts now defines a real "release" signing config when
ANDROID_KEYSTORE_PATH is exported (with the matching password +
alias env vars). Without those, falls through to the debug
keystore so local `flutter run --release` still works.
2. flutter.yml decodes ANDROID_KEYSTORE_B64 (CI secret, base64 of the
real release keystore) into a tmp path before the release-APK
build step, then exports ANDROID_KEYSTORE_PATH + the three
password/alias secrets as env vars. CI now hard-errors if the
keystore secret is missing on a tagged build — we don't want
another silent debug-keystore release.
OPERATOR ACTION REQUIRED before the next tag:
1. Generate a release keystore (one-time):
keytool -genkey -v -keystore minstrel-release.keystore \
-alias minstrel -keyalg RSA -keysize 2048 -validity 10000
2. Base64 it:
base64 -w0 minstrel-release.keystore > keystore.b64
3. In Forgejo, add four repo secrets:
ANDROID_KEYSTORE_B64 = contents of keystore.b64
ANDROID_KEY_ALIAS = minstrel
ANDROID_STORE_PASSWORD = the password you set
ANDROID_KEY_PASSWORD = same password (or different alias pwd)
4. Keep minstrel-release.keystore offline and backed up — losing it
means you can never sign an upgrade for any existing install.
5. Existing installs from prior debug-keyed releases must be
uninstalled before installing the first key-signed release. This
is a one-time transition; future tagged builds will upgrade
cleanly.
- artist_detail_screen.dart missed an `import '../models/artist.dart'`
for the ArtistRef seed parameter; analyze flagged undefined_class.
- radio.dart + player_provider.dart doc comments wrapped URL
parameters in <...> which lint reads as HTML. Switched to backticks.
Full player title is now centered absolutely via a Stack: title with
horizontal padding equal to the actions cluster width sits at the
optical center, while LikeButton + TrackActionsButton are pinned to
the right edge with Positioned. Fixed-height SizedBox(32) keeps the
row stable when the title wraps to a single ellipsized line.
Three issues, all related to the player surface:
1. Player UI didn't update on track change. audio_handler's
_onCurrentIndexChanged only kicked off the cover load — it never
pushed the new MediaItem onto the mediaItem stream. Title/artist/
cover stayed pinned to whatever setQueueFromTracks(initialIndex:)
set on first play. Now the listener pushes queue[idx] when the
index changes.
2. Player kebab "Go to artist" 404'd while the same item from
MostPlayed worked. Same TrackActionsSheet for both, but the
player's _trackRefFromMediaItem was hardcoding artistId: ''
because audio_handler's _toMediaItem never stashed it in extras.
Stash artist_id alongside album_id; player_bar +
now_playing_screen read it back. Both kebabs now navigate.
3. "Start radio" didn't exist on Flutter even though the server has
/api/radio?seed_track=<id>. New RadioApi (lib/api/endpoints/
radio.dart) wraps the endpoint; PlayerActions.startRadio(trackId)
fetches + plays the result via the existing playTracks path.
New menu item between "Add to playlist" and the divider above
"Go to album", calls startRadio with a snackbar error fallback.
System playlists now show in the home row (good!) but tapping them
landed on an empty playlist. Same root cause as the earlier album
bug: playlistsListProvider wrote the playlist *row* to drift but
never the tracks. playlistDetailProvider only triggered cold-fetch
when the row was missing, so it yielded with an empty track list.
Refactor playlistDetailProvider to mirror albumProvider's approach:
- fetchAttempted guard (one-shot per subscription).
- Detect "playlist row exists but trackRows empty" and trigger the
same fetchAndPopulate the cold-cache path uses. Drift watch
re-emits with populated tracks.
- 10s timeout on the API fetch + diagnostic prints so any failure
surfaces in logs.
While here, fix two adjacent issues:
- Wipe + re-insert this playlist's track positions on every fetch
so server-side deletions actually propagate. Without the wipe,
removed tracks would linger in drift forever.
- Write the artist + album rows referenced by the fetched tracks
into cachedArtists / cachedAlbums (deduped). Without this, the
joined artistName/albumTitle columns in the playlist track rows
surface empty.
Cover art for system playlists is a separate issue — server emits
coverUrl but it may be empty for system mixes; PlaylistCard falls
back to the queue_music icon. Will tackle in a follow-up.
Artists tab 404: client called /api/library/artists, but the server
mounts the artists list at /api/artists (handleListArtists). Albums
sit at /api/library/albums for historical reasons — the paths aren't
symmetric. Switch listArtists() to /api/artists with sort=alpha.
Albums tab grid: matched the responsive 3-up layout we built for
artist detail. LayoutBuilder computes cellW from available width;
AlbumCard sized to the cell with titleMaxLines: 2; mainAxisExtent
matches actual content height (cover + 2-line title + artist line +
fudge). No more wide-aspect cells with empty space below the card.
Also wired `extra: ref` on the artists/albums grids and the Liked
tab so detail-screen nav hydration kicks in here too — taps from
library screens get the same instant header that home tiles do.
Three changes addressing the cold-start spinner + stale-on-revisit pain.
A. SWR on remaining cacheFirst providers
- artistProvider, artistAlbumsProvider, artistTracksProvider all gain
alwaysRefresh: true. Cache hit still renders instantly; one
background refresh per subscription keeps the row from going stale
forever.
- albumProvider (inline async*) and playlistDetailProvider (inline
async*) now keep a `revalidated` flag and kick a one-shot
background fetch on the first complete cache hit. Same effect as
cacheFirst's alwaysRefresh, just inline.
- Three providers gained the connectivity timeout that
album/artist/playlist already had.
B. Navigation hydration
- AlbumDetailScreen accepts `AlbumRef? seed`; ArtistDetailScreen
accepts `ArtistRef? seed`. When the live provider is still loading,
the seed populates cover/title/artist immediately so the page
isn't blank.
- Routing wires `extra: AlbumRef|ArtistRef` from go_router into
the seed parameter.
- Call sites updated: home (Recently added, Rediscover albums +
artists, Last played), artist detail album grid. Where a ref isn't
available (track actions sheet), the screen falls back to the
spinner — no regression.
C. Cold-start home skeleton
- Replace the full-screen CircularProgressIndicator on /home with a
layout-preserving skeleton: 5 section titles + 6 grey card-shaped
placeholders per row. The page feels populated immediately;
sections fill in independently as data arrives via the per-section
providers.
- Drops the unused DelayedLoading import.
Net effect: re-visits to detail screens render instantly (cache hit
+ silent refresh); first visit from a tile shows the seed header
immediately while tracks load; cold-start home shows a layout
skeleton instead of a 30s blank spinner.
Web UI shows system playlists; Flutter shows placeholders. Wire shape,
adapters, drift schema, and filter constants all line up on inspection,
so adding instrumentation to pinpoint where the system rows fall out:
- Log what /api/playlists?kind=all actually returns (owned/public
counts, including how many of owned are system).
- Log what cacheFirst sees on each drift emit: total rows, filtered,
how many are system, how many landed in owned vs pub, plus the
user.id used for the owned-vs-pub split.
Also add the 3s connectivity timeout that albumProvider/artistProvider
got — keeps the alwaysRefresh path from blocking on a stalled
connectivity stream.
Three log lines from one home-screen visit will tell us:
- Does the server emit system rows? (wire log: system=N)
- Do they land in drift? (drift log: filtered system=N)
- Do they survive the user-id filter? (drift log: owned system=N)
Two test failures from the comparator rewrite:
1. "different unparseable strings → newer" was useful — branch-name
builds (e.g. PackageInfo reports 'main' while server reports
'dev') should still surface the banner so the operator sees that
something is misaligned. Restore that fallback: if both sides
parse to all-zero components (no numeric structure on either),
compare as strings.
2. "honors prerelease ordering" tested pub_semver behavior we no
longer use. Replaced with tests that cover what the new
comparator actually does: zero-padding for length mismatch,
leading-zero normalization, 4-part date+build comparisons, and
month-rollover correctness without leading zeros.
bvandeusen