A 404 on the media stream means the server resolved the item but could not
open the file behind it — moved, deleted, or on storage that went away.
Jellyfin maps the resulting FileNotFoundException to 404, and PlaybackInfo
never stats the file, so negotiation succeeds and only the stream request
fails. Playback then died with a snackbar reading "Failed to open
[REDACTED_URL]" before popping the route, which tells the user nothing and
leaves nothing useful in a bug report.
Generalize the HTTP-500 log probe into PlayerError.httpStatusFromLog and
latch every status in fatalPlaybackHttpStatuses. Each latches on its own so
the 503 that stream-lavf-o deliberately retries cannot mask the fatal status
behind it. A 404 now raises a dedicated modal naming the cause and the fix.
On Android a 404 previously failed the "Response code: 500" string test and
fell through to the ExoPlayer→MPV fallback, showing "switching to compatible
player" before failing again on the same request. Read the real status off
HttpDataSource.InvalidResponseCodeException instead and skip the fallback:
an HTTP status is not a codec problem.
The persist-then-flush model had four ways to lose or corrupt the record
it exists to protect.
A no-op hub — which is what a failed or timed-out crash-reporting init
leaves behind, because that phase is best effort — accepts an event and
returns an empty id without throwing. "Did not throw" was treated as
delivery, so the record was marked reported and suppressed forever.
Delivery now requires a non-empty Sentry id, and init completion is
tracked explicitly rather than assumed.
Opting out, and building without a DSN, are deliberate suppression
rather than delivery failure: both mark the record resolved so it is not
rediscovered every launch. Everything else stays pending, and
consumption no longer deletes an unreported record — deleting it ended
the only retry there was, which made "the next launch tries again"
false.
The write path is now a queue. Record writes were launched unawaited
from the failure path, so a fast retry could flush before the file
existed, consume before a late write landed, or run two writers against
one file and let the older one finish last. markReported joins the same
queue and compares record identity before rewriting, because reading and
writing outside it let a concurrent record land in between and be
overwritten by the record it had just superseded. Records carry an id so
that comparison is meaningful.
Consumption also waits on a registered flush, so the success path cannot
delete the file mid-send.
Also routes the tvOS recovery marker through the tolerant read.
reconcile() runs inside AppDatabase.open, a fatal gate step, so a
wrong-typed marker vetoed the launch outright on a first-class TV
target. Both new guards have regression tests verified to fail without
the fix.
Reporting the failure inline was wrong for the phase that matters most.
The gate opens preferences before SentryFlutter.init, so a corrupt or
unreadable store — the likeliest cause of #1732 — was captured by a
NoOpHub and silently discarded, which is exactly the telemetry gap the
previous commit claimed to close. Initialising the reporter earlier is
not an option either: `_beforeSend` reads the crash-reporting opt-out
from settings, so events raised before settings load would bypass a
user's choice.
Every failure is now persisted first and flushed once the reporter is up
with settings loaded, which in practice is the user's own retry seconds
later in the same process. Records carry a `reported` flag so a send
happens exactly once, and a failed send leaves the flag clear so the
next launch tries again. The flush reads without consuming, so the
record still reaches Settings > Logs.
Also routes the tvOS recovery marker through the tolerant read:
`reconcile()` runs inside `AppDatabase.open`, a fatal gate step, so a
wrong-typed marker vetoed startup outright on a first-class TV target
despite the new default-instead-of-veto behaviour. An unreadable marker
tells us nothing, which is the same position as an absent one.
The wrong-type recovery only covered reads that went through a
BaseSharedPreferencesService instance. The three stores that hold
credentials read the shared cache directly, so a mistyped value there
still threw a raw TypeError or, for Seerr, was swallowed by a catch-all
and reported as "no session" — the registry documented protection it did
not actually provide.
readPreferenceTolerantly now takes the cache, so CredentialVault,
TrackerAccountStore and SeerrSessionStore get the same classification as
the settings layer. CredentialVault's post-write re-read moves outside
its catch: a wrong-typed value written by another isolate was swallowed
there, and the process then returned a key that never durably landed,
making every ciphertext written under it unreadable on the next launch.
Those stores are consulted long after startup, where a throw is an
unhandled provider error rather than a repair prompt, so SettingsService
initialization now walks the cached key set once and reads every
sensitive key. That puts the failure inside a fatal gate step while the
store is still open and a surgical single-key repair is possible.
The remaining direct reads in settings and storage are routed too; the
only ones left are the library-density dual-type migration, which probes
both types deliberately, and an untyped switch that is type-safe by
construction.
Since 2.10.0 the whole app sits behind one all-or-nothing initialization
gate, and that gate discarded the only evidence of its own failure. It
caught the error, logged nothing but `error.runtimeType`, rendered an
icon plus the word "Error" plus Retry, and never reported the error
because catching it kept the crash reporter from ever seeing it. There
is no log file on any platform, the buffer is in memory only, a
double-clicked Windows release build has no console, and the log viewer
lives in Settings, behind the gate that just failed. #1732 is the result:
a Windows 11 user whose app will not boot and who cannot produce a single
byte of diagnostic detail.
The gate now names its phases. Each step is wrapped so a throw carries
the phase it came from, replacing a `Future.wait` that discarded every
error but the first and could not attribute it to any of four concurrent
steps. The failure screen renders the phase, the exception type, the
message and an expandable stack, plus copy and upload actions that reuse
the existing log-relay flow. The record is persisted next to the database
so the next successful launch can surface it in Settings > Logs, and it
is reported to the crash reporter explicitly.
Only preferences and the database still gate the launch. Window chrome,
locale, crash-reporting init, TV/performance detection, the image-cache
budget and download storage are best-effort and time-bounded, so a
stalled platform thread degrades instead of holding the splash forever.
Sentry no longer receives the startup work as its `appRunner`: that made
a startup failure indistinguishable from a Sentry failure, and the guard
would then have re-run migrations and the database open a second time.
The two remaining fatal steps become recoverable. Preference reads
tolerate a value whose stored type no longer matches, dropping the key
and defaulting instead of failing the boot. A store that cannot be parsed
is detected before either desktop plugin backend can memoise it, which is
what makes an in-process repair possible at all. Repair is never
automatic: it states what it will cost, salvages the credential-vault key
and every tracker and Seerr session it can validate out of the damaged
bytes, reseeds them, and moves the original aside rather than deleting
it. Servers and profiles survive a salvaged key because their tokens are
ciphertext in the database; tracker and Seerr sessions are plaintext
preference entries, so the copy says they may still need reconnecting.
Nothing derived from the store reaches a diagnostic. `FormatException`
prints an excerpt of whatever it failed to parse, and during startup that
document holds the vault key, refresh tokens and session cookies while
the redaction manager still has nothing registered, so the wrapper keeps
only the cause's type and offset and the record is an allowlist of
already-redacted fields. The quarantined copy is labelled as containing
credentials, is never offered for upload, and can be deleted from the
dialog.
Also self-heals orphaned WAL/SHM sidecars on desktop rather than only
tvOS, makes every `createTable` migration step idempotent, keeps MSVC
link by-products out of the Windows bundle, and asserts bundle contents
in CI.
Refs #1732
Apple TV single-line fields moved to the engine's UITextField proxy in
2.10.0 (71735354), which made three focus behaviours user-visible.
Submitting re-attached the input connection. EditableText schedules a
restart when a submit action fires with a non-null onSubmitted, and that
microtask runs before the setState flipping readOnly, so the field
re-showed a keyboard the form had just dismissed. The native path now
withholds onSubmitted from EditableText and invokes it from the host,
independently of onEditingComplete as _finalizeEditing does.
Auto-open fired on every focus entry, so D-pad traversal of a multi-field
form raised and dismissed the modal system keyboard on each step.
TvTextInputAutoOpenBehavior gains onFirstFocus, and the new `automatic`
default resolves to it on Apple TV: arriving at a field opens it once,
returning to it does not. Android TV keeps its docked-IME auto-open, and
explicit modes stay literal on both. The autofocused Jellyfin and Seerr
URL fields keep an explicit exception so entering the screen still does
not bury the form (#1217).
EditableText.connectionClosed unfocuses the field outright, so a UIKit
keyboard dismissal left nothing focused at all. The host takes focus back,
keyed on identity with the field's own enclosing scope so a dialog or
route claiming focus meanwhile is left alone.
close#1728
The Store's unpackaged EXE path would require Authenticode-signing the
installer and every PE file inside it. MSIX submissions are re-signed by the
Store instead, so this route needs no code-signing certificate. build-msix.ps1
mirrors build-installer.ps1 and consumes the same per-architecture build
artifacts, leaving the installer, portable archives and WinSparkle appcast
untouched.
One template generates the manifest for both architectures, carrying the
identity reserved in Partner Center. check_windows_msix.py recomputes the
package family name from the publisher DN, so a mistyped identity fails CI
rather than a submission, and it parses the script rather than running it
because root CI is Linux. Qualified logo assets are indexed into
resources.pri; without the altform-unplated variants the shell draws the
taskbar icon on an accent-coloured plate.
PlatformDetector.isPackagedInstall gates the in-app updater and the Liberapay
tile, which the read-only package directory and Store commerce policy
respectively rule out. Gating at runtime keeps one Windows build feeding both
the installer and the Store package.
Formatting was clean through 53288116 and then drifted across three commits on
2026-07-30: 1bf7aac7 left one source unformatted, f13f5af6 a second, and
daab4f1e four more. CI's Verify formatting job checks the whole tree, so it has
had six files to report ever since. No pre-commit hook is installed in this
checkout, so the aggregate check never ran locally to catch them.
Formatted with the dart_style revision Dart 3.12.0 bundles, which is what the
pinned Flutter 3.44.0 CI toolchain runs, rather than with a newer local SDK; the
two disagree about some argument-list splits. The current stable formatter
accepts this result as well, so both report the tree clean.
Trakt was the one service outside the tracker abstraction. TraktScrobbleService
re-implemented the whole playback lifecycle beside TrackerCoordinator, and
TraktSyncService pushed watched state from its own WatchStateNotifier
subscription, so the player called two objects at every lifecycle point and one
watch could be written twice. TraktTracker now implements RealtimeScrobbleTracker
like Simkl; the duplicated player call sites collapse to one each, and Trakt
shares the coordinator's ID resolver instead of re-fetching show ids every
episode.
Capabilities are split so a tracker declares what it is rather than being
special-cased: ScrobblePolicy carries each service's own resend/seek rules,
EpisodeHistoryTracker names the remote row a per-item history write targets, and
SeriesProgressTracker covers one-counter-per-series services. Writes from all
four trackers go through a shared TrackerWriteQueue, generalised from the
Trakt-only queue, with the legacy Trakt payload migrated on load. Trakt becomes
the fourth TrackersProvider slot and TraktAccountProvider is deleted, so one
object owns the active session per profile.
Two failure paths found while consolidating are fixed here too.
The queue's retries only ran on profile bind, connect and app foreground, so a
network blip mid-session left queued watches waiting for the next foreground.
OfflineModeProvider now notifies on connectivity changes, not just offline-state
or WiFi-flag changes, and main.dart flushes the queue when the network returns.
The queue also counted every failure toward the five attempts that permanently
drop an item, so a rate limit or a service having a bad hour could discard a
pending watch - the loss the queue exists to prevent. Only an answer about the
write itself now spends an attempt: 4xx counts, while rate limits, 5xx,
recoverable token-refresh failures and requests that never arrived do not. A
back-off answer also defers that service for the rest of the flush, so a queue
holding many rows does not fire all of them at a service that just asked for
quiet.
Simkl only heard about an item once playback crossed the media server's
watched threshold, so stopping partway recorded nothing at all: no resumable
position, no watch. Drive Simkl's /scrobble/start, /pause and /stop from the
player lifecycle instead, carrying the measured progress. Seeks report
nothing, as Simkl asks.
The terminal stop owns watched state for in-player playback, so real-time
trackers are excluded from the threshold markWatched fan-out and one watch
never produces two writes. Progress is reported as measured — it doubles as
the user's resume position — so when a server threshold configured below
Simkl's own 80% rule would leave the watch unrecorded, the tracker records it
through /sync/history rather than inflating progress. Manual, container,
offline-replay and external-player marks keep using /sync/history. Only
/scrobble/stop accepts a 409, which is the sole action documented to return
one.
Reports go out one at a time because Simkl serialises scrobble writes per
user and fails queued ones with a 400; overflow sheds the oldest non-terminal
report so an episode swap cannot drop the previous item's stop. A playback
session is pinned to the account bound when it began and every send re-checks
that binding, so a profile switch or a disconnect/reconnect can neither
redirect a queued report nor misfile the watched fallback.
Also close the paths that lost the terminal report entirely: app exit flushes
it instead of dropping it, the desktop window button goes through the app
shutdown rather than exit(0), a detached VOD player reports a stop, and a
finished item reports completion at EOF instead of waiting for teardown. A
session that opened at 0% is still closed on stop, or Simkl keeps showing the
item as playing until its runtime elapses.
close#1719
41ffaa7f2 gated picture-in-picture on FEATURE_AUTOMOTIVE and added a
settings case for it, but the assertion that case leads with — a stored
auto-PiP true surviving a read — needs supportsPictureInPicture() to be
true, and that gate ends in Platform.isAndroid || isIOS || isMacOS. The
term is false and unmockable on the Linux and Windows runners, so the
case passed on a macOS host and could never pass in CI: sanity checks
have been red for six commits on this one failure out of 4723. f13f5af6e
recorded it as a pre-existing Windows-host failure, but it entered in
this window and is red on Linux too.
Extract the gate's decision into a pure pictureInPictureAllowed that
takes the host's own capability as a parameter, the way
driver_distraction.dart already splits automotivePlaybackAllowed from its
ambient wrapper. The boolean algebra is unchanged, so the three callers
keep their behaviour; what changes is that the automotive and TV vetoes
become observable where every Platform branch is false, instead of being
vacuous on the host that gates the release.
The settings case keeps the pref-level contract on both host classes: a
stored true survives where the host supports PiP, and the gate pins it
off where it does not.
Verified with the host term forced false to emulate a Linux runner: both
files stay green, as does the full suite on macOS.
Plex treats a subtitle stream as forced when its title says "Forced" even
with the API flag unset. Every forced comparison now uses that effective
forced-ness on both sides: the match scorer, the low-metadata hard gate,
the Jellyfin OnlyForced/Smart profile modes, and stream-index negotiation.
Carrying a track choice into the next episode no longer reuses the
same-item identity matchers. A sealed SubtitlePreference (off / track
reference / semantic intent) replaces the id-'navigation' pseudo-track
through the whole preference channel, and cross-item intents hard-require
language and forced-class parity. When the next episode has no track of
the same class, the intent declines and selection falls through to the
server's own per-episode choice instead of latching onto a full track by
position and persisting that mistake back to the server.
Intents wait for pending native tracks under the same catalog-completeness
rule as source ids, so an early decline cannot retire the selection
listener before the real track arrives.
Ref #1716
Every artwork budget in the image pipeline was tuned for 1080p surfaces:
the transcode request clamp (1920x1080), the per-type decode caps
(poster 720x1080, thumb 960x540, heroLogo 1000x500, ...) and the TV
image-cache bytes. Those numbers are exact on phones and on the many TV
boxes that composite the app at 1080p, but a TV compositing at 4K
renders every capped image below its slot and GPU-upscales the result:
hero backdrops by 2x, hero logos by ~1.8x, wide episode thumbs by ~1.3x,
shelf posters by ~1.13x - the softness reported against the official
Plex client in #1697, and the class #860's min-2x-DPR fix could not
reach.
DevicePerformance now latches a display budget factor - the display's
shortest physical axis over 1080, capped at 2x - whenever the image
cache budget is applied (startup, post-mount, effects-setting changes).
The transcode clamp, the full-tier decode caps and the TV cache bytes
all scale by it, so a 4K surface fetches and decodes 4K backdrops and
proportionally larger cards. The reduced tier stays pinned to 1.0, and
sub-2.5GiB hardware holds the factor at 1.5 so full-budget 4K art
(~33MB per decode) cannot starve mid-RAM boxes; latching once per
session keeps transcode URLs - and with them the disk cache keys -
stable across rotation and rebuilds.
Whether a given TV composites at 1080p or 4K decides whether any of
this can help, and logs never recorded it: the startup banner and the
log-upload header now carry a display line (physical, logical, DPR,
latched budget) so uploaded logs answer that question directly.
The two pre-existing Windows-host test failures (automotive auto-PiP
gate, backdrop temp-dir teardown lock) reproduce unchanged on the base
commit.
Two defects sank Explore's Plex integration. Discover started rejecting
X-Plex-Container-Size=500 with a 400, so the watchlist membership
snapshot never loaded: hearts stayed unknown and toggles dead. The
snapshot now pages at 100, and getWatchlist refetches a rejected page in
chunks of the row fetch's field-proven 25, so the next cap drift degrades
gracefully instead of failing and callers' offset math survives either
way.
Worse, every Plex catalog item reached the library matcher carrying only
its Discover rating key: listings were fetched without includeGuids, so
the lookup rested entirely on exact plex:// guid equality between two
metadata universes (Discover duplicate entries break it, notoriously for
anime), and the title fallback can never confirm a candidate without
external ids to intersect - "Not in your library" for owned titles the
MAL provider matched fine. Discover listings now request Guids, the
detail screen re-runs the matcher when enrichment gains id forms
(generation-guarded so the slower bare lookup cannot overwrite the
richer verdict), the matcher keys its memo by id fingerprint so the poor
form's cached negative cannot answer for the rich one, and the Plex
client stops burning title requests that external-id verification is
guaranteed to reject.
Discover requests are now logged like every other API surface; this bug
shipped blind because they were not.
close#1715
Four sections of the catalog detail screen spent more room than their data
justified.
Franchise relations drew one hub shelf per label. Real payloads make that
absurd: MAL returns twelve relations for Attack on Titan across six labels,
and "Side story" and "Sequel" each hold exactly one title, so each spent a
header, a scroll row and one card. Flatten the labelled groups into one
"Related titles" section of compact rows — poster thumb, label, title and year
— that flow into columns on wide viewports. D-pad moves through the grid by
index and still hands off to the cast strip above and the recommendations
shelf below, which keeps its shelf because taste-based recommendations are
meant to be browsed.
Drop the MAL picture gallery. It was a horizontal strip of unfocusable poster
variants of the title you are already looking at, and it cost a page-height of
scroll; the `pictures` field comes back out of the detail request with it.
Draw attributed scores behind their own brand mark where the source has one,
the way the media detail screen already does: Rotten Tomatoes fresh/rotten and
upright/spilled, IMDb and TMDB, each on the scale that source publishes.
Sources with no mark (critic, audience, tracker scores) keep their written
label. Plex's own badge state is derived from the 60% tomatometer threshold it
encodes in `image.rating.ripe`.
Flow the definition rows — original title, studios, country, budget, box
office, crew — into two or three columns once the window is wide enough.
A 1440-wide window drew a 140-pixel label, a short value and 1,000 pixels of
nothing per fact.
Verified against live MAL and Plex Discover payloads on macOS: the Attack on
Titan page drops from 4,082 to 2,115 logical pixels, Dune: Part Two from 1,282
to 1,154.
The 1x1 keep-alive repaint loop was extended to Windows in a87aa296 to
paper over the legacy compositing path's resize desync (#227); the DComp
rework replaced that presentation path entirely. On the DComp engine the
100ms repaints become DirectComposition commits during playback, and
once fullscreen focus engages VRR (FreeSync/G-Sync) every commit forces
a scanout off the video's cadence - the micro-stutter of #1707.
The widget now owns the platform decision behind a test seam, and a new
quiescence test pins the hidden-chrome player UI to zero scheduled
frames so no future ticker can silently reintroduce the defect.
Explore shelf cards drew a poster, a title and a year. An audit of all six
catalog sources found the rest was lost at two boundaries — the wire-to-DTO
mapping and the DTO-to-CatalogItem mapping — and then simply not drawn: the
grid card fell through every branch of buildMetadataSubtitle to the year-only
case, while the list card used by search already composed certification,
runtime and rating from fields the synthesized MediaItem already held.
Extend CatalogItem with the neutral facts every provider had been dropping:
attributed rating sources, leaderboard ranks that keep their season window,
audience counters that keep their timeframe, broadcast slots, next-episode air
times, server availability and request state, exact release dates, alternate
titles, format, source material, studios, countries, languages, credits, tags,
links, artwork variants, play state, gallery art and background prose. Replace
fetchCast and fetchRelated with one fetchDetail returning the enriched item,
its cast, its recommendations and labelled franchise relations without adding
a request: sources needing two calls keep two and run them concurrently with
isolated failures.
Map those fields in all six sources, widening only field selections that cost
no extra round trip — MAL's fields list, AniList's selection set and a bounded
row cast that lets detail skip its character call, Trakt's guest stars, Seerr's
language parameter and TMDB size ladder, and Plex's includeUserState. Plex hub
artwork widens only on TV, where the spotlight is its only consumer, because it
doubles the payload.
Render them: a rating-first caption and bounded badges on the shelf card,
labelled sections on the detail screen, provider hub styles and result counts
on shelves, and logo, banner and accent art in the TV spotlight.
Verified against live Plex, AniList, Simkl and MAL responses, and on a Pixel 7.
The sheet collapsed an item to `Media.first` / `MediaSources.first` and
rendered a fixed set of rows, so split files, extra versions, per-track
properties, HDR classification and Dolby Vision were all invisible.
Model the payload the way both servers shape it — versions own parts,
parts own streams — and project every property either backend populates
onto `MediaStreamDetails`. The field set comes from sweeping both test
servers in full through the clients' own request shapes (Plex 6842
Media / 6842 Part / 39864 Stream entries, Jellyfin 6349 sources / 37763
streams / 61224 attachments), so file presence, Dolby Vision layers,
dynamic range, sample rate, spatial audio, sidecar provenance, embedded
attachments, rotation and the lyric stream type all survive. A coverage
test fails when a server key is neither carried, folded into a sibling,
nor excluded with a reason.
File Info also stopped trusting the shared `/library/metadata/{id}`
cache row: `getPlaybackExtras` writes it without `includeStreams` /
`checkFiles`, so the sheet could render with no stream table at all.
Detect that shape and refetch once under the request context captured
before the cache read, so the outgoing token and the cache namespace
stay on one profile.
Rework the layout to match: a summary chip row, then flat tonal cards
with a two-column field grid that collapses to one column on narrow
viewports, per-stream cards with flag chips, and a copyable monospace
path row. The card fill is a tonal step off the text colour rather than
the `bg` token, which is one shade from the sheet surface on OLED.
The skip badge doubled as an armed state: while it was up, any single tap
in the same-direction zone seeked again. The badge is also raised by
keyboard, D-pad, media-transport and live seeks, so one remote press armed
one-tap seeking on the touch surface with no double tap at all. It stayed
armed for 1200 ms plus the fade and renewed on every tap, leaving the side
zones - 35% of the width each, over 70% of the height - unable to raise
the chrome.
Pair taps off the pending single-tap timer rather than differencing
DateTime.now(). The window is then one deadline that a clock adjustment
cannot stretch, suppressing touch taps disarms a half-finished pair, and
_lastSkipTapTime belongs to the desktop double-click paths alone.
Consecutive completed skips still accumulate into one running badge total.
Pressing pause or seeking while the player's on-screen controls were hidden raised
the entire OSD, covering the subtitles the viewer was rewinding to read. Transport
keys now answer with a transient indicator and leave the chrome down; Select,
D-pad Center and a centre tap remain the deliberate way to bring the controls
back.
Play/pause confirms with an icon-only translucent disc at the centre of the frame,
72px around a 44px glyph, which grows and fades in, holds half a second at rest,
then runs the same motion in reverse. Seeking shows the amount plus a single
chevron on the same line at the edge it travels toward, with no backdrop at all:
anything large enough to read as a surface is large enough to cover picture and
subtitles, so legibility comes from shadows instead. Only the chevron moves, and
it eases outward across most of its cycle and returns briefly, holding a visible
opacity floor rather than blinking out. Type is scaled per platform, since a
television is read from across the room. The existing text pill stays for genuine
notices - rate changes, chapter titles, zoom, errors - because an earlier centred
pill overlapped ASS \an8 subtitle placement, which is the readability complaint
this feedback exists to answer.
Every relative seek entry point now shares one coalescing primitive. The keyboard
shortcuts fell through to KeyboardShortcutsService and previously reported
nothing, and both they and the remote's chapter fallback rebased each press off
player.state.position, so a burst against a slow backend pinned every request near
one step while the indicator climbed to a total that was never committed. A
released key commits its pending target immediately and resets the acceleration
tier, including on live TV where seeks bypass the accumulator. A chapter seek with
nowhere to go, past the last chapter or already at the start, no longer announces a
jump it does not perform.
Rewind-on-resume follows the resolved intent rather than the current state, so a
directed pause on an already-paused video neither resumes nor rewinds. Indicators
carry their own liveRegion semantics nodes: their labels previously merged into the
full-screen "show playback controls" target, corrupting its accessible name, and
they keep announcing "Paused"/"Playing" and the seek amount from icon-only visuals.
close#1676
MAL/AniList season entries never matched the library show they belong to.
Both backends use the catalog title as a server-side filter before verifying
external ids, and a title like "Mushoku Tensei: Jobless Reincarnation Season 2"
cannot reach a show stored as "Mushoku Tensei: Jobless Reincarnation". Measured
against a 267-show Plex library, 3 of 113 mapped sequel entries matched.
The reverse lookup now takes two ordered title candidates instead of one: the
entry's own title and its season-stripped form, with typographic punctuation
normalised because both servers miss on a curly apostrophe. That matches 77 of
113. Widening it further to romaji/native/synonym variants reached only 81, so
the cap stays at two rather than spending up to five more requests per lookup
that finds nothing.
A sequel's year is its own season's, not the parent show's, so the +/-1 year
window is dropped for one - it would exclude the very show being looked for.
That also keeps a miss at the two requests the single-title lookup already
spent. A Plex Discover item additionally skips the title search entirely by
filtering on the plex:// guid its own rating key already is, which costs no
extra request and needs no cloud lookup.
A season 2+ entry only matches when the server really has that season, which
costs one children fetch on a match. Only a season both TVDB and TMDB agree on
is gated: which provider a library numbers its seasons by is a server setting
no dataset supplies and none of it is inferable from the ids an item exposes,
so a disagreeing reference is left ungated rather than gated on a guess.
The match cache is keyed per source and per entry rather than by canonical id,
which every season of a series shares: all five Mushoku Tensei entries collapse
to imdb:tt13293588, so one season-gated result would have poisoned the rest.
Entries whose Fribb row carries no provider id at all remain unmatched. That is
an upstream mapping gap, not something to guess around with extra lookups.
close#1704
media3 reports STUCK_PLAYING_NOT_ENDING when the player sits in STATE_READY
past the declared duration with no renderer ending. On a tunneled MTK decoder
the clock ran a full minute past the last frame behind a black screen, so the
item never completed: no Play Next, no auto-play, and a "playing" timeline the
server kept extrapolating past the item duration.
Treat that report as the end of the file when the rendered-frame counter has
stopped as well, which separates a finished file from a container that
under-declares its duration and is still painting. The terminal event is shaped
like the STATE_ENDED one and pins the timeline at the duration first, so the
completion flow cannot mistake it for a stream that died mid-file. Shorten the
detection window to media3's stuck-playing default, and clamp a backend
hand-off to just inside the media so a fallback can no longer resume MPV past
the last frame and park there without reporting it.
close#1673
Plezy declares appCategory="video", so on Android Automotive OS it is a
parked app bound by car app quality DD-2/DD-3: audio must stop when the
vehicle starts driving and must not be resumable while driving. Two paths
kept audio alive. Music playback ran under a mediaPlayback foreground
service whose lifecycle observer was registered for Apple TV only, so it
never paused when Android backgrounded the app. Video pausing hung off
AppLifecycleState.hidden, which Flutter only synthesizes once Android
delivers onStop; a car without the Automotive compatibility mode delivers
onPause alone, which maps to AppLifecycleState.inactive and the player
ignored.
Gate every path that can start audio on a new lifecycle predicate,
automotivePlaybackAllowed, which permits playback on a car only while the
app is resumed and fails closed on an unknown lifecycle state. That covers
explicit play, gapless arming and track transitions, live retry and
channel switch, frame-rate-match resume, VOD/live startup, and the queue
navigation commands of the OS media session, plus a last-resort pause for
when the platform player resumes itself on native audio-focus regain.
Playback authority on the media-session router is deliberately left alone:
the router consumes a denied event, so gating it would swallow PauseEvent
and leave the OS unable to stop audio. Reacting to lifecycle callbacks is
the mechanism the platform documents as sufficient, so no android.car
dependency is added.
The music queue no longer requests POST_NOTIFICATIONS on a car, where the
foreground service and its notification never start: there is nothing to
authorize, and the prompt would take focus and make the gate discard the
first play intent.
Detect the form factor too: FEATURE_AUTOMOTIVE now vetoes the Android TV
verdict, so a rotary-only head unit no longer inherits the leanback
experience. Picture-in-picture is gated on FEATURE_PICTURE_IN_PICTURE,
which cars lack, so the app's UI cannot stay on screen while driving, and
nothing forces a preferred orientation on a fixed-orientation display.
Audio passthrough defaults on for Android TV, scoped to ExoPlayer because
mpv force-passes through every codec named in audio-spdif and has no
decode fallback. That scoping did not survive the ExoPlayer to mpv
handoff: PlayerAndroid queued the raw ac3,eac3,dts,dts-hd,truehd list as
a pending mpv property and prepareMpvFallback replayed it verbatim, so a
sink that bitstreams only Dolby formats was told to force TrueHD and
DTS-HD anyway. mpv selected spdif_truehd, the audio output never
initialised, and playback froze at its start position while still showing
a first frame — the stop timeline reported the position it opened with.
Treat passthrough as a request and resolve the codec list against the
route when mpv actually starts, so an HDMI or AVR change between
ExoPlayer startup and the handoff cannot replay codecs from the old sink.
Gate each codec on the exact advertised encoding rather than media3's
passthrough probe: that probe answers DTS-HD by downgrading to the DTS
core, and mpv reads "dts,dts-hd" as "dts-hd" alone, so accepting the
downgrade would name DTS-HD MA to a core-only receiver and lose DTS too.
The scoped lookups run in sequential batches, and MediaServerHttpClient
applies a per-call timeout to connect and receive separately. A silent
endpoint therefore cost up to two request timeouts per batch, and six
batches of that outlast the single request the scoped form replaced —
the enrichment could hold Continue Watching longer than the query it was
introduced to fix.
Give the pass one deadline instead of a per-batch check. It aborts the
in-flight batch and is also raced client-side, because aborting only asks
the transport to stop and not every client honours abortTrigger. Whatever
phase a lookup is stuck in — silent connect, delayed headers, stalled
body — the pass now ends at the deadline with whatever dates it has.
Plezy's device profile declares every subtitle format with
`Method: External`, so Jellyfin answers PlaybackInfo with
`DeliveryMethod: External` and a `DeliveryUrl` even for streams embedded
in a direct-played container. Direct play never fetches those URLs, but
the rows kept the delivery URL as `MediaSubtitleTrack.key`, and keyed
rows only match a native track loaded from the same URL. No embedded
track could satisfy that, so `selectSubtitleTrack` reported "still
pending" forever: playback started with subtitles off and logged the
five- and thirty-second waits, and the server's default subtitle had to
be picked by hand on every item.
Restrict sidecar identity to the rows an open actually fetched as
sidecars. A row that stays in the container loses `key` and
`usesExternalDelivery` and matches on metadata again; genuine
`IsExternal` files keep theirs, and remuxed or transcoded renditions
still resolve their sidecars by URL.
Also declare every subtitle format Embed-first so a direct-played
container reports embedded delivery in the first place, and make the
pending contract match its purpose on every backend. The
complete-catalog escape is no longer Plex-only, so a Jellyfin row the
native player has not produced keeps the pass pending instead of
committing an unrelated default and retiring the listener that was
waiting for the real track. A source id absent from the catalog no
longer defers a decision that can never change, and the thirty-second
deadline resolves from what has arrived instead of re-deriving the same
deferral and applying nothing.
close#1696
The Next Up shelf dated its rows from one server-wide
`/Items?SortBy=DatePlayed&Recursive=true` scan. Jellyfin 12.0-rc3 builds
that sort key by OR-ing an item's own progress with its alternate
versions' (`ItemId == e.Id || Item.PrimaryVersionId == e.Id`,
jellyfin/jellyfin#17044), which no index can serve, so the user's whole
UserData table is scanned per sorted row. Measured on identical
10,120-item libraries, that scan cost 25ms on 10.10.7 and 5.8-13.3s on
12.0-rc3 while pegging a core, so it blew the call's 10s budget and
starved every other client of the server for tens of seconds. Upstream
fixed the order mapper after rc3 in jellyfin/jellyfin#17422.
Ask each pending series for its own newest played episode instead:
`ParentId` bounds the sort input to that series, and the same 21 series
now resolve in 1.5s against the rc3 server with byte-identical dates.
The lookups run four at a time under a shared wall-clock budget and a
short per-request timeout, so a silent endpoint costs less than the one
default-budget request this replaced, and a `count: null` shelf can no
longer fan out one request per started series. Endpoint failover stays
off so a slow enrichment row cannot move the client off a working
endpoint.
close#1699
Cycling backdrops reach a fallback path only once every rotating path
has failed to load, but every hero passed the rotation-agnostic backdrop
list as the rotation set and the aspect-ordered candidates as the
fallback. One servable wide backdrop was therefore enough to hide the
square background for good, so phone detail and Discover heroes
cover-fitted a 16:9 backdrop into a portrait box instead of showing the
square image Plex supplies.
Give the rotation set the same aspect-aware preference the candidate
list already has: near-square containers rotate the square background
alone and keep the backdrops behind it as fallbacks.
close#1700
Explore only reached search through an app-bar icon that pushed a separate
screen. Touch and pointer builds now carry the field inline under the app
bar: results replace the shelves while the query is non-empty and the
shelves return when it clears. TV keeps pushing CatalogSearchScreen, since
a text field cannot share the spotlight scaffold with the bottom-pinned
browse rail and the on-screen keyboard.
Pull-to-refresh and the toolbar refresh action re-run the live query
instead of reloading hidden rows, and switching catalog source re-runs the
query against the new source rather than leaving the previous source's
results under its name.
Plex Explore showed only the Watchlist row. `/hubs/sections/watchlist`
answers with placeholder hubs — every entry carries `placeholder: true`,
`size: 0` and no `Metadata` — so `fetchHubs` mapped each one to an empty
page and dropped all of them. That is true no matter what the profile has
watchlisted; the shelves never rendered.
Read `/hubs/sections/home` instead, the section Plex's own web client
renders on its Home > Trending tab, and hydrate each placeholder from its
own key (six at a time). `directory` shelves list browse categories and
`clip` shelves list trailers, neither of which becomes a catalog item, so
they are skipped before spending a request. A shelf that fails degrades to
the ones that succeeded; a pass where every shelf failed still throws.
Discover ignores container offsets on hub keys and truncates with `limit`
instead, so a hub is one page: View All takes the whole shelf in a single
request rather than replaying page one, and hub requests drop `Media` and
`Image` elements the catalog layer never reads.
Since real device names started reaching the header, an accented one
made login impossible: dart:io refuses header values above 0x7F, and
CFNetwork puts the raw code unit on the wire as a Latin-1 byte, which
Kestrel rejects as a malformed request with 400 before Jellyfin routes
POST /Users/AuthenticateByName.
Encode every field the way the official Jellyfin SDK does; the server
already reverses it with WebUtility.UrlDecode, so the wire value stays
pure ASCII while the device list shows the real name. Quotes, commas
and `=` no longer need stripping either. sanitizeHeaderValue, which
still guards the Plex headers, now folds Latin letters to their base
form instead of emitting bytes no transport accepts.
close#1685
Detection now runs off the UI isolate and covers every platform where
the answer can be trusted.
Availability was a plain platform check, so Linux always listed VLC, mpv
and Celluloid, macOS always listed VLC and IINA, and Windows always
listed VLC and PotPlayer whether or not any of them existed. Each player
now has a detector that asks exactly the question its launcher asks:
`sh -c 'command -v'` for PATH launches so the kernel performs the
executable check, NSWorkspace/Launch Services for `open -a`, `where.exe`
plus the concrete install paths for Windows VLC, and the registered URL
handler for PotPlayer and the iOS players.
Detection is asynchronous and memoised behind KnownPlayers.probe rather
than a Process.runSync in a static initialiser, which forked three
shells on the UI isolate during ExternalPlayerScreen.build. It is
prewarmed from startup, fails open when a probe throws, and keeps the
selected player listed when a detector misses it so a false negative
cannot leave the list with nothing selected.
iOS and tvOS gained LSApplicationQueriesSchemes entries for vlc and
infuse. Without them canOpenURL returns false for both schemes, so
_launchUrlScheme was already refusing to hand off to either player.
Android keeps the platform check: package visibility needs native
declarations, and a wrong answer there hides a working player.
A room whose peers have all left is a code nobody is using, but the relay
kept it bound to the creator's reconnect capability and rejected every
other create with room_exists. The app compounded it: enterRoom only
promoted to host on room_not_found, so tapping a recent code landed the
user in the retained room as a guest of a host that was never coming
back, until the cleanup sweep finally dropped the room.
Create now replaces a room with no connected peers, and enterRoom hosts
the code when its probe join finds an empty room. An occupied room still
rejects create, including from its previous owner, and a host that is
merely disconnected still reclaims its peer ID through join with the
matching token.
The scan omitted six files the config itself cites, so jsdelivr.net,
api.github.com and image.tmdb.org were listed as system-only while no scanned
source referenced them; changing those hosts would have fallen through to the
base config and its user certificate authorities undetected.
Also assert the reverse direction, so a domain no scanned source produces
fails instead of silently losing its guard.
Persisting the deferred choice suspends, so the source switch can be
superseded before the pass is armed. Return early when the continuation is
stale, and refuse to arm a disposed or inactive TrackManager at all.
The per-callback generation checks only stopped the work; the subscription
and the five-second timer were still allocated on a manager whose dispose had
already run, so nothing would ever cancel them.
Return initialized show and season rules from the backfill query so their
coverage is recomputed from download ancestry. Rule execution links only the
unwatched episodes it inspected, so the cached flag let a sibling list
cleanup delete episodes the show rule covers.
Keep the action enabled and let a press retry the snapshot, as the media
detail action bar already does. A disabled sole action left the detail screen
with no initial D-pad focus on TV.
Distinguish an unresolvable URL from a failed load at the error-widget
boundary. A transiently null media client during a profile switch or
reconnect marked the primary poster dead in a process-global set, pinning the
item to fallback artwork for the rest of the session.
Scope user-installed CA trust to the user-entered server hosts that need it
and pin the hard-coded first-party hosts to system anchors. The base config
applied user CAs to every host, including plex.tv token exchange and the
OAuth proxy.
Only the load that commits a favorites set writes the loaded flag, so a
refresh keeps the previous set authoritative. Clearing it up front widened
the guide to the full lineup for the whole round-trip and moved the D-pad
cursor when it collapsed back.
Match the tolerant list predicate the import path already uses. The platform
preference cache returns List<Object?> after a restart, so the exact
List<String> pattern silently dropped tracker library filter ids and a
restored profile resumed scrobbling libraries the user had excluded.
Supplementary repair runs over downloads whose video is already complete, so
it no longer emits a downloading transition, and artwork updates carry the
row's real status instead of asserting downloading. Previously a reconnect
left completed downloads stuck at "downloading 0%" until the next DB read.
Storage exhaustion fails every active row in one transaction, so
failActiveDownloadsForStorageFull now returns the affected keys and each one
gets a failed event; only the triggering key was announced before.
The post-recovery database open also closes its handle before rethrowing. A
failing storage-full write abandoned a drift background isolate and its
SQLite handles on every retry.
Persist every user-entered URL that is not positively known to belong to a
different server, matching reconcilePreviouslyStoredBaseUrls. Requiring a
successful identity probe deleted a stored LAN endpoint whenever the box was
asleep or the user saved from outside the network.
Persist the choice before arming the deferred selection pass. The screen
callback routes to onSubtitleTrackSelectedByUser, which invalidates the
pending selection, so arming first retired the very listener that applies
the choice once mpv discovers the sidecar.
The existing test stubbed the persist callback and so could not observe the
invalidation; it now routes through the manager like production does.
The Dart suite spent 77% of its cost compiling one isolate per test file
while `flutter test` used half the cores, and every Maestro flow replayed
a full Jellyfin onboarding before its first real assertion.
- Add scripts/run_tests.sh, which runs `flutter test` with -j set to the
cores the process may actually use instead of the ncpu/2 default.
Measured on 8 cores: 190s -> 136s; -j 12 regresses to 165s, so it scales
to the core count rather than hard-coding one. CI and CONTRIBUTING use it.
A cgroup v2 quota, a cgroup v1 quota, and the cpuset/affinity nproc
reports can each be the binding limit independently, so the detector
takes the smallest; trusting whichever it found first would oversubscribe
4x on a container holding an 8-CPU quota while pinned to 2. Covered by
scripts/test_run_tests.py, which the ci_guard_checks.sh glob picks up.
- Add .maestro/subflows/ensure_onboarded.yaml: cold-start the app and only
onboard when no session is stored. Flows that just need a signed-in Home
use it; 02_onboarding_home, 08_logout, 09_download_offline_playback and
the profile regressions keep clearing state. 59s -> 16s per flow.
- Guard onboarding's two optional taps behind visibility checks. A missed
`optional: true` tap still runs the full element search, costing 3.0s
and 7.8s per onboarding to find nothing.
- Disable device animation scales in run_maestro.py, restored by the
existing cleanup path. CI's emulator got this from the runner flag;
physical devices never did.
- Shorten the watch_together setup-timeout replacement from 500ms to the
10ms the same file already proves sufficient, and shorten the retry
backoff at the one site that missed it: 8.04s -> 1.59s of execution.
- Make the LAN discovery waits deadline-based and resend the beacon while
polling. Loopback UDP drops datagrams under load, which timed out a
wait that could never be satisfied; this was the suite's one flaky test.
- Fix 08_logout, which searched for "Logout" and "Are you sure you want to
logout?" after both strings became "Log out". The flow had been failing
and aborting the suite before 09 ever ran.
flutter test 190s -> 131s. Maestro's Android suite 621s -> 385s across the
eight flows the baseline reached, and now runs all nine green.
Back left the Manage Libraries sheet open on Android with no way to
dismiss it. The host answered the platform pop with
`BackKeyCoordinator.consumeIfHandled()`, which dedups the focused key
path against the platform pop. Only TV routes one Back through both;
touch platforms never deliver Back to the sheet's key handler, verified
on device — a physical Back produced only `popRoute` and no key event.
So there was nothing to dedup against, and the global one-shot marker,
once set by any other handler, silently swallowed the only signal that
closes the sheet.
Scopes the dedup to TV. The TV regression that guarded this never set
the TV override, so it asserted the swallow on every platform and hid
the defect; it now enables the override and a touch counterpart pins the
dismissal.
Also blocks semantics behind the barrier. The barrier takes every
pointer event but left the screen underneath in the semantics tree, so
assistive tech and UI automation still saw rows that could not be
activated — Maestro read an occluded settings row as visible and tapped
its stale coordinates into the sheet. Flutter's own ModalBarrier blocks
semantics for the same reason.
Verified by replaying the failing Maestro sequence
(.maestro/subflows/settings_deep_checks.yaml lines 42-60) on a device:
back dismisses the sheet, Services opens, and back returns to settings.
On phone layouts main_screen pushes SettingsScreen as its own route, and
that route carried no OverlaySheetHost. showAdaptive could not find one
from the tile's context, so Manage Libraries fell back to
showModalBottomSheet. The sheet also owns a focused Back handler, so a
single Android Back arrived twice — once as a key event, once as
popRoute — and the two route-based paths raced, tearing down Settings
along with the sheet.
Installs one route-local host when no enclosing host exists, and opens
the sheet from a context below it. OverlaySheetHost then holds the route
while a sheet is open and deduplicates the key path, so one Back closes
only the sheet.
When a source advertises subtitles the native track list has not
produced yet, applyTrackSelectionWhenReady keeps an automatic selection
armed for up to thirty seconds. That late pass re-runs
TrackSelectionService against the stored preferences, so a track the
user picked in the meantime was silently reset. The Maestro codec suites
caught it: the English E-AC3 and Japanese DTS-HD flows select an audio
track, and fifteen seconds later the deadline puts the preferred
language back.
Adds explicit user-selection entry points that retire the pending
automatic selection first, and routes the sheet callbacks and the
remote's cycle shortcuts through them. Subtitles get the same treatment,
because the same pass re-selects them.
Bumping the generation is sufficient: TrackSelectionService re-checks it
in the statement immediately before each select call, and a mutation
already in flight was dispatched before the user's and so lands first.
`dart format --set-exit-if-changed` over lib and test rewrites these.
The analysis job never reached its formatting step, so the drift went
unnoticed. No behaviour changes.
edde746