Round two of review on the accounts layer.
The settings picker mapped its legacy instance-keyed selection straight
onto DataContext(legacy:), which turns an instance UUID into an account
UUID. That identity only holds for migrated accounts - a login created
after the revamp has a random ID - so selecting a freshly added account
named one that does not exist and left the provider nil. Selections now
resolve through AccountsManager to an actual account on that instance,
preferring a signed-in one. The picker still lists one entry per instance
and so cannot show a second login; the Phase 3 switcher replaces it.
BackgroundFeedRefresher was missed in the previous sweep because it
resolves credentials differently: it routed on the legacy setting and
looked up "first enabled Invidious instance" with an instance-keyed
credential. After any new-style login that silently found nothing and
skipped the refresh. It now routes on the selected context and takes the
credential from that account.
Subscription reads used the raw stored context, so an account waiting on
its instance or credential rendered empty instead of falling back to local
data as the tri-state design specifies. Reads go through the validator's
effective context; writes deliberately do not, so a subscribe fails
visibly rather than landing in local data the user did not choose.
Account statuses are cleared when the identity namespace reloads, but
nothing recomputed them once a deferred scope resolution succeeded - a
synced account kept showing "sign in required" until relaunch. The scope
change now triggers a refresh, deferred by one turn so it cannot depend on
observer registration order relative to the store's own reload.
Five findings against the accounts layer, all real.
A fresh login created a random-UUID account and then flipped
usesAccountLogin, which posts instancesDidChange and re-ran legacy
reconciliation immediately. That guard only looked for an account whose ID
equalled the instance ID, so it synthesized a second, credential-less
account alongside the real one. Signing in now marks the instance migrated,
and reconciliation treats any account on an instance as evidence its login
is already represented.
save() set hasStoredContext, so the save inside reconcileLegacyAccounts
marked the still-default .local context as explicitly chosen and the
legacy-context migration that ran immediately after did nothing. Upgrading
users with a selected server account silently lost it. Only setContext,
applySyncedContext and the migration itself record a choice now.
An Apple Account change re-resolved the credential scope but never told
AccountsManager to reload its identity-namespaced storage - the only reload
was a one-time launch task - so the previous identity's accounts, context
and outbox stayed live under the new one, as did cached per-account
subscription providers. The scope now posts when its resolved value
changes; the accounts store reloads and the provider cache is dropped.
Fourteen call sites still read the pre-revamp instance-keyed Keychain item,
which a new login no longer writes: reopening the source screen showed the
account as signed out, and feed, browse, home and import operations treated
it as unauthenticated. They now resolve the credential through the account
layer, preferring the selected context's account and otherwise any
signed-in account on that instance.
Signing in while the identity scope was still unresolved created the
account in a temporary namespace that reloadForCurrentIdentity later
discards, while credential storage silently refused the write and the UI
reported success. Login is gated on a resolved scope and verifies the
credential actually stored before reporting success, rolling back a
newly-created account if it did not.
SubscriptionService held one provider per backend and each resolved its
instance from settingsManager.subscriptionAccount, falling back to "the
first enabled Invidious instance" when that was unset. With several logins
possible on one instance that model breaks twice: two accounts would share
a single in-memory subscription cache, and the fallback could send an
account's request to a server it holds no credential for.
Providers are now built per account and cached by account ID, and each
resolves its instance and credential from the account it was built for -
no fallback. The protocol's accountType becomes a DataContext, which is
what actually identifies a provider now.
Views that branched on "is the subscription account local" read the context
instead, and the settings picker writes both stores. Keeping the legacy
setting in step is deliberate for now: several views still read it, and
letting the two diverge would show one account in the UI while serving
another's data. The Phase 3 switcher removes the last readers.
Signing out discards the account's provider so its cached subscriptions
cannot be served after the account is gone.
Yattee 2 could hold at most one login per instance: a Keychain item keyed by
the instance UUID and a Bool on the instance. This adds the account layer
the rest of the revamp builds on.
DataContext (.local or .account) replaces SubscriptionAccount as the thing
that selects a data set. The legacy value maps onto it as an identity
mapping, because migrated accounts deliberately reuse their instance UUID.
Contexts carry an app-managed changedAt, and a migrated one is stamped at
the epoch so any explicit choice on any device outranks it whatever order
they arrive in.
AccountCapabilities makes what a backend supports a single switch in
InstanceType.accountCapabilities, verified against each API surface -
Invidious has history and playlist privacy, Piped has neither. Call sites
ask for capabilities instead of switching on instance type, which is the
seam the planned per-user Yattee Server accounts land on.
AccountsManager owns accounts and the selected context, stored per identity
scope so a switch of Apple Account cannot show or upload the previous one's
list. Legacy single-logins become accounts through a repeatable
reconciliation rather than a one-shot at launch, because instances arrive
over iCloud whenever they arrive; a permanent marker set is what makes
re-running safe, since usesAccountLogin stays true on the instance forever
and would otherwise recreate a deliberately deleted account. A dirty-flag
outbox records what still needs uploading so an offline change survives a
crash. The legacy subscriptionAccount selection is migrated one way,
reading local defaults first and then the raw KVS value - the previous
release published that key to iCloud but never read it back, so on a fresh
install it exists only there.
DataContextValidator replaces SubscriptionAccountValidator, which corrected
any selection it could not immediately verify. That was safe when a context
was just "an instance I am logged into", but with credentials arriving
asynchronously it would discard the user's choice during a slow Keychain
sync. Validity is now tri-state: anything that might still resolve keeps
the selection and falls back to local for reads only, with server mutations
disabled rather than silently retargeted at local data; only an actual
deletion corrects the stored context.
EditSourceView's single log in/out pair becomes a list of accounts with
per-row status, sign-in-again and sign-out-and-remove, plus the explicit
"use saved login from this device" action - the one-tap adoption that
replaced automatic adoption after the Keychain provenance measurement. On
tvOS a credential-less account offers a local sign-in instead, since Apple
never syncs app Keychain items there and the state will not resolve on its
own.
InstanceLoginView now returns the identifier alongside the credential so
the caller can file it in the Keychain beside the secret, keeping login
identifiers out of synced metadata.
Seventh review pass. No correctness findings this round; both are
availability.
The retry budget was only reset inside the tokenless branch of an identity
change, so an account that exhausted its four attempts left the next one
with none - its queued resolve() did no lookup at all and credentials
stayed unavailable until another foreground transition. The budget belongs
to the identity being resolved, so it now resets on every observed change.
ubiquityIdentityToken was archived with requiringSecureCoding: true, but
Foundation guarantees only NSCoding for it. Measured: a class adopting only
NSCoding does throw under secure archiving, and the `try?` there would have
silently produced no fingerprint - losing the scope cache, forcing a
CloudKit lookup every launch and leaving credentials unavailable offline.
This Mac's token happens to support secure coding, so nothing was failing
yet; the archive is only ever hashed, never decoded, so secure coding buys
nothing regardless. Also measured: the non-secure archive is deterministic,
which is what the fingerprint comparison depends on.
Note the digest differs between the two archiving modes, so any scope
cached before this change resolves once more against CloudKit and re-caches.
Sixth review pass, both findings in the tokenless path the fourth pass
introduced.
The staleness guard compared fingerprints, which is useless when there is
no ubiquity token: nil before the switch and nil after it, so a user-record
lookup started under account A was accepted as current after CKAccountChanged
had already announced B. The scope then read as resolved, the pending-resolve
replay was skipped because it only runs while unresolved, and B was left
pointed at A's credential namespace. Every async path now captures a
generation counter that every observed account change bumps, and refuses to
apply a result across it.
Foreground revalidation also kept serving the old tokenless scope across
both awaits, and indefinitely when the account status came back
undetermined - so a switch that happened while the app was suspended could
hand A's retained credentials to B. The scope is now dropped before the
confirmation rather than compared after it, which also let the bespoke
revalidation logic collapse into resolve(), where the signed-out,
signed-in and undetermined cases are already handled. A tokenless setup
briefly reports "can't tell yet" on each activation; the tri-state UI
covers that, and the alternative traded a real cross-account disclosure for
the convenience.
Fifth review pass, which prompted measuring whether SecItemDelete can be
conditioned on a value. It cannot: a delete issued with a deliberately
wrong kSecValueData still returns 0 and removes the item, so there is no
compare-and-delete and every check-then-delete is a race by construction.
That measurement decides both fixes here.
Adoption no longer rolls back its own write when the verification read
comes back empty. The rollback could only ever be a blind delete of
whatever occupied the key by then - including a credential synced from
another device in the gap - and it would propagate. Nothing is deleted; no
marker is written, so the account stays signed out and an ordinary sign-in
recovers it.
Legacy deletion re-reads immediately before deleting and aborts if the
value moved, shrinking the window to a single call gap and catching a
replacement that landed while the fingerprint was being computed. A
residual race remains and is documented at the call site: closing it is not
possible with these APIs, and never deleting would leave a credential the
user explicitly removed sitting in their iCloud Keychain.
The debug provenance report repeated the same nil-token ambiguity the scope
logic was fixed for, reporting "no Apple account signed in" when iCloud
Drive is merely off. It now reports the token as unavailable and shows
CloudKit's account status separately, so a re-run cannot draw a false
conclusion from it.
Fourth review pass, on a hole the previous commit opened.
Making a nil ubiquity token mean "unresolved" fixed the shared-local-scope
collision, but left a scope resolved from CloudKit alone with nothing local
to watch: with iCloud Drive off the token stays nil across a switch from
account A to B, so NSUbiquityIdentityDidChange never fires and the scope
kept pointing at A - and its retained credentials with it - until relaunch.
Two additions. CKAccountChanged is now observed alongside the ubiquity
notification, and a tokenless scope is dropped outright when either fires
because there is no fingerprint to re-derive it from. Foreground activation
also revalidates a tokenless scope against CloudKit, covering a switch
whose notification never arrived; it compares before replacing so the
common case never blanks out credentials, and an unreachable account leaves
the scope standing rather than flapping on every transient failure.
A token-derived scope is untouched by all of this - it changes only via the
notification, and a test pins that down.
Third review pass.
adoptCurrentIdentity() treated a nil ubiquityIdentityToken as sign-out and
scoped credentials to "local". But Foundation also returns nil when iCloud
Drive is unavailable or switched off, and Yattee declares only the CloudKit
service with no ubiquity container, so the token tracks iCloud Drive state
rather than the account. In that configuration two different signed-in
Apple accounts both resolved to "local" and shared one credential
namespace - since Keychain items survive an account switch, account B could
then read account A's retained credential. That is the precise failure this
whole change exists to prevent.
A nil token now means unresolved, and CloudKit's account status settles it
during resolution: noAccount scopes to local, available proceeds to the
user-record lookup, anything else stays unresolved and retries. A scope
resolved without a token is not cached, because there is no key that could
be trusted to mean this account.
Adoption markers are now namespaced by Keychain service. Both credential
managers share one marker store and an instance UUID can outlive a change
of backend, so adopting an Invidious credential could suppress the Piped
offer for the same instance and point deletion at the wrong secret.
Writing the account-status branch also surfaced a Swift trap: an
unqualified `break` inside the switch left only the switch and fell through
to the user-record lookup, resolving a scope on an undetermined account
status. The loop is labelled now, and the test that caught it is included.
Two more issues from a second review pass.
Adoption verified its write by reading the key back and deleted it on any
mismatch. But a mismatch can also mean iCloud Keychain delivered a real
credential for that key between the add and the read - and since the delete
spans both synchronizable forms, it would propagate and revoke that
credential on every device. A readable value that is not ours is now left
strictly alone; only an unreadable key (the add reported success but
produced nothing usable) is rolled back so the offer can be retried.
An identity switch arriving while a resolution was in flight was dropped:
the observer's resolve() returned early on the isResolving guard, and the
in-flight one then bailed on its fingerprint check without starting a
lookup for the new identity, leaving credentials disabled until a
foreground transition or relaunch. Requests that arrive mid-resolution are
now replayed once the running one finishes.
Adds a fingerprint provider and an injectable backoff so identity
resolution is testable without a real Apple account or real seconds, plus
tests for retry, foreground recovery, mid-flight switching, signed-out
scoping and the cached-scope fast path.
Four issues from an external review of the previous commit.
The scope was resolved once at startup, so signing out or switching Apple
account while the app kept running left it pinned to the old identity -
reads and writes would then reach the previous account's retained
credentials, which is the exact leak scoping exists to prevent.
CredentialIdentityScope now observes NSUbiquityIdentityDidChange and
re-resolves.
Legacy deletion authority checked only the instance and the identity, not
the account. On an instance with several accounts, deleting any of them
satisfied the marker lookup and could retire the legacy item another
account had adopted, logging out devices still on the previous release.
The marker's accountID is now required to match.
A first-time resolution that failed (launched offline, CloudKit
unreachable) was never retried, disabling credentials for the rest of the
process. resolve() now retries with backoff, and resolveIfNeeded() runs on
foreground activation.
Adoption re-checked nothing at execution time, so a scoped credential
delivered by iCloud Keychain between rendering the offer and tapping it
would be overwritten with the older legacy secret. Adoption now re-verifies
eligibility and writes with a non-replacing add, leaving a duplicate
strictly alone.
Adds regression tests for the sibling-account deletion and the adoption
race.
Yattee 2 keys credentials by instance UUID with no notion of which Apple
account they belong to, and reads them back with SynchronizableAny. Apple
documents that Keychain data may stay on device when a user signs out and
chooses to keep it, so whoever signs in next would be served the previous
user's credential - guaranteed for migrated accounts, since local instances
survive the switch and the migrated account UUID is deterministically the
instance UUID.
Credentials are now keyed "<identityScope>|<accountID>", where the scope is
the CloudKit user record (identical across a user's devices, so iCloud
Keychain sync still works). Usernames get the same scope and live beside the
credential, keeping login identifiers out of synced metadata.
The legacy unscoped item is never adopted automatically. A strict
kSecAttrSynchronizable == true query was the candidate discriminator for
"this secret belongs to the current Apple account", but measured on iPadOS
27.0 a retained item keeps the attribute and still matches that query with
no Apple account signed in at all - the flag records what was once requested,
not live iCloud Keychain membership. Reads therefore never fall back to the
legacy key; adoption is an explicit user action that copies the item,
verifies the scoped read-back, then records a marker bound to the identity
and a salted fingerprint of the secret. The legacy item is retained, because
deleting a synchronizable item propagates and would log out devices still on
the previous release mid-rollout. Deleting it later needs all three of: a
marker on this device, matching identity, and the item still holding the
adopted secret.
The instance-keyed path is untouched, so no shipping call site changes
behavior. BasicAuthCredentialsManager stays instance-keyed - proxy auth
belongs to the instance, not to a user account.
While the scope is unresolved (first launch under a new Apple account,
CloudKit unreachable) reads report "can't tell yet" via isScopeResolved
rather than "signed out", and writes are refused rather than guessing a
scope. Login state is cached per scope so one identity's cache can never
answer for another.
Also adds the debug-only harness the measurement was taken with, under
Advanced > Developer > Experiments.
URLRouter.parseTimestampValue used TimeInterval(_:) for the plain-numeric
branch ("90", "90.5"). That initializer also accepts the special tokens
"inf"/"infinity" and "nan", and the value infinity compares as >= 0, so a
deep link or share URL carrying ?t=inf parsed to Double.infinity and was
forwarded to the player as a seek target. Seeking to infinity is undefined
and breaks playback startup for the affected link.
Reject non-finite values explicitly alongside the existing negative
check, so only finite non-negative seconds are accepted.
- Add isFinite guard to the plain-numeric branch of parseTimestampValue
- Cover inf/infinity/nan/negative rejection in URLRouterTests
PlaybackRate.displayText/compactDisplayText used String(format: "%.2gx"), whose significant-figure notation dropped meaningful digits (1.25 -> "1.2x") and rounded others (1.75 -> "1.8x"). Replace with a fixed two-decimal format + trailing-zero strip so whole rates have no decimal (2 -> "2x"), halves show one (1.5 -> "1.5x"), and quarter-steps are preserved (1.25 -> "1.25x"). Adds a Swift Testing suite (PlaybackRateTests) proving red->green.
Timed links arrive wrapped as yattee://open?url=... from the share
extension, but the timestamp was parsed from the wrapper URL, where a
single-? form like youtu.be/ID?t=N hides t inside the url query item.
With no startTime the player fell back to saved watch progress.
- Add URLRouter.unwrapped(_:) resolving the wrapper to the inner URL
(raw remainder after ?url=, since the share extension does not encode
& and URLComponents would drop &t= parts) and use it in handleDeepLink
- Thread forceStartTime through openVideo/playPreferringDownloaded/play
so explicit link timestamps beat the 90%-watched restart threshold,
clamped to just before the end; resume flows keep the old behavior
- Carry the parsed timestamp through OpenLinkSheet's play action, which
dropped it entirely
- Cover timestamp parsing and wrapper unwrapping in NavigationTests
When queue mode was set to repeat one, the "Playing in N seconds"
countdown still appeared after a video ended, showing the next queued
video - then restarting the current one anyway. The video-end path
never consulted the queue mode; only playNext() did, after the
countdown had already run.
- PlayerService.backendDidFinishPlaying: restart immediately in repeat
one mode, regardless of queue contents, player visibility, background,
or PiP. This also fixes repeat one never looping with an empty queue.
- ExpandedPlayerSheet.isAutoPlayEnabled: false in repeat one mode, which
gates all countdown triggers (start on ended, overlay render, re-arm
on appear) on iOS/macOS.
- TVPlayerView.handleVideoEnded: early-return in repeat one mode so
tvOS shows neither the countdown nor the replay controls.
The remote control loadVideo command only carried the raw video ID and
an instance URL, so extracted videos (Twitch live streams and other
yt-dlp sites) failed on the receiver: it tried to fetch the ID from the
/api/v1/videos endpoint, which rejects non-YouTube IDs.
The command now carries the full ContentSource and title. The receiver
rebuilds the VideoID from the source and opens a placeholder Video -
the player then re-extracts streams and full metadata from the original
URL, same as when opening the video locally. Media-source extractors
(WebDAV/SMB/local) keep using the existing UUID:path branch.
Both fields are optional for protocol compatibility: old receivers
ignore them, commands from old senders keep the legacy behavior.
Also stop sending a start time for live streams - there is no shared
timeline, the receiver joins at the live edge.
The previous fix rebuilt the quality chain when converting persisted
models back to videos, but many consumers discarded it again by
rendering bestThumbnail (usually a maxresdefault.jpg that 404s for
older videos) as a single URL with no fallback.
Extract the retry logic from VideoThumbnailView into FallbackLazyImage
and use it at every in-app site that can iterate: player thumbnails
(mini bar, expanded sheet loaders, autoplay previews, tvOS audio-mode
artwork - previously a silent black screen), video info card and tvOS
header, and the tvOS playlist cover.
Sites that fetch or send exactly one URL are rewritten to the
always-available hqdefault variant via Thumbnail.reliableURL (exposed
as Video.reliableThumbnailURL): Now Playing artwork, Top Shelf
snapshots, remote control state, the frozen transition thumbnail,
blurred info background, navigation covers, and playlist covers
derived from a video's first thumbnail in Invidious/Yattee Server
responses.
Also invert RecentPlaylist's upgrade helper, which rewrote covers *to*
maxresdefault, walk the quality chain when caching download thumbnails
for offline artwork instead of giving up after one 404, and expand the
remaining single-thumbnail Piped conversions into full chains.
Persistence models (local playlists, watch history, bookmarks,
downloads) store only the best advertised thumbnail URL - usually
maxresdefault.jpg, which doesn't exist for many older videos and 404s.
Live API results survive this because views fall back through the full
quality chain, but toVideo() rebuilt videos with that single dead URL,
leaving placeholder covers.
Reconstruct the quality fallback chain from the stored YouTube-style
URL at read time (Thumbnail.fallbackChain), and rewrite playlist list
covers to the always-available hqdefault variant
(Thumbnail.reliableURL), since those views render a single URL without
fallback. Also expand the fabricated maxres URL in PipedAPI into a
full chain.
With "Partially Watched Videos" set to Ask, opening a live stream showed
the resume sheet with a playback timestamp. A live position is relative
to the live edge, not a fixed timeline, so it is meaningless as a
playhead.
Live views are still recorded in history, but carry no resume position:
- WatchEntry gains a persisted isLive flag plus recordLiveWatch(), which
stores no position and clears state a drifting HLS duration may have
left behind. progress is always 0 and toVideo() propagates the flag so
history rows render the LIVE badge.
- from(video:) clamps duration, keeping Piped's -1 live sentinel out of
storage.
- DataManager routes live saves through recordLiveWatch() and clears the
flag on the non-live path (before updateProgress, so the auto-finish
check is not suppressed by the hard-0 live progress), letting an entry
heal once the stream becomes a VOD.
- PlayerService skips completion saves, always starts at the live edge,
and drops any startTime passed for a live video.
- The resume prompt, the "Continue at" button label, thumbnail progress
bars, "X remaining", Continue Watching and tvOS Top Shelf all exclude
live entries.
- CloudKit syncs isLive with the rest of the watch state: the conflict
resolver carries it with watchedSeconds/isFinished and the sync engine
applies it when merging into an existing entry. It is read leniently,
so the schema version stays at 2 and older clients keep parsing these
records.
Adds a "Load subtitle from file…" row to the Subtitles section on iOS
(document picker) and macOS (open panel), available when playing files
from local folder, SMB, or WebDAV sources. The picked file is copied
into the per-video temp subtitle directory, registered as a selectable
Caption (displayed by filename), and activated through the existing
loadCaption flow. The Subtitles tab and captions control button now
also appear for media-source files without any subtitles.
Streams built for local folder files carry no resolution/codec/fps
metadata (unknowable before demux), so the quality selector labeled
them "Unknown". Use the video track mpv reports (demux-w/h/fps, codec)
to fill those fields into a display-only copy of the stream — the
Video row and summary now show e.g. "720p · 30fps" with a codec badge.
Selection and tap handling keep using the original stream.
Files with multiple embedded tracks (e.g. MKV from media sources or
downloads) previously exposed none of them: mpv's track-list was never
read, aid was never set, and sid was only ever no/auto. The quality
selector showed no Audio or Subtitles tab for such files.
MPVClient now observes track-list/aid/sid and delivers a coalesced,
leniently decoded [MPVTrack] snapshot; selection state is derived from
the reported "selected" flags. Embedded tracks are switched live via
aid/sid (no reload), surfaced in the existing Audio/Subtitles sections
alongside external captions on iOS, tvOS and macOS. External sub-add/
audio-add tracks are filtered out to avoid double-listing. User picks
are sticky across same-video reloads (quality switch, audio mode,
retries) and preferred audio/subtitle languages auto-select a matching
embedded track once per load, with external captions taking precedence.
When an Invidious/Piped account is active, several views read the local
SwiftData subscription store instead of the account, so counts
contradicted each other (Home tile said 10 while the Channels list
showed the server's 2):
- Home Channels tile now uses the provider's count via new
SubscriptionService.cachedSubscriptionCount (fetched once in the
background when the server cache isn't populated yet)
- Export footer and export content use the active account's list;
export runs async with a spinner and surfaces fetch errors as a toast
- CSV/OPML import routes through SubscriptionService.importSubscriptions,
subscribing on the server for server accounts instead of silently
writing to the invisible local store
- ChannelView subscribe-state is corrected via the provider after the
optimistic local-store read; unused *Sync write helpers removed
- Server-account subscribe/unsubscribe/import now post
subscriptionsDidChange so other views refresh
- New "Delete Local Subscription Data" section in Subscriptions
settings (visible with a server account) clears the local store and
queues CloudKit deletions so iCloud doesn't restore it
Sideloading tools re-sign the app with a team that cannot register the
iCloud.stream.yattee.app container, and CKContainer(identifier:) fatally
traps when the entitlement is missing. Detect availability by parsing
the embedded provisioning profile, skip creating any CloudKit objects
when the entitlement is absent, and show an explanation in iCloud
settings instead of the sync toggle.
Two build 264 crashes on macOS 26.5.1 hit AppKit's non-finite-frame
trap (_NSViewValidateGeometry) on AVKit's PiP sample-buffer host view
during a routine layout pass. The NaN is produced inside AVKit from
geometry we feed it - either the manual PiP window resize around a
video switch or an infinite container rect copied into the sample
buffer layer frame.
Add a sanitizedGeometry helper that rejects NaN/infinite or degenerate
rects and apply it at every layer/window geometry write in the bridge
(updateLayerFrame variants, moveLayer, updateVideoAspectRatio,
resizePiPWindow, updateLayerFrameToMatchPiPWindow). Rejected writes are
logged with their call site so the actual source can be identified when
it reproduces, instead of crashing on the next CATransaction commit.
macOS 27 beta routes contentsScale changes (window moving between
screens) through Core Animation's implicit-animation path, which builds
a presentation copy of the layer via init(layer:). Our override re-ran
visual setters on that copy, and -[CAOpenGLLayer setColorspace:]
dereferences render state the shadow copy doesn't own, crashing with
EXC_BAD_ACCESS.
Drop the redundant setter block (super.init(layer:) already copies
presentation values) and disable implicit actions around the
contentsScale writes in MPVOGLView so the presentation-copy path is
never entered for scale updates.
Guard the deinit detach by superview so a container whose render view
was stolen by a newer, not-yet-windowed container cannot rip the shared
view out of it. Narrow the loading-gate heal to .ready only: the old
video still playing during the details/streams fetch can emit .playing,
which would open the gate to stale time updates and break the
stale-error suppression in play()'s catch.
After reinstalling the app and importing sources from iCloud, remote
server instances with Keychain-only credentials showed no indication
that login is needed, unlike WebDAV/SMB sources.
- Show the orange key icon for Yattee Server instances without stored
basic auth credentials (auth is always required for this type)
- Persist usesBasicAuth/usesAccountLogin flags on Instance (synced via
iCloud) so missing Invidious/Piped/PeerTube credentials are detectable
too; maintained on credential save/delete, login/logout, and legacy
migration
- Centralize the check in AppEnvironment.needsCredentials(for:) used by
SourcesListView, MediaSourcesView, and SourceRow
- Backfill the flags at launch from credentials currently in the
Keychain so existing setups publish them without a re-login
Fixes four bugs behind yattee#956 on Apple TV HD (AppleTV5,3, A8,
tvOS 26.5), where video played audio only with a permanently black
picture, never autoplayed, the clock stayed at 0:00, and navigating
during/after playback could freeze the whole app.
1. A8 GL driver deadlock (device-specific). glTexImage2D uploading
float32 data into a float16 texture hangs in libGLImage's
glgProcessPixelsWithProcessor (a dispatch_group_wait that never
returns), permanently wedging mpv_render_context_render. mpv 0.37
(commit 703f1588) turned on dithering (dither-depth=auto) and
LUT-based scalers by default; on the A8 (no GL_EXT_texture_norm16)
both upload float weights via that fatal conversion, and per-frame
CPU plane uploads (videotoolbox-copy / software decode) hit the same
path. Yattee 1.x shipped mpv 0.36 with these off, which is why the
device worked there. Worked around on AppleTV5,3 only: dither-depth=no,
hwdec=videotoolbox (zero-copy via IOSurface), and bilinear scalers.
Confirmed by two on-device lldb backtraces.
2. Swift cooperative-pool starvation (all platforms). MPVClient's event
loop ran as Task.detached and blocked a cooperative-pool thread for
the client's entire lifetime. With two clients alive (active +
pre-warmed) on a 2-core A8, both pool threads were held and every
await in the app starved: the load pipeline froze at its first
Task.sleep, killing autoplay, Now Playing and progress saving. Moved
the event loop to a dedicated Thread; EAGLContext creation now bridges
through GCD instead of Task.detached.
3. Leaked time-update gate (all platforms). PlayerService dropped all
time updates while loadingVideoID was set; when a load task died
mid-flight the flag leaked and the clock stayed at 0:00 during
playback. Heal the gate on the ready/playing transition and leak-proof
play() with a defer; log the previously silent load cancellation.
4. Orphaned player view (all platforms). Detach the shared render view in
MPVContainerView.deinit when no successor container exists, and skip
framebuffer recreation while the view is windowless, to stop a 100% CPU
SwiftUI trait-update loop when opening a settings detail during playback.
The verbose toggles (MPV, remote control) kept working with the master
"Enable Logging" toggle off: the settings UI hides them, but their
stored values - verboseMPVLogging is even iCloud-synced - were still
honored at runtime, producing console output and enabling the periodic
MPV stats collection while logging appeared disabled.
Enable Logging now takes precedence everywhere:
- LoggingService gates its DEBUG-build OSLog console output on the
master toggle (it previously ran before the isEnabled guard, so the
Xcode console was never silent)
- MPVLogging requires loggingEnabled && verboseMPVLogging, which also
gates the 10s playback stats task via MPVLogging.verboseEnabled
- rcDebug in LocalNetworkService and RemoteControlCoordinator requires
the master toggle alongside verboseRemoteControlLogging
Build 264 stutters for a split second every ~10 seconds. The periodic
playback-stats task fetches ~30 debug properties in one ~190ms block on
mpvQueue, and MPVClient.render() wrapped mpv_render_context_render in
mpvQueue.sync - so the render thread queued behind the fetch and stalled
~190ms, at exactly the stats interval. Build 261 predates the stats
task, which is why it was unaffected.
libmpv's render API is documented thread-safe relative to the client
API, so render-context calls no longer go through mpvQueue:
- New renderContextLock protects renderContext lifecycle vs use;
render(), renderWithDepth(), renderSoftware(), reportSwap(),
shouldRenderUpdateFrame() and hasRenderContext use it
- Render-update callbacks dispatch on a dedicated renderEventQueue so
frame-ready notifications can't queue behind property/command work
- destroyRenderContext takes mpvQueue then the lock, still excluding
in-flight renders during teardown
- The stats task skips its property fetch on iOS/tvOS unless verbose
MPV logging is on (macOS keeps it for the render watchdog)
- performRender slow-frame warnings now break down render vs present
time and report frame-delivery gaps for future triage
Hide the toggle on macOS, where apps are never suspended and playback
always continues regardless of the setting. Pause playback on iOS when
backgrounding with the setting disabled — the audio background mode
previously kept mpv playing, making the off state a no-op.
Move the continue-playing policy into PlayerService and make the
backend's background render pause unconditional (PiP-gated only):
it is a GPU optimization that should not depend on the setting.
Also ungate the player sheet visibility handlers, which manage
foreground mini player rendering unrelated to background playback.
Audio-only streams never report a video size, so the aspect ratio of
the last played video stuck to the player area indefinitely once audio
mode was enabled. Reset it in play() when the selected stream is
audio-only, and let the macOS player window fall back to 16:9 instead
of keeping the previous lock.
Downloads bypass stream selection and reach play() as ready-made local
file streams, so audio mode never affected them. Apply the mode at that
choke point instead: swap a local video stream for the separately
downloaded audio track, or play a muxed file with the video track
disabled (vid set per-load in MPVClient, never toggled live). The
transform is bidirectional so stored audio-only queue/history entries
restore full video when the mode is off, and setAudioMode now reloads
downloaded content at the current position in both directions.
CloudKit record names carry a source scope suffix so the same channel
or video ID can exist under different sources, but the apply path
matched local entities by bare ID: incoming records could merge into an
entity from a different scope, and a remote deletion of one scope
deleted every local entity with that ID and cascaded scoped deletions
for all other scopes back to CloudKit.
- Match local entities on the full mapper-derived record name when
merging incoming records, materializing records for upload, and
applying remote deletions.
- Remote deletions now remove only the matching entity and no longer
queue CloudKit deletions (no echo, no cross-scope cascade).
- Insert dedupe for CloudKit-applied records is scope-aware, so a
same-ID record from a different source inserts instead of being
silently dropped. Remotely added bookmarks also update the
fast-lookup cache immediately.
- When conflict resolution during a fetch keeps newer local data, the
merge result is queued for upload; previously the server (and other
devices) kept the stale version unless the entity was edited again.
Strict timestamp comparisons prevent re-upload ping-pong between
devices.
The recent-channel resolver copied avatarURLString/providerName and the
recent-playlist resolver copied authorID/providerName - fields that do
not exist in the records the mapper writes - while missing the real
ones (thumbnailURLString, subscriberCount, isVerified, videoCount). As
a result, when the local side won a conflict, those fields silently
kept the older server values.
Also drop the notificationsEnabled preservation in the subscription
resolver: the subscription mapper never writes that field since
notification preferences moved to their own record type.
Naturally finished videos only saved their 100% progress locally
(play() deliberately skips the sync-on-switch when videoEndedNaturally
is set), so the watched state never reached other devices in the
autoplay flow. Save completion through updateWatchProgress, which
queues the iCloud sync.
Also queue the current playback position when the app enters
background, so backgrounding mid-video hands the position off to other
devices instead of only syncing on explicit stop or video switch.
Register pending changes with the engine state instead of keeping
app-managed record arrays, and materialize records from local data at
send time in nextRecordZoneChangeBatch. This fixes several data-loss
paths:
- Pending changes now persist inside the engine state serialization,
so saves queued during the debounce window (or while offline) survive
app termination. Legacy record-name keys are migrated once.
- Transient CloudKit errors (zoneBusy, serviceUnavailable, rate limits,
network failures, limitExceeded, batchRequestFailed) re-register the
change for engine-scheduled retry instead of silently dropping it.
- Conflicts are resolved against the exact record that was sent; the
resolved record (carrying the server change tag) is cached and takes
precedence at the next send, and any newer local edit evicts it.
- Queue methods are now synchronous, removing a task-reordering race
where a fast delete/re-add could end up deleting the re-added entity.
- Records are built fresh at send time, so queued changes no longer
upload stale field snapshots, and batch size limits are handled by
the framework.
Persistent per-platform mode that plays only the audio track to save
bandwidth and speed up loads. When enabled, stream selection picks the
best audio-only stream (preferred language, codec, bitrate) and never
fetches the video URL; toggling mid-video reloads at the current
position using the quality-switch path (live vid=no toggling is avoided
due to A/V desync).
- Toggle row in the player quality/settings sheet on iOS, macOS and
tvOS, next to lock controls
- New audioMode control button type (red when active) for player
controls and mini player layouts
- Picking a video quality explicitly turns the mode off; audio track
picks keep it on and now work while audio-only is playing
- Stream URL refresh preserves audio-only playback instead of
resurrecting video
- Queue preloads and history entries resolved before a toggle are
discarded so selection re-runs with the current mode
- tvOS player shows the video thumbnail instead of a black screen
during audio-only playback
- PiP button hidden for audio-only streams
On macOS 27 beta, mpv's coreaudio AO fails to initialize (channel layout
set rejected with paramErr -50). The existing ao=coreaudio,avfoundation
fallback restored audio, but exposed an mpv bug: init() registers a HAL
hotplug listener (AudioObjectAddPropertyListener on the system object,
with the raw ao pointer as user data) before the step that fails, and
the coreaudio_error path never unregisters it. mpv then frees the ao
and falls back to avfoundation, leaving a dangling listener per
playback start. The next audio device change (AirPods connect,
sleep/wake device republish) invokes the listener with the freed
pointer and crashes with a use-after-free SIGSEGV on the
HALC_ProxyNotification Call Listener Queue (observed in build 262;
faulting address was reused "texture_" shader string memory).
Gate the AO order by OS version: on macOS 27+ use
avfoundation,coreaudio so the failing coreaudio init never runs; older
macOS keeps the native coreaudio first with avfoundation as fallback.
Update the read-only defaults mirror in MPV options settings to match.
Views without an explicit background sat on the translucent window
material, which picks up a wallpaper tint (bluish in dark mode) and
clashed with views drawing windowBackgroundColor, making the video
info header gradient fade visibly mismatch. Add opaqueWindowBackground()
and apply it centrally in NavigationDestinationHandlerModifier to both
stack roots and pushed destinations, plus VideoInfoView for
presentations outside navigation stacks.
Claude-Session: https://claude.ai/code/session_0154KH8RAVAvm6iVanhmoW8o
When "Play in a separate window" is disabled, the expanded player was a
native sheet attached to the main window. AppKit refuses fullscreen with
an attached sheet, so pressing F swapped the sheet for an in-window
overlay before fullscreening — a visible two-step transition.
Present the inline player as the full-window overlay from the start, so
fullscreen is just the window's native toggle. This removes the whole
sheet-swap machinery (toggleSheetFullScreen with its detach timing,
SheetWindowResizer, sheet sizing helpers).
- beginInlineOverlay/endInlineOverlay hide/restore the main window
toolbar and mirror the window's fullscreen state into the renamed
isMacInlinePlayerFullScreen flag via NSWindow notifications, so exits
through Esc/green button/Mission Control stay in sync
- Collapsing exits fullscreen only if it was entered for the video
- Esc in the windowed overlay collapses to the mini bar via a new
onCollapse callback (replacing the sheet's default Esc-dismiss)
- ContentView registers its hosting window (MainContentWindowReader) so
the overlay targets the right window even when the setting is toggled
while the Settings window has focus
macOS 27 beta's AUHAL rejects mpv's channel-layout property with paramErr
(-50), so the coreaudio AO fails to initialize and playback is silent.
Change the macOS ao option from "coreaudio" to "coreaudio,avfoundation"
so mpv self-heals: older macOS keeps using coreaudio unchanged, while the
beta drops to avfoundation (AVSampleBufferAudioRenderer) instead of playing
no sound.
The fullscreen button did nothing in sheet mode: AppKit ignores
toggleFullScreen on a window attached as a sheet, and the old fallback
targeted the key window — the sheet itself. Two other strategies were
ruled out first: fullscreening the sheet's parent is silently refused
while any sheet is attached, and resizing the sheet's backing window to
screen size NaN-asserts inside AppKit's sheet positioning even without
animation.
Working approach: swap the sheet for a plain overlay in the main window
and native-fullscreen the now sheet-free main window.
- isMacPlayerFullScreenOverlay on NavigationCoordinator dismisses the
sheet (presentation binding excludes it; the setter ignores the
dismissal write-back) and mounts ExpandedPlayerSheet as a black-backed
overlay in ContentView
- ExpandedPlayerWindowManager fullscreens the parent after the sheet
detaches, verifies AppKit accepted it (rolls back if refused), hides
the window toolbar (NSWindow keeps it as a strip above the overlay
otherwise) and restores it on exit
- a didExitFullScreen observer restores the sheet on any exit path
(button, Esc, green button, Window menu); closing the player while
fullscreen exits the main window's fullscreen too
- scheduleSharedViewAdoptionRetry on entry re-homes the shared render
view once the dismissing sheet's window loses visibility — its steal
is declined mid-dismissal and nothing else retried, leaving the
overlay black until the next video load
