Play rejected Twitch, 7TV, BTTV, and FFZ in the visible listing fields. Keep iOS aligned with the same card.
Co-authored-by: Cursor <cursoragent@cursor.com>
webview_flutter_android 4.14.0 is the only direct-ish change. Held
secure_storage / plus packages / wakelock 1.7 for the win32 pin.
Co-authored-by: Cursor <cursoragent@cursor.com>
5.1.0 already serialized useNativePlayer: false, so existing installs
need a one-shot migrate. WebView stays as a legacy opt-out.
Co-authored-by: Cursor <cursoragent@cursor.com>
The full release procedure was loading in every Frosty session. Same split
as the dotfiles resync playbook: always-on file stays gotchas, the ritual
lives in docs/release.md.
Co-authored-by: Cursor <cursoragent@cursor.com>
* build(deps): bump Flutter pin to 3.47.0 and Android Gradle stack
Pin Flutter 3.47.0 (Dart 3.13) as the CI source of truth and move the
Android toolchain to the 3.47 template versions: Gradle 9.3.1, AGP 9.1.0,
Kotlin 2.4.0.
App Gradle script matches the 3.47 template (drop kotlin-android /
kotlinOptions, use compilerOptions). Keep android.builtInKotlin=false and
android.newDsl=false so plugins that still apply KGP keep building.
AGP 9 requires libraries to compile against their dependencies' compileSdk;
force plugin modules to 36 via finalizeDsl. Ignore android/build/, which
AGP 9 now writes a problems report into.
analyze, test, and a release APK all passed. 16 KB zip alignment verified.
iOS UIScene is not in this commit: AppDelegate is customized
(CookieExtractorPlugin), so Flutter will not auto-migrate.
* ci: drop unused Claude Code GitHub Actions
Remove the PR review workflow and the @claude mention bot. Neither is
used anymore; the review job just failed this PR on an expired OAuth
token.
* feat(ios): adopt UIScene lifecycle for Flutter 3.47
Adopt UIApplicationSceneManifest and FlutterImplicitEngineDelegate in AppDelegate to support the UIScene lifecycle requirement on iOS.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Stop asking for mod scopes at normal login so existing core-only tokens
never hit a missing-permissions dialog. Gate moderated-channels Helix,
offer an Enable flow for mods/broadcasters, and upgrade atomically.
Co-authored-by: Cursor <cursoragent@cursor.com>
Tapping the display name or surrounding whitespace was opening the web-session dialog; only the checkmark/info icon should.
Co-authored-by: Cursor <cursoragent@cursor.com>
Existing tokens lack the newer mod scopes, so Helix 401s looked like a vague permissions failure. Cache scopes from /oauth2/validate and show which ones need re-authorization.
Co-authored-by: Cursor <cursoragent@cursor.com>
`--delete-conflicting-outputs` was removed in build_runner 2.15 and is
now ignored with a warning; conflicting outputs are handled
automatically. Documenting it kept sending agents down a no-op path.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The pubspec pinned Flutter 3.44.5 while the toolchain had moved to
3.44.9, so `flutter pub upgrade` failed version solving outright and no
dependency had been upgradable for a while. Bumping the pin unblocks
resolution (CI reads the same field via flutter-version-file).
Notable bumps: dio 5.9.2 -> 5.11.0, firebase_core 4.10 -> 4.13 (plus
crashlytics/analytics/performance), app_links 7.2.1, flutter_markdown_plus
1.0.12, intl, path_provider, webview_flutter, flutter_cache_manager.
dio 5.11 adds DioExceptionType.transformTimeout, which broke the
exhaustive switch in BaseApiClient._handleError. Grouped it with the
other timeout types and covered all four mappings with tests, which
were previously untested.
Raised the json_annotation floor to ^4.12.0 to silence a
json_serializable build warning; the resolved version was already 4.12.0.
flutter_secure_storage stays on v9 deliberately. v10.0.0 shipped in 5.0.0
and was reverted in 5.0.1 (ba28e09) while chasing the #484 startup freeze
on low-end Android 14, and upstream #1084 (unrecoverable NPE in
initialize() on Android 13) is still open. v9 pins win32 ^5, so
device_info_plus, package_info_plus and wakelock_plus are held at their
current majors too - their only breaking change was a win32 6.0.0 bump
that affects Windows only.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reverts the tracking-issue job from 601088b: a single maintainer watching
the Actions tab doesn't need a bot filing issues at them, and GitHub
already emails on scheduled workflow failure.
Also moves the heartbeat from weekly to monthly. Nothing it watches --
cert expiry, secret rotation, dependency drift -- moves on a weekly
timescale, and a dispatch run before cutting a release is the check that
actually matters.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
fastlane warns on every run that it will soon require Ruby 3.3.0 or
newer; both release workflows pinned 3.2. Bumping now means that break
lands on a quiet commit instead of mid-release. fastlane 2.234 already
supports 3.3 and no gem in either lockfile is version-sensitive, so this
stands alone — the lockfiles are unchanged.
Also gives the weekly dry-run somewhere to report. It broke on the
release after 5.1.0 and stayed red for weeks without surfacing, which is
the failure mode of any heartbeat nobody reads: it looks green by
default. Scheduled failures now open a tracking issue, or comment on the
existing one for each week it stays red. Dispatched runs are excluded —
those are already in front of whoever started them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Play rejects a version code it has already seen, and between releases
pubspec's code is by definition already published — so every scheduled
dry-run died on "Version code 69 has already been used" before reaching
any of the credential, signing, or store validation it exists to check.
It has been failing on that since 5.1.0 shipped.
Overrides the build number with epoch seconds, which is always unused and
always increasing. The bundle is validate_only, never published, so the
value is discarded. Only downside is the dry-run can no longer catch a
forgotten version bump, which the real release catches anyway.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
880a9c9 added the handleAndroidPipChanged and _setPipActive actions to
VideoStoreBase without regenerating the mixin, so neither was wrapped by
the MobX action controller. Both still ran, but their observable writes
landed as separate notifications instead of one batched transaction —
extra rebuilds on every PiP transition. Analyze and the test suite pass
either way; only CI's generated-file diff check catches it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Bumps better_native_video_player to the frosty-patches tip, which now
carries the Twitch low-latency work: #EXT-X-TWITCH-PREFETCH segments are
promoted into real segments, EXT-X-TARGETDURATION is rewritten to the true
~2s cadence so media3 stops starving on reloads, and the LoadControl no
longer parks playback ~5s behind after a rebuffer.
Takes native live latency on Android from ~9s to ~3s. iOS is unaffected —
AVFoundation parses HLS internally, so the same rewrite needs an
AVAssetResourceLoaderDelegate to intercept the playlist.
Note the readout: getLatencyToLive() measures distance to the playlist's
live edge, while Twitch's own "Latency To Broadcaster" stat adds ingest and
packaging on top, so the two aren't directly comparable.
Refs #573
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three related PiP gaps on Android:
`handleAndroidPipChanged` lived on `NativeVideoPlayerInterface`, and
`notifyPip` type-checked against it, so the WebView store never learned it
was in PiP — `_isInPipMode` stayed false for the whole session and the
overlay kept auto-hiding behind the PiP window. Moved to the shared
`VideoPlayerInterface` and implemented in both stores. The WebView store's
two JS PiP channel bodies collapse into one `_setPipActive` state machine,
now shared with the Android callback.
`NativeVideoStore` never called `_pip.setIsPlaying()`, so with the
`mediaOnlyPause` layout the PiP window's button was stuck on whatever glyph
it started with, and `onPipAction` — which toggles blind — could do the
opposite of what the button showed.
`VideoStore.handleAppResume()` did nothing in video mode, while the native
store calls `play()`. Backgrounding can suspend the webview's playback with
no way for the page to restart itself, leaving it dead until a manual
refresh. Resumes now unless the user was the one who paused.
Not fixed: PiP with the WebView player. Twitch's embed refuses to play
below 400x300 CSS pixels and a PiP window is far under that, Android
resizes the activity before firing onPictureInPictureModeChanged so any
correction arrives too late, and the embed latches once tripped. The native
player has no such constraint. Refs #563.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`_needsInit` replaced the JS-side `window._injected` guard in 9cec89a, but
the two have different lifetimes. `window._injected` died with the page, so
any page load re-ran `initVideo()`. `_needsInit` lives on the store and was
only re-armed by `handleRefresh()`.
The store outlives the Video widget, which loads `videoUrl` on every mount.
Those reloads hit `onPageFinished` with `_needsInit` already false and
returned early, so the player came back with no play/pause listeners, no
cached `_videoEl`, no autoplay call and no overlay CSS — visibly stuck
paused with a play button that did nothing, unrecoverable without an app
restart.
Re-arm on `onPageStarted` so every navigation re-initializes, restoring the
page-scoped lifetime while keeping the Dart-side control that fixed the
original race.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* feat(chat): add bubbles for shared chat participant channels
Twitch's native shared chat previously had no way to interact with
participant channels the way manually-added chats work. Adds
per-participant bubbles alongside manual tabs: tapping one spotlights
that channel (fading the rest), hiding one filters its messages from
the feed entirely. Defaults to unfaded, resets on rejoin (fresh
ChatStore per connect).
Manual "merged mode" fading is untouched and unaffected.
* refactor(chat): dedup shared-chat tag lookup and extract fade policy
- add IRCMessage.sourceChannelId, replacing three copies of the same
source-room-id/room-id tag lookup
- move the shared-chat spotlight/merged-mode fade decision out of
ChatMessage's build into ChatStore.shouldFadeMessage, matching the
extract-into-store convention already used for SharedChatBubbleState
- factor the tab-bar item's Padding/Center wrapper into a shared
_wrapItem helper instead of duplicating it per branch
* fix(chat): don't spotlight a hidden shared-chat bubble
toggleFocus didn't check hiddenChannelIds, so tapping a hidden
bubble's chip added it to the spotlight set — fading every other
participant to showcase a channel whose own messages were already
filtered out of the feed.
* fix(chat): stop shared-chat detection flapping on system messages
_isInSharedChatMode was recomputed from every PRIVMSG/NOTICE/USERNOTICE,
but only PRIVMSG reliably carries source-room-id for a session's
duration. A sub/raid/announcement without the tag would flip the flag
off and back on within a single live session. Previously this only
risked a redundant asset refetch (already guarded); now the shared
chat bubble bar reads the flag directly, so a flap would visibly
collapse and reopen it. Scope detection to PRIVMSG only.
- add `flutter: 3.44.2` to environment so `flutter pub get` fails locally
the moment the SDK drifts (was already drifting: local 3.44.2 vs the
3.44.1 pinned in CI)
- point all 8 workflow pins at `flutter-version-file: pubspec.yaml` so CI
reads the same field; bumping Flutter is now a one-line pubspec edit
instead of editing 5 workflow files
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
run build_runner in both ci.yml (push to main) and pr-check.yml (PRs),
then fail if git status is dirty. catches forgetting to regenerate a
.g.dart after editing a MobX store or @JsonSerializable model, which
analyze/test/build all miss (as #576 did).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- getModeratedChannels now follows the `after` cursor so moderators of
more than 100 channels aren't silently truncated at the first page
(mirrors getUserBlockedList's throttled recursion)
- add refreshModeratedChannels() and fire it on channel entry so mod
status reflects grants/revokes made since login
- regenerate user_store.g.dart, which also restores #576's moderation
action wrappers and the _moderatedChannels atom (its generated file
was never regenerated when #576 merged)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Freeze the message list on a stationary touch-and-hold (pointer-counted,
touch-slop aware) so an incoming message can't shift the row out from
under a long-press, while still allowing scroll to cancel the freeze.
Co-authored-by: Francis Lavoie <lavofr@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a reusable ModerationActions widget shown to channel moderators (and
the broadcaster) from both the message long-press menu and the user
panel. Supports delete message, timeout (10 min), ban, and remove
ban/timeout — the recovery action appears only when the target is
currently banned or timed out (tracked via CLEARCHAT). Surfaces Twitch's
own error reason when an action is rejected, and forces an immediate
chat re-render so a deleted/banned message reflects instantly.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Wrap the native player in a GlobalKey so an orientation change re-parents
the existing platform view instead of tearing it down and reloading the
stream.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- main: record transient I/O, plugin-teardown, and non-crashing UI errors
(network/websocket/timeout, MissingPluginException, image-load, layout
overflow) as non-fatal, and add a grouping reason to async errors so
distinct errors stop collapsing into one catch-all Crashlytics issue
- irc: guard IRCMessage.fromString against malformed/truncated lines that
threw RangeError, with regression tests
- chat: wrap the emote menu in OverflowBox so it no longer overflows
(RenderFlex) while the panel collapses
- chat: fix onBadgeError returning a List instead of a Map, which threw a
Future.catchError type error during channel asset loads
- video: skip setAutoPipMode on Android < 12 where it isn't implemented
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Squash-merge commit. Covers four areas: native video player, chat reliability, multi-tab, and hardening.
## Native video player
- Direct HLS playback via `better_native_video_player` (forked from `plug-and-pay/flutter-native-video-player`, patched on `frosty-patches` branch)
- Twitch GQL playback token fetching with platform cookie extraction (iOS Swift + Android Kotlin plugins) for authenticated streams
- Quality selection (source, 720p, 480p, 360p, 160p, audio-only) persisted per channel
- Picture-in-picture: iOS automatic inline PiP, Android auto-PiP on API 31+ guarded against older devices
- Latency polling, chat latency sync, stall detection/recovery, startup auto-pause recovery, watchdog timers
- **Twitch stitched-ad break detection** (HLS `EXT-X-DATERANGE`) on both platforms, emitting `adBreakChanged` so the host suspends latency/stall recovery and chat sync during ads (program-date-time is unreliable then). Matches `CLASS="twitch-stitched-ad"` / `ID="stitched-ad-"` **and** the `X-TV-TWITCH-AD-*` attribute prefix (the durable signal streamlink/yt-dlp use)
- Shared `VideoPlayerInterface` so WebView and native paths are interchangeable via settings toggle
- `PlaybackAccessToken` model, `TwitchGqlApi` client
## Chat delay & send reliability
- **Optimistic send-finalize**: Twitch acks a sent message with a per-PRIVMSG USERSTATE (not an echo), which it can intermittently drop — more so under a long chat delay. A missing ack no longer shows a false "message may not have been sent" error; the send is finalized optimistically (local echo rendered), and only genuine rejection NOTICEs surface an error. Ack timeout scaled by the chat delay.
- **Per-line chat-delay bypass**: control/state lines (USERSTATE/ROOMSTATE/GLOBALUSERSTATE), rejection NOTICEs, the reconnect notice, and untagged server lines bypass the delay so send-acks and PONGs aren't stalled; only displayable chat content is delayed, staying in sync with delayed video.
- **Measured chat bottom-bar height**: the message list reflows to the composer's real measured height, fixing the input overlapping messages when a long message expands it (replaces a fixed-constant estimate).
- IRC tag parsing handles valueless and escaped tags per the IRCv3 message-tags spec (fixes a recent-messages RangeError on certain tags).
## Chat / multi-tab
- Redesigned tab chips: unread dot indicator, chips dim until WebSocket connects
- Animated tab options popover replacing `MenuAnchor` (no animation hook in Flutter 3.44)
- Simplified `ChatTabs` from `StatefulWidget` to `StatelessWidget`
- Focus-current-channel fade bump to 0.55; section-header alpha lifted to match the muted-label tier
- `addTab` now calls `setActiveTab` to snapshot outgoing tab's `lastSeenMessageId` — fixes spurious unread dots after adding a tab
- `onActivateSourceTab` threaded into `ChatUserModal` so Reply in merged mode switches to the correct source tab's composer
## Bug fixes and hardening
- **Reverted the experimental iOS low-latency mode** (5s live-edge join + 1.05x catch-up): AVPlayer can't sustain <~18s on Twitch (declared `TARGETDURATION:6` paces playlist refreshes), and the offset starved the buffer into pause-on-launch. Default join distance restored.
- Fork (Android): **cap `BEHIND_LIVE_WINDOW` recovery** so a stream repeatedly falling off the live window can't silently re-prepare forever; emits `error` on exhaustion so the host recovery ladder takes over
- Guard all Android `setAutoPipMode` calls with `_autoPipAvailable` — prevents `PlatformException` on Android 8–11; check `_disposed` after `isAutoPipAvailable` await to prevent post-dispose auto-PiP re-enable
- Remove `_isInPipMode` guard from `_onPaused` so `_paused` tracks actual state, fixing PiP play/pause toggle getting stuck
- Tighten iOS cookie domain check to exact `twitch.tv` / `.twitch.tv` suffix
- Invalidate the stored web session token when Twitch rejects it with a 401 (profile card flips back to unlinked; future loads skip the doomed authenticated request), and re-extract the cookie on launch when logged in without a linked session
- Send `play_session_id` and `player_backend=mediaplayer` to Usher, matching the web client / streamlink
- Fork (iOS): restore program-date-time latency measurement (a prior rev measured distance to the manifest edge, reading ~0s and silently disabling high-latency recovery / chat-delay auto-sync); release platform views on dispose (per-stream view + AVPlayerViewController leak), track the observed AVPlayerItem for KVO removal (sibling-view crash), main-thread seek results, resolve superseded load futures
- Fork (Android): clean up method handlers orphaned by PiP-time view disposal (leaked audio-focus listeners), drop the live-offset re-anchoring seekTo on metadata updates (latency creep), restore pre-duck volume, respect user pause in the stall watchdog, resolve superseded load futures
- Deep review pass: expire the quality-switch flag so stall recovery isn't disabled for the session, cap latency-driven recovery (clock-skew loop), generation-guard the error overlay, two-strike offline confirmation, controller null-safety on quality switch
- Tighten `/me` local-echo match (no longer renders `/meow ...` as an action); clear pending send text on disconnect
## Web / docs
- Remove stale Cloudflare Images binding from `wrangler.jsonc` (covered by `images.unoptimized: true`)
- Correct `web/AGENTS.md` Cache Components claim (feature not enabled)
- Repo and agent config cleanup
## Test plan
- [x] `flutter analyze` — clean (frosty + plugin lib)
- [x] `flutter test` — 319/319 pass (adds `ChatLatencySync`, `StreamInfoPoller`, video timing invariants, `PlaybackAccessToken`, IRC tag parsing)
- [x] Android plugin Kotlin compiles against the bumped fork rev
- [ ] Native player plays live streams on iOS and Android
- [ ] Quality switching works and persists across channel reopens
- [ ] PiP enters/exits correctly; pause/resume toggle in PiP window works
- [ ] Audio-only mode does not crash on Android < API 31
- [ ] Sending under a long synced chat delay no longer shows a false "may not have been sent" error
- [ ] Long composer text reflows the message list (no input/message overlap)
- [ ] Multi-tab: adding a tab does not mark previous tab as unread
- [ ] Merged chat: Reply from user modal targets the correct tab's composer
🤖 Generated with [Claude Code](https://claude.com/claude-code)
The Crashlytics 'Upload Symbols' build phase was hardcoded to the CocoaPods
path ($PODS_ROOT/FirebaseCrashlytics/upload-symbols), but Flutter 3.44 +
firebase_crashlytics 5.x resolve Firebase via Swift Package Manager, so that
path no longer exists and every release build (including fastlane build_app)
failed with 'No such file or directory'. Fall back to the SPM checkout path
when the pod isn't present so the upload works under either integration.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a weekly + manual release-dryrun workflow that exercises the full store
pipeline without publishing: Android validates the AAB against Google Play
(validate_only), iOS builds a fully signed archive via match (readonly) and
skips the App Store upload. Both Fastlane lanes gain a dry_run option that
reuses the real release path so the dry-run stays faithful to it.
Also add release-mode compile checks (flutter build apk/ios --release) to CI
and PR checks to catch release-only breakage, pinning the JDK to match the
release toolchain. Trim ci.yml from a redundant 3-OS matrix to two
purpose-built jobs (analyze/test once on Linux, iOS build on macOS).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Release builds now use the real upload key when key.properties is present,
fall back to debug signing only when ALLOW_DEBUG_SIGNED_RELEASE is set (CI
compile-only checks), and otherwise fail loudly. Lets CI compile a release
APK without production secrets while ensuring a real release can never be
silently debug-signed if the keystore step goes missing.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When a shared chat session ended, the participants' merged emotes and badges
stayed in the channel asset maps, so the other streamer's emote set and badges
kept showing (#522). Only the enter transition was handled.
Detect the leave transition (was shared, no longer shared) and reset: clear the
loaded-participant tracking and re-run getAssets(), which replaces the channel
emote/badge maps with this channel's own assets. Guarded on having actually
loaded shared assets so a flapping source-room-id tag can't cause redundant
refetches.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Workflows used subosito/flutter-action with no version, silently tracking
GitHub's latest stable. When stable rolled to 3.44.1 it broke the build with
no code change. Pin explicitly so builds are reproducible and Flutter upgrades
are a deliberate bump rather than a surprise.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The CI/release toolchain uses the latest stable Flutter, now 3.44.1, which
broke two dependencies:
- simple_icons 14.6.1 extends IconData, now a final class -> bump to 16.23.0.
- extended_text_field 16.0.2 (latest pub release) vendors an EditableText copy
referencing ExtendSelectionByPageIntent, removed in 3.44 -> point at a fork
(tommyxchow/extended_text_field@76305e9) that drops the unused handler, pinned
by commit SHA.
Restores flutter analyze (no issues) and flutter test (all pass) on 3.44.1.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ReorderableListView's onReorder was deprecated after Flutter 3.41 and
fails flutter analyze on the CI's pinned 3.44.1 toolchain. Switch to
onReorderItem, which adjusts newIndex for the removed item itself, and
drop the now-redundant manual off-by-one correction in reorderTab.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The recent-messages API now serializes empty IRCv3 tags as bare keys
(e.g. `emotes`) instead of `emotes=`, which threw a RangeError in the
tag parser and silently dropped every recent message. Parse tags per the
IRCv3 spec: treat bare and empty values as equivalent, split on the first
`=` so values may contain `=`, and unescape values in a single pass
(\: \s \ \r \n). Also guard user extraction against prefixes without
`!`, and isolate per-message parse failures in _handleIRCData so one bad
message can no longer abort an entire batch.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
