Adapt typing/voice surfaces for iPad width and height, add the system globe
key and last-input editing flow, harden host-only Rime deployment, and remove
clipboard voice commands. Bump build to 61.
Keep marketing version at 1.6.6 and bump build to 59. Strengthen never-answer
polish safeguards, clipboard reply-intent continuity, voice undo UX, device
UITests harness, eval fixtures, and What's New assets.
Stabilize clipboard long-press prepare/resume across paste alerts and cold start, add an in-app release-notes sheet with remote bilingual HTML, localize typing input settings, and bump build to 55.
Long-press mic runs ASR as an instruction over eligible clipboard text,
with host-confirm recording UI, min-record gate, and light ASR prewarm.
Bump CURRENT_PROJECT_VERSION to 53.
Custom polish styles can opt in to emotion-matched emoji (default off), with
prompt-level opt-in detection so paste-only styles keep model-added emoji.
Also include Volcengine API-Key ASR auth, voice-processing capture, PiP flash
fix, and related keyboard Shift/haptics reliability work.
Retain the external target app for menu-bar paste and polish context, retry
failed middle ASR chunks once, and stop false-positive path/number violations.
Use redacted cursor neighborhood and pause-aware chunks for more natural polish,
validate protected terms with retry/local fallback, and structure bilingual prompts
for consistency and provider prefix caching.
Harden polish so question drafts stay questions, add local density
routing with style-specific degrade, expand fun style packs, and add
end-to-end FlowTrace logging plus offline guard eval scripts.
Unify Bailian/Volcengine/OpenAI realtime streaming, ABE polish routing with
fun styles, and a shared card-page Settings hierarchy; bump to 1.1 (build 32).
- Add FlowKeepAliveMode (Dynamic Island vs PiP) with iCloud sync
- Settings: keep-alive picker; inactivity timeout only for Live Activity
- PiP: waveform sample-buffer controller, mic released between utterances
- PiP sessions have no idle expiry; user closing PiP ends the session
- FlowSessionManager branches hostReady, session start, and utterance paths
- UIBackgroundModes picture-in-picture; bilingual strings and tests
Co-authored-by: Rocky <hkgood@users.noreply.github.com>
- Accumulate utterance PCM in FlowContinuousCapture for batch retry
- Run full-utterance transcribeChunk when stitched final lags partial
- Refactor Mac MLX tail drain to shared FlowUtteranceEndCoordinator
- Add FlowUtterancePCMStore, UtteranceBatchFallbackPolicy, and tests
Co-authored-by: Rocky <hkgood@users.noreply.github.com>
- Add FlowUtteranceEndCoordinator with 350ms silence drain and 150ms post-roll
- Extend FinalChunkRecovery for short/empty final chunks in chunked pipeline
- Snapshot partial at mic stop and guard final transcript in FlowSessionManager
- Unify tail drain presets (iosFlow/macMLX) and expand diagnostics
Co-authored-by: Rocky <hkgood@users.noreply.github.com>
Add shared prompt composition, custom style management, iCloud sync, and native iOS/macOS selection interfaces so users can keep a consistent writing voice across devices.
Use compliant permission prompts, align cloud ASR privacy disclosures, and default Flow sessions to a shorter inactivity window for safer resubmission.
Replace the Sherpa offline pipeline with native MLX streaming, resilient model downloads, live transcript previews, and supporting tests and documentation.
Add optional StoreKit support tip (features stay free), redesign the
landing page with zh/en + light/dark and product shots, and ship
iPhone/iPad/Mac App Store screenshot sets.
Fix cold-start overlay recursion that overflowed the main-thread stack when
recording began while the ready overlay was still up; also remove temporary
on-screen Flow DEBUG panels after the orange-mic investigation, and land the
macOS overlay/catalog/layout polish plus related Flow recovery hardening.
End Live Activities and release the audio session synchronously on
applicationWillTerminate, and continue macOS local-model install,
onboarding, and settings polish on this branch.
Adds a bundled local ASR model catalog for the macOS app with one-click
Sherpa Qwen3 / SenseVoice downloads (pause/resume, inline actions) and a
shared model storage directory used by MLX Qwen3. Fixes the light-mode
sidebar material and makes the menu-bar icon follow the system appearance
with a refreshed status mark. Renames the built product to OSGKeyboard.app.
Bumps version to 0.5.2 (build 19).
Introduce a standalone macOS menu-bar app (OSGKeyboardMac) that reuses the
platform-agnostic OSGKeyboardShared core: record -> cloud/local ASR -> polish
-> insert. Local mode uses Qwen3-ASR via mlx-swift-asr (macOS 15+, Apple
Silicon); iOS targets stay zero-SPM.
Harden iCloud sync for multi-device correctness:
- Per-field settings merge (appSettings.v2) so concurrent edits no longer
clobber each other's unrelated fields.
- Per-device usage statistics (G-Counter) that sum instead of max().
- Tombstoned dictionary/history merge so deletes propagate and entries can't
resurrect.
- API keys replicate via iCloud Keychain, never iCloud KVS JSON; pulling a
legacy blob without key fields no longer wipes local Keychain entries.
- Add a low-risk "Sync Now" action in Settings.
Fix Flow keyboard mic state: stay orange until the host publishes a real ready
contract, share a single MicVoiceAvailability gate, and self-heal stale
cross-process heartbeat jitter instead of getting stuck.
Extract shared storage (SpeechHistoryStore/UsageStatisticsStore,
ConfigurationStore) into OSGKeyboardShared and add tests for the new
sync/merge logic.
Mirror the personal dictionary through iCloud Key-Value Store with
deterministic merge rules, main-app-only sync UI, and App Group as the
keyboard runtime cache. Add cloud-engine ASR with dictionary bias and
expand the bundled custom language model lexicon.
- P0/P1: tail-drain state machine in FlowContinuousCapture with converter
flush; FlowSessionManager awaits drain before finalize
- P2: FlowCaptureTailDrain policy/tracker, pipeline diagnostics, unit tests
- P3: short final chunk merged re-transcription in ChunkedUtterancePipeline
- P4: UtteranceTranscriptStitcher composedSafely fallback; preview path parity
in LiveDictationController
Co-authored-by: Rocky <hkgood@users.noreply.github.com>
Reduce perceived latency from key release to final text:
- Adaptive chunking: 2.5s first chunk + 5s follow-ups so short
utterances start on-device recognition while still recording.
- Session-level ASR warmup and audio-format cache reuse to remove
per-utterance cold-start of SpeechAnalyzer.
- Mirror live pipelined partials to the keyboard transcript line via
a new flow.transcriptionPartial App Group key + Darwin ping.
Also commits the accumulated custom language model, Flow session,
keyboard extension restructure, and Xiaomi MiMo provider work in
progress on this branch.
Batch of in-progress app work from the working tree.
- feat(keyboard): CursorNavigation + CursorDragPad for caret movement;
KeyboardSoundFeedback for system key click sounds
- feat(dictionary): DictionaryAliasGenerator + PersonalDictionaryEntrySheet;
TranscriptPostProcessor quality gate; retire DictionaryLearner
- feat(ui): TabBarVisibility handling; drop PageHeaderRow /
PageHeaderConfirmButton; refresh views and localizable strings
- fix(security): move the hardcoded DeepSeek key out of
PreconfiguredKeys.swift into a gitignored PreconfiguredKeys.local.swift
(seeded from .example by generate-xcodeproj.sh)
- docs(agents): add Conventional Commits versioning + bilingual changelog rules
- chore(gitignore): ignore PreconfiguredKeys.local.swift, .cache/, pycache
Custom language model / lexicon work stays on
feature/custom-language-model-asr. Changelog bullets added under
[Unreleased]; no version bump.
Two long-standing UX papercuts, fixed without leaving the keyboard:
1. **First-launch onboarding inside the keyboard** — iOS keyboard
extensions *cannot* programmatically switch back to the previous
app after jumping out, so the old flow (jump to host app → user
has to manually navigate back) was 5+ taps of friction. The new
`KeyboardOnboardingOverlay` keeps the user inside the keyboard
for steps 1 (welcome), 2 (mic permission), 3 (speech permission),
and 5 (API key hint). The only step that *must* leave is step 4
("Enable Keyboard"), which jumps to `Settings.app` via
`UIApplication.openSettingsURLString` — on return,
`viewWillAppear` calls `autoAdvancePastKeyboardSetupStepIfNeeded`
which silently advances past that step if the keyboard is now
enabled. Net UX: user types in their app, keyboard walks them
through setup, normal UI appears as soon as setup is done.
2. **Per-app context chip on the keyboard top bar** — the v0.3.0
intelligent-prompt pipeline already adapted tone to detected
context (code/email/chat/document), but without a UI cue the user
had no way to know which mode was active or override the heuristic
when it guessed wrong. The new `AppContextChip` surfaces the
detected context; tap-to-override writes back to
`AppGroupStore.setDetectedAppContext` so the next LLM call
picks up the new tone immediately. Wired into `pressBegan` so
the chip updates in real time as the user types into different
fields.
### Files added
- `OSGKeyboardExt/Views/AppContextChip.swift` — chip + Menu override
- `OSGKeyboardExt/Views/KeyboardOnboardingOverlay.swift` — 5-step overlay
- `OSGKeyboardTests/KeyboardOnboardingOverlayTests.swift` — round-trip + enum surface tests
### Files modified
- `OSGKeyboardShared/Services/KeyboardState.swift`
+ `hasCompletedOnboarding`, `onboardingPage`, `appContext`
+ `setAppContext`, `advanceOnboarding`, `completeOnboarding`
+ `requestMicPermission`, `requestSpeechPermission`, `openSystemSettings`
- `OSGKeyboardShared/Services/AppGroupStore.swift`
+ `hasCompletedOnboarding` / `onboardingPage` accessors (mirror of
`ProviderConfig` keys, so the keyboard extension never has to
instantiate the host-app config)
- `OSGKeyboardExt/KeyboardViewController.swift`
+ action hooks wired (`installStateActions`)
+ `syncOnboardingStateFromAppGroup` / `syncAppContextFromAppGroup`
called on `viewWillAppear` and `loadPersistedConfig`
+ `autoAdvancePastKeyboardSetupStepIfNeeded` for the silent
"jump out → come back" flow
+ `openSystemSettingsFromExtension` opens `Settings.app` via
`UIApplication.openSettingsURLString` (the only system URL
the extension is allowed to open)
+ `detectAndStoreAppContext` mirrors to `state.appContext` so
the chip updates without waiting for `viewWillAppear`
- `OSGKeyboardExt/Views/KeyboardRootView.swift`
+ overlay mounted in `ZStack` over normal UI (animated)
+ AppContextChip in top bar (hidden during onboarding)
- `OSGKeyboardExt/{en,zh-Hans}.lproj/Keyboard.strings`
+ onboarding copy + chip labels (35 keys per language)
### iOS sandbox notes (kept here for posterity)
- Keyboard extensions **cannot** present AVAudioSession /
SFSpeechRecognizer permission dialogs directly. The overlay's
step 2/3 buttons optimistically advance; the actual permission
is granted when the user first opens the host app (which the
step-5 "Open OSGKeyboard" button triggers). This is the same
pattern the previous "jump to host app" flow used — just
without the broken return trip.
- `UIApplication.openSettingsURLString` is the only system URL
reachable from `extensionContext.open`. Both step 4 and the
"Open Settings" button route through `HostAppLauncher` so the
responder-chain fallback also kicks in if needed.
Co-authored-by: Mavis <Mavis@hkgood.dev>
v0.3.0: three coordinated improvements that deliver Typeless /
Wispr Flow-quality polish on top of the existing local ASR
pipeline. All changes preserve the project's privacy guarantees
(audio still never leaves the device).
## 1. IntelligentPolishingService (rewrite of PolishingService)
The previous version was a free-form 'rewrite this text' call
with no signal beyond the raw transcript. The new one is a
single LLM call that does three things in one pass, exactly as
Typeless and Wispr Flow do internally:
1. ASR error correction (homophones, near-misses, missing chars)
2. Polish (drop filler words, fix grammar, add punctuation)
3. Style adaptation per app context (code / email / chat / doc)
The merged-prompt design halves the round-trip vs the previously
proposed two-stage design (correction + polish separately) and
the academic literature confirms it performs equivalently for
everyday Chinese / English dictation.
## 2. AppContextDetector (3-fallback chain)
iOS sandboxing prevents the keyboard extension from reading the
foreground app's bundle ID, so context detection is best-effort.
The detector runs three fallbacks in order, with caching to
avoid the cold-start 'unknown' that would force a neutral-tone
LLM call every time the user opens a new field:
1. Heuristic on the text at the cursor (code / email / chat / doc)
2. 30-minute cache of the last successful detection
3. Time-of-day + weekend heuristic as a soft default
The keyboard extension runs the detector on every press of the
mic and persists the result to the App Group so the host app's
polisher picks it up.
## 3. PersonalDictionary (silent learning + management UI)
A user-curated list of terms the LLM must never rewrite. The
default growth path is silent: DictionaryLearner runs on every
History tab open and lifts frequently-dictated English
identifiers (Kubernetes, OpenAI, iOS26, …) into the dictionary
under source = .history. Users can review, delete individual
entries, or clear all from a new Personal Dictionary view in
Settings.
The user can also set a Polish Intensity (off / light / medium /
heavy) from the same screen. Default is medium, which is what
Typeless and Wispr Flow also use.
## Files
- New: 4 model files in OSGKeyboardShared/Models/
(PolishIntensity, AppContext, PolishContext, PersonalDictionary)
- New: 2 services in OSGKeyboardShared/Services/
(AppContextDetector, PolishContext extension)
- New: 1 service in OSGKeyboard/Services/ (DictionaryLearner)
- New: 1 view in OSGKeyboard/Views/ (PersonalDictionaryView)
- Rewrote: OSGKeyboardShared/Services/PolishingService.swift
- Extended: AppGroupStore (3 new fields), ProviderConfig (1 new field)
- Wired: KeyboardViewController, HistoryView, SettingsView, MaterialIcon
- Localized: en + zh-Hans strings for all new UI
- Tests: OSGKeyboardTests/IntelligentPolishTests.swift (16 tests)
## Verification
- All new code follows the existing Sendable / strict-concurrency
patterns (the keyboard extension stays within its 60MB sandbox;
the polisher remains an actor; @MainActor is applied to the
learner and the settings UI).
- Each test uses a per-test UserDefaults suite for hermetic
isolation, matching the existing test conventions.
- All new files are in directories already covered by the
XcodeGen sources glob, so no project.yml change is needed.
## Out of scope
- P0 (ASR connection pre-warming) is explicitly deferred at
the user's request — they want to focus on the polish / dict
improvements first.
- The Cloud polish (WebSocket) work is not touched.
## Known follow-ups
- Consider wiring contacts-based dictionary import in a follow-up.
- Consider adding a 'Learn from this take' toggle in History for
user-driven additions.
- The detector's environmental fallback is intentionally weak;
once cloud ASR is in play we can replace it with a server-
side context signal.
Add preset-driven polish scenarios (Settings, onboarding, ScenarioChip)
with ScenarioPrompt and style directives; drive keyboard height from
content (240pt) and use viewIsAppearing encapsulated-height offset for
smoother keyboard switches; remove redundant StatusBadge and hide system
dictation via hasDictationKey.
Rolls back the v0.2.0 Qwen3 CoreML on-device ASR stack and replaces the
'local engine' UX with iOS 26 SpeechAnalyzer + DictationTranscriber only.
The 'Cloud polish after ASR' toggle (ProviderConfig.localModeCloudPolishEnabled)
lets users opt into a post-ASR DeepSeek round-trip from the local engine.
Defaults to off so the local engine stays genuinely local. New PolishError.missingAPIError
surfaces an inline 'fill in your key' warning when the toggle is on but the
Keychain is empty. DeepSeek preset default model bumped to deepseek-v4-flash.
Deleted:
- OSGKeyboard/ThirdParty/Qwen3Speech/ (74 files, ~16k LoC)
- OSGKeyboard/Services/ModelManager.swift (492)
- OSGKeyboard/Services/OnDeviceModelWarmup.swift (197)
- OSGKeyboard/Services/Qwen3ASRService.swift (257)
- OSGKeyboard/Services/ModelDownloadSourcePicker.swift (126)
- OSGKeyboard/Views/OnDeviceModelsView.swift (184)
- OSGKeyboard/Views/DownloadConfirmSheet.swift (96)
- OSGKeyboardShared/Models/OnDeviceModel.swift (140)
- OSGKeyboardShared/Services/OnDeviceModelStatus.swift (104)
- Qwen3ASRServiceProvider registration in OSGKeyboardApp
- Qwen3Speech package declaration in project.yml
- 5 .qwen3ASR enum / branch reference sites in HomeView, OnboardingView,
LocalEngineSettingsRows, FlowSessionManager, ASRService, EngineServiceLabel
- Two pre-existing Swift 6 strict-concurrency errors in
LiveDictationController + FlowSessionManager (the weak [weak self] in
detached-task MainActor.run blocks) that were blocking clean builds
Added:
- LocalModelsGroup: 'Built-in iOS SpeechAnalyzer' badge + 'Cloud polish
after ASR' Switch toggle
- PolishingService: honour localModeCloudPolishEnabled; new .missingAPIKey
error case with localised warning
- AppGroupStore.localModeCloudPolishEnabled (mirrored into App Group
so the keyboard extension honours the toggle during live dictation)
- SettingsView: show provider/api sections when local-mode cloud polish
is on so the user can paste a DeepSeek key
- FlowSessionManager: route through PolishingService for local + polish-on
flow; translate missingAPIKey into a polished warning
- KeyboardViewController: handle PolishingService.PolishError.missingAPIKey
in the keyboard-side live polish path
- CHANGELOG v0.2.1: documents the rollback + new toggle
- README.md / README.zh.md: engine matrix section, data flow note
Verified: xcodebuild -scheme OSGKeyboard -destination 'generic/platform=iOS Simulator'
build succeeds under SWIFT_STRICT_CONCURRENCY=complete.
Replace MLX GPU inference with CoreML bundles so transcription continues
while the host app is backgrounded. Adds model download and warm-up,
vendored Qwen3Speech, and updates onboarding, settings, and copy for the
~1.6 GB CoreML package (iOS 18+).
Remove KeyboardL10n hard-coded fallbacks in favor of ExtL10n and extension
Localizable.strings. Add Flow session expiry hints, Darwin cross-process
notifications, session monitor on the keyboard, and app icon on GitHub Pages.
Migrate keyboard dictation to continuous Flow sessions with auto-start,
tap-to-toggle recording, 60s countdown, five-step onboarding, and App Group
IPC. Add docs/ GitHub Pages site with en/zh privacy policy for App Store compliance.
The previous code path for the keyboard preview's ASR controller
cancelled the consumer task at the exact moment it closed the
audio stream:
asrTask?.cancel() // ← kills the .final consumer
asrTask = nil
...
bufferContinuation?.finish() // tells ASR "no more audio"
The cancellation cascaded: the for-await on the events stream
exited → the AsyncStream's `continuation.onTermination` fired →
ASR.cancel() ran → producer task was marked cancelled → the
producer's `if !Task.isCancelled { yield(.final) }` guard
suppressed the .final event. Net result: nobody told the UI to
leave `.processing`, and the disc sat there forever.
Fix (4 changes):
1) `stop()` no longer cancels the consumer. The consumer task
exits naturally when the events stream finishes, sees the
`.final` event the producer still yields, and transitions
the phase out of `.processing`. This is the primary fix.
2) `start()` cancels any leftover `asrTask` at the entry point
as a safety net — covers the "user smashes the disc twice
quickly" race where a previous consumer is still draining.
3) `stop()` schedules a 3-second safety-net Task: if the ASR
pipeline never produces a `.final` (analyzer hang, system
glitch), force the phase back to `.idle` so the user isn't
stuck. Normal recordings complete well under 3 seconds, so
the timeout is only hit on the unhappy path.
4) `KeyboardPreviewSheet` adds `.onDisappear { asr.stop() }`
so closing the sheet mid-recording releases the
AVAudioSession and mic. `stop()` is idempotent (no-op on
non-recording phases), safe to call here.
State machine: `phase = .processing` now has TWO transition
paths out — the consumer receiving `.final` (fast path) and
the 3-second safety net (fallback). Both are required; the
fast path is the common case, the fallback is the
"guaranteed-progress" guarantee.
Testability: `asrTask` was `private`; relaxed to `internal` so
the regression test in
`OSGKeyboardTests/PreviewASRControllerStateTests.swift` can
install a known consumer task and assert `stop()` does not
cancel it. The class is `@MainActor` so Swift 6 isolation
rules still prevent production code outside the class from
racing on it.
Tests:
- `testStopDoesNotCancelConsumerTask` — primary fix regression.
- `testStopIsIdempotent` — `.onDisappear` after a manual stop
doesn't misbehave.
- 27/27 tests pass (25 existing + 2 new).
- BUILD SUCCEEDED.
🤖 Generated with Claude Code
The keyboard preview crashed on first record with a
`__abort_with_payload` deep inside Speech's
`DictationTranscriber`. The disassembly surfaced three
preconditions checked before a `brk #0x1`:
+620 "Audio sample data must be 16-bit signed integers"
+848 "Multi-channel audio is not supported"
+1072 "Client info not fully initialized"
We hit the first one. `DictationTranscriber` (iOS 26's new
`SpeechAnalyzer`-backed engine) is strict about its input
format: only Int16 PCM, not the Float32 PCM that the iOS 18
`SFSpeechRecognizer` path accepted. Our audio-tap and
`AudioBufferSnapshot.samples: [Float]` are Float32 all the
way down — that was the SFSpeech shape, and the previous
`AppleSpeechASR` adapted internally. With iOS 26 as the
deployment target, the only ASR backend is
`SpeechAnalyzerASR`, and the conversion needed to happen at
the `AnalyzerInput` boundary.
Fix:
- `transcribe` builds the `AVAudioFormat` as
`.pcmFormatInt16, 16 kHz, 1 ch, interleaved: true` (the
canonical layout for Int16 Speech input).
- `makeInputStream` runs the per-sample conversion
`Int16(round(clamp(s * 32767, -32768, 32767)))` into the
`AVAudioPCMBuffer`'s `int16ChannelData[0]`. The explicit
clip is required (a `s == 1.5` from a gain-overflow at the
audio-engine boundary would otherwise wrap to a negative
Int16 after the implicit truncation). `round()` (not
truncate) preserves DC balance — `0.5` quantises to
`+16384`, not `+16383`, matching what audio DAWs expect.
- The conversion helper is exposed as
`ASRServiceFactory.convertFloat32ToInt16` so unit tests
can lock the math without instantiating the full pipeline.
Why not change `AudioBufferSnapshot` to `[Int16]` instead
(see earlier first-principles discussion): the snapshot is a
transport format that both `AudioCaptureService` (in the
ext) and `PreviewASRController` (in the main app) produce.
Float32 is the natural shape coming out of `AVAudioEngine`,
and pushing the conversion to the ASR service keeps the
transport contract platform-agnostic — a future second
backend with different format needs can have its own
adaptation without dragging everyone else.
Tests:
- `testFloat32ToInt16EdgeCases` — 0, ±1, ±0.5, ±1.5
(gain-overflow case).
- `testFloat32ToInt16RoundTrip` — quantisation step is
1/32767 (so the asymmetric Int16 range is honoured: -32768
has no exact Float source).
- `testFloat32ToInt16Empty` — `sourceCount == 0` with nil
pointers is a no-op (function guards on count before
dereferencing).
- All 25 tests pass (22 existing + 3 new).
- BUILD SUCCEEDED.
🤖 Generated with Claude Code
1) Preview chips weren't actually buttons.
`modeChip` and `localeChip` in `KeyboardPreviewStub` were
decorative HStacks — no `Button`, no action, no callback. The
chevron-down glyph made them *look* like pickers, so a user
tapping them got nothing. The screenshot the user shared
("润色 ▾" / "中文(简体) ▾") shows exactly that surface.
Fix: wrap each chip in a `Button(action: ...)` with
`.buttonStyle(.plain)`. The stub now takes `modeId`, `localeId`,
`onModeCycle`, `onLocaleCycle` and the sheet's `cycleMode` /
`cycleLocale` advance the config:
- mode cycles [off → transcribe → polish] (mirrors Settings)
- locale cycles [auto → zh-Hans → zh-Hant → en-US → ja-JP → ko-KR]
Mid-recording locale switches call `asr.stop()` because ASR
sessions are bound to the locale they were started with.
`modeId == "off"` also stops any in-flight recording so the
disc isn't recording into a mode that won't insert.
The mode chip's icon also follows the mode (mic.slash /
mic / wand) as a redundant visual cue, and both chips get
accessibility labels (preview.modeChip.cycle /
preview.localeChip.cycle) so VoiceOver users can use them.
2) Onboarding's "Next" stays enabled when local engine is picked
but no API key is filled in. Root cause: `ProviderConfig.isConfigured`
checks `!apiKey.isEmpty && !baseURL.isEmpty && !model.isEmpty` —
it never asks whether the user *needs* a key. The local engine
(on-device ASR) doesn't round-trip through the LLM, so an
empty key on the local path is correct, not a configuration gap.
Fix: short-circuit `isConfigured` to `true` when
`engineMode == "local"`. The onboarding "Next" button is
already disabled on the API page when `!isConfigured`; this
just makes the gate respect the engine choice. New test
`testIsConfiguredTrueForLocalEngineWithoutAPIKey` locks the
behaviour in (local → true, cloud → false, flip back).
3) Add the two new accessibility keys to all four
`Localizable.strings` files (en + zh-Hans, main app + ext)
so VoiceOver and the cycle button labels resolve in both
languages.
Build: BUILD SUCCEEDED.
Tests: 22/22 pass (1 new).
🤖 Generated with Claude Code
Rocky
rocky