Rotate AI idle suggestions with optional remote packs, move history/dictionary onto self-sizing Home preview cards, harden clipboard capture/prompting, and simplify keyboard chrome by dropping most liquid-glass shadows.
Add the AI keyboard tab, Agent settings, and user-owned LLM key path for 1.7.0, including streaming answers and web-search transports without the built-in DeepSeek fallback.
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.
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.
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.
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+).
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
Security: API key moves from App Group UserDefaults (plaintext on disk)
to the iOS Keychain. The host app writes in Settings; the keyboard
extension reads before each request. Cross-process sharing is via a new
shared keychain-access-group declared in both targets' entitlements.
UX: when the user denies microphone or speech-recognition permission,
the message becomes a tappable row that opens the host app's settings.
Previously the message said "请到「设置」中允许" but the only way to
actually get there was a top-bar ⚙ button that wasn't obviously
related. Auto-clear (2.4s) is now suppressed for .denied so the user
has time to read it. Re-pressing the mic from .denied re-checks
permission so the user can simply press again after granting.
API Keychain migration
----------------------
- New `OSGKeyboardShared/Services/Keychain.swift` — minimal
`kSecClassGenericPassword` wrapper for one item
(service "com.osgkeyboard.apikey", account "current"), backed by
`kSecAttrAccessibleAfterFirstUnlockThisDeviceOnly` (no iCloud sync).
`setAPIKey("")` deletes the entry rather than storing an empty
placeholder so "stored but empty" stays distinguishable from
"not stored" for the noAPIKey error path.
- `ProviderConfig.apiKey` now reads/writes through Keychain instead
of UserDefaults. `didSet` skips the round-trip when oldValue equals
apiKey (init reads Keychain, then assigns — without this guard the
init write would silently re-write the same value).
- One-shot migration: on first `ProviderConfig.init` after upgrade,
a legacy `config.apiKey` UserDefaults entry is copied to Keychain
and removed from UserDefaults. The legacy key is renamed in code to
`apiKeyLegacy` so future reads of `config.apiKey` from UserDefaults
would be a bug.
- `AppGroupStore.apiKey` reads from Keychain (was UserDefaults).
- Cross-process sharing: both targets' entitlements gain
`com.apple.security.keychain-access-groups: ["com.osgkeyboard.shared"]`.
`com.osgkeyboard.shared` is the first entry in both, so it becomes
each process's default access group — Keychain queries don't need to
specify `kSecAttrAccessGroup`.
Permission-denied UX
--------------------
- `KeyboardViewController.pressBegan` now accepts `.denied` and
`.error` as starting states (previously only `.idle`), so pressing
the mic after returning from Settings re-checks permission
without waiting for an auto-clear.
- `scheduleAutoClearError` no longer clears `.denied` — only
transient `.error` is timed. `.denied` is sticky until the user
takes action (taps the row → settings, or presses mic → re-check).
- `TranscriptLine` `.denied` case now wraps the text in a Button
that calls `state.openSettings`, with a `chevron.right` to make
the affordance obvious. The text was shortened to
"麦克风被拒绝" / "语音识别被拒绝" so the chevron has room and
the action isn't implied twice (it was previously both in the
text and via the top-bar ⚙ button).
- VoiceOver hint on the button: "Opens the OSGKeyboard settings
page where you can grant microphone or speech recognition access."
Tests
-----
- New `OSGKeyboardTests/KeychainTests.swift` — 6 tests covering
round-trip, empty-string-deletes, idempotent-delete,
AppGroupStore-reads-from-Keychain, legacy UserDefaults → Keychain
migration, and "Keychain wins when both are present".
- `LLMClientTests` setUp/tearDown now wipes the Keychain
(`try? Keychain.deleteAPIKey()`) and clears `StubURLProtocolStorage`
so tests are independent across runs in the same simulator process.
- All 21 tests pass (6 new + 15 existing).
🤖 Generated with Claude Code
13 items, 555-line diff, build + 15/15 tests green.
ARCH-A3: Phase.error now carries ErrorKind (micDenied/speechDenied/asr/llm/
appGroupUnavailable/unknown) so the UI can pick icons/copy without parsing
free-form strings. Phase.ErrorKind, Phase, LLMError all Equatable.
ARCH-A4: Every TextField in APISettingsCard gets .keyboardType(.asciiCapable)
to defeat SwiftUI's iOS 18 system-keyboard hand-off that auto-suggests
Chinese/emoji and corrupts API keys / URLs / model names.
ARCH-A5 + DOC-3: PrivacyInfo.xcprivacy audited for honesty. Removed three
declared-but-unused APIs (FileTimestamp / DiskSpace / SystemBootTime) and
added ActiveKeyboards (DDA9.1) to the extension (it actually calls
advanceToNextInputMode in the tap path). Main App now declares only
UserDefaults (CA92.1). CHANGELOG updated.
ARCH-A6: Extracted PermissionManager (mic+speech permission flow, iOS 17
branching) and AppGroupPersistor (App Group load/persist) from the God
Object. KeyboardViewController drops 515 → 459 lines. KeyboardPipelineController
left in-place per risk plan — pressBegan state machine is too race-sensitive
to refactor in this pass.
RED-2: Deleted unused Theme enum (no call sites).
RED-3: Deleted unused cardStyle() alias (no call sites).
RED-7: Single source of truth for LLM timeout — LLMClient.requestTimeout +
LLMClientFactory.defaultRequestTimeout; PolishingService derives timeout from
defaultRequestTimeout+1 instead of hardcoding 15.
RED-8: ASRService.transcribe now emits .capability(onDeviceSupported:) as
first event per session; StatusBadge shows REC ⚠️ when the locale fell
back to cloud. New @Published var onDeviceSupported on State.
TEST-1: testPolishThrowsOnTransportTimeout now actually exercises
cancellation: StubURLProtocol delays response 5s, client.polish is
cancelled via Task.cancel(), test asserts the client throws .cancelled /
.transport / .decoding (was: silently passed).
TEST-2: New testPolisherSkipsNetworkWhenModeOff — PolishingService now
short-circuits when modeId == 'off' and returns trimmed input without
invoking LLMClient (proved via injected CountingLLMClient). Service was
moved to OSGKeyboardShared to be reachable from the test target.
TEST-3: KeyboardState (formerly KeyboardViewController.State) extracted
into OSGKeyboardShared so tests can @testable-import it. 5 phase/mode
tests in new KeyboardStateTests. Typealias preserves the old name.
TEST-4: New OSGKeyboardExtTests target with 6 tests covering State
initial values, phase transitions, structured-error round-trip, mode
switching, and InputMode rawValue round-trip.
- AppGroup.defaults: in DEBUG, missing App Group is a hard fatalError
with a precise remediation message (was a soft print + .standard
fallback, which desynced the keyboard extension from the main App).
Release keeps the fallback + NSLog so end-users still get a usable app.
- KeyboardViewController.loadPersistedLocale now prints a masked DEBUG
view of the live App Group config (provider, baseURL, masked key,
model, mode, locale) so the extension's view is visible in the
device console.
- KeyboardViewController.handleFinalTranscript now routes by typed error:
noAPIKey → red error '未配置 API Key · 请在主 App 设置中填写'
http 401 → red error 'API Key 无效 (401) · 请检查主 App 设置'
http 429 → red error 'API 限流 (429) · 请稍后再试'
other → insert raw transcript + generic error badge
- APISettingsCard gains a 'Test connection' button that runs a single
client.polish('ping') round-trip and surfaces the typed result inline.
- PolishingService.timeout raised 12s → 15s to match LLMClient.request
timeout (was racing and discarding successful responses in 12–15s).
- Tests: 4 new cases (HTTP 429, transport timeout, App Group cross-process,
AppGroupStore→LLMClient noAPIKey). All 8 tests pass on iPhone 16e sim.
xcodebuild iOS Simulator: SUCCEEDED
xcodebuild test: 8/8 passed
This is a major rewrite of OpenLessKeyboard, renamed to OSGKeyboard
and rebuilt end-to-end. 59 files changed (+3205/-1550).
Architecture
------------
- Rename project, targets, directories from OpenLess* to OSGKeyboard*
(OpenLess / OpenLessKeyboard / OpenLessShared / OpenLessTests).
- AudioCaptureService rewritten as @unchecked Sendable class with
OSAllocatedUnfairLock instead of an actor, so it survives Swift 6
strict-concurrency checks while still serialising engine + converter
state correctly.
- Single design system (Palette / Spacing / Radius / TypeStyle /
Motion) lifted into OSGKeyboardShared so the host app and the
keyboard extension stay in lock-step.
Push-to-talk — first-principles fix
-----------------------------------
- App Group + audio-input entitlements were stripped by Xcode's
Automatic Signing. They are now declared in project.yml so
'xcodegen generate' re-emits them every time. iOS Developer
Account is untouched; only the App Group capability was added.
- State machine uses a real stored `phase` (was a derived shim
that locked out every press after the first because
recordStream was never nilled after the pipeline finished).
- Microphone permission is requested inside pressBegan (async
Task) so the press flow optimistically enters .recording;
permission denial surfaces a short error and returns to idle.
- Replaced LongPressGesture(0.15s) with a DragGesture +
TapGesture pair separated by pressArmed, so a single tap no
longer fires both onPressBegan and onTap simultaneously.
- Real RMS / peak level meter from the AVAudioEngine tap (was a
pseudo-random walk); the visible waveform is now driven by
actual audio.
- SFSpeechRecognizer(locale:) with selectable ASR locales
(auto / zh-Hans / zh-Hant / en-US / ja-JP / ko-KR) for
first-class Chinese / English / Japanese / Korean dictation,
with on-device recognition when supported.
- AVAudioSession now deactivates on stop so other apps' audio
routing is restored.
Keyboard UI — Typeless-inspired layout
---------------------------------------
- Hero area is 280 pt with a 96 pt record disc, breathing outer
ring, and a 12-bar waveform driven by the real RMS.
- inputView.allowsSelfSizing + a heightAnchor constraint so iOS
no longer crops the keyboard under the Spotlight bar / home
indicator.
- Top bar: mode chip (Off / 转写 / 润色) + locale chip
(Auto / 简体 / 繁體 / EN / 日 / 한) + status badge + ⚙.
- Bottom bar: globe / delete / 空格 / return — all 40 pt and
balanced.
- RecordButton onPressEnded is now safe to fire from a quick
press; pressArmed prevents double-firing.
LLM / Polishing
---------------
- LLMClient: stopped leaking the server response body in errors
(server body is now logged at debug, never surfaced to UI);
added a dedicated .rateLimited case for 429.
- PolishingService timeout 8s → 12s to accommodate slower
domestic LLM providers.
- AppGroupStore.defaultSystemPrompt is now provider-aware
(Chinese for zhipu/moonshot/qwen/deepseek, English otherwise).
Onboarding & Settings
---------------------
- Re-themed OnboardingView / HomeView / SettingsView on the
new design system.
- ProviderPickerSection now shows 6 providers (OpenAI, DeepSeek,
Qwen DashScope, 智谱 GLM, 月之暗面 Moonshot, Custom) with
blurb + selected accent.
- PickerRow for Mode and ASR locale; System Prompt editor with
reset-to-default.
- API settings page "Get an API key" used SwiftUI Link, which
has a hit-test bug on iOS 18 that ate gestures from adjacent
TextFields (manifested as "typing jumps to a website"). It is
now an explicit Button + contentShape + .submitLabel(.done) on
the fields.
Polish & tests
--------------
- LLMClientTests: 4 unit tests passing (ProviderConfig
persistence + OpenAI request/response + HTTP error + missing
key); test App Group renamed to the correct identifier.
- ProviderConfig.apply now captures the previous provider id
*before* mutating, so switching providers actually resets the
system prompt to the new default.
Build
-----
- Swift 6 strict concurrency, iOS 18.0 deployment target.
- Tested on Xcode 26 + iPhone 17 Pro simulator. A real device on
iOS 27 beta aborts with __abort_with_payload (dispatch
library ABI mismatch); use an iOS 18 real device or the
iOS 26 simulator for now.
🤖 Generated with Claude Code
Rocky