<details>
<summary>Prompts Used</summary>
User: This checkout corresponds to https://github.com/jimrandomh/faceclaw/pull/27. Review this branch.
GPT-6 Astra:
> Found one issue:
>
> - **[P2] Wait for Whisper’s final result before completing dictation** — [FaceclawVoiceController.java:729](/Users/jbabcock/repositories/g2/faceclaw/App_Resources/Android/src/main/java/com/faceclaw/app/FaceclawVoiceController.java:729). Skipping partial decodes leaves short utterances without text until release, but the dialog still finalizes after its existing timeout. If decoding exceeds roughly 2.7 seconds (native stop’s 1.5-second wait plus the dialog’s 1.2-second timeout), auto-send can abandon a short command or send an incomplete longer one. Continue can append the late editing instruction literally instead of refining the message. I reproduced these outcomes with delayed transcript events. These flows should await recognition completion.
>
> Validation: TypeScript and a clean Android build passed. All 21 test failures also occur on `main`. Actual Whisper latency was not tested on hardware.
<hr>
User: Thanks, let's fix that issue.
GPT-6 Astra:
> Fixed. Dictation now waits for native recognition to finish before fallback timers run. Auto-send, Continue, and AI Chat preserve delayed Whisper results; cancellation ignores late transcripts.
>
> TypeScript and Android builds pass. All 11 new regression tests pass. The full suite still has the same 21 pre-existing failures.
</details>
sherpa-onnx's native library already reads an offline Whisper config
(verified against sherpa-onnx/jni/offline-recognizer.cc at v1.13.0);
only the Java wrapper was a package-private stub. This adds a public
OfflineWhisperModelConfig builder and wires base.en (int8) in as a
second on-device model next to Moonshine.
- FaceclawVoiceController: setOnboardModelKind() selects the model
before start("onboard"); defaults to Moonshine. Whisper skips the
700ms live-partial redecode (each call re-encodes the whole utterance
plus ~10s of tail padding) and gates near-silent segments before
decode.
- asr-model.ts: ASR_MODELS registry with per-model download state.
- Settings > Voice: "On-device (Moonshine)", "On-device (Whisper)",
and "OpenAI (Whisper)" for the existing cloud provider (stored
value unchanged); one model row per on-device model.
- Microphones live captions pass "moonshine" explicitly (unchanged).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Nightscout app: Alert thresholds for canula age, cartridge volume, pump
battery, time since last loop, shown in top bar. Option to always show
in top bar when window is closed.
Terminal app: Activity indicators in the app-switcher sidebar.
</details>
<details>
<summary>Prompts Used</summary>
Investigate and fix the bug in https://github.com/jimrandomh/faceclaw/issues/18
> Fixed. Here's what was going on and what changed.
>
> ## Root cause
>
> The issue's diagnosis was right, and the underlying flaw is a bit broader than the holder set: **`captureHolders` is phone-side bookkeeping, but the mic enable lives in the glasses' EvenHub session.** That enable dies — silently, from the phone's point of view — on a transport drop, the charging case, an EvenHub screen-off suspend, or a refused enable (`startG2AudioCapture` returns false when the display path isn't warm, and the Java worker then exits while the JS side still believes it started).
>
> `acquireCapture()` answered "is the mic running?" from the holder set, so a single stranded holder (a Transcribe window that was open when the link dropped) made every later request — wakeword assistant included — return early onto a stream that no longer existed, for the life of the process.
>
> The reported repro hits it via `resumeEvenHubSession()`, which resets `audioCaptureActive` and replays the prelude without re-enabling the mic; nothing on either side cleaned up.
>
> ## The fix
>
> **Java** — expose the real state of the stream instead of inferring it:
> - `FaceclawBleCommunicator.isAudioCaptureActive()` (`App_Resources/.../FaceclawBleCommunicator.java:363`) — true only while the session is up, not shut down, and the enable has been acked. Every teardown path already clears that flag.
> - `FaceclawVoiceController.isCapturing()` — `started` plus the communicator's view, with a lock-guarded `audioStarted` so the mic-warmup window doesn't read as dead.
>
> **`app/native/voice-control.ts`** — the bridge now asks `micIsLive()` rather than trusting the set, and tears a dead capture down before starting a new one. Added `handleSessionEnded()` (parks live holders, drops the local state) and `resumeCapture()` (restarts them on the next session); `releaseCapture`/`stop` clear the parked set. The raw-PCM tap for EvenHub mic apps had the identical staleness bug (`if (this.rawActive) return true`) and is covered the same way.
>
> **`app/g2/dashboard-controller.ts`** — wired the two session boundaries: any phase leaving `connected` parks the capture, and the first composited frame of the next session resumes it (a rendered frame is exactly the signal that a mic enable will be accepted). The screen-off EvenHub suspend is now deferred while a capture is held, so the screen timeout no longer kills a live transcription; a capture that races the suspend is parked by the same path.
>
> So Transcribe recovers on its own after a reconnect instead of sitting on a dead "Listening…", and the wakeword assistant is no longer blocked by anything left behind.
>
> ## Verification
>
> `./build.sh` (Java + webpack) clean, `tsc --noEmit` clean, `npm test` 61/61. Not committed.
>
> The behavior is all glasses-session state, which the test harness (pure modules only) can't reach, so it needs hardware. I wrote `notes/voice-capture-session-resume.md` with the five checks that matter — most importantly: Transcribe open → walk out of range → return → "Hey Even" should now get audio, and Transcribe should resume itself.
</details>
<details>
<summary>Prompts Used</summary>
Thanks. Next up, let's add a "Phone display" section to the Settings app (above the Watch section). We'll add a "Preview color" setting (options: White, Green; default White), which controls whether the mirror display is white/grayscale (best looking) or green (matches the physical glasses). We'll move the "Touch mirror" toggle setting (currently on the phone main screen) into the phone display settings menu. We'll add a "Show brightness control" toggle setting (default off) and a "Show screen size" toggle setting (default off), which control the visibility of the corresponding elements on the phone main screen.
Let's also make it so that double-tapping the watch (or simulated watch on the phone, mirror, etc) while the app-switcher is selected turns off the glasses display, in the same way that double-tappign the R1 ring does.
> All four pieces are in and everything builds (webpack, Gradle with the Java changes, and a clean `tsc`). Summary:
>
> **"Phone display" Settings section** — added between Roam and Watch in the on-glasses Settings app, with four entries:
> - **Preview color** (White/Green, default White) — new enum setting `phone.previewColor`. The colorization happens in the Java preview path: `PreviewBitmapUtil.fromGray` gained a `green` variant whose LUT maps luminance to green-on-black instead of grayscale, threaded through `getCompositePreviewBitmap` → `getCompositePreview(green)` → the controller, which reads the setting per refresh. A settings change also schedules an immediate preview refresh, so flipping the color updates the mirror right away instead of waiting for the next composited frame.
> - **Touch mirror** — moved from the phone main screen. It's now a proper `ConfigSettingBoolean` reusing the existing `phone.mirrorTouch` storage key, so anyone who already turned it off keeps that choice. The label+switch pair is gone from the ring row in `remote-controls.xml`, and the view model reads the setting per gesture.
> - **Show brightness control** and **Show screen size** (both default off) — new booleans controlling visibility of the brightness slider/Auto toggle and the display-mode picker on the phone main screen. The whole display-controls row collapses when both are off; the bindings refresh through the existing `onAnySettingChanged` → `refreshDisplayControls` path, so toggling them from the glasses updates the phone screen live. Note this means the main screen is more minimal out of the box — brightness and screen size remain reachable via the Settings app either way.
>
> **Watch double-tap at the app switcher** — the watch-scheme special case in `handleSidebarInput` is removed: a double-click from any source (watch, phone pad/d-pad, mirror double-tap — they all arrive as source `"watch"`) now turns the display off when the sidebar is selected, exactly like the ring's. The original code deliberately excluded the watch on the theory that sleeping was a trap since any watch gesture re-wakes; with your "Watch input wakes display" setting now gated (and double-tap waking a dark display regardless), the toggle behavior is symmetric: double-tap at the switcher sleeps, double-tap again wakes.
>
> Hardware-test notes: the green preview tint and the collapsed-controls layout are worth one visual check on the phone, and the double-tap-to-sleep from the watch is a quick on-device confirm.
</details>