connect() registers window surfaces in a loop that reads the foreground
window once, before the loop, then awaits each surface's configure call
before queuing its visibility. On a fresh process restoreOpenApps runs at
the same time and focuses windows as it relaunches them. The bridge runs
Java calls in order, so a visibility decided before the await lands after
the focus change's own calls and undoes them: the old foreground can stay
visible and cover the new one, and a later window switch doesn't clear
it, because a switch only hides the window being left.
configureWindowSurface now also accepts visibility as a function, asked
after the configure call, when the visibility call is queued. The display-
attach loops (glasses and preview), the display-mode, per-app layout and
vertical-position changes, and an in-process window's height-mode change
pass one. Launch paths that pass a literal false are unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Passing a JS ArrayBuffer to a Java ByteBuffer parameter leaks it in
@nativescript/android 9.0.4 through 9.1.1: JsArgConverter wraps the backing store with
NewDirectByteBuffer and takes a JNI global ref that it never deletes. The
ref keeps the ByteBuffer reachable, so makeInstanceWeakAndCheckIfAlive
always answers "alive", the JS finalizer re-arms, and V8 never frees the
ArrayBuffer. submitSurfaceFrame passed a fresh copy of each frame this way,
so every frame leaked a full surface (576x260, 640x452 or 640x480 bytes)
plus one global ref, including frames that Java then discards as unchanged.
The draw-list buffer and buzzer payloads leaked the same way.
Each payload now goes through a reused Java direct ByteBuffer
(app/native/java-direct-buffer.ts): JS writes into it through an
ArrayBuffer.from view, and the Java object itself is passed, which takes
NativeScript's ordinary object path. Every consumer (SurfaceCompositor
applyAndComposite, playBuzzerSequence) copies before returning, so one
buffer per JS context is safe.
Covered: the communicator bridge (frames, draw lists, buzzer, which also
carries EvenHub tones), the preview-only display, and all eight worker
apps. Workers submit through the Android getActiveDisplay() (frames) and
worker-buzzer (tones), so the reused buffers live there rather than in
each worker; the worker sources are shared with iOS, whose active display
takes the ArrayBuffer itself. The
smaller one-off crossings (glyph and image atlas registration, the gray
preview helper, firmware and file writes) are left as they are.
tests/java-direct-buffer.test.cjs covers the helper against a fake
ByteBuffer, and the Android worker-buzzer test stubs it.
(Rebased onto main by James Babcock: worker changes moved into
active-display.ts and worker-buzzer.ts.)
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
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>
awaitEvenHubSessionReady() polls fixedLayoutCreated/displayedFingerprint at up to 100ms
granularity because resolveAckLocked() never notified waiters when the relevant ack
actually arrived. Confirmed live on real hardware (2026-09-08): full resume (prelude +
create-layout + content frame) now completes in 449ms with a clean, tight ack sequence
and no regressions. Safe by construction: the one caller already holds the lock, and
every lock.wait() consumer in this file rechecks its own condition in a loop.
classifyOnboardingFirmware() (firmware-compat.ts) returns kind values of
"custom" | "flashable-stock" | "newer-stock" | "unknown", but
applyClassification()'s switch (onboarding-firmware-check-view-model.ts)
checked for "flashable" / "newer" -- names borrowed from the unrelated
CheckPhase type used a few lines later via setPhase(). Since kind was
typed as plain string, nothing caught the mismatch at compile time.
Effect: any device that successfully reported a firmware version (stock,
flashable-stock, or newer-stock alike) fell through to the switch's
default case and showed "Connected, but couldn't read a firmware
version" -- the hard-block message meant for when the probe gets no
data at all. Confirmed on real hardware (glasses on 2.2.9.22): the BLE
probe read the version correctly every time, logged it correctly, and
the UI still reported it as unreadable.
Fix: match the real OnboardingFirmwareKind strings, and tighten the
parameter's type from `string` to `OnboardingFirmwareKind` so this
category of typo fails to compile instead of failing silently.