Repository navigation
Release 1.1.67 fixes: integrate #1927 #1928 #1929 #1930 #1932 #1933 #1935 - #1934
Merged
Merged
Conversation
Right Option fires hands-free dictation on press. When M, E, or an arrow went down while it was held, the detector only sent .comboInterrupted if CGEventSource.keyState(.combinedSessionState) said Option was down. Our session tap consumes that Option flagsChanged, so the session state never saw it, the check was always false, and every Option+M left a stray dictation running until the next Option press pasted it. HandsFreeModifierComboTracker now follows the held key from the tap's own events with no key-state gate. It still ends on the key's own release and on tap-disable reset, and also once no Option key is left down, so Right Option let go under a held Left Option can't leave it armed.
Setup tells people Screen Memory "reads the window you're replying in", but full-display capture could pull text from other visible windows and apps, including apps outside the Writing app scope. Owner decision: read only the window the user is typing in. - New pure FocusedWindowCapturePolicy picks the capture target at capture time: the typing target's window, owned by the frontmost app, that app's frontmost normal window, not excluded (password managers included), in the app scope, and not contradicted by Accessibility's focused app. Anything unprovable refuses and nothing is captured. - Full-display capture, CaptureKindPolicy and multi-window attribution are gone. The AX reader reads only the app's focused window when its frame matches the chosen one; no other-window search, no unmatched fallback. - The service re-checks the app scope itself and drops held window text as soon as the user moves to another window or leaves the field. - Lock screen, secure input, the master toggle and the any-visible-window exclusion gate are unchanged.
… call timeout behind dictation
- Once "Who was on this call?" naming ends (Done, Later, a newer review) or
something covers it, the island hands key focus back to the app the person
was in, without hiding and without activating Transcripted.
- A dictation opens the island on the display with the focused text field
(one AX read per show, multi-display only), then the pointer's display,
then main.
- The call prompt's timeout holds while it waits behind a dictation
(CallPromptTimeoutClock), as it already did while hovered.
- Global mouse monitor hops with Task { @mainactor } instead of assumeIsolated.
- Clicking a dictation message's words no longer dismisses it.
- Pause for 1 hour is decided when a key is typed. The keyboard no longer captures while paused (it breaks the history segment), and the capture permit refuses during a pause, so text typed in a pause can't reach Save my writing or Personal History when the pause ends before delivery. WritingPausableIngest stays as the second layer. - The completion stream's response waiter always wakes: finish() decides the waiter's outcome under the lock, and a response after finish is refused. The network side sits behind LlamaStreamNetwork so tests can interleave cancel and the response callback deterministically. - The frontmost-window poll (1 Hz CGWindowListCopyWindowInfo) runs only with Autocomplete and Screen Memory on; the app-activation observer only with Autocomplete. Save-only users run neither. - writing-diagnostics.log rolls to .1 at 4 MB. - ModelManager records the launch check's fingerprint, taken from the same locked descriptor before and after the hash, so the first helper handoff skips a second full hash of the model. - Launch reaps an orphaned llama helper once setup is done, whatever the Autocomplete switch says, off the main thread. Same rule as before: only our own binary, re-parented to launchd, sole listener.
MeetingSessionController.init runs on main in applicationDidFinishLaunching, and it called SpeakerEmbedderFactory.makeEmbedder, which loaded ReDimNet2 (MLModel(contentsOf:), plus a GPU compile the first launch after an update) right there. With ReDimNet2 the default, every launch froze the menubar for that long. makeEmbedder now loads nothing. It picks the model, and so the speaker database, from model-file presence and returns Core's new BackgroundLoadedSpeakerSegmentEmbedder, which knows the model's id, size and thresholds up front and loads the real embedder on a utility queue the first time something waits for it: - DiarizationService awaits it during initialize() (the launch warmup), so "meeting models warm" still means the voiceprint is loaded, and again before re-embedding a meeting, so a meeting that races the load gets the right model's vectors. - SpeakerVoiceprintMigration.run awaits it before moving anyone. The gate still closes at launch, so writers stay held through the load. If the load fails, this launch keeps the model's own database and gets no voiceprints (nothing 256-d ever reaches the 192-d file), the migration fails with embedder_unavailable without writing the ledger, and SpeakerEmbedderLoadFailureMemory makes the next launch on the same app build and macOS version use WeSpeaker and speakers.sqlite, the old fallback. A new build or macOS update tries the model again.
- A review confirmation now releases a held voiceprint-migration person in the same transaction, instead of waiting for the next launch's run. Unreadable held ledger rows are logged as a count. - Separation step 1 folds a short voice only when it clears the model's microAbsorb bar; a voice with no fingerprint is never folded. - The invite cap folds a voice only when it clears separationMerge, so a 1:1 invite can't collapse a third person into the invitee. - The separation provider reads the diarizer's active backend per meeting, so a Nemotron load failure gets pyannote's settings.
…eason Review fixes for the focused-window change: - A nil keyboard-focus answer from Accessibility now refuses (keyboardFocusUnknown). Without it a non-activating launcher's typing could read the frontmost window behind it. - Each FocusedWindowCapturePolicy refusal logs a fixed reason string, and those plus no-target-window and target-changed are in the diagnostics allowlist, so skips no longer log as a redacted length. - Stale full-display wording in comments and test titles; ledger records the CaptureTriggerPolicy doc change and the owner decision as a deviation from Tilde.
A dictation error with an action button never auto-dismisses and a model download can run for minutes; holding the call prompt's timeout for all of that would bring back a stale prompt much later. Each prompt now gets two minutes off screen in total, then expires unanswered as before. Hover holds stay uncapped.
…ation' into claude/release-1.1.67-fixes
…floor' into claude/release-1.1.67-fixes
…' into claude/release-1.1.67-fixes
…dow' into claude/release-1.1.67-fixes
…into claude/release-1.1.67-fixes
13 tasks
… into claude/release-1.1.67-fixes
r3dbars
changed the title
build-beta.sh now refuses to build without the Nemotron 3 diarizer and the ReDimNet2 voiceprint model, and the Release Candidate runner has neither. scripts/release/provision-release-models.sh fills the runner's FluidAudio cache before build-beta, checking every file against pinned sha256s: - Nemotron fast128 from FluidInference/nemotron-3-diarization-coreml at 25a90f97 (the repo FluidAudio 0.17.0 downloads from), marker ga-2026-09-23. - ReDimNet2 b4 slim from the models-redimnet2-b4-slim-v1 release asset. - Either one from the previous release's app first when it matches the pins, so later releases reuse the same bytes.
…nto claude/release-1.1.67-fixes
r3dbars
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Integration branch for today's 1.1.67 release, so CI runs once on the combined result instead of queuing each PR. Merge commits keep each PR's history (they'll show as merged).
Island focus/screen and the voiceprint off-main load (#1932, under review) join when ready.
🤖 Generated with Claude Code