Feat/macos chat voice controls

[Yny4ii]

What changed and why

Add cancel-while-recording for chat PTT (mic still sends; cancel trails so the hit target does not jump), a stop-speaking control for spoken AI answers (composers, Escape, agent Stop), and a stable Idempotency-Key on action-item create keyed to the local SQLite row so sync retries do not duplicate tasks. Backend dedupe for that key lands with the web duplicate-task fix branch and is not in this PR.

Product invariants affected

• INV-AUTH-1
• INV-CHAT-1
• INV-VOICE-1
• INV-TASK-1

How it was verified

• Focused Swift tests: PushToTalkButtonTriggerTests, VoiceResponsePlaybackTests, ActionItemCreateIdempotencyTests (and Home ask-bar stop control coverage) — 14/14 PASS.
• Named bundle /Applications/omi-voice-controls.app: bridge ptt_start → ptt_cancel → cancelled=true, terminal_reason=cancelled.
• Pre-push / local PR preflight passed with this body (line-count exceptions + invariant citations).
• Not re-verified in this session: mid-TTS Stop in the live UI (spoken playback was already silent on this machine and on the site download build — treated as environmental, not introduced here). Manual cancel → re-record and live task create with Idempotency-Key still worth a quick pass on reviewer’s machine.

Tests

• Core cancel path + post-cancel restart: PushToTalkButtonTriggerTests (including cancel-trails-mic layout tripwire).
• Stop-speaking helper: VoiceResponsePlaybackTests.
• Idempotency-Key always sent / caller key reused / stable local-row key: ActionItemCreateIdempotencyTests.
• VoiceResponsePlayback.swift covered in floating-bar-functional.yaml for e2e flow coverage.

Failure class (fixes)

Failure-Class: none

Scoped cleanups (optional)

• Pre-push unblock commits: e2e covers for VoiceResponsePlayback, SwiftLint force-unwrap fix in the idempotency URL stub, source-inspection escape placement for the layout tripwire.

Line-Count-Exception: desktop/macos/Desktop/Sources/DesktopAutomationBridge.swift | 4679 -> 4687 | Register ptt_cancel automation action next to existing PTT bridge actions
Line-Count-Exception: desktop/macos/Desktop/Sources/FloatingControlBar/FloatingControlBarView.swift | 3012 -> 3038 | Add interactive cancel on floating/notch PTT recording overlays
Line-Count-Exception: desktop/macos/Desktop/Sources/FloatingControlBar/PushToTalkManager.swift | 2807 -> 2821 | Expose cancelPushToTalkForAutomation for the bridge cancel path
Line-Count-Exception: desktop/macos/Desktop/Sources/MainWindow/Pages/DashboardPage.swift | 4297 -> 4303 | Wire Home ask bar stop-speaking through existing trailing controls
Line-Count-Exception: desktop/macos/Desktop/Sources/Providers/ChatProvider.swift | 6858 -> 6860 | Interrupt spoken playback when the user stops an in-flight agent turn
Line-Count-Exception: desktop/macos/Desktop/Sources/Stores/TasksStore.swift | 3676 -> 3684 | Pass stable Idempotency-Key when syncing local-first task creates

[Git-on-my-level]

Thanks @Yny4ii — this is carefully built work: the cancel affordance shows real attention to hit-target stability, the idempotency key is threaded through the right seams, and the tests (including the source-layout tripwire) are unusually thoughtful. One verified gap needs fixing before merge; the rest is minor.

Required: the Stop-speaking control never appears for native realtime voice answers.

VoiceResponsePlaybackMonitor is only refreshed from the TTS lane (FloatingBarVoicePlaybackService.swift:526/559/763/771) and from native stop (RealtimeHubController+VoiceOutput.swift:96 inside stopNativePlayback). Nothing refreshes it when native PCM playback starts — RealtimeHubController+SessionDelegate.swift:711 acquires the .nativeRealtime lease and enqueues audio with no refresh() — nor when playback drains naturally (onPlaybackIdle → releaseOutput, RealtimeHubController+VoiceOutput.swift:16-31). Every composer gates the new button on the monitor's cached @Published isActive (ChatInputView.swift, QueryHeroBar.swift, AskAIInputView.swift, and DashboardPage.swift through HomeAskBarControls.resolve(isSpeaking:)), so the button will not render while a native realtime answer is speaking, and the cached value can go stale afterwards. Escape still works because FloatingControlBarWindow.handleEscapeKey() calls VoiceResponsePlayback.interrupt() directly — which is also why this escapes casual testing.

Suggested fix: call VoiceResponsePlaybackMonitor.shared.refresh() after a successful acquireVoiceOutput(.nativeRealtime, …) and in onPlaybackIdle after releaseOutput. That also matches the test gap: VoiceResponsePlaybackTests covers only the idle path (testInterruptWhenIdleReturnsFalse) — please add a test that drives isActive true (e.g. coordinator acquires a .nativeRealtime lease → monitor refresh → isActive == true), which would have caught the missing start-side refresh.

Everything else is in good shape:

• VoiceResponsePlayback.swift — good consolidation of stop into one entry point across TTS and native lanes; lease-aware interrupt() (native stop before TTS interrupt) is the right ordering.
• PushToTalkMicButton.swift / PushToTalkButtonTrigger.swift — cancel trailing the mic so the send hit-target doesn't jump is exactly the right call, and showsCancelRecordingControl(phase:) correctly limits it to .recording / .lockedRecording / .pendingLockDecision.
• FloatingControlBarView.swift — both overlays (chat surface and pill) share the same cancel action, keep the container accessibility labels, and the chat overlay correctly drops allowsHitTesting(false) so the cancel button is clickable.
• PushToTalkManager.cancelPushToTalkForAutomation() and the ptt_cancel action in DesktopAutomationBridge.swift — mirrors the ptt_stop pattern faithfully and resets automationCaptureBypass.
• APIClient.swift / APIClient+TaskCatalog.swift / TasksStore.swift — the Idempotency-Key chain is clean: always-sent key (web parity), stable desktop-action-item:<rowid> for durable local rows, and ActionItemCreateIdempotencyTests verifies all three properties via URLProtocol capture.
• ChatProvider.swift (~line 3261) — stopping an agent turn now also silences spoken audio after the owner check; the comment explains the intent.
• FloatingBarVoicePlaybackService.swift — the setFloatingPillResponseGlow refactor (if/else instead of early return) keeps both lanes publishing; note this file carries four of the five refresh sites, so centralizing the refresh (or deriving the monitor from the coordinator's output snapshot) might be a worthwhile follow-up.

Non-blocking, for a maintainer: the always-on Idempotency-Key header lands ahead of its backend dedupe counterpart (as the description notes), and the cancel-vs-send semantics on the floating bar are a product call worth a quick human pass.

Requesting changes for the native-lane refresh wiring plus the one test; voice-UX semantics and header sequencing also deserve human maintainer sign-off before merge.

---

🤖 Automated maintainer review on behalf of @Git-on-my-level. AI-generated feedback - human review required before merge.

---

by AI on behalf of David — if you need David’s attention urgently, please @Git-on-my-level and escalate with need human response.