[pineapplestrikesback]

Agent speech currently plays through a floating status pill without a way to pause, seek, change speed, or replay it there. This adds a compact player using the existing completed recording: pause/resume, position and duration, seeking, replay, stop, and a saved 0.75×–2× speed for agent speech.

The same widget mirrors the main waveform player and sends commands back to its existing audio source. Selecting a recording alone does not open the widget. Finished playback stays visible for five seconds for replay; History keeps the recording afterward.

Implementation

• Extract completed-file speech playback into a testable transport with media-clock progress, pitch-preserving browser speed, and cleanup of stale generation events and play promises.
• Retain the existing latest-speech-wins behavior and backend endpoints. No model, generation, streaming, Inbox, or attention-tracking changes.
• Keep the main player's existing WebAudio backend and normal speed. Native device routing exposes its existing limitations: no seeking/speed, and stop/start restarts playback.
• Serialize native playback starts/stops so cancelling an in-flight start cannot leave audio running or stop a newer recording. Pause the current waveform before awaiting native cleanup.
• Make space for the controls in the floating window; show it without activation on macOS and restore the preceding app only when the widget still owns focus. Preserve existing application-close and server behavior.
• Translate controls into all nine existing locales, add a usage/manual-test guide, and run playback tests in frontend CI. No new dependencies.

Validation

• 30 playback tests passed, 507 assertions: transport lifecycle, queued media-event races, native start/stop ordering, current-session widget commands, persisted speed, rendered controls, and locale parity.
• App/web TypeScript checks passed. Web production build and Tauri frontend production build passed.
• Targeted Biome checks passed for new playback modules, tests, and DictateWindow. Existing unrelated AudioPlayer lint diagnostics remain.
• macOS Rust cargo check passed with a temporary bundle.externalBin=[] override because the checkout has no built Python sidecar. Upstream's stale root-package version in Cargo.lock was resolved temporarily for the check and restored; no lockfile changes are included. This was a compile check, not a packaged build.
• An isolated Brave fixture exercised the real component and transport with a muted five-minute WAV: pause, keyboard seek, speed change, saved speed after reload, completion, and replay using the same audio instance. Observed seek/speed handlers took 0.1–0.3 ms in this fixture. This does not measure audible speech quality or native WKWebView behavior.
• Packaged macOS 27.0.1 arm64 pass: an isolated, signed release app used the branch backend on a separate port with seeded speech recordings. Native pause/resume, pointer seeking, speed, persisted speed after process restart, completion/replay, stop, main-player mirroring and replacement requests passed. A manual click focused the player; dismissal returned focus to the prior app, confirmed by the tester and a passive OS activation log. Playback/focus code matched this patch byte-for-byte. No additional code fixes were required.

The macOS package test used a separately supervised Python backend and bypassed synthesis only. It does not validate Python-sidecar bundling, installer/notarization/updater delivery, model inference, selected-device hardware routing, acoustic output quality, or Windows/Linux. Native keyboard seeking was not established by background automation. The manual checklist is in docs/PLAYBACK_CONTROLS.md.

macOS preview