[Juliusolsson05]

Summary

Holding a configured hotkey (default: bare Fn on macOS) records into the focused TileLeaf composer. Live transcript previews while you talk; on release the final, STT-tagged text replaces the preview as a normal editable draft. Multi-turn safe — pausing mid-recording does not lose earlier sentences.

What changed

Dictation feature (commit f8f3a9c)

• Native Fn capture. native/macos-hotkey-helper is a tiny Swift CGEventTap. Compiled-on-first-run, hashed by source so repeat boots are instant. Chromium does not deliver bare-Fn to renderer keydown and Electron's globalShortcut rejects it as "not an accelerator", so a side-channel was the only path. Streams hotkey-down/up JSON over stdout to main.
• Single dispatch registry. Per-hook IPC subscriptions silently dropped every press until a composer was focused. Now one shared registry owns the IPC, picks the focused composer (or most-recently-focused as fallback) at press time, and remembers the chosen target across press → release so a quick tap can never strand a recorder.
• Recorder-led socket lifecycle. MediaRecorder starts immediately. The Deepgram WebSocket only opens after MIN_HOLD_TO_TRANSCRIBE_MS = 180ms so accidental Fn taps never open and immediately cancel a handshake. Audio queues locally; on WS open, the queue is drained before publishing recording.id so concurrent dataavailable handlers cannot interleave with the flush and corrupt WebM order.
• Live preview chokepoint. Every interim/commit/restore write goes through one writeInput(next, source) so the composer's draft is always derivable from a single point.
• Microphone pre-warm. macOS's first getUserMedia per app boot is a multi-second cold start. We fire a one-shot warm-up the moment any composer hook becomes enabled and immediately stop the tracks. Subsequent presses are instant. Pre-warm failure is non-fatal.
• Build glue. electron.vite.config.ts aliases agent-voice-dictation to its dist and externalises ws (Vite otherwise inlines bufferUtil in a shape where mask is not a function and the first audio chunk crashes main). Matching path in tsconfig.node.json.

LSP destroyed-stream fix (commit e51c1cf)

Unrelated to dictation; the change was in the same working tree so it ships here. LSP notifications were calling connection.sendNotification directly and throwing ERR_STREAM_DESTROYED when a language server died mid-flight, leaving syntax highlighting stuck in an error state. Now tracks a closed flag per server, routes notifications through sendNotificationIfOpen that no-ops on a destroyed stdin, and the semantic-tokens read path returns null on stream-destroyed so the editor falls back to lexer highlighting cleanly.

Notes on the journey

This branch is the result of a long debug session that fixed several intertwined bugs:

1. Spurious unmount cancel — the unmount effect's deps were unstable callbacks, so React fired the cleanup on every render and cancelled the WS within ~1 frame of opening. Fix: prop callbacks behind refs, true unmount effect with [] deps.
2. Out-of-order chunks — recording.id was published before the queue drain, so late dataavailable handlers raced the flush. Fix: drain entirely BEFORE assigning recording.id.
3. 30s "no Fn after launch" — every TileLeaf hook gated on its own focus, so on cold boot with no focused composer every press was a silent no-op. Fix: shared registry with most-recently-focused fallback.
4. Cold mic = stuck "starting" — getUserMedia first-call took 4+ seconds on macOS; users released before recording began. Fix: pre-warm at boot.
5. Multi-turn audio loss — fixed in agent-voice-dictation@cd0f87f (cumulative interim) and @06cb096 (combined final + interim).

Test plan

• npm run dev boots cleanly; helper logs mac helper ready for "Fn" within ~1s.
• First Fn press is instant; live transcript appears in the focused composer.
• Speak, pause >1s, speak again, release: both sentences land in the commit (multi-turn safety from the package fix).
• Quick Fn tap (<180ms) does nothing user-visible — no orphaned WS, no stuck "listening".
• Press Fn, click another tile mid-record, release: recording stops cleanly in the original tile.
• Disable dictation in Settings → next press is a no-op; re-enable → works again.
• LSP regression: open a TS file, kill typescript-language-server externally — editor falls back to lexer highlighting instead of erroring.

Out of scope / follow-ups

• Packaging (electron-builder) currently does not ship native/** under the asar, so a packaged build would fail to find the Swift source. Dictation in dev only until that's wired.
• /tmp chunk dump is opt-in via CC_SHELL_DICTATION_DUMP=1; off by default.
• Verbose per-chunk traces gated on CC_SHELL_DICTATION_VERBOSE=1; off by default.

🤖 Generated with Claude Code