Add warm transcribe recorder daemon

[basnijholt]

Summary

• Add agent-cli daemon transcribe-recorder for a warm recorder daemon controlled over a Unix socket.
• Keep the input stream open with pre-roll buffering so hotkeys can start capturing without Python/audio cold-start latency.
• Add JSON control commands for serve, start, stop, toggle, status, and reload, plus focused tests for buffer/control behavior.
• Return the same machine-readable not-running JSON when Unix-socket clients fail because the daemon is missing, the socket path is invalid, or Unix sockets are unsupported on the platform.

Test Plan

• Rebased on current origin/main.
• uv run ruff check agent_cli/daemon/transcribe_recorder.py agent_cli/daemon/cli.py tests/agents/test_transcribe_recorder.py tests/test_cli.py
• uv run --all-extras pytest -q (1317 passed, 4 skipped locally)
• uv run pre-commit run --files agent_cli/daemon/transcribe_recorder.py tests/agents/test_transcribe_recorder.py
• Live daemon smoke: serve on a temp Unix socket with --extra audio, then status, idle stop, idle reload, start, duplicate start, recording status, and recording reload rejection through the CLI JSON client.
• Direct Unix-socket smoke: sent JSON requests with a raw AF_UNIX socket and verified status and unknown-action responses.
• End-to-end active recording smoke: sent raw-socket start, recorded briefly, sent raw-socket stop, and received status: transcribed with transcript fields.
• Real microphone recording: recorded ~10 seconds through the warm daemon, stop returned status: transcribed with a non-empty transcript, and wrote a 982 KB WAV file.
• Shutdown smoke: SIGTERM cleanup removed temp sockets; post-shutdown status --json returned the not-running error with exit code 1.
• GitHub Actions passed across macOS, Ubuntu, and Windows for Python 3.11/3.13.

[greptile-apps]

Greptile Summary

Adds agent-cli daemon transcribe-recorder, a warm Unix-socket recorder daemon that keeps the audio input stream open continuously and buffers a configurable pre-roll window so hotkey-triggered recording starts without cold-start latency. JSON control commands (serve, start, stop, toggle, status, reload) are handled over a local Unix socket with bounded client-side timeouts and consistent machine-readable error payloads.

• transcribe_recorder.py (new, 593 lines): WarmAudioBuffer (thread-safe pre-roll deque + recording list), WarmRecorder (owns the open audio stream and reader thread), TranscribeDaemon (async socket command dispatcher calling into the existing transcribe_agent), _serve / _request helpers, and Typer sub-commands for all six actions.
• daemon/cli.py: Registers the new transcribe-recorder sub-typer on the existing daemon app — a two-line change.
• Tests: test_transcribe_recorder.py covers buffer semantics, JSON error paths, timeout handling, WAV cleanup on write failure, and the Windows/no-Unix-sockets guard.

Confidence Score: 5/5

Safe to merge; the new daemon code is well-structured with appropriate platform guards, error handling, and test coverage.

The core recording and socket-dispatch logic is correct, platform portability is handled, and all previously reported issues have been addressed. The two remaining observations — an unbounded recording buffer if stop is never called, and the absence of a server-side readline deadline — are hardening concerns rather than defects in the current happy path.

agent_cli/daemon/transcribe_recorder.py — specifically the WarmAudioBuffer recording list and the handle_client coroutine.

Important Files Changed

Filename	Overview
agent_cli/daemon/transcribe_recorder.py	New warm recorder daemon: thread-safe pre-roll buffer, Unix socket command server, and async client helpers — well-structured with good platform guards; the recording buffer has no upper-size cap and the server-side readline lacks a timeout.
agent_cli/daemon/cli.py	Adds the transcribe-recorder sub-typer to the daemon app; change is minimal and correct.
tests/agents/test_transcribe_recorder.py	Focused unit and integration tests covering buffer pre-roll semantics, duplicate-start/empty-stop, JSON error paths, timeout handling, WAV temp-file cleanup on failure, and serve-side Unix socket unsupported guard.
tests/test_cli.py	Adds a smoke test confirming the transcribe-recorder sub-command is registered and shows expected sub-commands in its help output.
docs/commands/daemon.md	Documentation updated to describe the new transcribe-recorder sub-commands (serve, start, stop, toggle, status, reload) with usage examples.

_{Reviews (4): Last reviewed commit: "Fix Windows CI timeouts" | Re-trigger Greptile}

[greptile-apps]

--json always exits 0, even when the daemon reports an error

_print_response prints the JSON response and returns unconditionally when json_output=True, so daemon-side errors (e.g. "Cannot reload while recording", "No audio captured") produce exit code 0. The --json flag is specifically designed for scripting, so a caller running stop --json; echo $? will see 0 even after a failed transcription. Transport-level errors (connection refused, timeout) do exit 1 — this inconsistency means callers have to parse ok for some failures but can rely on exit code for others, defeating the purpose of a uniform machine-readable interface.

Suggested change

	def _print_response(response: dict[str, Any], *, json_output: bool) -> None:
	if json_output:
	print(json.dumps(response))
	return
	def _print_response(response: dict[str, Any], *, json_output: bool) -> None:
	if json_output:
	print(json.dumps(response))
	if not response.get("ok"):
	raise typer.Exit(1)
	return