Add live transcription preview

[basnijholt]

Summary

• Add rolling live transcription previews for Wyoming ASR with JSONL and terminal output support.
• Wire macOS menu bar app to display live preview text above the voice-level overlay while recording.
• Add regression coverage for CLI preview output, app preview parsing, command wiring, and app self-test path.

Test Plan

• uv run ruff check agent_cli/agents/transcribe.py agent_cli/services/asr.py agent_cli/opts.py tests/test_asr.py tests/agents/test_transcribe.py tests/agents/test_transcribe_recovery.py
• uv run pytest tests/test_asr.py tests/agents/test_transcribe.py tests/agents/test_transcribe_recovery.py tests/test_docs_gen.py -q
• swift build in macos/AgentCLI
• swift run AgentCLI --agentcli-live-preview-self-test in macos/AgentCLI
• Manual CLI smoke: uv run agent-cli transcribe --live-preview-stdout

[greptile-apps]

Greptile Summary

This PR adds rolling live transcription previews for Wyoming ASR: the Python side periodically re-transcribes a sliding audio window and writes JSONL events to a log file (and/or the terminal), while the macOS app's new LiveTranscriptionPreview singleton incrementally reads that log and publishes preview text above the voice-level overlay.

• Python layer (asr.py, transcribe.py, opts.py): adds LivePreviewConfig, LivePreviewStreamer with a cancel-safe loop, and four new CLI flags. live_preview_console_active now correctly gates Rich live-status suppression on both the Wyoming provider and non-file-transcription checks, fixing the previously reported regression where non-Wyoming users would lose the Rich status.
• macOS layer: panel sizing is now dynamic (compact vs. preview), LiveTranscriptionPreview reads the JSONL file incrementally via FileHandle offset tracking, and a new self-test validates the full argument-wiring and parsing path.
• Tests: cover JSONL write/read, unique-partial deduplication, stale-partial-after-stop guard, the cancel-race scenario, and Rich live-status suppression logic for all provider/file-path combinations.

Confidence Score: 5/5

Safe to merge; the feature is well-scoped to Wyoming ASR live recording, the cancel-race ordering is correct, and the previously reported Rich live-status suppression for non-Wyoming providers is now fixed.

The Python cancel-ordering (request_stop before task.cancel + await) correctly prevents stale partials from landing after the final event, confirmed by the dedicated race-condition test. The live_preview_console_active gate now includes both the Wyoming provider check and the file-transcription check, fixing the issue raised in an earlier review pass. The two P2 observations are both benign given current usage patterns.

No files require special attention; the incremental FileHandle-based reading in LiveTranscriptionPreview.swift and the cancel-safe run loop in asr.py are the most complex new paths and both have direct test coverage.

Important Files Changed

Filename	Overview
agent_cli/services/asr.py	Adds LivePreviewConfig dataclass and LivePreviewStreamer with a periodic emit loop; wires live_preview_callback into _send_audio; adds finally-block cleanup in _transcribe_live_audio_wyoming. The race-condition ordering (request_stop before cancel+await) matches the test coverage, and _transcribe_recorded_audio_wyoming returns str (never None) on failure, so emit_partial is safe.
agent_cli/agents/transcribe.py	Wires four new live-preview parameters through _async_main and transcribe; live_preview_console_active now correctly gates Rich live status on both the Wyoming provider check and the file-transcription check, addressing the previous concern about non-Wyoming providers losing the Rich status.
macos/AgentCLI/Sources/AgentCLI/LiveTranscriptionPreview.swift	New singleton that incrementally reads the JSONL log with FileHandle (offset-tracked), parses events, and publishes text via @published. The pendingLine double-parse edge case is low-risk given the Python writer always appends newlines.
macos/AgentCLI/Sources/AgentCLI/RecordingIndicatorController.swift	begin(for:) checks both supportsLivePreviewOverlay and isLivePreviewOverlayEnabled before calling start(), but end(for:) calls stop() on supportsLivePreviewOverlay alone — harmless today since stop() is idempotent, but asymmetric.
macos/AgentCLI/Sources/AgentCLI/VoiceLevelOverlay.swift	Panel sizing is now dynamic (compactPanelSize vs previewPanelSize) instead of always allocating the preview-space size; panel content is re-hosted only when showsPreviewSpace changes.
tests/test_asr.py	Comprehensive new test coverage: JSONL write, console print, unique-partial deduplication, stop writes final, stale-partial-after-stop guard, and the cancel race all covered.

Sequence Diagram

sequenceDiagram
    participant User
    participant RecordingIndicatorController
    participant LiveTranscriptionPreview
    participant VoiceLevelOverlayController
    participant CLI as agent-cli process
    participant LivePreviewStreamer
    participant WyomingASR

    User->>RecordingIndicatorController: begin(for: toggleTranscription)
    RecordingIndicatorController->>LiveTranscriptionPreview: start(logURL)
    Note over LiveTranscriptionPreview: Truncates JSONL, starts 250ms timer
    RecordingIndicatorController->>VoiceLevelOverlayController: show(showsPreviewSpace: true)
    RecordingIndicatorController->>CLI: spawn with --live-preview-log --live-preview-interval 1 --live-preview-window 10

    loop Every interval_seconds (1s)
        CLI->>LivePreviewStreamer: emit_partial()
        LivePreviewStreamer->>WyomingASR: _transcribe_recorded_audio_wyoming(rolling window)
        WyomingASR-->>LivePreviewStreamer: partial text
        LivePreviewStreamer->>LivePreviewStreamer: "_publish_event(type=partial)"
        LivePreviewStreamer-->>CLI: append to JSONL log
    end

    loop Every 250ms (Swift timer)
        LiveTranscriptionPreview->>LiveTranscriptionPreview: readNewContents(offset-tracked)
        LiveTranscriptionPreview->>LiveTranscriptionPreview: parse last valid JSON event
        LiveTranscriptionPreview-->>VoiceLevelOverlayController: "@Published text update"
    end

    User->>RecordingIndicatorController: end(for: toggleTranscription)
    CLI->>LivePreviewStreamer: finally: request_stop() then cancel task then stop(final_text)
    LivePreviewStreamer-->>CLI: append final event to JSONL
    RecordingIndicatorController->>VoiceLevelOverlayController: hide()
    RecordingIndicatorController->>LiveTranscriptionPreview: stop()
    Note over LiveTranscriptionPreview: Invalidates timer, clears state

Loading

_{Reviews (6): Last reviewed commit: "Signal live preview stop before cancella..." | Re-trigger Greptile}