Transcribe media

Turn one verified local media input into a traceable transcript artifact. Do not combine transcription with summarization or knowledge-base writes.

Platform boundary

• Mutating and file-verifying operations support Python 3.11–3.14 on macOS and Linux.
• They require POSIX fcntl, dir_fd, O_NOFOLLOW, O_DIRECTORY, directory fsync, atomic no-replace rename, and atomic exchange rename. Missing capabilities fail closed with UNSUPPORTED_PLATFORM; there is no weaker fallback.
• Input, model, adapter, output, state, and chunk paths must not traverse arbitrary symlinks. Managed files must be current-user-owned regular files with one link.

CLI protocol

Success prints exactly one JSON object to stdout, leaves stderr empty, and exits 0. An expected failure leaves stdout empty and prints exactly one redacted JSON error to stderr. Repository-wide failure codes are: 2 for arguments, schemas, or unsafe inputs/paths; 3 for unavailable dependencies, models, or platform capabilities; 4 for locks, conflicts, recovery conflicts, or stale plans; 5 for external-tool, network, or runtime I/O failures; 7 for contract, identity, or file-evidence integrity failures; and 130 for interruption.

Workflow

1. Determine whether the input is a URL or a local path.
   • For a URL, invoke $download-video, require a complete awesome-capture.artifact/v2 video artifact, and retain both the artifact path and media.path.
   • For a local path, inspect it directly. Do not invent or guess an adjacent artifact path.

2. Check the environment:

   python3 scripts/transcribe_media.py doctor \
     --model "<absolute-local-ggml-model-path>" \
     --whisper-cpp-bin "<absolute-whisper-cli-path>"

   --engine auto requires both this explicit binary path and the explicit local model path. An explicit --engine whisper-cpp may still discover whisper-cli on PATH when the binary option is omitted. For speech recognition without a sidecar subtitle, require ready_for_asr: true; status: ok only means the diagnostic command itself completed.

3. Inspect the media and confirm it has an audio stream:

   python3 scripts/transcribe_media.py inspect "<absolute-media-path>"

4. Prefer an exact-basename .srt or .vtt sidecar when present. Otherwise select an engine:
   • auto: whisper.cpp only when both --whisper-cpp-bin and --model explicitly name existing local files. It never discovers the binary from PATH or selects a Python engine.
   • whisper-cpp: first-class local engine. It requires an explicit local --model file and optionally --whisper-cpp-bin.
   • faster-whisper: explicit only; requires a fully local model directory.
   • mlx-whisper: explicit only; requires a fully local model directory and a compatible local MLX installation.
   • external: explicit only; requires --adapter, a local --model, and --trust-external-adapter.

5. For a local file, run:

   python3 scripts/transcribe_media.py transcribe "<absolute-media-path>" \
     --output-dir "<absolute-dir>" \
     --engine auto \
     --model "<absolute-local-ggml-model-path>"

   For a video produced by $download-video, pass the handoff explicitly:

   python3 scripts/transcribe_media.py transcribe \
     "<video-artifact.media.path>" \
     --source-artifact "<absolute-video-artifact-v2-path>" \
     --output-dir "<absolute-dir>" \
     --engine auto \
     --model "<absolute-local-ggml-model-path>"

   --source-artifact is never inferred. When present, its schema, contract digest, media path, bytes, SHA-256, duration, stream counts, and current ffprobe evidence are all revalidated.

   Other transcription options are:

   • --language <code> to request a language;
   • --ignore-sidecar to force ASR despite an exact-basename subtitle;
   • --chunk-seconds <n> (default 600, minimum 30);
   • --timeout <seconds> for each external process (default 3600);
   • --lock-timeout <seconds> for the persistent job lock (default 30).

   To bypass the initial whisper.cpp GPU subprocess, add --whisper-cpp-cpu-only. Normally leave it off: the first GPU failure is recorded, retried with -ng, and cached so the remaining chunks in that task go directly to CPU.

6. Require zero exit status and a complete awesome-capture.artifact/v2 transcript. Read output paths from artifact.outputs; do not derive them from filenames. Report the engine, detected language, duration, segment count, transcript Markdown path, and any empty/silent result.
7. If the transcript contains speech, ask exactly one follow-up: “是否需要把这份内容写入本地 Obsidian 知识库？” If yes, invoke $ingest-knowledge with transcript.json. If no, stop.

Recovery

Transcription uses a persistent job lock and private managed layout:

<output>/.awesome-capture-media/v2/
  locks/
  staging/
  transcriptions/<64-hex-job-id>/
  quarantine/

The source is copied to a private content-verified snapshot before ffmpeg or ASR reads it. Managed directories use mode 0700; locks, snapshots, chunks, state, outputs, pending artifacts, and final artifacts use 0600.

Re-running the same command acquires the same lock and either resumes a valid running state, publishes a verified pending artifact, or returns reused. For an explicit recovery pass:

python3 scripts/transcribe_media.py recover \
  --output-dir "<absolute-dir>" \
  --lock-timeout 30

Recovery never overwrites an existing artifact. Unknown workspace files, unsafe links, conflicting identities, or invalid hashes produce RECOVERY_CONFLICT or an integrity error.

Reliability rules

• Preserve source timestamps; never invent missing speech.
• Keep awesome-capture.transcription-state/v1, the complete awesome-capture.chunk-set/v1 manifest, and per-chunk results so an interrupted run can resume.
• Reject missing or extra chunks, filename gaps, changed hashes, non-mono 16 kHz PCM16 WAVs, inconsistent cumulative timing, foreign state entries, and incomplete result sets.
• Offset each chunk by the cumulative duration measured from the normalized WAV sample counts. Never multiply the nominal --chunk-seconds value by the chunk index.
• Do not download models. Every ASR engine requires an explicit local model. Record and bind the model, engine executable or adapter, package versions, settings, and source to the job identity.
• Re-hash the private source snapshot, model, executable/adapter, and the versioned content identity of the standalone transcription implementation before publishing. Also compare the execution metadata guard so a component changed and then restored to the same bytes still aborts with IDENTITY_CHANGED. Once the private snapshot is durably verified, recovery does not require the original source path or upstream video artifact to remain present.
• Write chunk results and derivative outputs while state remains running. After validating every result and output hash, durably write transcript.pending.json; only then advance state to ready_to_publish. Promotion revalidates source/model/adapter/chunks, advances state to complete, and refreshes the pending artifact's state descriptor. Publish transcript.json last without clobbering an existing entry.
• Reject artifact/v1, unversioned or unknown state, duplicate JSON keys, non-finite numbers, and unknown contract properties. Do not migrate or silently fall back.
• Treat a platform title, description, chapter summary, or LLM-generated reconstruction as metadata, not transcription.
• Do not summarize before the transcript artifact is complete.
• external executes explicitly trusted local code. It is not selected automatically and is not an OS-level network sandbox.
• External media tools run with a working directory pinned by directory FD; regular media/model-file inputs are passed through inherited read-only FDs where supported. Local model directories are still re-hashed before and after execution and are not an OS sandbox.
• Read engines.md before choosing models, handling long files, or adding an engine.
• Read artifact-contract.md when composing with another skill.