watch2

/watch2

Use watch2 to collect video evidence. The Rust binary downloads media, parses captions, detects scenes, and extracts frames only at timestamps selected by the agent. The agent performs selection, vision review, and conclusions.

Required workflow

# Pass 1: collect metadata, transcript, and scene boundaries.
watch2 "URL_OR_PATH" --out-dir /tmp/watch-XXX --output json

# Read report.json with jq. Do not use Python for this workflow.
jq '{title, uploader, language, duration, scene_count}' /tmp/watch-XXX/report.json
jq -r '.transcript[] | "[\(.start) → \(.end)] \(.text)"' /tmp/watch-XXX/report.json

# Select evidence timestamps, then extract only those frames.
watch2 "URL_OR_PATH" \
  --out-dir /tmp/watch-XXX \
  --keep-video \
  --timestamps "00:30,01:15,02:45" \
  --output json

# report.json is refreshed on every output mode.
jq '.frames[] | {path, timestamp, reason}' /tmp/watch-XXX/report.json

Rules

• Extract frames only with --timestamps; no timestamps means no frames.
• Select timestamps from transcript and scene boundaries. Add enough coverage for the video duration.
• Inspect every extracted frame before making visual claims.
• Use jq for report inspection, never Python helpers.
• Check report.json.analysis_capabilities.visual_verification before writing visual claims. It is true only when frames exist.
• On YouTube HTTP 403, do not retry or claim visual analysis. Verify yt-dlp -v URL lists a PO-token provider and use the mweb client; then pass --cookies-file (0600), use a local video, or explicitly use --allow-transcript-only.
• If watch2 fails, inspect its error; use ffprobe or ffmpeg only for diagnosis or a documented manual fallback.
• Return user-facing conclusions, not workflow logs or raw frame-by-frame notes unless requested.

CLI