Stage 2: Extract + YAML + Human Review

Rule

• •Use plugin tools only. Do not call scripts directly.
• •Keep execution in the main agent turn. Do not use subagents.
• •This stage must save program info to YAML after extraction.
• •This stage ends with human review. Do not continue to move/apply automatically.

Tool sequence

1. •Call video_pipeline_validate with {"checkWindowsInterop": true}.
2. •Call video_pipeline_reextract with:
   • •queuePath from Stage 1 summary (queue)
   • •optional batchSize / maxBatches
   • •keep default preserveHumanReviewed=true (reviewed rows are protected from overwrite)
3. •Call video_pipeline_export_program_yaml with:
   • •sourceJsonlPath omitted (latest extraction output auto-detected), or explicit path
   • •optional outputPath if user wants a fixed location
   • •default is ${windowsOpsRoot}/llm/program_aliases_review_YYYYMMDD_HHMMSS.yaml
4. •If user manually edits extracted JSONL, call video_pipeline_apply_reviewed_metadata:
   • •sourceJsonlPath: edited extraction JSONL path
   • •default markHumanReviewed=true
   • •this step writes reviewed metadata into DB (path_metadata)

Human review checklist

• •YAML file was generated successfully.
• •Program titles/aliases in YAML are acceptable.
• •Rows with needs_review are either fixed or intentionally kept for later.
• •Manual fixes are applied to DB via video_pipeline_apply_reviewed_metadata.
• •User confirms YAML should be used as the next review baseline.

Handoff

• •Return:
  • •extraction result summary
  • •YAML output path
  • •count of programs exported
• •Ask user whether to proceed to Stage 3 (Move + Human review).