| name | video-library-pipeline |
| description | Run and inspect the video library pipeline through the V2 run-based OpenClaw tool surface. |
| metadata | {"openclaw":{"emoji":"🎬","requires":{"plugins":["video-library-pipeline"]},"localReads":["~/.openclaw/openclaw.json"]}} |
video-library-pipeline
This skill is the V2 orchestrator for video-library-pipeline.
V2 Rules
- Use plugin tools only. Do not call Python, PowerShell, or shell commands directly.
- The active public tool surface is limited to:
video_pipeline_startvideo_pipeline_resumevideo_pipeline_statusvideo_pipeline_inspect_artifact
- Treat
WorkflowResult.nextActions and followUpToolCalls as the source of truth for the next operation.
- Do not infer "latest" JSONL/YAML/plan files. Use
runId, artifactId, ReviewGate.artifactIds, and artifact paths returned by the run.
- Human review is explicit. If a result has
requiresHumanReview: true or an open ReviewGate, stop and ask the user to review the referenced artifact before resuming.
- If
nextActions returns complete_empty_plan with requiresHumanInput: false, call the exact video_pipeline_resume params directly. It only completes a checksum-verified sourceRoot plan whose recorded count is zero and whose JSONL contains no operations.
- If
nextActions returns register_unregistered with requiresHumanInput: false, call the exact resume params directly. This action only registers missing paths/observations/events; it uses the non-moving registration mode and then reruns dry-run so the new files enter metadata preparation.
- Execute in the main agent turn; do not delegate to subagents.
Stale Run Supersession
- Never apply an old plan merely because its run is still active.
- When filesystem or metadata state may have changed, first run a fresh dry-run for the same flow and safe parent root.
- Compare the old artifact summary and preview with the fresh run: source existence, scanned count, planned count, metadata queue count, and destination layout.
- If the old snapshot is no longer valid, supersede it through the public resume tool:
video_pipeline_resume {
"runId": "<old-run-id>",
"resumeAction": "supersede_run",
"supersededByRunId": "<newer-same-flow-run-id>",
"reason": "<specific evidence that invalidated the old snapshot>"
}
supersede_run is state-only: it performs no file move. It requires an active old run, an existing newer run of the same flow, and a non-empty reason. It marks old artifacts and open gates superseded, completes the old run, and records the successor link in the manifest.- Keep the fresh run active when it represents current human-review work. Do not supersede the current review gate merely to make the active count zero.
Intent Mapping
| User intent | V2 action |
|---|
Process new recordings from sourceRoot | Read skills/inventory-review/SKILL.md; start flow: "source_root" |
| Reorganize/relocate existing library files | Read skills/relocate-review/SKILL.md; start flow: "relocate" |
| Continue an existing run | Call video_pipeline_status with runId and includeArtifacts:true; use returned nextActions to resume. |
| Inspect review YAML, plan, diagnostics, or apply log | Call video_pipeline_inspect_artifact with runId and artifactId |
| Metadata review handoff | Read skills/extract-review/SKILL.md; use the run's ReviewGate and artifacts |
| Apply/move after plan review | Read skills/move-review/SKILL.md; resume the run with the plan action returned by nextActions |
| DB maintenance, EPG ingest, dedup, title repair, folder cleanup | Not exposed through the V2 public surface in this issue. Report that the operation requires a future V2 workflow or an explicit maintenance/admin path. |
If the request targets an already-existing directory tree under the library, treat it as the relocate workflow, not the sourceRoot workflow.
Standard V2 Loop
- Start a run or inspect current state:
- New sourceRoot run:
video_pipeline_start {"flow":"source_root"}
- New relocate run:
video_pipeline_start {"flow":"relocate", "roots":[...]}
- Existing run status only:
video_pipeline_status {"runId":"<runId>", "includeArtifacts":true}
- Read the result fields:
runIdflowphaseoutcomeartifactsgatesnextActions / followUpToolCallsdiagnostics
- If a human review gate is present, inspect the referenced artifact and ask the user to review it.
- If
complete_empty_plan or register_unregistered is returned with requiresHumanInput: false, call its exact resume params and verify the run advances.
- After other review or approval, call
video_pipeline_resume only with exact params from followUpToolCalls[].params or nextActions[].params.
- Repeat until
phase is complete, blocked, or failed.
video_pipeline_status reconstructs actionable nextActions from the run manifest for non-terminal review and plan phases. If no action is returned, do not guess a latest JSONL/YAML/plan path.
Reporting
- Always state
runId, flow, phase, and outcome.
- For review gates, show
gate.id, gate.status, and the artifact IDs the user must review.
- For apply/move plans, summarize artifact IDs and paths; do not invent destination decisions outside the artifact content.
- Distinguish "review required", "plan ready", "applied", "complete", "blocked", "failed", and "superseded".
Legacy Tool Guardrail
Legacy public tools from V1 are hidden after #108. Do not instruct the user to call hidden legacy tools such as video_pipeline_analyze_and_move_videos, video_pipeline_reextract, video_pipeline_apply_reviewed_metadata, video_pipeline_relocate_existing_files, or video_pipeline_validate as active V2 steps.
