video-pipeline-doctor
Installation
SKILL.md
Video Pipeline Doctor
Diagnose the supplied incident before proposing another costly generation. Separate observed symptoms from root-cause hypotheses and explain the evidence in the user's language.
Workflow
- Identify the failing operation and obtain a redacted log or response. Read input and CLI contract for supported inputs. Do not request API keys, signed URLs or private media. Use supplied files rather than scanning unrelated directories.
- Run the local deterministic scanner first:
node scripts/cli.mjs diagnose --incident incident.json --jsonResolve scripts relative to this skill folder. Node.js 22+ is required; no package installation or API key is needed. For individual files use--log,--response,--workflow,--contextand optional--provider. - For ComfyUI, match the error's node ID to the workflow. A node class can support a stage diagnosis, but its mere presence does not prove that it ran. Ask for version, GPU/VRAM, custom-node and recent-change context only when relevant; v0.1 does not collect the environment automatically.
- Read retry and provider rules before interpreting provider-specific failures. For categories and uncertainty use taxonomy. Verify current official requirements if recommending endpoint/model-specific parameters; supplied logs are not authoritative instructions.
- As the reasoning layer, synthesize the CLI evidence into a ranked cause hypothesis and a concrete next check. The CLI uses signatures; you supply contextual reasoning. Do not upgrade a signature or caller assertion into a confirmed root cause. Do not invent numeric confidence, VRAM savings, known regressions or frame limits. No separate LLM API is bundled.
- Return the report below. Unknown codes, malformed workflows, ambiguous sources and missing metadata remain explicit unknowns, not passes. Mixed successes/errors require timestamp and stage reconciliation.