What problem does it solve?
When generating new meta-skills automatically, you need a way to verify that a candidate SKILL.md actually triggers on relevant inputs and rejects irrelevant ones before shipping it. This internal sub-agent runs the G3 (positive smoke) and G4 (negative smoke) validation gates as a DAG step inside the meta-skill-creator workflow.
Core Features & Use Cases
- Positive Smoke Gate (G3): Generates a positive fixture with an LLM and calls simulate_meta_resolution to confirm the candidate skill resolves correctly.
- Negative Smoke Gate (G4): Generates a cross-domain negative fixture to confirm the skill does not falsely trigger.
- Cross-Vendor Validation: Uses a different LLM for fixture generation than for classification to avoid self-confirmation bias.
- Degraded Fallback: Falls back to a deterministic fixture generator with a stub classifier when OPENROUTER_API_KEY or model configuration is missing, marking results with a degraded flag.
- Use Case: As part of an automated skill-authoring pipeline, invoke this step after drafting a candidate SKILL.md to receive a JSON report of G3/G4 gate results before accepting the new skill.
Quick Start
Invoke this as a kind: agent step within the meta-skill-creator DAG, passing skill_md, fixture_gen_model, and classifier_model in the step's with block.
