What problem does it solve?
When an AI agent harness routes a request incorrectly or loads the wrong context, teams often patch prose without a reproducible regression, so the same failure returns. This Skill turns each real failure into a versioned, schema-validated case with semantic assertions, a single knowledge owner, and before/after evaluation evidence.
Core Features & Use Cases
- Failure-to-case workflow: Capture a prompt, transcript, or PR as untrusted evidence, classify and deduplicate it, and reproduce it in a fresh pinned standalone scaffold.
- Semantic assertions: Reduce failures to routing, decision, and artifact assertions instead of whole-output or whole-file goldens, with calibrated context budgets.
- Single-owner knowledge changes: Select exactly one smallest owner (root invariant, router row, guide, skill reference, facts extractor, hook) and update only it, replacing duplicates with references.
- Gated release validation: Run catalog validation, writable-target generate/typecheck/lint/build gates, code review, an isolated judge lane, and the full harness release suite with sanitized reports.
- Use Case: An agent picks the wrong module guide for a request. Use this Skill to add an OMH-NNN case that fails before the fix, correct the router row, and prove the fix with the release suite.
Quick Start
Ask the agent to add a harness case for the routing failure you just observed and evolve the harness with before and after evaluation evidence.
