What problem does it solve? It turns recurring violations, blind spots, and stalled objectives found in agent trajectory findings into governed rule proposals, without letting machine-learned rules influence the current run's approval. ## Core Features & Use Cases - Failure-signature detection: Maps recurring violations, uncitable blind spots, tool-call loops, stalled objectives, unmeasured escapes, and dead rules to Create/Update/Delete proposals. - Blast-radius routing: Routes prose proposals to a rule-verifier for refutation and lint/test/gate-metric or high-risk-path proposals to an operator gate with literal text rendered. - Evidence-bound proposals: Requires source findings with file:line citations, conflict checks, refuted_by criteria, and records every outcome including rejections in refine-log.md. - Use Case: After a Judge phase completes, dispatch the refiner over the findings window so a recurring lint violation becomes a proposed rule update that the operator approves or rejects at confirmation. ## Quick Start Run the Refine phase after Judge by dispatching the refiner subagent with the findings window counts and per-rule proposeBar blocks from harness-findings.mjs.
