What problem does it solve? A green CI run does not prove an agent's work was good: it can pass while bending conventions, writing tautological tests, or repeating mistakes you already corrected. This Skill judges a finished factory run cold and turns the findings into durable rules and memories so the system improves over time. ## Core Features & Use Cases - Cold adversarial evaluation: Spawns a fresh evaluator subagent that judges from a compressed session digest plus the real git diff, never from the agent's own self-narration. - Seven-dimension rubric: Scores loop adherence, convention compliance, test integrity, ask-vs-proceed judgment, correction handling, verification honesty, and efficiency with file:line evidence. - Durable lesson routing: Routes RULE findings to /add-rule (conventions doc plus lint guard) and PROCESS findings to persistent memory, acting only on user approval. - Use Case: After a factory run converges on a green PR, ask for an evaluation to discover it mocked the database in a test that required real dependencies, then capture that as a permanent lint guard. ## Quick Start Ask the agent to evaluate the current factory run and grade how it went against your conventions.
