What problem does it solve? When an AI agent skill fails, gets corrected by users, regresses, or drifts out of alignment with neighboring skills, teams often lack a structured way to turn that failure evidence into concrete skill updates, regression tests, or retirement decisions. This Skill closes that failure-to-evolution loop with a repeatable, evidence-driven process. ## Core Features & Use Cases - Failure normalization and root-cause classification: Converts tool traces, user corrections, and regression reports into structured failure records classified by a nine-category taxonomy (trigger, workflow, tool contract, environment drift, safety, output contract, library drift, eval gap, stale knowledge). - Minimal compatible patch decisions: Produces an evolution decision record choosing among NO-CHANGE, PATCH, EVAL-ONLY, ROLLBACK, MERGE-SPLIT, RETIRE, or BLOCKED, with patch surface, compatibility risk, and rollback path. - Eval and regression gate generation: Drafts trigger evals (direct, indirect, negative) and functional evals (happy path, edge case, failure mode) plus regression gate thresholds. - Use Case: A PPTX skill ships slides that break after SVG-to-PowerPoint conversion. Feed the failure trace to this Skill and receive a PATCH decision adding a render QA gate, a reference rule, and a failure-mode eval fixture. ## Quick Start Ask the agent to analyze this failed agent run and turn it into a skill evolution decision with a regression eval draft.
