What problem does it solve? Agent workflows accumulate recurring friction and repeated failures that go unnoticed without a systematic review process. This Skill turns session episodes into structured reflexion and evidence-based proposals for improving skills and agent instructions over time. ## Core Features & Use Cases - L1 Reflexion Loop: Reflect on completed tasks, capture what worked and what failed, and record lessons as episodes in memory. - L2 Evidence-Based Refinement: Identify recurring patterns across 5+ episodes, root-cause them, generate A/B instruction variants, evaluate them against historical evidence, and propose the winner for human approval. - Governed Improvement Pipeline: Route approved proposals through an auto-composed pipeline with evidence review, ensuring no silent instruction changes and no kernel modifications. - Use Case: After several sessions show entropy alerts during auth module changes, mine the episodes, trace the cause to missing test-dependency discovery in the context-map skill, and propose a validated instruction fix. ## Quick Start Ask the agent to review recent session episodes for recurring friction and propose an evidence-backed improvement to the relevant skill instructions.
