What problem does it solve? Weekly agent work produces scattered sessions and plans, making it hard to tell which workflow habits actually helped, which frictions keep recurring, and which lessons deserve to become durable rules versus one-off incidents. ## Core Features & Use Cases - Evidence-graded learning loop: Runs an Evidence → Friction/Success → Root Cause → Candidate → Promote/Reject → Eval pipeline with E0–E4 evidence grading so only repeated, costly patterns become rules. - Skill routing and context hygiene audit: Separates Skill Selection failures from Execution failures and checks context pollution, session boundaries, and verification gaps. - Promotion discipline: Distinguishes cross-project reusable constraints from project-specific exceptions, and never edits global files like AGENTS.md or CLAUDE.md without explicit approval. - Use Case: After a week of AI-assisted development, run a retrospective that reconstructs 1–3 workstreams, lists top frictions with root causes, proposes 2 promotable workflow constraints with triggers and scope, and defines next-week verification hypotheses. ## Quick Start Run a week retrospective over my sessions and plans from the last 7 days and report frictions, root causes, and candidate workflow rules.