What problem does it solve? After an automated work cycle finishes, it is hard to tell where money was wasted, which steps looped or failed, and what work leaked out into manual fix-sessions. This Skill performs a structured post-mortem on a completed cycle and turns the findings into concrete, evidence-backed proposals for improving the skills and cycle definitions that caused each problem. ## Core Features & Use Cases - Deterministic forensics: Runs cycle-forensics.py to extract the step ledger, attempt tables, cost rollups, session maps, rework records, and mechanical flags without spending model tokens on re-derivation. - Leak detection: Identifies out-of-cycle fix-sessions, live rework, and uncommitted changes that reveal what the cycle failed to deliver, with correct attribution to the current or previous cycle. - Parallel investigation: Fans out up to 6 subagent investigations that read step transcripts by path and return root causes classified as skill-defect, cycle-definition, environmental, or human-choice. - Gated proposals: Produces a ranked proposal artifact (max 8) with quoted evidence and exact edits, applying changes only after explicit approval. - Use Case: After a cycle cost $32 and you hand-fixed the loading bar twice, run the retro to discover which step should have owned that work and get a minimal edit proposal for the responsible skill. ## Quick Start Ask the AI to analyze the latest cycle and propose skill improvements based on what went wrong.
