What problem does it solve? It automates the self-improvement loop of a skill-based system by diagnosing six quality dimensions, selecting the lowest-scoring dimension, and delegating reasoning-heavy improvement tasks to an isolated headless agent instead of relying on template-based batch scripts. ## Core Features & Use Cases - Diagnose and Dispatch: Runs a diagnostic script to score six dimensions, picks the lowest-ROI dimension, and generates a single-dimension task prompt with an explicit output contract for the dsh headless agent. - Independent Verification: Re-runs diagnostics after execution and compares measured scores against the agent's claimed results, flagging self-deception risk and rolling back only the files the agent changed via git diff. - Git-as-Memory Recording: Updates evolution-state.json with structured Python reads/writes, appends to evolution-log.md, and commits with selective git add under commit-scope discipline. - Use Case: When 20 SKILL.md files lack semantic IO_CONTRACT sections, dispatch one headless task to generate them, independently verify coverage improved, then record the cycle with kept/discarded markers. ## Quick Start Run one dsh-based evolution cycle by diagnosing the current skill scores and dispatching the lowest-scoring dimension as a headless task.
