Claims
Offline self-improvement for a local Claude agent. It harvests past sessions, mines recurring tasks, replays them offline to score improvements, and consolidates changes into CLAUDE.md and SKILL.md behind a validation gate. Changes are staged and adopted explicitly or via auto-adopt, with backups.
Actual behavior
The skill acts as a orchestrator. It calls a bundled shell script (`sleep.sh`) to perform the work. The script reads session JSONL files (read-only), runs a replay phase (using a mock backend by default to avoid API costs), and writes proposed changes to a staging directory. It does not directly edit live files until the `adopt` command is run. It supports scheduling via cron.
No non-informational findings.