AutoHarness
A self-learning skill layer for Claude Code that distills reusable skills from a developer's real sessions, merges near-duplicates, updates them in use, and prunes the ones that stop being used.
AutoHarness distills a coding episode into a skill from the session already underway, firing on its own once a session has done enough work, or on demand with /learn. A reflector compares a new episode against existing skills and folds same-scenario ones together instead of accumulating near-duplicates, recording which skill absorbed which so a merge is never mistaken for a deletion.
Every session opens with a grouped index of the skills it wrote, so recall does not depend on the host surfacing them, while the host's native recall is left untouched. A skill survives by being followed in later turns rather than by a held-out benchmark score, and every create or update logs its scenario and decision to a per-skill ledger; it only touches skills it generated itself, leaving anything written by the user or installed separately alone.
It runs entirely as Python with no third-party dependencies, requiring python3 on PATH for its hooks and MCP server; it installs as a Claude Code plugin and is licensed under MIT.