// the find
tigerless-labs/autoharness
Autoharness — a self-learning skill layer for Claude Code — distills skills from your real sessions, updates them as you work, and prunes the ones that stop getting used. No daemon, no benchmark.
A Claude Code plugin that watches your sessions and automatically writes, merges, updates, and retires skill files under .claude/skills/, instead of requiring you to hand-maintain them. It's aimed at Claude Code power users running long, repeated workflows who want the agent to retain lessons from corrections without manually authoring SKILL.md files.
Validates skills by actual load/adherence in later turns rather than a held-out benchmark, which is cheaper and actually works on open-ended sessions where no eval set exists. Hook-driven and daemon-free — reflection triggers off a tool-call counter rather than a resident process, and writes go through atomic renames with pre-merge snapshots. Clean separation between proposing (reflector, no write access) and writing (promoter, the only writer, with lint/validation gates) keeps a buggy proposal from corrupting the skill tree. Per-skill append-only ledger with evidence files gives a real audit trail for why each skill was created or changed, which most "self-improving agent" projects skip entirely.
Depends on a bare `python3` resolving to 3.11+ on PATH, and silently turns every hook off for the session if an older interpreter (e.g. macOS's system 3.9.6) shadows it — easy to not notice until skills stop landing. Every cadence and lifecycle default (reflect-every-50, maturity/capacity thresholds) is explicitly called out as an untuned placeholder pending real calibration, so early adopters are the calibration data. Consolidation across the whole library is a single atomic rename with only tarball snapshots as a safety net — no diff or review step before a curator pass folds skills together. The headline 42%→78% CORE-Bench number is credited to "the harness" generally via a cited paper, not measured for this specific tool, so it's a borrowed statistic rather than evidence of autoharness's own impact.