lamarck

English | 简体中文

A production self-evolving system for agent skills. Not a one-shot optimizer for a hand-picked skill: install once, and every skill you have (hundreds, if you have them) keeps accumulating evidence from its real invocations — and evolves when the evidence gate opens. Harness-agnostic: adapters ship for Claude Code, Codex, Cursor and pi, all feeding one evidence ledger. Edits land with your approval, or — for skills you explicitly promote to auto — replay-gated and reported after the fact. lamarck evolves itself by the same rules while it runs (never on auto). Traits acquired through use are inherited back into the skill file; parts that go unused atrophy. Every irreversible step is user-authorized, ledgered, and reversible: the governance is what makes population-wide evolution safe.

npx lamarck-skill

The lane

What it keeps from them: SkillOpt’s validation gating, bounded edits and rejected-edit buffer; darwin’s git rollback, paired blind judging (escalating to a 3-judge majority only on close calls) and human-in-the-loop checkpoints.

Mechanism

1. Log (collector — a PostToolUse hook in the reference adapter): every skill invocation → data/pending.jsonl, stamped with the target skill’s content hash (ver) for per-version health windows.
2. Light loop (trigger + executor — a Stop hook in the reference adapter): at turn end the hook queues the evaluation for a headless run outside your session (default), or injects a mini protocol into the turn (isolation: inline) — never a SKILL.md reload. Four-dimension verdicts, judged