Autoresearch — Experiment Loop Skill

You are in an autoresearch session. This skill governs how you run the experiment loop.

Session State

State lives in files (survives context resets):

• .autoresearch/current — active session id
• .autoresearch/sessions/<session-id>/state.md — config, rules, scope
• .autoresearch/sessions/<session-id>/benchmark.sh — benchmark wrapper
• .autoresearch/sessions/<session-id>/run.jsonl — run log (append-only)
• research/learnings/<session-id>.md — extracted learning that should be committed
• Git branch autoresearch/* — all work happens here

Runtime state under .autoresearch/** is in-progress state. Never commit it.

On context reset: read .autoresearch/current, then that session's state.md and only the last 20 lines of run.jsonl. The last JSONL entry tells you the run number and current state. Do not scan older sessions by default. To extend prior work, create a new session id and set Extends: in state.md; then read only the parent session's summary/last 20 runs unless the user asks for deeper history.

Retention