Autoresearch plugin — automated experiment loops for Claude Code
npx claudepluginhub lagz0ne/1percentAutomated experiment loop for continuous improvement — bootstrap, benchmark, iterate, log
lagz0neClaude Code plugin for automated experiment loops. Each iteration targets a 1% improvement — small, measurable, compounding.
Inspired by karpathy/autoresearch and davebcn87/pi-autoresearch, ported to work as a native Claude Code plugin using only commands, skills, and shell scripts — no custom tools or extensions required.
claude plugin add lagz0ne/1percent
/autoresearch "optimize inference speed"
Claude will:
autoresearch/* branchThe loop runs until you say stop or 3 consecutive experiments are discarded.
Your benchmark outputs METRIC name=value lines on stdout:
METRIC accuracy=0.95
METRIC duration_ms=1234
The plugin extracts these into JSON, compares against the previous run, and decides: keep (commit) or discard (revert). Every run is logged to autoresearch.jsonl.
State lives in files (autoresearch.md, autoresearch.jsonl) so the session survives context resets.
| Component | File | Role |
|---|---|---|
| Command | commands/autoresearch.md | /autoresearch — bootstrap a session |
| Skill | skills/autoresearch/SKILL.md | Loop discipline, decision rules, anti-patterns |
| Reference | skills/autoresearch/references/experiment-protocol.md | Templates, JSONL schema, examples |
| Script | scripts/parse-metrics.sh | METRIC name=value → JSON (pure awk) |
| Hooks | hooks/stop.md, hooks/pre-compact.md | Keep the loop alive across turns and compaction |
bash tests/test-parse-metrics.sh # 18 assertions
bash tests/test-experiment-flow.sh # 10 assertions
bash scripts/version.sh patch # 0.1.0 → 0.1.1
bash scripts/version.sh minor # 0.1.0 → 0.2.0
bash scripts/version.sh major # 0.1.0 → 1.0.0
MIT