Skip to content

README.md

Latest commit

 

History

History
95 lines (68 loc) · 2.59 KB

README.md

File metadata and controls

95 lines (68 loc) · 2.59 KB
Raw

OpenClaw plugin

OpenClaw is the default harness for SkillAdaptor (workspace / PinchBench / Claw-Eval). This repo ships the Python evolution engine; the TypeScript UI plugin lives in a sibling package.

Prerequisites (live runs)

npm install -g openclaw
openclaw gateway start
openclaw gateway status

Windows: if openclaw is not on PATH, set in secrets/.env:

OPENCLAW_CLI=C:\Users\you\AppData\Roaming\npm\openclaw.cmd

Load secrets from repo root:

. .\scripts\load_secrets.ps1

Required for live evolve: chat + embedding keys. Claw-Eval also needs official judge credentials (CLAW_EVAL_JUDGE_* or OPENROUTER_API_KEY serving google/gemini-3-flash-preview).

Skill inject surface (written by harness):

~/.openclaw/workspace/skills/skill-adaptor-evolved/SKILL.md

Install (UI plugin)

TypeScript plugin (UI):

openclaw plugins install <ABS_PATH>/SkillAdaptor-openclaw/src/skill-adaptor-ts
openclaw plugins enable skill-adaptor

Python bridge — copy into the TS package:

skill-adaptor/plugin/python/run_openclaw_evolve.py
  → SkillAdaptor-openclaw/src/skill-adaptor-ts/python/run_openclaw_evolve.py

The bridge calls run_plugin.py and prints EVOLVE_OUTPUT_FILE=... for the UI parser.

Configure

openclaw.json or plugin UI:

{
  "skillAdaptorRoot": "<ABS_PATH>/skill-adaptor/skill-adaptor",
  "pythonCommand": "python",
  "benchmarkEnv": "pinchbench",
  "allAsTest": false,
  "maxIterations": 2
}
Field Meaning
skillAdaptorRoot Directory containing run_plugin.py
benchmarkEnv pinchbench, claw-eval, workspace, or webshop
allAsTest false for train/val split; true for quick probes only
maxIterations Evolution loop cap

CLI (no TS UI required):

cd skill-adaptor
python run_plugin.py init --workspace ../my-workspace --harness openclaw
python run_plugin.py --workspace ../my-workspace --harness openclaw --dry-run
python run_plugin.py --workspace ../my-workspace --harness openclaw --max-iterations 2

Benchmarks via OpenClaw harness

Env Extra env Notes
workspace none Tasks in input_task/*.md
pinchbench PINCHBENCH_PATH Gateway required
claw-eval CLAW_EVAL_PATH + judge keys + --task-manifest OpenClaw inject + claw-eval CLI grade

Do not collapse Claw-Eval to “OpenClaw chat only” — graders need claw_eval.cli run + official judge.

Secrets

See root .env.example. Never commit secrets/.env or local task-ID manifests.