菜单
Run the evaluation suite on the current harness or a specific candidate run. Reports deterministic check results and LLM-judge assessment.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Evolve repo-local Claude Code harness assets through a 5-phase pipeline — harvest context, propose candidate, evaluate with evidence, audit regressions, report results.
Full Meta-Harness status view — Pareto frontier, recent runs, regressions, eval health, installed plugins.
Analyze the current project and generate initial eval tasks for harness optimization. Creates regression and capability eval tasks based on project structure.
Summarize the current Meta-Harness-style frontier of harness candidates, including quality, cost, latency, and safety notes.
Audit recent regressions in the harness search and identify likely confounds, brittle edits, and safer next-step ideas.
| name | eval |
| description | Run the evaluation suite on the current harness or a specific candidate run. Reports deterministic check results and LLM-judge assessment. |
| disable-model-invocation | true |
| allowed-tools | Read Grep Glob Bash(python3 *) Bash(git *) |
Run the evaluation suite to measure harness quality.
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/eval_runner.py --eval-dir ${CLAUDE_PLUGIN_ROOT}/eval-tasks --cwd . 2>&1 || echo "Eval runner not available"
llm_judge criteria, evaluate the criteria manually:
If a specific run_id was provided as $ARGUMENTS, evaluate that candidate's artifacts in runs/{run_id}/. Otherwise, evaluate the current harness state.