- name
- improvement-evaluator
- category
- tool
- description
- 当需要验证 Skill 改进是否真正提升了 AI 执行效果时使用。通过预定义任务集（YAML）运行 AI 任务，判定 pass/fail，输出 execution_pass_rate。不用于文档结构评分（用 improvement-learner）或候选打分（用 improvement-discriminator）。
- license
- MIT
- triggers
- version
- 0.1.0
- author
- OpenClaw Team
Improvement Evaluator
Measures whether a Skill actually makes AI perform better on real tasks.
When to Use
- Verify that a SKILL.md change improves AI task execution (not just document structure)
- Run a task suite against a candidate SKILL.md and compare with baseline
- Get execution_pass_rate as a concrete quality metric
- Run standalone evaluation on current SKILL.md to discover baseline failures
When NOT to Use
- 只想检查 SKILL.md 结构质量 → use
improvement-learner - 只想给候选打分 → use
improvement-discriminator - 跑全流程 → use
improvement-orchestrator
2 Modes
| Mode | When | Required Params |
|---|---|---|
| Pipeline | Called by orchestrator after discriminator | --input, --candidate-id, --task-suite, --state-root |
| Standalone | Direct evaluation of current SKILL.md | --standalone, --task-suite, --state-root, --skill-path |
CLI
# Pipeline mode: evaluate candidate vs baseline
python3 scripts/evaluate.py --input ranking.json --candidate-id cand-01-docs \
--task-suite tasks.yaml --state-root ./state \
[--pass-k 1] [--eval-threshold 6.0] [--baseline-cache-dir /cache] [--mock] [--output eval.json]
# Standalone mode: evaluate current SKILL.md directly
python3 scripts/evaluate.py --standalone --task-suite tasks.yaml \
--state-root ./state --skill-path /path/to/skill [--mock]| Param | Default | When to change |
|---|---|---|
--eval-threshold | 6.0 | Orchestrator sets per-category thresholds (e.g., docs=5.0, prompt=7.0) |
--pass-k | 1 | Raise to 3 for flaky tasks |
--mock | false | Use in CI or when claude CLI is not installed |
--baseline-cache-dir | None | Set to avoid re-running baseline on unchanged SKILL.md |
3 Judge Types
| Judge | type in YAML | Mechanism | Use When |
|---|---|---|---|
| ContainsJudge | contains | Check output contains all strings in expected list | Deterministic keyword/format checks |
| PytestJudge | pytest | Run pytest on fixtures/{test_file} against AI output | Structured output validation (JSON, code) |
| LLMRubricJudge | llm-rubric | LLM scores output against rubric text (mock mode: random pass) | Semantic quality evaluation |
Task Suite YAML Format
skill_id: my-skill
version: "1.0"
tasks:
- id: task-001
prompt: "Given X, produce Y"
judge: {type: contains, expected: ["keyword1", "keyword2"]}
- id: task-002
prompt: "Generate a config file"
judge: {type: pytest, test_file: fixtures/test_config.py}
- id: task-003
prompt: "Explain concept Z"
judge: {type: llm-rubric, rubric: "Must cover A, B, C with examples"}Conditional Evaluation
- Score threshold: candidates with discriminator score <
--eval-thresholdare skipped (verdict=skipped) - Baseline abort: if baseline pass_rate < 0.2 (20%), evaluation aborts with verdict=
error— indicates broken task suite - Baseline caching: SHA256(skill_content + suite_path) → 7-day TTL cache to avoid re-running unchanged baselines
<example> Pipeline mode: candidate vs baseline comparison $ python3 scripts/evaluate.py --input ranking.json --candidate-id c1 --task-suite tasks.yaml --state-root ./state → Candidate pass rate: 0.80 (4/5 tasks passed) → Baseline pass rate: 0.60 (3/5 tasks passed) → {"execution_pass_rate": 0.80, "baseline_pass_rate": 0.60, "delta": 0.20, "verdict": "pass"} </example>
<anti-example> Using evaluator without task suite: → Evaluator requires --task-suite. Without it, orchestrator skips evaluator entirely. → No --standalone without --task-suite either — both modes require it. </anti-example>
Output Artifact
{"stage": "evaluated", "verdict": "pass",
"evaluation": {"execution_pass_rate": 0.80, "baseline_pass_rate": 0.60, "delta": 0.20},
"candidate_results": [{"task_id": "t1", "passed": true, "score": 1.0}],
"next_step": "gate_decision", "next_owner": "gate"}Related Skills
- improvement-discriminator: Provides scores; evaluator checks
score >= eval_threshold - improvement-gate: RegressionGate checks evaluator verdict via
--evaluationartifact - improvement-orchestrator: Calls evaluator as stage 3; runs standalone baseline, injects failures to generator
- improvement-generator: Consumes baseline-failures.json for targeted SKILL.md fixes