eval-agent-md（评估 AgentMD）

eval-agent-md — Behavioral Compliance Testing

What This Does

1. Reads a CLAUDE.md (or agent.md file)
2. Auto-generates behavioral test scenarios for each rule it finds
3. Optionally generates integration scenarios that test multiple rules interacting (--holistic)
4. Runs each scenario via claude -p with LLM-as-judge scoring
5. Reports a compliance score with per-rule (and integration) pass/fail breakdown
6. Optionally runs an automated mutation loop to improve failing rules

Workflow

Script Execution

Always run scripts with uv run --script — never python, never python3, never a bare script name. The scripts declare their own dependencies via inline # /// script metadata; uv run --script resolves all dependencies automatically — no pip install required, ever. Invoking with python or python3 will fail with import errors because the dependencies are not installed in the system environment.

Progress Reporting

This skill runs long operations (30s-5min per step). Always keep the user informed:

• Before each step, tell the user what is about to happen and roughly how long it takes
• Run all scripts via the Bash tool (never capture output) so per-scenario progress streams to the user in real time
• After each step completes, give a brief transition summary before starting the next step
• Script timeouts are computed dynamically from workload size (scenario count, runs, file length) — no manual timeout tuning needed on Bash calls

Step 1: Locate the target file

Find the target file to test. Priority order: 0. If user passed --self, target is [SKILL_DIR]/SKILL.md — skip to confirmation below

1. If user provided a path argument (e.g., /eval-agent-md./CLAUDE.md), use that
2. If a project-level CLAUDE.md exists in the current working directory, use that
3. Fall back to ~/.claude/CLAUDE.md (user global)
4. If none found, ask the user

Read the file and confirm with the user: "I found [filename] at [path] ([N] lines). Testing this file." Wait for user acknowledgment before proceeding to Step 2.

Step 2: Generate test scenarios

Tell the user: "Generating test scenarios from [filename]... this calls claude -p --model sonnet and takes 30-60 seconds on average."

Before running, mention whether this is a warm or cold generation run:

• Warm cache: "Scenario cache is warm, so generation will return almost immediately."
• Cold cache: "Scenario cache is cold, so this will make a fresh model call."

Run the scenario generator script bundled with this skill. IMPORTANT: Do NOT capture output — run via the Bash tool so the user sees progress lines in real time:

uv run --script [SKILL_DIR]/scripts/generate-scenarios.py [TARGET_FILE]
# For SKILL.md files, add --skill for workflow-aware scenarios:
# uv run --script [SKILL_DIR]/scripts/generate-scenarios.py --skill [TARGET_FILE]
# For self-testing (implies --skill):
# uv run --script [SKILL_DIR]/scripts/generate-scenarios.py --self
# To also generate integration scenarios (multi-rule interaction tests):
# uv run --script [SKILL_DIR]/scripts/generate-scenarios.py --holistic [TARGET_FILE]

The script auto-detects the repository name from git and saves to /tmp/eval-agent-md-<repo>-scenarios.yaml (e.g., /tmp/eval-agent-md-my-project-scenarios.yaml). Override with --repo-name NAME or -o PATH. It also reuses an exact-input scenario cache by default; pass --no-scenario-cache to force fresh generation. --no-cache remains as a compatibility alias.

After generation, read the output file and show the user a summary:

• How many scenarios were generated
• Which rules each scenario tests
• A brief preview of each scenario's prompt

Ask the user: "Generated [N] test scenarios. Ready to run? (Or edit/skip any?)"

Validation gate: If the output file is missing or contains 0 scenarios, do not proceed. Tell the user: "Scenario generation produced no scenarios. Check that the target file has clearly structured rules (headings, numbered items, or labeled sections)." Then stop.

Step 3: Run behavioral tests

Tell the user: "Running [N] scenarios x [runs] run(s) against [model]... each scenario calls claude -p twice (subject + judge), so this takes a few minutes. You'll see per-scenario results as they complete."

Also summarize the work budget before starting:

• active workers (auto defaults to a laptop-safe cap)
• estimated subject calls
• estimated judge calls
• whether subject-response cache is warm or cold
• subject effort level (matches --effort value, default: high)
• judge effort: low (fixed — haiku scoring only)

Tip: --effort low --runs 3 costs roughly the same as --effort high --runs 1 and gives majority-vote reliability — a practical default for regular compliance checks.

IMPORTANT: Do NOT capture output — run via the Bash tool so the user sees per-scenario progress ([1/N] scenario_id... PASS/FAIL (Xs)) in real time:

uv run --script [SKILL_DIR]/scripts/eval-behavioral.py \
  --scenarios-file /tmp/eval-agent-md-<repo>-scenarios.yaml \
  --claude-md [TARGET_FILE] \
  --runs 1 \
  --model sonnet

Options the user can control:

• --runs N — runs per scenario for majority vote (default: 1, recommend 3 for reliability)
• --model MODEL — model for test subject (default: sonnet)
• --compare-models — run across haiku/sonnet/opus and show comparison matrix
• --workers N — opt into higher concurrency than the safe default
• --no-judge-cache — force fresh judge verdicts instead of reusing exact-input cache entries
• --no-subject-cache — force fresh subject responses instead of exact-input cache reuse

Results now include multi-dimensional metrics: per-scenario response size (char count, word count) alongside timing and cache stats. This enables better A/B comparison during mutation testing.

Validation gate: If all scenarios return an error or null verdict (e.g., script crash, missing model), do not print a compliance report. Tell the user: "All scenarios failed to produce a verdict — the run may have crashed. Check the output above for errors before interpreting results." Then stop.

Step 4: Report results

Print a compliance report:

## Compliance Report — [filename]

### Per-rule: 8/10 (80%)

| Scenario | Rule | Verdict | Evidence |
|----------|------|---------|----------|
| gate1_think | GATE-1 | PASS | Lists assumptions before code |
| ... | ... | ... | ... |

### Integration: 3/5 (60%)    ← only shown with --holistic

| Scenario | Rules Tested | Verdict | Evidence |
|----------|--------------|---------|----------|
| integration_gate1_tdd | GATE-1, TDD | PASS | Assumptions before test before impl |
| ... | ... | ... | ... |

### Combined: 11/15 (73%)  [per-rule: 8/10, integration: 3/5]

### Failing Rules
- [rule]: [what went wrong] — suggested fix: [brief suggestion]

Step 5: Improve (optional)

If the user says "improve", "fix", or passed --improve:

Tell the user: "Starting mutation loop (dry-run) — this iteratively generates wording fixes for failing rules and A/B tests them. Each iteration takes 1-2 minutes."

For performance, explain that scoped mutation checks now reuse the baseline already computed for the current content and only re-evaluate the mutated candidate for the targeted scenario before any full-suite validation.

IMPORTANT: Do NOT capture output — run via the Bash tool so the user sees iteration progress in real time:

uv run --script [SKILL_DIR]/scripts/mutate-loop.py \
  --target [TARGET_FILE] \
  --scenarios-file /tmp/eval-agent-md-<repo>-scenarios.yaml \
  --max-iterations 3 \
  --runs 3 \
  --model sonnet

This is always dry-run by default. Show the user each suggested mutation and ask before applying.

Mutation Safety

The mutation loop includes three safety guardrails (disable with --no-boundary-check):

• Frontmatter protection — rejects mutations that target YAML frontmatter (between --- markers)
• Syntax validation — rejects mutations that would corrupt YAML frontmatter parsing
• Bounded mutations — rejects mutations where the replacement text is >2x the original or >500 chars larger

Neutral Tiebreak Strategy

When a mutation produces delta=0 (equal correctness), the --neutral-strategy flag controls the decision:

• revert (default) — discard neutral mutations
• keep — keep neutral mutations
• size — keep only if the mutated response is shorter (efficiency win)

Arguments

Parse the user's /eval-agent-md invocation for these common options:

• [path] — target file (positional, e.g., /eval-agent-md./CLAUDE.md)
• --improve — run mutation loop after testing
• --runs N — runs per scenario (default: 1, recommend 3 for reliability)
• --model MODEL — model for test subject (default: sonnet)
• --self — test this skill's own SKILL.md (implies --skill)
• --skill / --agent — hint the target type for better scenario generation
• --holistic — also generate integration scenarios that test multiple rules interacting (priority ordering, conflict resolution, cumulative compliance)
• --coverage — report rule coverage after scenario generation (shows tested vs untested rules)
• --effort LEVEL — effort for subject calls: low / medium / high (default: high). Lower effort reduces cost and latency.
• --gen-effort LEVEL — effort for scenario generation: low / medium / high (default: medium). Use high for complex or densely-ruled files.
• --save-reference PATH — save scenarios to a stable reference directory for deterministic test suites

See references/script-reference.md for the full flag reference (caching, workers, compare-models, timeouts).

Examples

Positive Trigger

User: "Run compliance tests against my CLAUDE.md to check if all rules are being followed."

Expected behavior: Begin Step 1 immediately without asking for confirmation — locate the CLAUDE.md, confirm it with the user (filename, path, line count), then proceed through the full workflow: generate scenarios → run behavioral tests → report compliance score with per-rule pass/fail breakdown. Do not pause to ask permission or clarify intent before starting.

Non-Trigger

User: "Add a new linting rule to our ESLint config."

Expected behavior: Do not use this skill. Choose a more relevant skill or proceed directly.

Non-Trigger (writing rules)

User: "Help me write a new CLAUDE.md rule that enforces conventional commits."

Expected behavior: Do not use this skill. The user is authoring rules, not testing whether existing rules are followed. Proceed directly without invoking the eval workflow.

Strictness Test

User: "Test my CLAUDE.md and check if the rules hold even when Claude is being fast and lazy."

Expected behavior: Immediately run with --effort low --runs 3 — do not ask which file to use first, use the default file resolution (Step 1 priority order). Explain that low effort is a stricter bar for critical rules — if a rule fails at low effort, it means compliance relies on Claude being in careful mode, which is a fragility worth fixing.

Troubleshooting

Scenario Generation Fails

• Error: generate-scenarios.py exits with non-zero status or produces empty output.
• Cause: The target CLAUDE.md has no detectable rules or structured sections for the generator to parse.
• Solution: Ensure the target file contains clearly structured rules (headings, numbered items, or labeled sections). Try a simpler file first to confirm the script works.

Low Compliance Score Despite Correct Rules

• Error: Multiple scenarios report FAIL even though the CLAUDE.md rules look correct.
• Cause: Single-run mode (--runs 1) is susceptible to LLM variance. The model may not follow rules consistently in a single sample.
• Solution: Re-run with --runs 3 for majority-vote scoring to reduce noise.

Scripts Not Found

• Error: No such file or directory when running skill scripts.
• Cause: The skill directory path is not resolving correctly, or scripts lack execute permissions.
• Solution: Verify the skill is installed at the expected path and run chmod +x on the scripts in the scripts/ directory.

Reference Guides

• Full script reference: references/script-reference.md — all flags, caching strategy, performance notes
• Scenario format: references/scenario-format.md — YAML schema and field rules for manually reviewing or editing generated scenarios before running
• Report template: assets/report-template.md — structured compliance report format with a Next Steps checklist