| name | test-skill |
| description | Test-drive a skill by spawning one sub-agent per incomplete task in PRD.md. Each sub-agent executes the skill for the task's Test Case and judges output against its Evaluation Criteria. |
Invocation: Test skill [SKILL-NAME] — reads PRD.md in the current working dir.
Shared files: test_log.txt (actual outputs + verdicts + fix recommendations).
Lead Agent (loop controller)
- Read
PRD.md → collect all tasks where Test Passed: false.
- Spawn one Tester Sub-Agent per task in parallel via Task tool. Pass each: skill name, skill path, task N, task goal, test case input, evaluation criteria.
- Wait for all. Read
test_log.txt. Print summary: N passed, M failed.
- If any failures: spawn Advisor Sub-Agent once. Wait. Then stop.
Sub-Agents (spawned via Task tool, general / mode: subagent)
Tester (parallelizable, once per task — do NOT use Task tool):
- Read
SKILL.md of the target skill. Understand its steps and expected outputs.
- Execute the skill for the assigned test case input: follow SKILL.md steps directly, run any tools it requires. Do not skip steps.
- Judge actual output against the evaluation criteria: PASS if criteria met, FAIL otherwise.
- If PASS: set
Test Passed: true for this task in PRD.md.
- Append to
test_log.txt: --- Task <N> ---\nInput: <input>\nActual: <output>\nVerdict: PASS|FAIL\nNotes: <notes>.
- Stop.
Advisor (once, only if failures exist):
- Read
PRD.md and test_log.txt.
- For each FAIL: identify root cause (wrong output, missing step, bad tool usage, etc.).
- Append
## Fix Recommendations to test_log.txt with one concrete fix suggestion per failure.
- Stop.
Stop Conditions
- All tasks evaluated + summary printed (+ advisor done if needed)
- Sub-agent fails repeatedly on same task → mark
[ERROR] in test_log.txt, continue
- User cancels