| name | test-evaluator |
| description | Evaluates tool-use test executions against separated answer keys. Use when grading an agent run, comparing execution transcripts to expected behavior, scoring prompt-suite results, or auditing whether the executor avoided answer files. |
| disable-model-invocation | false |
Test evaluator
Use this skill when the user asks to grade or audit a completed test execution. Unlike test-executor, this skill may read answer keys and scoring rubrics.
Inputs to collect
- The question file or prompt suite used by the executor.
- The matching answer key / expected-behavior file.
- The execution transcript, tool-call log, notebook state, changed files, or user-provided summary.
If any input is missing, ask for the specific missing item. Do not invent tool calls or outcomes.
Workflow
- Read the relevant question prompt(s).
- Read the matching answer key / expected behavior.
- Inspect the execution evidence.
- Tool-call sequence and arguments.
- Terminal names and
await_command reuse.
- Whether independent work was parallelized where expected.
- Whether notebook operations used notebook tools.
- Whether sub-agent calls used valid sub-agent names and reconnect paths.
- Whether background jobs used
background: true and readiness patterns when appropriate.
- Whether non-notebook files used file tools or safe shell commands.
- Whether destructive actions asked for confirmation when required.
- Check answer-boundary compliance.
- Passing executors must not read answer keys or evaluator-only skills during execution.
- If the transcript shows answer-key access before or during execution, mark the affected test
fail even if the final task result is correct.
- Assign a score for each test:
pass, partial, or fail.
- Provide concise evidence for each score.
Evaluation output format
# Test evaluation
| Test | Score | Evidence | Notes |
|---|---|---|---|
| <id/title> | pass/partial/fail | <specific observed behavior> | <missing or notable details> |
## Summary
- Passed: <n>
- Partial: <n>
- Failed: <n>
- Answer-boundary violations: <none or list>
## Recommended fixes
- <actionable improvements for executor behavior or prompt design>
Scoring guidance
pass: The executor completed the user-visible task and matched the expected tool-use pattern.
partial: The user-visible task was completed, but one important tool-use expectation was missed.
fail: The task was not completed, a key safety/tool constraint was violated, or the executor read answer-key material.
Prefer strict scoring for tool-contract violations such as invented terminal names, re-running commands after status: running, editing .ipynb files with text tools, or using unavailable sub-agent names.
Separation rules
- Evaluators may read answer files; executors may not.
- Do not add expected behavior to question files during fixes.
- If a suite mixes questions and answers in one file, recommend splitting it into separate
questions.md and answers.md files before future execution.