task-autoresearch-verify

Task Autoresearch — Verify

Run the narrowest useful verification for the in-flight task-autoresearch task and capture its result in a form the report step can consume. This skill does not modify code.

Procedure

1. Pick the narrowest useful check — the one that exercises the touched surface and nothing else. Examples:

   • Unit test for the changed function: pytest path/to/test_file.py::TestClass::test_case
   • Type check on changed files only: mypy path/to/changed.py / tsc --noEmit path/to/changed.ts
   • Specific lint rule on changed lines.
   • Replay of the captured HAR/log entry that originally failed.

   If the user passed an explicit command, run that instead.

2. Run it and capture the full output (stdout + stderr + exit code).

3. If it passes, broaden one step only if the touched surface is shared (e.g., a util used by many callers). Otherwise stop — broader is not better.

4. If it fails, do not widen the verification. Diagnose the failure first; the original task loop resumes.

5. Record the result in a form the report step can consume:

   • command run
   • exit code
   • one-line summary of the outcome
   • any failing assertion or stack frame, quoted

Handoff

• Verification passed → next natural step is the task-autoresearch-report skill.
• Verification failed → return control to the main task loop with a clear handoff line: Verification failed at <where>; resuming task loop to fix.

For verification with many steps or a large output, delegate to the verifier subagent so its context stays separate from the implementation context.