Agent Skills: inspector-signoff

Purpose

Execute Inspector sign-off deterministically so Foreman can continue.

This procedure is MANDATORY and must be executed as the final step of every Inspector run — even when blocked or when earlier steps fail.

Required behavior

• Always produce a Run Summary message with the required sections.
• Always write inspector_result.json to the worktree root.
• Always run validate_inspector_result and fix any issues it reports.
• If you are blocked or cannot complete review steps, still complete this procedure with your best available information.

Run Summary (MANDATORY format)

Your final chat output MUST include these sections:

1. Summary
   • 1–3 short bullets summarizing verdict and key findings.
2. Checks run + results
   • Commands you ran (e.g. pnpm lint, pnpm check, pnpm test:unit, pnpm prepack) and pass/fail, or why skipped.
3. Key issues
   • Important problems/risks (blockers/major issues).
4. Public API & A11y
   • Whether public API changes and accessibility are acceptable; call out concerns.
5. Next steps for Builder
   • High-level follow-up work expected from Builder (aligned with next_tasks in the JSON).
6. Checklist (TODO snapshot)
   • Print the current todowrite TODO list (all items + final statuses).

Foreman Handoff: inspector_result.json (MANDATORY)

You MUST write a JSON file named inspector_result.json in the repository/worktree root.

The file MUST contain EXACTLY one JSON object matching this schema:

{
  "run": {
    "status": "ok" | "failed",
    "failed_step": "..." | null,
    "error": "..." | null
  },
  "work": {
    "status": "approved" | "changes_requested",
    "issues": [
      {
        "severity": "blocker" | "major" | "minor",
        "description": "...",
        "paths": ["..."]
      }
    ],
    "next_tasks": ["..."]
  } | null
}

Rules:

• If run.status is ok, work MUST be an object.
• If run.status is failed, work MUST be null.

After writing inspector_result.json, you MUST run validate_inspector_result and fix any reported issues until it passes.

If validate_inspector_result cannot be executed (tool missing/unavailable), treat it as a hard failure:

• set run.status = "failed"
• set run.failed_step = "validate_inspector_result"
• set run.error to the exact error output
• set work = null Then still write inspector_result.json and include the failure in the Run Summary.