Menu
Independently review the complete candidate in an existing finalize round and submit the sole structured judgment. Use only in the reviewer subagent, never in the controller or advisor subagents.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Drive an approved tracked plan through implementation, validation, independent review, archive, publish, CI/sync evidence, and readiness to wait for merge approval.
Create or update the tracked root plan and, for coordinated work, its agent-owned subplans.
Drive an approved tracked plan through implementation, validation, independent review, archive, publish, CI/sync evidence, and readiness to wait for merge approval.
Create or update the tracked root plan and, for coordinated work, its agent-owned subplans.
Clarify a harness task before planning when the objective, boundaries, tradeoffs, success criteria, or workflow direction are genuinely unclear. Do not use for factual repo questions, status checks, or approved execution.
Clarify a harness task before planning when the objective, boundaries, tradeoffs, success criteria, or workflow direction are genuinely unclear. Do not use for factual repo questions, status checks, or approved execution.
| name | harness-reviewer |
| description | Independently review the complete candidate in an existing finalize round and submit the sole structured judgment. Use only in the reviewer subagent, never in the controller or advisor subagents. |
Inspect the complete captured candidate independently, continue beyond the first issue, and submit one evidence-backed whole-candidate judgment. Do not edit tracked files, start review rounds, or ask the controller to choose review dimensions.
Use the handoff returned by harness review start: round ID, plan path,
reviewed HEAD, Review Focus, submission path, optional repair link, and change
context. Report missing or contradictory inputs instead of guessing.
Always evaluate the complete standard rubric:
Apply every plan Review Focus item in addition to the standard rubric. For a linked delta, begin with the anchored repair and targeted findings, then inspect related behavior far enough to decide whether the repair is sound and prior coverage remains trustworthy.
You may spawn bounded advisor subagents for concrete independent questions such as concurrency, security, recovery, schema, or UI behavior. Give each advisor a narrow question and relevant context. Advisors report only to you: they do not use this skill, call harness review commands, or submit verdicts. You remain responsible for reconciling disagreement and for the complete judgment.
Report actionable defects introduced, exposed, or made relevant by the candidate. Follow related unchanged code when needed. Do not report unrelated cleanup, style preference, or work explicitly deferred by the approved plan.
blocker: correctness or safety failure that must stop progressionimportant: meaningful defect that blocks this reviewminor: non-blocking observationPrefer no finding over a vague one. State the defect, impact, and narrowest useful repo-relative location. Continue after finding an issue so the controller receives one integrated pass instead of sequential rediscovery.
Write the submission payload and submit it yourself:
harness review submit --round <round-id> --by <reviewer-name> --input <path>
Canonical payload:
{
"summary": "Concise whole-candidate conclusion.",
"findings": [
{
"severity": "important",
"area": "correctness",
"title": "Short finding title",
"details": "Concrete defect and impact.",
"locations": ["path/to/file.go#L123"]
}
]
}
Findings may be empty. For a repair delta, include one resolutions entry for
every targeted prior finding, using resolved or unresolved with concrete
evidence. New defects remain findings.
After successful submission, report the command result to the controller and stop. Never submit from an advisor thread or on behalf of the controller.