| name | subagent-review |
| description | This skill should be used when the user asks for "multi-plane review", "review with different LLMs", "pi/opencode review panel", "review code/spec/idea through subagents", or wants independent reviewers for code, product, UX, evidence, security, or requirements. |
| version | 0.1.0 |
Subagent Review
Run independent review planes through codex-subagent. Keep this skill read-only by default. Use it for code, specs, product ideas, prose, implementation evidence, and PRs where one blended opinion would hide disagreement.
Contract
- Use one
context-pack/v1 for the shared target.
- Use one
role-card/v1 per review plane.
- Require
subagent-result/v1 structured output for every reviewer.
- Keep each reviewer independent. Do not include other reviewers' conclusions in the first pass.
- Treat missing tests, CI, screenshots, or runtime proof as
not_assessed.
- Do not treat reviewer output as approval to merge, ship, or widen scope.
- Preserve reviewer conflicts instead of flattening them into a false consensus.
For context and role details, read ../references/context-and-roles.md when needed.
Review Plan
- Identify the target: diff, branch, files, spec, PR, idea, or artifact.
- Define planes. Prefer 3-5 focused planes:
requirements: match against stated goal and acceptance criteria.code: correctness, maintainability, edge cases.ux: user path, wording, friction, expectation mismatch.evidence: tests, CI, logs, screenshots, reproducibility.security: secrets, command execution, auth, filesystem boundaries.
- Build a context pack:
codex-subagent context build \
--subject "<target>" \
--mode review \
--goal "<review goal>" \
--file <path> \
--rule AGENTS.md \
--diff \
--out context.json
- Run the panel with built-in advisory role templates:
codex-subagent panel run pi \
--context-pack context.json \
--role requirements-reviewer \
--role code-reviewer \
--role evidence-reviewer \
--role security-reviewer
- Aggregate structured results:
codex-subagent panel results <panel-id> --structured
- Inspect child runs when a role is failed, partial, empty, or surprising:
codex-subagent inspect <run-id>
codex-subagent logs <run-id> --stream stderr
codex-subagent result <run-id> --structured
- Synthesize findings by severity and evidence. Mark unusable, empty, hung, or off-task reviewer output as
not_assessed.
Output
Return:
- review status:
assessed, partial, or not_assessed
- reviewer table: role, model/runtime, status, run id
- findings:
critical, major, minor
- conflicts between reviewers
- evidence gaps
- recommended next action
