review（审查）

Review

Independently audit existing code with concern-specific lenses and decide whether it is safe to ship. Review is the gate after verify — the builder proves the change works on the real surface, then review decides whether the change is *good*.

Principles

• Prefer parallel reviewer personas when the concerns are independent
• Evidence beats taste
• Load shared doctrine from the target repo's guidance files such as AGENTS.md, CLAUDE.md, or repo rules
• Keep the final verdict tied to concrete evidence, not reviewer instinct alone
• Keep findings risk-focused; do not drown the user in low-value nits
• Track reviewer personas internally; include them visibly only when asked or when the harness has compact metadata
• If runtime proof for your own completed change is the goal, hand off to verify

Handoffs

• Self-checking a change you just authored, before handing it off for review → use verify
• Review is blocked because the repo cannot be booted or exercised reliably → use agent-readiness
• Main problem is stale AGENTS.md, README, specs, or repo docs → use docs

Before You Start

1. Define the scope: file, diff, branch, commit range, or PR
2. Load the target repo's guidance files such as AGENTS.md, CLAUDE.md, or repo rules, when present
3. Choose reviewer personas from references/reviewer-selection.md
4. Decide which personas can run independently in parallel

Default personas:

• reviewers/general.md
• reviewers/tests.md
• reviewers/silent-failures.md

Add conditional personas only when they earn their keep:

• reviewers/types.md
• reviewers/cleanup.md
• reviewers/comments.md

Persona shortcuts:

• doc-only or comment-only diffs: use general plus comments; skip tests, types, silent-failures, and cleanup unless the diff actually justifies them
• type-shape or schema changes: add types
• dead files, deprecated paths, or obviously unused helpers: add cleanup and call out deletion explicitly when warranted
• mock-heavy or shallow tests around risky behavior: make that a finding rather than treating test presence as proof

Workflow

1. Scope nearby risk

Review the requested code, but inspect adjacent behavior when the risk leaks past the named diff.

2. Run reviewer personas

Use parallel subagents when available. Keep each persona concern-focused and independent.

Concrete starting points:

• git diff --stat <base>...HEAD to size the change
• git diff <base>...HEAD -- <path> to inspect risky files
• targeted tests such as pnpm test path/to/spec when behavior claims need proof

3. Collect evidence

• Cite exact file references for static findings
• Run the smallest runtime check that changes the verdict when the repo supports it
• If something is unverified, say so explicitly instead of bluffing
• If legacy or dead code is still present, say whether it should be deleted or why it must stay
• If tests mock the main integrations or boundaries, say that the behavior is still unverified on the real surface

4. Synthesize the verdict

Order findings by severity. If no findings are discovered, say that explicitly and mention any residual risk or testing gap. Choose exactly one verdict: ship it, needs review, or blocked.

Output

After review, report a tiny verdict footer:

• verdict
• evidence summary: exact command names or runtime surfaces, not full logs
• unverified areas or readiness gaps
• next: implementation, verify, agent-readiness, or docs

Use those labels explicitly. Do not replace them with softer prose like "safe to merge" or "do not ship today".

Prefer the active harness's best native review representation instead of a prose-heavy wall of text.

Keep the final answer short:

• Put detailed issue text, file references, and line numbers in native findings or the fallback findings list
• Do not repeat native finding details in the verdict block
• Keep the verdict footer to 4 labeled lines or fewer
• Keep each label to one sentence; use comma-separated command names instead of log excerpts
• Omit scope and personas from the footer unless the user asked for them or the scope would be ambiguous without one short reviewed: line
• If there are no findings, say findings: none and keep the rest equally compact

Harness-specific presentation rules:

• Prefer the strongest structured finding format available: Codex/OpenAI native P0 / P1 / P2 / P3 cards, or a compact table in Claude/Anthropic harnesses
• If no richer primitive exists, use a short severity-ordered findings list with file/line, issue, impact, and evidence
• Never hide actionable findings inside the footer or a long prose recap

Example:

verdict: needs review
finding: high — src/auth/session.ts:42 fallback returns an anonymous session when token parsing fails
evidence: pnpm test src/auth/session.test.ts
unverified areas: runtime behavior for malformed OAuth callbacks
next: implementation

References

• references/reviewing.md — reviewer persona workflow, evidence expectations, and verdict synthesis
• references/reviewer-selection.md — which reviewer personas to run for which change shapes
• reviewers/ — specialized review lenses