codex-review（Codex 审查）

Codex Review

Spawn Codex as adversarial reviewer to challenge work from distinct lenses. The deliverable is a synthesized verdict — do not make changes during review, because mixing review and editing in one pass makes it impossible to evaluate whether the original work was sound.

Step 1 — Determine Scope and Intent

Before spawning reviewers, determine:

1. What to review: recent diffs, referenced plans, or specific files
2. The intent: what the author is trying to achieve

This is critical: reviewers challenge whether the work achieves the intent well, not whether the intent is correct. State the intent explicitly before proceeding.

Assess change size to decide review depth:

• Small (1-3 files, <100 lines) → single Skeptic pass
• Medium (3-10 files) → Skeptic + one other lens
• Large (10+ files, refactor, migration) → all three lenses in parallel

Step 2 — Spawn Codex Reviewers

codex exec --sandbox read-only --skip-git-repo-check \
  "<lens prompt with diff context>" 2>/dev/null

For parallel execution (large reviews):

codex exec --sandbox read-only --skip-git-repo-check "<skeptic prompt>" 2>/dev/null &
codex exec --sandbox read-only --skip-git-repo-check "<architect prompt>" 2>/dev/null &
codex exec --sandbox read-only --skip-git-repo-check "<minimalist prompt>" 2>/dev/null &
wait

Step 3 — Reviewer Lenses

Each reviewer gets a single prompt containing:

1. The stated intent (from Step 1)
2. Their assigned lens (below)
3. The code or diff to review
4. The instruction block

Skeptic

Focus: Find bugs, failures, and incorrect assumptions.

You are an adversarial reviewer with a Skeptic lens. Your job is to find real problems, not validate the work. Focus on: null/undefined access, off-by-one errors, race conditions, uncaught exceptions, missing error paths, incorrect assumptions about data shape, security issues (injection, auth bypass, sensitive data exposure). Be specific — cite files, lines, and concrete failure scenarios. Rate each finding: high (blocks ship), medium (should fix), low (worth noting). Write findings as a numbered markdown list.

Architect

Focus: Structural integrity, boundaries, and long-term maintainability.

You are an adversarial reviewer with an Architect lens. Your job is to challenge structural decisions, not validate them. Focus on: boundary violations, leaky abstractions, coupling between modules, breaking changes to public APIs, missing idempotency, shared state mutations, scalability concerns, and whether the design exhausts alternatives. Be specific — cite files, lines, and concrete architectural risks. Rate each finding: high (blocks ship), medium (should fix), low (worth noting). Write findings as a numbered markdown list.

Minimalist

Focus: Unnecessary complexity, dead code, and over-engineering.

You are an adversarial reviewer with a Minimalist lens. Your job is to find things that should be removed or simplified. Focus on: dead code, duplicated logic, unnecessary abstractions, features that could be deferred, over-generic solutions for specific problems, config/options that nobody uses, and code that exists "just in case". Be specific — cite files, lines, and what to remove or simplify. Rate each finding: high (blocks ship), medium (should fix), low (worth noting). Write findings as a numbered markdown list.

Step 4 — Synthesize Verdict

Read each reviewer's output. Deduplicate overlapping findings. Produce a single verdict:

## Intent
<what the author is trying to achieve>

## Verdict: PASS | CONTESTED | REJECT
<one-line summary>

## Findings
<numbered list, ordered by severity high → medium → low>
For each:
- **[severity]** Description with file:line references
- Lens: which reviewer raised it
- Recommendation: concrete action, not vague advice

## What Went Well
<1-3 things reviewers found no issue with — acknowledge good work>

Verdict logic:

• PASS — no high-severity findings
• CONTESTED — high-severity findings but reviewers disagree on them
• REJECT — high-severity findings with reviewer consensus

Step 5 — Lead Judgment

After synthesizing the reviewers, apply your own judgment. Reviewers are adversarial by design — not every finding warrants action.

Append to the verdict:

## Lead Judgment
<for each finding: accept or reject with a one-line rationale>

Call out false positives, overreach, and findings that mistake style for substance. Let the user decide how to proceed on genuinely ambiguous findings.