Bububuger/claude-review-codex

Review Gate (Codex subagent fallback)

Temporarily replace the local Claude CLI gate with a Codex subagent review. This is a fallback second-pass review, not true heterogeneous-model review. Use it until the local Claude path is restored.

Prerequisites

• Code is implemented and passes local validation (npm run check && npm test)
• Changes are committed to the current branch
• origin/main is available for diffing
• spawn_agent, send_input, wait_agent, and close_agent are available in this session

Invocation Timing

Invoke after WORKFLOW.md Step 3a (auto-review) self-review passes and before approving the PR.

Input Collection

Collect the following context before invoking:

# 1. diff (relative to main)
DIFF_FILE=$(mktemp /tmp/cross-review-diff-XXXXXX.md)
git diff origin/main...HEAD > "$DIFF_FILE"

# 2. AGENTS.md constraints
AGENTS_FILE="AGENTS.md"

# 3. exec-plan (parse path from issue description, if present)
# EXEC_PLAN_FILE="docs/exec-plans/bub-xxx.md"

Review Loop (max 3 rounds)

Maintain state:

• round: current round number (1–3)
• review_history: accumulated results from all previous rounds
• fix_summary: summary of changes made this round
• review_agent_id: spawned Codex subagent id

Round 1 — Initial Review

Spawn one Codex subagent and wait for its result.

Recommended spawn settings:

• model: gpt-5.3-codex
• reasoning_effort: high
• fork_context: true

Prompt the subagent with:

You are a senior code reviewer. You are reviewing code produced by another AI
agent (Codex/GPT).

Project constraints:
<contents of AGENTS.md>

Change diff:
<contents of git diff origin/main...HEAD>

Please review comprehensively:
1. Correctness — Is the logic correct? Are edge cases handled?
2. Architecture compliance — Does it follow AGENTS.md constraints
   (Contract-First, immutable design, field registry, adapter isolation, etc.)?
3. Type safety — Is TypeScript strict mode satisfied? Does `any` leak?
4. Cross-package boundaries — Are there direct internal file imports that
   violate rootDir rules?
5. Test coverage — Are new logic paths covered by tests? Consider whether
   `npm run check && npm test` is sufficient.
6. Security — Are there secret leaks or injection risks?

Classify each issue as P0 (blocking) / P1 (important) / P2 (suggestion).
Attach the specific file path and line number for each issue.
If there are no P0 issues, write NO_P0 at the end.

Output format:
### Findings
- [P0/P1/P2] file:line — issue

### Summary
- short overall assessment

If there are no P0 findings, append `NO_P0` as the final line. Otherwise omit
it.

Read the subagent result directly; no extraction script is needed.

Round 2/3 — Incremental Review

Reuse the same subagent with send_input when possible. If the agent is no longer available, spawn a fresh one and provide the full context again.

Send:

You are conducting round {round} of review.

Previous review results:
{review_history}

Changes made this round:
{fix_summary}

Incremental diff:
<contents of git diff origin/main...HEAD>

Please complete:
1. Verify whether each previous P0/P1 issue has been fixed — mark each as
   [FIXED] or [NOT FIXED]
2. Check whether the new changes introduced any new issues
3. Keep the same P0/P1/P2 classification and file:line precision

Output format:
### Previous Issue Verification
- [FIXED/NOT FIXED] issue

### New Findings
- [P0/P1/P2] file:line — issue

### Outstanding Issues Summary
- unresolved issues

If there are no outstanding or new P0 findings, append `NO_P0` as the final
line. Otherwise omit it.

Wait for completion and parse the returned text directly.

Judgment Logic

• No P0 (result contains NO_P0) → PASS — record P1 issues in workpad for reference
• P0 present and round < 3 → fix issues, commit, proceed to next round
• P0 still present after round 3 → abort, escalate to Human Review
• Convergence check → if P0 count increases compared to the previous round, abort immediately (wrong direction)

Fix Principles

• Only fix the issues that were flagged — do not take the opportunity to refactor
• Re-run local validation after each fix (npm run check && npm test) to confirm no regressions
• Record fix_summary for use in the next review round

Actions on Pass

After review passes:

1. Record in workpad under ### Cross-Review section:

   - Reviewer: Codex subagent
   - Model: gpt-5.3-codex
   - Rounds: {n}/3
   - Result: APPROVED
   - P1 remaining: {list or "none"}
   

2. Close the review subagent
3. Continue with the approve → merge flow in WORKFLOW.md

Actions on Failure (after 3 rounds)

1. Record review failure and outstanding P0 issues in workpad
2. Close the review subagent
3. Escalate to Human Review

Per-Round Output

Output in Chinese for the human user

--- Round {n}/3 · Review Gate ---
Reviewer: Codex subagent (gpt-5.3-codex)
Result: P0:{x} P1:{x} P2:{x}
{Previous round fix verification (from round 2 onward)}
{Issue list}
Changes: {fix_summary}

Rules

1. Self-review must pass first — this skill is the second gate after self-review, not a replacement for it.
2. Temporary fallback only — use this path until local Claude review is restored.
3. One dedicated subagent per review loop — reuse it across rounds when possible via send_input.
4. Hard 3-round limit — hand off to Human Review after 3 rounds to prevent infinite loops.
5. Convergence first — abort immediately if P0 count increases.
6. Small targeted fixes — only fix what was flagged; do not expand scope.
7. P0-free = pass — P1 issues are listed for reference but do not block.
8. Agent failure is blocking — retry once; if subagent review still cannot run, stop and escalate instead of skipping the gate.
9. Close the agent when done — do not leave the review subagent running.

Codex subagent Notes

| Issue | Mitigation | |-------|------------| | spawn_agent fails or no subagent is available | Retry once; if it still fails, escalate to Human Review | | Diff too large for context window | Pass only the diff for changed files, not the entire repo | | Same issue repeats across rounds | Stop at round 3 and escalate | | Agent output is verbose | Normalize it into the per-round Chinese summary before reporting |