Dual-AI Pull Request Review

Use AskUserQuestion to ask the user two questions in a single prompt:

1. Which Codex model: gpt-5.4 (recommended), gpt-5.3-codex (complex software engineering), or gpt-5.4-mini (fast/lightweight)
2. Which reasoning effort: xhigh, high, medium, or low

Phase 3: Run Both Reviews in Parallel

Launch Claude and Codex reviews at the same time — don't wait for one before starting the other.

Claude Review

Use the feature-dev:code-reviewer agent with these inputs:

• Full contents of all changed files (not just diffs)
• PR description and commit messages for context
• Focus areas: bugs, logic errors, security (XSS/injection), accessibility, code quality, project convention adherence
• Confidence-based filtering: only report high-priority issues

Codex Review

Run Codex in read-only sandbox:

echo '<review prompt>' | codex exec --skip-git-repo-check -m <model> --config model_reasoning_effort="<effort>" --sandbox read-only -C <repo-root> 2>/dev/null

The review prompt should include:

• The list of changed files with instructions to read each one
• Request for structured review with severity levels
• Focus on high-confidence issues only
• Same focus areas as the Claude review

Always append 2>/dev/null to suppress Codex thinking tokens.

Phase 4: Merge and Present Results

Combine both reviews into a unified report. Deduplicate findings that overlap (even if worded differently) and note which reviewer(s) detected each one.

Output Format

## PR #<number> Unified Review: Claude + Codex

**PR**: <title>
**Changes**: +<additions> / -<deletions> (<file count> files)

---

### Findings

| #   | Severity      | Finding           | Detected By           |
| --- | ------------- | ----------------- | --------------------- |
| 1   | **Important** | Brief description | Claude / Codex / Both |

---

### Detail per Finding

#### 1. [Important] Finding title — Detected by X

**`file:line`**

Explanation and suggested fix.

---

### Checked and Clean

- **XSS**: no issues found
- (other areas checked)

Sorting: Critical > Important > Medium > Low.

Phase 5: Fix Issues (when instructed)

Only fix issues when you determine they are warranted OR the user explicitly asks.

1. Present a concise fix plan (one line per change)
2. Apply fixes incrementally — one issue at a time, minimal changes
3. Run relevant tests after each fix
4. Run the full test suite after all fixes

Phase 6: Post-Fix Re-Review via Codex Resume

After fixes are applied, use Codex resume to re-review. This preserves the original session context so Codex can verify its earlier findings are resolved and catch any new issues from the fixes.

echo 'I have applied fixes for all issues you identified. Please re-review the changed files to verify the fixes are correct and check for any new issues. The fixes were: <brief summary>' | codex exec --skip-git-repo-check resume --last 2>/dev/null

The resume command must not include model/effort/sandbox flags — those are inherited from the original session.

If the re-review finds new issues, present them and fix if needed (repeat Phase 5-6).

Key Principles

• Codex is a peer, not an authority. If Codex flags something you believe is wrong, say so with evidence and let the user decide.
• Do not post to GitHub unless the user explicitly asks. Default is local-only output.
• Parallel execution matters. Both reviews should run simultaneously to minimize wall-clock time. Use the Agent tool for the Claude review and Bash for the Codex review in the same message.
• Explain disagreements. When Claude and Codex disagree, present both viewpoints and your assessment of which is correct.