Codex Code Review (Iterative)

Review code changes made in the current Claude Code session by sending them to OpenAI Codex. Claude and Codex iterate until the code passes review. Max 5 rounds.

---

When to Invoke

• When the user runs

  /codex-review:code

  after making code changes
• When the user wants Codex to review their implementation before committing

Agent Instructions

When invoked, perform the following:

Step 1: Generate Session ID

REVIEW_ID=$(uuidgen | tr '[:upper:]' '[:lower:]' | head -c 8)

Use this for all temp file paths:

/tmp/claude-code-${REVIEW_ID}.md

and

/tmp/codex-code-review-${REVIEW_ID}.md

.

Step 2: Gather Changed Files

Identify all files changed in the current session:

1. Run

   git diff

   (unstaged) and

   git diff --cached

   (staged) to get the full diff of changes.
2. Run

   git diff --name-only

   and

   git diff --cached --name-only

   to get the list of changed files.
3. Also check

   git status

   for any new untracked files that were created in this session.

If there are no changes detected, ask the user which files or changes they want reviewed.

Step 3: Build Review Context

Write a review document to

/tmp/claude-code-${REVIEW_ID}.md

containing:

# Code Review Request

## Changed Files
- [list of changed files with brief description of what each does]

## Intent
[Summarize the purpose of these changes based on the conversation context -- what was the user trying to accomplish?]

## Diff
[Full git diff output]

## New Files (if any)
[Full content of any newly created files not captured in the diff]

Step 4: Initial Review (Round 1)

Run Codex CLI in non-interactive mode:

codex exec \
  -m gpt-5.3-codex \
  -s read-only \
  -o /tmp/codex-code-review-${REVIEW_ID}.md \
  "Review the code changes described in /tmp/claude-code-${REVIEW_ID}.md. Focus on:
1. Bugs - Logic errors, off-by-one, null/undefined issues, race conditions
2. Security - Injection, auth issues, secrets exposure, OWASP top 10
3. Performance - N+1 queries, unnecessary allocations, missing indexes
4. Error handling - Uncaught exceptions, missing validation, silent failures
5. Readability - Unclear naming, overly complex logic, missing context

For each issue found, specify:
- File and line number (or range)
- Severity: CRITICAL / WARNING / SUGGESTION
- What the problem is
- How to fix it

If the code is solid and ready to ship, end with: VERDICT: APPROVED
If changes are needed, end with: VERDICT: REVISE"

Capture the Codex session ID from the output line that says

session id: <uuid>

. Store this as

CODEX_SESSION_ID

.

Notes:

• Use

  -m gpt-5.3-codex

  as the default model. If the user specifies a different model (e.g.,

  /codex-review:code o4-mini

  ), use that instead.
• Use

  -s read-only

  so Codex can read the codebase for context but cannot modify anything.
• Use

  -o

  to capture the output to a file for reliable reading.
• Do NOT pipe the command through

  tail

  or any other filter — let the full output be visible so the user can see Codex's progress.
• Set a timeout of at least 10 minutes (600000ms) for the Bash tool call, as Codex reviews can take a while.

Step 5: Read Review & Check Verdict

1. Read

   /tmp/codex-code-review-${REVIEW_ID}.md

2. Present Codex's review to the user:

## Codex Code Review -- Round N (model: gpt-5.3-codex)

[Codex's feedback here, organized by severity]

3. Check the verdict:
   • If VERDICT: APPROVED -> go to Step 8 (Done)
   • If VERDICT: REVISE -> go to Step 6 (Fix & Re-submit)
   • If no clear verdict but no actionable issues -> treat as approved
   • If max rounds (5) reached -> go to Step 8 with remaining concerns noted

Step 6: Fix the Code

Based on Codex's feedback:

1. Apply fixes -- address each CRITICAL and WARNING issue Codex raised. Actually edit the files to fix the problems.
2. For SUGGESTION items, apply them if they're clearly beneficial; otherwise note them for the user.
3. Summarize what you fixed:

### Fixes Applied (Round N)
- [File:line] [What was fixed and why]

4. Inform the user: "Sending updated code back to Codex for re-review..."

Step 7: Re-submit to Codex (Rounds 2-5)

Update the review document with the new diff:

1. Regenerate the diff (

   git diff

   +

   git diff --cached

   ) and update

   /tmp/claude-code-${REVIEW_ID}.md

Resume the existing Codex session:

codex exec resume ${CODEX_SESSION_ID} \
  "I've fixed the code based on your feedback. The updated diff is in /tmp/claude-code-${REVIEW_ID}.md.

Here's what I changed:
[List the specific fixes applied]

Please re-review. If the code is now solid and ready to ship, end with: VERDICT: APPROVED
If more changes are needed, end with: VERDICT: REVISE"

Note:

codex exec resume

does NOT support

-o

flag. Read the Codex response directly from stdout. Do NOT pipe through

tail

or any filter — show full output. Set a timeout of at least 10 minutes (600000ms).

Then go back to Step 5 (Read Review & Check Verdict).

Important: If

resume ${CODEX_SESSION_ID}

fails, fall back to a fresh

codex exec

call with context about prior rounds.

Step 8: Present Final Result

Once approved (or max rounds reached):

## Codex Code Review -- Final (model: gpt-5.3-codex)

**Status:** Approved after N round(s)

**Files reviewed:**
- [list of files]

**Issues found and fixed:** X critical, Y warnings, Z suggestions

[Final Codex feedback]

---
**Code has been reviewed and approved by Codex. Ready to commit.**

If max rounds were reached without approval:

## Codex Code Review -- Final (model: gpt-5.3-codex)

**Status:** Max rounds (5) reached -- not fully approved

**Remaining concerns:**
[List unresolved issues with file:line references]

---
**Codex still has concerns. Review the remaining items and decide whether to proceed.**

Step 9: Cleanup

rm -f /tmp/claude-code-${REVIEW_ID}.md /tmp/codex-code-review-${REVIEW_ID}.md

Rules

• Claude actively fixes the code based on Codex feedback between rounds -- not just relaying messages
• Default model is

  gpt-5.3-codex

  . Accept model override from user arguments (e.g.,

  /codex-review:code o4-mini

  )
• Always use read-only sandbox mode for Codex -- Codex reviews but never writes files
• Max 5 review rounds to prevent infinite loops
• Show the user each round's feedback and fixes so they can follow along
• Prioritize CRITICAL issues over WARNINGs over SUGGESTIONs
• If Codex CLI is not installed or fails, inform the user and suggest

  npm install -g @openai/codex

• If a fix would contradict the user's explicit requirements, skip it and note it for the user
• Do NOT commit changes -- leave that to the user