codex-code-review-loop

name

codex-code-review-loop

description

Narrow, single-reviewer variant that drives the review loop through the Codex MCP tool (`mcp__codex__codex`) instead of the `codex review` CLI, and self-reviews when running inside Codex itself. Do not use for general review requests — the two-reviewer finishing workflow in `code-review-loop` supersedes this. Use only when the user names this skill explicitly, or explicitly asks for a Codex-MCP-only loop with no Claude pass.

Codex Code Review Loop

Codex reviews; you fix. Repeat until the review comes back clean.

Step 0: Determine execution context

Check whether mcp__codex__codex is available in your tool list.

• Available → you are running inside Claude Code. Follow the External Codex review path below.
• Not available → you are running inside Codex itself. Follow the Self-review path below.

Also note the repository root:

git rev-parse --show-toplevel

Step 1: Determine mode

git status --short

• Pre-commit mode — uncommitted changes exist (staged or unstaged)
• Post-commit mode — working tree is clean

External Codex review path (running in Claude Code)

Look up the current frontier model

Fetch https://developers.openai.com/api/docs/guides/latest-model.md and extract the exact model ID of the current frontier model. Note it — you'll use it for every Codex call and for the Assisted-by trailer.

Pre-commit mode

Start the Codex review session — call mcp__codex__codex:

Parameter	Value
model	the frontier model ID from above
cwd	absolute path to the repository root
sandbox	"read-only"
prompt	see below

Please review the uncommitted changes in this repository (both staged and unstaged).

Focus on:
- Correctness and potential bugs
- Security vulnerabilities
- Code quality and maintainability
- Edge cases

For each issue found, state:
1. File and line number (if applicable)
2. What the problem is
3. How to fix it

If you find no issues, respond with exactly: "No issues found."

Record the threadId from the response.

Fix → re-review loop — if issues were found:

1. Apply the fixes yourself (do not commit yet)
2. Continue the same Codex session via mcp__codex__codex-reply:

   • threadId: the thread ID recorded above

   • prompt:

     I've applied the fixes you suggested. Please re-review the uncommitted changes.
     Check whether the previous issues are resolved and look for any new issues.
     If everything looks good, respond with exactly: "No issues found."
     

3. Repeat until Codex responds with “No issues found.”

Track whether any fixes were applied during the loop.

Commit — commit all changes. If any fixes were applied, add this trailer (substituting the actual model ID):

Assisted-by: Codex:gpt-5.5

Post-commit mode

Determine review scope:

BRANCH=$(git branch --show-current)

• Feature branch (not main/master):

  BASE=$(git merge-base HEAD main 2>/dev/null || git merge-base HEAD master 2>/dev/null)
  # Scope: commits from $BASE to HEAD
  

• On main/master:

  git log --oneline @{u}..HEAD 2>/dev/null || git log --oneline origin/$(git branch --show-current)..HEAD 2>/dev/null
  # Scope: commits since last upstream push
  

Start the Codex review session — call mcp__codex__codex with the same parameters as above, but use this prompt (fill in the scope before sending):

Please review the code changes in this repository.

Review scope: [e.g. "the commits on branch 'feature/auth' since it diverged from main" or "the 3 commits on main since the last upstream push"]

Focus on:
- Correctness and potential bugs
- Security vulnerabilities
- Code quality and maintainability
- Edge cases

For each issue found, state:
1. File and line number (if applicable)
2. What the problem is
3. How to fix it

If you find no issues, respond with exactly: "No issues found."

Fix → commit → re-review loop — if issues were found:

1. Apply the fixes

2. Commit with a message describing what was fixed, plus the trailer:

   Assisted-by: Codex:<model-id>
   

3. Continue the Codex session via mcp__codex__codex-reply with the same re-review prompt as above

4. Repeat until Codex responds with “No issues found.”

Self-review path (running inside Codex)

When you are Codex, review the code yourself — no external MCP call needed.

Pre-commit mode

1. Read the uncommitted changes: git diff HEAD (or git diff --cached for staged-only)

2. Review the code yourself, applying the same criteria: correctness, security, quality, edge cases

3. If you find issues, fix them directly (do not commit yet), then re-read the diff and review again

4. Repeat until your own review finds no further issues

5. Commit all changes. If any fixes were applied, include the trailer using your own model ID:

   Assisted-by: Codex:<your-model-id>
   

Post-commit mode

Determine review scope the same way as in the external path above.

1. Read the relevant commits: git diff <base>..HEAD or git show as appropriate

2. Review the code yourself

3. If you find issues, fix and commit them with the trailer:

   Assisted-by: Codex:<your-model-id>
   

4. Re-read the full diff and review again

5. Repeat until your review finds no further issues

Key rules

• Never paste diffs or file contents into the Codex prompt (external path). Pass only cwd and a description of the scope — Codex reads the repository itself.
• One Codex thread per session (external path). Reuse the same threadId across all review rounds so Codex retains context about what has already been reviewed and fixed.
• Assisted-by trailer format: Assisted-by: Codex:<exact-model-id> — use the model ID exactly as it appears (e.g. Assisted-by: Codex:gpt-5.5).
• Pre-commit: accumulate all fixes before committing; the trailer goes on that single commit.
• Post-commit: commit after each fix batch; every such commit gets the trailer.