Codex Review Loop

Iterative review-fix cycle: Codex reviews the diff, Claude Code fixes, repeat until LGTM.

Works with any language or framework — TypeScript, Python, Go, Rust, etc.

Prerequisites

• codex CLI installed and authenticated

  npm install -g @openai/codex
  codex --version  # verify
  

• Inside a git repo with changes on the current branch vs a base branch

Parameters

Workflow

Step 0: Auto-detect Typecheck Command

If TYPECHECK_CMD was not provided by the user, detect it before starting:

1. Check package.json scripts for typecheck or type-check → use npm run typecheck
2. Else if tsconfig.json exists → use npx tsc --noEmit
3. Else if mypy.ini or pyproject.toml exists → use mypy .
4. Else if cargo.toml exists → use cargo check
5. Otherwise → no typecheck, skip the gate

Step 1: Run Codex Review

bash ~/.claude/skills/codex-review-loop/scripts/codex-review.sh <BASE_BRANCH> /tmp/codex-review-findings.md "<FOCUS>"

Script handles: preflight checks, empty diff detection, timeout, partial output warnings.

Then read /tmp/codex-review-findings.md.

Step 2: Parse Findings

Codex output varies — it may be structured (file:line bullets) or prose. Handle both:

• Clean: output contains "LGTM", "no issues", "looks good", or has zero actionable items → go to Step 5
• Structured findings: extract file, line (if present), severity, description
• Prose findings: extract the core issue description and the file mentioned

For each finding, build a signature: file:line:first_30_chars_of_description (use unknown:0: prefix if no file/line).

Write all signatures for this round to /tmp/codex-review-seen.txt (append, one per line).

Skip any finding whose signature already exists in /tmp/codex-review-seen.txt from a prior round.

Step 3: Fix Findings

For each new (non-repeat) finding:

1. Read the file at the specified location using the Read tool
2. Understand the issue in full context
3. Apply the fix using the Edit tool
4. Skip if: subjective style preference, false positive, or outside scope of the diff
5. Note skipped findings with reason: repeat | style | false-positive | out-of-scope

After fixing all actionable findings, go to Step 4.

Step 4: Verify & Loop

If TYPECHECK_CMD is set (or was auto-detected):

<TYPECHECK_CMD>

If typecheck fails → fix the type errors before proceeding. Do not re-review with broken types.

Check loop safety:

• If ALL findings this round were repeats → stop immediately (infinite loop guard)
• If round >= MAX_ROUNDS → go to Step 5
• Otherwise → increment round counter and go back to Step 1

Clean up temp files at the start of each round:

rm -f /tmp/codex-review-findings.md

Step 5: Final Report

Output a summary table:

## Codex Review Loop — Complete

| | |
|---|---|
| **Rounds** | <N> |
| **Base branch** | <BASE_BRANCH> |
| **Findings fixed** | <count> |
| **Findings skipped** | <count> (repeat: X, style: Y, false-positive: Z) |
| **Status** | ✅ LGTM / ⚠️ Stopped at max rounds |

### Files changed
- `path/to/file.ts` — <brief description of what was fixed>
- `path/to/other.py` — <brief description>

### Skipped findings
- `file:line` — <reason>

Then clean up:

rm -f /tmp/codex-review-findings.md /tmp/codex-review-seen.txt

Loop Safety

• MAX_ROUNDS cap: Never exceed configured max (default 5).
• Repeat detection: Signatures are written to /tmp/codex-review-seen.txt and persist across rounds. If a finding's signature is already in that file, skip it. If ALL findings in a round are repeats, stop immediately.
• Type check gate: Always fix type errors before re-reviewing. Broken types produce noisy findings.
• Empty diff guard: Script exits early if there's nothing to review vs the base branch.
• No commits: Never commit or push. Leave that to the user.

Codex CLI Quick Reference

# Review diff vs base branch
codex exec review --base <branch>

# With focus prompt (focus goes BEFORE flags)
codex exec review "Focus on security vulnerabilities" --base <branch>

# Review only uncommitted changes
codex exec review --uncommitted

# Review a specific commit
codex exec review --commit <sha>

# Shorthand
codex review --base <branch>

Timeout: 1–5 min per review depending on diff size.