Codex Review

Independent cross-model code review using OpenAI Codex (GPT-5.6 Sol, xhigh reasoning). Run from inside Claude Code to get a second opinion from a different model family, optionally alongside Claude-side persona reviewers.

Why This Exists

Codex reviews from a different model family than Claude, which means different blind spots. It does NOT see Claude's planning notes or prior review comments — only the code and diff. This anti-anchoring property is what makes the ensemble valuable.

Instructions

Step 1: Determine Review Scope

Check what the user wants reviewed:

• If $ARGUMENTS contains a file path or directory: Review that specific path
• If $ARGUMENTS contains "pr" or "branch": Review all changes on the current branch vs main
• If $ARGUMENTS is empty: Review all uncommitted changes (staged + unstaged)
• If $ARGUMENTS contains "panel": Review all changes since the last commit on main

Step 2: Collect the Diff and Changed Files

# For uncommitted changes:
DIFF=$(git diff HEAD)
CHANGED_FILES=$(git diff HEAD --name-only | grep '\.py$')

# For branch changes:
DIFF=$(git diff main...HEAD)
CHANGED_FILES=$(git diff main...HEAD --name-only | grep '\.py$')

# For specific path:
DIFF=$(git diff HEAD -- "$PATH")
CHANGED_FILES="$PATH"

Optional (requires the Serena MCP server — not part of this kit): if Serena is installed and there are more than 15 changed files, you can use it to build a focused review package instead of sending full files:

1. get_symbols_overview for each changed file
2. find_symbol with include_body=True for changed symbols only
3. find_referencing_symbols for each changed symbol (1 level deep)
4. Package the symbol bodies + caller context as the review content

Without Serena, the script below simply sends the diff (and full file contents for small change sets).

Step 3: Run the Codex Review

Execute the review script, passing the diff and file contents via stdin:

~/.claude/skills/codex-review/scripts/codex_review.sh

The script handles:

• Assembling diff + file content
• Piping to codex exec -m gpt-5.6-sol -c model_reasoning_effort=xhigh
• Passing the structured review prompt

IMPORTANT: Do NOT include any of Claude's reasoning, planning notes, or prior review comments in the Codex prompt. Codex must form its own independent judgment.

Step 4: Parse and Present Findings

Read the Codex output. It will be structured as:

## Codex Review — Verdict: <PASS / CONDITIONAL PASS / FAIL>

| Finding | Severity | File:Line | Description |
|---------|----------|-----------|-------------|

Present this to the user as-is. If running as part of the full review panel, this output feeds into the Panel Synthesis step where Claude merges all four reviewer outputs.

Step 5: Panel Synthesis (if running with full panel)

When Codex review runs alongside a Claude panel (personas such as Architect, CTO, and Red Team — or any personas specified for the goal):

1. Agreement matrix: Findings flagged by 2+ reviewers → bump severity (high confidence)
2. Disagreements: Findings where reviewers conflict → flag for human judgment
3. Codex-only findings: Issues only Codex caught → highlight as "cross-model catch"
4. Deduplicated findings table: Single table with source column showing which reviewer(s) flagged each issue

Codex Review Dimensions

The review prompt instructs Codex to evaluate these dimensions:

Examples

# Review uncommitted changes
/codex-review

# Review a specific file
/codex-review src/scoring/engine.py

# Review current branch vs main
/codex-review branch

# Review all changes since the last commit on main (panel scope)
/codex-review panel

Troubleshooting

• "codex: command not found": Install Codex CLI: npm install -g @openai/codex
• Authentication error: Run codex login to authenticate with OpenAI
• Timeout on large diffs: For very large diffs, raise CODEX_TIMEOUT, narrow the scope to a path/branch, or (if Serena is installed) pre-build a symbol-only review package as described in Step 2
• Empty output: Check that there are actual changes to review (git diff HEAD should show content)

Intent and focus (automatic)

Derive these from the user's request — they should not need to set anything. From how they phrase the ask, infer intent and focus and pass them via CODEX_INTENT / CODEX_FOCUS. For example, "codex-review this branch, focus on the auth logic and skip style" -> CODEX_INTENT="auth change" CODEX_FOCUS="auth logic, skip style".

The text is appended to the prompt as author framing, never as another reviewer's findings. A guardrail still surfaces every CRITICAL/HIGH finding even when it falls outside the focus.