claude-codex-review

English | 中文

claude-codex-review

A Claude Code skill that hands code or plans Claude just produced to Codex (the codex CLI) for an independent second-opinion review.

Why: the author and the reviewer shouldn't be the same model. Claude can have blind spots on its own output; Codex runs a different model and training path, so it often catches what Claude misses on self-review.

Requirements

• Claude Code
• codex CLI, authenticated via codex login. Tested with codex-cli 0.135.0.

Install

In Claude Code, add this repo as a plugin marketplace, then install the plugin:

/plugin marketplace add maoningge/claude-codex-review
/plugin install claude-codex-review@claude-codex-review

Restart Claude Code (or run /reload-plugins) so the skill is picked up.

Usage

The skill triggers when Claude has just written/edited code or drafted a plan and a cross-check adds value — or when you explicitly ask: "have codex review this", "cross-check with codex", "second opinion", "codex 审一下", "codex 过一遍"…

Claude asks before running (a review costs codex tokens); you confirm, then it runs the right command and evaluates the findings critically.

Path A — review code (git diff)

Preferred, with focus instructions (pure-prompt mode):

codex review "Review the current uncommitted changes only. <focus instructions>"

Flag mode (no focus prompt):

codex review --uncommitted        # staged + unstaged + untracked
codex review --base <branch>      # whole feature branch
codex review --commit <sha>       # a specific commit

⚠️ --uncommitted / --base / --commit are mutually exclusive with a PROMPT. To pass focus instructions, use the pure-prompt form above and let codex locate the changes itself.

Path B — review a plan / doc (non-diff)

codex exec --skip-git-repo-check "Critically review the plan at <path>. Identify correctness gaps, \
unstated assumptions, risky/out-of-order steps, missing edge cases, and simpler/safer alternatives. \
Tag each issue with a severity (P1/P2/P3). Do not rubber-stamp." </dev/null

• codex exec requires a git repo by default → add --skip-git-repo-check for non-git dirs.
• It reads stdin → </dev/null prevents hanging.
• When reviewing plans, codex may search the web and cite sources (deeper, but more tokens).

Parsing codex output

codex stdout is mostly noise; the real review is after the final codex marker. Findings use [P1]/[P2]/[P3] severity + file:line, and are printed twice (dedupe). Ignore the metadata block, the exec process trace, and macOS git warnings like couldn't create cache file '/tmp/xcrun_db-...'.

Core principle: no blind trust

Codex is a second opinion, not a verdict. Claude evaluates each finding: adopt the valid ones (and fix), push back on the wrong ones (with reasoning), and surface trade-offs for you to decide. It never silently applies or silently ignores everything.

License

MIT