← ClaudeAtlas

codex-reviewlisted

Get an independent cross-model code review from OpenAI Codex (via the codex MCP server) on the current changes or a given scope, then synthesize it with Claude's own review into one actionable report. Use when the user asks to review / double-check / get a second opinion on the current work, or invokes /codex-review.
dougefla/claude-codex-review · ★ 0 · AI & Automation · score 70
Install: claude install-skill dougefla/claude-codex-review
# Cross-model review with Codex Codex is a *different model family*. A Claude subagent reviewing Claude inherits the same blind spots; Codex's mistakes don't correlate with Claude's. That uncorrelated check is the whole point — but **Claude stays the final judge**. Never forward Codex's verdict verbatim; verify each claim against the actual code before reporting it. Requires the `codex` MCP server (already configured globally) and a logged-in Codex CLI. ## Step 1 — Determine scope In order of precedence: 1. Arguments name files/dirs → review those. 2. Arguments name a PR → `gh pr diff <n>`. 3. Otherwise → the current working set: `git status --short` plus `git diff HEAD` (fall back to `git diff HEAD~1` if the tree is clean). 4. Not a git repo / no diff → review the files just written or discussed in this session. Record the absolute repo root — Codex needs it as `cwd`. Read the changed files **in full** before reviewing; a diff alone hides the surrounding invariants. ## Step 2 — Claude primary review Rate each dimension Pass / Concern / Fail, skipping ones that don't apply: 1. Bugs & correctness 2. Security 3. Code structure & maintainability 4. Performance 5. Type safety 6. Error handling & edge cases 7. Tests & coverage of the change 8. UI / accessibility (only if the change touches UI) For every Concern/Fail, note `file:line` and a concrete failure scenario (inputs/state → wrong output). No vague "consider refactoring". ## Step 3 — Independent Codex review C