route-codex-review

Family: routing Status: Stable

Purpose

The execution arm of the no-self-review rule. Hands Claude's output to Codex for an independent, adversarial pass — bugs, security risks, missing edge cases/tests, broken invariants, unclear logic. Mandatory for substantial output heading to a client or production, or whenever the user asks to verify Claude's work.

Triggers

• "review the code you wrote" / "review what you just did"
• "check your work" / "look over what you wrote" / "give it a once-over"
• "second opinion" / "sanity-check what you wrote" / "double-check this"
• "adversarial review" / "have codex review" / "is this right"

Scoped to Claude's own output. Bare "audit this" / "audit my site" belongs to the SEO/QA families, not here (see references/ENGINE-CONTRACTS.md §3).

Inputs

• The output to review (diff, files, content, or plan)
• Review focus (bugs / security / correctness / completeness / all)

Steps

1. Check Codex is available: codex --version. If absent, tell the user it's off until installed and offer a structured manual review from a different lens (never a same-model self-review).
2. Bundle the context — the diff or the specific files, plus what was tried and the relevant constraints.
3. Invoke Codex:

   git diff | codex exec --skip-git-repo-check "Review this diff. Find bugs, security risks, missing tests, edge cases. Be specific and adversarial."

   For untracked code, pipe the file(s) instead of git diff. For a harder pass: ...codex exec ... "Adversarial review. Challenge the design. Prove what's broken."
4. Parse Codex's findings into severity-tagged items (Critical → nits).
5. Integrate, don't rubber-stamp. For each finding decide fix / defer / dismiss (with a reason). Apply the fixes, then re-verify.

Outputs

• A severity-tagged findings list from Codex
• The integrated outcome: what was fixed, deferred, or dismissed (with reasons)

Dependencies

• The codex CLI (or the adjacent openai-codex tool) (optional — adds an independent GPT-5.5 adversarial pass; free path: a structured manual review from a different lens, never a same-model self-review).

Notes

Claude reviewing its own substantive output repeats its own blind spots; this skill operationalizes routing that review to a different model before it ships.