Codex Review

Runs a code review through the OpenAI Codex CLI and returns findings. Its value is being a different model — it exists to catch what Claude's own review misses, not to be a better Claude review. That is the whole scope:

• Use it as /deep-review's second reviewer, or when the user explicitly wants a non-Claude read on a diff.
• Don't use it for general review, child self-review, or as a default reviewer anywhere. The built-in /code-review is faster, runs in its own context, and needs no external CLI.

Codex runs read-only. It never edits files, and this skill never commits — it reports findings to whoever called it. When the user invokes it standalone with no caller to route to, apply the clearly-correct fixes yourself and commit.

Concurrency guard

Every launch below is wrapped in bash $HOME/.claude/scripts/codex-guard.sh --wait <secs> -- <launch...> — a machine-wide flock semaphore with 2 slots, so a burst of parallel agents each reviewing can't stampede the machine. Never pass a per-caller --slots; all callers depend on the shared default.

The guard sits outside the whole timeout … node "$CODEX_COMPANION" … invocation, because the companion's broker-busy retry path can spawn an extra app-server under contention.

Guard-timeout → fallback contract: if no slot frees within --wait, the guard exits 75 without running codex, which drops to the fallback like any other nonzero exit. The manager launch waits 300s (it wants the cross-model read); a child waits 90s and takes the lighter fallback rather than piling on load.

Process