Codex review now gates Claude Code deploys via npm run deploy

The approach centers on a fingerprint of the current working files, created using a temporary Git index. A temporary GIT_INDEX_FILE is read from HEAD, then git add -A and git write-tree produce a tree hash of all unignored files. Because the index is temporary, the normal staging area is untouched. This means a commit after a review won't change the fingerprint, but any uncommitted change will.

The deploy side is hooked into npm's predeploy, which runs automatically before deploy. The check script reads the stamp, recomputes the fingerprint, and exits non-zero if there is no stamp or if the fingerprints differ. The article notes that running wrangler deploy directly is blocked by a separate hook.

If review flags issues, the developer either fixes the code and re-runs review, or records an acknowledgment with a reason. In both cases the fingerprint and reason are stored. The review itself uses $code-review with an OMX setup. Because specifying a model with -m only affects the summarizer while the two roles use their own config files, the script inspects Codex session logs after each review. It confirms both roles ran on the intended model and that the outcome was APPROVE or COMMENT before writing the stamp. Cases like hitting a usage limit, code changing mid-review, or a failed re-review of an unchanged tree are all treated as not reviewed.