Claude Code gets a one-time review gate before it stops

The rule the developer started from was simple: after implementing, have a different AI review the work. Written down as prose, it was not followed, so the rule became a Stop hook (Hook-CodexGate.js) that reads the conversation log and returns decision: block once with a reason. The gate counts file edits and writes, shell rewrites, and requests to implementation subagents as changes, while leaving HANDOFF.md, memory files and temp folders out. Only a successful Codex command counts as a review, and a stop already blocked once is always let through, so the loop cannot spin.

The misjudgments the developer reproduced on 2026年9月24日 all came from reading the log too loosely. The list of changes shown on a block covered the whole conversation rather than only what came after the last Codex pass. A background Codex run counted as done at launch, so the fix was to judge by the completion notice and to ignore notices quoted inside tool output. The phrase "usage limit" appearing in the gate's own text made it look satisfied, so the fix was to count only the limit display from the script that calls Codex. Together with the token cost of stacking block messages, these point at the same lesson: a deterministic check over log strings and order is only as good as the spans it reads.

The second problem was parallel hooks. A completion-notice hook and the gate hook do not know each other's results, so completion could fire while the gate was still blocking. The developer's answer was a shared StopVerdict.js component that dry-runs the three gates (Codex, quality, Japanese) with a 12-second cutoff and withholds completion while background work has no end notice yet. Some Codex review findings were adopted; a rare case where background work longer than 3 hours could yield early completion was not.