A Claude reviewer and a Codex reviewer take the identical brief, in parallel. Every claim is verified against the real code before it reaches you. Weak findings get pushed back to the reviewer that made them. Live disputes are settled by the other vendor. Distilled from a review loop run 186+ times in my own sessions.
Both reviewers report; nothing is taken on faith. You get four buckets — confirmed issues, disputed-and-settled (who said what, who won, why), rejected false positives, and genuinely open questions. You decide what to fix.
The loop closes. Confirmed issues get fixed, both reviewers re-examine in their original sessions, and rounds continue until both return VERDICT: READY. Reviewers stay read-only throughout. Hard cap of five rounds.
no plugin system? copy skills/fusionthink out of the repo into ~/.claude/skills — the folder is self-contained.
Codex CLI support — same repo, ~/.codex/skills — coming next.