Using Codex CLI as an adversarial reviewer in Claude Code with a convergence loop

Short summary

The author wired OpenAI's Codex CLI into Claude Code as an adversarial reviewer with a convergence loop: fresh reviewer sessions critique work until no new sound findings emerge. When pointed at its own implementation, the loop caught three real bugs in three rounds—including broken output capture, false convergence logic, and Windows command-line length limits. The shipped version uses a lean findings ledger with rules: fresh sessions only, adopt only sound findings, cap rounds, and report NOT CONVERGED honestly.

• •Codex CLI integrated as adversarial reviewer in Claude Code with convergence loop until no new sound findings
• •Self-review caught three real bugs in three rounds: broken output capture, false convergence logic, and Windows CLI length limits
• •Lean rules: fresh sessions only, adopt only sound findings, keep adopted-but-unfixed findings open, cap rounds and report honestly

Generated with AI, which can make mistakes.