Auto Review gates AI agent code edits

Employing an independent model to review code changes serves as a vital safeguard against AI hallucination and sloppy edits, but it burns tokens and should remain strictly advisory.

• –Advisory evaluation: All findings are advisory and must be verified by reading the real code path and adjacent files.
• –Engine flexibility: Supports Codex (GPT) as the default with fallback options, as well as optional engines like Claude.
• –Scope governor: Freezes a baseline to prevent agents from spiraling into endless refactoring cycles.
• –Security-minded: Safe large diffs are scanned in full, and the tool fails closed if it detects secrets or sensitive paths.