OpenAI describes a new feature letting Codex Code Review read custom repository rules from AGENTS.md files so it can catch recurring, project-specific issues that generic review misses. In testing, rules written this way achieved a 98% recovery rate for identifying violations of a given convention, compared to 58.3% when Codex was given no custom guidance. The post gives concrete authoring guidance: scope rules to specific code areas, state both the invariant being protected and any permissible alternatives, and focus on consequential checks that human reviewers would otherwise have to explain repeatedly.
