Developers using OpenAI’s Codex for code review are trapped in a frustrating loop where the tool continuously reports new issues, even after fixes are applied. A deep dive into the mechanics of AI-driven code auditing reveals that this isn't a bug, but a fundamental design flaw in how these reviewers handle diffs and convergence. The core issue lies in the fact that every fix modifies the codebase, creating a new diff that the reviewer treats as fresh ground, while the underlying logic often re-flags pre-existing problems.

The Illusion of New Bugs

The root cause is that the next review pass reads different code. As developers fix initial findings, the diff changes, and the review rubric—which flags bugs "introduced in the commit"—now scans the modified lines. Anthropic’s Claude Code team provided hard data on this phenomenon on October 15, 2026. By running git blame on their own internal pull requests, they discovered that approximately 59% of the issues flagged as "new" in subsequent review rounds were actually present in the codebase since the very first pass. The reviewer wasn't finding new bugs; it was finding old ones again, simply because the context had shifted.

Lack of Convergence Controls

Current documentation for major AI review tools, including Codex, GitHub Copilot, and CodeRabbit, lacks explicit stopping rules for repeated reviews. For instance, Codex’s open-source code review rubric instructs the model to "Do not stop at the first qualifying finding," yet there is no documented limit on how many times a review can iterate on the same pull request. While GitHub Copilot attempts to mitigate this by resolving comments when a commit addresses them, its documentation admits it may still repeat dismissed findings. The absence of a configurable "number of issues" limit in Codex, as noted in open request #5547, forces developers to manually manage the noise.

Four Levers to Damp the Loop

To break this cycle, engineers must implement four specific controls: a written bar for what constitutes a reportable issue, a clear definition of done, controlled timing for when the reviewer runs, and a hard stopping rule. Anthropic’s internal practice includes a "round cap" where the reviewer asks the author to reconsider the approach after three rounds. Similarly, Cursor’s Bugbot offers a oncePerPr frequency setting, and CodeRabbit pauses automatic reviews after five commits. Without these guardrails, the cost of noise multiplies, particularly in agent-driven workflows where an AI fixes code and immediately triggers another review cycle.

Key Takeaways

  • 59% of "new" findings in Claude Code reviews were traced to pre-existing code via git blame, proving that AI reviewers often re-flag old issues in new diffs.
  • OpenAI’s Codex currently lacks a documented stopping rule or configurable limit for repeated review passes, as of October 2026.
  • Effective convergence requires explicit engineering controls, such as round caps, diff-only re-reviews, and bug-class flagging, rather than relying on the AI’s default behavior.

The Bottom Line

If your AI code reviewer doesn't have a hard stop, it's not a tool—it's a noise generator. Stop letting agents loop on their own mistakes and enforce a round cap today.