OpenAI Codex CLI v0.151.0 Changelog

This update is all about making the Codex CLI more customizable and secure. It matters most to developers who use this tool to build apps, because it gives them better ways to manage external tools, stops security rules from accidentally breaking, and ensures the AI keeps working smoothly even when switching between different models.

What's new

• Adjustable wait times for external tools: The system can connect to optional external tools (called MCP servers). You can now decide exactly how long the system should wait for these tools to load before moving on.
• Extensions can edit AI answers: If you use add-on extensions, they can now look at the information an external tool brings back and change it *before* the AI sees it. This lets you filter or fix data on the fly.
• Smarter plugin lists: If you have plugins set up for different code projects, the system now combines those settings correctly. It will also warn you if a project's plugin store is broken, without hiding the plugins that actually work.

Improvements

• Better security in remote sandboxes: A "sandbox" is a safe, locked-down space where the AI runs code so it can't mess up your computer. The system now does a much better job enforcing these rules remotely by paying attention to the actual operating system and file paths being used.
• Smarter helper bots: Sometimes the main AI creates "subagents" (smaller helper bots) to do extra tasks. The system now correctly counts the text used by these helpers against the main AI's overall limit, so things stay on budget.
• Behind-the-scenes upgrades: The developers added invisible tracking tools to monitor how the system handles remote tools, and they made their automated tests more reliable when the system gets really busy.

Fixes

• Fixed a security flaw with the /cd command: Previously, using a command to change folders (/cd) could accidentally weaken the security sandbox. The system now makes sure your security permissions stay exactly the same when you move around.
• Fixed tool and brainpower glitches: When switching from one AI model to another (or falling back to a weaker one if the main one is busy), the AI would sometimes lose track of what tools it could use or how hard it should be thinking. This is now fixed.
• Fixed broken error messages: If an external tool failed, the detailed error messages were getting lost. Now, those structured errors are safely passed back so you can see exactly what went wrong.
• Fixed stale security approvals: The "Guardian" (the system's security bouncer) used to sometimes approve an action based on old rules, even after you changed the permissions. It now double-checks the current rules before letting the AI do anything.

Breaking changes

No breaking changes in this release.