Codex update is a terrible regression

What the hell happened to Codex?

The recent updates have made the product significantly worse. A tool that used to let users freely go back and continue from previous points now forces users to deal with a bunch of ridiculous internal concepts like “steer”, “branch”, and whatever other garbage state management you decided to expose.

I don’t care what your engineers call these things internally. If I typed a message, it is my conversation history. If I made a mistake, I should be able to go back and undo it.

Instead, Codex now says:

“Failed to branch because the selected prompt is a steer and cannot be branched independently.”

This is an unbelievably stupid design decision.

You built a coding agent that cannot properly undo a user’s own instructions. Think about how absurd that is. Developers constantly experiment, make mistakes, change direction, and try different approaches. Undo and rollback are not optional features. They are the foundation of a usable development tool.

Then you made it even worse by creating endless useless branches after retries and rollbacks. The conversation history becomes a mess full of garbage entries that nobody wants.

This is exactly the kind of update that makes users wonder whether anyone actually uses the product for real development before shipping changes.

The AI model can be powerful, but if the surrounding product experience keeps getting worse, the whole tool becomes frustrating to use.

Stop making the product follow your internal implementation decisions. Users don’t need your architecture. Users need a tool that works.

  • Bring back proper rollback.
  • Let users branch from where they want.
  • Stop creating useless conversation clutter.

Right now Codex feels less like a professional development tool and more like an unfinished experiment.

This update is garbage.

This is what I’m wondering about — when tools silently regress after an update, how do people catch it before all their users do?

For our own features we keep a small manual test set, but it’s not automated. Curious how others are handling this at scale.