Codex CLI Fixes Auth Leakage, TUI Crashes, and Worktree Safety

Ethan Park

Imagine switching from your personal ChatGPT account over to your company enterprise login, thinking you are completely isolated, but under the hood, your terminal daemon is still holding open live connections and model permissions from your personal session.

Maya

Wait, hold on. You mean like... background tasks running under the wrong credentials entirely?

Ethan Park

Exactly that. In the Codex CLI app server architecture, the background daemon was caching things like model catalogs, active WebSockets, and remote control sessions globally, without strictly tying them to your active auth identity.

Maya

Wow, so if you toggled between personal ChatGPT Pro and an enterprise OIDC login, those persistent remote control handles could just... sit there. According to recent release notes, switching accounts now invalidates remote control sessions, cached WebSocket state, and model catalogs belonging to the previous identity.

Ethan Park

Yeah, that fix landed in Codex zero point one fifty five point zero across pull requests forty three nine zero six and forty three eight forty four. Basically, model catalog caches are now compound keyed by provider and auth identity together. And when you trigger an account logout or switch, it forces an immediate teardown.

Maya

Which means developers can finally stop running pkill codex in their terminal every time they switch client projects. That was such a messy workaround.

Ethan Park

Right, um, the important operational caveat now is that if you do switch identities while a background turn is actively running under the old profile, Codex will abort that turn cleanly instead of letting it execute with mismatched authority.

Maya

Now, speaking of stability fixes, right after zero point one fifty five point zero came out, they had to drop zero point one fifty five point one almost immediately because of a terminal user interface issue with reasoning summaries.

Ethan Park

Oh, right! The four hundred Bad Request errors when people connected to third party providers like Amazon Bedrock.

Maya

Exactly. The release notes for zero point one fifty five point one specifically state that new local TUI sessions now leave reasoning summaries disabled by default. So unless a provider explicitly supports reasoning summaries, it won't crash your API requests.

Ethan Park

And then came the big surprise in the zero point one fifty six point zero alpha releases. Alpha six explicitly reverts the experimental voice chat feature due to instability and fixes a TUI rendering issue.

Maya

Man, that live WebRTC voice mode in the terminal was so cool on paper, with the split flap animations and microphone controls. But the audio sink desyncs and redrawing bugs were just tearing up terminal performance.

Ethan Park

It is definitely better for them to yank it back to alpha and stabilize the core TUI first. Meanwhile, we did get some great agent management upgrades in pull requests forty three nine forty two and forty four four twenty four.

Maya

Oh, the worktree safety features?

Ethan Park

Yes! The agents overview in the terminal now shows clear worktree ownership metadata, and it requires explicit confirmation before purging clean managed worktrees. No more accidentally nuking active agent workspaces.

Maya

That is a huge relief for anyone running multi agent workflows. Special thanks to Jellypod for helping make this daily changelog deep dive possible. If you are managing multi tenant developer environments, make sure you upgrade to zero point one fifty five point one today.

Ethan Park

Catch you all next time.