Codex 0.155.1 Fixes: Proxies, Bedrock, and TUI Cleanup
This episode breaks down the Codex 0.155.1 fixes, including the reasoning summary default change that stopped proxy-related 400 errors, plus new ways to enable summaries when you want them.
It also covers Bedrock credential refresh via shell commands, safer Guardian authorization handling, and faster workspace cleanup shortcuts in the agents overview TUI.
Chapter 1
Codex 0.155.1 Reverts Reasoning Summaries to Fix Provider Request Rejections
Ethan Park
Big thanks to Jellypod for helping us bring you the show today. So, Maya, if your terminal started throw, throwing random HTTP four hundred errors yesterday morning out of nowhere, you were definitely not alone.
Maya
Oh, I saw the panic online! People were, like, tweaking their config files, rolling back versions... what actually broke?
Ethan Park
So, when OpenAI dropped Codex version zero point one hundred fifty five point zero, they enabled streaming TUI reasoning summaries by default. Which sounds great on paper, right? Live summaries in your terminal status row as the model thinks.
Maya
Right, but if you're not pointing directly at OpenAI's native endpoints...
Ethan Park
Exactly. If you're routing through Azure APIM, or Bedrock proxies, or local custom gateways, those strict API proxies had no idea what to do with that new reasoning summary payload parameter. So they just slammed the door shut with a four hundred Bad Request.
Maya
A classic schema mismatch! The proxy sees an unexpected field in the JSON body and goes, nope, I don't know you, rejected.
Ethan Park
Yep! Which brings us to pull request forty six thousand four hundred sixty seven in zero point one hundred fifty five point one. Celia from OpenAI pushed a quick fix that restores none as the default reasoning summary setting for new local TUI sessions.
Maya
So it stops automatically injecting that reasoning summary field into every outbound request.
Ethan Park
Right, exactly. It leaves reasoning summaries disabled by default, fixing request rejections by providers that don't support them. But, and this is crucial, it still respects explicit reasoning summary settings if you configured them.
Maya
That's huge for enterprise setups. Because if you're on a corporate gateway, you don't want to spend three hours writing custom wrapper scripts or hacky CLI flags just to keep your agent from failing mid turn.
Ethan Park
Yeah.
Maya
Wait, so if someone is hitting native OpenAI models directly and they actually want those live reasoning summaries back, how do they turn them on?
Ethan Park
Super simple. You can open up your config point toml file in dot codex and set reasoning summary equal to auto, or enabled. Or if you just want it for a single run, pass the flag double dash reasoning summary auto when you launch the CLI.
Maya
Auto just detects if the underlying model supports it?
Ethan Park
Exactly. If you're targeting native OpenAI reasoning models, auto will stream the summaries down nicely without breaking third party proxies.
Chapter 2
Bedrock Auth Helper Commands and Agent Overview Management
Maya
Okay, but zero point one hundred fifty five point zero wasn't just about bug fixes. There was a massive quality of life update for Amazon Bedrock users, right?
Ethan Park
Oh, PR forty four thousand twenty eight is a game changer if you run on AWS. Amazon Bedrock can now obtain AWS credentials directly from configured shell commands.
Maya
Wait, like aws vault or custom SSO helper scripts?
Ethan Park
Yes! You set up a bedrock point credentials section in your configuration, and tell Codex what command to run to fetch fresh keys. It handles the command execution, caches the tokens, watches for expiration, and automatically refreshes them.
Maya
No way! So if your SSO token expires thirty minutes into a long, autonomous coding task...
Ethan Park
It doesn't crash the whole thread! It catches the authentication failure, re-executes your helper script in the background, grabs fresh temporary credentials, and recovers mid turn.
Maya
That is, uh, that is going to save so many engineers from coming back to a dead agent after lunch!
Ethan Park
Seriously. And speaking of keeping long agent turns from dying unexpectedly, they also patched Guardian authorization reviews in PR forty four thousand four hundred eighty two.
Maya
Guardian is the security reviewer that checks proposed actions before executing them, right?
Ethan Park
Right. Previously, if Guardian had a transient network blip or a temporary timeout while trying to review an action, it treated the review failure almost like an unsafe action finding and locked up the turn.
Maya
Wow. So a tiny hiccup in your network connection got misclassified as a security violation?
Ethan Park
Exactly. Now, automatic approval reviews distinguish transient network inspection failures from actual unsafe action findings. It retries transient errors gracefully and preserves complete action evidence so authorization isn't lost.
Maya
That makes total sense. Don't punish the agent just because the wifi dropped for two seconds.
Ethan Park
Precisely.
Maya
Now, what about workspace clutter? Because whenever I run multiple background agents, my git tree looks like a disaster zone after a few hours.
Ethan Park
Well, PRs forty four thousand four hundred twenty four and forty four thousand four hundred thirty three fixed that in the agents overview TUI. They added clean single key workspace management shortcuts.
Maya
Oooh, tell me the keys!
Ethan Park
Press h to hide a task from your list, a to archive it, and d to delete.
Maya
Wait, if you press d, does it just wipe the directory?
Ethan Park
It checks worktree ownership first! It will prompt for confirmed deletion of clean managed worktrees, ensuring you don't accidentally blow away uncommitted work while cleaning up background tasks.
Maya
Safety rails on workspace cleanup. Love to see it. Alright, go grab zero point one hundred fifty five point one and update those configs!
Ethan Park
Catch you all next time.