Xcode 26.4: Regressions in Intelligence features

One more issue: I'm seeing Claude Agent just randomly stop in the middle of working. After submitting a prompt, it will occasionally respond with "Let me start by understanding the current (whatever) implementation." And then do nothing. This is different than the OAuth issue, as the OAuth token is still valid in this case.

Yes I must admit the same issues. I have claude and anthropic signed in. I want to select a model and then just use that every time I click new message. I don't like the mental burden of having to select the working model every time I click a new message. I have the other issues listed as well, sometimes it just stalls and I need to send another message to get it going again.

First, thank you for taking the time to post these. We really love hearing from our developers because this helps us make the tools better.

Re: OAuth — hmm, OK we're investigating. Is there any pattern to this? Do you see this after a certain amount of time?

Re: Thinking — interesting, this is good feedback about how you're using thinking as progress tracking. We know about the issue where you can't open the popover until the thinking is done. But, sounds like you really want a "don't put thinking in a bubble" entirely because that's part of your tracking of the agent progress. Makes sense. It'd be great if you could file a feedback request specifically for this.

Re: Slowness to start — yeah, I agree that sounds weird. The only way we have to debug this at the moment is if you attach the contents of your conversation via the "Bug" button at the bottom of the transcript.

Re: Codex & Claude showing — Have you downloaded both agents using the Intelligence settings even if you haven't logged in? We support talking to the agent once you've downloaded without logging in since some enterprise configurations automatically handle authentication without logging in via Xcode.

Thanks for the detailed response! To answer your questions:

1. OAuth - I don't see a pattern other than time. It seems to lose the OAuth token 2 or 3 times per day. It also doesn't seem to depend on idle time. I can be in the middle of working and it will lose the token.

2. Yes, especially since thinking can go off the rails or go into a nearly-endless loop, this serves as a way to monitor and guide agent progress. I did file a feedback on all the issues, btw! (FB22310171 on the whole set of issues, and I also just filed FB22316631 per your request on this specific one.)

3. Next time this occurs, I will attach it to the Feedback report.

4. No, I have never downloaded or setup Codex.

5. I especially want to highlight the additional bug I reported in my Feedback report but did not post here -- the agent will take action even if I specifically tell it to ask for approval before proceding. This is concerning because there is now some internal state where the agent believes it has gotten permission even though it hasn't. (It literally says this in its thinking bubble.)

Thanks!

Hi, I have also not determined a pattern. It's often if I pause for a while. Guaranteed to need a long overnight. But also happened during a session.

I agree about the thought bubbles. They are almost pointless if you can't read them.

Claude has got slower more recently, may be due to the influx of users due to political affairs.

Re: Codex & Claude showing: Up until 26.4 beta 3 and before you selected your agent to use and then from then on you just hit new session/chat. In 26.4 rc there is no choice of active agent. All you do is click new session/chat. When you tap the button a pop over appears requesting the agent you wish to use. It's a very jarring workflow...

Hopefully this isn't coming across as too negative. The agentic development flow is amazing, it just needs the warts removed :)

I can't reply inline because it limits the characters. Maybe that needs feeding back to the website team. It feels like "we don't want people to engage when the replies are artificially curtailed"

Chiming in to say three things:

1. The Oauth token expiration happens to me as well. I can't find an exact pattern, but it seems to happen when I leave Xcode idle for a while. I'll go have lunch, come back, and suddenly I have to sign out and sign back in to Claude Agent.

2. The AI seems to "lose itself" a lot. It'll begin responding to a prompt, I'll see output in the chat, but then it'll hang at "Thinking..." for a couple minutes. Suddenly, it'll pick up again and it always starts with: "I should resume..." or "I'll resume...", as if the original session got interrupted. My internet connection is rock-solid; that is NOT the issue. This happens very frequently and it's pretty annoying because it adds minutes of overhead to what should be very quick, easy requests.

3. Unlike Xcode 26.3, version 26.4 now puts much of the AI's "thinking" into chat bubbles that cut off the text and are not expandable. This is very dumb. I WANT to see what the AI is thinking. Xcode 26.4's UI is actively harmful and in the way. The only bubbles that can be expanded are plans and checklists. There also doesn't appear to be much rhyme or reason for what goes in a bubble and what doesn't—some "thinking" text appears in the chat normally and other text is cut off in bubbles. Lay off the bubbles. Just give us the chat.

** All of the above applies to Claude Opus. I don't use Sonnet or Codex.

I can already tell you, nothing works for me anymore. I updated to macOS 26.4. This morning, it updated Xcode to 26.4 as well. As a result, I had to reinstall Codex, which I did. I logged in, and now nothing works at all. I can submit a request in the chat, sometimes I get a short response, and then it starts thinking about something and never stops. I think it is amazing what kind of unfinished mess Apple is delivering once again. This is absolutely unbelievable. I am furious that such a broken system is being provided to us developers. It would be the easiest thing for Apple to simply test Codex. If they are not even providing us with the latest Codex versions and instead releasing outdated versions with their Xcode, it is outrageous that we are given something so completely unfinished and nonfunctional. With macOS 26.3 and Xcode 26.3, I was at least able to program perfectly fine with Codex and developed several apps. Now I wake up in the morning, want to continue working on my apps, and great job, Apple, thank you for this massive bug. I cannot develop anything anymore. Codex no longer works with Xcode 26.4. I warn everyone against installing Xcode 26.4. You can assume that Codex will no longer work. At least it does not work for me. There is no way to check why it is not working. Apple blocks the entire reasoning process. We are not shown any logs, we cannot see anything. Nothing is possible. I can even specify a log path in the Codex config file, yet no logs are written there. Once again, I find it incredible how Apple treats us. They simply do not give a damn. The main thing is that Apple gets to preconfigure everything, and we developers have no control over how Codex functions within Xcode. Once again, I say thank you, Apple, for restricting us in every possible way. I am just furious. I do not understand how something this broken can be released. And I am curious how many weeks we will have to wait before Apple finally takes action. Just look at which Codex version is included in the latest Xcode release. We are now on Codex version 1.0.6. Just think about that. An ancient version with no Fast Tier available. But okay, Apple, you must know what you are doing. I am furious, just completely annoyed, and that is all I have to say.

Zorast's post comes off very angry. I definitely understand why though: Apple steals a LOT of money from its developers and they justify that by saying, "It costs money to provide all the tools and frameworks." In that context, there is no excuse for shipping broken, hamstrung tools.

Oauth token expiration is a decades-old solved problem. It was solved in Xcode 26.3, even! Allowing that to slip through QA is not excusable.

These tools aren't free. They aren't volunteer projects. There is a standard that is expected and Apple is failing to meet it.

Awhile back I used OpenAi for intelligence, though recently switched to using Anthropic Claude. Now, when launching Xcode 26.4 I am prodded with a prompt stating "Codex is not installed. Get Codex to resume the conversation". In my Xcode Settings I have ChatGPT in Xcode switched off. Likewise I have not downloaded the Codex Agent - the Get button appears which I ignore.

In my Xcode Settings I have also signed into my Anthropic account, i.e. the Sign Out button is showing.

It appears I am completely unable to use Claude for intelligence unless I download OpenAI Codex - though other users have done this, and nothing works following.

This release appears severely broken. It would be easiest and best to revert to 26.3 to allow developers continue development.

Seconding everything that's been said in this thread so far. The agentic coding is much laggier and overall unpredictable. I'm waiting 5-10 minutes sometimes for a response. It also seems to eat up my usage limits much faster.

I'm seeing the same issues with authentication. I have to sign in to Claude multiple times per day.

And I agree that hiding the thinking while it's happening leads to a lot of wasted time and tokens because I have no idea if the agent has lost its way.

Appreciate you looking into this and hopefully releasing a 26.4.1 to address these issues!

Just wanted to chime in and co-sign the issues OP brought up. The OAuth and "thinking is in a bubble" issues being the most significant for me.

Update:

I've reverted to using Claude in the Terminal. The implementation in Xcode 26.4 is just so godawful buggy that it's unusable. (I've discovered that if you quit and restart Xcode, the chat history suddenly contains TONS of text that wasn't showing up previously!)

Whatever the reason, Xcode 26.4 is just crippled. I haven't tested the new 26.5 beta because you can't ship apps from it yet.