My Cursor agent has been broken all day. I am not using any Cursor extensions. The issue presents in two distinct ways:
Agent hangs on “Thinking” — I give the agent a prompt, it starts normally (thinking, reading files, etc.), then gets stuck in the Thinking state with no visible output. If left alone, it eventually times out and returns garbled output (see below).
Repeated token loop after stopping and re-prompting — If I stop the stuck agent and prompt it again, it gets stuck on Thinking again, but this time I can see its thought stream. The entire output is a repeated string: begin_of_sentence | >%< | begin_of_sentence | >%< | begin_of_sentence | >%< | …
This repeats indefinitely until it times out.
Additional Anomalous Behavior
After restarting Cursor with --disable-extensions (no extensions were installed to begin with), I navigated back to the problematic agent and prompted it to summarize the chat history. The response it generated was unlike anything I’ve seen — it appeared to hallucinate/fabricate an entirely different conversation history, claiming I had sent empty messages, and the output text was broken mid-word (e.g., “Auto\nCID”, “co\ndebase”, etc.), suggesting some kind of token/rendering pipeline failure. The output was also formatted in the chat window differently than I’ve ever seen. The transcript also could not be exported to .md normally.
Steps to Reproduce
I have not been able to identify a specific trigger — the issue started mid-session and has persisted across: Full PC restart, Cursor restart, Cursor restart with --disable-extensions.
The bug report submission would permit only one attached image, so below is another. I can also provide the .md of the chat transcript, but the forum will not accept upload.
This is a known bug with Auto mode. The team is already tracking a few similar reports.
As a workaround, try switching from Auto to a specific model, like Composer 1.5 or GPT-5.2. It looks like the issue is tied to Auto routing, so picking a specific model should bypass it.
Also, try starting a new chat instead of continuing the broken one. That session is probably already in a bad state.
Dean, correct me if I’m wrong, but doesn’t selecting a specific model use API credits while Auto does not? I still have Auto credit available on my plan but I’m out of API credits.
@MTARK - got it, that’s helpful. Can you share the Request ID from one of those Opus 4.6 sessions? (Chat menu in the top right > Copy Request ID.) That’ll help us investigate further.
@bunst - you’re right. Auto and specific models use different usage pools. Auto has its own separate quota, and picking a model like Opus 4.6 will use your API credits. So this workaround won’t help if your API credits are depleted. Sorry for the confusing suggestion.
Since this happens with both Auto and specific models, it looks like a broader issue, not just Auto routing. I’ve noted your Request ID along with the extra context from this thread.
For now, if the issue keeps happening, try starting a completely new chat session and don’t continue the broken one. Let me know if that changes anything.
If you can share a Request ID of a case like that (and a screenshot of the issue) that would be great! Having Opus affected is unexpected based on our investigation so far.
getting the exact same issue with auto – starting new chats help in the short term but it sometimes goes back stuck on thinking and “begin_of_sentence” after a couple exchanges
