Agent is stuck in Sumarizing chat context - Planning next moves, then An unexpected error occurred on our servers. Please try again, or contact support if the issue persists.
Steps to Reproduce
Running any prompt with Composer 2
Expected Behavior
any actual response
Operating System
Windows 10/11
Version Information
3.3.30
VSCode Version 1.105.1
For AI issues: which model did you use?
Composer 2
For AI issues: add Request ID with privacy disabled
Request ID: 3d4f5d75-8947-47a7-bbac-200d7a5bb86a
[internal] Failed to run step, exceeded max retries
EEe: [internal] Failed to run step, exceeded max retries
at cVy (vscode-file://vscode-app/d:/Programas/Cursor/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:29387:25583)
at aVy (vscode-file://vscode-app/d:/Programas/Cursor/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:29387:24382)
at gVy (vscode-file://vscode-app/d:/Programas/Cursor/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:29388:6490)
at d9d.run (vscode-file://vscode-app/d:/Programas/Cursor/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:29388:11311)
at async zfs.resume (vscode-file://vscode-app/d:/Programas/Cursor/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:41721:18428)
at async h1h.streamFromAgentBackend (vscode-file://vscode-app/d:/Programas/Cursor/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:41789:12518)
at async h1h.getAgentStreamResponse (vscode-file://vscode-app/d:/Programas/Cursor/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:41789:19638)
at async ryt.submitChatMaybeAbortCurrent (vscode-file://vscode-app/d:/Programas/Cursor/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:29503:16976)
Hey, this is a known bug. When the chat context grows to the model’s token limit, the blocking summarization can’t compress it in a single pass, the retry loop runs out, and you get Failed to run step, exceeded max retries. We’re tracking the issue, but I can’t share an ETA for a fix yet.
For now, the workaround is to start a new chat. This usually shows up in long multi-turn sessions with Composer 2 where the context is already very large. Starting a new conversation, you can paste a short summary of what you were doing, avoids the issue until it’s fixed.
If the error still happens in a fresh chat after a couple of simple messages, let me know and I’ll ask for another Request ID to check.
Thanks for the update. If it crashes even in a new chat with a simple request, that’s not quite the scenario we expected. We’ll need to look at a fresh Request ID.
Could you please:
Open a new chat in Composer 2, send something simple like “Hey”, and wait for the error.
Before you do that, make sure Privacy Mode is turned off in Cursor Settings > Privacy, otherwise the Request ID won’t show much.
Copy the new Request ID from the chat menu top right > Copy Request ID and paste it here.
In parallel, can you check if another model works in the same workspace, like Claude 4.6 Sonnet or GPT-5.5? If other models work fine and only Composer 2 crashes, that’s an important signal.
It did, then I updated and it worked again for a couple of days. Today it started with the same behavior, where any prompt (even in new chat) gets stuck with summarizing context and planning next steps until the error message shows up.
Thanks. The screenshot where Chat context summarized shows up right after “hey” in a new chat is an interesting signal. Usually this bug only happens after a long session. That suggests something is blowing up the context even before your first message.
A couple things to check:
Take a look at Cursor Settings > Rules. Do you have big user rules there, or lots of accumulated rules and skills? They get mixed into every request and can overflow the Composer 2 window.
Try opening a completely different workspace any small folder and send “hey” in Composer 2 there. If it works in a clean workspace, the issue is likely in your main project like .cursorrules, a large AGENTS.md, or auto-attached skills.
While we dig in, other models Claude 4.6 Sonnet, GPT-5.5 do not block, so you can use them as a workaround.
We are already tracking an issue for this behavior. I’ll add your new Request ID to it, it will help the team reproduce this specific variation. I can’t share an ETA for a fix yet.
yes, for this project in particular I have a lot of rules and skills. I have a simpler project where I am not finding the issue. I tried turning skills off and changed the rules to be very simple and still get the same error
Useful signal: since the clean project works, and in this one the error still happens even after simplifying rules and turning off skills, it means the context isn’t being inflated by those. Something else in the project is getting mixed in before the first message and it’s overflowing the Composer 2 limit.
Let’s narrow down what it is. Try these one by one:
MCP servers. If this project has MCP servers connected, temporarily disable all of them in Cursor Settings > MCP, then send “hey” in a new Composer 2 chat. A lot of MCP tools can bloat the context even before your first message.
Check if there’s a large AGENTS.md or .cursorrules in the project root. Those get added to every request in addition to what you see in the Rules UI.
If the chat still has pinned context or large @-attached files or folders, remove them.
Also, please send a new Request ID from this problematic project after you’ve simplified rules and turned off skills. Your Privacy is set to Share Data, so everything will be included. That’ll help the team see your exact case.
While they’re fixing it, you can use Composer 2 in this project with Claude 4.6 Sonnet or GPT-5.5. Those aren’t blocked. I can’t share an ETA for the fix yet.
I deleted my .cursor folder with .skills and .agents and it seems to work now. I will continue to test for a couple of days and let you know if the issue comes back. Thanks!
