Used Gemini to help me track everything I tried to do to fix on my own. Here is a rundown. Please help. I’m almost at the give-up-on-cursor stage.
Final Bug Report for Cursor Support
Where does the bug appear (feature/product)?
[X] Cursor IDE
[ ] Cursor CLI
[ ] Background Agent
Describe the Bug Consistent [internal] Failed to run step, exceeded max retries and Internal Error when using Auto or Composer modes.
Critical Context: This occurs globally across all projects, including brand-new empty directories.
Individual models (e.g., GPT-4o, Claude 3.5 Sonnet) work perfectly, proving the account and internet connection are valid. The failure is specific to the “Auto” routing logic.
Steps to Reproduce
Open Cursor in any folder (even an empty one).
Open Chat (Ctrl+L) or Composer (Ctrl+I) and set the model to Auto.
Type a basic prompt like “test” or “hi.”
The Agent attempts to initialize and immediately throws the retry error without streaming any tokens.
Expected Behavior The “Auto” model should negotiate the usage pool and initiate a response stream.
Operating System
[X] Windows 10/11 (Version: Windows_NT x64 10.0.19045)
Version Information
IDE:(Please paste the text from: Menu → Help → About Cursor → Copy)
Model used: Auto / Composer
Request ID:ee4b5864-9437-4ca4-8646-1025881faad9
Additional Troubleshooting Already Performed I have already performed the following “Deep” troubleshooting steps with no success:
Network: Verified DNS resolves correctly to Cursor APIs; toggled HTTP/1.1 Compatibility Mode; confirmed no VPN/Proxy/Zscaler.
Account: Logged out/in (IDE and Web); verified subscription is active and not maxed out.
Cache Wipe: Deleted state.vscdb in %APPDATA%\Cursor\User\globalStorage and renamed the %APPDATA%\Cursor folder to force a total fresh start.
Privacy/Rules: Toggled Privacy Mode OFF and cleared all Global AI Rules.
Logs: Diagnostic logs show “Success” for all endpoints, but the Agent stream consistently fails with a MODULE_NOT_FOUND or max retries error in the background.
Does this stop you from using Cursor?
[X] Yes - Cursor is unusable (Core AI features are broken).
One Final Technical “Punch”
If you want to be extra thorough, you can add this line to the Additional Information section:
“My initial network diagnostics showed api2.cursor.sh resolving to 127.0.0.1. While I have corrected this locally and now get a ‘Success’ on diagnostics, the Agent stream still fails specifically for Auto/Composer. This suggests a server-side session stickiness or a corrupted handshake for my Account ID.”
For AI issues: which model did you use?
Model name (e.g., Sonnet 4, Tab…)
For AI issues: add Request ID with privacy disabled
Request ID: f9a7046a-279b-47e5-ab48-6e8dc12daba1
For Background Agent issues, also post the ID: bc-…
Additional Information
Add any other context about the problem here.
Does this stop you from using Cursor?
Yes - Cursor is unusable
Sometimes - I can sometimes use Cursor
No - Cursor works, but with this issue
The more details you provide, the easier it is for us to reproduce and fix the issue. Thanks!
Hey, thanks for the detailed report and for trying so many things already. That really helps narrow it down.
The key detail here is that individual models work, but Auto and Composer don’t. That’s because Auto mode sends more context to the server, like skills summaries and orchestration metadata, and that extra overhead pushes the request size over the 50 MB limit.
A few questions to help diagnose:
Do you have any Skills installed in Cursor? If yes, about how many? We’ve seen the exact same issue for other users when a large number of skills bloated every request. Even in an empty project, skills summaries still get sent.
What’s your current HTTP/2 setting? You said you tried both options, but which one are you using now? If Disable HTTP/2 is enabled, meaning it’s using HTTP/1, requests are encoded less efficiently and can end up taking about twice as much space.
Can you share your Cursor version? Go to Menu > About Cursor > Copy.
If you have Skills installed, try removing them or cutting the number down a lot, then test Auto mode again. That will very likely fix it.
I found the issue. My /skills folder is to large. I moved it out of .cursor and now it’s working. How can I use my skills tho. Can I just use a reference .md file and keep the skills folder out of the .cursor? Where do I save the skills folder too?
Glad you found the cause, sounds like that was exactly it. About your question:
You don’t need to remove skills from .cursor completely. You can add them back, but disable automatic detection for most of them. In each skill’s SKILL.md, add this to the frontmatter:
disable-model-invocation: true
This means the skill summary won’t be sent with every request, which is what was blowing up the payload. You’ll still be able to run the skill manually with /skillname in chat.
Suggestion: keep disable-model-invocation: false (or just omit it) only for the skills you use most often and want the model to pick up automatically. Set all the others to disable-model-invocation: true. That way Auto mode still works, and all skills stay available via slash commands.
