Claude confused with paths, and Gemini refuses to serve requests

Bug Reports
1

Describe the Bug

Claude-4-sonnet has been confused (?) with one of my projects and it keeps searching for zebrd instead of zebrad, when doing file searches or tries to apply changes to a recent file it was able to find. The model realizes it is doing it the wrong way, and it tries to to “fix itself” but it keeps calling the tool with the wrong zebrd path.

No matter how many times I try to indicate it how zebrad it’s spelled, it might do it correctly once and then continues with the misspelled zebrd word. This seems to be an issue with the tool being used for search.

Another issue is with all Gemini models which refuse to serve requests when asking just to do simple things, like replacing text:

Request ID: 2a62c203-a1cf-4e8e-b0ed-0d8918dcd419
{"error":"ERROR_UNAUTHORIZED","details":{"title":"Unauthorized request.","detail":"The provider refused to serve this request based on the content","isRetryable":false,"additionalInfo":{},"buttons":[],"planChoices":[]},"isExpected":true}
ConnectError: [unauthenticated] Error
    at mel.$endAiConnectTransportReportError (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:4918:251962)
    at Zlr._doInvokeHandler (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:493:20860)
    at Zlr._invokeHandler (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:493:20602)
    at Zlr._receiveRequest (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:493:19367)
    at Zlr._receiveOneMessage (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:493:18189)
    at TSt.value (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:493:16283)
    at ve._deliver (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:49:2962)
    at http://ve.fire (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:49:3283)
    at http://Mst.fire (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:4905:12154)
    at d.onmessage (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:7030:18167)

Both of these issues have been happening with so much recurrence recently that I’ve decided to directly use Claude Code instead (although I prefer Cursor’s UX)

Considering these are the 2 models I use the most, this makes Cursor unusable for me.

Steps to Reproduce

You can try cloning GitHub - ZcashFoundation/zebra: Zcash - Financial Privacy in Rust 🦓 and reproducing Request ID: c4ba6dae-a62f-4fa7-b699-bf4b30ea37c3

Or indicating to make a lot of search and replace of text, as it commonly gets triggered after a while.

Expected Behavior

No errors.

Screenshots / Screen Recordings

Claude 4 sonnet confused.jpg
Claude 4 sonnet confused.jpg1262×690 76.6 KB

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.4.2
VSCode Version: 1.99.3
Commit: 07aa3b4519da4feab4761c58da3eeedd253a1670

Does this stop you from using Cursor

Yes - Cursor is unusable

2

Hi, are you still facing this issue?

We’ve seen some issues with LLMs doing things not as expected. This is more of a model issue than a Cursor issue.
Potential workarounds:

  • Try starting a completely new chat session when this happens

  • Explicitly include the correct spelling “zebrad” in your initial context or instructions

3

I still have this issue, even when applying those potential workarounds.

When using the models outside Cursor, this does not happen.

4

This is still an issue in the latest version. I can’t practically use any Anthropic model nor Gemini in this repo.

5

Thanks for reporting this. Two separate issues here that need different approaches.

First, please update to v1.5.5 (latest) and retest both issues.

For the Claude path confusion: Can you share a sample chat showing the misspelling? Does this always happen, or just in certain contexts? Try starting a fresh chat that mentions “zebrad” from the beginning to see if it still gets confused.

Consider adding project-specific context using Cursor. Create a .cursor/rules file explaining your project structure and the correct “zebrad” spelling. This gives Claude persistent context about your codebase.

For the Gemini authorization errors: This is Gemini’s content filtering being overly strict. Try rephrasing your requests differently, or use Claude models for those specific text replacement tasks. More info on this issue here

Please let us know if these issues persist after updating and trying these approaches.

6

First, please update to v1.5.5 (latest) and retest both issues.

In my last reply, this was in version v1.5.5 (latest).

For the Claude path confusion: Can you share a sample chat showing the misspelling?

I shared an image in the first chat, it’s the same behavior. Or do you need me to export the chat?

Does this always happen, or just in certain contexts? Try starting a fresh chat that mentions “zebrad” from the beginning to see if it still gets confused.

A fresh chat works for a while, then it goes back and starts using zebrd. This can happen either in the first iteration or after 2 or 3 more. It doesn’t matter how much I try to make it use zebrad, it might use it once in a response, and then continue with the wrong zebrd

I’ve done this, and still it doesn’t work once it starts using the wrong word.

7

Thanks for the shared details.

Can you check if “zebrd” is getting into your context somewhere? Is that mentioned anywhere in any files or context you pass on? (I did a search on the repo you mentioned and couldn’t find anything.)

Would be great if you can share an exported chat and a request ID.

8

This topic was automatically closed 22 days after the last reply. New replies are no longer allowed.