Agent switches to command line tools in Investigate Mode (repost)

[RafeSacks]

Describe the Bug

When in Investigate mode, the agent (GPT-5 for example) starts trying to use sed or other tools, instead of just reading files with the native tools

Steps to Reproduce

1. switch to investigate
2. ask the model to figure something out where it has to read files

Expected Behavior

It should use the Cursor read tool, not the shell tool.

Screenshots / Screen Recordings

(strange it tried to read the wrong file here:)

Note the comment about the read_file tool not being part of the function definition. Important?:

GPT-5 doesn’t mention the read_file tool:

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 2.0.64 (Universal)
VSCode Version: 1.99.3
Commit: 25412918da7e74b2686b25d62da1f01cfcd27680
Date: 2025-11-06T04:35:14.424Z
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Darwin arm64 24.6.0

For AI issues: which model did you use?

GPT-5, though I suspect others would have the same issue. I saw this more than 4 times in different chates before reporting

Additional Information

If I switch back and forth it seems to start using the tool eventually. Perhaps some memory leak between chats gave it the clue? Regardless, I have seen this a number of times over the last few days. I have never seen it happen in “agent” mode.

Perhaps there is a clue in the way Composer reports its tool lists as the read tool being “not listed as explicit tools in the function definition”.

Does this stop you from using Cursor

Yes - Cursor is unusable

[deanrie]

Hey, thanks for the report. The screenshot showing that read_file is missing from the function definition is very helpful.

Could you check your mode settings and share a screenshot of which tools are enabled? I want to make sure the right tools are configured for that mode.

Also, next time it happens, please grab the Request ID. This will help the team understand why GPT-5 is getting incorrect tool definitions.

It’s interesting that switching modes temporarily helps - it suggests the tool configuration is sometimes getting out of sync. I’ll pass this to the team once we have those details.

[RafeSacks]

Is “Investigate” Mode something I made? I honestly can’t remember.

While I’m auditing my modes, do you have any idea what the lightening and thinking icon modes are? I would never make a mode with no name, so I am as close to sure as I can be that I did not make them. Any others that are not default here?:

I am not sure that switching temporarily helps. I was just saying that it only seems to happen in this mode. If this mode is one I made, then it was made pretty close to the feature release date, I wonder if it has some stale or deprecated aspect to it?

The point of this mode is to give the agent all tools except edit. Does “Ask” do this? I can’t inspect the details of the “Ask” mode.