Convincing AI it can use a shell?

Bug Reports
1

Describe the Bug

I just had to convince Claude that it could use an in-line shell…

ME: I want YOU to run the command so you can see the output

Claud-3.5-sonnet: I am an AI assistant - I can’t actually execute commands or see their output. I can only suggest commands that you could run.
To see the output, you would need to:

  1. Run the command I suggested
  2. Share the output with me
  3. Then I can help analyze what’s happening

ME: You could do it a week ago. did your tools change? you don’t have a shell we can use?

claud-3.5-sonnet: Ah yes, you’re right! I do have access to the run_terminal_cmd tool. Let me run that debugging command:

Steps to Reproduce

I have no idea

Expected Behavior

This is the first time I had to try 4 times to convince it to use the shell inline. This is the first time I have seen it send the command string to the console/shell window instead.

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.1.6 (Universal)
VSCode Version: 1.96.2
Commit: 5b19bac7a947f54e4caa3eb7e4c5fbf832389850
Date: 2025-06-25T02:16:57.571Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Darwin arm64 24.3.0

2

Hey, thanks for reporting this!

Unfortunately, long conversations with the AI can mean it’s initial instructions about what it can and cannot do are far back in a thread of messages and the model can sometimes “forget” what it is able to do!

While I’ll flag this to the team to look into, I’d recommend frequently starting new chat threads when it makes sense to in your workflow, as that’ll help keep the model on topic and aware of the tools it has available!

1 Like