Model is some how able to execute commands without approval?

[edelauna]

App Version

3.22.5

API Provider

xAI (Grok)

Model Used

grok-3-mini

Roo Code Task Links (Optional)

Grok is somehow able to execute a command without approval. The command is not showing up in my terminal, so I'm not sure where this command is getting generated. I thought maybe it was hallucinating, but it displays the results with accurate timestamps as if they were somehow generated on the command line.

My auto-approve is set to READ

Whereas this is what I was expecting (and it sometimes does this)

🔁 Steps to Reproduce

Just regular use, was having it write some tests for me.

💥 Outcome Summary

Expected to be asked for approval to run a command, but somehow the model was able to run the command some other fashion and read the results.

📄 Relevant Logs or Errors (Optional)

[daniel-lxs]

Hey @edelauna,

To me, it looks like the model is hallucinating and generating a command output without actually running it. If you want to double-check, you can export the task history like this:

Feel free to share it here and I can take a closer look, but I'm fairly confident no command is actually being executed.

[MuriloFP]

The current timestamp is always included into the system instructions, so it might be just hallucinating the output.

Closing this one for now.

If it's still an issue and if you get more info, please feel free to re-open it!