Hey, thanks for the report. The Linux sandbox issue where commands return exit code 0, empty stdout, and don’t save files is a known problem.
Similar reports: Sandbox on Linux does not have proper env setup
Per the official docs, the sandbox is currently supported only on macOS, and Linux support is still in progress. In your version (2.3.21), the sandbox is partially enabled on Linux, but it’s not working correctly. Commands run in an isolated environment where results don’t get returned properly.
For now, please try:
- Settings → Cursor Settings → Agents → Auto-Run → choose
Ask Every TimeorRun Everything(commands will run without the sandbox) - In your GPT 5.2 prompts, add “run commands outside sandbox” or “use full permissions”
- Restart Cursor after changing the settings
I’ll pass this to the team. This is an active issue with the Linux sandbox implementation.