[BUG] Windows: Terminal commands fail due to macOS sandbox-exec after recent Antigravity update

Google Antigravity
1

Summary

After the latest Antigravity update, terminal commands executed by the agent fail on Windows because the system incorrectly injects the macOS-only command sandbox-exec. Since this utility does not exist on Windows, every automated terminal action fails immediately.

This started happening after updating the application and resetting the app data while trying to fix a model availability issue.

Environment

  • OS: Windows 11

  • App: Antigravity (latest update)

  • Model: Gemini models (3.1 ), GPT-OSS, Claude Opus … every model

  • Terminal: PowerShell / CMD

Issue 1 — Model availability problem

After updating Antigravity:

  • The model selector still shows Gemini 3.1

  • But when sending a prompt the app returns:

Gemini 3 is no longer available

Reinstalling the app did not fix this.

What I tried

To resolve the model issue I cleared the Antigravity AppData directory.

After that:

  • The app launched like a fresh install

  • The model issue partially reset

  • But a new bug started appearing with terminal execution.

Issue 2 — Windows commands wrapped with macOS sandbox

Whenever the agent tries to execute a terminal command, it generates something like:

sandbox-exec -f C:\Users\...\antigravity-sandbox.sb /bin/bash -c 'python --version'

PowerShell returns:

sandbox-exec : The term 'sandbox-exec' is not recognized as the name of a cmdlet

Steps to Reproduce

  1. Use Antigravity on Windows.

  2. Ask the agent to run any command (example: check Python version).

  3. Agent attempts to execute command via sandbox wrapper.

Example request:

python --version

Generated command:

sandbox-exec -f ... /bin/bash -c 'python --version'

Result: command fails.

Expected Behavior

On Windows the agent should execute commands directly, for example:

python --version

or

cmd /c python --version

without macOS sandbox wrappers.

Technical Impact

Severity: High

This breaks core functionality including:

  • Automated terminal tasks

  • project setup commands

  • debugging workflows

The agent repeatedly retries the same failing command which also wastes tokens and time.

Possible Cause

This may be related to two things happening together:

  1. Model migration issue (Gemini 3 → newer versions).

  2. Clearing AppData causing Antigravity to initialize with incorrect OS defaults.

It appears the Windows client may be incorrectly initializing a macOS terminal environment.

Temporary Workaround

For anyone encountering this issue:

  1. Open Settings

  2. Disable shell integration features.

  3. Restart the application.

    After doing this the agent stop wrapping commands with sandbox-exec.

    Additional Context

    This bug started only after the recent update and after trying to resolve the “Gemini 3 is no longer available” error. Before that, terminal commands worked normally.

2 Likes
2

Same issue. Connecting from macos to linux XFCE.

zsh: command not found: sandbox-exec
global_ai> sandbox-exec -f /tmp/antigravity-sandbox.sb /bin/bash -c ‘ls -R
/home/xyz// | head -n 2
0’
zsh: command not found: sandbox-exec

3

Same issue here on my RHEL 10 host

```
(base) ➜ terraform git:(branch-name) ✗ sandbox-exec -f /tmp/antigravity-sandbox.sb /bin/bash -c ‘terraform fmt -recursive’
zsh: sandbox-exec: command not found…
```