# Auto (preview) approval mode is highly ineffective **URL:** \ **Category:** Google Antigravity\ **Tags:** issues\ **Created:** [October 13, 2026, 7:21pm UTC](https://discuss.ai.google.dev/t/auto-preview-approval-mode-is-highly-ineffective/188925 "2026-10-13T19:21:06Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![bandsaws](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/bandsaws/32/148974_2.png) [@bandsaws](https://discuss.ai.google.dev/u/bandsaws)\ **Post date:** [October 13, 2026, 7:21pm UTC](https://discuss.ai.google.dev/t/auto-preview-approval-mode-is-highly-ineffective/188925/1 "2026-10-13T19:21:06Z")
I’m trying Antigravity for the first time as it appears to be the only way to use Argon for software development via an Ultra subscription. (I’d prefer to use another harness, but that doesn’t seem to be an option.) I tried to enable Auto Mode in Settings for the app, but there is no Auto mode entry there at all. There _is_ an Auto (preview) toggle in the actual conversation view, but it seems to barely work. I created a new session and added the prompt “Pull the latest code from the remote”, but auto approval failed - I got a permissions prompt saying “Allow pulling latest code from remote? the auto-run safety check did not reach a decision”. Followups kept prompting me for additional approvals on clearly innocuous work, like running ‘npm ci’ to install dependencies to complete a task I had specifically requested. Is this expected behavior? Why are the Auto settings misaligned between app-level settings and conversation-level settings? And why does auto mode not seem to actually be able to automatically approve basic requests? ---
**Author:** ![Rebbanpalli\_Naveen](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/rebbanpalli_naveen/32/120945_2.png) [@Rebbanpalli\_Naveen](https://discuss.ai.google.dev/u/Rebbanpalli_Naveen)\ **Post date:** [October 15, 2026, 1:50pm UTC](https://discuss.ai.google.dev/t/auto-preview-approval-mode-is-highly-ineffective/188925/2 "2026-10-15T13:50:45Z")
Hi @bandsaws, Thank you for reporting this. Here is how permissions, the terminal sandbox, and **Auto** mode work across Antigravity, along with how you can configure commands like `git` and `npm` to run without prompting: ### 1. Why Commands Like `git pull` and `npm ci` Prompt for Approval - On **macOS and Linux** , Antigravity runs terminal commands inside an isolated **[Terminal sandbox](https://antigravity.google/docs/sandbox)** by default, which allows access to your project folders and temporary directories but blocks outbound network access. - When a command requires network connectivity (such as pulling from a remote Git repository or fetching packages with `npm`), it cannot run inside the default sandbox unless network access to that domain is allowed, so the agent requests to run the command outside the sandbox. - **Auto** mode approves file edits and safe commands without asking while still prompting before irreversible actions, and commands that would leave the terminal sandbox are first reviewed by a separate agent (**[Antigravity CLI `v1.3.5` Release Notes](https://github.com/google-antigravity/antigravity-cli/releases/tag/1.3.5)**). When that review does not auto-approve an unsandboxed command, Antigravity falls back to prompting you for confirmation. ### 2. Preset Picker vs. Application Settings - **[Antigravity `v2.23.0`](https://antigravity.google/docs/changelog)** added the permissions preset picker to the conversation message box, while **Settings \> General \> Permission Settings** (and **Settings \> Projects** ) currently lists the **[Default, Request Review, and Turbo presets](https://antigravity.google/docs/permissions#permission-presets)**. - We have shared your feedback regarding the misalignment between the conversation-level `Auto (preview)` option and **Settings \> General \> Permission Settings** , as well as the auto-run safety check behavior on `git pull` and `npm ci`, with the engineering team. ### 3. How to Allow Trusted Commands Without Prompting Your configured **Allow** , **Deny** , and **Ask** rules are layered on top of your permission preset and **always take precedence** (**[Agent permissions](https://antigravity.google/docs/permissions#permission-presets)**): 1. **Add `command(...)` rules to your Allow list (covers execution both inside and outside the sandbox):** Under **Settings \> General \> Permission Settings** (or per project under **Settings \> Projects** ), you can add command prefixes or regular expressions to your **Allow list** : 2. **Allow outbound network domains inside the sandbox with `read_url(...)`:** Any domain granted under `read_url` in your **Allow list** is added directly to the terminal sandbox’s outbound network allowlist, permitting tools like `npm` or `curl` to connect to authorized hosts inside the sandbox (**[Terminal sandbox — Permissions integration](https://antigravity.google/docs/sandbox#permissions-integration)**): ### 4. Using Gemini 4 Argon in Other Environments If you prefer working in your existing terminal or code editor with your **Google AI Ultra** subscription, **Gemini 4 Argon** is supported across the **Antigravity app** , **[Antigravity CLI](https://antigravity.google/product/antigravity-cli)**, and **[Antigravity IDE extensions](https://antigravity.google/docs/ide/extensions)** (Visual Studio Code, Visual Studio, JetBrains IDEs, Zed, and Xcode) as listed on the **[Models](https://antigravity.google/docs/models)** page.