Auto-review, sandbox and the Docker

Support Help
,
1

I recently received email about auto-review mode (“Recommendation: Auto-review mode”) and decided to give Auto-review a try. Unfortunately, I quickly ran into what seems to be its biggest limitation.

The main issue is that the agent runs in a sandbox environment where it has no access to Docker. As a result, even though my project rules and AGENTS.md explicitly instruct the agent to communicate with containers using Docker commands, it consistently falls back to running tools directly on the host machine because it reports that Docker and the containers are unavailable from within the sandbox.

This leads to a cascade of problems. Many of the required tools either don’t exist on the host system or are installed in different versions than those available inside the containers. The agent then starts reporting errors and, even worse, attempts to modify dependencies to match the versions it discovered on the host instead of the versions that are actually used by the project inside Docker.

I don’t use the Docker extension for Cursor (or VS Code). I had various issues with it in the past, and I also work across a large number of projects, including some that use Vagrant, so it’s not a practical workflow for me. My primary IDE handles native Docker integration very well, and Cursor agents work correctly as well - as long as they’re not running in sandbox mode.

Given these limitations, is there any way to use Auto-review effectively with Docker- or Vagrant-based projects? For example, is it possible to either:

  • use Auto-review without sandbox mode, or
  • allow sandboxed agents to access Docker (and Vagrant) so they can interact with the actual development environment?
4

Hey, thanks for the detailed write-up. The issue is how the sandbox works. It intentionally isolates the agent terminal with access only to the workspace plus network filtering, so it can’t see the Docker socket or your containers. Because of that, the agent thinks Docker isn’t available and falls back to host tools, and then you get a mess with versions and dependencies. There isn’t a separate toggle to give the sandbox access to Docker or Vagrant right now. That isolation is the whole point of the mode.

A working option is to run commands outside the sandbox so the agent can see your real Docker or Vagrant environment:

  • In Cursor Settings > Agents, switch Auto-Run Mode from Auto-Run in Sandbox to allowlist mode. Commands in the allowlist will run automatically, everything else will ask for approval, and all commands run on the host with Docker access.
  • Or enable Legacy Terminal Tool in Cursor Settings > Agents > Inline Editing & Terminal. The terminal runs on the host and respects the allowlist.

Tradeoff: outside the sandbox you lose auto-running everything, so you’ll need to maintain the allowlist and approve some commands. But since the agent was changing dependencies based on host versions, approvals are probably a plus. It’s less likely to break something.

Try it and let me know how it goes with your Vagrant projects.