How to Run Commands from the Chat

Estimated time: 4 minutes
Difficulty: Beginner

Why you'd use this

Fixing a failing test or getting a build green means running things, reading the output and changing code in a loop. The assistant proposes shell commands, you approve them, and the output goes back to it so it can act on the result. Long-running commands move to the background so the conversation is not blocked, and every command is cut off by a timeout so nothing runs forever.

Before you start

Permissions required:

• Ability to run the project's tooling from a terminal on this machine (the same shell and PATH VS Code uses).

You'll need:

• A project with a test or build command, for example npm test or pytest.

Steps

1. Ask for something that needs execution, for example "Run the tests and fix any failures."
2. A card appears with the proposed command and a Run button. Read the command, then click Run, or Reject to refuse it. A rejected command is reported to the assistant as not executed.
3. A live output box shows the command's output as it streams, with a running timer.
4. When the command finishes, its exit code and output are sent back to the assistant, which proposes a fix if something failed and then offers to run the command again.
5. To stop approving each command on a trusted project, open the / menu and tick Auto-run commands. Commands then run as soon as they are proposed, except ones classed as dangerous, which still ask.
6. To interrupt at any point, click Stop next to the composer. Every command the assistant started, including background ones, is killed.

What you should see

A command that runs longer than 20 seconds shows a "Backgrounded" line; the assistant continues with the output so far and receives the final result when it lands. If the assistant runs more than 25 commands in one task, it pauses with a Continue button so a loop cannot run unattended. Commands are generated for your operating system and shell, including remote hosts when you use Remote-SSH, WSL or a dev container.

Troubleshooting

• "command not found" for a tool that works in your terminal: the tool is on a PATH your shell sets up in its profile. The extension augments PATH from common locations, but a tool installed somewhere unusual needs a symlink or an absolute path.
• A command timed out: raise ellmCode.commandTimeout (seconds) in settings. Timeouts are treated as unresolved, never as passes.
• A result arrives after you pressed Stop: it is shown as "Result held" and included with your next message rather than restarting the run.
• The assistant claims tests pass without running them: it is nudged to run them. If you see a "Nudge" line about an unverified claim, that is this check working.

Related guides

• How to Stop, Pause and Continue a Run
• How to Understand the Verification Gates
• How to Tune Generation Settings for Your Model