Subshell Docs

Use a reusable prompt

Give an agent instructions from your prompt library.

Use at launch

  1. In the server dashboard sidebar, open Subshells → New subshell.
  2. Choose the agent, optional preset, machine, and working directory.
  3. Under Add prompts, select Add prompt.
  4. Search your library in Yours, or choose Shared for other users' shared entries. Select a prompt or stack.
  5. Add more entries if needed and use the up and down controls to order them.
  6. Review the task and launch.

The form combines the selected text in order, separated by blank lines. The complete task must fit the 20,000-character limit. Task text selected here is a copy; later library edits do not rewrite it.

For task defaults you use repeatedly with the same configuration, follow Set launch defaults.

Type into a running session

You need edit access to the subshell, and its process and execution machine must be available.

  1. In the dashboard sidebar, open Subshells.
  2. Open the session card or row's ⋯ menu and choose Inject prompt....
  3. Choose a saved prompt or stack, or select Write your own... to enter custom text.
  4. Review Send prompt to agent? and select Type into pane.
  5. Open the session's terminal, inspect the typed text, and press Enter when the agent is ready.

Type into pane does not submit the task. It sends text to the existing process without creating another subshell or changing the preset. Check the agent's state first; text sent during a turn or to another CLI screen may not become a new task.

When writing custom text, Save to my prompts can also save it to your private library. Provide a description if you enable that option.

Check delivery

For a new session, confirm that the agent received the initial task. If Subshell reports missed delivery, inspect the terminal before retrying to avoid duplicating input. With MCP, check promptDelivered when you requested a task.

For injection, confirm the text is present and submit it yourself. An agent response, not a successful typing request alone, tells you the task was processed.

Next steps

Follow Build prompt stacks to reuse a sequence of instructions.

Edit on GitHub

Last updated on

On this page