Terminal

The Terminal tab of a workspace cockpit (/dashboard/workspaces/<id>/terminal) opens an interactive bash shell inside that workspace's executor. It is a real terminal: full-screen programs, colours, job control, and window resizing behave as they would in a local terminal.

The shell runs in the same sandbox as agent commands, with the same user, limits, network profile, and environment. It does not run on the host or in the harness's own services, and it gives you no more access than an agent working in that workspace has.

Opening and re-attaching

Visiting the tab never starts a shell by itself. Choose New session to open one in the workspace root. If the workspace already has running terminals, the tab lists them with Re-attach; re-attaching replays the recent output it still holds. When earlier output has already been dropped, the terminal says how many bytes were skipped instead of showing an incomplete screen as if it were whole.

A workspace can hold only a limited number of live terminals at once. Opening one more is refused with a message naming the limit; close a terminal you no longer need first.

Keyboard

Keys typed into the terminal go to the shell, including Tab and Escape. Press Ctrl+Shift+F6 to move focus out of the terminal to its toolbar, then use Tab and Shift+Tab to move through the page as usual.

The terminal follows the dashboard theme, including a switch between light, dark, and system while it is open.

Closing and limits

Close ends the shell and every process it started. The workspace and its files are unchanged. A terminal also ends when it has been idle too long, when it reaches its maximum lifetime, when the shell exits, or when the workspace stops. Its status pill says which, and New session starts a fresh one.

While a terminal is running in an open tab, the workspace counts as active, so it is not expired for idleness. It still ends at the workspace's maximum lifetime; the idle and lifetime expiry the dashboard shows are never later than the workspace's own. If the workspace is closed or expires, the terminal shows No longer available and stops polling.

Terminals are not resumed. If the harness restarts, every open terminal ends and shows No longer available; start a new session to continue. Use a background task (for example tasks_run from an AI tool) for long-running work that must survive a restart.

All terminals

/dashboard/terminals lists your terminals across workspaces with their size, age, and idle expiry. Ended terminals stay listed briefly with their final status. Each row links to the workspace's Terminal tab, and Close ends a running terminal after a confirmation.

Privacy and audit

Secret values that the harness injected into the workspace are redacted from terminal output before it is kept or shown, the same as for agent command output. Redaction matches exact values only, so avoid printing secrets in an encoded or partial form. If a typed character only appears after the next key (for example g when a ghp_… token is injected), that is redaction holding back text that could begin a secret until the following bytes show it is not one.

The audit log records only that a terminal was opened and how it ended. What you type and what the shell prints are never written to the audit log.

For operators

The live-terminal limit, idle timeout, maximum lifetime, and retained output size are the TERMINAL_* runner settings in the environment-variable reference. Terminals are a dashboard feature only; there is no MCP tool for them. For errors, see Troubleshooting.