Decisions

Why Link Desktop gives an agent a virtual desktop of its own.

Why a virtual display rather than the owner's screen?

An Xvfb display works on a box with no screen and touches nothing of the owner's, so an agent can use it unattended. The tool never reaches the owner's own screen, and no setting turns that on: a real screen would need the owner's permission for every action, and the desktop portal and capture interfaces it would use cannot be exercised on a cloud machine.

Why is there no window manager?

A window manager draws frames and title bars and moves, raises, and closes windows. The desktop runs none, so it stays one Xvfb with nothing more to confine or keep running. Programs' windows have no frames and no key closes one, so the tool says so rather than suggest alt+F4. A minimal window manager can come later if a program needs one.

The display starts with an MIT cookie, so no other user of the machine draws on it. The cookie is renewed at each start, and sent only to an X server run by your user, since another user's server started with -ac takes any client and would be handed it, and a root one, such as the computer's own X server, would see the agent's screen. Each program the agent's run tool starts gets that display's DISPLAY and XAUTHORITY, and its sandbox reads that display's socket and the cookie, never the whole of /tmp/.X11-unix.

Link selects a number with no X server socket or lock file when the desktop is added, reserves it in your user's runtime folder, and keeps it for later starts, so none of your other agents' desktops takes it. Another user's X server can still take the number meanwhile: the tool asks the kernel who serves the socket before sending the cookie, and refuses a server that isn't your user's. Troubleshooting explains how to select another number.