Skip to main content
Monitor is a tool, not a command: the model reaches for it when you ask it to watch something — a dev server, a build, a log file, a flaky test loop. The command runs in the background and each output line arrives as a notification while the model keeps working, so “start the server and tell me when it errors” is one instruction, not a babysitting job.
A watch ends when the process exits, on timeout (5 minutes by default, up to an hour), when the agent cancels it, or when the output floods (more than 200 lines in 10 seconds). For a log tail or anything that should keep going, ask for a watch that lasts the whole session: it has no timeout and still ends when the process exits, floods, or the session closes. Ask the agent what it is watching and it can list, inspect or cancel running watches. Inspecting a watch also shows how many lines it has sent, so a missing notification is visible rather than silent. Monitor renders in the transcript like any built-in tool, and sub-agents can use it too.