What broke
The human's words: "when the agent works on the app's geometry, does it have to manipulate the app's windows in the foreground? It floods my screen and often I can't do anything else." And a mission that was foreground for a reason no window rule would have found — it drove the physical device through a capture harness and needed the operator's undivided attention.
The rule it paid for
The Screen demand field, three values and a duration, classified by what the mission takes from the human and not by whether the toolkit draws a window.