A practitioner-grade thought experiment, not a shipping product. The author wants an infinite horizontal canvas as the desktop, with each physical display as a viewport onto it — scrollable, session-persisted, task-organized, essentially "tmux for normal people." Docking and undocking become deterministic state replay: window order, scroll position, and task membership never change; only column widths reflow. Focal vs. glanceable monitors are inferred from which display receives ~90% of keystrokes, no eye tracking or config. Browser tabs detach into first-class windows belonging to tasks instead of the browser. A local LLM proposes layouts via a constrained MCP verb list, shows a preview, and deterministic code executes only what you confirm — the model never touches the window manager. Privacy is a per-window flag gated by machine-observable facts (connected displays, capture sessions), never inferred attention. The genuinely useful part is the honest failure section: the privacy feature hides content from its owner and shows it to the threat, because reading is an idle state and a pedestrian is undetectable. And the infinite append-only canvas is a garbage collector nobody built — he solved the messy desk by buying it an infinite desk. That's the part worth stealing for your own design review. Also of note: the cited literature alone is worth the click — Henderson and Card's 1986 Rooms paper, WindowScape, and the 2004 paper legitimately titled "Stuff goes into the computer and doesn't come out." Windows Sets (2018–19, canceled) remains the closest thing anyone ever shipped to this, and it died of strategy, not merit.