ROOK separates work from judgment. Work is the coding CLI planning, changing and checking a project. A decision is a real fork where the result depends on the operator's intent. The work loop carries the first; the decision loop carries the second.
The operator is the person responsible for the project. ROOK does not remove that person's judgment. It moves the moment that judgment is needed out of the terminal and into a decision card that can be answered from another surface.
[Visual 1.1 — The crossing point between the two loops enlarged. On the left, a wide lane labelled "work" carries plan, change and verify blocks toward a narrow junction. At the junction the coding CLI holds up a card labelled "a real fork". One path continues as runnable work; the other crosses to the operator as "decision" and returns as "recorded answer". The driver appears below the lanes as a small transport rail, never inside the decision diamond.]
What stays in the work loop
The coding CLI authors the work and chooses what is runnable from documented project state. A non-thinking driver starts fresh sessions and relays their output. When one session finishes with recorded writeback, the loop can advance to the next instructed step.
That division matters. The driver does not interpret a report, rank tasks or invent a next move. It carries instructions written by the component that has the project context.
What crosses to the operator
A choice crosses only when the coding CLI needs intent it should not supply for itself. It writes the situation as a short card, gives concrete options and their consequences, and recommends one. The operator answers that bounded question; the recorded answer later returns to the work that raised it.
If the card blocks one thread, that thread can park while other runnable work proceeds. This does not mean every project always has spare work. When nothing else is runnable, the machine can still wait.
The limit of the separation
The two-loop model describes responsibility and movement, not a guaranteed productivity result. ROOK cannot make an unclear choice clear, ensure that another task is runnable or turn a mechanical driver into a substitute author. Its boundary is smaller: work continues where the coding CLI can act, and judgment returns to the operator where it cannot.
Next
Follow the autonomous side of the boundary through the work loop.