The LLM could use a different session to respond to each command.

That would keep its behavior passive and restricted to simply the current state, since it would retain no memory of previous actions.