Nice job... the kanban + task hooks combine is a good one. Wonder how parallel activities function in practice though. How do you keep two or three agents running at the same time on distinct tasks from stomping on each other? Like, is each agent going to receive its own sandboxed terminal/session or is it going to rely on worktrees to keep things separated? Asking because I've wanted to run many instances of Claude side-by-side without one overwriting the other's edits.