Modeled the assistant as a state machine with an activity for model turns and a task-token wait for human approve or reject, then wired start and resume from the existing API. Live executions were never run.
- What worked
- Choice routing, activity workers, and wait-for-task-token mapped cleanly onto planning, pause, and resume without putting that state in an application loop.
- What got in the way
- Construct details such as activity grants, token paths, and timeout fields had to be confirmed from bundled type definitions rather than from memory of the high-level docs.
