Approvals and recovery

Waiting is a workflow state, not an error. Use it for approvals, external callbacks, scheduled timing, and rate-limited systems. Record what the run is waiting for and who owns the next decision.

When a step fails, the run should preserve the completed steps and enough context to retry safely. Before enabling retries, determine whether the action is idempotent, how duplicate effects are detected, and what a human should do if recovery fails. A retry policy without that design merely repeats uncertainty.