Should budget checks happen only at safe halt boundaries?

[Keesan12]

A naive budget cap can interrupt an agent mid-edit and leave the repo inconsistent.

One design direction is to only check budget or policy at clean state transitions:

Task → Plan → Attempt → Verify → Record → Halt or Continue

Curious how others think about this.

Should budget caps interrupt immediately, or only at safe halt boundaries?