Progress and outcomes
Human waits are specific
A question requests information. A Verdict requests a configured decision or form. A shared review assesses an exact subject version. Explorer plan acceptance starts Build work. These decisions have separate authority and records. Typing a message does not settle a pending Explorer plan. Run completion does not establish output approval.Retry and cancellation
A retry uses the run’s selected versions and records another attempt. Check which outputs already exist before repeating a step. A whole model-step retry can repeat tool effects. Inspect calls that collected data, wrote resources or contacted a destination. Reconcile uncertain external outcomes before starting a replacement action. Cancellation stops new work and requests child cancellation. It does not remove bytes already delivered externally or promise automatic compensation. Accepted evidence and incurred usage remain.Budgets and actual usage
The context meter describes the size of the latest model context. It is not the turn’s total billed token count.
The root-run budget governs its configured work and child work. Attempts and revisions use remaining allowance. Inspect the configured limit and actual usage rather than assuming a retry has a fresh budget.
An app’s model-spend cap is a soft stop between blocks. A running block can exceed it before the next check. It does not by itself cap every storage or compute charge.
A human wait with no allocated compute creates no runtime charge. Retained evidence continues to use storage. Failed work retains incurred usage under the billing policy.
Use Set budgets and billing ownership for resource attribution.