Budgets and stopping states

Build and runtime phases within one command share model call count, cumulative tokens, and a deadline. Set budgets explicitly before resolution and retain observed usage. Locked replay consumes no model requests.

uv run nexa build examples/open_programs/sum.nx --environment examples/open_programs/aliyun.json --state .nexa/sum.db --max-model-calls 4 --max-tokens 16000 --timeout 90 --json

Control	Default and rule
--max-model-calls	4; hard cap; implementation and model review requests both count
--max-tokens	20000; cumulative provider-reported input/output tokens
--timeout	120 seconds; shared deadline; each OCI call is limited to 30 seconds
ResolutionBudget.max_output_tokens	SDK default 4096; each request is also capped by remaining token budget

Input usage becomes observable after the provider returns. Missing usage blocks admission with usage_unreported; excess cumulative usage returns model_token_budget. A failed request still consumes its initiated call. Calculate cost separately against the actual model and price baseline.

Handle a stop

model_call_budget, model_token_budget, and resolution_timeout identify the exhausted budget. Preserve existing candidates, evidence, and activity records before choosing the next operation. clarification_required needs an author answer followed by source synchronization. cancelled follows cooperative cancellation of an exact activity ID.

Cancellation is durable. An ongoing provider/OCI call checks it after returning or timing out, then stops subsequent steps. Check client process termination and the runtime's durable activity state separately; see recovery.

---

Source and validation · open_execution