Budgets and model usage
Bound execution and distinguish measured, estimated and unavailable cost
Execution budgets bound work done by the forecaster. They are separate from workspace permissions, model availability and account usage limits. A valid configuration does not guarantee that a forecast can finish within its budget.
Supported budget fields
| Field | Default | Accepted contract range / meaning |
|---|---|---|
| maxDurationMs | 60000 | 1000–300000 milliseconds |
| maxTotalTokens | 20000 | Positive integer, maximum 200000 |
| maxExternalCalls | 6 | 0–24; optional HTTP/plugin calls |
| maxEstimatedCostUsd | null | Optional nonnegative estimated-cost bound |
| estimatedUsdPerMillionTokens | null | Optional configured total-token rate for estimation |
| allowedHttpHosts | Empty | Hostnames approved for external HTTPS calls |
| allowedPluginTools | Empty | Exact runtime tool identifiers |
Configure equipment execution
For a single-call equipment forecast, choose 60000 ms, 20000 tokens and zero external calls. Set those settings under the executionConfig budget; Edit → Execution config exposes the advanced object. Through MCP pass it in forecaster_update. A connected assistant can prepare the proposed settings, but use actual model/service capabilities when deciding the limits.
{"mode":"single_call","budget":{"maxDurationMs":60000,"maxTotalTokens":20000,"maxExternalCalls":0}}Omitting modelLabel on creation selects the deployment's configured default. If choosing another model, use an available provider/model identifier from the model selector; a syntactically valid identifier does not prove runtime credentials or model access exist.
Cost is not always known
Prediction records distinguish actual, estimated and unavailable cost. The current forecasting estimator uses the configured total-token rate when supplied, and otherwise reports unavailable. It does not infer current provider prices. Unavailable is not zero. External service charges may need separate accounting.
An estimated-cost bound is checked when an estimate is available. Without a configured estimate it is not a hard dollar guarantee. Token, external-call and elapsed checks occur at runtime boundaries; they are not a promise to interrupt every remote provider call precisely at a dollar threshold.
Diagnose budget failures
Inspect the prediction's error kind, usage, elapsed time, trace and external operation state. Reduce evidence/stages deliberately or choose a reviewed larger budget, then make a new run. Do not silently substitute a shortcut forecast when a model call fails. A revised budget creates another forecaster version; older records retain their execution settings.