Skip to main content
A run can contain multiple model calls and tool roundtrips. Inspect the sequence below to see why model completion, tool completion, and run completion are different events.

​
The boundaries that matter

run() gives you a final result. stream() gives you an iterator that may span several model calls. Consume it to completion or explicitly cancel/close it when its owner is done.

​
Configure stopping behavior

Set tool-roundtrip and token/cost limits appropriate to the task. Pass a cancellation signal for connection-owned work. Tools that call external services should forward the signal where supported. Cancellation, cost limits, and execution policy cover the controls.

​
Inspect failures at the right layer

A provider can reject a request before a tool runs. Tool validation can reject arguments. Approval can deny the action. A tool can fail after contacting a remote service. These require different responses; retrying the whole run indiscriminately can repeat a side effect. For diagnostics, start with run status, correlation IDs, model configuration, tool outcome, and timing. Observability defaults to metadata; configure content capture deliberately. See recovery for outcomes that cannot be inferred from a local error alone.