Cancellation and Errors

A successful prompt() returns the final assistant message. Failures reject the promise; Agent always clears its running state in finally.

Cancel a run

Each prompt has an independent AbortController. agent.abort() signals it and returns immediately. Observe the rejected prompt promise before starting another prompt.

The signal reaches the model runtime and local tools. The loop checks it before each model request and tool execution. Waiting for stream acquisition, iterator events, and the final result is abortable. Tool execution must cooperate with the signal; there is no forced interruption of arbitrary local code.

Completion rules

Condition	Behavior
stop or toolUse without calls	Return the final assistant.
stop or toolUse with calls	Execute the batch and continue.
error or aborted	Reject and execute no tools from that response.
length	Reject as truncated; do not execute potentially incomplete calls.
deferred or pending	Reject as unsupported; no polling.
Reached maxTurns	Reject before another model request; never silently report success.
Tool lookup, validation, or execution failure	Append an error tool result and let the model continue unless cancelled.
Unexpected stream or listener exception	Reject and release running state.

Errors use ordinary Error values, including a provider's errorMessage where available. There is no exported structured error-code hierarchy; inspect the failure and completed history rather than assuming every failure has a particular string.

History after failure

Completed messages remain accessible through agent.messages. Draft deltas are not appended individually. A provider error/aborted assistant may appear in display history; request normalization filters such messages before the next model request.

When an assistant's batch is interrupted, the loop adds error results for calls not completed, preserving IDs. It does not execute skipped calls or start another model request. Error/aborted assistants are excluded from this pairing because request normalization removes them on replay.

History retention does not undo tool side effects. For durable session snapshots and storage failures, use coding-agent sessions.

Source

agent.ts, agent-loop.ts, cancellable waits, stream cleanup, and cancellation cases in agent.test.ts.