Cancel a hosted agent turn (preview)

Cancel an in-flight turn on a long-running hosted agent, understand how cancellation differs between the Responses and Invocations protocols, and how it differs from stopping a session.

Cancel is a route on the agent itself - /responses/{response_id}/cancel or /invocations/{invocation_id}/cancel - scoped to one specific in-flight turn. It’s different from steering, which redirects a turn by queuing a replacement, and different from stopping a session, which is a session-management API, not a per-turn one - see How stopping a session affects an in-flight turn for the side effect stopping a session has on any work in flight.

Long-running agents are in preview. APIs and package versions can change.

Cancel in the Responses protocol

The framework exposes a built-in cancel endpoint - you don’t implement it yourself:

POST {endpoint}/responses/{response_id}/cancel

Calling it only works on a response created with background=true; calling it on a synchronous response returns an error. It sets the same cooperative cancel signal your handler already observes as its third positional parameter, and stamps context.client_cancelled = True so the handler can tell a genuine cancel apart from steering pressure:

@app.response_handler
async def handler(request, context, cancellation_signal):
    while not cancellation_signal.is_set():
        ...
    if context.client_cancelled:
        # Explicit /cancel call. Distinguish from steering pressure, which
        # sets cancellation_signal without this flag.
        return
    # A newer turn is queued (steering) - wind down and return partial
    # output instead, per the steering pattern.

For a multi-turn conversation, canceling a turn doesn’t end the conversation - the underlying chain stays alive and resumable for the next turn (see What happens at the task-store level).

Cancel in the Invocations protocol

The Invocations protocol registers the /invocations/{invocation_id}/cancel route for you, but doesn’t wire it to anything automatically - the framework has no built-in disconnect monitoring or automatic cancellation for this protocol. Without a registered handler, calling it returns 404 not_found.

If you’re using the task primitives (@task / @multi_turn_task) directly, the way to actually end an in-flight or suspended multi-turn chain is delete(), called on the decorated function itself:

@multi_turn_task(name="conv", steerable=True)
async def conv(ctx: TaskContext[dict]) -> dict:
    ...

@app.cancel_invocation_handler
async def cancel_invocation(request):
    invocation_id = request.path_params["invocation_id"]
    await conv.delete(invocation_id)
    return Response(status_code=202)

delete() always removes the chain’s persisted record, for a running turn or a suspended one waiting on its next turn. This behavior is a real difference from the Responses protocol’s /cancel: Responses cancels the current turn but keeps the conversation resumable; delete() ends the conversation - there’s no next turn to resume. There’s no built-in “cancel this turn only, keep the chain alive” primitive for direct task-primitive use; if you need that, build it as an app-level convention (for example, a flag in your own task state that the handler checks and reacts to by returning normally, leaving the chain in its usual suspended state).

One-shot tasks (@task) have no equivalent method - they’re always deleted automatically on their own completion, failure, or cancellation, so there’s nothing separate to call.

What happens at the task-store level

The outcome depends on which mechanism ends the turn. The two protocols don’t always produce the same result:

Task shape	Effect
One-shot (@task)	The system deletes the persisted record. This outcome matches any other terminal result for a one-shot task.
Multi-turn (@multi_turn_task) via the Responses protocol’s built-in /cancel	The chain transitions to suspended, exactly like a normal return from the handler ends a turn. The chain stays alive; a turn that was queued behind the canceled one is promoted and runs next.
Multi-turn (@multi_turn_task) via delete()	The system removes the persisted record outright. The chain isn’t resumable afterward.

How this process differs from stopping a session

Cancel is an agent-level route scoped to one turn. Stopping a session is a session-management API that tears down the whole container. It has no response_id or invocation_id and isn’t part of the cancellation contract described on this page. Because it ends the container rather than targeting a turn, it still affects any turn that happens to be running. To learn more, see How stopping a session affects an in-flight turn.

Related content

• Steer an in-flight agent turn
• Recover long-running work after a crash
• Manage hosted agent sessions
• Resilience for long-running hosted agents