Stale usage statistic in the AfterModelCallEvent

[almyazin]

Problem

As a Strands Agent SDK user I'd like to have usage statistic for current execution loop via event.agent.event_loop_metrics.latest_agent_invocation property.

But according to the current implementation, the hook is invoked before the statistic is updated:
hook invocation:
https://github.com/strands-agents/sdk-python/blob/e7d3eb97c81eb45da486cb8266464ee60d76b20f/src/strands/event_loop/event_loop.py#L346-L355
usage update:
https://github.com/strands-agents/sdk-python/blob/e7d3eb97c81eb45da486cb8266464ee60d76b20f/src/strands/event_loop/event_loop.py#L412-L414

Question

Is there a way to access used tokens count of the current loop from the callback?

[yonib05]

Hey @almyazin, good question, and the line references made this easy to confirm. Sorry it took so long to get you an answer; this one slipped through the cracks.

You're right about the ordering: event_loop_metrics is still updated after AfterModelCallEvent fires, so latest_agent_invocation won't include the call that just finished at that point.

There is a direct way to get the current call's tokens now, though. Since strands-agents 1.36.0 the event loop attaches usage to the assistant message's metadata before the hook is invoked, so you can read it straight off the event:

def on_after_model_call(event: AfterModelCallEvent) -> None:
    if event.stop_response:  # None when the model call raised
        usage = event.stop_response.message["metadata"]["usage"]
        print(usage["inputTokens"], usage["outputTokens"], usage["totalTokens"])

That gives you exactly the usage of the model call the event describes, no accumulation involved. The aggregated view via agent.event_loop_metrics keeps working as before once the cycle completes. Hooks docs are here: hooks.