context.usage returns 0 in streaming mode

[brodguez]

Bug description

When using openai-agents-python with Runner.run_streamed, the context.usage values in the hooks (input_tokens, output_tokens, total_tokens, requests) are always 0.

This issue does not happen when using Runner.run (non-streaming mode), where the usage values are correctly populated.

Debug information

• Agents SDK version: v0.0.12
• Platform: macOS

Repro steps

# Tested with multiple models (gpt-4o, o4-mini, o3-mini)
run_config = RunConfig(
    model="o4-mini",
    model_settings=ModelSettings(include_usage=True)
)

class AIAgentsHooks(RunHooks):
    def __init__(self):
        self.event_counter = 0

    def _usage_to_str(self, usage: Usage) -> str:
        return f"{usage.requests} requests, {usage.input_tokens} input tokens, {usage.output_tokens} output tokens, {usage.total_tokens} total tokens"

    async def on_agent_start(self, context: RunContextWrapper, agent: Agent) -> None:
        print(f"Start: {self._usage_to_str(context.usage)}")

    async def on_agent_end(self, context: RunContextWrapper, agent: Agent, output: Any) -> None:
        print(f"End: {self._usage_to_str(context.usage)}")

hooks = AIAgentsHooks()

# Run in streaming mode
result = Runner.run_streamed(
    starting_agent=agent,
    input=input,
    run_config=run_config,
    hooks=hooks
)

async for event in result.stream_events():
    if event.type == "raw_response_event" and hasattr(event.data, "delta"):
        delta = event.data.delta
        if delta:
            yield StreamingChunk(
                data=ChunkData(delta=delta, finish_reason=None)
            )

Output (streamed)
The output is always: 0 requests, 0 input tokens, 0 output tokens, 0 total tokens

If the same config is run with Runner.run, usage works correctly:

result = await Runner.run(
    starting_agent=agent,
    input=input,
    run_config=run_config,
    hooks=hooks
)

Output (non-streamed):
requests, 123 input tokens, 56 output tokens, 179 total tokens

Expected behavior

In streaming mode, the context.usage values should reflect actual usage data just like in non-streaming mode.

[rm-openai]

@brodguez PR #595 fixes this. Note that streaming usage is only available in the very last chunk of the LLM response, so in your example it would be present in on_agent_end but not necessarily before that.

[brodguez]

Thanks! 🙌

I just need it in on_agent_end so that works perfectly for my use case.

Appreciate the support!

[rm-openai]

Will be available in next version 0.0.14