Duplicate Events for Handoffs

[bonk1t]

Describe the bug

When using streaming with handoffs, the SDK emits both ToolCallItem and HandoffCallItem with the same call_id for a single handoff. If frameworks persist both events (as they naturally would when processing the stream), this creates duplicate function_call entries in conversation history, causing API errors on subsequent turns.

Debug information

• Agents SDK version: 0.3.3
• Python version: 3.13

Repro steps

"""
Minimal reproduction: SDK emits duplicate events for handoffs with same call_id
Run: export OPENAI_API_KEY=your_key && python repro_duplicate_handoff_events.py
"""
import asyncio
import os
from agents import Agent, Runner, set_tracing_disabled

if not os.getenv("OPENAI_API_KEY"):
    exit("Set OPENAI_API_KEY")

set_tracing_disabled(disabled=True)

worker_agent = Agent(name="Worker", instructions="You are helpful.")
coordinator_agent = Agent(
    name="Coordinator",
    instructions="When asked to delegate, transfer to Worker.",
    handoffs=[worker_agent],
)

async def main():
    # Turn 1: Collect streaming events
    persisted_history = []
    result = Runner.run_streamed(coordinator_agent, "Transfer to worker")
    
    async for event in result.stream_events():
        if event.type == "run_item_stream_event":
            persisted_history.append(event.item.to_input_item())
    
    # Turn 2: Use persisted history (will fail with duplicate error)
    persisted_history.append({"role": "user", "content": "What did I ask?"})
    result2 = await Runner.run(coordinator_agent, persisted_history)
    print(result2.final_output)

if __name__ == "__main__":
    asyncio.run(main())

Output:

openai.BadRequestError: Error code: 400 - {'error': {'message': 'Duplicate item found with id fc_... Remove duplicate items from your input and try again.', 'type': 'invalid_request_error', 'param': 'input', 'code': None}}

Root cause: In agents/_run_impl.py:498-505, both ToolCallItem and HandoffCallItem are appended for handoffs, sharing the same call_id.

Expected behavior

Only one event should be emitted per handoff, OR the SDK should deduplicate when converting to input history. Currently, frameworks must implement workarounds to skip one of the events.

[ihower]

@bonk1t This isn’t a bug, you’re just using it incorrectly. You shouldn’t convert each item individually inside the loop.
Instead, after the stream ends, use to_input_list() to collect the full persisted history:

async for event in result.stream_events():
    pass

persisted_history = result.to_input_list()

The reason is that loop events can include multiple entries referring to the same call. And to_input_list() is the method specifically designed to produce the correct input for the next turn.

[seratch]

@ihower thanks for answering the question (as always)!

[bonk1t]

Thank you!