stream_mode="messages" streams restored history first (from checkpointer) before new LLM tokens - how to stream only new messages?

[VettHor]

When I use stream_mode="messages" with a checkpointer and stable thread_id, the stream first emits old Human/AI messages restored from persistence, and only after that it starts streaming new LLM tokens. I’d like to stream only the new tokens for the current turn. The only workaround I see is filtering out already-known message IDs, but that feels inconvenient. Is there a built-in way to avoid replaying persisted messages and stream only new output?

agent = create_react_agent(
    model="openai:gpt-4.1",
    tools=[fetch_some_data],
    pre_model_hook=pre_model_hook,
    state_schema=UserContext,
    checkpointer=checkpointer,
    store=store,
)

config = {"configurable": {"thread_id": chat.id}}

async for stream_item in agent.astream(
    {
        "messages": [HumanMessage(content=request_text)],
    },
    config=config,
    stream_mode="messages",
):
    if isinstance(stream_mode, AIMessageChunk):
        yield ...