Tool-Calls not saved in session

[Maxi-J]

Describe the bug

When using sessions (no matter which type), the tool calls are not saved, which results in repeated tool calls after each turn, as the LLM lacks the previous tool call's output and awareness that the tool has already been called.

In case it matters, I am using the streamed runner. The expected tool call events in result.stream_events() are properly present though, just not automatically persisted in the session, which I would expect.

Using my own custom session doesn't help either, as the add_items() function doesn't receive the tool call entries in the first place.

Handoffs have no input filter set.

Debug information

• Agents SDK version: 0.8.0
• Python version: 3.14.2

Repro steps

Seems a general issue to me, happy to supply an example if it works for others.

Expected behavior

Tool calls including outputs are persisted in the session, so the following turns can use that info.

---

If you can't reproduce or need additional info, let me know. Thanks!

[seratch]

Thank you for writing in. If possible, could you please share repro steps?

[Maxi-J]

Sure thing. I created a simplified and generalized version of my setup.
I could verify, that only the streamed version (Runner.run_streamed) has the issue of not persisting tool calls in the session.

Code

Python code

from openai import AsyncAzureOpenAI
from config import Config
from agents import (
    Agent,
    AgentUpdatedStreamEvent,
    ModelSettings,
    OpenAIChatCompletionsModel,
    RawResponsesStreamEvent,
    RunItemStreamEvent,
    Runner,
    ToolCallItem,
    ToolCallOutputItem,
    function_tool,
    set_tracing_disabled,
)
import asyncio
from agents.extensions.handoff_prompt import prompt_with_handoff_instructions
from openai.types.shared import Reasoning
from agents.extensions.memory import EncryptedSession, SQLAlchemySession
from openai.types.responses import ResponseTextDeltaEvent, ResponseContentPartDoneEvent

client = AsyncAzureOpenAI(
    api_key=Config.AZ_OAI_API_KEY,
    azure_endpoint=Config.AZ_OAI_ENDPOINT,
    api_version=Config.AZ_OAI_API_VERSION,
)
model = OpenAIChatCompletionsModel(Config.AZ_OAI_DEPLOYMENT_NAME, client)
default_model_settings = ModelSettings(
    extra_args={"max_completion_tokens": 1000}, reasoning=Reasoning(effort="low")
)
set_tracing_disabled(True)


@function_tool
def get_knowledge() -> str:
    """Fetch knowledge base content to assist with user queries."""
    return "The IT is located in CITY. The phone number for the IT is 123-456-7890."


knowledge_agent = Agent(
    "Knowledge Agent",
    handoff_description="This agent handles knowledge-related tasks to answer the user's questions.",
    tools=[get_knowledge],
    model=model,
    model_settings=default_model_settings,
    output_type=str,
    instructions="You are a helpful assistant that provides information based on a knowledge base. Use the provided tool to fetch relevant information to answer the user's questions. Only answer what the user specifically asked for.",
)
coordinator_agent = Agent(
    "Coordinator Agent",
    handoffs=[knowledge_agent],
    tools=[],
    model=model,
    model_settings=default_model_settings,
    output_type=str,
    instructions=prompt_with_handoff_instructions(
        "You are a coordinator agent that routes user queries to the appropriate tool agents. If a query is related to knowledge or a question, always immediately route it to the Knowledge Agent. Any general chatting like greeting is your responsibility."
    ),
)
current_agent: Agent = coordinator_agent


def get_session(session_id: str) -> EncryptedSession:
    underlying_session = SQLAlchemySession.from_url(session_id, url="sqlite+aiosqlite:///test.db", create_tables=True)
    session = EncryptedSession(
        session_id=session_id,
        underlying_session=underlying_session,
        encryption_key=Config.AGENT_SESSION_ENCRYPTION_KEY,
        ttl=3600,
    )

    return session


async def handle_turn(prompt: str, session_id: str):
    session = get_session(session_id)
    global current_agent

    # Debug: Print all items in the session before processing the turn
    print("Session items: " + str(await session.get_items()))

    result = Runner.run_streamed(current_agent, prompt, session=session)  # type: ignore

    async for event in result.stream_events():
        if isinstance(event, RawResponsesStreamEvent):
            data = event.data
            if isinstance(data, ResponseTextDeltaEvent):
                print(data.delta, end="", flush=True)
            elif isinstance(data, ResponseContentPartDoneEvent):
                print("\n")
        elif isinstance(event, AgentUpdatedStreamEvent) and event.new_agent != current_agent:
            print(f"Handoff from {current_agent.name} to {event.new_agent.name}")
            current_agent = event.new_agent
            continue
        elif isinstance(event, RunItemStreamEvent):
            new_item = event.item
            agent_name = current_agent.name

            if isinstance(new_item, ToolCallItem):
                item = new_item.raw_item
                print(f"{agent_name}: Calling tool '{item.name}' with input: {item.arguments}")  # type: ignore
            elif isinstance(new_item, ToolCallOutputItem):
                print(f"{agent_name}: Tool call output: {str(new_item.output)[:100]}...")


def main():
    while True:
        user_input = input("User: ")
        asyncio.run(handle_turn(user_input, session_id="test"))


if __name__ == "__main__":
    main()

Usage

• Add deps to venv with uv add aiosqlite openai-agents[encrypt,sqlalchemy]
• Execute with python main.py

Steps to reproduce

I tested it with the following inputs:

1. Hi
2. Where is the IT department located? Just the location
3. Whats the phone number?

Output

See session debug print to see that indeed the tool calls are not persisted and the LLM has to re-call the tool every time.

User: Hi
Session items: []
Hi! 👋  
How can I help you today?

User: Where is the IT department located? Just the location
Session items: [{'content': 'Hi', 'role': 'user'}, {'id': '__fake_id__', 'content': [{'annotations': [], 'text': 'Hi! 👋  \nHow can I help you today?', 'type': 'output_text', 'logprobs': []}], 'role': 'assistant', 'status': 'completed', 'type': 'message', 'provider_data': {'model': 'gpt-5.1', 'response_id': 'chatcmpl-D6FSAd1qlUfk7fTb7ppKuBiiY8auo'}}]
Coordinator Agent: Calling tool 'transfer_to_knowledge_agent' with input: {}
Handoff from Coordinator Agent to Knowledge Agent
Knowledge Agent: Calling tool 'get_knowledge' with input: {}
Knowledge Agent: Tool call output: The IT is located in CITY. The phone number for the IT is 123-456-7890....
The IT department is located in CITY.

User: Whats the phone number?
Session items: [{'content': 'Hi', 'role': 'user'}, {'id': '__fake_id__', 'content': [{'annotations': [], 'text': 'Hi! 👋  \nHow can I help you today?', 'type': 'output_text', 'logprobs': []}], 'role': 'assistant', 'status': 'completed', 'type': 'message', 'provider_data': {'model': 'gpt-5.1', 'response_id': 'chatcmpl-D6FSAd1qlUfk7fTb7ppKuBiiY8auo'}}, {'content': 'Where is the IT department located? Just the location', 'role': 'user'}, {'arguments': '{}', 'call_id': 'call_riwmOw9JXeHwdeQaoUfsu3tJ', 'name': 'transfer_to_knowledge_agent', 'type': 'function_call', 'id': '__fake_id__', 'provider_data': {'model': 'gpt-5.1', 'response_id': 'chatcmpl-D6FSLhqiNULGNxrGuz6g6wa8o41Bc'}}, {'call_id': 'call_riwmOw9JXeHwdeQaoUfsu3tJ', 'output': '{"assistant": "Knowledge Agent"}', 'type': 'function_call_output'}, {'id': '__fake_id__', 'content': [{'annotations': [], 'text': 'The IT department is located in CITY.', 'type': 'output_text', 'logprobs': []}], 'role': 'assistant', 'status': 'completed', 'type': 'message', 'provider_data': {'model': 'gpt-5.1', 'response_id': 'chatcmpl-D6FSNt7XDBZA4CRZfdtoA2hMwxSBx'}}]
Knowledge Agent: Calling tool 'get_knowledge' with input: {}
Knowledge Agent: Tool call output: The IT is located in CITY. The phone number for the IT is 123-456-7890....
123-456-7890

[seratch]

Thank you so much for detailed report. #2433 resolves this issue and the fix will be included in the upcoming patch release.