Tool not invoked on consecutive calls when using SQLiteSession in OpenAI Agent SDK

API Bugs
1

When using SQLiteSession with Runner.run() in the OpenAI Agent SDK, tools are not invoked on every call if the same input is repeated. Instead, the second, fourth, sixth (even-numbered) calls return a cached/remembered answer directly without calling the tool.

If I remove the session parameter from Runner.run(), the tool is invoked every time as expected.

Code to Reproduce

from agents import Agent, Runner, SQLiteSession, function_tool

session = SQLiteSession("conversation_123")

@function_tool
async def find_prime_minister(name: str) -> str:
    """
    Find the current Prime Minister of a given country.
    
    Args:
        name (str): The name of the country.
    
    Returns:
        str: A sentence stating the current Prime Minister of the specified country.
    """
    print(f"++++++++Finding Prime Minister for {name}++++++++")
    if name.lower() == "canada":
        return "Prime Minister of Canada is Justin Trudeau"
    elif name.lower() == "uk":
        return "Prime Minister of UK is Rishi Sunak"
    elif name.lower() == "pakistan":
        return "Prime Minister of Pakistan is Mian Shahbaz Sharif"
    else:
        return "Prime Minister not found for this country."

agent = Agent(
    name="Assistant",
    instructions="Reply very concisely.", 
    tools=[find_prime_minister],
    tool_use_behavior="stop_on_first_tool",
)

# First call
result = await Runner.run(
    agent,
    "Who is the current Prime Minister of Canada?",
    session=session
)
print(result.final_output)

# Second call with the same input
result = await Runner.run(
    agent,
    "Who is the current Prime Minister of Canada?",
    session=session
)
print(result.final_output)

Expected Behavior

The tool find_prime_minister should be invoked on every call, regardless of session state, so that:

++++++++Finding Prime Minister for Canada++++++++
Prime Minister of Canada is Justin Trudeau

is printed on every run.

Actual Behavior

  • First call: Works as expected (tool is invoked)
  • Second call: Tool is not called; answer is returned directly from somewhere in the session history.
  • Third call: Works again
  • Fourth call: Skips tool call again

This alternates (tool called on odd calls, skipped on even calls).

Notes

Removing session=session from Runner.run() makes the tool run every time. This suggests that the session memory mechanism is returning a previous answer without re-triggering the tool, possibly as a caching or reasoning optimization. However, for tools with side effects or real-time data, skipping the tool can lead to incorrect or outdated results.

2

This is expected behavior. If a tool was run previously and that run and its results were saved in your session, the API will use those previous results instead of calling the tool again.

There doesn’t seem to be a setting per tool that will force it to be called every time. It’s frustrating.

3

Same thing here, for whatever real time information you need to get openAI Agent SDK is not a usable solution.