OpenAIConversationSession doesn't work with reasoning models

[jacksonrnewhouse]

Please read this first

• Have you read the docs? Yes, although of minimal use as they point to a nonexistent doc page that it seems like codex just invented, the link at https://openai.github.io/openai-agents-python/sessions/#openai-conversations-api-memory points to https://platform.openai.com/docs/guides/conversational-agents/conversations-api.
• Have you searched for related issues? Yeah, no other mention of the session.

Describe the bug

You get an opaque error like

Error: Error code: 400 - {'error': {'message': 'Invalid item', 'type': 'invalid_request_error', 'param': 'item', 'code': 'invalid_item'}}

It turns out to be due to sending reasoning contents, which the conversation API does not like. I patched the library to print out the items and this was problematic

[{'content': 'Hello', 'role': 'user'}, {'id': 'rs_68c1c3694bf88195bbb478a8bc9b059f07382c9e416afbdf', 'summary': [], 'type': 'reasoning'}, {'id': 'msg_68c1c36a1b88819593c15df26da4a17007382c9e416afbdf', 'content': [{'annotations': [], 'text': 'Hello! How can I help you today?', 'type': 'output_text', 'logprobs': []}], 'role': 'assistant', 'status': 'completed', 'type': 'message'}]

Debug information

• Agents SDK version: v0.2.11
• Python version Python 3.12.6

Repro steps

#!/usr/bin/env python
import asyncio
from agents import Agent, Runner, OpenAIConversationsSession

async def test():
    agent = Agent(name="test", model="gpt-5", instructions="You are a helpful assistant.")
    session = OpenAIConversationsSession()
    runner = Runner()
    
    try:
        result = await runner.run(agent, "Hello", session=session)
        print(f"Success: {result.final_output}")
    except Exception as e:
        print(f"Error: {e}")

asyncio.run(test())

Expected behavior

It should work like the other sessions.

[neuromaxer]

I would add that docs for OpenAIConversationsSession are empty (linked at the bottom of Sessions docs)

[Silve2]

I noticed the same bug in my code, are there any walkarounds ?

[slyapustin]

Same issue

[quartzjer]

Also encountering this. Chat suggested this workaround and it seems to work for now until this is fixed (I’d like/expect to have tools/thinking in the history since it’s server-side w/ this session):

from agents import Session, TResponseInputItem

class SafeConversations(Session):
    def __init__(self, inner: Session):
        self.inner = inner

    async def get_items(self, limit=None):
        return await self.inner.get_items(limit=limit)

    async def add_items(self, items: list[TResponseInputItem]) -> None:
        cleaned = []
        for it in items:
            if it.get("type") == "message" and it.get("role") in ("user", "assistant"):
                parts = it.get("content") or []
                text_parts = [p for p in parts if p.get("type") == "text" and isinstance(p.get("text"), str)]
                if text_parts:
                    cleaned.append({"type": "message", "role": it["role"], "content": text_parts})
        if cleaned:
            await self.inner.add_items(cleaned)

    async def pop_item(self):
        return await self.inner.pop_item()

from agents import OpenAIConversationsSession
base = OpenAIConversationsSession()
session = SafeConversations(base)

[cablecreek]

Same issue for me

OpenAIConversationsSession --> none of the gpt-5 models work only gpt-4 related models.

I noticed the SQLiteSession is also temporary, so just use this if you are ok for ephemeral sessions.

class SQLiteSession(SessionABC):
    """SQLite-based implementation of session storage.

    This implementation stores conversation history in a SQLite database.
    By default, uses an in-memory database that is lost when the process ends.
    For persistent storage, provide a file path.
    """

[ihower]

@cablecreek This issue should have been fixed in #1883
Please try upgrading to version v0.4.0 and check if it works now.

[cablecreek]

I've upgraded and the issue seems to be resolved. My bad! thanks @ihower