“Unknown parameter: 'input[1].status'” when using GPT-5 with OpenAIConversationsSession #1911
Description
Bug Summary
When using the gpt-5 model with the OpenAI Agents SDK (openai-agents==0.3.3) and storing conversation sessions with the Conversations API, a follow-up message in the same conversation fails with a 400 error:
{'error': {'message': "Unknown parameter: 'input[1].status'.", 'type': 'invalid_request_error', 'param': 'input[1].status', 'code': 'unknown_parameter'}}
This occurs consistently on any second turn (follow-up message) within a conversation session.
Debug information
- Agents SDK version:
v0.3.3 - openai version:
1.109.1 - Python version:
3.11.11 - OS: MacOS Sequoia 15.6.1
- Model: gpt-5
Repro steps
Script to reproduce issue:
"""
Reproduce GPT-5 'Unknown parameter: input[n].status' error when using OpenAIConversationsSession.
Requires:
pip install openai==1.109.1 openai-agents==0.3.3
Set your key in OPENAI_API_KEY.
"""
import asyncio
from openai import OpenAI
from agents import Agent, Runner, OpenAIConversationsSession, ModelSettings
from pydantic import BaseModel, Field
import os
# Ensure OPENAI_API_KEY is set in environment (run `export OPENAI_API_KEY=your_key` in terminal)
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
if not OPENAI_API_KEY:
raise ValueError("OPENAI_API_KEY environment variable must be set")
class AgentOutput(BaseModel):
response: str = Field(description="Agent response text")
async def run_repro():
client = OpenAI(api_key=OPENAI_API_KEY)
# --- Create a conversation (this is where the issue happens) ---
conv = client.conversations.create()
conversation_id = conv.id
# --- Create the agent ---
agent = Agent(
name="ReproAgent",
instructions="You are a helpful assistant.",
model="gpt-5", # This triggers the 400
output_type=AgentOutput,
)
# --- First message: works fine ---
session = OpenAIConversationsSession(conversation_id=conversation_id)
print("Running first turn...")
result = Runner.run_streamed(agent, input="Hello!", session=session)
async for event in result.stream_events():
pass # consume stream
print("First turn complete.")
# --- Second message: reproduces 'input[n].status' error ---
print("Running follow-up turn (expecting 400 error)...")
try:
result2 = Runner.run_streamed(agent, input="Tell me more.", session=session)
async for event in result2.stream_events():
pass
except Exception as e:
print(f"\n❌ ERROR: {e}\n")
if __name__ == "__main__":
asyncio.run(run_repro())
Actual Output
Running first turn...
First turn complete.
Running follow-up turn (expecting 400 error)...
Error streaming response: Error code: 400 - {'error': {'message': "Unknown parameter: 'input[1].status'.", 'type': 'invalid_request_error', 'param': 'input[1].status', 'code': 'unknown_parameter'}}
❌ ERROR: Error code: 400 - {'error': {'message': "Unknown parameter: 'input[1].status'.", 'type': 'invalid_request_error', 'param': 'input[1].status', 'code': 'unknown_parameter'}}
Expected behavior
The follow-up message should continue the conversation seamlessly (like GPT-4 agents), not trigger an invalid parameter error.
Observations
- This issue only starts happening when using a gpt-5 model. I did not experience any issues while using the 'gpt-4.1' model.