Responses API: multi-turn conversations 400 on turn 2 when passing response.output back as input

[achandmsft]

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

Passing response.output items back as input via model_dump() on reasoning models compiles fine but 400s on turn 2:

400: Unknown parameter: 'input[1].status'

model_dump() fabricates status: None, encrypted_content: None for fields the API never returned. The API rejects null unknown parameters.

Separately, there's an undocumented pairing constraint: reasoning+message must be consecutive pairs in input. Filtering to only messages (dropping reasoning) also 400s. model_dump(exclude_none=True) fixes the null fabrication but not the pairing constraint.

Root cause is the spec (openai/openai-openapi#536). Same root cause as #2561 (open since Aug 2025, no fix).

Either model_dump() should produce valid input items, or an .as_input() helper should strip output-only fields.

Workaround: previous_response_id instead of manual history.

Related:

• Responses API: undocumented reasoning+message pairing constraint breaks multi-turn conversations across all SDKs openai-openapi#536 (spec root cause)
• Responses API: undocumented reasoning+message pairing constraint breaks multi-turn conversations #3009 (pairing constraint, separate from this null fabrication bug)
• Responses API: undocumented reasoning+message pairing constraint breaks multi-turn conversations openai-node#1791 (same pairing bug, TypeScript)
• openclaw/openclaw#49167 (OpenClaw, 64.9k forks, production breakage)
• GPT-5 + tool calls: Error code: 400 - Item 'rs_...' of type 'reasoning' was provided without its required following item. openai-agents-python#1660 (Agents SDK)
• ModelHTTPError caused by OpenAI client: Item of type 'function_call' was provided without its required 'reasoning' item pydantic/pydantic-ai#3230 (downstream breakage)

To Reproduce

from openai import OpenAI

client = OpenAI()
conversation = []
for msg in ["What is the capital of France?", "And Germany?"]:
    conversation.append({"role": "user", "content": msg})
    response = client.responses.create(
        model="o4-mini", input=conversation, max_output_tokens=200,
    )
    for item in response.output:
        conversation.append(item.model_dump())  # compiles, 400s on turn 2

Dropping reasoning items and keeping only messages also 400s (pairing constraint).

OS

Windows 11, also reproduced on Linux

Python version

Python v3.13

Library version

openai v2.29.0

[nomiveritas]

Recommended practical solution

It is recommended to use the previous_response_id parameter to continue the conversation instead of manually passing back response.output items (e.g., via model_dump()).
This approach avoids invalid input structures generated by model_dump() (such as status=None, encrypted_content=None, and other output-only fields) and eliminates errors caused by the undocumented reasoning + message pairing requirement.
Using previous_response_id, the API manages the conversation state internally and reliably prevents the 400 error that can occur on the second request when output items are reused as input.

model_dump() should only be used for reading, debugging, logging, or special processing purposes.
For continuing a conversation, never use it as input — in that case, previous_response_id is always the recommended solution.

Respectfully,
Major Noémi Nomiveritas