Fixes #3008 Responses API: multi-turn conversations 400 on turn 2 when passing response.output back as input

[Kilros0817]

Problem

Reusing response.output items as input via model_dump() caused 400 errors in the OpenAI API:

• ResponseOutputMessage included status (required field, not valid input)

• ResponseReasoningItem included status: None and encrypted_content: None

• ResponseFunctionToolCall also included status

• This made the API reject requests with:

  400: Unknown parameter: input[i].status

Solution

This PR introduces:

• as_input() on:
  • ResponseOutputMessage (strips status)
  • ResponseReasoningItem (strips status and encrypted_content)
  • ResponseFunctionToolCall (strips status)
• output_as_input() on Response as a convenience method that:
  • calls as_input() on supported items
  • falls back to to_dict() for others

Example Usage

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,
    )
    conversation.extend(response.output_as_input())  # no more 400s

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0faec06edd

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Strip output-only fields in output_as_input fallback

The new fallback path serializes unhandled output variants with item.to_dict() unchanged, which can still emit output-only keys and trigger the same 400 errors this change is trying to prevent. For example, ResponseToolSearchCall includes created_by (response_tool_search_call.py:30), but the corresponding input shape ToolSearchCall does not accept that field (response_input_item_param.py:153-170), so response.output_as_input() can produce invalid follow-up input whenever such items appear in response.output.

Useful? React with 👍 / 👎.

[Kilros0817]

@chatgpt-codex-connector
pls check again, I fixed

[Kilros0817]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0faec06edd

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Sanitize fallback items before returning input-ready output

output_as_input() currently uses item.to_dict() for every output variant that does not implement as_input(), which leaks output-only fields for some union members. For example, ResponseFunctionToolCallOutputItem can include created_by, but the corresponding input shape (FunctionCallOutput) does not define that field, so replaying response.output_as_input() into responses.create(input=...) can still produce 400 unknown-parameter failures for non-covered item types. This makes the new helper unreliable outside the three explicitly handled classes.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 57d56ff892

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Strip output-only fields before using fallback serialization

output_as_input() still appends raw item.to_dict() for any output item without as_input(), which means output-only fields leak back into next-turn input. This is still reproducible for items like ResponseToolSearchOutputItem, ResponseFunctionShellToolCallOutput, and ResponseApplyPatchToolCallOutput that include created_by, while their input counterparts (ResponseToolSearchOutputItemParam, ShellCallOutput, ApplyPatchCallOutput) do not define that field; in those tool-enabled conversations, conversation.extend(response.output_as_input()) can still trigger 400 unknown-parameter errors despite this fix.

Useful? React with 👍 / 👎.

[Kilros0817]

@chatgpt-codex-connector
fixed

[chatgpt-codex-connector]

Note

To use Codex here, create a Codex account and connect to github.