fix(responses): guard parser/output_text when response.output is null

[jackpalm88]

Fix parse_response crash when response.output is None

Summary

This patch hardens the OpenAI Python SDK Responses parsing path against output=None payloads.

Observed failure (openai==2.33.0):

• openai/lib/_parsing/_responses.py iterates response.output unconditionally.
• If a Responses backend returns output: null, parser raises:
  • TypeError: 'NoneType' object is not iterable

Patch behavior:

• Treat response.output is None as an empty list in parser loop.
• Apply same guard in Response.output_text property for consistency.

Why

In rare edge cases, a Responses payload may contain output: null; the SDK should handle this defensively.

Changes

1. src/openai/lib/_parsing/_responses.py

- for output in response.output:
+ for output in (response.output or []):

2. src/openai/types/responses/response.py

- for output in self.output:
+ for output in (self.output or []):

Expected result

• parse_response(...) no longer crashes when response.output is None.
• Parsed response behavior is equivalent to an empty output list.
• Response.output_text returns "" instead of crashing when self.output is None.

Suggested tests

• Add regression test where Response.model_construct(..., output=None) is passed to parse_response.
• Assert parsed response returns output=[].
• Assert Response(..., output=None).output_text == "".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 946947ce15

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Make output nullable for strict validation

When the API returns output: null and the client is configured with _strict_response_validation=True, this guard is never reached: _process_response_data calls validate_type() before constructing the Response, and Response.output is still declared as List[ResponseOutputItem], so Pydantic rejects None up front. This means strict-validation users still get an APIResponseValidationError for the null-output response this change is trying to tolerate; the field type needs to allow None as well as the runtime guard.

Useful? React with 👍 / 👎.

[jackpalm88]

Happy to adjust the regression tests if maintainers prefer a different test location or setup.

[jnohclee-rgb]

AI-assisted independent offline check at 946947ce15c6a4740626e2166b3fc57ef926720f against parent 09ece2ee0aa73077d04ac3669d101120076b52fb. Six constructed Response shapes exercise actual output_text and parse_response:

• output=None: both helpers raise TypeError before; this head returns empty text / empty parsed output.
• Empty list and one valid text message: unchanged results; status='incomplete' and fictional metadata remain preserved.
• output=False or 0: previously TypeError, now silently treated as empty, because or [] handles all falsy values rather than only None.
• Message with content=None: both helpers still raise TypeError on both sources; this PR is a top-level output guard, not a general null-tolerant parser.

If only JSON null is intended to be recoverable, [] if response.output is None else response.output (and the equivalent property check) narrows the behavior and leaves malformed falsy values distinguishable. Suggested controls: null, valid empty/list output, metadata/status retention, and an invalid falsy value. The invalid shapes below are deliberately constructed with SDK construct(), bypassing validation; they demonstrate helper behavior, not documented valid API responses.

Scope: six fictional constructed cases, actual helpers, Python3.14/Pydantic2.13.5 and shared local dependencies with archived sources. No provider responses, streaming or full SDK suite; network denied. This is not evidence that the provider emits False,0 or null nested content.

Standalone reproducer

import json
from openai.types.responses import Response,ResponseOutputMessage,ResponseOutputText
from openai.lib._parsing._responses import parse_response
from openai import omit
rows=[]
message=ResponseOutputMessage.construct(type='message',id='fictional_msg',role='assistant',status='completed',content=[ResponseOutputText.construct(type='output_text',text='hello',annotations=[])])
for label,out in [('null',None),('empty',[]),('message',[message]),('false',False),('zero',0),('missing_content',[ResponseOutputMessage.construct(type='message',id='fictional',role='assistant',status='completed',content=None)])]:
 response=Response.construct(id='fictional_response',object='response',created_at=0,model='fictional_model',output=out,status='incomplete',metadata={'fictional':'retained'},incomplete_details={'reason':'max_output_tokens'})
 row={'case':label}
 try:row['output_text']=response.output_text
 except Exception as e:row['text_error']=type(e).__name__
 try:
  parsed=parse_response(response=response,text_format=omit,input_tools=[])
  row.update(parsed_count=len(parsed.output),status=parsed.status,metadata=parsed.metadata)
 except Exception as e:row['parse_error']=type(e).__name__
 rows.append(row)
print(json.dumps(rows))