fix: preserve reasoning from compatible Chat Completions providers

[carey-bk]

Related Issues

• Fixes OpenAIChatGenerator drops reasoning / reasoning_content from OpenAI-compatible providers (streaming and non-streaming) #13100

Proposed Changes

Preserve non-empty string reasoning / reasoning_content response extras in the existing ChatMessage.reasoning and StreamingChunk.reasoning fields. Prefer reasoning, fall back to reasoning_content, and ignore empty or non-string values.

Split mixed reasoning/text/tool deltas because a StreamingChunk permits only one payload type. Signal finish/usage on the last derived chunk. Once reasoning appears, track content-block indices separately from provider tool indices: ToolCallDelta.index stays unchanged so argument reconstruction and final tool ordering remain stable, including sparse or out-of-order indices. Without reasoning, preserve the existing chunk indices and start markers.

Update the docstring, component docs, and release note. No schema, dependency-bound, outgoing reasoning-replay, or Responses API changes.

How did you test it?

Offline HTTP JSON/SSE fixtures exercise real SDK parsing and public sync/async generator calls, with network sockets blocked.

• Relevant OpenAI, Azure, and generator-utils tests: 203 passed, 20 integration tests deselected in each environment: Python 3.12.14 / OpenAI 3.24.0; Python 3.12.14 / OpenAI 2.6.0; Python 3.10.21 / OpenAI 2.6.0.
• The 88 new cases on unchanged 87dba65f: 54 failed, 34 passed. All 22 ordinary-stream compatibility cases also pass on that baseline.
• Changed-file mypy, Ruff lint/format, applicable pre-commit hooks, and reno lint passed. Full-tree mypy has the same three baseline errors (yaml.py:40, test_device.py:60, document_splitter.py:276).

These checks used focused Hatch environments. Live providers, the full unit suite, and the full docs build were not run.

Notes for the reviewer

Please review whether these non-standard fields belong in the generic generator or provider-specific integrations, given the historical discussion in #8776. A provider delta may now produce multiple callbacks when reasoning is present. #13079 addresses the separate Responses API path.

The implementation and tests were generated with AI assistance; the assistant performed the local validation above. Personal contributor attestations are left unchecked for human review.

Checklist

• I have read the contributors guidelines and code of conduct.
• I have updated the related issue with new insights and changes.
• I have added unit tests and updated the docstrings.
• I have used a conventional commit type for the PR title.
• I have documented my code.
• I have added a release note.
• I have run pre-commit hooks and fixed applicable issues.

[vercel]

@carey-bk is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
haystack/components/generators
utils.py
haystack/components/generators/chat
openai.py					712, 817
Project Total

This report was generated by python-coverage-comment-action