Greptile Summary

This PR fixes a field name mismatch in the Ollama native API stream handler: the code was reading message.reasoning to capture thinking/chain-of-thought output, but the Ollama /api/chat endpoint actually returns this content under message.thinking. This caused thinking-capable models (Qwen 3, DeepSeek R1, etc.) to silently discard all thinking content when streamed through the native Ollama endpoint.

The fix is surgical and correct:

• OllamaChatResponse interface type updated: reasoning? → thinking?
• buildAssistantMessage fallback updated to read response.message.thinking
• The streaming accumulator's else if branch updated to check chunk.message?.thinking
• All affected tests updated to match the corrected field name

The logic is sound: the streaming accumulator only falls back to thinking content when content is empty in a given chunk, which matches Ollama's actual streaming behavior where thinking and response content arrive in separate chunks. After streaming completes, the fully accumulated string (including any thinking content) is placed into finalResponse.message.content before buildAssistantMessage is called, so the thinking fallback in buildAssistantMessage serves the non-streaming (single-response) path.

Confidence Score: 5/5

• This PR is safe to merge — it's a minimal, well-tested fix that corrects a clearly-documented API field name mismatch with no behavioral regressions.
• The change is a one-field rename throughout a single file and its tests. The fix directly matches the Ollama native API specification (message.thinking). The 20 existing tests all continue to pass, and the streaming accumulation logic (using else if to fall back to thinking only when content is empty) correctly reflects how Ollama streams thinking model output. No new edge cases are introduced.
• No files require special attention.

_{Last reviewed commit: 1bdbc24}

Closing this as superseded by #39292, which merged on 2026-03-08 and covers the same native Ollama message.thinking path while preserving message.reasoning as a compatibility fallback.

Thank you for the original fix direction here. The merged path kept the broader compatibility handling and added follow-up coverage for mixed fallback/content streaming.

If there is behavior in this PR that you believe did not land in #39292, reply here and we can reopen or spin out the missing piece explicitly.