Greptile Summary

This PR completes the fix for native Ollama reasoning leakage by removing the fallbackContent streaming path in createOllamaStreamFn() that still promoted thinking/reasoning chunks into the final assistant reply when message.content was empty. After the change, only message.content is ever used as the visible assistant text; reasoning fields are silently dropped at both the buildAssistantMessage layer and the streaming accumulation layer.

Key changes:

• buildAssistantMessage() — fallback chain content || thinking || reasoning || "" replaced with content || ""
• createOllamaStreamFn() — fallbackContent accumulator and sawContent flag removed; finalResponse.message.content is set to accumulatedContent (never fallbackContent)
• Tests — "falls back to …" expectations updated to content: []; test descriptions accurately describe the new "drops … when no content" semantics

The implementation is minimal, well-scoped, and fully covered by the updated test suite.

Confidence Score: 5/5

• This PR is safe to merge — it removes a narrow, well-understood fallback code path and backs the change with targeted tests.
• The change is small and focused: two code sites touched, clear root-cause description, and every affected behaviour has a corresponding updated or new test. No new dependencies, no interface changes, and the logic correctly handles all known streaming combinations (thinking-only, reasoning-only, mixed thinking+content, content following reasoning).
• No files require special attention.

_{Last reviewed commit: af15e4f}

Merged.

Landed commit: 7778627
Source head: af15e4f
Original fix commit carried forward: 9504281

This landed via replacement PR flow to preserve attribution while fixing the remaining Ollama streaming-path leak and updating the tests/changelog.