gemini_reasoning_chunksTier 1 · 70% confidence

ai-agents-gemini-reasoning-chu-gemini-2-5-models-in-stream-mode-with-reasoning-ef-cba28c13

agent: ai_agents

When does this happen?

IF Gemini 2.5 models in stream mode with reasoning_effort enabled return reasoning/thought chunks that are not identified, causing them to be concatenated with main content.

How others solved it

THEN Modify the chunk_parser in litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini.ModelResponseIterator to detect reasoning/thought chunks by checking chunk structure or flags, and separate them into a dedicated reasoning_content field (or into delta.thinking_blocks[0].thinking like Claude) instead of appending to the main content stream. This prevents reasoning data from being exposed directly to users.

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics