New Features
- Add Gemini 3.1 Flash Image Preview model — Support for gemini-3.1-flash-image-preview, a fast and low-cost image generation model. Web Search grounding is also supported.
- Enhanced Langfuse tracing — Added per-round spans (round-N) to the tool-call loop for better visibility into each round's execution. RAG usage is now tracked with a dedicated retriever:file-search span recording source count and toolUsePromptTokens. Added ragTokenRatio to generation metadata for easier RAG token cost analysis.
Changes
- Deprecate Gemini 3 Pro Preview — Removed gemini-3-pro-preview in favor of gemini-3.1-pro-preview. Existing settings are automatically migrated.
- Generalize Web Search support for image models — Web Search support for image generation models is now determined generically (all except gemini-2.5-flash-image) instead of being hardcoded to a specific model.
Bug Fixes
- Fix incomplete toolUsePromptTokens accumulation — RAG token counts were not accumulated into totalUsage when function call limits were reached, resulting in inaccurate ragTokenRatio. Now accumulated correctly in all code paths.