Gemini streaming: narration before a tool call is shown live but never saved to history

[AmirF194]

Found this reading _query_stream in astrbot/core/provider/sources/gemini_source.py, not from a running deployment, so most of the bug report fields below don't really apply.

What happens: when Gemini streams narration before a tool-call chunk (for example "Sure, let me check that for you." followed by a separate chunk carrying the function_call), that narration is sent to the user live as streaming deltas, but the turn's final saved response is built only from the function-call chunk. The narration never reaches conversation history.

Why: the function-call branch (around line 769) does llm_response = LLMResponse(..., is_chunk=False), builds it from that one chunk, yields it, and returns as soon as any chunk carries a function_call part. That skips the accumulated_text/accumulated_reasoning recovery block further down (around lines 793-826), which is what #2429 added for the plain-text-finish path. The tool-call branch never got the same treatment.

Reproduction: fed three chunks (two narration, one function_call) into the real _query_stream, mocking only self.client.models.generate_content_stream. What the user sees live: ["Sure, ", "let me check that for you."]. What ends up in the final saved result_chain: empty.

AstrBot Version: N/A, found in source at HEAD 763a5ee, not a running deployment
Operating System / Deployment Method / Message Platform Adapter: N/A, code-level finding
Error Log: none, no runtime crash involved

Happy to send a PR extending the accumulated_text/accumulated_reasoning recovery to the tool-call branch if that's useful.

[KashmirAwana]

It seems like the narration part of the streaming isn't being saved to the conversation history when there's a tool call. I'd take a closer look at how _query_stream handles the accumulated_text for these cases.