Gemini 2.5 Flash returns empty streaming responses intermittently , error generating content no response generated

[ZafeerKhan]

Bug Description

Gemini 2.5 Flash intermittently returns complete streaming responses where every chunk has either no candidates or empty candidate.content.parts. The stream completes without producing any content, and finish_reason stays None. This causes google/llm.py:493 to raise APIStatusError("no response generated").

The issue is intermittent — it occurs on roughly 50% of first-turn requests in our testing, but can also occur mid-session.

Expected Behavior

Gemini 2.5 Flash should return at least one chunk with non-empty content

Reproduction Steps

1. Create an AgentSession with google.LLM(model="gemini-2.5-flash") as primary LLM in a FallbackAdapter
2. Start a voice agent session (or console test with python main.py dev --mode console)
3. Send the first user message (triggers ~10,500 token context: system prompt + greeting + user turn)
4. Observe that the streaming response contains zero usable chunks — all chunks hit continue at google/llm.py:449 (not response.candidates) or google/llm.py:459 (not candidate.content or not candidate.content.parts)

  livekit.agents._exceptions.APIConnectionError: gemini llm: error generating content 
    no response generated
...
- Sample code snippet, or a GitHub Gist link -

Operating System

macOS

Models Used

Primary LLM: gemini-2.5-flash (via livekit-plugins-google). Fallback LLM: gpt-4.1-mini (via livekit-plugins-openai)

Package Versions

livekit-agents==1.3.12
livekit-plugins-google==1.3.12
google-genai==1.55.0

Session/Room/Call IDs

No response

Proposed Solution

Additional Context

No response

Screenshots and Recordings

No response

[ZafeerKhan]

Looks like a Streaming issue. Have you tried running npx ai-patch doctor to sanity-check stalls/partial output and common SSE buffering pitfalls?

I dont believe this is the issue as i faced this Gemini issue on 2 different computers

[aalkhulaifi605]

i am facing the exact issue on 1.4.2 and it only happens with gemini flash 2.5.

[davidzhao]

this would be an issue with the model. if a generation does not produce any results, it's considered an error because the agent cannot continue

[ZafeerKhan]

@davidzhao is there any known workaround for this to use gemini flash 2.5 on livekit?