Thanks — I’m sorry you’re blocked by that error. I’m responding about the Gemini 3 Flash function-calling issue you described (first call succeeds, then immediate 400: "Please ensure that function call turn comes immediately after a user turn or after a function response turn", repeated thereafter). I don’t have a confirmed known-bug entry for Gemini 3 Flash + livekit-agents 1.3.9 / livekit-plugins-google 1.3.9. To help narrow this down quickly, please provide the following so I can reproduce and pinpoint the cause: 1. A minimal reproducible conversation log (full sequence of chat messages and role labels — user/assistant/function — from session start through the failing turn). Redact any sensitive data. 2. The exact request payload sent to Google for the first successful function call and for the failing subsequent call(s) (headers + body or the equivalent client call). 3. The code snippet showing how you format and send chat context to the Google plugin/LLM (especially how you append function responses, user turns, and any “thought” or signature metadata). 4. Which function-calling mode you’re using (explicit function call vs. model-invoked function call) and whether you pass an explicit

function_call

parameter. 5. Whether you see any differences in message ordering or role names between the OpenAI case (works) and Gemini case (fails). Quick troubleshooting steps you can try now (these often reveal what’s malformed): • Verify role/order strictly: ensure the immediate previous message before the function call is a user message or a function response message (no assistant non-function messages, no system messages between them). Gemini enforces this ordering. • Strip any extra metadata from assistant messages (e.g., thought signatures, custom fields). Try sending only {role, content} entries to see if the model accepts the function call. • Compare the exact JSON sequence you send to OpenAI vs Google for the same interaction. Look for differences in role naming (

assistant

vs

tool

), extra assistant messages, or duplicate assistant turns. • If you include a function response message, ensure it is labeled with role

function

and includes content exactly where the next call is expected. • Temporarily disable any plugin-specific wrappers that mutate the chat history (one at a time) to isolate whether the plugin formatting causes the issue. What I’ll do once you provide logs/snippets: • Reproduce locally against Gemini 3 Flash with the same request format you’re using. • Check whether the plugin adds assistant-thought metadata or transforms roles that could break Gemini’s strict turn-order rule. • Suggest a minimal patch or workaround (strip/translate roles, reorder messages, or adjust function_call parameter) you can test immediately. If you can paste the minimal log + request payload next, I’ll start reproducing and give a concrete fix.