BYOK Gemini 3.x models fail on tool calling — thought_signature not included in function call responses

[akashyxp]

Does this issue occur when all extensions are disabled?: Yes/No

• VS Code Version: 1.122.1
• OS Version: Ubuntu 22.04

Problem: Gemini 3.5 Flash and 3.1 Flash-Lite work for basic chat, but fail on the second turn of any agent interaction with:

"Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly."

The first request succeeds with finish_reason: tool_calls, but when VS Code sends the tool result back, the request is rejected because thought_signature (returned by Gemini in the first response) is not echoed back.

Expected: VS Code should preserve and echo back thought_signature fields from model responses when sending tool results in subsequent turns, per Google's docs.

Impact: Gemini 3.x models cannot be used in Agent mode via BYOK. They work in Ask mode (no tools).

Steps to Reproduce:

[HackintoshwithUbuntu]

A small note, 3.5 flash is not accessible for me via the default Google provider. This issue occurred when I used my own key with Google's openai compatible endpoint. Didn't test extensively but the normal Gemini 3 worked when used via the Google auto-model configuration. An easy solution would just be to add 3.5 Flash as a possible model under the Google provider (or switch the underlying to flash latest)

[vijayupadya]

gemini 3.5 flash is now enabled

[dafo-formacio-oberta]

@vijayupadya @pwang347 The main thing here with BOYK and 400 responses if Agent mode is enabled (with tools) is really done?

Thank you!

[HackintoshwithUbuntu]

Not 100% sure if you are asking a question but seems to be ok for me now. This works if you use it via the preconfigured Google provider instead of putting in the google open compatible endpoint as a custom server (i.e. not typing in https://generativelanguage.googleapis.com/v1beta/openai/ as a custom openai endpoint). Of course, thanks to the devs for the quick response