Thought_signatures are, as i understand, cryptographed thinking tokens. I don’t know why Gemini 3.1 Flash can’t accept thoughts from 3.1 Pro and vice-versa, but it is what it is. I am doing my best to store them in the chatbot i’m making, the docs could be better in this regard, i had to dump a json and look at the response to map the schema and get the signatures of tool_calling.
What i do wonder is how Antigravity handles those thoughts. Since they have not only Gemini 3.1 Flash and Pro but also GPT-OSS. I’m excluding Claude because the limits get exceeded after you say “hi”. Does A.G only include the thoughts of the Model being used at the moment? say, you use Pro, than Flash, than Pro again it will not use the Flash thoughts. Than if i switch to Flash again it will be a chat where the messages have the thoughts included on-and-off again?