>>101707944
Wait wtf, it works?? It shouldn't work, it wasn't working just a few days ago.
Oh I see the misunderstanding. If, because of this last prompt, the AI feels like just writing an end sequence token, GMS returns this:
MakerSuite API returned no candidate { usageMetadata: { promptTokenCount: 28, totalTokenCount: 28 } }
aka nothing instead of an empty string.
Which ST throws an error for, unlike Claude, where it would just be an empty successful message. And again the errors only appear aith streaming off which is super scuffed, I don't know if that's a ST thing or the API is just like that.
Anyway, it's not prefill, it's just the same as OpenAI's. You are allowed to end turns on assistant but the response forcibly starts on a new message. Pseudofill, I call it. I don't find it very effective on (newer) OpenAI models but Gemini seems easily influenced by it, might be worth playing with.