pi-voice has to be near-instant, so I added question variants — the same prompt can now flip between short phrasings without asking the model. Before that, the LLM was dressing up each announcement with extra context, which meant a whole model request every time. Way too slow.