The specific example I'm interested in is weather. Many weather apps in HA are pretty lacking so I've been using a tool call through my LLM fallback. While the performance is much better (I can ask it for forecasts anywhere, and often further out than HA apps), Sophia treats all weather questions as climate questions and won't forward them to the LLM.
Another potential solution would be to differentiate between weather and climate as domains. I often ask what the weather is outside only for Sophia to tell me one of the temperatures reported by a thermostat in my home (or that the question couldn't nail down a specific entity, presumably because it matches multiple thermostats).
I know this could probably be solved with an LLM trigger phrase, but I don't like the idea of having to "train" my home users :P