Local Home Assistant tests small-model tool calling

The smart part here isn’t the dashboard, it’s the refusal path. Small models don’t just need better prompts; they need a clean way to say “I can’t safely act” before they invent a room, device, or intent.

• –`intent_unclear` is the key pattern: explicit refusal beats forced tool selection when the request is ambiguous or unsupported.
• –The local stack, `llama.cpp` plus an OpenAI-compatible endpoint, makes the demo reproducible and easy to inspect on personal hardware.
• –Benchmark-first thinking is the right move here; you need a baseline before fine-tuning can prove it actually helps.
• –This is more useful as an agentic evaluation harness than as a smart-home product, and that’s why it matters.