Local small models need a lighter agent mode

[PL-zengying]

I tried running with a smaller local model. Normal chat was okay, but multi-step ReAct/tool calling became fragile pretty quickly. A lightweight mode that uses fewer tool rounds and falls back to plain explanation when parsing fails would make local deployment much more practical.

[Gelisi]

To ensure retrieval quality, the current version prioritizes more capable models. Optimizations for smaller local models are planned for a future update.