STS2 Local Agent maps tool-calling lessons

This is less an LLM breakthrough than a systems-design case study: the agent gets better when the action space gets smaller and the game loop gets more deterministic. For small local models, reliability is being engineered, not prompted.

• –State-based tool routing and single-tool mode are the strongest ideas here because they shrink the decision surface and avoid stale-state mistakes after each action.
• –The energy guard, smart-wait loop, and parser fallback are exactly the sort of hardening local agents need when both the host API and the model are a little messy.
• –If "block first" is non-negotiable, it should be enforced outside the prompt; a validator or rule gate will beat hoping a 27B model remembers the priority.
• –A rolling combat summary makes more sense than raw chat history, but it should be structured state, not prose recap.
• –Fair A/B testing will need fixed seeds, archived state traces, or matched encounter sets, otherwise you're mostly measuring RNG.