dsh-skill-router

Most skill loading still relies on the model to decide every step, which often means late loads, wrong picks, or missed opportunities entirely. dsh-skill-router steps in before the model starts thinking: it runs deterministic YAML rules against the latest user message, pours matched skill bodies into the current step on a hit, and stays completely silent on a miss. Zero LLM calls, zero token cost, zero added latency, and each skill pours at most once per session. A broken policy file simply falls back to bundled defaults without breaking the session.

The design philosophy is simple: rules are the policy, the plugin is the execution. Add a URL-path rule and a Feishu document routes to the right skill in microseconds. Write a short trigger phrase in a skill's whenToUse field and it becomes a secondary match. No embeddings, no model judge, fully reproducible every time. It complements skill-bartender's judgment layer or works perfectly standalone.

Best for developers who already maintain a growing skill catalog in DeepSeek Harness and want loads to be faster, more accurate, and token-free. If you prefer tuning routing by editing data rather than writing code, this YAML file is your only entry point.