dsh-skill-lazy

By default, DeepSeek Harness injects the full skill catalog into every new session context: each installed skill name plus a description of up to 500 characters. With several dozen skills accumulated, that single item costs roughly 3,000 tokens per session, even though the model rarely touches most of them. dsh-skill-lazy replaces the full dump with a compact catalog where each skill appears as just its name and a one-line summary (40 characters or fewer by default), with full descriptions available a single skill_search call away whenever needed.

The core capabilities revolve around on-demand loading. The compact catalog is sorted by recorded call frequency so the most-used skills always appear first. skill_search supports exact-name matches, description substrings, tokenized keywords, and Chinese 2-gram overlap, returning hits grouped by MECE domains with full descriptions and a domain-coverage report. skill_mece_check verifies mutual exclusivity and collective exhaustiveness before you create or evaluate a skill, recommending whether to build a new one or extend an existing one. cost_audit acts as a token-budget watchdog, flagging any tool or skill description that exceeds its limit so you can slim or collapse it.

This plugin is built for DSH users who have installed a dozen or more skills and want to minimize the per-session context overhead while keeping full skill descriptions retrievable on demand. Once installed in an existing DSH environment, it takes effect immediately without requiring you to modify any existing skill files.