dsh-skill-search

By default, DeepSeek Harness injects the name and description of every skill into the model context at startup. A few hundred skills already cost tens of thousands of tokens, and a library of thousands can blow the context window entirely. dsh-skill-search replaces that upfront load with a zero-preload, search-on-demand architecture: nothing is injected at boot; when the AI needs a capability, it runs a keyword search over the SKILL.md files on disk (ripgrep fast path with a pure-Node fallback) and reads only the single file that matches. The context cost is always one SKILL.md, regardless of how large the library grows. The plugin is purely read-only, has no runtime dependencies, exposes no HTTP routes and emits no custom events, so it stays completely non-intrusive. Because skills are shared by file path rather than copied, any update made elsewhere (for example in Codex) is visible immediately without a sync step. Ideal for users whose skill libraries are large, who want to keep fixed context overhead minimal, and who still need flexible, on-the-fly retrieval.