boomzikazita/dsh-skill-lazy

README 摘要

dsh-skill-lazy Lazy skill catalog loading for DeepSeek Harness (DSH). Save tokens on every session by injecting only skill names + one-line summaries instead of the full catalog, and look up full descriptions on demand via skill search . Why By default, DSH injects the full skill catalog — the name plus up to a 500-character description of every installed skill — into the context of every new session. With dozens of skills installed this costs roughly 3K tokens of persistent context per session , even though the model rarely touches most of those skills. dsh-skill-lazy replaces that full dump with a compact catalog: just the skill name and a one-line summary (≤ 40 characters by default) . Full descriptions stay one skill search call away and are only fetched when actually needed. How it works First pass: compact catalog injection - On every agent pre-step, the plugin rewrites the official skill-catalog message into a block where each skill appears as name : one-line summary . - Entries are sorted by call frequency (most-used first) , falling back to alphabetical order for ties. - The underlying source.entries of the catalog message is left untouched, so the harness's own digest/ded…

分类