dsh-lazy-skill

In DeepSeek Harness, every installed skill contributes a summary entry to the model catalog even when the current task has nothing to do with it. When a family of related skills, such as a toolchain, a phased workflow, or a multi-task project, is installed as flat siblings, their catalog summaries permanently consume prompt budget, pollute the context window, and tempt the model into asking which one should I use instead of simply working. dsh-lazy-skill is built specifically to eliminate that structural waste.

The plugin wraps a group of related skills into a single Bundle Box. Externally it exposes one compact root skill while sub-skills remain hidden from the model catalog by default. The frontmatter field loadSubskills on the root skill controls expansion: load every child, load a declared subset, or return only the root body and let the model read and decide on its own. Together with the skill_browse and skill_load tools, the model can inspect and precisely load hidden children on demand, removing the fixed token overhead of carrying every sibling summary in every prompt.

Built for developers who maintain multi-step toolchains, phased workflows, or project-level sub-task sets inside DeepSeek Harness. If your model catalog keeps growing, per-turn prompt budget is being eaten by irrelevant summaries, or the model keeps prompting you to choose which skill to call, this plugin was made for that situation.