Introduction

DeepSeek Harness (DSH) uses a plugin-based architecture. As the Skill library expands, carrying the full directory with every request consumes large amounts of context and makes Skill selection unpredictable. SkillFlux acts as a runtime layer located between the Agent and its Skill pool. It keeps the full Skill pool out of the model-visible directory, selects only a small number of relevant Skills to mount for each task, and releases them at the end of a turn.

Plugin Overview

yiyuzh/dsh-skillflux is the dynamic Skill runtime manager for DeepSeek Harness.
* Maintainer: YiyuZh
* License: MIT
* Purpose: Automatically discovers, routes, mounts, and unmounts Agent Skills.

Installation and Activation

Using this plugin requires the following environment:
* Node.js 22.20.0 or higher
* DeepSeek Harness profile version 0.1.1-rc.2

  1. Install the plugin into the specified profile (e.g., web):
    dsh plugin --profile web add github:YiyuZh/dsh-skillflux

Replace `web` with another profile name (e.g., `headless`).
  1. Restart the Harness profile.
  2. Verify the running status in a DSH session:
    /skillflux status

For reproducible deployment, lock a specific GitHub commit SHA:

dsh plugin --profile web add github:YiyuZh/dsh-skillflux#<commit-sha>

How It Works

The core workflow of SkillFlux is as follows:
1. Routing and Discovery: Initial matching is performed based on ordered rules and deterministic lexical scoring (supports English and Chinese). If local matching is insufficient, adaptive online discovery is enabled to fetch candidates from the DSH Registry, skills.sh, and authenticated GitHub SKILL.md code search.
2. Reranking: Remote matches are reranked based on relevance, market adoption, repository activity, Stars, Forks, license, and source.
3. Metadata Publishing: Only metadata summaries of selected Skills (name and description) are exposed to the model until the maxActiveSkills limit is reached.
4. Lazy Loading: Registry Skills are pre-mounted, while the full content of cached and remote Skills is downloaded, validated, and loaded only when the model invokes skill <name>.
5. Validation: Cache content is validated using a SHA-256 manifest before loading.
6. Unmounting: The catalog is released at the end of a turn, without deleting cached files.

Core Features

  • Routing Strategy: Supports routing based on ordered rules and lexical scoring (English and Chinese), with optional Embeddings (Ollama or OpenAI-compatible endpoints) for semantic matching.
  • Catalog Limiting: Limits the size of the model-visible catalog via maxActiveSkills.
  • Multi-source Discovery: Supports discovering Skills from the DSH Registry, local cache, skills.sh, and the GitHub SKILL.md index.
  • MCP Support: Supports the MCP Skills extension (io.modelcontextprotocol/skills), allowing skills to be bound from any registered MCP source.
  • Quality Assessment: Provides explainable evidence-level scoring based on dimensions such as relevance, activity, and Stars.
  • Immutability: Resolves remote candidates to immutable GitHub Commit SHAs.
  • Adaptive Discovery: Dynamically decides whether to perform online search based on local confidence.
  • Model Tools: Exposes model tools for loading, searching, and mounting Skills.
  • Management Commands: Provides the /skillflux command for viewing status, routing explanations, usage, and cache cleanup.

Notes

  • Bundle Patch: This plugin patches Cordis to disable the official tool-skill consumer and mounts it under the unique skillflux loader ID. It preserves the active status of the official Skill Registry and the skill-filesystem provider.
  • Security: The plugin runs with the current DSH process permissions. Review the source code and license before installation.

Typical Usage

  • View the current routing status:
    /skillflux status
  • Use a mounted Skill (for example):
    /pdf-reader
  • Use the /skillflux command group to view routing explanations, usage guides, or clean up the cache.

Conclusion

SkillFlux manages the Skill lifecycle through a runtime layer, addressing context inflation caused by large catalogs. It provides fine-grained control, verifiable quality scoring, and lazy-loading mechanisms, making it suitable for scenarios that require managing a large number of Skills in an Agent.

Project URL: github.com/YiyuZh/dsh-skillflux
Catalog Page: skillhub.cn/plugins/YiyuZh/dsh-skillflux