Introduction

The core philosophy of DeepSeek Harness (DSH) is plugin-based extension. When using DSH for conversational development, developers often repeat specific tasks across different conversations. Manually writing Agent Skills is time-consuming and makes it difficult to ensure standardized and stable outputs. dsh-skill-forge is a plugin designed to address this pain point. It uses a multi-agent collaboration pipeline to transform accumulated conversational experience into verifiable, reusable Agent Skills.

What Is This

dsh-skill-forge is a multi-agent skill forging system that belongs to the workflow category of the DSH ecosystem. It is maintained by developer Epiphany-Leon and is aimed at automatically converting reusable tasks from conversations into quality-controlled, standardized Agent Skills.

Core Features

The plugin provides the following capabilities through engineering-driven quality control and automated workflows:

  • Seven-stage forging pipeline: Skill generation must pass seven quality gates: trigger, extraction, generation, validation, iteration, audit, and approval.
  • Multi-agent collaboration: It includes six roles—extractor, generator, validator, refiner, adversarial-test generator, and auditor—each performing its specialized function.
  • Validation-driven iteration: Based on SkillOpt-style textual gradient descent, when validation fails, the refiner locates the issue and performs a targeted rewrite instead of full regeneration.
  • Three evolution modes: Supports Darwin (single-skill optimization), Taotie (cross-skill fusion), and co-evolution (alternating evolution of skills and their adversarial test suites).
  • Dream Engine: Supports off-peak batch optimization, performing health checks, bulk optimization, and pattern fusion on the skill library when DSH is idle.
  • Skill orchestration: Supports task decomposition and multi-skill workflow composition, with up to 5 skills per task.
  • Four-layer security defense: Covers the full lifecycle of generation, validation, storage, and use, and defaults to a zero-trust posture for auto-generated content.
  • Full provenance tracking: Records the source conversation, iteration history, version changes, and lineage of each skill.
  • Intelligent recall engine: Dynamically selects and injects the most relevant skill into the conversation using multi-dimensional weighted scoring (e.g., validation score, usage frequency, keywords), instead of relying on simple registration.

Installation and Enablement

  1. Environment preparation:
    Ensure that DeepSeek Harness (DSH) 0.1.0-rc.5 or higher, and Node.js 20+, are installed.

  2. Install the plugin:
    Run the installation command in the DSH profile directory:

    pnpm add dsh-skill-forge@github:Epiphany-Leon/dsh-skill-forge
  1. Enable and configure:
    Configure via the DSH settings UI or by directly editing the profile’s cordis.patch.yml file.

Typical Usage

  1. Complete a reusable task in the DSH conversation window.
  2. Click “Forge Skill” in the right sidebar, or directly use the forge_skill tool in the conversation.
  3. Provide the forging rationale and confirm.
  4. Wait for the pipeline to complete (typically 30 seconds to 2 minutes).
  5. Review the generated skill content in the approval modal.
  6. Approve the skill. It will automatically enter the registry and can be invoked in later conversations through the intelligent recall engine.

Use Cases and Notes

This plugin is suitable for scenarios where high-frequency conversations need to be converted into standardized Agent Skills. It runs inside the DSH process. Before installing, it is recommended to review the source code and license (MIT). Configuration options can be dynamically adjusted via the DSH settings UI or by editing cordis.patch.yml.

Summary

dsh-skill-forge introduces multi-agent collaboration and an engineering-driven forging process to solidify conversational experience into high-quality Agent Skills. For full documentation and source code, see GitHub.