What It Is

DeepSeek Harness (dsh) is currently in developer preview. It provides strong execution capabilities (agent loop, tools, sandbox, etc.) but lacks some software-development process semantics, such as phase gates, acceptance evidence, immutable Spec, and process discipline. dsh does not yet have a mechanism to keep project iterations within controlled boundaries and to distill process knowledge into assets.

Maglev for DSH is a standalone plugin for dsh, intended to add this missing layer. Through skill injection, GUI interaction, and mechanical constraints, it helps development teams and individuals achieve long-term, controlled project iterations in dsh, and it reduces project dependence on specific people or AI stacks.

What It Is

This is a self-contained dsh plugin. After installation, dsh gains the following capabilities:

  1. 29 Maglev skills: Skills are automatically injected via the maglev-bundled provider. They include the main chain of convergence → design → verification → crystallization, as well as governance discipline and knowledge asset distillation skills. Any project (no local .agents/skills required) can discover and use them.
  2. Maglev GUI: Provides “Truth Cards” and “Crystallization Cards.” Truth Cards display the current project state as read by the AI (capability domains, in-progress topics, vision, contracts); Crystallization Cards display the conclusions distilled from each iteration.
  3. Spec tools and verification gates: Provides spec completeness checks and sets mechanical gates before crystallization, turning “verification” from model self-awareness into mechanical constraints.

Core Features

29 Maglev Skills

The plugin includes 29 built-in skills covering the main chain of the full R&D process. These skills are automatically registered in dsh’s skill discovery mechanism; no manual configuration is required.

Maglev GUI

The GUI provides two card views:
* Truth Card: Visible to humans, it displays the project capability domains, in-progress topics, vision, and contract state as read by the AI.
* Crystallization Card: Displays what was distilled in each iteration and appears in the conversation stream as a card.

Spec Tools and Verification Gates

  • Spec completeness check: Performs a mechanical completeness check on spec files.
  • Crystallization gate: Before calling the crystallization tool, the spec completeness gate must be passed, ensuring that only compliant conclusions are distilled.

Installation and Activation

  1. Prerequisites
    dsh is currently in developer preview and has no global dsh command. All dsh operations must be run with pnpm dsh <subcommand> from the dsh source checkout directory.

  2. Install the plugin
    In the source checkout directory, run the following command to install the plugin into the web profile:

    pnpm dsh plugin --profile web add @idea-maglev/maglev-for-dsh
  1. Start Web
    After installation, start the web service:
    pnpm dsh web

After startup, open the page (default http://127.0.0.1:3080), and the Maglev cards (Truth Card + Crystallization Card) will appear.

Typical Usage

In a conversation, simply ask the AI to call the tools to complete the main operations:

  • maglev_reality_status: Reads the current project state (capability domains, in-progress topics, vision, contracts) and generates a Truth Card.
  • maglev_crystallize: Crystallizes verified conclusions into the specs/ directory and generates a Crystallization Card (must pass the spec completeness gate first).
  • maglev_spec_check: Performs a mechanical completeness check on the spec.

Dependencies and Relationships

  • Independence: This plugin is derived once from the Maglev source repository and has zero runtime dependency on the Maglev source.
  • Tool dependencies: The plugin’s host side does not directly depend on @deepseek-ai/dsh-tools (tool definitions are constructed manually), avoiding tool dispatcher symbol identity issues caused by multiple dsh-tools instances.
  • Skill injection: Skills are injected via ctx.skills.registerProvider; their rank is lower than project skills, so project skills take priority.

Summary

By introducing 29 skills, GUI cards, and spec verification gates, Maglev for DSH adds R&D process semantics and discipline to dsh. It helps users achieve controlled iteration in dsh and ensures that knowledge assets are properly distilled and traceable.