In the DeepSeek Harness (DSH) plugin ecosystem, when handling complex engineering tasks, a single model is often limited by a narrow perspective, architectural blind spots, or hallucination issues. dsh-moa introduces the Mixture of Agents (MoA) architecture, improving the reliability and quality of code generation through multi-model parallel collaboration and review mechanisms.

Plugin Overview

This plugin is maintained by GooDAnDReaREADY and provides multi-agent collaboration capabilities for DSH. It offers file workspace support via the /moa slash command and integrates with Live Canvas. The plugin aims to address the poor performance of a single model under ambiguous or complex instructions.

Core Features

1. Slash Command and Autocompletion

The plugin integrates the /moa slash command with autocompletion support. Users can trigger the command directly in the chat input box or specify a preset configuration using the --preset parameter.

2. Adaptive Survey Gating

For ambiguous or insufficiently defined prompts, the plugin automatically triggers “adaptive survey gating.” The Advisor model generates 2 to 4 structured questions and confirms key details with the user (such as UI style and backend choices) before code generation.

3. Parallel Dispatch and Heartbeat

The system invokes multiple agents in parallel to propose solutions. Each agent’s processing status is displayed with real-time heartbeats (such as ⏳ [3s] Processing...), while redundant system prompts and tool schemas are automatically filtered out to avoid token waste.

4. Disk-Level Candidate Isolation and Promotion

Unlike pure Chat scenarios, the plugin writes candidate solutions into isolated disk workspaces (such as .moa/candidate-N/), preventing cross-contamination between proposals. The Judge model ultimately selects the winner, and the files are directly promoted to the project root directory, with temporary directories automatically cleaned up.

5. Native Settings Card and Presets

The plugin provides a native settings card for configuring the Proposer model, Judge model, and named presets (such as default, fast, deep-reasoning). Custom judging criteria and temperature parameters are supported.

6. Candidate Diff Viewer

A visual diff comparison tool is provided, allowing developers to inspect code differences between different agent proposals line by line.

7. Optimization Mode (Incremental Editing)

In optimization mode, the plugin automatically detects the context of the existing codebase and generates precise incremental modifications rather than destructive full-file rewrites.

8. Fast Mode and Custom Judging Criteria

“Fast mode” is supported for handling simple tasks, and custom Judge criteria are allowed, offering high flexibility.

9. Run History and Win-Rate Leaderboard

The plugin persistently records the type of each run (synthesis, fast mode, survey, etc.) and provides a win-rate leaderboard to help track model performance.

10. Live Canvas 1-Click Preview (Optional)

If the @goodandready/dsh-live-canvas plugin is installed, the promoted HTML is pushed to the Canvas sandbox, and a one-click preview link is provided.

11. Token-Saving Chat Summary

In chat, the plugin replaces large code outputs with concise file lists and architecture summaries, reducing token consumption.

12. Single-Session Model Recovery

The plugin runs as a single-session modifier and automatically reverts to the user’s main session model after completion, preserving context continuity.

Installation and Enablement

No specific installation command is currently provided. The plugin is distributed as an npm package. When installing, note the following dependency requirements:

  • Node.js: Node.js 20 or later is required.
  • Peer Dependencies: The following dependencies must be satisfied during installation:
  {
    "@deepseek-ai/cordis": "^4.0.1",
    "@deepseek-ai/dsh-credentials": "^0.1.0-rc.6",
    "@deepseek-ai/dsh-host-webserver": "^0.1.0-rc.6",
    "@deepseek-ai/dsh-settings": "^0.1.0-rc.6",
    "@deepseek-ai/dsh-tools": "^0.1.0-rc.6",
    "@deepseek-ai/schemastery": "^3.18.1"
  }

Typical Usage

The plugin is started with the /moa command and supports multiple parameter combinations:

/moa build a real-time reactive dashboard with charts and websocket updates

/moa code-review audit the auth middleware and security boundaries

/moa --preset=deep-reasoning solve this math problem step by step

Notes and Limitations

  • Live Canvas Integration Requirement: To use the Live Canvas 1-Click Preview feature, install the @goodandready/dsh-live-canvas plugin in the same Profile.
  • Directory Permissions: The plugin generates a .moa directory and cache files in the workspace.
  • External Dependencies: The plugin uses OpenRouter’s public catalog (model pricing and categories), and this data is cached in ~/.dsh/storages/dsh-moa-catalog.

Summary

dsh-moa is a tool focused on multi-model collaboration and code review. Through disk isolation, parallel proposals, and incremental editing, it helps developers overcome single-model limitations in complex engineering tasks.