Introduction

DeepSeek Harness (DSH) adopts a plugin architecture and is designed to enable functional integration through extensions. In local development environments, developers usually already have multiple code assistance tools installed (such as OpenCode, Claude Code, etc.). Connecting DSH with these existing agent CLI tools is key to improving development efficiency.

What Is This

dsh-agent-bridge is a data-driven agent bridging plugin. It was developed by maintainer mike-sl-ig to solve the problem of difficult integration between DSH and existing programming agents.

The plugin describes agent behavior through JSON configuration files (Recipes) instead of hard-coding vendor logic. This means adding a new agent or fixing configuration changes only requires modifying data files, without modifying plugin code.

Core Features

The plugin provides the following core capabilities:

  • Cross-platform launcher: Supports locating and launching local agents on different operating systems.
  • JSON Recipes: Defines an agent’s runtime parameters, output parsing, and version probing through JSON files.
  • Session header interaction: Provides buttons in the DSH session header for switching the candidate agent pool and viewing the previous run result.
  • Toolset: Three tools that the model can call directly:
    • agent_list: Probes available agents and views their location, cost, and supported capabilities.
    • agent_run: Runs an agent as an independent worker and returns the answer, session ID, duration, tool calls, and run report path.
    • agent_mode: Reads or changes the state currently displayed by the button.
  • Run reports: Generates reports in both Markdown and JSON formats for each run.
  • Wizard and self-test: Provides a configuration wizard and self-test feature to assist with configuration.
  • Registry import: Supports importing configurations from a community registry.

Installation and Enablement

Before installation, make sure the environment meets the requirements:

  • Node.js version >= 20
  • Required DSH version: @deepseek-ai/dsh >=0.2.0-rc.2 <0.3.0

Install the plugin using the following command:

dsh plugin --profile <your-profile> add dsh-agent-bridge

Typical Usage

1. Switching and Probing Agents

In the header of a DSH session, you can switch the candidate agent pool using buttons. The model can use the agent_list tool to probe available agents in the current environment:

  • agent_list: Used to probe available agents, including location, cost, and supported capabilities. Using the probe: true parameter additionally runs version detection.

2. Executing Agent Tasks

The model can use the agent_run tool to execute tasks with a specific agent as an independent worker. The tool returns the answer, session ID, execution time, tool call records, and the file path of the run report.

3. Reading and Writing State

Use the agent_mode tool to read the current agent state or manually modify the configuration displayed by the button.

4. Writing Configuration

Agent behavior is defined by JSON configuration files, which need to be placed in the lib/recipes/ directory. Refer to the examples in opencode.json, claude.json, or workbuddy.json to write your own configuration.

5. Using Role Modes

The plugin defines four role modes to control agent behavior strategies:

  • default: Standard mode with no additional instructions.
  • verify: Independent verifier from a different model family, tool use is prohibited, single pass execution.
  • bulk: Low-cost mechanical workflow, prefers direct answers, and prohibits self tool checks.
  • second: Independent second opinion, tool use is prohibited, single pass execution.

Notes and Use Cases

  • Vendor neutrality: The plugin itself does not hard-code any vendor-specific CLI; all vendor logic is defined by Recipes.
  • Tool restrictions: In verify and second modes, the plugin forcibly disables tool calls to prevent unexpected long-running processes or infinite loops.
  • Dependencies and lifecycle: The plugin has no runtime dependencies and no lifecycle hooks during installation.
  • Environment check: Before use, check the source code and license (MIT).
  • Use cases: Suitable for development scenarios requiring cross-model verification, low-cost batch task processing, or obtaining an independent second opinion.

Conclusion

dsh-agent-bridge provides a standardized way to connect DSH with existing local agent tools. Through the JSON Recipe mechanism, it maintains high flexibility and maintainability. To view the full documentation or source code, refer to its table of contents page or GitHub repository.

Table of Contents | GitHub Repository