DeepSeek Harness’s Agent Teams module allows building collaborative agents, but directly controlling the Provider, Model, and reasoning effort used by each teammate usually requires configuration in code. The dsh-agent-team-model plugin provides a visual configuration solution. It allows developers to specify model parameters by default rules or teammate name through the interface, and to view the actual routing assigned to each teammate in the current session. It should be clear that this plugin is only responsible for configuring teammate routing; the teammates themselves are provided by DeepSeek Harness’ official experimental Agent Teams layer.

Core Features

  • Visual configuration: Configure default teammate routing in the Settings page or the teammate button next to the input box.
  • Rule overrides: Supports precise overrides by teammate name in spawn_teammate, and also supports configuring global default rules.
  • Real-time status: View the leader, teammate list, running status, and actual routing used by each teammate in the current conversation.
  • Reasoning effort control: Supports selecting a reasoning effort, or letting the model use its own default value.

Prerequisites

This plugin does not replace the official Agent Teams layer. The following dependencies must be installed for it to work properly:

  1. DeepSeek Harness: Verified version 0.1.5-alpha.2.
  2. Official Agent Teams layer: @deepseek-ai/dsh-experimental-agent-team-profile. The profile bundled with DSH does not enable this feature by default and must be added explicitly.
  3. Web Profile: The host side must register /agent-team-model/api, and the profile must provide webServer and webRuntime. Headless / SDK / ACP profiles cannot activate this plugin.

Installation and Activation

Ensure the above dependencies are ready before installation.

# 1. 安装官方 Agent Teams 层（必须）
dsh plugin --profile web add @deepseek-ai/dsh-experimental-agent-team-profile

# 2. 安装本插件
dsh plugin --profile web add github:XMeowchan/dsh-agent-team-model

After installation, restart DeepSeek Harness and refresh the Web page. The configuration file is saved by default at $DSH_HOME/agent-team-model.json.

Usage Steps

  1. Open Settings → Teammate Models, or click the teammate button next to the input box.
  2. Turn on the enable switch.
  3. Select the default provider, model, and reasoning effort.
  4. If special rules are needed, add a teammate name and select a separate routing.
  5. New teammates created afterward will use the corresponding rules.

Configuration Notes

The configuration is saved at $DSH_HOME/agent-team-model.json and supports the following fields:

{
  "enabled": true,
  "default": {
    "provider": "deepseek-official",
    "model": "deepseek-flash",
    "reasoningEffort": "high"
  },
  "overrides": {
    "researcher": {
      "provider": "deepseek-official",
      "model": "deepseek-flash"
    }
  }
}
  • enabled: false: Fully inherits the leader routing.
  • default: The routing used by teammates without a specific rule.
  • overrides[teammate name]: Exact match by teammate name.
  • Missing provider or model: Not injected; inherits the leader.

Notes

  • Only affects new teammates: Configuration only takes effect for newly generated teammates. Existing teammates do not hot-switch models and must be recreated.
  • Toolset unaffected: The plugin only changes LLM routing; teammate toolsets still follow the leader’s preset.
  • Permission scope: The configuration API only accepts authority from loopback or authorities explicitly trusted by the host; it is a local configuration interface.

Summary

This plugin provides visualized routing management for DeepSeek Harness’s Agent Teams, solving the cumbersome problem of manually writing configuration code. It is suitable for developers who need fine-grained control over model invocation strategies in Agent Teams.