Introduction

In the native Team formation interface of DeepSeek Harness (DSH), spawn_teammate only provides four parameters: name / description / prompt / context, and does not include a model field. This means that teammates within a formation cannot be assigned individual models and can only inherit the default configuration of the main session. This plugin wraps the internal DSH runtime interfaces, filling this gap by allowing users to configure models and reasoning effort levels for specific teammates on the Settings page.

Plugin Overview

The core function of this plugin, sequoiayunus-hue/dsh-subagent-model-config, is to configure individual models and reasoning effort levels for teammates in native DSH Team formations. It provides a graphical settings interface, does not modify any official files, supports hot activation, and does not require restarting DSH.

Installation and Enablement

Installation is completed through the DSH plugin marketplace:

  1. Open DSH Settings.
  2. Go to “Plugin Marketplace”.
  3. Search for dsh-subagent-model-config.
  4. Install and enable it.

Core Features

The core logic of the plugin is based on a rule table, matching teammates by name and injecting the corresponding configuration.

Model and Reasoning Configuration

  • Glob matching: Supports * (any string) and ? (single character) wildcards. Matching is case-insensitive. Rules are evaluated in order, and the first matching rule takes effect.
  • Reasoning effort: Supports selecting from a dropdown the reasoning effort levels supported by the target model.
  • Safe when unset: If only the teammate name is filled in but the model is left empty, the rule does not take effect. The teammate inherits the default model, with no side effects.

Diagnostics

When debug mode is enabled, each time a teammate is spawned, one line of log is appended to the .probe.l file to help with troubleshooting.

Video Preset Integration

When the parent session preset is video, the plugin automatically applies two configurations:
* Teammate name hard gate: If a teammate name starts with video- or video_ but does not match any rule, launching that teammate is rejected, an error is reported, and the current list of allowed naming patterns is shown.
* Capability restrictions: Automatically injects request.persona (masking the main window identity card) and request.toolFilter.allow (video analysis tool whitelist, fail-closed). The whitelist includes watch_video, read, write, edit, glob, grep, pwsh, job_list, job_output, job_kill, skill, and todo_write.

Typical Usage

Add rules under “Subagent Models” on the Settings page. The following is a configuration example:

{
  "enabled": true,
  "rules": [
    {
      "match": "verifier*",
      "note": "独立验证员——便宜快的",
      "provider": "…",
      "model": "…",
      "reasoningEffort": "low"
    },
    {
      "match": "researcher*",
      "note": "检索员——角色先建，模型后补"
    }
  ]
}

Notes

  • Environment requirements: engines.dsh >= 0.1.7-alpha.1, Node >= 22.
  • Compatibility testing: Tested on DSH 0.1.7-alpha.2 and 0.2.0 desktop editions; all 143 npm test test cases pass.
  • Interface dependency: The plugin hooks into DSH internal runtime interfaces. Future official changes to these interfaces may cause it to stop working.

Summary

This plugin achieves fine-grained control over native DSH formation models by wrapping internal interfaces. It requires zero official modifications and is easy to roll back. For users who need precise management of subagent models and reasoning effort levels, this is a practical tool.