In the development workflow of DeepSeek Harness (DSH), subagents typically inherit the model configuration of the main conversation by default. When a different model or Reasoning Effort needs to be assigned to a subagent, the existing mechanism lacks flexibility. To address this scenario, the community maintains the dsh-subagent-model-picker plugin.

This plugin provides independent model and Reasoning Effort control for DSH subagents, without changing the model configuration of the main conversation. It is maintained by user chr003 and is open source under the MIT License.

Core Features

This plugin primarily provides the following capabilities:

  1. Independent model selection: Provides an independent Sub: selector next to the main conversation model selector for specifying the model and Reasoning Effort for new subagents.
  2. Session-level configuration: Supports per-session settings. The default policy is Inherited (inheritance), meaning it retains DSH’s normal selection (usually the parent model).
  3. State persistence: Existing subagents retain their bound model and Reasoning Effort when the conversation continues or after DSH restarts.
  4. Read-only view: The view for subconversations is read-only; configuration changes can only be made in the root conversation.

Installation and Enablement

Installing this plugin requires the following environment requirements:
* DSH Web version: 0.1.2-rc.1
* Node.js version: 22.19+ (22.x series) or 24+
* The system must have Git and npm installed

On the machine running DSH, enter the target directory and run the following command (assuming the Web profile name is web):

git clone --branch v0.1.0 --depth 1 https://github.com/chr003/dsh-subagent-model-picker.git &&
cd dsh-subagent-model-picker &&
npm run setup &&
npm run check &&
dsh plugin --profile web add "link:$PWD"

After installation is complete, restart the DSH instance and refresh the Web page. You will then see the Sub: option next to the main conversation model selector.

Typical Usage

  1. Configure a new subagent: In the main conversation, select the provider, model, and Reasoning Effort, then choose a different model or Reasoning Effort in Sub:.
  2. Use the inherited policy: Select Inherited, and the subagent will use DSH’s normal selection (usually the parent model), but tool call options generated by the parent model may still override this setting.
  3. Override default behavior: An explicit selection in Sub: overrides tool call options generated by the parent model.
  4. Descendant inheritance: New descendants will use the latest explicit policy of the current root conversation; if the policy is Inherited, they will use DSH’s normal selection.

Use Cases and Notes

The plugin controls standard in-process DSH subagents, not standalone Codex or Claude CLI products. Before deployment, please note the following:

  • Platform compatibility: The plugin has currently passed testing on Linux (test environment Node 22.22.3 / 24.19.0 and DSH 0.1.2-rc.1); macOS and Windows have not been verified.
  • Permissions and context: The plugin runs with the permissions of the current DSH process and is an in-process plugin.
  • Data privacy: The plugin stores relevant configuration data.
  • Known limitations: If saving the initial binding for a new subagent fails, that subagent may not be managed after restarting DSH and must be recreated.

Summary

dsh-subagent-model-picker provides DSH users with fine-grained control over subagent models. With the independent Sub: selector and persistent configuration, developers can flexibly allocate subagent resources without interfering with the main conversation model.

  • GitHub Repository: https://github.com/chr003/dsh-subagent-model-picker
  • Plugin Directory: https://www.skillhub.cn/plugins/chr003/dsh-subagent-model-picker