AI Agent Hub
Back to plugins
🤖

dsh-subagent

Model Inference Updated 2026.08.27

Run the following command in DeepSeek Harness:

dsh plugin install xie-tj/dsh-subagent

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install xie-tj/dsh-subagent in DeepSeek Harness; the source code is hosted at https://github.com/xie-tj/dsh-subagent.

About this plugin

In DeepSeek Harness, an Agent often needs to offload subtasks to differently configured child models—a quick low-effort draft, a deep high-effort review—yet the core package only exposes the low-level ctx.subagents service without a user-facing configuration layer. dsh-subagent fills that gap: it turns any named subagent profile into a model-visible subagent_ tool that every live root Agent can call directly during a conversation.

Each profile independently binds a backend (spawn for an isolated child context, or fork to inherit the completed parent conversation), a model route (follow the current conversation or pin a specific Provider and Model), and a reasoning effort (inherit the default or set a custom value), all manageable under Settings → Subagents. The plugin also ships a custom_subagent management tool and a default-subagent mechanism: once a default is designated, unnamed delegation requests are nudged toward the matching tool, reducing Agent hesitation.

It suits developers running the DeepSeek Harness Web UI who want their Agent to automatically dispatch subtasks under different reasoning strategies and need fine-grained control over child-model routing.

Use Cases

  • Dispatch drafting and review subtasks to child models with different reasoning efforts
  • Pin a specific Provider and Model route for specialized tasks instead of following the main conversation
  • Set a default subagent so unnamed delegation requests automatically match the right tool

Best For

  • Developers using the DeepSeek Harness Web UI
  • Agent builders who need fine-grained control over child-model routing and reasoning policy
  • Developers who want the Agent to auto-dispatch subtasks without manually naming tools