Introduction

DeepSeek Harness (DSH) provides native subagent invocation capabilities, including subagent and subagent_fork. Native implementations usually inherit the parent-level model routing policy and lack fine-grained control for a single invocation, such as specifying a specific model or reasoning effort (e.g., xhigh).

The dsh-task-models plugin aims to enhance DSH’s native subagent tools by allowing a specific model and reasoning effort to be specified at invocation time, while maintaining compatibility with DSH’s native behavior.

Plugin Overview

  • Name: dsh-task-models
  • Maintainer: TTTPOB
  • License: MIT
  • Core Value: Enhances subagent and subagent_fork by supporting on-demand selection of model and reasoning effort.

Installation and Enablement

Use the following command to install the plugin. After installation, restart dsh web to load the new configuration.

dsh plugin --profile web add dsh-task-models@https://github.com/TTTPOB/dsh-task-models/releases/download/v0.2.0/dsh-task-models-0.2.0.tgz

Core Features and Usage

The plugin overrides the native tools and adds the model and reasoning_effort parameters.

subagent Tool

Uses the spawn provider to create a child session. By default, it is a resumable background session, but it can also be configured as a foreground one-shot execution.

subagent({
  description: "Review authentication",
  prompt: "Review the authentication flow and report correctness issues.",
  model: "opencode-go/deepseek-v4-pro",
  reasoning_effort: "xhigh",
})

subagent_fork Tool

Uses the fork provider to inherit the completed conversation turns from the parent session. The model and reasoning effort selection logic is the same as subagent.

subagent_fork({
  description: "Verify prior analysis",
  prompt: "Check the previous conclusion and identify counterexamples.",
  model: "opencode-go/deepseek-v4-flash",
  reasoning_effort: "high",
})

task_models Discovery Tool

Used to query available models under a provider and the reasoning efforts they support.

task_models({ provider: "opencode-go", query: "deepseek" })

Behavior and Compatibility

  • Parameter Selection: The plugin supports the model and reasoning_effort parameters. If these parameters are not provided, it inherits the caller’s current routing and explicit effort. If reasoning_effort: "default" is specified, it clears the explicit effort in the fork history and re-resolves the model default value.
  • Dependency Limitation: Reasoning-effort injection depends on the local DSH Agent Loop, so it applies to spawn and fork. For external process providers such as ACP, Codex, and Claude Code, each must implement remote protocol support separately.
  • Version Change: In v0.2.0, the standalone foreground task tool was removed; this functionality has been merged into and enhanced within subagent and subagent_fork. It is recommended to use subagent directly and set run_in_background: false.
  • Official Status: This plugin has been officially deprecated. DeepSeek Harness now natively supports similar functionality. Please verify whether this plugin is still needed based on your actual DSH version.

Summary

dsh-task-models provides fine-grained model control capabilities for subagents. Although it has been replaced by official support, it remains a practical utility for enhancing subagent behavior under specific versions or configurations.