Preface

The plugin-based architecture of DeepSeek Harness (DSH) allows complex workflows to be extended. When handling complex requirements that need to be decomposed into multiple subtasks, manually assigning dependencies and tracking state is error-prone. dsh-task-orchestrator aims to address this pain point through automated processes, closing the loop from plan creation to subtask execution.

What This Is

dsh-task-orchestrator is a DSH plugin designed to provide a “plan-first” role delegation mechanism. It allows the parent model to create and assign a TODO plan before launching isolated child workers. The plugin includes a dependency-aware work queue and can review final results.

Core Features

  • Plan-first delegation: Before starting execution, the parent model must create and assign a TODO plan.
  • Dependency-aware queue: By default, independent work is executed in parallel using two model streams. Additional tasks enter the dependency queue and wait until resources are released.
  • Automatic routing: Short queries follow the normal single-agent path without starting a team.
  • Graph structure validation: The plugin validates the complete task graph and writes it to the native TODO before creating the first subtask.
  • Isolated execution: Ready tasks run in a new spawn context. Subtasks receive only their assigned data package and dependency reports.
  • State snapshots: Each native TODO or written snapshot contains the task list and versioned plugin checkpoint metadata for execution recovery.

Installation and Enablement

The plugin is distributed as an npm package. During installation, you must specify a DSH profile (for example, web). After installation, restart DSH to load the new configuration fields.

dsh plugin --profile web add https://github.com/hasan-aghayev/dsh-task-orchestrator/releases/download/v1.2.0/dsh-task-orchestrator-1.2.0.tgz

After installation, enable the task-orchestrator-suite group and the task_orchestrate component in the Plugins settings. Configuration and component enablement are profile-scoped.

Typical Usage

For significant work, the parent model must provide an explicit plan argument to the task_orchestrate tool. The plan consists of multiple task nodes, and each node includes:
* A unique id and owner
* Task goal and role
* Dependencies
* Read/write policy

For short queries, there is no need to invoke the orchestration tool; use the single-agent path directly.

Use Cases and Notes

  • Use cases: Suitable for complex long-chain workflows that need to be decomposed into multiple independent subtasks with dependencies among them.
  • Operating mechanism: The plugin directly uses the subagent service. The accompanying workflow components are only auxiliary and do not handle execution.
  • Context checks: Hard sub-context checks must be executed via in-process spawn. If using a remote runtime, configure the corresponding resource policies.
  • Mode control: The mode option in the configuration controls behavior:
    • off: Disables automatic planning instructions and safeguards but keeps the explicit tool available.
    • suggest: Asks whether to use the saved plan.
    • hybrid and auto: Instruct the parent model to plan and execute significant work under existing authorization.
  • Resource limits: By default, at most two logical subtasks run concurrently. When a task returns NEED_* (such as files, history, or budget), it does not automatically retry; the parent model must handle it.
  • Version compatibility: v1.2.0 targets the DSH 0.2 API line and was tested with version 0.2.0-rc.2.

Closing

By introducing a planning stage and a dependency queue, this plugin enhances DSH’s ability to handle complex collaborative tasks. For more configuration details and source code, refer to the plugin catalog page and the GitHub repository.