Introduction

When performing multi-step tasks in DSH, fixed pipelines or fixed agent topologies often mismatch specific requests. The approach of dsh-meta-orchestrator is to let the model select a working mode based on the request at runtime, record the plan, and then continue execution using DSH primitives.

Overview

jiruidai/dsh-meta-orchestrator is a DeepSeek Harness plugin maintained by jiruidai under the MIT license.

It can be positioned as a “Model-Native Meta-Orchestrator Plugin”: synthesizing task-specific workflows at runtime without relying on fixed pipelines, templates, or hardcoded routing.

Core Features

  • Synthesize task-specific workflows at runtime based on the request, rather than using hardcoded pipelines or fixed agent topologies.
  • Supports five modes: prompt-chaining, parallel-workers, router, supervisor, and evaluation-loop.
  • Adds approximately 330 words of instruction blocks to the system prompt.
  • Provides three recording tools: orchestrate, adapt_workflow, and complete_workflow.
  • Saves the plan as a persistent record, versioning revisions and recording success criteria results.
  • Does not have a built-in runtime; execution uses DSH primitives: subagents, todos, workflow scripts, and plan mode.

Installation and Usage

The plugin requires a profile with a storage stack. The standard web profile (provided by the dsh-web-app bundle) offers this stack; the headless profile does not, so the web profile should be used.

Install using npm:

dsh plugin --profile web add dsh-meta-orchestrator

Install from git and pin a version:

dsh plugin --profile web add github:jiruidai/dsh-meta-orchestrator#v0.2.0

Git installation requires allowing pnpm prepare builds; the configuration is as follows:

allowBuilds: { dsh-meta-orchestrator: true }

After add or remove, you need to restart dsh.

Runtime Environment

  • Node: ^22.19.0 || >=24.0.0
  • peerDependencies: @deepseek-ai/* ^0.1.0-rc.5; documentation states it has been verified with 0.1.0-rc.6
  • Only uses public plugin surfaces:
systemPrompt, skills, tools, storageDomain, agent/pre-step

Typical Usage

After installation, the plugin only records the plan and does not execute it. The typical workflow is:

  1. The model selects one of the five modes.
  2. Call orchestrate to save the plan.
  3. Execute using DSH primitives: subagents, todos, workflow scripts, plan mode.
  4. When the plan needs to change, call adapt_workflow to save a revision.
  5. When the task is complete, call complete_workflow to record the success criteria results.

Applicable Scenarios and Notes

Suitable for scenarios where multi-step tasks are performed in DSH and the plan needs to be recorded, versioned, and persisted.

The plugin does not have a built-in runtime and does not provide a standalone executor; execution relies on DSH primitives.

The plugin runs with the permissions of the current dsh process. You should check the source code and license before installing. The license listed in the current documentation is MIT.

Links

Community Directory Page: https://www.skillhub.cn/plugins/jiruidai/dsh-meta-orchestrator

GitHub: https://github.com/jiruidai/dsh-meta-orchestrator