Preface

Running multi-agent collaboration in DeepSeek Harness (DSH) commonly involves opening multiple sessions manually or using scripts to pass tasks between roles. As roles multiply, task states, routing, and session registrations can become scattered across chat logs, making it hard to see the overall progress or pinpoint where bottlenecks occur.

This article introduces not-big-dog/dsh-pipeline-kernel (Community Directory: SkillHub Page). It consolidates “task boards, ledgers, deployment, registries, inspections, packaging reuse, and control panels” into a single Cordis plugin, using configurations to define pipelines rather than hardcoding specific business logic into the core.

What Is This

pipeline-kernel is a pipeline-agnostic management core: A pipeline corresponds to a row in the pipelines table, defined via pipeline_define; the core does not bind to specific businesses like image or video generation, nor does it preset any pipelines—users define and deploy their own.

Maintained by not-big-dog. The plugin has approximately 36 stars on GitHub, current version 0.9.0, licensed under MIT. The UI and documentation follow DSH language settings, supporting both Chinese and English.

The core handles three responsibilities (see docs/ in the repository for details): communication, ledger management, and UI. Business capabilities reside within the roles and presets of pipelines, not in the core.

Core Features

Pipeline Definition and Deployment

Use pipeline_define to describe roles (roles), execution chains (chain), gates (gates), etc.; pipeline_deploy initiates the chain, and pipeline_seed creates sessions for unregistered roles and writes them to the registry.

Task Flow and Ledger

pipeline_push delivers tasks; roles view tasks via pipeline_list, claim them via pipeline_claim, and complete them via pipeline_done while declaring the next route. The core automatically routes tasks to the next step based on the result. Use pipeline_status to view the overall status.

Task labels follow conventions like target:<role>, task:<ID>, attempt:1, etc., for easy routing and tracking.

Web Control Panel

After restarting dsh web, a pipeline capsule button appears at the bottom right of the interface, expanding into a full-height sidebar on the right:

  • Activity: Summary of pipeline runs (role status, in-progress tasks, recently completed, failures highlighted)
  • Pipelines: First-class pipeline entities (Graph thumbnails, KPIs, packaged artifacts); clicking a task opens the Task Inspector
  • New: Define a new pipeline (work folder is required, using the system directory chooser)
  • Archive: Cold storage; archiving does not delete data and is recoverable; permanent deletion only within the archive bin requires inputting the pipeline ID for confirmation

Watchdog (Inspection)

Enabled by default since v0.8, requiring no additional configuration. Scans every 5 minutes: sends wake-up calls for open tasks (with a 10-minute cooldown per task), self-heals zombie tasks that are marked done but lack a complete route, and reports to the supervisor after the entire chain has been idle for 30 minutes. It produces no noise when there are no pipelines or tasks.

Packaging and Reuse

The “Pipelines” view in the sidebar supports viewing packaged artifacts, making it easy to reuse defined pipeline configurations in other environments (see docs/ in the repository for details).

Installation and Activation

Dependencies (v0.9+): Host deepseek-harness 0.1.0-rc.8 or above; plugin pinned to ^0.1.0-rc.8, Cordis 4.x. Node engine requirement: ^22.19.0 || >=24.0.0.

Method 1: Edit the package.json in your profile, add the dependency and register it in the bundles list:

// C:\Users\<you>\.dsh\profiles\web\package.json → dependencies
"dsh-pipeline-kernel": "link:<path-to-this-package>"
// Add "dsh-pipeline-kernel" to the bundles list

Method 2: Use the DSH CLI:

dsh plugin --profile web add <path-to-this-package>

After installation, restart dsh web. New sessions will have the pipeline_* tools available; existing sessions have a fixed tool directory and require recreation.

The plugin runs with the current DSH process permissions. Before installing, review the repository source code and MIT license to ensure it meets your environment and security requirements.

Typical Usage

Follow the README for a quick start—six steps to run a custom pipeline:

  1. Define the pipeline:
pipeline_define {id:"my-pipe", roles:{supervisor:{preset:"pipeline-supervisor"}, ...}, chain:["supervisor",...], gates:{...}}
  1. Deploy:
pipeline_deploy {pipelineId:"my-pipe"}
  1. Seed role sessions:
pipeline_seed {pipelineId:"my-pipe"}
  1. Push a task:
pipeline_push {title, description, tags:["target:<role>","task:<ID>","attempt:1"]}

  1. Flow: pipeline_list → pipeline_claim → pipeline_done {result:"done route:<role|UID> artifact:..."}

  2. View overall status: pipeline_status

For field and topology specifications, see docs/管线定义规范.md (Pipeline Definition Specifications); for tools and HTTP endpoints, see docs/内核接口文档.md (Core Interface Documentation); for architecture details, see docs/内核设计文档.md (Core Design Documentation).

Watchdog Configuration (Optional)

Enabled by default. To adjust or disable it, override in cordis.patch.yml:

# c:\users\<you>\.dsh\profiles\web\cordis.patch.yml
pipeline-kernel:
  watchdogEnabled: false          # Disable watchdog (default true)
  watchdogIntervalMs: 300000      # Inspection interval (default 5 minutes)
  watchdogStallMs: 1800000        # Chain idle threshold (default 30 minutes)
  watchdogWakeCooldownMs: 600000  # Wake-up cooldown per task (default 10 minutes)
  watchdogReportCooldownMs: 1800000

Use Cases and Notes

Who is this for: Developers who need to orchestrate multi-agent unidirectional task chains in DSH—for example, a supervisor with multiple worker roles, requiring a task board and automatic routing, while wanting business logic to remain in pipeline configurations rather than plugin code.

Not suitable as: An out-of-the-box business plugin for image/video generation; the core does not provide preset pipelines—users must define their own.

Notes:

  • The host version must meet deepseek-harness 0.1.0-rc.8+; install after aligning with the plugin’s peer dependencies.
  • Existing sessions will not automatically gain pipeline_* tools; create a new session after installation.
  • The community directory SkillHub is an independent site with no official affiliation to DeepSeek or High-Flyer; the DSH ecosystem philosophy is “everything is a plugin,” and this plugin is one of the community-maintained Cordis extensions.

Conclusion

dsh-pipeline-kernel extracts the general management capabilities of multi-agent pipelines (tasks, ledgers, deployment, registration, inspections, UI) from specific businesses: you define configuration lines, and the core handles chain execution and visualization. If you’re setting up multi-role collaboration in DSH, it’s worth starting with the README’s quick start and the docs/ specifications to run your own pipeline.