Overview

DeepSeek Harness (DSH) supports extending functionality through Bundles. yanxu-dsh-plugin addresses how to integrate Yanxu Dev’s evidence-bound AI coding delivery workflow into the DSH environment. It does not change DSH’s core permission model; instead, it loads a skill named yanxu-delivery-workflow to guide the agent in following specific process standards at each stage of code generation and delivery.

Plugin Positioning

This is a Bundle plugin for DSH. It adds the yanxu-delivery-workflow skill to DSH, enabling the agent to perform evidence-bound AI coding delivery. The plugin uses DSH’s existing bash or pwsh tool to execute commands, ensuring all operations remain within Harness’s permission and sandbox model.

Core Functionality

After the plugin is loaded, the agent will follow the workflow guidance below:

  1. Repository Readiness and Contract: Before AI code generation, first run doctor and create a requirement contract.
  2. Code Implementation Standards: When using implement, an explicit source file whitelist and test commands must be provided.
  3. Isolation and Review: Keep generated changes in an isolated archive until they pass manual review.
  4. State Management: Treat draft PRs, CI success, merging, and deployment as distinct states.
  5. Efficiency Measurement: Use matched, quality-passing real-time tasks to measure development efficiency claims.
  6. Evidence Collection: Guide the agent to complete the collection of PR/CI evidence.

Installation and Enablement

Add version v0.1.0 from the GitHub repository to a specified profile using DSH’s official installation mechanism:

dsh plugin --profile web add github:guannan1031/yanxu-dsh-plugin#v0.1.0

The profile must be restarted after installation. DSH parses cordis.patch.yml to add the plugin layer to the profile’s bundle list. For local debugging during development, installation by absolute path can be used.

Typical Usage

After loading this skill in a DSH session, you can begin working within authorized repositories. The agent will automatically perform the following steps:

  1. Run environment checks and establish a task contract.
  2. Implement code according to the whitelist and test commands.
  3. Maintain the isolated state of the code.
  4. Manage PR states across stages (draft, CI, merge, deployment).
  5. Output efficiency assessments based on real task data.

Applicable Scenarios and Notes

  • Prerequisites: Yanxu Dev itself must be installed in the target project’s environment.
  • Permission Boundaries: The plugin does not bypass DSH permissions, does not run arbitrary host processes from the plugin process, does not store credentials, and does not automatically merge PRs or deploy applications. All actual execution is performed by DSH’s existing sandbox tools.
  • Licenses: The plugin itself is licensed under the MIT License. DeepSeek Harness and Yanxu Dev are separate projects, each with its own license and certification requirements.

Verification

After installation, you can verify that the plugin was integrated correctly with the following commands:

npm test
dsh --profile web --dump-config

The first command verifies the contract exported by the Bundle, and the second command confirms that DSH correctly composes the plugin layer. A normal dsh session should list yanxu-delivery-workflow in the skills directory.