Introduction

The core philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When using Agents for repetitive tasks (such as web patrol or desktop automation), asking the model to replan steps every time leads to high cost and instability. DSH Patrol decomposes this process into two stages: first use an Agent to teach in natural language and solidify the workflow, then use a deterministic Runner to repeat execution.

Plugin Introduction

DSH Patrol is a cross-application browser and Windows desktop patrol plugin for DeepSeek Harness. It allows developers to describe patrol workflows in natural language, solidify them, and then have the system automatically replay them, without relying on the model to guess steps on the spot each time.

  • Name: qigelunbiya/dsh-patrol
  • Maintainer: qigelunbiya
  • Category: Network tools
  • License: MIT
  • Current Status: Alpha / GitHub-first

Core Features

1. Natural Language Teaching and Deterministic Replay

Directly describe “where to open, what to click, what to check, and what to screenshot.” After teaching is complete, save it as a Runbook; subsequent repetitive patrols do not rely on the model to replan the path.

2. Browser and Desktop Dual-Mode Support

  • Browser: Automatically finds Chrome / Edge / Chromium, uses an independent persistent Profile (preserving login state), and automatically loads built-in extensions.
  • Windows Desktop: Prioritizes Windows UI Automation; when insufficient, falls back in order to keyboard shortcuts, Windows OCR, and coordinate clicking.
  • Cross-application workflow: Supports a single workflow that first patrols a web page and generates screenshots, then activates a desktop application (such as WeChat or WPS) to continue operations.

3. Security and Self-Repair

  • Secure credentials: Runbooks store only ${credential:REF}; plaintext passwords or Cookies are not stored.
  • Conservative self-repair: Selector self-repair performs one retry only when there is a unique, exact match, and requires explicit confirmation.
  • Knowledge base: Built-in 微信.md, WPS.md, 百度网盘.md, with workspace-level custom overrides supported.

4. Output and Checkpoints

  • Checkpoint / Resume: Supports pausing at steps requiring manual tokens, QR code scanning, etc., and resuming the original Run after manual completion.
  • Cross-step references: Supports using ${artifact:last-screenshot} in desktop steps to retrieve a screenshot generated earlier.

Installation and Enablement

The project is currently in the Alpha phase. It is recommended to clone the source code directly from GitHub and use the bundled installation script in the repository.

Prerequisites

  • Git installed.
  • Node.js >= 22.
  • pnpm installed.
  • A working DeepSeek Harness source-code environment on the local machine.
  • Windows PowerShell / PowerShell 7.

Installation Steps

  1. Clone the repository:
    git clone https://github.com/qigelunbiya/DSH-Patrol.git
    cd DSH-Patrol
  1. Run the installation script (assuming Harness is located at D:\deepseek-harness):
    .\scripts\install-local.ps1 -HarnessRoot "D:\deepseek-harness"
  1. Start Harness:
    cd D:\deepseek-harness
    pnpm dsh web
  1. In the DSH interface, create a new session and select “Patrol Mode”. Patrol automatically starts the managed browser and connects the extension.

Typical Usage

Web Patrol Example

Create a web patrol that does not require login:

“Help me create a web patrol. Patrol name: Example Domain test patrol. Address: https://example.com. Open the page, confirm that ‘Example Domain’ exists, read the page content, take a screenshot, and generate a report and page summary.”

Desktop Automation Example

Execute operations via Windows UI Automation:

“Windows UI Automation -> keyboard shortcuts/keys -> screenshot+OCR -> coordinate click”

Cross-Application Workflow Example

After a web patrol, send a screenshot to a desktop application (such as WeChat):
1. browser_screenshot
2. ${artifact:last-screenshot}
3. desktop_set_clipboard_files
4. desktop_paste

Notes

  • Current Status: The project is in the Alpha phase. Source installation is recommended rather than the npm package.
  • Permission Control: Desktop Automation in TEST MODE and NORMAL MODE currently does not perform action-level permission tiering.
  • Data Security: Web page content is always treated as UNTRUSTED DATA.
  • Installation Advice: Do not use the npm package as the default installation method. Use the provided PowerShell script to avoid build and integration issues.

Summary

DSH Patrol addresses the nondeterministic behavior of Agents in repetitive patrol tasks. It separates “teach once” from “run many times,” and through browser and desktop dual-mode support, achieves full-chain automation from web inspection to desktop archiving.

  • GitHub: https://github.com/qigelunbiya/DSH-Patrol
  • Plugin directory: https://www.skillhub.cn/plugins/qigelunbiya/DSH-Patrol