Introduction

DeepSeek Harness (DSH) is a plugin-extensible agent development framework. When building agents capable of handling office documents, developers often need to manually write file parsing logic or integrate third-party libraries. dsh-doc-toolkit is a document read/write toolkit designed for DSH and maintained by developer qingmumingyang. By providing standard read/write tool interfaces, it addresses the need for AI assistants to directly read and generate common document formats such as PDF, Word, Excel, and CSV.

Core Features

This plugin provides the following core capabilities:

  1. Multi-format support: Supports reading and writing PDF, DOCX, XLSX, and CSV formats.
  2. Large file handling: Controls the number of returned rows using limit and offset parameters to prevent context overflow.
  3. Chinese PDF export: Supports generating PDF documents. Chinese content automatically locates system CJK fonts and embeds them in subset form, ensuring the text remains copyable and searchable.
  4. AI-friendly interface: Parsing results are returned directly as plain text or TSV, making it easier for large language models to understand and process.
  5. Built-in Skill: When the plugin is loaded, it automatically registers the doc-toolkit-usage skill, guiding the AI on when and how to call the read/write tools.

Installation and Enablement

Before installing, make sure your DSH version meets the compatibility requirements (>=0.1.0-rc.12 <0.2.0). The plugin registers under the Web runtime.

Run the following command to install it:

dsh plugin --profile web add https://github.com/qingmumingyang/dsh-doc-toolkit

After installation is complete, the plugin automatically registers the corresponding tools and Skill.

Typical Usage

After installing and enabling the plugin, the AI assistant can invoke the following functions using natural language:

  • Read documents:
    > “Read the contents of D:\report.pdf for me”
    > “Read the first 50 rows of D:\data.xlsx”
  • Write documents:
    > “Generate a sales report DOCX, including a title and three paragraphs, and save it to D:\sales.docx”

At the underlying level, these requests are resolved into structured read/write interfaces.

Notes

  • Version compatibility: Only supports DSH versions >=0.1.0-rc.12 <0.2.0.
  • PDF limitations:
    • When reading PDFs, the file must contain a text layer. Scanned-document OCR (Optical Character Recognition) is not supported.
    • When writing PDFs, titles, paragraphs, and tables are supported. Chinese text automatically embeds fonts.
  • Node.js environment: Node.js v18+ is only required for installation or development. The plugin runtime is provided by DSH.

Resource Links