Preface

The core design philosophy of DeepSeek Harness (DSH) is plugin-based. If you need to connect an agent to a local knowledge base containing a specific domain (such as Conjugate Spectral Geometry, CSG), you don’t need to build a retrieval system or manage a vector database yourself—just install dsh-geometry-knowledge. This plugin encapsulates the knowledge base into 31 native DSH tools, allowing the Agent to directly retrieve source text, verify formulas, and manage state during conversation.

Plugin Overview

dsh-geometry-knowledge is a knowledge base plugin for Conjugate Spectral Geometry, maintained by sdoygb and licensed under the MIT License.

Core capabilities include:
* Offline retrieval: based on the BM25 algorithm, including full text of 251 articles, 4,204 chunks, and 869 truth records.
* Native tools: provides 31 native DSH tools (geo_*) covering retrieval, validation, auditing, and file operations.
* Symbolic verification: supports strict symbolic formula verification via SymPy.
* Service reuse: the same kernel can be reused as an MCP service by other clients.

Installation and Configuration

The plugin requires zero configuration by default; after installation, it can be used directly by reading the package-built-in data/ index.

  1. Install via npm:
    dsh plugin add dsh-geometry-knowledge
  1. Or install from GitHub:
    dsh plugin add github:sdoygb/dsh-geometry-knowledge

If you need to customize the data path, you can override the configuration in DSH’s cordis.patch.yml:

- id: geometry-knowledge
  config:
    dataDir: /absolute/path/to/codex++-data
    sourceArticlesDir: /absolute/path/to/app/articles
    stateDir: /absolute/path/to/.geometry-knowledge
    allowShell: false

Core Features

The plugin mainly provides three types of capabilities: offline retrieval, deep validation, and state management.

Offline Retrieval and Reading

  • BM25 retrieval: supports semantic retrieval over article chunks (scope=articles) and the truth layer (scope=truth).
  • Article reading: the geo_read tool supports reading specified articles, with precise section positioning or viewing a structural summary.
  • Truth query: geo_truth is used to retrieve the currently valid truth layer, which includes permanent numbers and proof summaries.

Deep Verification and Auditing

  • Symbolic verification: geo_verify supports symbolic simplification verification of formulas (requires local installation of Python and SymPy), or can perform numerical sampling checks.
  • Dependency graph audit: geo_claim can query a typed claim dependency graph to analyze citation relationships and logical consistency.
  • Consistency check: geo_audit is used to detect definition conflicts, dangling references, and logical cycles.

Workspace and State Management

  • File operations: supports listing, reading (single file limit of 2.5MB), and writing local project files.
  • Git integration: provides the geo_git_* series of tools for commit history and state management.
  • Personal library: supports reading and writing personal memories, emotions, and corrected erroneous statements.

Typical Usage

The following are examples of common commands:

# 检索文章分块
geo_search "弱混合角"
geo_search "谱刚性" scope=truth

# 读取文章
geo_read "10.8" section="Strouhal"
geo_read "7.5" whole=true

# 数学计算
geo_calc "57.93+26.16+5.91"
geo_calc "D=10/7821; D*2"

# 公式验证（需 SymPy）
geo_verify lhs="sin(x)^2+cos(x)^2" rhs="1" variables='["x"]'

Notes

  • Zero-dependency operation: By default, core functions such as retrieval, reading, and writing do not depend on external APIs or the network.
  • Security restrictions: geo_shell_execute prohibits reading .md articles or executing high-risk blacklisted commands.
  • Dependency requirements: geo_verify requires local installation of Python and SymPy to perform strict symbolic verification; otherwise, it will report an error.
  • Network and permissions: geo_web_search requires external network access; geo_chat_history only reads the webui.db of Open WebUI (read-only).

Ecosystem Positioning

This plugin is part of the DeepSeek Harness ecosystem. The DSH philosophy is “everything is a plugin”; the community directory provides a plugin index and has no affiliation with DeepSeek official.

Appendix

  • Source code: https://github.com/sdoygb/dsh-geometry-knowledge
  • Plugin directory: https://www.skillhub.cn/plugins/sdoygb/dsh-geometry-knowledge