Introduction

The design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When building complex workflows or agents, a core pain point in development is how to enable AI to continuously maintain a knowledge base and avoid losing context. dsh-llm-wiki aims to solve this problem. It is a workspace-scoped, local-first LLM Wiki plugin.

Plugin Overview

Name: dsh-llm-wiki
Maintainer: wangning19940904
Category: Memory
License: MIT

The plugin builds a local persistent index by treating the directory structure of the current workspace as a knowledge source. It does not rely on external services. All data is stored locally and isolated by the workspace’s absolute path.

Core Features

  1. Data Isolation and Source Files
    The plugin isolates data based on the canonical real path of exec.agent.session.header.cwd. Markdown and TXT files are the only data sources. Index files maintained by the plugin can be deleted and rebuilt at any time.

  2. Hybrid Retrieval
    The search engine combines 384-dimensional feature-hashing embeddings, PostgreSQL full-text search, CJK substring fallback, and [[wikilink]] graph signals, and uses Reciprocal Rank Fusion (RRF) to generate final results.

  3. Knowledge Graph
    The plugin deterministically builds graph edges from Wiki links and same-page entity co-occurrence, supporting exploration of entities, relations, and paths of up to three hops.

  4. Safe Maintenance
    The plugin provides Agent-based maintenance tools. Write operations use SHA-256 hash verification to prevent accidental overwrites. All maintenance operations are visible in normal session logs, with no hidden nested model calls.

Installation and Enablement

Before installation, ensure the environment meets the following requirements:
* Node.js ^22.19.0 or >=24.0.0
* pnpm 11
* DeepSeek Harness 0.1.0-rc.6 or a compatible version

Use the official command to install the plugin:

dsh plugin --profile headless add github:wangning19940904/dsh-llm-wiki

Note: This plugin uses TypeScript source code and relies on the prepare script during installation. If the installation reports that the plugin build was ignored, add the following to the profile’s pnpm-workspace.yaml:

allowBuilds:
  dsh-llm-wiki: true

Typical Usage

The plugin provides a set of tools for Agent invocation:

  • llm_wiki_search: Retrieve the existing knowledge base before the Agent answers.
  • llm_wiki_import: Copy relative .md, .markdown, and .txt files from the workspace into the raw/docs directory.
  • llm_wiki_ingest: Update the derived index, with logging for new or modified files.
  • llm_wiki_read: Read source file content (limited by readMaxBytes).
  • llm_wiki_write: Write source pages into the wiki/sources, wiki/entities, or wiki/concepts directories.
  • llm_wiki_graph: Explore entities, relations, and paths.

Use Cases and Considerations

This plugin is suitable for scenarios that require persistent local knowledge and strict requirements for privacy and offline behavior.

  • Limitations: The current version does not include a Web UI, MCP server, Trace system, background model tasks, PDF/image conversion, or a standalone CLI.
  • File Support: Only locally stored UTF-8-encoded Markdown and text files are supported.
  • Vector Strategy: Uses feature hashing rather than neural embedding models, prioritizing predictable offline behavior over semantic quality.

Summary

dsh-llm-wiki provides a toolchain that allows the Harness Agent to autonomously manage a local Wiki. It addresses workspace-level data isolation and retrieval, making it suitable as a memory-layer component in DeepSeek Harness development.

For more details and source code, visit: GitHub / Skill Center Catalog