Introduction

The DSH (DeepSeek Harness) plugin ecosystem emphasizes that “everything is a plugin.” dsh-llm-wiki is a DSH plugin designed to let the agent directly manage the LLM-Wiki knowledge base.

It wraps retrieval, page reading, statistics, validation, fixing, error book inspection, and ingestion into 7 agent tools, suitable for scenarios where you need to query the knowledge base during a session, ingest source text into the knowledge base, and continuously check the knowledge base structure.

What This Is

The owner of @detpecca/dsh-llm-wiki is detpecca, and the license is MIT.

It invokes the LLM-Wiki engine CLI’s --json channel through DSH’s subprocess service. Retrieval, compilation, validation, and fixing are still handled by the LLM-Wiki engine.

The plugin itself has zero runtime dependencies, is pure ESM, has no build step, and requires Node ≥ 18.

Core Features

The 7 tools are described below:

  • wiki_search: Structured signal scoring retrieval (CJK tokenization).
  • wiki_read: Batch page reading / directory indexing, following [[wikilink]].
  • wiki_stats: Page/category/digest/error book statistics.
  • wiki_validate: 4 types of deterministic structure validation.
  • wiki_fix: Deterministic fixing; finalize:true appends an LLM repair round.
  • wiki_errorbook: View the Error Book (self-correction records).
  • wiki_ingest: Compile source text into the knowledge base (full flow of Algorithm 1).

Only wiki_ingest and wiki_fix with finalize:true require LLM configuration; query-focused tools do not require a second set of LLM keys.

Installation and Activation

First, check the environment:

  • DeepSeek Harness is installed, and the dsh command is available;
  • Python ≥ 3.10;
  • Optional: OpenAI-compatible LLM API key.
  1. Install the Python engine

This step installs the LLM-Wiki engine.

pip install git+https://github.com/detpecca/LLM-Wiki.git
  1. Install the plugin

This step adds the plugin to a specified DSH profile.

dsh plugin --profile web add @detpecca/dsh-llm-wiki
  1. Configure the knowledge base path

After installation, you need to write the actual path into the cordis.patch.yml file of the profile. This file is located at:

$DSH_PROFILES/<name>/cordis.patch.yml

If you override the configuration by ID, you must restate all keys. Note: the patch is a whole-line replacement, not a deep merge.

  1. Configure LLM (optional)

Only wiki_ingest and wiki_fix with finalize:true require LLM configuration. The configuration priority is:

  • Explicit values in cordis.patch.yml;
  • Environment variables LLM_WIKI_BASE_URL / LLM_WIKI_API_KEY / LLM_WIKI_MODEL;
  • Not set; in this case, wiki_ingest/finalize will fail with a clear error.
  1. Restart DSH

After the steps above, the agent will be able to call all 7 tools.

Typical Usage

The following phrases can be given directly to the agent:

  • Add D:\notes\transformer.md to my knowledge base, then validate it.
  • What does my knowledge base say about attention mechanisms? Provide the source pages.
  • Check the knowledge base structure for issues and fix them if found.
  • What has been recorded recently in the error book? Why does it keep appearing?

Applicable Scenarios and Notes

  • Suitable for users who already use DSH and want the agent to directly manage the LLM-Wiki knowledge base.
  • Query-focused tools do not require a second set of LLM keys; wiki_ingest and wiki_fix with finalize:true require LLM configuration.
  • The plugin runs with the current dsh process permissions. Check the source code and license before installation.
  • Subprocess output limits are stdout 2MB and stderr 200KB.
  • DSH-Wiki is deprecated; please do not use it further.

Conclusion

The value of dsh-llm-wiki lies in integrating the 7 types of LLM-Wiki knowledge base operations into the DSH agent, allowing querying, ingestion, validation, and fixing to be completed consecutively within the same session.

GitHub repository:

https://github.com/detpecca/dsh-llm-wiki

The community directory page URL is not provided in the verified materials, so it is not listed in this article.