Preface

A common issue encountered when building long-term projects for agents is knowledge retention: entities, concepts, and decisions are scattered across sessions. Next time work begins, either the model has to hard-recall them, or humans have to manually move notes. The claude-obsidian plugin in the Claude ecosystem validated the feasibility of “letting the agent maintain its own wiki” through a prompt-centric vault workflow, but dsh users cannot directly reuse it. The dsh-plugin-wiki-skills introduced below does exactly this port: it packages the knowledge-base skill set derived from claude-obsidian into a DeepSeek Harness plugin, enabling the dsh agent to have a complete pipeline for building wikis, ingesting data, querying and answering, health checking, and writing back insights.

What is this

dsh-plugin-wiki-skills is maintained by Lion-1209, current version 0.2.2, MIT license. The core consists of five skills: wiki, wiki-ingest, wiki-query, wiki-lint, save, plus four supporting skills: think, obsidian-markdown, obsidian-bases, defuddle. The vault design follows Andrej Karpathy’s described LLM Wiki pattern.

Engineer-wise, the plugin mounts as a read-only bundled provider to the ctx.skills registry, inserting a line wiki-skills via the dsh.bundle patch, making no invasive modifications.

Core Features

The five core skills are divided by responsibility:

  • wiki: Vault setup, scaffolding, cross-project referencing, hot-cache management;
  • wiki-ingest: Ingest source into vault, extract entities/concepts, build cross-references, record logs;
  • wiki-query: Answer based on vault, retrieval order is hot cache → index → pages, with citations;
  • wiki-lint: Health check, covering orphans, dead links, frontmatter gaps, stale index entries;
  • save: Write conversation insights back to vault.

Regarding supporting skills: think provides a 10 principles reasoning loop (OBSERVE…GROW), and all wiki skills map onto this loop; obsidian-markdown is a reference for Obsidian Flavored Markdown syntax (wikilinks, callouts, embeds); obsidian-bases is an Obsidian Bases reference, used for data views on note attributes; defuddle is responsible for cleaning scraped webpages before ingestion, outputting ad-free readable text.

Three implementation details are worth noting:

  1. Register a single provider named wiki-skills, candidate rank is bundled-rank (600);
  2. Each skill declares resourceBase: directory, the body can reference sibling files under references/;
  3. Report errors directly during discovery if frontmatter format is wrong, rather than silently skipping a skill.

Origin and Scope of Adaptation

The five skills are adapted from AgriciDaniel’s (AI Marketing Hub) claude-obsidian (MIT license, see ORIGINAL_LICENSE in the repo), based on its v1 prompt-centric version, adapted again to dsh by ZCode.

Modifications in this adaptation version: Removed upstream promotion footer (Community Footer) and route lines for unbundled sub-skills (autoresearch, canvas); Added dsh adaptation notes for each skill. Upstream autoresearch, canvas, wiki-cli, wiki-fold, wiki-mode, wiki-retrieve skills are not bundled; the research and canvas workflows remain upstream functions. Upstream product scripts (transport detection, wiki-mode.py, wiki-lock.sh, DragonScale address allocation, .vault-meta/) are also not bundled; the mechanical core is handled by native tools in the accompanying package dsh-plugin-wiki-tools.

Installation and Enablement

Runtime environment requires Node.js >=22.19.0 (engines constraint), dependency yaml ^2.8.0.

Installation is a single command:

dsh plugin --profile web add dsh-plugin-wiki-skills

This command installs the plugin into the web profile. The package declares a dsh.bundle patch; upon loading, it inserts a line wiki-skills into the ctx.skills registry, providing skills in the form of a read-only bundled provider.

Typical Usage

It is recommended to pair this with dsh-plugin-wiki-tools, which provides native wiki_query / wiki_write / wiki_lint tools, handling the mechanical core of these skills.

According to the routes of each skill, a typical pipeline is: first use wiki to build the vault and do scaffolding; sources like webpages go through defuddle cleaning first, then use wiki-ingest to extract entities and concepts and build cross-references; when an answer is needed, wiki-query retrieves it following the hot cache → index → pages order with citations; run wiki-lint periodically to check for issues like orphans and dead links; insights from the conversation are written back using save.

If participating in development, first install dependencies, then run tests:

npm install
node --test

Applicable Scenarios and Notes

Suitable for agent developers maintaining long-term projects on dsh, especially scenarios where knowledge is expected to be preserved as an Obsidian vault (the skills directly reference Obsidian Flavored Markdown and Bases). Note three points before using:

  1. bundled-rank (600) means skills with the same name for project, runtime, or user take priority over this suite; to override default behavior, one can write skills with the same name;
  2. the research and canvas workflows are not included in this plugin;
  3. the plugin runs with the current dsh process permissions; it is recommended to check the source code and license yourself before installing. Regarding the license, this package is MIT, and the packaged skill content is derived from claude-obsidian (also MIT), see ORIGINAL_LICENSE for details.

Conclusion

After the steps above, the dsh agent can possess a persistent knowledge base with clear sources, queryable, and health-checkable, and the entire port retains the upstream attribution. Project address and listing page:

  • GitHub: https://github.com/Lion-1209/dsh-plugin-wiki-skills
  • Community Directory: https://www.skillhub.cn/plugins/Lion-1209/dsh-plugin-wiki-skills

The community directory is an independent site and has no official affiliation with DeepSeek or 幻方.