dsh-llm-wiki: Let DeepSeek Harness Agents Directly Manage an LLM-Wiki Personal Knowledge Base

dsh-llm-wiki is a DeepSeek Harness (DSH) plugin (DSH plugin) maintained by detpecca, licensed under MIT, last updated in 2026-08. It acts as a thin adapter layer, enabling agents to directly manage an LLM-Wiki personal knowledge base, providing 7 tools: search, read, stats, validate, fix, errorbook, and ingest. It solves the problem of lacking structured interaction between agents and knowledge bases. The plugin itself has zero runtime dependencies (pure ESM, no build step), and invokes the DSH-Wiki engine CLI via DSH's subprocess service with the --json channel. The engine remains the single authority for structured signal retrieval (weighted scoring of page names/aliases/tags/summaries) and the paper's Algorithm 1 compilation flow.

What core tools does the dsh-llm-wiki DSH plugin support?

The plugin exposes 7 tools: wiki_search (structured signal scoring retrieval with CJK tokenization), wiki_read (batch page reading/directory index, follows [[wikilink]]), wiki_stats (page/category/digest/errorbook statistics), wiki_validate (4 types of deterministic structural validation), wiki_fix (deterministic fixes, with finalize:true adding LLM repair rounds), wiki_errorbook (view Error Book self-correction records), and wiki_ingest (compile source text into the wiki, requires LLM key). In wiki_search, score is an integer weight accumulation (page name 8 > alias 6 > tag 4 > summary 2 > body 1), higher is stronger, not similarity. wiki_validate checks dangling links, incomplete pages, malformed source references, and index inconsistencies; on errors the engine exits 1, but the plugin treats it as a result, not a failure.

How to install and configure dsh-llm-wiki?

Prerequisites: DeepSeek Harness installed (dsh command available), Python ≥ 3.10, optional OpenAI-compatible LLM API key (only needed for wiki_ingest and wiki_fix finalize). First install the Python engine DSH-Wiki, then install the plugin, then configure the knowledge base path. Installation commands:

pip install git+https://github.com/detpecca/DSH-Wiki.git
# only runtime dependency is pyyaml

dsh plugin --profile web add github:detpecca/dsh-llm-wiki

After installation, the default points to ./wiki. Write the actual path into the profile's cordis.patch.yml ($DSH_HOME/profiles/<name>/cordis.patch.yml, $DSH_HOME defaults to ~/.dsh), overriding by id and restating all keys (patch is line-based replacement, not deep merge). Optional LLM config (llmWikiBaseUrl, llmWikiApiKey, llmWikiModel) can be set; explicit config takes precedence over environment variables. Restart DSH for the agent to use all tools. Windows users can run scripts/install.ps1 for one-click setup (automatically creates venv with uv, installs engine, runs dsh plugin add, writes config).

How to initialize the knowledge base for dsh-llm-wiki?

If the knowledge base is empty, you can have the agent directly call wiki_ingest to compile your first note, or use the engine CLI to compile:

python -m llm_wiki --wiki ./wiki ingest my_notes.txt

Note that --wiki must be placed before the subcommand. See the DSH-Wiki README for more CLI usage.

How to uninstall dsh-llm-wiki?

Uninstall command:

dsh plugin --profile web remove @detpecca/dsh-llm-wiki

What DeepSeek Harness versions and platforms does dsh-llm-wiki support?

The plugin requires Node ≥ 18 and Python ≥ 3.10 (for the engine), supports Windows (PowerShell install script) and Unix-like systems. It interacts with the engine via DSH's subprocess service, does not depend on a specific DSH version, but requires the dsh command to be available.

View more DSH plugins at dsh-plugin.org