dsh-plugin-knowledge-graph

A DeepSeek Harness (DSH) plugin that exposes a read_graph tool backed by a codebase knowledge graph. The graph tracks every file, directory, exported symbol, and their relationships so an agent can answer structural and dependency questions faster than grepping or reading files.

Status: infancy. This project is in its infancy stage. Execution behavior, the query surface, and compatibility may change while the project is iterated on and tested by the open source community.

Derived work. This project is derived from the knowledge graph tool in Luke-Yong/H, adapted to DeepSeek Harness's "everything is a plugin" model.

What it does

Given a workspace root, the plugin builds an in-memory graph and answers queries such as:

• structure — print the full directory tree
• exports <file> — list every symbol exported by a file
• imports_of <file> — list symbols a file imports (with their source files; stdlib/third-party imports are tagged)
• exporters_of <symbol> — find which files export a symbol
• dependents <file> — find which files import from a file
• unused_imports <file> — list imports never referenced in the file

Symbols are parsed from TypeScript/JavaScript (via the TypeScript compiler API) and Python (module-level def, class, and name = value bindings).

Graph model

The graph uses three node types and four edge types:

Nodes	Meaning
dir	A directory
file	A source file
symbol	An exported symbol (function, class, const, type, interface, enum, default)

Edges	Meaning
CONTAINS	Structural parent → child (dir → dir/file, file → symbol)
EXPORTS	File → symbol it exports
IMPORTS	File-level dependency (file → file)
IMPORTS_SYMBOL	Precise symbol-level dependency (file → symbol)

Installation

Install with the DSH CLI into a profile:

dsh plugin --profile web add github:Luke-Yong/dsh-plugin-knowledge-graph

Pin to a specific commit for reproducibility:

dsh plugin --profile web add github:Luke-Yong/dsh-plugin-knowledge-graph#<commit-sha>

Usage

The plugin registers a single tool, read_graph, on ctx.tools. Its parameters:

Parameter	Type	Required	Description
query	string	yes	Query string in <query_type> <target> form
root	string	no	Absolute workspace root to graph. Defaults to process.cwd()

Root hygiene: the root default (process.cwd()) may be a parent directory containing several independent projects. Always pass an explicit root pointing at the single workspace you are working in, so structure doesn't descend into sibling trees.

Example queries:

exports src/index.ts
exporters_of buildKnowledgeGraph
imports_of src/query.ts
dependents src/knowledgeGraph.ts
unused_imports app.py
structure

Project structure

src/
  index.ts         Plugin entry — registers the read_graph tool
  knowledgeGraph.ts  Graph builder (filesystem walk, import/export parsing, serialization)
  query.ts         Query engine (runGraphQuery)

Building

npm run build   # tsc -p tsconfig.json

Output is emitted to dist/ (dist/index.js is the package entry point).

License

MIT