ListenJ/axiom-kb-dsh

Axiom Knowledge Base (Vault memory + knowledge graph) as a DeepSeek Harness (dsh) plugin.

The plugin bundles a Vault-memory and knowledge-graph backend and exposes memory, graph and unified-query tools to dsh under the kb__ prefix.

• License: MIT
• Hot-pluggable: dsh plugin add/rm
• Runtime: requires Bun

Install

No extra setup — the plugin carries its own backend (Vault memory + knowledge graph + MCP server):

dsh plugin --profile web add github:ListenJ/axiom-kb-dsh

Restart dsh (dsh web). The tool list then includes the kb__* tools and the kb_plugin_status diagnostic tool.

Uninstall

dsh plugin --profile web rm axiom-kb-dsh

Architecture

The plugin bundles a knowledge-base backend (backend/server.js, a single-file Bun build) and spawns it over stdio:

dsh (Node) ── axiom-kb-dsh ──stdio──▶ built-in backend (Bun) ──▶ Vault memory + knowledge graph
                 │                            │
          filter + register          deterministic search / SQLite

• A writable data/ directory is created automatically: Vault notes (axiom-memory/) and knowledge-graph SQLite (data/kg.db).
• Web-search tools (web_fetch / web_search) are not part of this plugin.

kb_plugin_status (always available)

A diagnostic tool reporting MCP bridge state (connected, tool count, server name) and the effective config summary.

Configuration

Overridden via cordis.patch.yml under line id kb (overriding the whole section replaces all keys).

Key	Default	Description
dataDir	<plugin>/data	Backend data directory (Vault notes + KG SQLite). Created automatically.
mcpEnabled	true	Launch and bridge the built-in backend.
mcpCommand / mcpArgs	bun / <plugin>/backend/server.js --stdio	Backend launch command.
mcpServerName	kb	Public tool prefix (<serverName>__<tool>).
mcpToolCallTimeoutMs	60000	Per-tool call timeout (ms).
mcpFailOnStartupError	false	false = tolerate startup failure (warn only); true = fail the fiber on initial connect error.
toolFilter	[]	Empty = built-in KB allow-list; explicit array fully replaces it (prefix ending _ or exact name).

Tools (prefix kb__)

• Vault memory: memory_search memory_read memory_write memory_atomic memory_browse memory_network memory_stats code_index
• Knowledge graph: kg_stats kg_entities kg_entity_detail kg_traverse kg_build kg_search kg_graph kg_add_node kg_add_edge kg_search_nodes kg_subgraph kg_shortest_path kg_detect_communities kg_echarts_data kg_d3_data kg_nl_query kg_enhanced_stats
• Unified query: kal_query kal_references
• Document pipeline: dip_ingest_document dip_query_ast
• Diagnostic: kb_plugin_status

Quick start

After install and restart, the LLM can call kb__* tools directly. Write and retrieve a memory note:

kb__memory_write(path="00-Meta/example.md", content="# Example", title="Example", overwrite=true)
kb__memory_search(query="Example", limit=5)

Ingest a Markdown document into the knowledge graph:

kb__dip_ingest_document(markdown="...", title="Doc title")
kb__kg_search_nodes(query="function name")

License

MIT