@deepseek-ai/dsh-knowledge

中文 | English

Bridge the agent into the user's global knowledge base at D:\knowledge — the same Markdown + YAML frontmatter library the Codex kb.cmd CLI writes. Knowledge accumulated by either tool is visible to both: one library, two agents.

Tools

Tool	Behavior
kb_add	Write one typed note (lesson/pattern/preference/project/reference/concept/moc/decision/note), byte-compatible with the Python CLI: slugified unique filename, typed directory, YAML frontmatter (title/type/tags/aliases/status/created/updated + project/domain).
kb_search	Keyword scan over notes (title > tags > content), returns path/title/type + snippet.
kb_show	Locate a note by title/path/tag and return its full text.
kb_timeline	Append a timestamped line to notes/99_时间线.md — mandatory after every writeback ("失联等于白写").

The heavy read side (ask/reflect/map/ingest) stays on the Python CLI; this package owns the deterministic write + quick-read surface so no shell encoding ever touches your Chinese content.

Config

Field	Default	Meaning
kbRoot	D:\knowledge	Knowledge base root.
maxSearchResults	8	Search result cap.

Companion discipline: the knowledge-writeback skill (writeback three questions + retrieval protocol + causal-paradigm worldview).

Notes on compatibility

• slugify replicates kb/ingest.py: NFKC normalize, collapse chars outside [\w\u4e00-\u9fff], 60-char cap, untitled fallback.
• unique_path replicates stem.md → stem-2.md dedup.
• Frontmatter is emitted as quoted-YAML scalars and flow arrays, readable by python-frontmatter. If kb/frontmatter.py changes, update this package to match.

Install

Not on npm yet - install from this repository:

npm install github:ICCuse/dsh-knowledge
# or: pnpm add github:ICCuse/dsh-knowledge

Then mount the bundle (declared in package.json 'dsh.bundle') and point it at your knowledge base:

- id: knowledge
  name: 'dsh-knowledge'
  config:
    kbRoot: 'D:\knowledge'

Companion discipline: the 'knowledge-writeback' skill (writeback three questions + retrieval protocol).