dsh-obsidian-agent-wiki
SQLite-indexed Obsidian memory for DeepSeek Harness — gives the agent a searchable view of a local Obsidian vault, with no frontend (browse and edit notes in Obsidian itself).
Built on Node 22's built-in node:sqlite — zero runtime dependencies.
Why
Plain file-system memory plugins scan every file linearly and can't filter by directory. This plugin indexes the vault into an in-memory SQLite table at startup and keeps it fresh on every write, so the agent can:
- full-text search any substring (including 2-character Chinese words),
- filter by
path / category / project / tag,
- get results ranked by match count.
Install
dsh plugin add dsh-obsidian-agent-wiki
# or from source:
dsh plugin add CagierAsh123/dsh-obsidian-agent-wiki
Configure
Point the plugin at your vault in your profile's cordis.patch.yml:
- id: dsh-obsidian-agent-wiki
name: dsh-obsidian-agent-wiki
config:
vaultPath: B:/path/to/your/vault
Or set the OBSIDIAN_VAULT_PATH env var. Restart dsh.
Tools
| Tool | Parameters | Purpose |
|---|
wiki_read | file_path | Read a note |
wiki_list | path? | List a directory |
wiki_write | file_path, content | Write / overwrite, then re-index |
wiki_append | file_path, content | Append, then re-index |
wiki_search | query, path?, category?, project?, tag?, limit? | Indexed full-text + structured search |
wiki_search returns path, title, category, project, matches, and a snippet for each hit, ordered by match count.
Vault layout
category and project are inferred from each note's path:
代码Wiki/<topic>/<note>.md → category 代码Wiki (shared technical notes)项目Wiki/<project>/<note>.md → category 项目Wiki, project <project>
tag filtering matches the frontmatter tags: value.
How indexing works
- At startup the plugin scans the vault (skipping
.obsidian/.git/node_modules/.trash) and builds the in-memory index.
- On
wiki_write / wiki_append the touched note is re-indexed immediately.
- Manual edits made in Obsidian are picked up on the next dsh restart.
Full-text matching uses LIKE '%query%' rather than FTS5 trigram because trigram returns zero hits for queries shorter than 3 characters — which would drop the 2-character Chinese words that dominate a Chinese vault. At tens-to-hundreds of notes, a LIKE scan is sub-millisecond.
Requirements
- Node ≥ 22.5 (for
node:sqlite)
- dsh ≥ 0.1.1-rc.2
License
MIT
