dsh-session-fulltext-search

Enable full-text search over DSH session history in the Web sidebar — and make it actually find Chinese text. 中文说明

The DeepSeek Harness web UI ships with a content-search backend (@deepseek-ai/dsh-session-query-sqlite, SQLite FTS5) that is disabled by default (openAt: never): the sidebar magnifier then matches only session titles and workspace names. This plugin is the deployment override the backend's documentation prescribes — openAt: first-search with a persistent derived index under $DSH_HOME.

That alone has a catch for CJK users: FTS5's unicode61 tokenizer treats a run of continuous Chinese text as one token, so a query only matches when it aligns with token boundaries — searching 福州大学 does not hit 已保研至福州大学. This plugin therefore also mounts a substring fallback: queries containing CJK are answered with a literal substring scan over the same derived index, so any Chinese phrase inside message bodies is found, exactly as you typed it. Non-CJK queries use the tokenizer unchanged.

What you get

• The sidebar magnifier searches message bodies (user messages, assistant replies) across all workspaces, with a content snippet per hit.
• Chinese searches are substring-based — no whole-token boundary problem.
• The index updates incrementally as sessions grow.

Install

Into a profile (e.g. desktop):

dsh plugin --profile desktop add github:amd47802574/dsh-session-fulltext-search

dsh plugin add also registers the package in dsh.profile.bundles (the patch-layer stack). Restart DSH afterwards. Manual equivalent:

// <profile>/package.json
"dependencies": {
  "dsh-session-fulltext-search": "github:amd47802574/dsh-session-fulltext-search"
},
"dsh": { "profile": { "bundles": ["…existing layers…", "dsh-session-fulltext-search"] } }

then pnpm install inside the profile directory and restart DSH.

Uninstall

dsh plugin --profile desktop remove dsh-session-fulltext-search

After restarting DSH the sidebar falls back to title-only search. The derived index file ($DSH_HOME/session-query-index.sqlite3) can be deleted.

Notes

• CJK queries take the substring scan; other queries keep the tokenizer's word/phrase matching (so pytho still does not match python — by design, that is how FTS5 works). Wrap a query in quotes ("two words") for exact-phrase search; FTS5 operators (OR, *) are treated as plain text.
• The fallback reads the derived index with its own read-only SQLite connection; the owning engine keeps writing (WAL). A missing or rebuilding index simply returns fewer results for that one query.
• This layer only overrides the session-query-sqlite row's config (path, openAt) and adds its own wrapper row. Your own profile cordis.patch.yml sits later in the stack, so a manual override there still wins over this plugin.
• The index is derived data — delete the file to force a rebuild. Never point path at the session persistence database.

License

MIT