dsh-session-keyword-search

Keyword search across every DeepSeek Harness session — the kind of search you want when you remember a phrase but not the conversation.

It takes over the sidebar search box instead of adding a second entry point: type 大麦 x-sign and the existing search lists the sessions that talked about it.

Why this exists

The harness ships an SQLite FTS5 search behind ctx.sessionQuery.searchSessions. Its tokenizer is unicode61, so one contiguous CJK run is a single token: a document containing 大麦协议劫持已在真机验证通过 does not match 大麦, and 会话搜索 does not match 搜索. This plugin replaces those two search methods with a substring index, so Chinese, English and mixed queries all behave the same.

What it does

• Substring recall. LIKE over a lowercased text column: no tokenizer, no phrase quoting rules to remember, no minimum token length.
• Session-level AND. Whitespace separates terms; every term must appear somewhere in the session, but not necessarily in the same message. A quoted "phrase" must appear verbatim inside one message.
• Ranked sessions. More matching messages first, then recent activity, with the current workspace winning ties.
• Provenance decides, not guessing. Every user/message carries data.source.kind: user is a human turn, while dsh-mnemon, runtime-context, time-context, skill-catalog and tool-jobs are injections. Only kinds listed in userSourceKinds (default ["user"]) are indexed, and assistant/message contributes text blocks only — reasoning is never searchable.
• Paragraph units. A message is split at blank lines (and long paragraphs at line boundaries) before indexing, so a snippet points at one paragraph rather than a whole message.
• Fast warm path. The first search builds a derived SQLite index (one table of message rows); later searches only re-read sessions whose revision changed.
• Marks in the result list. Query terms are wrapped in <mark> inside the sidebar rows.
• Opens at the match. After a result is opened, the transcript scrolls to the matched message and flashes it; older pages are pulled first, up to revealAttempts times, and it quietly stays put when the text cannot be found.

Install

# from this repository
node scripts/install.mjs --tgz ./dsh-session-keyword-search-0.1.0.tgz
# then restart DeepSeek Harness

The script backs up <profile>/package.json, adds the dependency and the dsh.profile.bundles entry, and runs the package manager. --dry-run prints the planned manifest change. Nothing under the application directory is touched, so a harness update cannot eat the plugin.

Uninstall:

node scripts/uninstall.mjs            # keep the index
node scripts/uninstall.mjs --drop-index

Configuration

Every field is optional; the bundle's cordis.patch.yml carries the same defaults.

Field	Default	Meaning
path	<dataHome>/storages/session-keyword-search/index.db	Derived SQLite index.
dataHome	$DSH_HOME, else ~/.dsh	Base directory for the default path.
takeover	auto	off leaves the host's own search untouched (and never builds an index).
includeToolEvents	false	Also index tool/call and tool/result text.
injectedPrefixes	built-in list	Leading text that marks a machine-injected user message.
injectedPatterns	[]	Extra regular expressions (source strings) treated the same way.
excludeSessionIds	[]	Sessions that must never enter the index.
snippetChars	240	Snippet length in code points.
maxTokens	8	Query token budget.
refreshDebounceMs	5000	Minimum gap between two corpus reconciliations.
uiHighlight	true	Mark query terms inside sidebar result rows (browser enhancement).
uiReveal	true	Scroll to the matched message after a result is opened.
revealAttempts	20	How many older-page pulls the reveal may attempt.
userSourceKinds	["user"]	source.kind allow-list for human turns; injections stay out of the index.

Override the row by id in the profile's own patch:

- id: session-keyword-search
  config:
    takeover: off

How it stays compatible

• Corpus access uses public services only: ctx.sessionQuery.listSessions(), ctx.sessionQuery.readSession(id), and the optional ctx.sessionPersistence.stat(id) revision. The plugin never learns where session logs live, never decodes the zstd log format, and never parses paths.
• It augments the live sessionQuery instance (ctx.effect restores the original methods on dispose) instead of replacing the service row, so an incompatible harness keeps its own behaviour instead of failing to boot.
• Every search degrades through three stages: our index → the original provider → session-title/workspace substring, and records the failure for the settings panel.

Settings panel

Settings → Session keyword search shows the index path, session/message counts, indexed text size, the last sync counters, the last error, and offers Rebuild / Delete index / Refresh. It talks to the host over the /dsh-session-keyword-search RPC channel.

Channel registration is recorded in rpc-registration.json next to the index. If the panel reports HTTP 405, read that file first: it states whether connection and webServer were visible, whether the route registered, and why it failed.

The panel is not required for maintenance:

node scripts/status.mjs            # index state plus the registration record
node scripts/status.mjs --rebuild  # rebuild from the on-disk sessions
node scripts/status.mjs --clear    # drop every indexed row

Verification

scripts/verify.mjs is a standalone harness: it feeds the indexer from the on-disk store and prints the hits for a set of queries, so recall can be checked without a running harness.

node scripts/verify.mjs --rebuild --queries "槟榔 价格" "沉静式翻译 气泡" "大麦 x-sign"

Measured on a real profile (106 sessions, 48.5k events): 2,114 indexable messages, 1.25 MB of text, full cold build ≈ 2.5 s, warm queries 16–60 ms.

Inside a live harness, one boot-time probe reports what the plugin sees through the real services and writes it to %TEMP%/dsh-session-keyword-search-selftest.json:

DSH_SESSION_KEYWORD_SEARCH_SELFTEST="槟榔 价格" dsh <profile> ...

Known limits

• A session's last turn or two may not be on disk yet, so a just-written message can be missing until the harness flushes.
• Sessions written before data.source existed are treated as human turns (an empty userSourceKinds list disables the allow-list).
• Highlight and reveal are heuristics over DOM contracts ([role="treeitem"], [data-chat-node-key], [data-conversation-scroll]): losing them never affects search, and a message rewritten substantially by markdown may not match its stored prefix.
• A session log the harness itself rejects as corrupt (for example step/start does not match the open turn) is skipped on its own without affecting the rest; the count shows in the settings panel as "Unreadable sessions skipped", and the same revision is not retried.
• Nothing is encrypted: the index holds message text in plain SQLite. Delete it from the settings panel or with --drop-index.
• Search results come from the harness's own result row component; the plugin supplies session id plus snippet, exactly like the stock provider.

License

MIT.

Publishing to the plugin market

DSH ships no central submission queue: the market is a reader for index JSON documents the user configures (plugin-market.indexes). This repository therefore carries a format-1 index:

• index: market/index.json
• install source: git, pinned to the release commit (marketInstallTarget produces git+https://…#<sha>)
• mirror for networks where raw.githubusercontent.com is unreachable: https://cdn.jsdelivr.net/gh/JimmyRowe13/dsh-session-keyword-search@master/market/index.json

Enable @deepseek-ai/dsh-experimental-plugin-market-bundle, add the URL as a source, and the plugin shows up under the market's discover tab.

Release checklist: bump version, commit, git tag vX.Y.Z && git push --tags, then point market/index.json at the new commit (versions[0].commit, version, publishTimestamp).