dsh-session-grep

Substring search across every message in a DeepSeek Harness workspace — the kind of search where 搜索输出 finds the sentence …只能搜索输出, and 插件 finds words inside longer words.

Why this exists

The built-in session search (sidebar Ctrl+K) runs on an SQLite FTS5 index created with the unicode61 tokenizer. That tokenizer matches whole tokens and phrases, not arbitrary substrings:

• Contiguous Chinese has no spaces, so an entire clause is one token. 搜索输出 is a fragment of the token 只能搜索输出 and therefore matches nothing.
• The upstream documentation says it plainly: "unicode61 tokenizer matches tokens and phrases rather than arbitrary substrings" (@deepseek-ai/dsh-session-query-sqlite, README line 63).

This plugin keeps its own literal index instead: it reads the session logs, decompresses them and does a case-insensitive indexOf. Any fragment matches, Chinese included.

What it looks like

• A 搜索 / Search pill in the conversation header (the same slot the session title sits next to).
• Click it, type, and get one row per matching message: role, title, session suffix, snippet with the match highlighted.
• Results rotate across sessions, so one very long conversation cannot crowd out the others.
• ↑ ↓ to move, Enter to open, Esc to close.
• Picking a row opens that conversation and flashes the matched text for ~2.6 s. A hit in the conversation already on screen only scrolls and highlights — no pointless navigation.
• Toggle 工具输出 / Tool output to also search tool calls, tool results and reasoning.
• Results are scoped to the current workspace, 40 rows max.

Install

From npm:

dsh plugin --profile <profile> add dsh-session-grep

From a clone:

git clone https://github.com/chenshanhan0711/dsh-session-grep
dsh plugin --profile <profile> add ./dsh-session-grep

Or in the app: Plugins → Add plugin and point it at that folder. The bundle adds one Host service and one client slot; the pill appears after a restart.

Remove it the same way you removed any plugin; it never touches the DSH install directory.

How it works

Host (index.js)

• Finds logs at $DSH_HOME/sessions/<cwd-key>/<session-id>/session.v4.jsonl.zstd.
• Each file is a sequence of concatenated zstd frames with no length prefix: split on the zstd magic, inflate each frame, and merge a slice into the next candidate when inflation fails (so a stray magic inside compressed bytes costs a retry, not an event).
• Parses the JSONL, extracts text per event type (user/message, assistant/message, optionally tool/call / tool/result / reasoning), and caches parses by mtime + size (max 64 sessions).
• One hit per event with an "N matches" count; each session collects up to max(10, limit/2) first, then the sessions are merged round-robin by their newest match — a long conversation cannot swallow the list.
• Serves three read-only routes under /dsh-session-grep/api: search, sessions, health.
• Scanning a 900-event / 3 MB session takes ~20 ms; a query over it is single-digit milliseconds.

Client (client.js)

• Registers into the conversation.header.leading slot.
• 200 ms debounce, one in-flight request at a time (the previous one is aborted).
• Results panel is position: fixed from the input rect, so it is never clipped by the header.
• Clicking a result calls uiWorkspace.openSession(...); for a side thread it first retains the session and opens through the durable subagent.address from its snapshot. If the view has not switched after 1 s it retries once, and only then reports failure.
• Then it walks the rendered text nodes to wrap the match in a <mark>, scrolling it into view; if the message is not rendered yet it retries for ~9 s and then says so.
• Selection runs on mousedown, not click, so a blur cannot unmount the row before the handler runs.

Limits

• Matches text, not rendered Markdown: a match split across **bold** boundaries may be found by the scanner but not highlighted on screen.
• Tool output, reasoning and JSON payloads are off by default; switch on Tool output to include them.
• Only sessions of the current workspace directory are scanned.
• Live sessions are re-parsed when the file changes; there is no incremental tailing yet.
• Very large histories are scanned fully per query (no persistent on-disk index), which is why results cap at 40.
• Changing the Host half needs an app restart; changing only client.js needs a page refresh (the Host hot-loads client bundles).
• Archived sessions still match; the app refuses to open them.

Privacy

Zero network requests. The Host reads local session files and answers its own loopback route; the client only calls that route. No telemetry, no remote endpoints, no hardcoded credentials.

License

MIT