search

name	search
description	Search your past Claude Code conversations. Use when the user asks to find, search, look up, or recall something from past sessions, transcripts, chats, or history.
argument-hint	["query"]
allowed-tools	Bash(node *)

Search past Claude Code conversations

Search the user's local Claude Code transcripts (everything under ~/.claude/projects/**) and return ranked, resumable sessions. Indexing is incremental — only changed files are re-read.

When the user gives a query

Run the search CLI with the user's query. Pass everything they typed as one argument.

node "${CLAUDE_PLUGIN_ROOT}/scripts/search.mjs" $ARGUMENTS

The CLI prints ranked results with:

• Date and project for each match
• A highlighted snippet
• A claude --resume <id> command to resume that session

Paste the CLI output into the chat reply, verbatim, inside a fenced code block. Raw shell output is hidden from the user — only Claude's text response shows. Without echoing the output, the ranked list, snippets, and resume commands are invisible. ANSI color codes are auto-stripped when the CLI runs under the bash tool, so the captured stdout is already clean text.

A short one-line lead-in before the code block is fine (e.g. "Top results for <query>:"). After the code block, a brief 1–3 line summary that groups or highlights notable matches is welcome — but never replace the echoed results with a summary alone.

When the user wants to browse in a web UI

If the user asks to "open", "browse", "show me a UI", or otherwise wants a visual interface, pass --web:

node "${CLAUDE_PLUGIN_ROOT}/scripts/search.mjs" --web $ARGUMENTS

This starts a local server on 127.0.0.1:<port> and prints the URL. Echo the URL in the chat reply and tell the user to open it in their browser. The server keeps running until they stop it with Ctrl+C.

When the user gives no query

Tell them what to do: /claude-search:search <words to find>. Mention they can also add --web to browse in a browser, or --project <name> to filter by project.

Notes

• All searching is local. Nothing leaves the machine.
• The index lives at ~/.claude-search/ and is refreshed automatically.
• If a session shows up but resume fails, the transcript was likely deleted — tell the user to run the search again to refresh the index.