session-index（会话索引）

Session index skill

You have access to a session index — a SQLite database with FTS5 full-text search across all Claude Code sessions. Use it whenever the user asks about past sessions, previous conversations, or wants to know what they've tried before.

The user's interface is this conversation. They ask naturally ("Didn't we discuss browser control recently?"), you translate to CLI commands, and present results conversationally. They should never need to know the CLI syntax.

How to handle session queries

1. Extract keywords from the question

The user says: "Didn't we discuss browser control recently?" You search: sessions "browser control"

The user says: "What approaches have I tried for form automation?" You search: sessions "form automation"

The user says: "How much time did I spend on Acme this week?" You run: sessions analytics --client "Acme" --week

2. Run the right command via Bash

Search — find sessions by topic:

sessions "relevant keywords"
sessions "relevant keywords" --context    # includes conversation excerpts

Context — read the actual conversation from a session:

sessions context <session_id> "search term"   # exchanges matching a term
sessions context <session_id>                 # all exchanges

Analytics — effort, time, tool usage:

sessions analytics                    # overall
sessions analytics --client "Acme"    # per client
sessions analytics --week             # this week
sessions analytics --month            # this month

Filter — find sessions by metadata:

sessions find --client "Acme"              # by client
sessions find --tool Task --week           # by tool + date
sessions find --project myapp              # by project
sessions recent 20                         # last N sessions

3. For synthesis ("what worked?", "what have I tried?")

This is the most valuable capability. When the user asks a question that spans multiple sessions:

1. Search: sessions "topic" -n 10
2. For the top 3-5 results, extract context: sessions context <id> "topic" -n 3
3. Spawn a Task with model="haiku" to synthesize:

- What approaches were tried? - What worked / what failed? - Recurring patterns? - Current state?

1. Present the synthesis conversationally with claude --resume <id> links for each source session

This uses an in-session Haiku subagent — no external API key needed.

4. Present results conversationally

Don't dump raw CLI output. Summarize:

• "You discussed browser control in 3 sessions last week..."
• "The main approach that worked was..."
• Include claude --resume <session_id> links so they can jump back in
• If context is relevant, quote key exchanges

Installation

pip install claude-session-index

First run of any command auto-indexes all existing sessions.

Data location

• Database: ~/.session-index/sessions.db
• Topics: ~/.claude/session-topics/
• Config: ~/.session-index/config.json (optional)