sessiondb

Full-text search over past Claude Code sessions stored as JSONL in ~/.claude/projects/.

When to invoke

• User wants to find a past session by topic or keyword ("あの話どこでしたっけ", "find the session about X").
• User asks to (re)build or update the session index.
• User asks aggregate stats about session history that would be slow with grep.

Do not invoke for lookups inside the current session — use Read/Grep.

Database location

~/.claude/sessiondb/sessions.db (override with env var SESSIONDB_PATH).

CLI

sessiondb index                       # incremental rebuild
sessiondb index --full                # drop and rebuild from scratch
sessiondb search "<query>" --format=jsonl [--project P] [--role R] [--limit N]
sessiondb search                      # query 省略 = セッション一覧 (新しい順)
sessiondb stats                       # content totals (sum of source JSONL bytes)
sessiondb size [--format=json]        # on-disk DB usage (.db + WAL/SHM + page stats)
sessiondb show <session_id> [--line L] [--context K]

重要: エージェントが呼ぶときは必ず --format=jsonl を付ける。CLI の 既定は --format=table (人間向け TSV) なので、明示しないと下記の JSON Lines 形式は得られない。

If bin/sessiondb is not installed, fall back to invoking sqlite3 directly:

SELECT m.session_id,
       snippet(messages_fts, 0, '<b>', '</b>', '…', 32) AS snippet,
       bm25(messages_fts) AS score,
       m.ts, m.role, m.source_path, m.line_no
  FROM messages_fts JOIN messages m ON m.id = messages_fts.rowid
  WHERE messages_fts MATCH :query
  ORDER BY score LIMIT 20;

Response format (search)

JSON Lines, one hit per line, followed by a summary line:

{"session_id":"…","project":"…","ts":"…","role":"…","snippet":"…","score":-4.21,"source_path":"…","line_no":142,"resume":"claude --resume …"}
{"_summary":{"hits":17,"shown":10,"query":"…","ms":34}}

--format=table (CLI 既定) is TSV for human reading; do not parse from agent.

Build / schema / ingest details

See how-to-create-db.md. Read it only when you need to implement, modify, or debug the indexer; not on every search.