6fc564278d
Adds a core sessions_search agent tool backed by a SQLite FTS5 index that lives next to the transcript rows in the per-agent database. Indexing is transactional: user/assistant text is indexed inside the same write transaction that persists the event, deletes drop index rows in the same transaction, and branch rewinds mark the session dirty so the next search rebuilds it from the same visible-path resolution sessions_history uses (which also lazily backfills doctor-migrated databases). Search executes gateway-side behind an additive sessions.search method with a bounded session-key allowlist; visibility, sandbox restrictions, and snippet redaction mirror sessions_history. Closes #100978
2.1 KiB
summary, title, read_when
| summary | title | read_when | ||
|---|---|---|---|---|
| Search past session transcripts and reopen the matching context | Session search |
|
Session search
sessions_search searches the user and assistant text in your own past sessions. Each result
includes a sessionKey, timestamp, role, and a short matching excerpt. Pass the returned
sessionKey to sessions_history when you need the surrounding conversation.
Visibility and output
Search uses the same session visibility rules as sessions_history. Results outside the caller's
visible session tree are removed before result limits are applied. Sandboxed agents remain limited
to sessions they spawned when spawned-session visibility is enabled.
Excerpts are redacted before they return to the model. Results are also bounded by count, excerpt length, and total response size.
Index lifecycle
OpenClaw stores a full-text index next to the transcript rows in each agent's SQLite database. New user and assistant messages are indexed in the same transaction that persists them, so the index never lags live conversations; tool results, reasoning blocks, and images are excluded. Only the transcript's active branch is searchable.
Transcripts that predate the index (for example, sessions imported by openclaw doctor) and
sessions whose active branch was rewound are reindexed by a background reconciliation that starts
with the next search. A response with indexing: true can therefore be incomplete; retry after
indexing finishes. Deleting a session removes its index entries in the same transaction.
Search currently uses SQLite's Unicode word tokenizer with diacritic removal. Trigram tokenization for CJK substring matching is a future improvement.
Session search vs. memory search
Use sessions_search for exact words or phrases from raw session transcripts. Use
memory_search for durable memory files and semantic recall. The
experimental session-memory corpus is the semantic complement to this exact transcript search.