Search Chat Messages

Search the CONTENT of the caller's own conversations (TOS-2330).

GET /api/chats?q= only substring-matches chat titles and participant display names, so a word the user remembers from inside a conversation finds nothing there no matter how far they page. This exposes the corpus search that already backs the agent-facing conversation_search MCP tool (TOS-477) to the user's own UI.

Scope is agent = person = caller, exactly as conversation_search's self-recall path does (TOS-844 P1): _scope_chat_ids then resolves to chats where the CALLER is an active participant. Nothing outside the caller's own participated chats is reachable and no new privacy SQL is introduced. include_excluded=True keeps extraction_excluded messages, which govern Mind-extraction subject scope rather than the user's own visibility - without it the user could not search their own captured Claude Code / Codex sessions.

Lexical only (TOS-3138). The semantic leg is deliberately off for this surface. Measured on prod over a 16,885-chat corpus, it returns a full page of unrelated segments for a query that matches nothing, at cosine scores ABOVE those of genuine matches for real queries - so no threshold separates them and the "No message matches" empty state was unreachable. The lexical legs fired 4x for each real query and 0-1x for each nonsense one. A search box wants that precision; agent recall keeps the full hybrid, where a loose hit is cheaper than a miss. Consequence: no query embedding is computed, so this endpoint makes no AI call at all.