feat(core): add search({ excludeSession }) and derive missing session titles

[Cyb3rN8]

Fixes the two behavioural items from #16. Documentation-only items from that issue are not included here.

search(text, { excludeSession })

Searching history from inside a session reliably hits that session's own prompt, because the query terms came from it. Filtering afterwards in the query script does not recover the loss: limit is applied in SQL, so the self-hits have already displaced evidence from other sessions. excludeSession accepts one session ID or an array and filters before limit.

Derived session titles

overview() and sessions() are advertised as a navigation map, but on a current index most rows come back with title: null — 792 of 821 Claude sessions on the index I measured. Both title sources are usually absent now: ai-title transcript rows have become rare (most recent one in a 264-transcript project directory: 2026-07-21), and history.jsonl entries no longer carry a title field at all, so the historyTitles path in providers/claude.ts cannot populate. Neither is an indexing failure — every title that exists is picked up.

overview(), sessions(), and search() now fall back to the first 80 characters of the opening user message, skipping <command-…> envelopes so the label describes the task instead of naming the slash command. A session with no user text keeps title: null rather than getting an invented one.

Implemented as a SQL expression in the query layer (titleExpr()), for two reasons: it covers sessions indexed before this change without a rebuild, and it leaves sessions.title meaning "a title the transcript actually provided". The tradeoff is that a derived title is not distinguishable from a stored one in the returned row — the docs call it a navigation label. Happy to switch to a separate field, or to compute it at index time, if you would rather have the provenance explicit.

Notes

• skill-doc changes are limited to what documents the new behaviour: the search() opts list, an Exclude Yourself entry in the Retrieval Contract, the sessions() note about derived titles, and the api-reference.md rows. The broader documentation suggestions stay in First-pass retrieval wastes most of its budget: untitled sessions, self-hits, and known-empty memory recall #16.
• Tests: two added to tests/query.test.mjs (SQL-level exclusion including the limit: 1 case, and the title fallback including the command-envelope skip and the null case). On this machine npm test goes from 207 passing to 209 with an unchanged set of 12 failures — all in the app/session suites, which need Electron deps that are not installed here.
• Branched from current main, so this may touch api-reference.md near feat(core): index failed tool results for retrieval #15. Say the word if you want it rebased onto either of the open branches instead.