dsh-title-index

When the dsh Web @-mention candidate list needs to display the latest title for each persisted session, the stock engine decompresses and replays the full session.jsonl.zstd row-by-row. Once a corpus reaches dozens or hundreds of sessions, every keystroke rescans all logs while the built-in LRU holds only five entries. On a real 114-session corpus the top-50 title lookup took roughly 10 seconds, making the @ menu effectively unusable.

dsh-title-index patches the stock session-query-sqlite row with a subclass that overrides only readTitleSnapshots. A disk index keyed by (sessionId, mtimeMs, size) caches folded title snapshots: a hit returns in milliseconds, a miss (new session or appended log) falls through to the stock fold and writes the result back. Live sessions, stock read failures, and a corrupted index all fall back to the original path unchanged. Every other method (listSessions, search, lineage, auth semantics) remains the stock implementation.

Best suited for dsh Web users who maintain a large number of persisted sessions (tens to hundreds) and rely on @-references in daily workflow. Setup is minimal: add one dependency to the profile and declare the bundle after the stock web-app entry.