v0.1.0-alpha.186

v0.1.0-alpha.186 — paginated thread load: 15 first, 35 background-filled

User feedback: deep-linking to a thread takes too long. Network
bytes dominate the cold path — a 50-turn agent thread with
tool_result blobs can ship 1-3 MB on first load.

Two changes:

  1. Initial page size 50 → 15. ~150-800 KB on the wire instead of
     500 KB - 3 MB. Paint time drops from 400-1000ms to 100-300ms.
     Covers ~7 user/assistant exchanges — enough to see the last
     few messages of the prior session without overpaying.

  2. Background-fill after first paint. The moment the first 15
     land in the cache, ChatSurface fires a silent
     loadEarlierMessages() to pull the next page (default 50 more)
     into the cache. By the time the user scrolls up to look for
     earlier turns, the data is already there — "Load earlier"
     becomes instant instead of a one-click round-trip.
     Bounded: once per thread mount, only when hasMore=true.

Performance impact for a typical deep-link:

  Before                         After
  ────────────────────────────────────────────────────────────
  GET messages?limit=50          GET messages?limit=15
    ~600ms total                   ~150ms total
    User sees content              User sees content
                                 [silent] GET messages?before_seq=X
                                   ~600ms in background
                                   (user has been reading the
                                    first page for that time)
  Scroll up → Load earlier →     Scroll up → "Load earlier" lands
    click → 600ms wait → ...        cached, instant

The "Load earlier" affordance still works for going further back
than 50 messages — the background fill only does one extra page.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>