Load long Conversation transcripts around the read position #2568

The dashboard Conversation page loads and renders every transcript row from the latest event backward. With read positions (#2513), it often opens in the middle of the transcript, but it still loads and renders from the end.

Current behavior

• The detail request returns the latest 500 events (limit default in conversationDetailQuerySchema).
• Older pages load only when the reader scrolls near the top, through an IntersectionObserver sentinel (useAutoLoadPreviousPage). Nothing loads in the background.
• Every loaded row stays mounted. There is no list virtualization, so render and layout work grows with the transcript.
• The event API pages only backward (before cursor). It has no page that starts at a given seq.

Gap

• A reader whose read position is far behind still waits for the latest page, then scrolls back through it to reach the "New" marker.
• If the read position is older than the first loaded page, the marker shows at the top of the loaded rows, not at the first unread row.
• Long-lived Conversations keep all loaded rows in the DOM, which can cause jank.

Related code

• Scroll anchoring for older pages (fix(dashboard): Hold the reader's place when transcript layout changes #2511) and bottom pinning (transcriptBottomPinning.ts).
• In-page transcript search searches only loaded rows (transcriptSearch.tsx).

Proposal (from BYK's review of #2513)

• Cap how far toward the present the first load goes, based on the known read position. Load the rest piece by piece in the background after first paint.
• Cap how far back into the past the page loads, and use a virtual list.

via @BYK.

--

View Session in Sentry