Chat backward pagination

Overview

Opening a long, tool-heavy chat no longer blocks on a full transcript transfer. The chat loads only the newest page of messages, pins to the bottom, and then eagerly fetches every older page in the background until the whole transcript is in memory. The first screen (and the head summary, which previously iterated every message) is usable immediately on chats with hundreds of rows, and the user can scroll to any point without waiting for a per-page load.

Usage

No user-visible control — the behavior is automatic.

• On open, the client fetches GET /api/chats/:id/messages?limit=100 and paints the newest page, pinned to the bottom.

• Immediately after that first page paints, the client keeps fetching older pages (beforeSeq windows) in the background and prepends them above the loaded content, preserving the reading position, until the entire transcript is resident.

• If a turn is running or the chat changes mid-load, the background drain stops; a manual scroll to the top still falls back to loading a page on demand.

• Reaching the very first message sets a "no more" state and stops further requests.

  The transcript cursor is the same stable per-chat seq used by the append-only tail sync, so the paginated view and the streaming/append path never disagree.

Related

• Chat load performance — the earlier load-time work (cost aggregation, revision cursor, lazy tool results).

• Chat UI — the transcript shell and /messages endpoint.

• Chat streaming performance — the per-token streaming hot path.