feat(tui): page older session messages

[kitlangton]

What

Load only the newest 20 messages when opening a V2 TUI session, then page older history into the transcript as the user scrolls upward.

On the measured 200-message session, the initial API request dropped from 1.66 MB / 111.9 ms median to 72 KB / 12.7 ms median. Across four open tabs, that is about 86% fewer bytes and 5.3x less sequential server time before client decoding, row reduction, and rendering.

Before / After

Before: Every session switch fetched up to 200 projected messages before reducing the transcript. Opening several long tabs repeatedly paid for history that was not visible.

After: A session opens from its newest 20 messages. Reaching the top fetches the next page through a private cursor, prepends it chronologically, and preserves the visible message anchor. Short pages continue loading until the viewport is filled.

Commands that require complete history still behave completely: First message, previous-message navigation, timeline, fork, transcript copy, and Markdown export load the remaining pages before acting.

How

• packages/tui/src/context/data.tsx keeps the opaque server cursor private and adds data.session.message.older(sessionID, count).
• Older requests coalesce per session, deduplicate overlap, prepend in chronological order, and rebuild the message index.
• Exhaustion stays internal: older(...) returns the number of newly loaded messages and becomes a zero-result no-op once the cursor is exhausted.
• packages/tui/src/routes/session/index.tsx requests history at the viewport top and serializes each fetch through layout restoration.
• Anchor restoration prefers the existing visible message boundary, so concurrent bottom appends do not affect compensation; total-height growth remains the fallback for notice rows.
• The existing full rows.ts reduction runs after each prepend so reasoning, exploration, and usage groups may cross page boundaries safely.

Scope

• No protocol, server, generated client, or plugin Data API changes.
• No incremental row-prepend optimization or transcript virtualization.
• JSON export retains its existing direct ascending API pagination.

Testing

• cd packages/tui && bun run test (570 passed, 5 skipped)
• cd packages/tui && bun typecheck
• Repository-wide push-hook typecheck (33 tasks passed)
• Changed-file oxlint (0 errors; baseline warnings remain)
• OpenCode Drive: seeded a 15-turn session, opened it in a fresh TUI, verified newest-20 startup, upward prepend with request 6 held at the same viewport position, and First message landing on request 1

Demo

The recording uses OpenCode Drive's simulated model. It opens the fresh 20-message tail, pages upward while request 6 remains anchored, then invokes First message and lands on request 1.

recording-c26e375c-8eb1-4876-956b-b30012647881.mp4

Anchor after the prepend:

Flow

sequenceDiagram
    participant View as Session view
    participant Data as TUI data layer
    participant API as Message API

    View->>Data: sync(sessionID)
    Data->>API: newest 20, order=desc
    API-->>Data: messages + opaque older cursor
    Data-->>View: chronological loaded window
    View->>View: user reaches transcript top
    View->>Data: older(sessionID, 20)
    Data->>API: cursor page
    API-->>Data: older messages + next cursor
    Data->>Data: dedupe, prepend, rebuild index
    Data-->>View: enlarged chronological window
    View->>View: full row reduction + anchor restore

Loading