feat(tui): make session tab switching fast for long transcripts - #39568

What

Switching session tabs could visibly stall on long transcripts. This PR makes tab switching roughly constant-time — the visible switch mounts a fixed-size tail regardless of transcript size.

Front-end only by design: an earlier revision also added data-context prefetch and sync-dedup plumbing, but paging/prefetch on the data layer deserves its own API design (and is being looked at separately), so this version touches nothing outside the session route.

Before / After

Measured with an isolated opencode-drive harness (three sessions of 51 / 123 / 243 transcript rows with realistic markdown + code-fence content, 160×45 viewport; sync = blocked main thread inside the switch, paint = next flushed frame):

Warm-switch numbers were measured with an earlier revision that also skipped a post-sync re-reduce; this version keeps that re-reduce (it needs a data-context change), adding ~6–23ms of CPU after the switch paints — the visible switch is unaffected.

Before: every switch tore down and rebuilt the entire <Session /> tree under the keyed route Show, eagerly mounting every transcript row — markdown parsing, tree-sitter highlighting, wrapping, and layout for up to 200 messages, scaling ~0.65ms/row.

After: a switch mounts only the newest 40 rows and backfills the rest in 60-row chunks shortly after, invisibly (bottom-anchored readers are held by sticky scroll; scrolled-up readers get scroll compensation). The switch cost no longer depends on transcript length.

Cold first visits still fetch inside the switch gesture (unchanged from before); hiding that behind prefetch is the data-layer follow-up.

How

• packages/tui/src/routes/session/index.tsx — tail-first transcript mounting. hidden derives from the row count until backfill pins it (no effect-ordering dependence), clamped so a shrinking re-reduce always leaves a tail visible. A 120ms timer drains 60 rows per chunk, pausing while streaming appends reset it. afterLayout consolidates the existing double-rAF “wait for layout” pattern (also adopted by alignMessage). jumpToMessage/scrollToMessage wrap in ensureAllRows, which force-mounts the full transcript before message navigation walks or jumps.
• packages/tui/src/routes/session/rows.ts — defers the first run of the four reactive re-reduce effects, which each duplicated the mount reduce.

sequenceDiagram
    participant U as user
    participant T as SessionTabs
    participant S as Session route
    participant D as DataProvider

    Note over U,S: on tab switch
    U->>T: leader+digit / click
    T->>S: route.navigate (keyed remount)
    S->>D: rows from store (cached if visited)
    S->>S: mount newest 40 rows (~40ms)
    loop every 120ms until drained
        S->>S: mount 60 older rows above viewport
    end

Scope

• No data-context changes: no prefetch, no sync return-type change, no plugin Data type change. Paging/prefetch belongs in a designed data-layer API (follow-up).
• No virtualization: rows above the viewport still mount (after backfill). Bounded by the 200-message fetch limit, so permanent windowing wasn’t warranted.
• One accepted edge: switching into an actively-streaming session before its rows settle can mount fully once (the pre-change behavior).
• The boundaryID index mapping stays positional; backfill briefly re-evaluates it per mounted row per chunk (bounded small at the message cap).
• The measurement harness (perf marks + opencode-drive benchmark scripts) is deliberately not included.

Testing

• bun typecheck in packages/tui
• bun run test in packages/tui: 562 pass, 5 skip, 1 fail — restores queued compaction from durable pending input fails identically on clean v2 HEAD (3c259fc552), pre-existing and unrelated
• oxlint + repo-pinned Prettier clean on changed files
• Benchmarked end-to-end with an isolated opencode-drive instance and JSONL perf marks (numbers above; light/heavy/scale content sweeps)
• Dogfooded against a live six-tab setup with 200–450-message sessions — switches read as instant
• Screenshot-verified: switch paints the bottom-anchored tail immediately and identically through backfill; a full row re-reduce mid-backfill renders correctly
• Review-verified (not exercised e2e — drive couldn’t transmit the needed keys): scroll-compensation during backfill for scrolled-up readers, and ensureAllRows under message navigation

Demo

https://assets.kitlangton.dev/anomalyco/opencode/pr-39568/tab-switch-demo-e5c2cbf8eb06.mp4

Recorded from the PR branch: cycling <leader>1/2/3 between three sessions (150 / 60 / 250 tool-call transcripts with heavy markdown) — every switch paints instantly.