Key results

Merge comparison

Summary

• Replaces the long-session Virtua timeline with TanStack Virtual using chronological semantic rows, stable keys, measured navigation caches, end anchoring, and a fixed 64px composer gap.
• Preserves visible keyed DOM coordinates while scrolling, loading older history, measuring very large rows, switching session tabs, expanding content, and dragging the custom scrollbar.
• Moves streaming Shiki tokenization into a dedicated worker and updates Markdown/code DOM incrementally without replacing stable token nodes.
• Separates timeline concerns into timeline/model.ts, timeline/projection.ts, timeline/rows.ts, and timeline/message-timeline.tsx.
• Coalesces adjacent provider text deltas within the existing 16ms event flush and applies realistic provider fragments without unnecessary character-by-character catch-up.

User-visible changes

Stability

• Long mixed-content sessions no longer paint blank ranges, overlap rows, reverse scroll direction, or move retained rows by unrelated distances.
• Loading older history preserves the latest visible keyed anchor, including when the user keeps scrolling while delayed pages load.
• Large asynchronous row measurements preserve visible row identity during the correction window.
• Cold and cached titlebar tab switches paint the latest message first without flashing earlier destination rows or stale source-session content.
• Cached titlebar returns reuse parsed Markdown synchronously, avoiding a plaintext-first paint followed by an equivalent rich-DOM repaint.
• Custom scrollbar dragging remains monotonic while content height changes.
• Collapsed tool sections retain their state when later assistant content streams.

Streaming

• Shiki/WASM highlighting runs in a dedicated worker; worker failure degrades to plaintext without moving highlighting onto the UI thread.
• Streaming code blocks retain their shell and stable token DOM as worker results arrive.
• Closing backtick and tilde fences are recognized even when the marker is split across separate provider deltas, so following prose cannot remain trapped inside the code block.
• Completed worker token state is released only after the latest response, while explicit disposal clears both token and request-tracking state.
• Text deltas append to existing part text instead of collapsing the part to only the latest delta.
• Adjacent deltas for the same message, part, and field are combined within the existing event flush while preserving non-delta event boundaries.
• Normal 3–4-word provider fragments render as one update; only large backlogs use paced catch-up.

Architecture

session.tsx
  -> timeline/model.ts       session synchronization, filtering, pagination policy
  -> timeline/message-timeline.tsx
       -> timeline/projection.ts  reactive message-to-row projection and identity reuse
            -> timeline/rows.ts   pure semantic row construction, keys, equality

message-timeline.tsx now owns browser presentation concerns only: TanStack Virtual, row rendering, measurement, gestures, sticky UI, and anchor registration.

Performance methodology

• Tab timing starts inside the browser during capture-phase dispatch of the actual titlebar click.
• Correct paint requires the latest destination message and bottom spacer to be painted at the viewport bottom.
• Stable confirmation requires three consecutive correct painted samples.
• Cold means the destination has not been rendered or measured in the current document.
• Hot means the destination was visited and measured, switched away from through the SPA titlebar, then revisited.
• Windows native results do not call the CPU-throttling API.
• Mac streaming stress uses explicit 30x CPU throttling.
• Busy-stream FPS is sampled only while applied renderer progress changes, not over trailing idle time.
• Streaming fragments contain roughly 3–4 words and deliberately split bold markers, fences, table cells, links, and TypeScript tokens across updates.
• Historical stress content contains Markdown tables, blockquotes, TSX/Bash fences, reasoning, realistic edit/write before-after data, and apply-patch payloads.

Performance results

MacBook Air: session tab switching

Native speed, median of five fresh browser contexts.

The earlier PR implementation measured approximately 330ms cold; the staged bottom-first mount reduced the current committed result to 54.2ms.

Windows: session tab switching

Native Chromium, no CPU slowdown, median of five fresh browser contexts.

MacBook Air: realistic busy streaming at 30x CPU stress

Exact pushed head 411847bb0, 320 heavy historical turns, provider-like fragments.

This heavier diagnostic fixture exceeds the old 1,000ms long-task assertion on both branches. The assertion is not claimed as a passing CI performance gate; the relative measurements and visual invariants are the comparison.

Windows: realistic busy streaming, native capped UX

360 provider-like fragments delivered in bursts of eight. No CPU slowdown.

At 40 turns, upstream produced 22 frames below 30 FPS, two frames below 20 FPS, and 53 dropped-frame equivalents. The PR produced zero for all three and applied all 360 fragments.

Windows: exploratory uncapped Chromium

--disable-frame-rate-limit --disable-gpu-vsync is retained only as secondary throughput evidence because rAF pacing remains machine/driver dependent.

The capped UX table is the authoritative Windows result.

Stability verification

Committed coverage

• packages/app/e2e/smoke/session-timeline.spec.ts
• Delayed history prepend preserves visible part coordinates.
• Exact 64px bottom spacer and bottom pinning.
• Cached session tabs paint the latest message first.
• The first visible cached Markdown subtree remains intact for 200ms after the correct paint; visible first-paint nodes may not be removed and replaced asynchronously.
• Cold session tabs paint the latest message first.
• Full seeded timeline remains ordered while paging history.
• packages/app/test-browser/solid-virtual.test.ts
• Reactive count updates retain measured row sizes.
• Logical scroll offset includes pending TanStack measurement adjustments.
• packages/app/src/pages/session/timeline/model.test.ts
• Revert filtering and pagination continuation/stopping rules.
• packages/app/src/context/server-sdk.test.ts
• Delta coalescing, event boundaries, field separation, and stale-delta suppression.
• packages/ui/src/components/markdown-stream.test.ts
• Incomplete emphasis/links, whole and split fences, tables, replacements, references, metadata, trailing newlines, and append projection.
• packages/ui/src/components/markdown-worker-protocol.test.ts
• Token accumulation, generations, stale responses, replacement streams, and latest-only completion cleanup.

Ignored local diagnostics

These high-volume tests live under ignored packages/app/e2e/local/ and are intentionally not included in the PR diff.

• session-timeline-scroll-stability.spec.ts
• Traverses the 72-turn mixed fixture bottom-to-top, down, and up again.
• Uses keyed visible DOM movement as the assertion oracle, never raw scrollTop.
• Asserts wheel-to-DOM error, differential row movement, real node remounts, blanks, overlap, gaps, and idle drift.
• Includes a non-virtualized edit diff approximately 77,000px tall.
• Result: 0px error, 0 remounts, 0 blanks, 0 gap/overlap, 0 idle drift.
• session-timeline-jump-matrix.spec.ts
• Cold up/down/up traversal, 24px high-resolution movement, rapid reversals, 600px flings, repeated settle cycles, and rapid painted wheel bursts.
• Result: 6/6 pass.
• session-timeline-stability.spec.ts
• Repeated delayed pagination with assistant-only page boundaries and continued upward input.
• Asserts the latest visible key remains fixed and index 0 never appears while older history remains.
• Result: pass; manual snap-to-top symptom not reproduced after the geometry fixes.
• session-timeline-cross-branch-stability.spec.ts
• Identical framework-neutral traversal on Virtua and TanStack branches.
• Mac PR: 0px error, 3/3.
• Windows PR: 0px error, 3/3.
• Windows upstream: 0.5px error, 3/3.
• session-tab-switch-benchmark.spec.ts
• Five fresh cold and five fresh hot browser contexts with browser-owned click timing and painted-frame sampling.
• session-timeline-benchmark.spec.ts
• 30x stress and native Windows busy streaming, realistic Markdown/code/tools/diffs, applied-fragment progress, FPS/tail latency/long tasks, visual identity, and bottom-follow checks.

Root causes fixed during validation

• Cold overscan 50 synchronously mounted expensive semantic rows and caused approximately 330ms tab switches.
• Cached Markdown initially painted plaintext fallback and then asynchronously replaced it with equivalent cached rich HTML. Exact parsed cache hits now seed the resource synchronously with owner-stable block keys.
• Lower overscan exposed invalid offscreen file geometry because nested content-visibility:auto reported 0px wrappers around rendered Pierre content.
• Very large asynchronous row growth could temporarily evict visible identities; visible indexes are pinned for two frames only when a single measured delta exceeds one viewport.
• Windows missed a deterministic 16px correction because the app predicate compared against raw scrollOffset while TanStack had pending batch adjustments. getLogicalScrollOffset() exposes the exact logical value used by the core.
• The text pacer split realistic provider fragments into many extra character updates; normal fragments now render together while only large backlogs are paced.
• The server event queue previously reduced every delta separately even inside one 16ms flush; adjacent compatible deltas are now applied once.

Dependencies and upstream work

• TanStack Solid Virtual 3.13.28 with a lifecycle patch that publishes reactive count changes without destructive measure() resets.
• TanStack Virtual Core 3.17.0 with a narrow public logical-offset accessor used by the stability predicate.
• TanStack issue: solid-virtual: reactive option updates clear measured item sizes TanStack/virtual#1197
• Pierre Diffs 1.2.10.
• Shiki and @shikijs/stream 4.2.0.

Verification summary

• App unit tests: 407 passed.
• UI tests: 57 passed.
• Timeline smoke tests: 5/5 passed.
• Solid virtualizer browser contracts: 2 passed.
• Mac cross-branch stability: 3/3, 0px error.
• Windows cross-branch stability: 3/3, 0px error.
• Strict giant-row traversal: pass with zero visual failures.
• Rapid-scroll matrix: 6/6 passed.
• Repeated delayed history regression: passed.
• Repository-wide push-hook typecheck: 23 tasks passed.
• git diff --check: passed.

Remaining performance headroom

• The pushed PR sustains native Windows 60 FPS through 640 heavy historical turns.
• Under synthetic 30x CPU stress, slow provider cadence reaches approximately 51 FPS and fully saturated bursts remain below 60 FPS.
• Profiling shows the remaining stress cost is primarily browser layout/paint and active-response DOM application, not Shiki worker CPU.
• No stability requirement was relaxed to improve these numbers.