[Raing5Days]

Summary

Optimize Web UI flow_chat session scrolling and streaming projection performance during long multi-round turns (10 -> 300+ ModelRounds), and add a categorized benchmark harness (src/web-ui/src/flow_chat/perf/longRoundPerfHarness*) covering the 5 frontend projection/rendering stages.

Key changes:

• Store & adjacent group projection (roundGroups.ts, lifecycle.ts, groupProjection.ts, modernFlowChatStore.ts):
  • Cache getProjectedModelRoundGroups, getModelRoundFlowGroups, getCachedInlineFlowGroupData, flowGroupLifecycle, and mergeFlowGroupRun via WeakMap so settled ModelRounds do not recompute buildFlowItemGroups or computeExploreStats on every streaming flush.
  • Add prefix-aware turn projection (projectTurnItemsWithPrefixCache, canStartFreshFlowGroupRunAt, reconcileTurnVirtualItems, reconcileProjectedGroupsInPlace) so unchanged settled rounds before the last hard group boundary skip projectAdjacentFlowGroups and preserve exact === VirtualItem and FlowGroupData object identities (97.7% settled VirtualItem reuse, 94.6% settled TimelineBlock reuse).
• Container group indexing, receive feedback & search (selectors.ts, receiveFeedback.ts, useFlowChatSearch.ts):
  • Cache per-item flow groups (getVirtualItemFlowGroups) and incrementally reuse prefix maps in indexFlowGroups.
  • Replace JSON.stringify([turnId, itemId]) allocations in receiveFeedback.ts with cached per-item keys, skip completed turns in captureFlowGroupFeedbackSnapshot, and short-circuit collectFlowGroupReceiveFeedback when previous.activeMembers.size === 0.
  • Cache pre-folded search sources and per-item matches in WeakMap<VirtualItem, ...>.
• Retained memory budget & background event guard (resourceBudget.ts, EventHandlerModule.ts):
  • Cache byte estimates for completed ModelRounds and completed-round prefix sums per turn in estimateRetainedBytes, turning the 500ms budget walk on active turns from O(rounds) to O(1) incremental tail evaluation.
  • Ignore missing_state_machine data events for unloaded internal background sessions (e.g. MemoryPhase2) instead of logging warnings and triggering requestRuntimeProjectionRepair hundreds of times per second.
• Virtualizer & message list (useFlowChatVirtualizer.ts, VirtualMessageList.tsx):
  • Avoid scheduling an extra RAF re-render when synchronous virtualizer commits stay within MAX_SYNCHRONOUS_VIRTUALIZER_COMMITS.
  • Skip scheduling viewportSnapshot RAF and full DOM geometry reads when onViewportSnapshot is not provided, and lazily sample readFlowChatViewportGeometry in native scroll handlers.

Type and Areas

Type:

Performance / refactor / test

Areas:

Web UI (src/web-ui/src/flow_chat, src/web-ui/src/shared/utils/resourceBudget.ts)

Motivation / Impact

During long multi-round agent turns (100–300+ rounds), every 32ms streaming flush previously re-projected all completed rounds in the active turn and allocated fresh VirtualItem / FlowGroupData references. That invalidated downstream WeakMap caches in ConversationDocumentProjection and React.memo on VirtualItemRenderer, causing O(rounds) main-thread work per flush and unnecessary settled-row re-renders.

Measured benchmark improvements (src/flow_chat/perf/longRoundPerfHarness.test.tsx, 10 -> 300 rounds, 20 flushes):

• Category 1 (bench:store-projection): 111.79 ms -> 7.82 ms (~14x faster), settled VirtualItem reference reuse 6.8% -> 97.7%.
• Category 2 (bench:timeline-document): 17.16 ms -> 3.12 ms (5.5x faster), settled TimelineBlock reference reuse 62.8% -> 94.6%.
• Category 3 (bench:group-state-search): 14.12 ms -> 3.74 ms (3.8x faster).
• Category 4 (bench:budget-persistence): estimateRetainedBytes per-call at 300 rounds 1.56 ms -> 0.01 ms (100x+ faster).
• Category 5A (bench:react-streaming, 120 rounds, 10 tail chunks): VirtualMessageList commits 23 -> 14, wasted settled-row renders 805 -> 2.
• Category 6 (bench:e2e-pipeline): Total 5-stage pipeline time 139.67 ms -> 12.38 ms (avg 0.39~0.62 ms/flush).

Verification

Automated checks run locally (all passed):

• pnpm --dir src/web-ui exec tsc --noEmit
• pnpm --dir src/web-ui exec eslint src/flow_chat/grouping/roundGroups.ts src/flow_chat/grouping/lifecycle.ts src/flow_chat/grouping/groupProjection.ts src/flow_chat/grouping/selectors.ts src/flow_chat/grouping/receiveFeedback.ts src/flow_chat/store/modernFlowChatStore.ts src/flow_chat/components/modern/useFlowChatSearch.ts src/flow_chat/components/modern/useFlowChatVirtualizer.ts src/flow_chat/components/modern/VirtualMessageList.tsx src/shared/utils/resourceBudget.ts src/flow_chat/services/flow-chat-manager/EventHandlerModule.ts src/flow_chat/perf/longRoundPerfHarness.ts src/flow_chat/perf/longRoundPerfHarness.test.tsx
• pnpm --dir src/web-ui exec vitest run src/flow_chat/perf/longRoundPerfHarness.test.tsx
• pnpm --dir src/web-ui exec vitest run src/flow_chat/store/modernFlowChatStore.test.ts src/flow_chat/grouping/activityGrouping.test.ts src/flow_chat/grouping/receiveFeedback.test.ts src/flow_chat/timeline/document.test.ts src/shared/utils/resourceBudget.test.ts src/flow_chat/services/flow-chat-manager/EventHandlerModule.test.ts src/flow_chat/components/modern/useFlowChatSearch.test.ts
• pnpm --dir src/web-ui exec vitest run src/flow_chat/components/modern/useFlowChatVirtualizer.test.ts src/flow_chat/components/modern/useFlowChatVirtualizer.cascade.test.tsx src/flow_chat/components/modern/useFlowChatVirtualizer.measurement.test.tsx src/flow_chat/components/modern/useFlowChatVirtualizer.follow.test.tsx src/flow_chat/components/modern/useFlowChatVirtualizer.initial-window.test.tsx src/flow_chat/components/modern/useFlowChatVirtualizer.aim.test.tsx src/flow_chat/components/modern/useFlowChatViewportAnchor.test.tsx src/flow_chat/components/modern/useFlowChatViewportOwner.test.tsx src/flow_chat/components/modern/useFlowChatFollowOutput.test.ts src/flow_chat/components/modern/useFlowChatFollowOutput.hook.test.tsx src/flow_chat/components/modern/useFlowChatFollowOutput.test.tsx src/flow_chat/components/modern/useFlowChatLeadingExtent.test.tsx src/flow_chat/components/modern/VirtualMessageList.session-boundary.test.tsx src/flow_chat/components/modern/VirtualMessageList.resize.test.tsx src/flow_chat/components/modern/VirtualMessageList.history-prepend.test.tsx src/flow_chat/components/modern/VirtualMessageList.timeline-prepend.test.tsx src/flow_chat/components/modern/VirtualMessageList.edgeFade.test.ts src/flow_chat/components/modern/VirtualMessageList.layout.test.ts src/flow_chat/components/modern/modelRoundItemMemo.test.ts
• pnpm run check:repo-hygiene
• node scripts/check-git-object-sizes.mjs --base origin/main --head HEAD

Remote scenarios & manual checks:

• Exercised in local unit/DOM fixtures; no wire protocol, persisted schema, or remote surface table changes. Manual native WebView scrolling checks in FLOWCHAT_VERIFICATION.md remain pending human confirmation.

Reviewer Notes

• AI-assisted PR; fully tested against the FlowChat unit and virtualizer/scroll-stability test suites (22 test files, 575 tests passing).

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.