[Atom112]

Summary

Field report: switching from a new topic to the topic where AIO was working on the AIO project itself (9 rounds / 18 messages) loads visibly slower than other switches, and expanding its hundreds of intermediate steps collapses the frame rate.

I measured the real DB before touching anything — that topic carries 16.53 MB of agent_steps_json / ~1674 step cards:

message	steps	tool calls	steps JSON
cb4fb055	588 (201 thinking + 381 tool)	381	8.01 MB
33669af3	483	282	3.89 MB
4ba5b3d6	348	226	3.32 MB
86fb63f5	78	49	0.80 MB
ea217061	91	78	0.48 MB

Three stacked costs, all fixed:

1. Collapsed historical rounds still mounted every card (the switch cost)

max-height: 0 hides without unmounting — its own comment block in this file already warned about it. A topic switch therefore built up to 100 step cards per past round (~469 across this topic's five heavy rounds), each with its own signals/effects/memos and a 920 ms entrance timer, plus one linear stepById scan per card. The body is now gated by <Show when={isExpanded()}>: a collapsed round mounts zero cards (Solid's false branch builds no DOM, components, or effects), and the id-sync effect short-circuits while collapsed, resolving lazily on expand.

2. "Show earlier" mounted every remaining card in a single frame (the expand cost)

The showAllSteps boolean flipped to true and revealed all 588 at once. Replaced with paged reveal: each click adds one STEP_WINDOW (100) page, the button counts down the remainder — bounded DOM growth per frame.

3. O(n²) step lookup

stepById was timelineSteps().find(...), and timelineSteps() re-sliced the whole array on every call — roughly 345k element copies for a full reveal. Now a memoized timeline + id → step Map: O(n) once, O(1) per card.

Supporting changes

• content-visibility: auto + contain-intrinsic-size: auto 56px on the step card — offscreen cards skip layout/paint (the pattern already proven on MessageBubble rows, where this exact problem was solved before)
• Paged-in historical steps are pre-marked as animated (same contract as ChatInterface's load-earlier) so expanding does not fire a hundred simultaneous entrance animations
• Window arithmetic extracted to utils/stepWindow.ts (pure, 6 unit tests)

Deliberately untouched

• The active round's streaming render queue and auto-follow behaviour
• DiffView — measured: this topic has 1 deduplicated changed file, it is not a bottleneck
• The legacy no-agentSteps ToolCallBubble fallback (older data, not exercised here)

Verification

• npm run verify fully green: 93 FE tests (6 new covering active pass-through, tail-window bounds, page increments, short/empty/negative inputs, countdown), gates, lint, tsc, build; Rust suite green
• Honest scope note: the frame-rate/switch-time measurement needs the installed build — this machine has neither the reporter's profile nor a real WebView timer. What is asserted here is the mount arithmetic: 469 → 0 cards on switch, 588 → 100 per reveal click, plus the O(n²) → O(n) lookup