Performance Regression: Auto-Expanded Thinking

[nabilfreeman]

App Version

3.28.2

API Provider

Anthropic

Model Used

Claude 4 Opus Max Thinking Tokens

Roo Code Task Links (Optional)

ReasoningBlock timer forces 1 Hz re-renders & CPU spikes (introduced in bbd3d98)

Problem
After bbd3d98 the chat view re-renders every second while the assistant is streaming, even if the reasoning block is never expanded. On long conversations this chews through CPU, blows up paint/layout costs, and feels janky. As the conversation lengthens the problem becomes worse and worse. Locked CPU means Roo is able to apply changes more slowly causing productivity to grind to a halt.

What changed

File & line	Old behaviour	New behaviour	Cost
ReasoningBlock.tsx#L28-L33	⬅️ no timer	setInterval(tick, 1000) inside the component	Triggers a state write → React commit once per second, dragging the whole markdown subtree along for the ride.
ReasoningBlock.tsx#L53-L57	Body was hidden unless user expanded (isCollapsed)	Markdown now always mounted – even when the user never opens it	Unnecessary parsing/layout on every commit; for huge chains-of-thought this dominates the main thread.
ChatRow.tsx#L1088-L1092	reasoningCollapsed defaulted to true and could be toggled	Collapse flag removed; block is always rendered	Eliminates the cheapest escape hatch (not rendering at all).

How to reproduce

1. Open any long conversation (>= 1000 tokens of reasoning)
2. Submit a message
3. Watch CPU usage spike

Proposed fixes

• Lift the timer into a tiny <ElapsedTime> child wrapped in React.memo so only the counter re-paints.
• Re-apply the old 160-char debouncer when streaming partial chunks.
• Collapse thinking by default again (probably not desired based on the issues I read and recent change).

🔁 Steps to Reproduce

1. Open any long conversation with a model that has very high thinking tokens
2. Submit a message
3. Watch CPU usage spike

💥 Outcome Summary

Expected Roo to stay snappy on long thinking workflows, but now it blows up my M1 Max at 100% CPU usage whenever the model is thinking.

📄 Relevant Logs or Errors (Optional)

[nabilfreeman]

I think debouncing the rendered content will probably be the most impactful. I've just realised on fast output models like Claude this re-render chain will probably happen many, many more times than 1 per second

[nabilfreeman]

I've opened PR #8000 with a fix for the performance regression.

The solution isolates the timer updates into a memoized component, preventing the entire ReasoningBlock from re-rendering every second. This should resolve the CPU spike issues you were experiencing with long reasoning chains.

The implementation:

• Extracts timer logic into a separate ElapsedTime component with React.memo
• Only re-renders the timer display, not the entire component tree
• Conditionally renders markdown content only when it exists

All tests are passing and the CI checks are in progress. Feel free to review and test the changes!

@roomote-agent What do you think about debouncing the streamed content as well? That way the maximum number of renders per second can be capped in a deterministic way.