When an agent session is producing output, the view does not stick to the bottom — the latest (bottom-most) output scrolls out of view and I have to manually scroll down to keep following along. It seems to stop tracking the bottom once terminal/code-block widgets get appended to the conversation.
Start an agent task that produces a lot of streaming output (including terminal/code blocks).
Keep the conversation scrolled to the bottom.
Observe: as new output arrives (especially after a terminal/code widget), the view stops following the bottom and you must scroll down manually to see the latest output.
Expected Behavior
While the conversation is scrolled to the bottom, it should stay pinned to the bottom as new output streams in, so the latest output remains visible without manual scrolling.
This is a known one: while the agent is streaming, the chat can lose its “stick to bottom” lock when a terminal or code block expands and changes the layout mid-stream, so new output lands below the fold until you scroll down. Our team is actively working on this - a few fixes have already shipped and more are in progress - though it’s a stubborn one that’s had several causes, so it isn’t fully resolved yet.
There’s no clean workaround for now beyond using the scroll-to-bottom control (or scrolling manually) while it streams.
One thing that would help us route this to the right fix: are you seeing it with local agents, cloud agents, or both in that window? The streaming-tail behavior has a slightly different cause for cloud agents, so knowing which one you’re on helps us pin it down.
The chat panel does not reliably auto-scroll as the AI responds. Sometimes it works, but often it stops partway through and I have to manually scroll down to see the latest content.
This seems most noticeable when the AI is doing extended work with many MCP tool calls. I’ve also seen it when the chat was auto-scrolling, the AI enters a thinking phase, and then continues — at that point auto-scrolling often stops.
The first message in a new conversation seems more likely to scroll correctly, but follow-up messages are where I notice the problem most.
Steps to Reproduce
Open Cursor IDE and start a new chat conversation.
Send a follow-up message that causes the AI to make many MCP tool calls.
Watch the chat panel while the AI works.
Observe whether the view stays scrolled to the latest content.
Result: Auto-scrolling often stops partway through. When the AI finishes, I frequently have to manually scroll down to see the end.
I’ve also noticed a similar pattern when the AI moves from thinking into continuing its response — scrolling that was working can stop at that transition.
Expected Behavior
The chat panel should reliably keep the latest content in view as the AI streams, thinks, and executes tool calls. I shouldn’t have to manually scroll to catch up.
We’ve made a number of improvements to how the chat follows streaming output since you posted! Could you give it another try on the latest version? Quick check:
Start an agent task that produces a lot of streaming output, including terminal/code blocks.
Keep the conversation scrolled to the bottom.
Watch whether the view keeps following as new output streams in after a terminal or code block appears.
Let me know if the chat still stops following the bottom. cc @KamilTheDev