Message Scroller

A chat scroll container that anchors turns, opens saved transcripts, follows streamed responses, loads history without jumping, and jumps to any message.

New Chat
I'm building a chat for our app and the scroll behavior is driving me nuts. Every time the AI streams a reply, the whole thread jumps around.
Demo is read only. Press send to send messages.

What Makes a Great Streaming Chat Experience

Building a chat interface used to be simple. You create an inverted list with an input. Type a message, it appends at the bottom. When a reply comes in, the list grows and scrolls. Done.

Streaming breaks that model. Messages arrive in chunks while you may still be reading, scrolling, or looking somewhere else entirely.

Now the challenge is preserving the reader’s place while the conversation keeps changing. Get that wrong and the experience feels jumpy: people are pulled to the bottom, lose context, and have to find their way back.

In practice, this comes down to scroll: when to follow, when to hold, and when to let the reader decide. A great streaming chat should:

  1. Move only when the reader asked to move. If someone is reading, don’t pull them somewhere else. Auto-scroll should never be the default.
  2. Follow only while they’re following. If they’re at the live edge, keep the stream in view. If they scroll away, leave them there.
  3. Every interaction is a signal. Scrolling is not the only one. Selecting text, using the keyboard, opening a link, or searching should all stop the interface from moving.
  4. Start a new turn near the top of the viewport. This gives the new turn somewhere it can be read from the beginning.
  5. Then stream in the answer. The answer should grow into the screen, not immediately push everything away.
  6. Keep part of the previous conversation in context. The prompt and reply should stay visually connected, and enough of the previous turn should remain visible so the reader knows where they are.
  7. Let new content arrive offscreen. The conversation can keep streaming without changing what the reader is looking at.
  8. Show what’s happening out of view. Make it clear when a response is still streaming or when new messages have arrived.
  9. Make it easy to return to the latest reply. A “Jump to latest” action should bring the reader back and resume following.
  10. Let people jump anywhere in the conversation. Long threads need message links, search, unread markers, and direct navigation.
  11. Reopen where the reader left off. A saved conversation should open at the last meaningful turn. Often this is the last user message. Not the absolute bottom.
  12. Keep the reader’s place when layout changes. Images load. Markdown expands. Code blocks render. Older messages appear above. None of that should make the reader lose their place.
  13. Handle interruptions without stealing position. Stopping, retrying, regenerating, branching, or errors should not unexpectedly move the conversation.
  14. Stay responsive in long threads. Streaming text, markdown, code, images, and long history should still feel responsive.
  15. Be accessible without the noise. Keep the transcript navigable, preserve keyboard focus, and announce important events at a comfortable pace.

Never move the reader against their intent.

MessageScroller

MessageScroller is a chat transcript scroller built for these behaviors. MessageScroller.Provider owns the scroll state and transcript-row behavior: opening position, streamed output, new-turn anchoring, prepended history, visibility, and scroll controls. MessageScroller.Root is the styled frame that renders inside it.

MessageScroller is scoped to the scroll viewport. It does not own messages, AI state, transport, persistence, branching, or model state. Your product code stays focused on composing Message rows, Marker events, tools, attachments, and prompt inputs.

It gives you the scroll behavior that chat needs, without taking over the rest of the chat UI. And it stays fast, even in long conversations with rich markdown.

Installation

import { MessageScroller } from "@o/pipeline/experimental/components/chat/message-scroller";

Usage

import { Message } from "@o/pipeline/experimental/components/chat/message";
import { MessageScroller } from "@o/pipeline/experimental/components/chat/message-scroller";
<MessageScroller.Provider>
  <MessageScroller.Root>
    <MessageScroller.Viewport>
      <MessageScroller.Content>
        {messages.map((message) => (
          <MessageScroller.Item
            key={message.id}
            messageId={message.id}
            scrollAnchor={message.role === "user"}
          >
            <Message.Root />
          </MessageScroller.Item>
        ))}
      </MessageScroller.Content>
    </MessageScroller.Viewport>
    <MessageScroller.Button />
  </MessageScroller.Root>
</MessageScroller.Provider>

MessageScroller.Root fills its parent, so place it inside a height-constrained container.

<div className="flex h-screen flex-col">
  <MessageScroller.Provider>
    <MessageScroller.Root className="flex-1">
      {/* transcript */}
    </MessageScroller.Root>
  </MessageScroller.Provider>
</div>

Composition

<MessageScroller.Provider>
  <MessageScroller.Root>
    <MessageScroller.Viewport>
      <MessageScroller.Content>
        <MessageScroller.Item>
          {/* a message, marker, or row */}
        </MessageScroller.Item>
        <MessageScroller.Item />
        <MessageScroller.Item />
      </MessageScroller.Content>
    </MessageScroller.Viewport>
    <MessageScroller.Button />
  </MessageScroller.Root>
</MessageScroller.Provider>
  • MessageScroller.Provider — the headless root. Owns scroll state and the behavior props for opening position, auto-scroll, anchoring, scroll commands, and visibility tracking.
  • MessageScroller.Root — the styled frame. Lays out the viewport, content, and controls inside the provider.
  • MessageScroller.Viewport — the scrollable element. Receives native scroll events and preserves the visible row when older messages are prepended.
  • MessageScroller.Content — the transcript container. Holds the rows and provides the live-region defaults for new messages.
  • MessageScroller.Item — the transcript row boundary. Wrap every direct child of the content so the scroller can measure, anchor, preserve position, track visibility, and jump to it. An item can be a message, marker, typing indicator, separator, join/leave event, or “load earlier” row.
  • MessageScroller.Button — the scroll control. Scrolls to the start or end of the transcript and is inert until there is content in its direction.

Core Concepts

Anchoring Turns

A turn is the part of the conversation that starts a new exchange. In a simple AI chat, that is usually the user’s message and the assistant reply that follows.

An anchor is the row the viewport should treat as the start of that turn. Mark that row with scrollAnchor. When a new anchor is appended, the viewport moves it near the top and keeps a peek of the previous item above it, so the new turn does not feel detached from its context.

// This tells the scroller to anchor the user's message for the next turn.
<MessageScroller.Item
  messageId={message.id}
  scrollAnchor={message.role === "user"}
/>

Scroll anchors are not tied to message role. You can turn any row into an anchor: a user message, a system marker, a handoff event, or anything else that starts a meaningful turn. MessageScroller only needs to know which row should anchor the viewport.

In the following example, the user’s message is anchored. When you send a new message, the viewport anchors it near the top and appends the assistant reply below it. Toggle the anchor to the assistant’s message to see the difference.

Anchoring Turns
Can you show me how anchoring behaves when a new prompt starts the turn?
Toggle the anchor role, then send messages to compare where turns settle.

Group Chat

In a group chat, the turn boundary is more specific than “the user message”. It is often the message that asks the model to respond, or a marker like “Marcus joined the chat”. Typing indicators and history controls usually should not anchor.

Because anchoring is role-independent, you can anchor a Marker just as easily as a message.

<MessageScroller.Item messageId="marcus-joined" scrollAnchor>
  <Marker.Root variant="separator">
    <Marker.Content>Marcus joined the chat</Marker.Content>
  </Marker.Root>
</MessageScroller.Item>
Group Chat

This will create a marker and make it the anchor

When a user joins, a marker is created. scrollAnchor on the marker marks it as the next turn

Keeping Context Visible

When a new turn starts, it should still feel like part of the same continuous thread. scrollPreviousItemPeek keeps a slice of the previous item visible above the anchor, so the reader keeps their context instead of feeling like the conversation restarted on a blank page.

// Keep 64px of the previous turn visible above the newly anchored row.
<MessageScroller.Provider scrollPreviousItemPeek={64}>
  <MessageScroller.Root>{/* anchored turns */}</MessageScroller.Root>
</MessageScroller.Provider>

Adjust the peek amount in the example below to see how it affects the conversation.

Keeping Context Visible
Okay, but when someone sends a new message the view still feels jarring — like the whole conversation reloads from the top.
Adjust the peek preset and send. Observe the previous message peak

Following the Live Edge

When the reader is at the live edge, either because they stayed there or returned there, autoScroll keeps streamed replies in view as they grow. Scrolling away from the live edge releases the view, whether by wheel, touch, keyboard scroll keys, or dragging the scrollbar. An explicit message jump releases it too. New chunks can then arrive without moving the reader.

autoScroll composes with turn anchoring. When a new turn anchors near the top, the view stays put while the reply streams into the room below it. Once the reply fills the viewport, the reader is back at the live edge and follow-output takes over from the anchor.

<MessageScroller.Provider autoScroll>
  <MessageScroller.Root>{/* streamed turns */}</MessageScroller.Root>
</MessageScroller.Provider>
Streaming Messages
I'm building a chat for our app and the scroll behavior is driving me nuts. Every time the AI streams a reply, the whole thread jumps around.
Streaming is simulated. `autoScroll` is enabled.

Calling scrollToEnd, or pressing MessageScroller.Button, re-engages follow-output when autoScroll is enabled, so a reader who scrolled away can return to the live edge and keep following. The root and viewport expose data-autoscrolling while that programmatic scroll to the latest message runs, so you can conditionally apply styles during the transition.

Opening Saved Threads

It can seem reasonable to reopen a saved thread at the absolute end of the transcript, but that often drops the reader into the conversation without enough context. A better default is "last-anchor": show the last meaningful turn, like the user’s latest message, with the reply below it.

That gives the reader an immediate place in the thread. They can see what they asked, where the answer starts, and continue from there without reconstructing the conversation from the bottom edge.

<MessageScroller.Provider defaultScrollPosition="last-anchor">
  <MessageScroller.Root>{/* transcript */}</MessageScroller.Root>
</MessageScroller.Provider>
Opening Position
Toggle the defaultScrollPosition to see where the transcript starts when you open the thread

"last-anchor" is keyed on scrollAnchor, not message role. If no anchor exists, or the last anchored turn already fits in the viewport, it falls back to "end".

Use "start" when you want to resume at the beginning of a conversation, or "end" when the absolute latest message is the right place to land.

Loading Earlier Messages

Loading earlier messages should not move the conversation the reader is already looking at. When older rows are prepended above the current transcript, MessageScroller.Viewport preserves the visible row so the reader stays in the same place while history loads above them.

This is enabled by default through preserveScrollOnPrepend.

Load History
End of Conversation

Restore earlier messages while keeping your place.

Click Load History to load the entire conversation

Use stable messageId values for message rows. That gives the scroller a specific row to preserve instead of guessing from whichever pixel happens to sit at the viewport edge.

Animating New Messages

MessageScroller.Item can be animated directly. Create a motion version of the item, keep messageId and scrollAnchor on it, and use transform and opacity for the entrance.

A common chat pattern is to animate the user’s message when it is sent, then let the assistant reply stream into a regular row below it. Start the user row below its final position so it feels like it rises from the live edge of the viewport.

const MotionMessageScrollerItem = motion.create(MessageScroller.Item);
Animation
Can user messages pop in like iMessage without breaking anchoring?
Select an animation then click send to see it in action.

Avoid animating height, margin, or padding for row entrances; those changes can fight the scroller’s positioning work. If the reader prefers reduced motion, skip the entrance animation and keep the scroll behavior the same.

Jumping to Messages

Search results, permalinks, outline items, and toolbar buttons often need to drive the transcript from outside the message list. Use useMessageScroller for those controls. Because the hooks read from MessageScroller.Provider, they work in any component inside the provider, including controls rendered outside the MessageScroller.Root frame.

import { useMessageScroller } from "@o/pipeline/experimental/components/chat/message-scroller";
const { scrollToMessage, scrollToEnd, scrollToStart } = useMessageScroller();
Commands
Use the controls to jump to any message in the conversation.

scrollToMessage targets the messageId on MessageScroller.Item, so rows that need to be addressable should have stable ids. scrollToMessage returns false when the target is not mounted and cannot be queued.

scrollToMessage can queue a target before items exist, which covers client-resolved permalinks while the transcript mounts. After rows have mounted, a missing id returns false instead of starting a guessed retry loop. A true result means the scroll ran or was queued, not that the row is already in view.

Tracking the Reader’s Position

Use useMessageScrollerVisibility to track the reader’s position in the conversation. A common example is a table-of-contents or a jump menu that highlights the current anchored turn.

import { useMessageScrollerVisibility } from "@o/pipeline/experimental/components/chat/message-scroller";
const { currentAnchorId, visibleMessageIds } = useMessageScrollerVisibility();
Transcript Outline
Open the outline to jump between anchored turns as you read.

currentAnchorId answers “where am I” by reporting the current anchored turn, and it stays set after that anchor scrolls above the viewport. visibleMessageIds answers “what is on screen”, in document order.

Visibility is pay-for-what-you-use. Tracking only runs while something subscribes to useMessageScrollerVisibility, and rows need a messageId to participate.

Reading Scroll State

Use useMessageScrollerScrollable when you need scroll state in JavaScript, such as a status indicator or a custom “jump to latest” control. It reports which edges the viewport can still scroll toward; “at the start/end” is the negation (!start / !end), and “scrollable at all” is start || end. For styling the scroller itself, prefer the data-scrollable attribute.

import { useMessageScrollerScrollable } from "@o/pipeline/experimental/components/chat/message-scroller";
const { start, end } = useMessageScrollerScrollable();
Scroll Status
All messages fit in the viewport.
Scroll the transcript to see the footer update.

Performance

Use the large-transcript benchmark to compare mount cost, DOM size, and scroll frame time with the same composed rows. Run it in development, where React Profiler records render work, and compare relative measurements from the same browser and device.

Large transcript
Compare rendering strategies with the same composed rows. This microbenchmark does not reproduce AgentChat behavior. Development builds are intentionally noisy; compare relative results on the same device.
React commit render work
React base duration
Benchmark start to rows ready
DOM nodes
Mean scroll frame
Worst scroll frame
Frames over 20 ms

Our performance goal for MessageScroller is to keep the scroll hot path outside of React state: no React rerenders for transcript rows, no forced layout on every scroll, and as little off-screen paint work as the browser can avoid.

Scroll position, anchoring, and follow-output are tracked imperatively and mirrored onto the root and viewport through data-* attributes, so scrolling and streaming do not rerender transcript rows.

The benchmark’s contained mode applies content-visibility: auto and an intrinsic block-size fallback to its rows. This stays opt-in because skipped, variable-height rows can change exact jump and anchor alignment when the browser resolves their real size.

Visibility tracking is pay-for-what-you-use. A jump menu or active turn indicator costs nothing until something subscribes to useMessageScrollerVisibility.

Measure the actual transcript before choosing a more complex rendering model. Row containment does not reduce React construction, DOM node count, or client memory, so those costs can still justify virtualization for exceptionally long or rich transcripts.

Virtualization

Virtualization is intentionally left outside the primitive. Prefer real DOM rows when the benchmark is within the product’s performance budget because they preserve browser find, selection, copy, SSR, and the full accessibility tree.

When a transcript is large enough to need virtualization, use MessageScroller.Viewport as the scroll element and let the virtualizer own the rows.

The performance story’s virtualized mode isolates rendering cost. It is not a behavior-equivalent MessageScroller integration: virtual rows are not direct children of MessageScroller.Content, so item-aware anchoring, prepend restoration, visibility, and message jumps need an explicit adapter before production use.

Virtual rows are absolutely positioned, so the gap on MessageScroller.Content does not apply to them. Pass MESSAGE_SCROLLER_GAP as the virtualizer’s gap so virtual and DOM transcripts space rows the same way.

import * as React from "react";
import { useVirtualizer } from "@tanstack/react-virtual";

import { Message } from "@o/pipeline/experimental/components/chat/message";
import {
  MESSAGE_SCROLLER_GAP,
  MessageScroller,
} from "@o/pipeline/experimental/components/chat/message-scroller";

function VirtualizedTranscript({
  messages,
}: {
  messages: Array<{ id: string; content: React.ReactNode }>;
}) {
  const viewportRef = React.useRef<HTMLDivElement>(null);

  const virtualizer = useVirtualizer({
    count: messages.length,
    getScrollElement: () => viewportRef.current,
    estimateSize: () => 86,
    gap: MESSAGE_SCROLLER_GAP,
    getItemKey: (index) => messages[index]?.id ?? index,
    overscan: 8,
  });

  return (
    <MessageScroller.Provider>
      <MessageScroller.Root>
        <MessageScroller.Viewport ref={viewportRef}>
          <MessageScroller.Content className="block min-h-full">
            <div
              className="relative w-full"
              style={{ height: virtualizer.getTotalSize() }}
            >
              {virtualizer.getVirtualItems().map((virtualItem) => {
                const message = messages[virtualItem.index];

                if (!message) {
                  return null;
                }

                return (
                  <div
                    key={virtualItem.key}
                    ref={virtualizer.measureElement}
                    data-index={virtualItem.index}
                    className="absolute start-0 top-0 w-full"
                    style={{
                      transform: `translateY(${virtualItem.start}px)`,
                    }}
                  >
                    <Message.Root>{message.content}</Message.Root>
                  </div>
                );
              })}
            </div>
          </MessageScroller.Content>
        </MessageScroller.Viewport>
        <MessageScroller.Button />
      </MessageScroller.Root>
    </MessageScroller.Provider>
  );
}

Accessibility

MessageScroller keeps the scroll container keyboard reachable and the transcript announceable without forcing a specific message UI.

MessageScroller.Viewport is a labelled, keyboard-focusable scroll region by default. It uses role="region", aria-label="Messages", and tabIndex={0}, so keyboard users can focus the transcript and scroll it directly.

MessageScroller.Content marks the transcript as a live region with role="log" and aria-relevant="additions". New rows can be announced, but streamed text mutations do not have to be announced token by token.

<MessageScroller.Content aria-busy={status === "streaming"}>
  {/* messages */}
</MessageScroller.Content>

Pass aria-busy while a turn streams if announcements should wait for the completed message row.

MessageScroller.Button renders a real button. When there is nothing to scroll toward, it sets inert, uses tabIndex={-1}, and exposes data-active="false" so inactive scroll controls do not create extra focus stops.

Unstyled

The behavior in MessageScroller comes from the @shadcn/react package. To use it directly with your own markup and styles, see Message Scroller under @shadcn/react.

API Reference

The props, data attributes, and hooks for every part are documented on the @shadcn/react Message Scroller page. They are identical for the styled component and the unstyled parts.