Show the beginning of a new assistant response instead of following its tail

[artile]

Before submitting

• I searched existing issues and discussions and did not find a duplicate.
• I checked the current scroll behavior in web, desktop, Android, and iOS code.
• I am describing a concrete reading problem.

Area

apps/web, apps/desktop, and apps/mobile

Problem or use case

When an LLM produces a long answer, T3 Code follows the live edge and leaves the viewport at the bottom of the answer. I then have to scroll back to the beginning before I can read it in its natural top-to-bottom order.

The useful starting position for a completed answer is its first line, not its last line. This is especially noticeable on a phone, where only a small part of a long response fits on screen.

I checked the current implementation before filing this request:

• Android and iOS share the React Native thread feed. Its follow latch starts enabled and maintainScrollAtEnd follows data and item-layout changes.
• Web calls scrollToEnd for normal sends and keeps live follow enabled. The first turn in a new thread has partial send-time anchoring, but streaming adjustments still move the viewport to reveal the growing end.
• Desktop wraps the web chat, so it inherits the web behavior.

Proposed solution

Add a reader-first response anchor across all clients:

1. When the first visible assistant answer text for a turn appears, place the beginning of that answer near the top of the usable viewport, below any fixed header.
2. Keep that reading position stable while the rest of the answer streams. Do not chase the growing tail by default.
3. If the answer fits in the viewport, avoid unnecessary movement or blank space.
4. Keep the existing Jump to latest control. Reaching the bottom or pressing that control may re-enable live-tail following.
5. Never override deliberate user navigation. If the user has already scrolled elsewhere, preserve that position.
6. Apply the same behavior to web, desktop, Android, and iOS.

Tool activity may continue to use the current live-follow behavior before answer text exists. The important transition is when the readable assistant answer begins.

Why this matters

LLM answers are written top to bottom. Starting at the bottom forces every long response into the same repeated interaction: receive answer, scroll upward, find the first line, then begin reading. A response anchor removes that work and makes completion notifications useful because opening the thread lands at the place where reading should start.

Smallest useful scope

For long answers, anchor the first assistant text block when it appears and stop automatic tail-follow until the user explicitly returns to the bottom.

Related work

• [Feature]: Stop auto-scrolling when the user reads earlier streamed output #3925 fixed forced auto-scroll after a user manually scrolls upward. This request changes the default position for a newly arriving answer, before manual scrolling is needed.
• Discussion [Feature]: Pin the latest user message to the top of the chat viewport (two-line sticky prompt, with a setting to disable) #6946 proposes a persistent preview of the latest user message. This request concerns the initial reading position of the assistant answer itself.

Acceptance criteria

• A long response finishes with its beginning visible, not its final lines.
• The reader can start from the first line without scrolling upward.
• Streaming content does not move that first line away once it is anchored.
• Manual scrolling always wins over automatic positioning.
• Jump to latest still works and can resume end-follow.
• Behavior is consistent across web, desktop, Android, and iOS.

Contribution

• I would be open to helping implement this.

[Rumi-sketches]

+1, and I'd like this to be a choice in Settings rather than only a new default.

On desktop (Windows) this happens on almost every turn: when the agent finishes I'm at the bottom of its reply and have to scroll back up to find where it starts before I can read it. It's the action I repeat most in T3 Code.

A detail from the current code (main, Oct 19) that explains why it happens so often: web already anchors the sent message near the top for a new turn, but shouldReleaseTimelineAnchorForToolActivity in apps/web/src/components/ChatView.logic.ts releases that anchor and goes back to following the end as soon as the running turn has any tool or command activity. With Claude Code and Codex nearly every turn uses tools, so the anchor rarely survives until the answer text arrives.

Suggested setting:

While a response streams

• Follow the latest output (current behavior)
• Keep the start of the response in view

With the second option, tool activity wouldn't release the anchor, the first line of the answer would stay on screen while the rest streams, and Jump to latest would still be there to return to the tail.