[BUG] Scroll position constantly snapping to end of conversation preventing scrolling to view previous output

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

I have reproduced this in the latest Claude Code in iTerm2, Ghostty, and Warp. I'm trying to read something the LLM output earlier but Claude Code keeps insisting on snapping the scroll position to the end of the conversation making it impossible to read. Once the LLM finishes output it will stop hijacking the scroll and then it lets me see the previous output. This is extremely problematic because I can't tell if the agent is going down a wrong path because it keeps preventing me from seeing it's earlier planning / thinking.

What Should Happen?

Rule 1: When I scroll it should scroll and stay where I leave the scrolling
Rule 2: If I am scrolled to the very bottom, then auto-scrolling is acceptable.
Rule 3: It should NEVER snap to the bottom of the conversation if I am scrolled up.

Error Messages/Logs

n/a

Steps to Reproduce

1. Type a prompt that generates a decent amount of output text that fills more than the screen has vertical room for.
2. While the agent is still going attempt to scroll up.
3. Notice the constant hijacking that snaps you back to the bottom that prevents you from reading the conversation above.

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.15

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

_No response_