claude-code - 💡(How to fix) Fix [BUG] Agent View silently rebinds my mouse wheel -- please, I'm begging, give it back

Preflight Checklist

• I have searched existing issues
• This is a single bug report
• I am using the latest version of Claude Code

What's Wrong?

I come to you not with a stack trace, but with a plea.

In classic claude, my mouse wheel scrolls the conversation. Up to read what was said, down to come back. This is the natural order of things. I did not choose it; the universe did.

In the new Agent View, once I drill into a single agent's conversation, the mouse wheel quietly changes religion. It no longer scrolls the conversation -- it now cycles through my previous inputs, i.e. the exact thing the Up/Down arrow keys already do, and have always done, and will presumably continue to do until the heat death of the terminal.

So now I have two ways to do the input-history thing and zero ways to do the scroll thing, using the same finger, in the same app, depending on which screen I'm on.

The cruelty is in the subtlety. A dead scroll wheel I could mourn and move on. But a scroll wheel that does something plausible-but-wrong -- that's psychological. Every ten seconds I spin the wheel expecting to review text, and instead my last three prompts flash past like my life before my eyes. My finger memory, lovingly trained over months of classic CLI use, has been declared legally invalid without notice.

Related but distinct:

• #59901 -- Agent View attached-session scroll is fully broken on macOS. Mine is worse in spirit: my wheel isn't dead, it's been reassigned.
• #12953 -- same symptom (wheel cycles input history) but caused by a stray watch command, not Agent View.

What Should Happen?

Inside an Agent View agent conversation, the mouse wheel should scroll the conversation, exactly like it does in classic claude. Input-history navigation already has a perfectly good home on the Up/Down arrow keys; it does not need a summer house on my scroll wheel.

Steps to Reproduce

1. Run claude agents (or press Left arrow on empty input in a classic session to enter Agent View).
2. Drill into a single agent's conversation.
3. Generate output longer than the viewport.
4. Scroll the mouse wheel expecting to review the conversation.
5. Observe: the wheel cycles through your previously typed inputs instead.

A Humble, Knees-On-The-Floor Appeal

This issue is about one mouse wheel. The plea is bigger than one mouse wheel.

Please -- never ship finger-memory-breaking changes. I am begging. Not "rarely." Not "only when justified." Never.

Here is why this category of change is uniquely cruel. Muscle memory is the whole point of a CLI. We spend months training our fingers to act without thought -- that unconscious fluency is the productivity. It is the entire deal. When you silently remap a trained interaction (a key, a wheel, a chord), you don't just change a behavior -- you revoke the unconscious and force every action back through the slow, deliberate, thinking part of the brain. Every ten seconds I now have to stop and think to do something I used to do for free. You have not added a feature; you have added a tax, and you charge it on every single scroll, forever.

A bug breaks the app once and we file an issue. A finger-memory change breaks the user continuously and there's no issue to file -- just a quiet, permanent decline in everyone's flow that never shows up in your metrics.

So, the ask, as general as I can make it:

1. Treat trained interactions as a contract, not a design surface. What the wheel/arrows/Ctrl-keys do is owned by our fingers now. It's not yours to silently reassign.
2. Never let the same input mean different things in different views. Per-screen remapping is the worst case -- it punishes exactly your most experienced users, the ones who scroll without looking.
3. If a change to a core interaction is truly unavoidable, make it loud and opt-in -- announce it, don't default into it, let us keep the old binding.
4. When in doubt: consistency over cleverness. Always.

Thank you for reading all of this. I will now go scroll something and be disappointed -- but hopeful.

Is this a regression?

Yes -- Agent View introduced the behavior change. Classic mode still scrolls correctly.

Claude Code Version

2.1.143 (Claude Code)

Platform / Terminal

Windows 11, running Claude Code inside the VS Code integrated terminal -- which is where the mouse-wheel behavior originates, and which is also one of the most common ways people run a CLI day-to-day, not an exotic setup. Behavior is still view-specific, not terminal-specific: #59901 reproduces the same Agent View scroll breakage on macOS in standalone terminals (Ghostty, Warp), so this is Agent View's rendering path, not VS Code's.