Scroll direction is inverted on KDE Plasma Wayland

[Cowation]

What happened?

On KDE Plasma Wayland, the scroll MCP tool scrolls vertically in the opposite direction from the requested direction:

• direction: "down" scrolls upward.
• direction: "up" scrolls downward.

This occurs when input is sent through the XDG RemoteDesktop portal.

Tracing v0.4.1 shows that computer-use-linux maps the directions in src/remote_desktop.rs as:

ScrollDirection::Up => (AXIS_VERTICAL, steps.max(1)),
ScrollDirection::Down => (AXIS_VERTICAL, -steps.max(1)),

The KDE portal passes this signed discrete value directly to the Wayland vertical-scroll axis:

m_fakeInput->axis(
    WL_POINTER_AXIS_VERTICAL_SCROLL,
    wl_fixed_from_double(delta)
);

KDE's continuous-axis implementation, however, explicitly negates the vertical value:

m_fakeInput->axis(
    WL_POINTER_AXIS_VERTICAL_SCROLL,
    wl_fixed_from_double(-y)
);

The observed reversal therefore appears to be a polarity mismatch in the discrete portal path.

The ydotool fallback uses Linux REL_WHEEL semantics and should be handled separately. An unconditional global sign flip could break that backend.

Environment:

computer-use-linux: 0.4.1
Installation: prebuilt x86_64 release packaged through Nix
NixOS: 26.11.20260719.241313f (Zokor)
KDE Plasma: 6.7.3
xdg-desktop-portal: 1.20.4
xdg-desktop-portal-kde: 6.7.3
Session: Wayland
Portal pointer forced: COMPUTER_USE_LINUX_FORCE_PORTAL_POINTER=1

Steps to reproduce

1. Run KDE Plasma on Wayland with computer-use-linux configured to use the portal pointer backend:

   COMPUTER_USE_LINUX_FORCE_PORTAL_POINTER=1
   

2. Start the MCP server:

   computer-use-linux mcp

3. Open an application containing enough content to scroll vertically.

4. Place the pointer over its scrollable area, or target the window through the scroll tool.

5. Call:

   {
     "direction": "down",
     "pages": 1
   }

6. Observe that the application scrolls upward.

7. Call:

   {
     "direction": "up",
     "pages": 1
   }

8. Observe that the application scrolls downward.

The tool reports success through the RemoteDesktop portal in both cases.

Expected behavior

The direction names should describe the resulting viewport movement consistently across input backends:

• direction: "down" should move the viewport downward through the document.
• direction: "up" should move the viewport upward through the document.

Portal and ydotool backends should normalize their native axis conventions internally so callers do not need backend-specific reversed directions.

Version

agentsys and agnix are not installed because computer-use-linux is installed directly as a NixOS package.

Platform

Codex CLI

OS

NixOS 26.11 (Unstable)

[avifenesh]

Thanks for the excellent write-up, @Cowation — the discrete-vs-continuous KDE path comparison made this straightforward to pin.

Triage: confirmed as a portal-backend polarity bug on KDE only. The ydotool path already uses Linux REL_WHEEL signs (up → positive, down → negative); the portal path used the same signs, which is correct for GNOME/generic but inverted under xdg-desktop-portal-kde's discrete NotifyPointerAxisDiscrete handling.

Fix: PR incoming — invert vertical discrete portal steps when the session looks like KDE/Plasma. Horizontal stays standard. ydotool untouched. Override if needed:

COMPUTER_USE_LINUX_PORTAL_SCROLL_INVERT=1   # force invert
COMPUTER_USE_LINUX_PORTAL_SCROLL_INVERT=0   # force standard (match REL_WHEEL)

Would appreciate a quick verify on your NixOS Plasma 6.7 setup once the PR lands / with a local build of the branch.

[Cowation]

Issue seems to be fixed with the new branch on KDE.

[avifenesh]

Shipped in v0.4.2 — https://github.com/agent-sh/computer-use-linux/releases/tag/v0.4.2

Available now on crates.io (computer-use-linux 0.4.2), npm (@agent-sh/computer-use-linux@0.4.2), and as prebuilt x86_64/aarch64 binaries on the release.

@Cowation — when you get a chance, please confirm on your NixOS Plasma 6.7 setup with COMPUTER_USE_LINUX_FORCE_PORTAL_POINTER=1: direction: "down" should now move the viewport down. If your environment somehow needs the old polarity, COMPUTER_USE_LINUX_PORTAL_SCROLL_INVERT=0 restores it. Thanks again for the discrete-vs-continuous KDE source comparison — that's what made this a 10-minute diagnosis instead of a guessing game.