vinzenz/hyperhive
vinzenz/hyperhive
Watch
0
0
Fork
You've already forked hyperhive
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

refactor(#2305): drop mouse tools + ydotool, switch key_press to wtype

Browse source 
- remove mouse_move and mouse_click (no Wayland-native alternative on Weston
  without /dev/uinput; follow-up filed for future investigation)
- replace key_press from 'ydotool key' to 'wtype -k': parses mod1+mod2+key
  into -M mod1 ... -k key ... -m mod1 sequence via virtual-keyboard protocol
- remove dest_path parameter from screenshot: always writes to /tmp/ (fixes
  arbitrary write-path concern from security review)
- simplify screen.nix: drop screenInput option, ydotoold systemd unit, ydotool
  package; only grim + wtype remain (both compositor-mediated, no /dev/uinput)
- update module header comment to reflect three-tool surface

Addresses mara's /dev/uinput veto (PR #2617 comment #40524).
This commit is contained in:
iris 2026-07-20 20:41:58 +02:00 • committed by mara
commit 7fa7e2bdfd
2 changed files with 55 additions and 154 deletions
Download patch file Download diff file Expand all files Collapse all files

50
nix/agent-modules/screen.nix
View file

@ -1,15 +1,13 @@
# Screen MCP — screenshot + input injection for GUI agents.
# Screen MCP — screenshot + keyboard input for GUI agents.
#
# Auto-activated when `hyperhive.gui.enable = true`. Wires the
# `hive-screen-mcp` stdio bridge as `extraMcpServers.screen` so claude
# gets five tools: `screenshot`, `type_text`, `key_press`, `mouse_move`,
# and `mouse_click`.
# gets three tools: `screenshot`, `type_text`, and `key_press`.
#
# `screenshot` and `type_text` work out of the box (grim + wtype, both
# pure Wayland clients). `key_press`, `mouse_move`, and `mouse_click`
# require the ydotoold daemon, which injects events via `/dev/uinput`
# at the kernel level — enable it by setting
# `hyperhive.gui.screenInput = true`.
# All tools use Wayland protocols mediated by the Weston compositor —
# no `/dev/uinput` or kernel-level injection needed. `grim` takes
# screenshots; `wtype` handles both text input and key/modifier combos
# via the `zwp-virtual-keyboard-unstable-v1` protocol.
{
pkgs,
lib,
@ -17,20 +15,6 @@
...
}:
{
options.hyperhive.gui.screenInput = lib.mkOption {
type = lib.types.bool;
default = false;
description = ''
Enable mouse and keyboard injection via ydotool + the ydotoold
daemon. Requires `/dev/uinput` device access inside the container
(the host must bind it in via `extraSystemdProperties` or
`systemd.nspawn.<name>.filesConfig.Bind`). When false,
`screenshot` and `type_text` still work; `key_press`,
`mouse_move`, and `mouse_click` return an error from ydotool
until ydotoold is running and `/dev/uinput` is accessible.
'';
};
config = lib.mkIf config.hyperhive.gui.enable {
# Register the screen MCP bridge so claude gets the screen tools.
hyperhive.extraMcpServers.screen = {
@ -38,27 +22,11 @@
args = [ ];
};
# grim: Wayland screenshot; wtype: text/key input (no daemon).
# ydotool: mouse + key injection via uinput (needs screenInput).
# grim: Wayland screenshot; wtype: text + key input via virtual-keyboard
# protocol (compositor-mediated, no /dev/uinput required).
environment.systemPackages = [
pkgs.grim
pkgs.wtype
]
++ lib.optional config.hyperhive.gui.screenInput pkgs.ydotool;
# ydotoold — uinput event injection daemon. The socket lands at
# /tmp/.ydotool_socket by default; ydotool picks it up
# automatically. Only started when screenInput is enabled.
systemd.services.ydotoold = lib.mkIf config.hyperhive.gui.screenInput {
description = "ydotool input injection daemon";
wantedBy = [ "multi-user.target" ];
after = [ "local-fs.target" ];
serviceConfig = {
ExecStart = "${pkgs.ydotool}/bin/ydotoold";
Restart = "on-failure";
RestartSec = "2s";
SyslogIdentifier = "ydotoold";
};
};
];
};
}