desktop

Use the AT-SPI accessibility bus to read and activate widgets by name instead of pixel coordinates. The gui script lists apps on the bus, dumps interactive widget trees with coordinates and state flags, and clicks or types semantically without moving the pointer or stealing focus.

Run the program in a headless tmux session, send input with tmux send-keys, and read output as text with tmux capture-pane. Use capture-pane -e with reverse-video regex to detect highlighted rows before destructive keystrokes, and always kill the session when done.

wlrctl only supports relative movement, so absolute positioning requires pinning to an inert corner first. The mouse script pins to the bar's right end, clamps targets 2 px inside the screen, and uses ydotool for held-button drags, avoiding niri's hot corners and edge strips.

Yes, playwright-cli attach --extension connects to the running browser through the Playwright Extension with real cookies and logins. Set PLAYWRIGHT_MCP_EXECUTABLE_PATH=/usr/bin/chromium-browser, and always detach rather than close so the user's tabs survive.

niri's scrolling layout does not expose window x-positions through IPC, so clicks computed from assumed tile coordinates miss silently. Fullscreen the window first with niri msg action fullscreen-window, then click coordinates taken from a grim -s 1 region capture.

Use tesseract OCR whenever the answer is text on screen, since it returns plain text without vision tokens. Capture at native scale rather than -s 1 because tesseract needs the full-resolution pixels for small fonts.