GIMP has merged an automated test suite that starts the real application in a headless Linux session and drives its menus and dialogs through accessibility data, and the suite now runs in CI on every merge request. The mechanism is documented and reproducible, but coverage stands at 13 short checks, and the project’s own developer guide says mouse drags on the canvas, which is to say paint strokes, do not yet work reliably.

What was merged
The work is Harsh Verma’s, reviewed on the GIMP side by maintainer Jehan and merged through merge request !2909, which adds the UI testing platform. The tests are Python scripts built on dogtail, a framework that locates buttons and menu items by role and name in the AT-SPI accessibility tree instead of matching pixels. Verma also wrote a dogtail feature for running an application on a machine with no screen, which was contributed upstream and reviewed there by dogtail developer Michal Odehnal.
According to the developer guide for running and writing GIMP UI tests, each test starts a private D-Bus session, a private accessibility bus and a headless mutter, launches GIMP inside it, and tears everything down afterwards. Tests run one at a time with a default timeout of 240 seconds, are invoked with meson test -C _build --suite ui, and execute in the gimp-debian-nonreloc CI job. They are switched off automatically on Windows and macOS, because dogtail supports neither.
What the tests check
The checks cover startup and the main window, the Welcome dialog, creating an image at a set size, exporting a PNG, swapping and typing in colours, adding and deleting a layer, undo and redo, zoom, selecting a tool and using Bucket Fill, saving a Preferences change, and flattening an image. The GIMP announcement of improved graphical interface testing counts 13 tests and itemises 12; the authoritative list is the ui_tests array in app/tests/uitests/meson.build.
What is new compared with GIMP’s older tests
GIMP has had a C file called test-ui.c since 2010, and it was still in the 3.0 source tree. Those tests ran inside a partially initialised GIMP, synthesised events with GTK’s test helpers, and were wrapped in xvfb-run for headless use. The new suite works from outside the process against the binary a user would launch, the direction a 2023 issue in the project’s tracker had proposed. There is a side effect: a widget dogtail cannot find is a widget a screen reader cannot find either, so each new test doubles as an accessibility check. Projects that ship frequent bug-fix updates, as LibreOffice did with the 45 fixes in 26.8.3, know how much of that work is catching regressions in dialogs nobody reopened.
Support for both X11 and Wayland is claimed in the announcement. The guide describes only the headless mutter session and does not say how an X11 path is exercised. Nothing has been measured in public: there are no pass rates, run times or flakiness figures.
A whole-program test suite that cannot paint a stroke
The announcement says UI tests check the program as a whole, the way a person uses it. For an image editor, that means dragging a brush across a canvas, and the guide states plainly that drags do not work reliably; a single click does, which is why Bucket Fill is the only canvas operation in the list. Plug-in dialogs, including the PNG export options, run in a separate process and do not