383. The desktop is asked whether to move less

Date: 2026-09-17

Status

Accepted. Closes book/src/TODO.md’s “Reduced motion is obeyed but not detected”; the density half of that entry is answered separately below.

Context

§13 names four accessibility switches and the toolkit obeys all four. Three are the application’s to set and one is not: reduce motion is a desktop setting, and renderer.reducedMotion(true) was the only way in — so every application had to find the answer for itself, on three platforms, which is exactly the work a toolkit exists to have done once.

ADR-0322 answered the other desktop preference with one SDL call and was glad not to write three platform integrations for a boolean. There is no SDL_GetReducedMotion, so this is those three integrations — taken because the alternative is every application writing them.

Decision

Ask each platform directly, through FFM, against a library the process already has; cache the answer; obey it.

• No native build. Nothing is compiled, nothing is added to the superbuild and nothing new ships: each library is dlopened by name at run time, which is what SDL itself does with D-Bus.
• Linux asks the specified key first — org.freedesktop.appearance’s reduced-motion, a uint32 whose 0 means “no preference” — and falls back to GNOME’s org.gnome.desktop.interface/enable-animations, which is the same question the other way round.
• dbus_message_append_args is variadic and is not used: the iterator API says the same thing with fixed signatures. The reply is a variant wrapping a variant on every portal implementation tried, so the reader unwraps until it finds a value.
• Three answers, not two. UNKNOWN is what a missing library, a missing portal, a missing key, a refused call and an unexpected type all produce, and it is not an instruction: a renderer built from it animates normally.
• Backend.reducedMotion(), Window.reducedMotion() and Host.reducedMotion() carry it, and Launcher applies it — unlike the theme, which the toolkit deliberately does not act on. A colour scheme is a matter of taste; an accessibility preference is not.
• -Dgoldberry.motion.reduced=reduce|full overrides it, for a test, a screenshot, and a desktop whose answer is wrong.
• The holders live in a package that is not exported, which is ADR-0173’s rule and what ExportedSurfaceTest enforces: what leaves :natives is an enum with three constants in it.

Consequences

• Asked once per process, and a change mid-session is not obeyed until the application restarts. Listening means a D-Bus main loop on Linux and an observer on each of the others — a much larger thing than one 200ms read, and worth its own decision if anybody asks for it.
• The Linux path is verified against a real portal on this machine, in all four of its outcomes: a uint32 preference, a uint32 zero, a GNOME boolean and a namespace nobody serves. The Windows and macOS paths are written from their documented APIs and are exercised by CI on those runners; each of them fails to UNKNOWN, which is the behaviour the toolkit had before this existed.
• Density is answered, not built. The TODO entry pairs the two, and they are not alike: no desktop has a density setting to read — not GNOME, not Windows, not macOS — and §1.3’s compact mode is an application’s own choice about its screens. There is nothing to detect.