Library launches opened on the operator's head — nothing ever focused the streamed one #283
Assignees
FernandoMX
HitFrostbite
enricobuehler (Enrico Bühler)
kszlim
luxus
nannku
neoney
sassycrown
Clear assignees
FernandoMX
HitFrostbite
enricobuehler (Enrico Bühler)
kszlim
luxus
nannku
neoney
sassycrown
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: unom/punktfunk#283
Reference in New Issue
Block a user
Reported on Hyprland: "anything from the library opens on my main display instead of the virtual screen."
The mechanism
Hyprland and wlroots/sway are the two EXTEND-only backends: the per-session headless output is added beside the operator's heads and nothing promotes it (KWin and Mutter promote theirs primary; gamescope nests the app in its own compositor). Both compositors open a new window on the focused monitor — and
hyprctl output create headless/swaymsg create_outputdoes not focus what it creates.So focus stayed on the physical head, and every library launch opened where the client could not see it, with the stream showing a bare desktop.
Nothing in the tree moved or focused a window before this: zero hits for
focusmonitor,movewindow,windowrule,activatewindow, and no post-map correction anywhere.The launch path's own doc comment asserted the opposite, which is probably why this survived so long:
False for wlroots, which is extend-only and never promoted — and Hyprland was not in the list at all. Corrected.
The fix
hyprland.rs/wlroots.rs—focus_output()claims focus for the head we are about to stream, at create.routing.rs—focus_streamed_output(compositor, name), split out solibrary.rsdoesn't reach into a private backend module. Its match is exhaustive on purpose (no_arm) so a backend added later has to decide rather than silently inherit "no focus". Gated on each backend's own managed-name predicate, which is what keeps the mirror-pin case out: there the streamed head is a physical monitor the operator is using, and stealing its focus is a change nobody asked for.library/launch.rs— re-asserts focus immediately before the spawn. Create-time focus alone isn't enough: the portal handshake (≤20 s), the encoder build and the first frame all sit between capture bring-up and the launch. The name comes from the same slot the absolute-input pointer is bound to, so focus and cursor land on one head by construction rather than by two independent guesses.Validation
Full gate green —
clippy --all-targets -D warnings, a plain build,pf-vdisplay239 tests,punktfunk-host library::56 tests,cargo fmt --check.On glass against real Hyprland 0.56.2 (headless instance in a transient scope, 2026-08-17):
WAYLAND-1 focused=Trueoutput create headless PF-probe-1ok— butfocused: false← the bug, reproduced directlydispatch focusmonitor PF-probe-1ok, andfocusedmoves to itPF-probe-1← end-to-end proofControl worth keeping: the bare
hyprctl focusmonitor <name>(nodispatch) answersunknown requestat exit 0.focusmonitoris a dispatcher.hyprctl_dispatch'sunknownmarker is what turns that into an error — without it the wrong argv would have read as success. Both argv shapes are pinned by tests, because that is exactly where this can silently regress: sway spells its commandfocus output <name>with the noun second, unlike every other call in that file (output <name> mode|enable|unplug).⚠️ The sway half is not on-glass — its shape is from sway's documented command surface, and no box in the fleet runs sway. It is the safer of the two to get wrong (sway exits non-zero on an invalid command, so a bad shape surfaces as the logged warning rather than a silent success), but the shape wants an on-glass confirmation.
What this deliberately does not do
primary/exclusivetopology remain no-ops on both backends — accepted, echoed by the management API, dropped with a warning. That is sweep 13.18, whose shipped fix was the warning only; the behaviour was never implemented. Noteresolve_topologymakesAutoresolve to Exclusive on any host without aPUNKTFUNK_COMPOSITORpin, Hyprland included — so the default policy on a Hyprland box is an Exclusive that does nothing. Focus is the extend-topology answer to window placement and leaves the operator's monitors lit; the real thing is specified indesign/display-management.md§5.2 (wlroots:primary→ log + extend,exclusive→ disable physicals, re-enable on teardown).Also still open, and documented rather than fixed:
library/launch.rs'splugin_recipe(&entry).map(|l| l.command)drops the plugin's resolvedcwd—launch_session_commandhas nocwdparameter at all — so emulator plugins that resolve cores relative to their install dir get the host's cwd.Docs
hyprland.mdandsway.mdgain a Window placement bullet; thevirtual-displays.mdtopology table gains the note that "treated as Extend" is backed by a focus claim (and what follows from it being focus rather than promotion);troubleshooting.mdgains an entry in the reporter's own words.Reported on Hyprland: "anything from the library opens on my main display instead of the virtual screen." Hyprland and wlroots/sway are the two EXTEND-only backends: the per-session headless output is added *beside* the operator's heads and nothing promotes it (KWin and Mutter promote theirs primary; gamescope nests the app in its own compositor). Both compositors open a new window on the FOCUSED monitor, and `output create headless` / `create_output` does not focus what it creates — so focus stayed on the physical head and every library launch opened where the client could not see it, with the stream showing a bare desktop. Nothing in the tree moved or focused a window: no `focusmonitor`, no window rule, no post-map correction. The launch path's own doc comment asserted the opposite ("KWin / Mutter / wlroots ... the per-session virtual output is promoted primary, so a plain spawn lands the app on the streamed output") — false for wlroots, and Hyprland was not listed. Both backends now claim focus for the head they are about to stream, and the launch path re-asserts it immediately before the spawn: the portal handshake, the encoder build and the first frame all sit between capture bring-up and the launch, so a create-time focus alone can be lost before the app ever starts. The name comes from the same slot the absolute-input pointer is bound to, so focus and cursor land on one head by construction. `focus_streamed_output` is gated on each backend's own managed-name predicate, which keeps the mirror-pin case out — there the streamed head is a physical monitor the operator is using, and stealing its focus is a change nobody asked for. Its match is exhaustive on purpose so a backend added later has to decide rather than inherit "no focus" silently. This does not implement the `primary`/`exclusive` topology levels on these two backends — they are still accepted, echoed by the management API and dropped with a warning (sweep 13.18, whose shipped fix was the warning only). Focus is the extend-topology answer to window placement and leaves the operator's monitors lit. Tests pin the two argv shapes, which is where this can silently regress: `focusmonitor` is a dispatcher and needs `hyprctl dispatch`, while sway spells its command `focus output <name>` with the noun second, unlike every other call in that file. Both wrong forms are rejected at exit 0 / with an error whose only field symptom is the bug itself.