The post-animation #238 package was also captured after all 13 checks passed
and it merged as 27fd8bf. Raw paired frames and reports
use candidate 571ff66, run 37177834905, against the same niri/Fedora fixture.
Endpoint client sizes remain 616×736 / 826×736, with 0.0% differing client
content pixels. Whole-work-area rest/preset/settled results remain
6.7% / 5.3% / 5.5%; the nominal 100ms transient differs by 11.5% (mean 9.48).
At that sample niri captures during 104–153ms with the right edge at 493px;
Tuna Desktop captures during 101–135ms with it still at 422px. Both first reach a
one-pixel settle in the nominal 400ms sample: niri 401–446ms, Tuna Desktop 415–467ms.
The nominal 300ms niri capture took 300–537ms, illustrating why these are
interval samples rather than precise animation-duration claims. Both are
settled by 600ms. This records a visible transient difference after #238;
it does not claim exact timing parity. The current comparison PR still
requires its independent CI recapture and full acceptance gate.
The retained pre-animation baseline follows for comparison.
The comparison uses niri 26.04 from Fedora 45 and the fully green Tuna Desktop
PR #233 package, before #238's column animation changes. It is a measured
baseline, not evidence for the later animation implementation. The package
came from run 37169850678,
head 5b76d44d9e6282b0b33d47e74eb01ed52394689e; its check, GTK proof,
scroll proof and packaging jobs all passed. The exact provenance and raw
frames are in baseline-233.
Both compositors run nested in the same Fedora container at 1280×800,
scale 1, with the same GTK 4.24.1 / libadwaita 1.10 runtime and the same
tuna-test-window.py windows: Scroll One, Two and Three. Each uses the
same header, text and list content. Niri reserves 32px at the top to match
Tuna Desktop's shell panel, uses 16px gaps, half-width columns, a 4px #7fc8ff focus
ring and center-focused-column "never". Its default spring is retained.
The reference config is validated by
niri before captures.
At rest, both configure columns to 616×736. After the focused column's next preset, both configure it to 826×736. Following focus back to the middle column leaves that widened column intact. Tuna Desktop's state confirms 16px top/bottom/side gaps and offsets 632 → 842 → 632; niri's IPC confirms matching client sizes, and the frames confirm the positions. No frame is resized to make this comparison.
scripts/lib/tuna-parity-compare.py measured the work area, excluding only
the 32px shell panel:
| State | Mean difference | Pixels off >24 |
|---|---|---|
| Strip at rest | 1.95 | 6.7% |
| Wider preset, settled | 1.56 | 5.3% |
| Mid-scroll, nominal 100ms | 6.46 | 8.6% |
| Scroll settled | 1.58 | 5.5% |
The main settled difference is the desktop visible through the gaps: niri's gray backdrop and Tuna Desktop's GNOME blue backdrop. Tuna Desktop also leaves a GTK keyboard-focus outline on the header menu button. Inside the settled window content, excluding the header and gaps, the three measured crops have 0.00 mean difference / 0.0% off. This confirms matching dimensions, content layout and fonts in that runtime; it does not claim the complete desktops have identical pixels.
The difference frames place niri above Tuna Desktop and amplify their difference four times below. The raw niri frame and Tuna Desktop frame are also retained, with IPC/state JSON alongside each frame.
Timing samples repeat a fresh right-to-left transition for each delay; PNG compression cannot push a later sample past its deadline. They retain both the capture start and completion timestamps. In this software-rendered baseline, the nominal 100ms captures span 100–153ms for niri and 100–117ms for Tuna Desktop. Their visible middle-column right edges are 536px and 494px respectively, starting at 422px and ending at 632px. A one-pixel settle is first observed in the nominal 300ms niri sample (300–383ms) and 400ms Tuna Desktop sample (400–432ms); both settle by 600ms. These are empirical sampling intervals under load, not precise animation duration measurements. The transient pixel difference includes that input/render timing uncertainty. Raw samples remain available for checking the interpretation.
The deterministic spring tests separately check the formula and settle threshold: critically damped, stiffness 800, epsilon 0.0001, matching niri 26.04's defaults.
238 adds that spring to column width and position. The post-animation
package measurements above establish its observed behavior; #233's captures remain evidence only for the earlier baseline.
The GTK proof now checks insertion right of focus, column movement and closure, wheel scrolling, overview previews, workspace thumbnails, Alt+Tab and client maximize/fullscreen/edge-snap requests. The output-failover wire test checks that surviving columns resize for the new primary output without losing focus or workspace membership. Current differences remain explicit: wheel/touchpad scrolling is direct; maximize, fullscreen and edge tiling keep a window in the strip; there is one strip using the primary output area, rather than niri's independent layout on every output.
Columns, drops and resizing
Columns hold several windows, following niri 26.04's
src/layout/scrolling.rs
and its default binds:
- Super+[ / Super+] (
consume-or-expel-window-left/right): a lone window joins the bottom of the neighboring column on that side; a window sharing a column leaves it for a new column on that side with the column's width (RemovedTile::width). Super+, / Super+. take the next column's top window into the focused column and send the bottom one out to the right. Super+Up / Super+Down move focus inside a shared column, with Shift moving the window (niri binds these on Mod+Ctrl; Tuna Desktop keeps Shift for its existing column moves). GNOME 51 binds none of<Super>bracketleft,<Super>bracketright,<Super>commaor<Super>period(gsettings-desktop-schemas 51.0 wm keybindings, mutter and gnome-shell 51.0 schemas); its<Super>Up/<Super>Down(maximize / unmaximize) and<Super><Shift>Up/<Super><Shift>Down(move to monitor) are only replaced in strip mode, inside a shared column, where forced strip membership already made them no-ops. - Heights follow
update_tile_sizes: windows share the column's height by weight, a window whose minimum is not met takes it, and one resized window keeps a fixed height clamped to what the others' minimums leave. The first fixed height turns the others into weights from their current heights over the median (convert_heights_to_auto). Column widths clamp to the largest member minimum and smallest member maximum, the minimum winning; a client changing its limits re-sizes its column. - A drop lands at the boundary nearest the pointer, aiming at the middle of a
gap (
insert_position): a column's edge makes a new column, a window's top or bottom edge inside the column under the pointer joins that column. - An xdg resize, or the GNOME modifier resize (
mouse-button-modifierwith the middle button, or the right one underresize-with-right-button) on the pointer's outer third of a window (niri's Mod+right-drag,resize_edges_under), changes the column width from one side and a window's height from its top or bottom inside a shared column; the topmost window's top edge does nothing (interactive_resize_update). The strip keeps GNOME's modifier and buttons rather than niri's fixed Mod+RMB, so the two modes agree.
Remaining differences: a lone window stays full height (niri lets a vertical
resize shorten it, and has no reset bind here); the dragged window keeps its
slot until the drop instead of leaving the layout, so both sides of that slot
name it; drop bars are 8px rather than niri's 300px hint areas; dialogs with
attached children only drop as new columns; interactive sizes floor at
120×80px; and there is no tabbed column display, window height presets or
Mod+Ctrl+R height reset.
To reproduce without compiling a local GTK shell:
scripts/tuna-niri-reference --out /tmp/fresh-niri-reference
scripts/tuna-niri-compare --candidate /tmp/extracted-tuna-package --out /tmp/fresh-niri-pair
The candidate directory must contain usr/bin/tuna-compositor and
usr/bin/tuna-shell-gtk. CI extracts its current Arch package, captures
both compositors, checks endpoint geometry/content and records timing and
pixel reports in the niri-parity artifact. The broader interactions remain
in gtk-shell; both jobs must pass before the comparison change is merged.