Restores Archie, the server monitor and the full-width bar surface. None of
today's changes fixed the stutter, and the surface refactor regressed it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reverting the diagnostic. The narrow surface is still the right destination
but it needs the frame band drawn without adding a second full-screen
surface, which is what the reverted attempt got wrong.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The bar's PanelWindow was screen-sized because the frame band and every
panel drew into it. Hyprland blurs that namespace, and blur is recomputed
whenever what is behind the surface changes — with a game underneath, that
is every pixel every frame. Measured at 3440x1440 it cost a third of the
card; the diagnostic shrink took game-running GPU 57% -> 38% and made the
stutter unprovokable.
So the surface is now barWidth at rest and grows to the screen only while a
dropdown, the toast or the launcher is up. exclusiveZone is unchanged, so
windows never reflow on resize. chromeSdf's cutout derives from
screen.width rather than the surface width, so the shader is identical at
both sizes and the narrow surface simply clips it.
The frame band moves to its own background-layer surface, below windows and
matched by no blur rule. Its shader hides (rather than the window
unmapping, which would flicker) while the bar is wide, since chromeSdf then
draws the frame itself and melts panels into it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Archie goes for good — he'd have had to move to the background layer with
the frame band, and that isn't where he belongs. History has him.
The implicitWidth change is a diagnostic, not a design: it clips the frame,
panels, toast and launcher, which all draw into this one screen-sized
surface. Every cost measured today (blur 84->62, SDF 27->17) scales with
that area, so this tests whether shrinking it is worth a real refactor
before committing hours to one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The SSH qs-stats stream and the router ping both tick every 2s, and each
tick rewrote six Text items in the bar. That surface is the whole screen,
so every tick repainted 3440x1440, reran the SDF shader and made Hyprland
recompute the blur behind it — one late frame, twice a second. Suppressing
both processes was the only change all day that held the frametime graph
flat under deliberate provocation.
Archie was measured innocent and comes back. The OSD anchored its y to the
glance widget, so it now derives from barBottom instead.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
As a user sheet it sorted below userChrome.css and lost every !important to
the copy parsed at startup. Agent sheets sort above ua.css; @-moz-document
keeps that origin off web pages.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The watcher not landing left Zen with no theme at all. Keep the startup copy
so the floor is the old behaviour, and let the watcher fight for the live
update on top.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gecko has no userChrome.css reload, so wrap zen with a mozilla.cfg script
that registers chrome/zen-theme.css through nsIStyleSheetService and
re-registers it under a cache-busting URI when its mtime moves.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
theme-apply sent reload-config to com.mitchellh.ghostty only, and its
comment claimed the scratchpad 'is the same process'. It is not — that
is the whole point of the --class in settings/hyprland.nix. Ghostty's
--class IS the GTK application id, which is how it forces a separate
instance, and a separate instance owns its own bus name and object path.
So the scratchpad never got the reload and kept the old palette until it
was killed.
Now both are told, scratchpad first and best-effort, since it only
exists once SUPER+Return has been pressed this session.
info/ok/alt were fixed blue/green/purple pulled only 25% toward the
accent, so ram, down and up looked imported on any warm palette. They
exist to tell six stats apart, not to warn, and their captions already
say which is which — so they now tint 55% toward the accent and trade
separation for family. err and warn keep the old 0.25: a critical
temperature is not the place to prioritise harmony.
Not hue-rotation off the accent, which was the obvious approach and is
unstable: it inherits the accent's saturation, collapsing the three into
near-identical greys on a muted wallpaper (minsep 0.107 on 5y39p9) and
blowing them out to neon on a vivid one (0.455 on d8wxw3). Tinting holds
~0.175 on every palette tested.
cpu drops the raw accent for base05 — it was the one fully saturated
value in the column.
base03/base04/base05 were three independent slots off the wallpaper, not
one ink at three strengths. On wallhaven-k8dqr6 the raw ladder is red,
blue and cream; on 5y39p9 base03 is a saturated gold against pale cream.
That is why muted labels read as a second accent and why the shell frame
never quite matched the window border beside it.
base04 and base03 are now base05 mixed toward base00. The two factors and
the mix live in settings/ui.nix and are used by all three consumers:
Theme.qml, theming.nix's hypr.lua (applied at runtime) and hyprland.nix's
build-time fallback. Same trick base01/base02 already used for surfaces.
Verified on 5y39p9: hypr.lua now emits aca888 / 727761 from base05 f1e2b6
against base00 0a1f1c, matching what inkStep computes in the shell.
All three were base03, which is its own wallpaper slot rather than a
step of the ink: on 5y39p9 it is a saturated gold against pale cream,
on k8dqr6 an outright red against cream. So muting the speaker changed
the icon's colour rather than its weight, and the launcher's host and
placeholder read as a second accent.
Dimmed base05 instead, same idiom as the workspace dots.
Focused was base05 while the rest were base03/base04 — three palette
slots for one control, which reads as a second colour in a bar that no
longer has one. All base05 now; unfocused dots step back to 0.4 opacity
and hover to 0.8, so state still shows without a colour change. Shape
already carries focus at 18px against 6px.
Username, category row icons, the pinned-app fallback glyph and the
pin/unpin menu row were all accented. Same problem as the bar column:
chrome in a list of chrome, picked out for no reason.
Left the calculator's equal and the answer card's sparkles accented —
those mark what KIND of result a card is, which is the one job the
accent still has in there.
Logo, workspace pill and wallpaper button are three buttons in a column
of buttons; two of them were accented, which put the only saturated
colour on screen in the corner. All base05 now.
Focus does not need the colour: the focused pill is 18px against 6px,
so shape already says it. Launcher header logo follows the bar's, since
it is the same mark.
Targeting a fixed saturation was the wrong variable. The generator gets
the accent's STRENGTH right — it scales it to the palette, so a cyan
scheme whose base05 is 55ffff wants a vivid accent and one whose base05
is the cream f1e2b6 wants a quiet one. Forcing S=0.60 is what made the
amber bar shout with a sharp orange logo in a row of muted cream.
What the generator gets wrong is the HUE: it fills each slot
independently, so on wallhaven-5y39p9 every accent slot comes back
orange except base0D, which comes back blue-grey. Accent now takes the
hue from whichever slot carries the most chroma and keeps base0D's own
saturation and lightness. 5y39p9 becomes a muted tan; k8dqr6 goes back
to its rose instead of the coral the last rule picked; the rest are
unchanged or nudged a few degrees.
Wallpaper button idles at base05 like every other opener in the column,
instead of being the one dim icon that brightened on hover.
Every icon in the bar was already bound to a Theme colour — the problem
was which one. base0D is the accent for 33 sites and every status colour
harmonises toward it, but the generator fills each slot independently:
on wallhaven-5y39p9 all six accent slots come back orange (H=0.08)
except base0D, which comes back blue-grey. Hence a blue NixOS logo and
a blue workspace pill in an amber bar.
Accent now auditions all six slots and takes the one nearest S=0.60.
Nearest, not max: max picks the neon fe5162 on wallhaven-k8dqr6, close
enough to err that errors stop reading as errors. Only 5y39p9 and
k8dqr6 change; the other four keep the colour they already had.
Renamed base0D -> accent since it no longer names the slot it comes from.
Toast is back to a fixed 320 — it is the same card wherever it appears
and should not restyle itself around whatever menu is open.
The residual indent was never the width step: two rounded boxes stacked
on one edge each contribute a corner that curves away over exactly one
radius, so the outline pinches and bulges at the joint. The curves only
cancel at 2*radius of overlap, where the corner circles coincide.
Burying that deep would drive the cards together, so a fused panel grows
by what it buries. Content is pinned 12px from the panel top, not
centred, so the extra height lands entirely below the cards and the
clearance stays 12px. Geometry now derives from baseH, keeping height
(which depends on fused) out of the position that decides it.
The warp was a width step. Panel and toast are both left-anchored to the
bar, so different widths put a step on the shared right edge, and smin
fillets a step into an S-curve waist.
Toast now takes the open dropdown's drawn width, keyed off its declared
fullWidth so nothing closes a loop through the toast's own height. Below
a 200px floor there is no width worth matching (the power menu is a 64px
icon strip), so those stand off with a gap as before.
The bump was two animations stacked: _dropdownRect is positioned off
chrome.y, so chrome already eased the panel to wherever y resolved, and
the Behavior added last commit put an OutExpo behind an OutExpo.
Panels now land ON the toast rather than a panelGap short of it, buried
half a radius so the shader smins them into one outline with a shared
waist. Cards inside still clear each other by ~14px.
The toast dodging open dropdowns meant it drifted around the column
depending on which menu was open. Invert it: the toast keeps its band
and melts into the frame as before, and a dropdown treats a live toast
as a shortened screen — its clamp ends short of the toast, so a menu
that would collide opens higher instead.
Snapping still tests against the frame, so a panel stopping above a
toast floats clear of it rather than fusing to its edge.
