Audit P1: 5 of 7 layout-property animations were transition-[width]
or transition-[left] — these trigger reflow + repaint every frame.
Convert to transform-based equivalents that run on the compositor.
Subscription.tsx — 3 progress bars + 1 toggle:
• Device usage bar (track w-16) — width % → scaleX with a 0.0625
floor (= 4px on 64px track) preserving the minWidth: 4px guard
when connectedDevices > 0.
• Traffic-purchase progress — absolute inset-0 + width % →
transform: scaleX with origin-left, gradient stretches with the
fill same as before.
• Subscription period progress — same conversion.
• Autopay toggle thumb — left: 3↔26px → translateX(0↔23px) with
fixed left: 3px. role='switch' + aria-checked unchanged.
globals.css — .progress-fill utility class:
• transition-property: width, background-color → transform,
background-color. Width set to w-full + origin-left so consumers
just set transform: scaleX(0..1). Currently has no .tsx consumers
(grep clean) but the API stays valid for future use.
Skipped (bounded one-shot, transform alternative has tradeoffs):
• Info.tsx + InfoPageView.tsx accordion height. Animating from 0
to content-driven scrollHeight needs layout one way or another —
grid-template-rows 0fr/1fr and max-height also layout-class
properties; transform: scaleY squishes text contents. Cost is
bounded: one animation per user click, not continuous, not on
critical path. Separate work if priority changes.
Visual parity: scaleX renders identical geometry to width % on the
inner fill. box-shadow on the device bar (0 0 8px 40)
scales proportionally — tiny visual difference at full width, none
visible at partial widths.
The #FFB800 (23 hits) and #FF3B5C (11 hits) raw hex literals across
Subscription / SubscriptionPurchase / TrialOfferCard / useTrafficZone
all represent the same semantic: 'expiring soon' and 'expired'.
Distinct from warning-500 (#f59e0b) and error-500 (#ef4444) — the
subscription timeline needs a sharper, hotter signal than the
system-wide warning/error roles.
Add two semantic tokens:
--color-urgent-400: 255, 184, 0 (#FFB800)
--color-critical-500: 255, 59, 92 (#FF3B5C)
Exposed via Tailwind as urgent.400 / critical.500. Sweep all
single-and double-quoted hex literals to rgb(var(--color-...))
form so style attrs, SVG stroke=, and ternary string values all
resolve through CSS vars (auto-respects future light-theme
variants).
Left as-is: 4 hex appearances inside gradient pair strings —
linear-gradient(135deg, #FFB800, #FF8C00) — where the partner
color (#FF8C00 / #FF6B35) isn't in the token set. Partial-replace
inside gradient strings would just add noise; full extraction of
the orange-ramp companion is out of scope for this pass.
Per /impeccable audit: Toast.tsx had a 4px border-l accent on top of
a 1px full border — the explicit 'side-stripe' absolute ban. Toast is
on screen after every notify.* call, so the violation repeated
constantly.
Polish pass:
• Replace 'border border-l-4 border-dark-700 ${style.border}' with
a single tinted full border (border-success-500/40, etc). Semantic
still carries through the colored icon box + matching border tint.
• Drop bg-tint flood — the icon box at 40px is enough at this scale.
• Reduce shadow weight (shadow-2xl shadow-black/50 → shadow-xl
shadow-black/30). Floating layer needs elevation, not drama.
• Soften interaction scale (1.02 hover → 1.01, 0.98 active → 0.99)
to match a calmer overall feel.
• Thinner progress bar (h-1 → h-0.5) — the countdown is ambient, not
primary content.
A11y:
• Container: role='region' + aria-label + aria-live='polite' so SR
announces incoming toasts without stealing focus.
• Toast: role='alert' for type=error (interrupts SR — appropriate
for failures), role='status' for everything else.
• Toast: tabIndex=0 + Enter/Space to activate, Escape to dismiss —
was click-only before, useless on keyboard.
• focus-visible: 2px ring at accent-500/60 with offset for visibility
on the dark backdrop.
• Icon box aria-hidden — it's decorative reinforcement, the role +
text content is what SR reads.
• Progress bar aria-hidden — visual countdown only.
shrink keyframe already used transform: scaleX (no layout thrash)
and slide-in-right keyframe already used ease-out-expo — both
align with motion laws as-is, kept unchanged.
Per /impeccable audit findings: admin overview pages concentrated
AI-aesthetic tells. Apply the 'quieter' refinement — strip the loudness,
keep the structure, preserve state-conveying motion.
AdminPanel.tsx
• Hero h1 — drop bg-clip-text + 3-stop gradient. Solid text-dark-50
(the absolute ban applied; type weight + size carry hierarchy now).
• 'Online' dot — drop the 10px-radius glow shadow; the dot + pulse
animation alone signal liveness.
• GlassCard — drop the 3D perspective-tilt mouse-tracking (decorative
motion that didn't convey state), drop the staggered enter
animation (no orchestrated page-load), drop the top hairline-glow
gradient, drop the inner shine overlay + drop-shadow-sm on the
count badge. Keep backdrop-blur on the card body itself — the
component is literally named GlassCard, the glass identity is the
point. Hover state stays.
• Stats-bar pills (2 places) — drop backdrop-blur. Tiny data cells
don't need glass.
• Search empty-state icon container — drop backdrop-blur. Decorative
on a non-interactive surface.
AdminDashboard.tsx
• Strip backdrop-blur from all 9 stat / node / panel cards. They
were using it as ambient noise; solid bg-dark-800/{30,50} reads
better and renders faster.
• Daily-sales progress bar — drop the from-accent-600 to-accent-400
gradient (and hover gradient swap). Solid bg-accent-500, hover
bg-accent-400. Width still animates (functional state change).
• Animate-pulse on the connected-node status dot kept — it conveys
'live' state, which is exactly what motion is for in product UI.
Last and biggest tab — extracted via view-facade pattern (like Info).
JSX moves into components/admin/userDetail/SubscriptionTab.tsx;
state, queries, mutation handlers stay in parent because most are
shared with other tabs (panelInfo + devicesQuery used by Info,
inline-confirm armer used by destructive actions, etc.).
The new file is 1223 lines — over the soft cap, but cohesive: one
tab with multi-sub list, sub-detail view, traffic packages,
devices CRUD with inline rename, panel info/config, live traffic,
node usage, and request history table. Sub-extracting (Devices /
RequestHistory / PanelLinks) is the next step but not free —
each adds prop-passing boilerplate.
Wide prop interface (~50 props) but every prop is a single pass-
through line at the call site. No state moves; just JSX reroute.
Parent imports trimmed: DEVICE_ALIAS_MAX_LENGTH, getFlagEmoji,
createNumberInputHandler, getCountryFlag, PlusIcon, MinusIcon,
StatusBadge — all gone, subscription tab was their last consumer.
RefreshIcon kept (page header reload button still uses it).
Net to AdminUserDetail.tsx: down to 970 lines (-1022). Total
across the 7 tab extractions: 3820 → 970 (-2850, -75%).
The info tab needed a different shape than the prior 5 — its
panelInfo, promoGroups, editingPromoGroup, action handlers
(reset-trial, disable, full-delete) are SHARED with the subscription
tab and with the inline-confirm armer up in the parent. Moving any
of that into the tab would either duplicate the query/state or
fork the confirm state. So the new InfoTab is a 'view facade':
state and handlers stay parent-owned, the JSX moves out.
Prop surface is wide (~25 props) but every prop is a single
'pass through' line at the call site — no logic shifted. Each
handler / state slot still lives in exactly one place; the tab
only renders.
Bonus: the parent's formatWithCurrency / useCurrency import is
dropped — info tab was the last consumer in the page body.
Net to AdminUserDetail.tsx: down to 1992 lines (-403). Total
journey across the 6 tab extractions: 3820 → 1992 (-1828, -48%).
Only the subscription tab JSX (~1043 lines) remains inline.
Pulls the tickets tab into components/admin/userDetail/TicketsTab.tsx:
- ticketsQuery (list)
- selectedTicketId / selectedTicket state + detail loader
- reply form state (replyText, replySending)
- status-change action handler + actionLoading
- auto-scroll messagesEndRef
- JSX is split into ChatView + TicketsList + EmptyState helpers
inside the new file (file stays under 400 lines despite the
split-view UI)
Parent now passes only userId + formatDate. Imports for adminApi,
MessageMediaGrid, linkifyText, AdminTicket / AdminTicketDetail types
all drop from AdminUserDetail.tsx — tickets are the only consumers.
Net to AdminUserDetail.tsx: down to 2395 lines (-295). Same DOM,
same event handlers, same queryKey ['admin-user-tickets', userId].
Pulls the balance tab — current balance, add/subtract form, active
promo offer summary + deactivate, send-offer form, recent
transactions list — into components/admin/userDetail/BalanceTab.tsx.
Form state (balanceAmount, balanceDescription, offerDiscountPercent,
offerValidHours), mutation handlers (handleUpdateBalance,
handleDeactivateOffer, handleSendOffer), and the inline two-click
confirm armer all live local now.
Tab takes a small prop surface — user, userId, hasPermission,
onUserRefresh, formatDate — and owns its own actionLoading /
offerSending flags so the deactivate-offer confirm doesn't dim
buttons on other tabs.
Net to AdminUserDetail.tsx: down to 2690 lines (-212). The
promoOffersApi import is dropped from the parent — it's only the
balance tab that talks to that endpoint now. PlusIcon / MinusIcon
are still used by the subscription tab in the parent, so they stay
there too.
Pulls the entire referrals tab into components/admin/userDetail/
ReferralsTab.tsx — state, useQuery, debounced search effects,
click-outside listeners, mutation handlers, and the 295-line JSX
block all live there now. The tab owns its referralsListQuery and
calls onUserRefresh() (parent's loadUser) when a mutation needs to
re-fetch the user object.
Net to AdminUserDetail.tsx: down to 2902 lines (-479). Local
'actionLoading' inside the tab means referral-mutation clicks no
longer dim balance/sync buttons. Same DOM, same handlers, same
queryKey ['admin-user-referrals-list', userId].
AdminUserDetail.tsx was 3820 lines — a single page bigger than the
project's 800-line soft cap by 4.7x. Split the two most self-contained
tabs (no mutations / setState from within the tab body) into their
own files under src/components/admin/userDetail/:
GiftsTab.tsx (299 lines)
- GiftStatusBadge (badge for one of 6 lifecycle states)
- GiftCard (single row, sent or received)
- GiftsTab (sent + received sections + counters)
props: giftsLoading, giftsData, locale, onNavigateToUser
SyncTab.tsx (207 lines)
- ArrowDownIcon, ArrowUpIcon (used only by sync action buttons)
- SyncTab (bot vs panel comparison + 2-way sync)
props: user, syncStatus, userSubscriptions, activeSubscriptionId,
onActiveSubscriptionChange, actionLoading, onSyncFromPanel,
onSyncToPanel, locale
Net: AdminUserDetail.tsx down to 3381 lines (-439). No behaviour change:
the tabs render the exact same DOM, same SVG paths, same event handlers
just routed through props.
TelegramRedirect already had a local getSafeRedirectUrl helper that
collapsed protocol-relative URLs, absolute URLs, exotic schemes, and
URL-encoded forms down to '/'. TopUpAmount.handleSuccess was navigating
straight to a user-supplied returnTo query param without that filter —
not externally exploitable through react-router's navigate() (it doesn't
trigger an external nav), but a crafted link could produce ugly path
artefacts ('?returnTo=https://evil.com' would land the user at
/balance/top-up/<method>/https://evil.com).
Hoist the helper to src/utils/safeRedirect.ts, rename to
getSafeRedirectPath, reuse it in TelegramRedirect, and wrap TopUpAmount's
returnTo through it before navigate().
Four image sites that render thumbnails/previews inside lists or tab
panels now hint loading=lazy so the browser defers the network fetch
until the element is near the viewport:
- Support.tsx attachment preview grid
- AdminTickets.tsx reply attachment grid
- BroadcastPreview.tsx photo preview
- BrandingTab.tsx logo preview in admin settings
Top-of-page logos (AppHeader/AppShell/DesktopSidebar/Login/Telegram
redirect/DeepLink) intentionally stay eager — they're above the fold
and should paint on first frame. Payment-method icons (6x6 px) are
already negligible. Compose attachment previews use blob: URLs where
loading=lazy is a no-op.
initialize() had six inline set({ accessToken, refreshToken, user,
isAuthenticated, isLoading }) blocks — three success variants and three
failure variants. They were byte-identical aside from which access
token they passed through, so any future field addition would have
required updating six places in lockstep, with no compile-time
guarantee.
Extract two closure-scoped helpers:
applySession(token, refresh, user) — the success shape
clearSession() — the failure shape, with the
tokenStorage.clearTokens()
call bundled in
State writes are byte-for-byte identical to before; this is a
maintenance-only refactor. The accessToken param keeps the same
'string | null' typing the inline set() previously accepted, so behavior
on edge cases (null tokens reaching success branches via type-flow gaps)
is unchanged.
The page maintained two parallel cache mechanisms:
L1 — adminTrafficApi.getCached + Map<key, {data, timestamp}> with 5min
TTL and a 20-entry LRU eviction
L2 — manual setState + initialLoading/loading flags + cancelled-flag
cleanup in useEffect
React Query covers both with stronger guarantees:
- dedup of in-flight requests with identical queryKey (the prefetch
+ main-fetch race could double-fire previously)
- staleTime: 5 min matches the old TTL
- gcTime: 5 min matches the eviction behaviour
- background refetch on stale + tab refocus disabled globally
trafficQuery owns the main list; enrichmentQuery is gated on
!trafficQuery.isLoading && items.length > 0 (preserves the 'don't tag
before tags have anyone to tag' guard). Existing setItems / setNodes /
setLoading / setEnrichment state is fed by sync useEffects so all
downstream selectors, the table, derived memos, and the column defs
are untouched.
Prefetch adjacent periods now goes through queryClient.prefetchQuery
(same warmup cadence, no separate path). handleRefresh invalidates
both queries + the internal Map cache and refetches enrichment
imperatively.
adminTrafficApi's own Map cache stays in place as a harmless L2 — it
backs queryFn calls and any non-RQ caller (none in src/ today, but
the API surface is shared).
TelegramRedirect schedules up to four setTimeouts (loading-screen-delay,
already-authenticated-redirect, not-in-telegram-redirect, post-login
redirect) inside a useEffect whose deps include isAuthenticated /
authLoading — both of which change during loginWithTelegram, re-running
the effect mid-flight. None of the timers were tracked, so the previous
run's pending navigate() callbacks fired after the new run started,
double-triggering setState on the new closure.
Route every schedule through a local timers[] that the effect cleanup
flushes. Same UX on the happy path; no stray late navigations on the
re-entrancy edge.
handleSubmit schedules a navigate('/login') via setTimeout 2 s after a
successful reset to show the success state first. If the user navigated
away before the timer fired (back button, manual URL change), the
callback still ran setState/navigate on an unmounted component — React
warns, and the redirect could yank a user who deliberately left.
Park the timer in a ref and clear it from a mount-time useEffect's
cleanup. Functionally invisible on the happy path; quiet on the edge.
translateAction took 't: any' with an eslint-disable on top — the only
explicit 'any' annotation in the entire src/ tree. Inline TFunction from
i18next; same call sites, full type safety on the i18n lookup. Now zero
'any' annotations in src/.
React 19's @types/react flags FormEvent as deprecated — quote: 'FormEvent
doesn't actually exist. You probably meant to use ChangeEvent, InputEvent,
SubmitEvent, or just SyntheticEvent'. All 17 call sites in this repo
typed form onSubmit handlers and only called e.preventDefault(), so
SyntheticEvent is the correct general replacement.
No behavior change — pure type-level cleanup that clears the chronic
deprecation hint that has been showing up after every edit to any
form-bearing page.
PermissionRoute unconditionally sent permission-failing users to /admin.
The current router wraps /admin in AdminRoute (no permission check), so
this never triggered. But if /admin ever picks up a PermissionRoute
guard — even by accident in a future change — the user would loop on
itself with replace history (no back-button escape).
Detect that case explicitly: if we're already on /admin and still failing
the permission check, fall back to / instead. Costs nothing today,
prevents a hard-to-diagnose lock-out later.
index.html ships <html lang='ru'> and that attribute never changed at
runtime, regardless of the user's selected language. Consequences:
- screen readers pronounced en/zh/fa content using Russian phonics
- browsers offered to translate already-translated content
- the Telegram login button and an OAuth widget read documentElement.lang
to set their own locale and got the wrong one
LanguageSwitcher had a half-fix that only touched dir for fa, and only
locally — i18n changes from anywhere else (initial detect, Telegram
adopt) were not covered.
Move both lang and dir updates into a single languageChanged subscriber
inside i18n.ts, run once for the initial detected language, and drop
the now-redundant dir-setting in LanguageSwitcher. Layout direction
flips automatically for the full RTL set (fa + future ar/he/ur).
Six routes were wrapped in <ErrorBoundary level='app'> around LazyPage.
After the previous commit moved a level='page' boundary inside LazyPage,
the outer 'app' boundary became unreachable — the inner one catches
first.
Strip the dead wrappers (PurchaseSuccess, QuickPurchase, AutoLogin,
TopUpResult, GiftSubscription, GiftResult). Behavior is identical, the
route blocks are 25 fewer lines, and there's only one boundary to
reason about per route.
A render error in any route used to bubble all the way up to the single
page-level boundary at AppWithNavigator, replacing the entire shell
(sidebar, header, blocking overlays) with the contained error UI — even
if only one sub-page was broken.
Embed an ErrorBoundary inside LazyPage so the smallest meaningful unit
(a route's lazy chunk + its rendered tree) is the boundary scope. Shell
chrome stays alive, the user can navigate elsewhere, and chunk-load
failures still get the lazyWithRetry reload path because the boundary
sits outside Suspense.
queryClient.invalidateQueries({ queryKey: ['user'] }) appears twice in
the codebase but no useQuery with that key is registered anywhere — the
auth user is a zustand store, not a React Query cache entry. The
invalidations were silent no-ops that suggested wiring that did not
exist.
The actual refresh path (await getMe + setUser) was already in place
right above each invalidate, so the cleanup is purely cosmetic +
documentation: replace with a comment explaining where the user state
actually lives, so the next reader doesn't try to invalidate it again.
The scope add-popover wrapped its tabs/search/list in role=dialog with
no aria-modal and no focus trap. That promises modality to AT but
delivers a non-modal popover users can click past — confusing.
Remove the wrapper role. The trigger button already exposes the popup
via aria-haspopup=listbox + aria-expanded, and the inner div carries
role=listbox aria-multiselectable. Nothing for AT to misinterpret now.
Two admin paths previously had empty catch blocks around copyToClipboard:
the user-detail copy helper and AdminLandings handleCopyUrl. The adapter
already attempts the legacy execCommand fallback, so reaching the catch
means the operation truly failed — staying silent left admins thinking
their click worked when it hadn't.
Both now fire notify.error so the failure is visible.
Bulk-action modal had role=dialog on the overlay but no focus trap,
no aria-labelledby, no scroll lock, and an ad-hoc Escape listener
that conflicted with anything else listening to document keydown.
Move role=dialog/aria-modal onto the focus-trapped content element,
wire aria-labelledby to the h3 title, and replace the ad-hoc
keydown effect with useFocusTrap (Tab cycle + Esc + body scroll
lock + focus restore on close). Loading state suppresses Esc but
keeps focus trapped, so Tab still cycles inside the progress view.
Header checkbox for 'select all subscriptions' read allVisibleSubscriptionIds,
subscriptionSelection, isMultiTariff, toggleAllSubscriptions from a stale
closure because those values were declared AFTER the columns useMemo. The
warned-about deps were absent, so the header only refreshed by accident
when expandedRows or getFilteredSubs changed.
Relocate filteredUsers / allVisibleSubscriptionIds / toggleAllSubscriptions
above the columns block, list them as deps. Header now reflects current
selection without piggybacking on unrelated re-renders.
Same commit clears the matching AdminInfoPageEditor warning with an
explicit eslint-disable-line comment — the activeLocale omission was
already intentional (initial-content lock), just not silenced.
60s setInterval becomes refetchInterval, dropping the cancelled-flag
ceremony and the manual loading/setState plumbing. Two independent
queries — one failure no longer poisons the other (Promise.all
previously rejected the whole batch on a single 5xx).
User list query keys all 8 filter inputs so pagination/search/filter
changes auto-refetch with proper dedup. Tariffs/promoGroups/campaigns/
partners become long-staleTime lookup queries — fired once and cached
across remounts of the page. Existing setUsers/setTariffs/etc. setters
stay; sync useEffects copy query.data into them so downstream selection
and handler code is untouched. loadUsers becomes a thin refetch wrapper
for handleRefresh and mutation handlers.
Drops manual activeTab/userId loader triggers — query enabled
gating now handles all per-tab fetching. Wrappers around refetch
keep mutation handler call sites unchanged. Removes 4 dead
wrappers (loadTariffs/Referrals/Gifts/PromoGroups) only called
by the now-deleted activeTab useEffect.
Replace the 13-line manual fetch with a useQuery hook keyed on userId. loadUser is
kept as a thin wrapper around userQuery.refetch() so all 25+ existing call sites in
mutation handlers continue to work unchanged. Sync useEffects copy data → setUser,
isFetching → setLoading, and isError → navigate, preserving the original
'load fails → redirect to /admin/users' behavior.
Remaining 12 loaders (loadSyncStatus/loadTariffs/loadTickets/loadReferrals/loadPanelInfo/
loadNodeUsage/loadDevices/loadSubscriptionData/loadGifts/loadPromoGroups/loadTicketDetail/
loadRequestHistory) NOT migrated in this commit — some are parameterized, some chain via
Promise.all in loadSubscriptionData; they need a dedicated refactor with functional
testing across all 7 tabs and the mutation flows. This commit ships the lowest-risk,
highest-leverage piece (userQuery has the most call sites).
All 10 tab data fetches + the initial status load move from manual useState +
useEffect + loadTabData(switch) pattern to per-tab useQuery hooks with enabled
gating. Each query syncs into the existing state vars via useEffect so the JSX
and mutation handlers stay unchanged; handleSearch/handleUnban/handleToggle/handleSet
now call query.refetch() directly; refresh button uses refetchActiveTab helper.
Net effect: tab switches return to cached data instantly (background revalidate),
no duplicate fetches, no manual loading/error wiring. The reports query has
reportHours in its queryKey so changing the period auto-refetches without the
extra useEffect that previously handled it.
Wrap the error message blocks in role=alert so assistive technology announces
authentication failures and reset-password errors when they appear (WCAG 4.1.3).
Subscription copy-URL button only had a title attribute; back and QR-open buttons in
InstallationGuide had no accessible name. Add aria-label so screen readers announce
the action (WCAG 2.4.6).
Replace manual fetchStats/fetchExtendedStats (useState + useEffect + setInterval(30s)
+ console.error) with two useQuery hooks using refetchInterval. Keeps existing variable
names (stats/loading/error/referrers/campaigns/payments/systemInfo) so JSX is unchanged.
handleRestartNode/handleToggleNode/retry button now refetch via the query.
Replace manual useState + useEffect + useCallback + console.error fetch pattern with
useQuery for the users list and stats. Adds caching, dedupe, stale-time, automatic
loading/error state — and the refresh button now uses refetch() (single source of truth).
Wrap template-selection and send-mode button groups in role=radiogroup +
aria-labelledby; mark each option button as role=radio with aria-checked
(WCAG 4.1.2). Labels get matching id for the labelledby relationship.
Sweep non-brand blue across admin pages, AnalyticsTab, TvQuickConnect, and menuLayout
style picker to accent tokens. Telegram-brand blue kept in BroadcastPreview (TG chat
mockup), the blocking screens (TG channel/bot CTAs), and the SuccessNotificationModal hero.
Sweep orange across SubscriptionPurchase/GiftSubscription/PromoOffersSection/
AdminTrafficUsage/AnalyticsTab/withdrawalUtils to warning-* tokens. Stars brand
yellow→orange gradient kept in TopUpAmount.
Map text/bg/border-gray-N → dark-N across the four blocking screens (Maintenance,
Blacklisted, ChannelSubscription, AccountDeleted) so muted text + decorative surfaces
follow the design tokens. BroadcastPreview's email mockup and MessageMediaGrid's
fullscreen overlay intentionally keep their literal gray/white (mock chrome).
Replace gradient CTAs that fade between same-token shades with flat backgrounds
(TopUpAmount accent submit, ChannelSubscriptionScreen single-channel button, three
SuccessNotificationModal action buttons). Stars yellow→orange gradient kept as a
Telegram brand cue; decorative card-header gradients left as visual identity.
Both pages were rendering with raw Tailwind gray-* + a non-existent border-primary-600
class (invisible spinner). Map: bg-gray-50 → bg-dark-950, text-gray-900 → text-dark-50,
text-gray-500 → text-dark-400, border-primary-600 → border-accent-500. Pages now match
the dark theme and the spinner is visible.
