homarr/REVIEW.md

PR #6502 widget-modernization ultra-review

Executive result

Merge recommendation: changes required. No P0 was found, but the current branch still has five P1 issues: a pre-existing Custom API authorization/redirect weakness exposed by this subsystem review, an inaccessible manual advanced-focus interaction, a likely fixed-position/transform geometry defect, systemic false empty states for failed queries, and unvalidated integration-provided navigation URLs. Eight P2 issues cover raw error disclosure, unsafe runtime-ref interfaces, localization, responsive/theming drift, test coverage, and inaccurate PR copy.

Audit health (Impeccable rubric): accessibility 1/4, performance 2/4, responsive design 2/4, theming 2/4, anti-patterns 3/4; 10/20 (Acceptable, significant work required). The UI does not look generically AI-generated; its problem is inconsistent technical behavior beneath an otherwise coherent Mantine vocabulary.

Reviewed revision and environment

Item	Value
PR	#6502, feat: modernize dashboard widgets
Reviewed head	96e92485a5fd45f3648cfac0453e614779c414ba
Base branch / fetched base	dev / d58a47b776a55d04dee2220f59d072b9af8bbed3
Merge base	d58a47b776a55d04dee2220f59d072b9af8bbed3
Diff	359 files, +15,721/-4,478
PR state	open, non-draft, BLOCKED; no approval decision
Hosted checks at audit start	Fast gate, Container and E2E, CodeQL, JS/TS analysis, Actions analysis, title, auto-fix: green
Inline threads	34 total, 7 unresolved; only 2 remain substantively valid
Mantine	repository lock: @mantine/core, hooks, charts and dates 9.4.1
Local tools	Node 24.16.0 and pnpm 9.5.0; repository requires Node >=24.18.0 and pnpm >=11.15.1

The former red baseline 7b7045202d289b7d83946befcee5d2845ea4a706 had a green Fast gate and a failed Container/E2E run (30748255760, job 91498308523). The latest head has a green Container/E2E job (30756067198, job 91519039504). This proves the current suite passed once; it does not satisfy the requested three-run flake characterization.

PR-description traceability

Claim	Result	Evidence / correction
“all 54 widget renderers with compact and advanced modes”	Inaccurate	54 WidgetKind values exist. displayMode="advanced" reaches 53; app is deliberately excluded in item-content.tsx:51 and documented as compact-only. Change copy to “53 widget renderers; app tiles remain compact,” or implement app advanced mode.
advanced focus: hover, keyboard, touch, geometry	Partial	Entry paths, Escape ownership, reduced motion, close focus, touch button and zero-dimension guards exist. WID-002/003 remain.
richer widget context actions	Partial	Menu.ContextMenu is a good Mantine choice and query scope is item-aware. Dispatch still crosses any; Custom API action authorization is not represented.
standardized query identity/cache/status	Partial	Query keys, matchers, runtime capture, recovery and tests exist. Thirty-nine of 53 renderer query calls do not read errors; runtime ref registration is render-time mutable state.
safe polling	Mostly confirmed	defaults are limited to widget namespace; Custom API resume immediately refetches; per-widget intervals exist. State/ref typing still needs redesign.
partial-failure rendering	Partial	settlement helper and several routers are strong; many renderer-level total failures still appear empty/disconnected.
bounded integration responses	Mostly confirmed	common bounded reader, Custom API 2 MiB, timetable 512 KiB, Traefik bounds and Beszel queue bounds exist. Custom action execution remains outside the hardened fetch path.
timetable hardening	Confirmed	board/item permission checks, saved-value comparison, public-address resolution, private ranges, per-request DNS re-resolution, pinned dispatcher, redirect rejection, timeouts, response bounds and schemas are present.
precise placement	Mostly confirmed	mouse, touch tolerance, keyboard, occupied-cell validation, dynamic bounds, nearest fit and immediate persistence exist. Failure/race/live responsive coverage remains insufficient.
widget catalog documentation	Partial	all kinds have pages and shared interactions; wording overclaims focus behavior and validation counts are stale.
validation: 40 files/162 tests, 6/6 E2E	Stale	current diff contains 57 changed unit-spec files (including three E2E specs); repository has 42 widget spec files and six E2E specs. Update from the exact final run.
independent reviews “pass”	Not supported	unresolved valid findings, current P1s, and unavailable live QA contradict an unqualified pass.

Mantine suitability ledger

Legend: U already used appropriately; C consolidation/replacement candidate; R relevant but current implementation can stay; N not relevant to widget runtime; D dependency/version addition. Exact 9.4.1 package archives were inspected in /tmp; live docs/Context7 include 9.5.0.

Core catalog

• U: ActionIcon, Alert, Anchor, Avatar, Badge, Box, Button, Card, Center, Code, Divider, Flex, Grid, Group, Image, Loader, Menu (Menu.ContextMenu), Overlay, Paper, Popover, Portal, Progress, RingProgress, ScrollArea, SimpleGrid, Skeleton, Stack, Table, Tabs, Text, ThemeIcon, Title, Tooltip, UnstyledButton.
• C: Accordion (dense integration sections), DataList (semantic label/value stats), EmptyState (replace bespoke empty state), FocusTrap and Modal compound API (WID-002), LoadingOverlay (background refetch without deleting content), NumberFormatter, OverflowList/Scroller (responsive action and chip rows), Table.ScrollContainer, Timeline (activity/feed histories), VisuallyHidden and Kbd, CloseButton, Dialog/HoverCard/Notification where persistent errors/actions need semantics.
• R: Affix, AppShell, AspectRatio, BackgroundImage, Breadcrumbs, Burger, Checkbox, Chip, Collapse, ColorInput/ColorPicker/ColorSwatch/AlphaSlider/AngleSlider/HueSlider, Combobox/ComboboxPopover, Container, CopyButton, Drawer, Fieldset, FileButton/FileInput, FloatingIndicator/FloatingWindow, Highlight, Indicator, Input/JsonInput, List, Mark, Marquee, MaskInput, Menubar, MultiSelect/NativeSelect/NumberInput/PasswordInput/PillsInput/PinInput/Radio/RangeSlider/Rating/SegmentedControl/Select/Slider/Switch/TagsInput/Textarea/TextInput, Pagination, Pill, SemiCircleProgress, Space, Splitter, Spoiler, Stepper, Transition, Tree/TreeSelect, Typography. Use only where an existing widget requirement matches; do not churn working layouts.
• N: Alpha/color/angle primitives outside configuration forms, AppShell/navigation primitives, TableOfContents, CorePackage documentation.
• D (9.5, not 9.4.1): Cascader. RollingNumber and Splitter are present in 9.4.1.

Hooks catalog

• C: useFocusReturn, useFocusTrap, useReducedMotion, useRovingIndex, useLongPress, useScroller, useElementSize, useViewportSize, useHotkeys, useMediaQuery, useResizeObserver.
• U/R: focus-within, hover, disclosure, click-outside, merged-ref, interval/timeout, local-storage, event/window listeners, throttled/debounced values and list/set/map/state helpers are available in 9.4.1 and should replace local mechanics only when behavior and tests improve.
• N: document title/favicon/hash/headroom/page-leave/OS/eye-dropper/fullscreen/network/orientation and other page-level utilities have no direct widget-remediation need. Every hook listed under ## Hooks in MANTINE.md is covered by this rule; no hook requires 9.5 for the recommendations above.

Form, dates, charts, schedule, extensions, theme and styles

• Form: existing @homarr/form should remain the abstraction over all cataloged form APIs. No widget-runtime replacement; use current validation/status APIs for widget editors (R).
• Dates: all catalog entries are in installed 9.4.1 dates. Calendar, MiniCalendar, and TimeValue are candidates; other picker/input entries are editor-only (R). Importing dates directly into @homarr/widgets requires adding that workspace dependency (D).
• Charts: 9.4.1 exports Area, Bar, BarsList, Bubble, Composite, Donut, Funnel, Heatmap, Line, Pie, Radar, RadialBar, Sankey, Scatter, Sparkline and Treemap (C/R). Prefer Sparkline for tiny trends, BarsList for ranked values, Heatmap for calendar/activity density, Composite for mixed analytics. BulletChart and SunburstChart are 9.5-only (D). Replacement requires accessibility and bundle comparison, not blanket Recharts removal.
• Schedule: all AgendaView, day/week/month/year/resource views and Schedule pages require new @mantine/schedule dependency (D). AgendaView is the only plausible calendar-widget candidate; its weight and semantics must beat the existing grouped list before adoption.
• Guides/FAQ/Other: server components, portal testing, scroll lock, accessibility, nested popovers, zoom, responsive styling and version changelogs are relevant references (R); framework/setup/contribution/old migration entries are N. The 9.5 changelog is informational only.
• Theming/styles: Colors, ColorSchemes, DefaultProps, MantineProvider, ThemeObject, CSS variables, data attributes, CSS Modules, responsive/RTL/style props/Styles API/performance are U/R. WID-009 requires shared layer/color tokens. Emotion, Sass, Vanilla Extract and unrelated global setup are N.
• Extensions: existing modals, notifications, spotlight and rich-text abstractions are U; Carousel could consolidate Immich only after behavior parity; Dropzone/CodeHighlight/NProgress are N for this audit.

Complete 54-widget coverage matrix

“Test” means a focused widget-local spec exists, not that all required states are covered. Every row still requires the common contract suite in WID-013.

WidgetKind	Family	Advanced	Focused test	Principal audit gap
clock	basic	yes	no	query error, locale-format options
weather	basic/feed	yes	yes	query error; dayjs locale/time patterns
app	basic/link	no	no	PR wording; URL contract
iframe	custom content	yes	yes	manual focus/iframe trapping; URL privacy thread now fixed
video	custom content	yes	no	media/focus/live behavior
notebook	notes	yes (inner)	yes	editor focus inside manual surface
anchorNote	notes	yes	no	query error, save/error regression
dnsHoleSummary	network	yes	no	total query error, hard-coded chart colors
dnsHoleControls	network/security	yes	no	total error and permission matrix
smartHome-entityState	smart home	yes	no	raw state in aria; action permissions
smartHome-executeAutomation	smart home	yes	no	mutation/permission/edit-mode E2E
stockPrice	analytics	yes	yes	total error
mediaServer	media	yes	no	total error and table a11y
calendar	calendar	yes	yes	total error; fixed HH:mm; schedule tradeoff
downloads	media	yes	yes	four-state/component coverage, arbitrary z-index
mediaRequests-requestList	media	yes	no	total error; external URLs/actions
mediaRequests-requestStats	media	yes	no	total error; links/colors
mediaTranscoding	media	yes	yes	total error and responsive table
mediaMissing	media	yes	yes	total error and external URL
minecraftServerStatus	infrastructure	yes	no	total error/malformed payload
networkControllerSummary	networking	yes	no	total error; WAN/unknown localization
networkControllerStatus	networking	yes	no	responsive rule duplication
rssFeed	feed	yes	yes	strong stale/partial state; locale relative time
bookmarks	basic/link	yes	yes	recomputation; URL contract
indexerManager	media	yes	no	total error; runtime action ref
healthMonitoring	infrastructure	yes	yes	integration query error
releases	infrastructure/feed	yes	yes	secret/item authorization needs E2E
mediaReleases	media	yes	no	total error; external URLs
dockerContainers	infrastructure	yes	yes	total error; z-index/table semantics
firewall	networking/security	yes	yes	partial-query matrix; loose router generics
notifications	infrastructure	yes	no	total error; untrusted href
systemResources	infrastructure	yes	yes	total error; previous network thread fixed
coolify	infrastructure	yes	yes	total error; URL validation
systemDisks	infrastructure	yes	yes	total error
timetable	calendar	yes	no	backend strong; component locale/live tests absent
immich-serverStats	media	yes	yes	two ignored query errors
immich-albumCarousel	media	yes	yes (query)	query error and nested overlay/focus E2E
paperlessNgx	analytics	yes	no	former error thread fixed, no regression spec
patchmon	infrastructure	yes	yes	total error
bazarr	media	yes	yes	total error
tracearr	analytics	yes	yes	total error
speedtestTracker	analytics	yes	yes	total error
uptimeKuma	infrastructure	yes	yes	total error
audioStats	media	yes	yes	good dual-error handling; partial-state tests needed
umami	analytics	yes	no	runtime query registration; chart a11y
vpn	networking/security	yes	no	total error; status/permission matrix
archiveTeamWarrior	infrastructure	yes	yes	suspense path; docs present
ups	infrastructure	yes	no	total error
beszelSystemTable	monitoring	yes	yes	nested chart/focus and live subscription
beszelSystemGrid	monitoring	yes	no	live/error UI; no focused renderer test
beszelAlerts	monitoring	yes	no	partial error and status semantics
beszelSystemStats	monitoring	yes	yes	runtime refs, locale tooltip
traefik	networking	yes	no	total error; bounded backend positive
customApi	custom content/security	yes	yes	WID-001/007/008; polling resume fixed

Severity-ranked findings

WID-001 — P1 confirmed, pre-existing: Custom API actions lack resource authorization and hardened fetch policy

• Location/claim: packages/api/src/router/custom-widget/custom-widget-router.ts:57-102,392-451; “harden Custom API” and action-permission semantics.
• Evidence/reproduction: any authenticated account can call customWidget.all, learn every definition ID and URL, then call customWidget.execute({definitionId}). The mutation verifies neither board nor item nor an action permission. It decrypts administrator-provided secrets, uses new URL() only, and redirect: "follow". The PR hardened widget.customApi.getData, but not this sibling execution path.
• Impact/rule: unauthorized side effects against internal services and authenticated redirect chains; broken object-level authorization. A view-only account can invoke administrator-created controls outside any board context.
• Correction: require {boardId,itemId,definitionId}; load the item and verify it is customApi, belongs to the board, references the definition, and the caller has an explicit execute policy. Until that policy exists, restrict to admin. Share validateCustomApiUrl, timeout, redirect rejection, bounded disposal and redacted errors with getData. Compatibility: old direct callers must send item/board context.
• Tests: procedure tests for anonymous, unrelated authenticated, view-only, modifier, admin, mismatched item/definition, disabled definition, redirects, credential URL, timeout and secret non-disclosure; E2E action button across roles.

WID-002 — P1 confirmed, PR-introduced: manual advanced focus is not an accessible modal

• Location: advanced-focus/context.tsx:194-209, items/item-content.tsx:113-189.
• Evidence: manual mode renders an Overlay plus a role="region" Card. There is no FocusTrap, aria-modal, background inertness, scroll lock or backdrop dismissal. Only the close button is initially focused; Tab can leave the surface and activate dimmed board controls.
• Impact/rule: keyboard/screen-reader users lose the interaction boundary; WCAG 2.1.2, 2.4.3 and ARIA dialog pattern. Iframes/editors make escape especially likely.
• Correction/Mantine: render manual mode through Modal.Root compound parts (or ModalBase), using its portal, FocusTrap, scroll lock, Escape/return-focus and semantic content. Preserve preview mode as non-modal. Configure nested overlays and removeScrollProps; keep a 44px Modal.CloseButton and explicit title via VisuallyHidden if needed.
• Tests: RTL portal/focus-cycle/return tests; Playwright Tab/Shift+Tab, nested Menu/Select/Popover/editor/iframe, backdrop and every close path; reduced motion and 200% zoom.

WID-003 — P1 probable, PR-introduced: fixed advanced surface remains under transformed Gridstack ancestry

• Location: advanced-focus.module.css:35-42; Card remains inside the Gridstack item while only the backdrop is portaled.
• Evidence: CSS fixed descendants establish their containing block at a transformed ancestor. Gridstack uses transforms for positioned items, so viewport-derived left/top can be interpreted relative to the item on affected layouts. A Portal wrapper exists only for Overlay.
• Impact: off-screen, clipped or incorrectly scaled focus surface across breakpoints/remounts.
• Correction: portal the focused surface itself (naturally achieved by WID-002) and calculate geometry in the same viewport coordinate system. Do not compensate with transform-specific offsets.
• Tests: real Gridstack at every layout/viewport, scrolled page, header, dynamic section and resize/remount; assert bounding rect within viewport and finite scale.

WID-004 — P1 confirmed, regression-exposed/systemic: query failures still become valid empty/disconnected UI

• Location: 39 of 53 renderer query calls do not read error, including bookmarks, weather, calendar, media server, system resources, Coolify, DNS Hole, Tracearr, notifications, Immich, VPN and others listed in the matrix. Query defaults do not set throwOnError (trpc.tsx:69-93).
• Evidence: TanStack Query returns data=undefined on total failure. Components default to [], null, or empty branches, so the board ErrorBoundary is never reached. Network Controller and Paperless were fixed, proving the pattern; it was not completed.
• Impact: operators are told “empty” or “disconnected” when monitoring failed, violating the product’s truthful-state requirement.
• Correction/Mantine: introduce a shared WidgetQueryState contract distinguishing initial Skeleton, background LoadingOverlay, stale Alert, partial warning, true EmptyState and terminal error. Either explicitly throw only when no usable cached data or render a shared error Alert. Preserve stale data.
• Tests: parameterized contract tests for every queried widget: initial, success-empty, total failure, stale+failure, partial failure, refetch, remount/persisted cache and sibling isolation.

WID-005 — P1 confirmed, PR-introduced/systemic: integration-provided links bypass a shared URL policy

• Location: examples include media-missing/component.tsx:291, notifications/component.tsx:77, media requests list/stats, Coolify resource/server links, releases, bookmarks and iframe. Only RSS and custom JSX visibly use explicit protocol sanitizers.
• Evidence: dynamic strings are assigned directly to href/new-tab controls. Integration APIs and remote feeds are trust boundaries; no renderer-wide http/https guard is enforced.
• Impact: unsafe schemes, credential-bearing URLs, internal detail disclosure and inconsistent opener behavior.
• Correction: add one typed application URL helper returning a sanitized URL/undefined, with allowed schemes per use case, credential stripping/rejection, consistent noopener noreferrer, and disabled non-link presentation. Validate again server-side for action endpoints.
• Tests: javascript:, data:, malformed, credential, relative, http/https and missing URLs for each shared link primitive.

WID-006 — P2 confirmed, PR-introduced: runtime query/action state mutates untyped refs during render

• Location: definition.ts:121-162; runtime registration calls in media missing/transcoding, audio, calendar, Umami, Beszel and Immich; widget-context-menu.tsx:157-174.
• Evidence: setWidgetRuntimeQueries mutates a shared object during render; concurrent/aborted renders can publish uncommitted query identity. Action dispatch casts the definition to Record<string,unknown> and any, then casts its result back.
• Correction: make definition generics preserve kind/options/action types end-to-end; replace stringly opaque ref slots with a typed per-item registration API updated in a committed effect (or use useSyncExternalStore). Separate query registration from imperative actions.
• Tests: StrictMode, interrupted/remount, options change and cleanup; compile-time tsd/type assertions rejecting wrong action props.

WID-007 — P2 confirmed, regression-exposed: raw backend errors are exposed to users and assistive text

• Location: Custom API UI component.tsx:561-576; API custom-api.ts:79-85; several partial fallbacks use error.message.
• Impact: internal hostnames, request paths and implementation details can appear in tooltips/aria labels and logs.
• Correction/Mantine: return stable error codes plus redacted public messages; keep detailed causes server-side with correlation IDs. Render localized Alert/Notification details, never raw throwable text.
• Tests: errors containing credentials, query strings, private hosts and multiline backend bodies must not reach tRPC payload, DOM, accessibility tree or notifications.

WID-008 — P2 confirmed, PR-introduced/regression-exposed: locale handling remains incomplete

• Locations: Beszel chart/tooltip dayjs formats, calendar event HH:mm, downloads YYYY-MM-DD HH:mm, timetable HH:mm:ss, weather day/month/time formats; _inputs/widget-location-input.tsx:199 hardcodes Intl.NumberFormat("en").
• Also: beszel/_shared/disk-usage.tsx:41, stats loading aria label, Network Controller WAN/unknown are English/raw accessible strings.
• Correction: centralize app-locale date/time/number/duration/status helpers; load canonical dayjs locale only at the boundary; use localized translation keys for aria/status. NumberFormatter is suitable for simple display but application helpers should own locale.
• Tests: at least en-US, fr-FR, de-DE, RTL locale, 12/24-hour preference and long translations, including accessibility names.

WID-009 — P2 confirmed: layer and color tokens are inconsistent

• Locations: downloads z-index 1000 and rgba row colors; item title z-index 4; item menu/dynamic menu 10; media server white rgba; Indexer Manager hex colors; DNS Hole rgba palette; Umami hex tick colors.
• Correction: define shared semantic layer tokens relative to Mantine CSS z-index variables and semantic status/chart tokens respecting light/dark/forced colors. Do not replace intentional service branding.
• Tests: light/dark/forced-colors with nested overlays and tooltip ordering.

WID-010 — P2 confirmed: responsive decisions are duplicated and under-tested

• Location: Network Controller Status component.tsx:44,59-78 (open valid thread); Bookmark component.tsx:94 recomputation (valid but low-impact thread); similar width/height literals across renderers.
• Correction: small typed breakpoint lookup helpers per family, memoized only for genuinely expensive layout. Prefer container-query Grid/SimpleGrid where structural. Do not cargo-cult useMemo for a cheap pure calculation; benchmark Bookmark before accepting the review suggestion.
• Tests: boundary tables for minimum/typical/wide/short/advanced sizes and 200% zoom.

WID-011 — P2 confirmed: PR/docs overstate behavior and validation

• Location: PR body, board/getting-started copy “advanced preview in place,” validation numbers, and unqualified review pass.
• Correction: state 53 advanced renderers plus compact-only app; call manual focus an overlay only after WID-002/003; publish exact final-SHA check/test counts and disclose unsupported/live-untested states.

WID-012 — P2 confirmed: focused automated coverage is far below the 54-widget/data-state claim

• Evidence: 42 widget-local spec files are helper-heavy; 20+ renderers have no focused spec. Six E2E files exist, but only board interactions and Custom JSX directly exercise this modernization. The earlier Custom JSX login wait failed once and current exact-route wait remains.
• Correction: add shared renderer contract harness plus family E2Es. Replace Custom JSX’s exact post-submit URL race with a shared login helper that waits for authenticated session/stable destination, then run it three times in CI.

WID-013 — P2 confirmed: open review-thread state no longer matches code

Thread	Classification
iframe complete URL	already fixed/false positive at current line: only host is rendered
system resources network gate	already fixed, outdated
RSS total error	already fixed: no-data error throws; stale data retained
bookmarks memoization	valid P3/P2-low; benchmark first
network-controller total error	already fixed, outdated
network-controller breakpoint consolidation	valid P2 maintainability
Paperless total error	already fixed, outdated

Resolve the five stale threads with evidence; address or explicitly disposition the two live suggestions. Re-run review because CodeRabbit was paused during active development.

Systemic remediation roadmap

1. Security/authorization: WID-001 and WID-005; add negative-role/protocol tests before merge.
2. Interaction accessibility: replace manual focus with portaled Mantine modal semantics (WID-002/003), then test nested widget content.
3. Truthful data state: build the shared query-state contract and migrate all 39 ambiguous sites; keep valid stale data and partial results.
4. Typed interfaces: replace widgetStateRef slot mutation and any context dispatch with committed typed registrations.
5. Widget families: media/downloads, infrastructure, networking/security, smart-home, analytics/calendar/feed, then Beszel; apply URL, state, locale, responsive and a11y contracts in that order.
6. Design-system cleanup: semantic layers/colors and restrained Mantine consolidation. Do not add Schedule or upgrade to 9.5 without a separate dependency/bundle decision.
7. Docs/PR: correct renderer count, behavior and exact validation evidence.
8. Final proof: frozen install on supported Node/pnpm; format, lint, full typecheck, affected and full unit tests, production and docs builds; Custom JSX three times; complete production-container E2E; desktop/tablet/mobile keyboard/touch/light/dark/reduced-motion/zoom/a11y-tree QA. Push nothing until all refer to one unchanged SHA; close every review thread and obtain required approval.

Positive findings

• Timetable is the strongest security work in the PR: saved resource authorization, private-range rejection, DNS rebinding resistance through pinned lookup, redirect rejection, bounds and schemas are coherently tested.
• Bounded response helpers, Beszel bounded queue/cancellation, and partial-integration settlement materially reduce failure blast radius.
• RSS preserves stale data and distinguishes partial feed failure; Network Controller and Paperless now correctly escalate total failures.
• Item-scoped query matchers and tests address sibling-widget refresh isolation; global polling defaults are restricted to the widget namespace.
• Touch long-press has movement tolerance; focus scale guards prevent Infinity/NaN; reduced-motion and remount cleanup are present.
• Existing Homarr theme, Mantine components and shared UI vocabulary are generally preserved. Most proposed Mantine work is consolidation, not redesign.

Validation evidence and uncertainty

• MANTINE.md was fetched from https://mantine.dev/llms.txt and passed byte-for-byte cmp; llms-full.txt stayed in /tmp.
• Context7 current Mantine docs were queried for overlay/focus, responsive/data-state/accessibility, charts and Schedule. Static 9.4.1 archives verified exact exports; Cascader, BulletChart and SunburstChart are 9.5-only.
• Local frozen install could not start because host Node/pnpm are below repository engine requirements. No dependencies, source files or machine toolchain were changed.
• Docker API access is denied, so production-container E2E could not be rerun locally. Hosted exact-head Fast gate and Container/E2E are green.
• agent-browser is not installed. Temporary npx execution was rejected by the environment’s security reviewer, so live exploratory QA/a11y-tree inspection was not performed.
• External integrations were not available; malformed/partial/permission combinations are static-analysis conclusions unless a focused test already exists.
• This review intentionally changes no production code and makes no commit, push, thread resolution or PR edit.

Finalization must re-query the PR head/checks/threads. If the head differs from the reviewed SHA, this report is a snapshot and the delta requires review before merge.

Final refresh (2026-08-03): head remained 96e92485a5fd45f3648cfac0453e614779c414ba; all reported hosted checks remained successful; merge state remained BLOCKED with no approval decision; unresolved threads remained 7 of 34. The branch did not change during the audit.