Skip to content

fix(web): polish window capture settings

MacroscopeApp / Macroscope - UI Consistency failed Aug 24, 2026 in 6m 23s

UI Consistency: 2 issues found

Two findings posted as inline review comments.

  • apps/web/src/components/chat/ChatComposer.tsx (line 3224): the window-capture attachment card is pinned to w-56, leaving roughly 50px for the app name and window title after the w-24 thumbnail, pr-8 remove-button gutter, icon, and gap, so both labels truncate after a few characters. The sibling wide attachment card (ComposerPreviewAnnotationCards) sizes intrinsically with a max cap; suggested dropping the fixed width and capping the label instead.
  • apps/web/src/components/desktop/WindowCaptureCoordinator.tsx (line 164): the t3:focus-composer event name is a bare literal duplicated in ChatView.tsx with no shared constant, and diverges from the existing t3code:-namespaced typed bus (components/preview/previewActionBus.ts) used for the same root-level to ChatView hop.

Previously reported findings that are now resolved: keyframe ownership for window-capture-card-enter, data-keybinding-capture on the recorder button, DialogPanel composition in the onboarding dialog, capability gating of the onboarding dialog via getDesktopWindowCaptureBridge(), and disabling the sound/flash/animation switches without a capture bridge.

Details

Note

Your check run agent prompt is: .macroscope/check-run-agents/ui-consistency.md
More information about how Check Run Agents work can be found in our Help Center.

Scope reviewed: apps/web/src/components/chat/ChatComposer.tsx, chat/MessagesTimeline.tsx, chat/WindowCaptureAttachmentDetails.tsx, desktop/WindowCaptureCoordinator.tsx, desktop/WindowCaptureOnboardingDialog.tsx, settings/WindowCaptureSettings.tsx, settings/SettingsSidebarNav.tsx, settings/settingsSearch.ts, lib/desktopWindowCapture.ts, routes/settings.window-capture.tsx, routes/__root.tsx, CommandPalette.tsx, and index.css.

Checks performed:

  • Compared the new capture cards against the existing wide composer attachment card (ComposerPreviewAnnotationCards) and computed available label width from the emitted Tailwind geometry in both the composer (w-56/w-24/pr-8) and timeline (col-span-2/w-28/px-3) variants.
  • Verified the window-capture-card-enter keyframe follows the existing prompt-stash-count-enter convention (global keyframe + call-site animate-[...] with motion-reduce:animate-none).
  • Verified Kbd, Button, Switch, Dialog*, SettingsRow/SettingsSection, and ITEM_ICON_CLASS usage matches existing call sites; text-secondary-label and formatShortcutLabel are established tokens/helpers.
  • Confirmed direct window.desktopBridge reads during render are an established pattern in this codebase (ConnectionsSettings, ConnectOnboardingDialog), so the settings-page read is not a finding.
  • Traced the cross-module t3:focus-composer contract against the existing previewActionBus pattern and confirmed no global composer-handle mechanism exists (CommandPalette uses a locally scoped context ref).

Not re-posted (already flagged on this PR and unchanged): Tab is preventDefault()-ed while recording in WindowCaptureSettings.tsx, so focus cannot leave the recorder button; and WindowCaptureAttachmentDetails anchors its title to text-foreground while rendering on bg-background/70 inside the text-message-foreground bubble in MessagesTimeline.


πŸ” Agent activity β€” 28 tool calls

πŸ’­ I'm checking the current state of the changed web files to see whether the previously flagged issues were addressed β€” things like the Tab key preventDefault in WindowCaptureSettings, sound/animation switches not being disabled without the bridge, contrast issues in WindowCaptureAttachmentDetails, onboarding dialog gating on desktopBridge, DialogPanel usage, and the index.css animation styles β€” and whether any new problems have surfaced. Looking at the current head, most of these appear resolved: the sound/flash/animation switches are now disabled when there's no bridge, the keybinding capture data attribute is present, the onboarding dialog now correctly checks the desktop window capture bridge, and DialogPanel is in use. The text-foreground contrast issue in WindowCaptureAttachmentDetails still looks unaddressed, and the Tab key still isn't skipped in the recording handler, but since those were already flagged in earlier comments I shouldn't duplicate them. Now I need to scan the latest changes for anything genuinely new rather than re-flagging what's already been raised. I'm going through the ChatComposer window-capture card sizing, checking for potential overlap between the remove button and content in WindowCaptureAttachmentDetails, and looking at whether the MessagesTimeline grid column spans still make sense. I also want to confirm whether text-secondary-label is actually a valid design token in this codebase before relying on it in my review.

1. browse_code β€” apps/web/src/components/chat/WindowCaptureAttachmentDetails.tsx

2. git_tools β€” secondary-label

πŸ’­ Good, that's valid. Now I want to check the ChatComposer attachment card region and the MessagesTimeline grid, so I'll look at ChatComposer around lines 3200-3290.

3. browse_code β€” apps/web/src/components/chat/ChatComposer.tsx

4. browse_code β€” apps/web/src/components/chat/MessagesTimeline.tsx

πŸ’­ The text-foreground inside bg-message bubble issue was already flagged and remains unchanged, so I won't repeat that note. Now I want to check what's newly changed since the last run's commit to focus my review only on the latest diff.

5. git_tools β€” aee6932..a4d9dda

πŸ’­ The branch history was likely rewritten via force-push, so I'll skip tracing commits and instead review the current head files for new issues, making sure not to re-flag things already covered like Tab preventDefault, sound/animation, text-foreground, onboarding gating, DialogPanel, and the CSS/keybinding fixes. Of the earlier findings, only the Tab preventDefault issue and the text-foreground usage in WindowCaptureAttachmentDetails remain unaddressed. Now I want to check some new areas: whether SettingsRow renders correctly when rows lack a title prop versus using searchableSetting, how the Kbd component and DialogPopup className are used across other dialogs, the formatShortcutLabel helper, and the remove button positioning on the composer window-capture card relative to its padding. I'm going to look at the SettingsRow and searchableSetting signatures to understand this better.

6. browse_code β€” apps/web/src/components/settings/WindowCaptureSettings.tsx

7. browse_code β€” apps/web/src/components/settings/settingsLayout.tsx

πŸ’­ I want to compare this against KeybindingsSettings, the closest existing pattern for a shortcut recorder, to see if there's a shared component like a KeybindingCaptureButton that this page is reinventing locally. That would be a notable finding if so, so I should check KeybindingsSettings.tsx.

8. git_tools β€” data-keybinding-capture

πŸ’­ I want to check the KeybindingsSettings recorder's Tab handling against this to see if there's a variant/size inconsistency, but since that code hasn't changed and I already flagged it in a prior comment, I won't re-post it. Now I'm moving on to check the composer/timeline and CommandPalette additions for new issues, starting with the CameraIcon usage and toastManager error handling, which look consistent, and now examining how actionItems are being pushed. The OpenCommandPaletteDialog function calls getDesktopWindowCaptureBridge during render, which seems fine. I'm now looking at WindowCaptureCoordinator, which plays sound via useClientSettings and dispatches a custom "t3:focus-composer" window event that ChatView listens for β€” I want to check whether the app has an existing convention for custom window events with a "t3:" prefix before treating this as a new pattern.

9. git_tools β€” "t3:

10. git_tools β€” addEventListener("t3code|dispatchEvent(new (Custom)?Event

πŸ’­ There's already a shared previewActionBus module for typed events, but the new code duplicates a raw window event string across ChatView and Coordinator instead of reusing it. I'm weighing whether this counts as a genuine UI-consistency violation or just an architecture nitpick outside the scope of the check.

11. browse_code β€” apps/web/src/components/preview/previewActionBus.ts

12. browse_code β€” apps/web/src/components/ChatView.tsx

πŸ’­ I notice a real inconsistency here: there's an untyped magic-string window event duplicated across two modules, while the repo already has an established typed bus pattern with a consistent namespace prefix. This new event uses a mismatched prefix, which creates a cross-module contract risk that could silently break focus behavior β€” worth flagging even though it leans more architectural than purely UI-related. I'll set that aside as a low-severity note and focus on clearer UI issues first: checking sizing consistency between the window-capture card and other attachment cards, verifying the remove button placement doesn't overlap content, and reviewing the MessagesTimeline grid layout where the fallback branch (no preview URL) seems to be missing a width constraint that other branches have. Next I want to check the WindowCaptureAttachmentDetails text color usage against the already-flagged foreground issue, then move to the onboarding dialog to verify its DialogPopup and Kbd usage stays consistent and whether the open state properly handles hydration timing.

13. browse_code β€” apps/web/src/components/desktop/WindowCaptureOnboardingDialog.tsx

14. browse_code β€” apps/web/src/components/ui/dialog.tsx

πŸ’­ The onboarding dialog now uses DialogPanel correctly, and the header icon spacing looks fine. I want to check the composer card's remove button positioning around lines 3300-3360 to see if it overlaps with the wide card layout.

15. browse_code β€” apps/web/src/components/chat/ChatComposer.tsx

πŸ’­ I'm noting that the upload progress overlay spans the full width of the wide window-capture card, which means it covers the bottom details text including the window title while uploading β€” but since window captures also go through the upload queue, this is a minor, likely pre-existing overlap rather than a new regression. Checking the timeline layout, the window-capture card fits fine within its grid span. Then I'm looking at the animation gating logic β€” since the card is keyed by image.id and remounts on navigation, revisiting a thread draft will retrigger the enter animation, which is a minor but acceptable side effect, especially since reduced-motion is already handled separately. I also confirm the animation follows the existing enter-animation convention used elsewhere, then move to checking where the new keyframes were added in index.css to see if it matches the existing organization of keyframe definitions there.

16. browse_code β€” apps/web/src/index.css

πŸ’­ Since the two previously flagged items shouldn't be reposted, I'm looking for new findings, examining the CommandPalette's use of the desktop window capture bridge and its action item pattern to see if it properly includes keywords or search terms. Looking at WindowCaptureSettings, the "Capture now" row lacks a searchable setting id, but that's probably fine since it's an action rather than a setting. The bigger issue is that captureStatus reads window.desktopBridge directly during render without the SSR guard that the helper function uses β€” this inconsistency could throw during server-side rendering or node-based test rendering, unlike the properly-guarded helper.

17. git_tools β€” window.desktopBridge

18. git_tools β€” isDesktopApp|desktopBridge

πŸ’­ Using window.desktopBridge directly in render is already a common pattern elsewhere in the codebase, so that's not really a concern. I want to check the CommandPalette addition against how other action items are structured, like whether they use await with try/catch or rely on a shared helper for closing the palette.

19. browse_code β€” apps/web/src/components/CommandPalette.tsx

πŸ’­ Checking the palette action, it seems capture doesn't gate on the windowCaptureEnabled setting, consistent with the "capture once" behavior described in the docs. Looking at WindowCaptureCoordinator now, which handles logic and plays a sound but similarly doesn't check that setting. I want to compare the new window-capture card in ChatComposer against ComposerPreviewAnnotationCards.tsx, which already implements a similar thumbnail-plus-label card pattern for preview annotations. It looks like the window-capture card was built inline with its own 224x80 layout rather than reusing that existing shared component, which could be a case of a locally reconstructed primitive worth flagging -- I should check the geometry and structure of the annotation cards to confirm.

20. browse_code β€” apps/web/src/components/chat/ComposerPreviewAnnotationCards.tsx

πŸ’­ I notice the new window-capture card duplicates the existing wide attachment card pattern used for preview annotations, but uses a fixed width instead of intrinsic sizing. The title styling with text-foreground does match the annotation card's convention though, so that part is consistent. Doing the math on the fixed width, the text area for app name and window title comes out to roughly 50px after accounting for the thumbnail, padding, and remove button β€” that's cramped enough to truncate to just a few characters, which seems like a real layout issue worth flagging. With only ~50px left for the title, both lines truncate almost immediately, while the sibling wide attachment card sizes itself intrinsically with a max-width cap instead. This seems worth flagging as a real usability issue since truncating app names to ~7 characters undermines the whole point of attachment identification, diverging from the pattern used elsewhere. I'm also checking the upload progress overlay in the wide composer cardβ€”it spans the bottom of the card and could overlap with the window title text, since the overlay sits right where the details line is positioned. Looking at the fallback rendering in the timeline for window-capture without a preview, the container flexes without a fixed width, so the name text can stretch into leftover spaceβ€”a minor visual quirk rather than a real bug. I also want to verify the focus effect in ChatViewContent, since a global window listener there could cause every mounted ChatView to steal focus if multiple panels exist, and I should confirm the animation gating setting doesn't affect the timeline card since it has no animation anyway. Now I'm checking how Kbd renders when nested inside a Buttonβ€”since Kbd likely carries its own padding and border, stacking it inside an outline button could create a boxed-in-a-box look. I want to grep for other <Kbd usages to see if this nesting pattern is already established elsewhere, like in KeybindingsSettings which instead uses a font-mono Input.

21. git_tools β€” <Kbd

22. browse_code β€” apps/web/src/components/ui/kbd.tsx

πŸ’­ My grep for the Kbd tag came back empty since it was escaped as literal text. I should search for "Kbd" more broadly instead.

23. git_tools β€” Kbd

πŸ’­ I notice KeybindingsSettings renders shortcuts by breaking them into individual Kbd chips grouped together, while the new code just uses a single Kbd with a formatted label. I want to check if there's an existing reusable component for rendering keybinding shortcuts that I should use instead of introducing a different pattern.

24. browse_code β€” apps/web/src/components/settings/KeybindingsSettings.tsx

25. git_tools β€” formatShortcutLabel

πŸ’­ The shortcut label formatting looks fine as is. The main remaining issue is the composer card's fixed w-56 width, which combined with truncation likely makes the text feel cramped compared to the annotation card that doesn't have a fixed width -- I'm leaning toward proposing removing that constraint. Working through the actual pixel math: the 224px card width minus the thumbnail (96px), padding, icon, and gap leaves only about 50px for text, which at typical font sizes truncates the app name to roughly 8 characters. That's not enough for common window titles like "Google Chrome," so the title line ends up nearly unreadable, while the timeline variant has noticeably more room (~248px) for the same content. The fix is to remove the fixed w-56 width and let the card size intrinsically via max-w-full truncation, matching how the sibling annotation card behaves β€” since with min-w-0 and truncate inside a shrink-to-fit flex container, the text block will expand to whatever's available rather than being artificially constrained. This actually touches two separate spots in the same fileβ€”the container sizing and the label's classNameβ€”so since it's a multi-hunk change, I'll skip providing a formal diff and just describe the fix in prose instead. I'm double-checking the MessagesTimeline layout for window-capture images: the img fills its parent with object-cover inside a fixed-width button, and the grid's col-span-2 behavior looks fine even with mixed image heights. Now I'm noticing an inconsistency between the event name used for focus-composer and the existing typed event bus naming convention, which seems worth flagging. That untyped duplicated string is a real cross-module risk since the codebase has an established typed bus pattern for this exact use caseβ€”different namespace prefix and no shared contract to prevent drift. I'll suggest adding a small typed helper to keep both sides in sync, posting it as a low-severity note on the WindowCaptureCoordinator line. It's a bit borderline for the UI consistency check's scope, but since it touches behavioral consistency between components, I think it's defensible to include alongside my main finding. Before finalizing, though, I want to check whether ChatView already has an existing mechanism for focusing the composer from outside triggers, like a keybinding command or shared refβ€”if so, this new window event would be a duplicate mechanism, which is a stronger finding. Let me search the codebase for that.

26. git_tools β€” focusComposer|focusAtEnd

πŸ’­ I notice CommandPalette accesses the composer through a shared composerHandleRef mechanism, which seems to be a shared ref used by non-ChatView components to focus the composer. I want to trace where CommandPalette actually gets that ref from.

27. git_tools β€” composerHandleRef

πŸ’­ That local ref within CommandPalette isn't a global mechanism, so a window event approach is the pragmatic route from a root-level coordinator. There's existing precedent for a typed bus pattern with a namespaced prefix, so this stands as a consistency finding: the untyped magic string duplicated across modules diverges from that established convention. I'll draft two comments: one flagging the fixed width in ChatComposer that leaves too little room for the label lines, and another noting the untyped duplicated event name in WindowCaptureCoordinator that should use a typed bus like the existing one. Both files are in scope for UI consistency review, and the duplicated literal across files is a genuine drift risk since focus could silently break if the strings diverge. I need to verify the exact diff line numbers for each comment to make sure they land on added lines in the diff. I'll keep the diff simple, changing just the container line's width class, and mention adding a max-width cap on the details className as an optional follow-up to prevent long titles from stretching the card unbounded. For the second comment, I'm noting that t3:focus-composer is a bare string duplicated between this file and ChatView.tsx with nothing linking the two occurrences together, which is fragile if one side changes. I'm pointing to the existing previewActionBus.ts pattern as a model for a similar typed dispatch/subscribe module that would keep the contract centralized and consistent with the current event namespace. I'll post this as a failure-state review since there are actionable findings, and briefly note in the summary that earlier findings (Tab handling, text-foreground) remain open without re-posting those comments.

28. modify_pr β€” Reviewed the web UI surface of this feature (composer/timeline capture cards, capture settings page, onboarding dialog, command palette entry, keyframe ownership). The earlier findings on the window-capture-card-enter keyframe ownership, data-keybinding-capture, DialogPanel composition, capability gating of the onboarding dialog, and the disabled state of the sound/flash/animation switches are all resolved. Two remaining items below. Still open from earlier reviews (not re-posted): the shortcut recorder still preventDefault()s Tab so focus cannot leave the button, and WindowCaptureAttachmentDetails still anchors its title to text-foreground while rendering on bg-background/70 inside the text-message-foreground bubble.


Agent Credits: 238 credits