Agentic Ui Walkthrough

# iOS/iPad simulator
xcrun simctl io DEVICE_UUID screenshot /tmp/ui-baseline.png

# watchOS simulator
xcrun simctl io DEVICE_UUID screenshot /tmp/watch-baseline.png

Step 2: Walk Through Every State

Document each state you can reach:

• Default/empty state
• Loading state
• Populated/data state
• Error state
• Edge cases (no data, very long text, single item, 100+ items)

Step 3: Analyze Against Heuristics

Use the UX Heuristic Checklist (below) to systematically evaluate each screen.

Step 4: Document Findings

Create a prioritized report with embedded screenshots showing issues.

Step 5: Fix ALL Findings

[!CAUTION] Every P-level finding must be fixed — not deferred, dismissed, or rationalized. Agents must NEVER classify findings as "acceptable," "cosmetic-only," or "not the major issue." UX Sandbox checks are explicitly refined by the project owner and are intentional. If the NRQA or visual review flags it, fix it. The only exception is a single finding estimated at >2 hours of work — document it as a follow-up issue with user confirmation.

• P0-P1: Fix immediately — blocks ship
• P2: Fix in the same session — do NOT defer
• P3: Fix in the same session — do NOT log for "future iteration"

UX Heuristic Checklist

Rate each item ✅ (good), ⚠️ (needs work), or ❌ (broken):

Visual Design

• Visual hierarchy — most important information is prominent
• Consistent spacing — margins and padding follow a system (4/8/12/16/24px)
• Typography — readable font sizes, clear hierarchy (h1 > h2 > body > caption)
• Color contrast — text is readable against background (WCAG AA minimum)
• Alignment — elements are properly aligned (no off-by-1 pixel issues)
• Empty states — empty views have helpful messaging, not blank screens
• Loading states — spinners/skeletons show when data is loading
• Error states — errors are clearly communicated with actionable recovery

Interaction Design

• Responsive — layout works at all supported sizes
• Touch targets — buttons/links are at least 44×44pt (mobile) or clearly clickable (web)
• Feedback — user actions have visible responses (hover, press, loading)
• Navigation — user always knows where they are and how to go back
• Affordance — interactive elements look interactive
• Consistency — similar actions behave the same way everywhere

Content & Data

• Data freshness — displayed data reflects current state
• Overflow handling — long text truncates gracefully (ellipsis, not clipping)
• Number formatting — numbers, dates, times are formatted for readability
• Localization-ready — no hardcoded strings that would break in translation

Performance & Polish

• Smooth transitions — animations are fluid, no jank
• No layout shifts — content doesn't jump around during load
• Image quality — images are sharp on retina displays
• Scroll performance — long lists scroll smoothly

Bug Classification Matrix

Review Report Format

# UI Review: [Feature/Screen Name]

**Date:** YYYY-MM-DD
**Platform:** [Web/iOS/iPadOS/watchOS/Electron]
**Reviewer:** [Agent name]

## Summary
[1-2 sentence overall assessment]

## Screenshots
![state-name](path/to/screenshot.png)

## Findings

### P1 — Must Fix
1. **[Title]** — [Description]
   - Screenshot: [reference]
   - Location: [file:line]
   - Suggested fix: [approach]

### P2 — Should Fix
...

### P3 — Nice to Have
...

## UX Scorecard
| Category | Score |
|----------|-------|
| Visual Design | ✅ / ⚠️ / ❌ |
| Interaction | ✅ / ⚠️ / ❌ |
| Content & Data | ✅ / ⚠️ / ❌ |
| Performance | ✅ / ⚠️ / ❌ |

Platform-Specific Patterns

Electron/Web (Two-Tier Testing via UX Visual Reviewer)

[!IMPORTANT] CC agents run visual reviews by calling the MCP API via the visual-review skill (.agent/skills/visual-review/SKILL.md). CC agents must NOT run Playwright or testing tools directly.

Tier 1 — UX Visual Reviewer Playwright (layout, structure, styling):

• UX Visual Reviewer loads via file:// and http://127.0.0.1:3050/gui/ — tests HTML/CSS/DOM
• Tests layout, responsive breakpoints, DOM structure, form input styling
• Uses injected click handlers to test sidebar view toggling
• Executed by: ux-visual-reviewer infrastructure only (CC agents call MCP API, consume artifacts)

Tier 2 — Live Browser Evidence (data, functional proof):

• http://127.0.0.1:3050/gui/ serves the full GUI with browser-renderer.js for HTTP-based data loading
• UX Visual Reviewer captures screenshots showing real data (workspaces, machines, agents)
• Exercise EVERY sidebar/navigation view — 📂, 📡, 🎯
• Verify data loads: workspace cards, machine list, agent activity
• Check styled components: form inputs, buttons, selects matching dark theme
• Test at multiple viewport sizes (1280×800, 1440×900, 1728×1117)
• Executed by: ux-visual-reviewer infrastructure only

Both tiers required. Tier 1 alone proves only structure. Tier 2 alone proves only one moment in time. Together they provide comprehensive evidence. CC agents verify the returned artifacts meet the gate checklist.

SwiftUI (iOS/iPad/Watch)

• Test both light and dark appearance
• Test Dynamic Type (accessibility font sizes)
• Check Safe Area handling (notch, home indicator)
• Verify NavigationStack push/pop animations
• Test rotating orientation (iPad)

watchOS-Specific

• Verify Digital Crown scrolling
• Check text fits on small screens (40mm vs 49mm)
• Test complications if applicable
• Verify data loads before app is suspended

Anti-Patterns to Catch

1. "It works on my screen" — always test multiple sizes/states
2. Invisible errors — errors that log but don't inform the user
3. Phantom loading — loading indicator that never resolves
4. Dead-end states — no way to recover or navigate away
5. Data staleness — showing cached data without indicating age
6. Truncation without affordance — text cut off with no way to see full content
7. Inconsistent spacing — mixing 10px, 12px, 15px instead of a spacing scale
8. Missing empty states — blank screen when there's no data