mobile-sweep

Open a running web app in headless Chromium at three phone/tablet breakpoints (360×640, 390×844, 768×1024), walk every interactive element + every visible dialog, and flag the four high-impact mobile-layout failure modes:

1. Off-screen overflow — an element extends past document.documentElement.clientWidth.
2. Tiny touch targets — interactive elements under 44×44 CSS px on touch viewports.
3. Clipped text — scrollWidth > clientWidth on a button / input / link (visible label is being cut).
4. Modal width / off-screen modals — dialogs wider than 100vw - 16px or that hang off either side.

Plus an optional static CSS scan that flags grid-template-columns patterns packing 3+ fixed-pixel widths summing to >200px without a @media (max-width: ...) override — the exact pattern that caused today's rewards-edit-row clipping bug.

The actual sweep is done by a bundled Playwright script (scripts/sweep.py). The skill body just orchestrates running it, interprets the findings, and proposes fixes the user can apply directly.

What the scanner exempts by default

These elements are intentionally tiny or off-screen and should not be flagged. The scanner skips them automatically; pass --no-default-ignores to see them anyway.

• Visually-hidden a11y elements — .skip-to-content, .sr-only,

.visually-hidden, .screen-reader-text, anything with aria-label starting with "skip", and structurally-detected SR-only elements (1×1 absolutely-positioned with overflow: hidden / clip-path).

• Inline text links in prose — anchors whose nearest block ancestor

is <p>, <li>, <td>, <dd>, <blockquote>, or any element with a class containing prose. WCAG 2.5.8 explicitly exempts these from the 24×24 target size rule. Exempted only from the small-touch-target check; off-screen and clipped-text still apply.

============================================================ === PRE-FLIGHT === ============================================================

Before running, verify:

• [ ] The user can name a URL where the app is currently reachable — a

Firebase Hosting URL, a localhost:NNNN dev server, or a file:// path to a static index.html. If the user only said "check mobile" with no context, look in the conversation: is there a recently deployed URL, or a firebase deploy output? Use that.

• [ ] Python 3.8+ is available (python3 --version).
• [ ] Playwright + a Chromium browser are installed. Quick probe:

  python3 -c "from playwright.sync_api import sync_playwright" 2>&1

If that errors, run:

  pip install playwright && playwright install chromium

• [ ] If the app requires sign-in to reach the screens being reviewed, the

user must say so — Playwright will land on the sign-in modal otherwise. Options: (a) sign in once with a saved storage state, (b) point at a deep-linked screen, (c) use --open-selectors to dismiss the sign-in modal in-page.

Recovery:

• URL missing → ask for it once. Do not invent one.
• Playwright missing → run the install commands above. The skill is

worthless without it.

• Auth gate → tell the user; offer to test only the sign-in surface and

the public landing.

VALIDATION: python3 scripts/sweep.py --help exits 0. FALLBACK: if Playwright install fails (network, sandbox), stop and tell the user. Do not fake findings.

============================================================ === PHASE 1: SWEEP THE APP === ============================================================

Run the bundled scanner:

python3 ~/.claude/skills/mobile-sweep/scripts/sweep.py <URL>

Useful flags:

The script writes:

• <out>/mobile-sweep-report.md — the formatted report
• <out>/finding-<width>-<state>-<i>.png — a tight screenshot crop of each finding

Exit code: nonzero if any errors are found, so this can gate CI.

VALIDATION: the report file exists and contains at least the header. FALLBACK: if the script crashes (timeout loading the URL, browser launch failure), capture the stderr and surface it to the user before guessing what's wrong.

============================================================ === PHASE 2: INTERPRET AND FIX === ============================================================

Open the generated report. Each finding has:

• The selector + label + breakpoint + state (which click triggered it)
• A 1-line measurement (scrollWidth 320 > clientWidth 80, clipped by 240px)
• A concrete fix suggestion
• A screenshot crop of the offending area

For off-screen and clipped-text findings, the dominant pattern is a rigid CSS grid or flex layout. Don't paper over with overflow: hidden. Instead:

• Switch grid-template-columns: <pile of fixed px> to grid-template-areas

that stack rows under @media (max-width: 640px) so the wide field gets its own row.

• For inputs, ensure min-width: 0 on the parent and width: 100% on the

field so the input fills the available row width.

For small-touch-target findings, add:

@media (hover: none) {
  .my-btn {
    min-width: 44px;
    min-height: 44px;
  }
}

44px is the iOS HIG minimum. 48px is Material's guidance — both are safe.

For modal-too-wide findings, change the modal width to:

width: min(<desktop-width>, calc(100vw - 16px));

For the rigid-grid static warnings, those are the _latent_ version of the runtime issues — the user might not have hit them yet because the modal isn't open, but they will. Apply the same grid-template-areas mobile override.

VALIDATION: a second run of the sweep on the fixed code shows the specific findings resolved. (Don't claim a fix worked without re-running.) FALLBACK: if a fix you applied moved a finding instead of removing it, that's the same bug in a different place — keep iterating until the report is clean.

============================================================ === PHASE 3 (OPTIONAL): WIRE INTO CI === ============================================================

If the user wants this to gate deploys, the script's nonzero exit makes that easy:

# .github/workflows/ci.yml
- name: Mobile sweep
  run: python3 ~/.claude/skills/mobile-sweep/scripts/sweep.py "$DEPLOY_PREVIEW_URL"

Only suggest this when the user has indicated CI is set up.

============================================================ === SELF-REVIEW === ============================================================

Score the output (1–5):

• Complete: Did the sweep run successfully across all three breakpoints?

Did the report surface findings or confirm clean?

• Robust: Did --open-selectors reach the modals the user actually

cares about? Were findings de-duplicated across breakpoints (the same element clipped at 360 and 390 should not double-count)?

• Clean: Is the report concise? Each finding actionable? No noise like

decoration-only elements being flagged as touch targets?

If any score < 4:

• Add --open-selectors to reach unseen modals, or --wait-selectors

if content was being captured before it loaded.

• If touch-target warnings include obvious non-interactive icons, suppress

them by adding their parent button class to a future ignore list (for now, mention them as "likely false positives" in the report).

============================================================ === LEARNINGS CAPTURE === ============================================================

After completing, append one entry to ~/.claude/skills/mobile-sweep/LEARNINGS.md:

## <YYYY-MM-DD> — <project + scan scope>

- **What worked:** <concrete fix suggested by the report that landed
  immediately, or a `--open-selectors` value that reached the right modal>
- **What was awkward:** <false positives, scroll-clipped inputs that were
  intentional, modals that needed a multi-step click sequence to open>
- **Suggested patch:** <one concrete improvement — "default `--wait-selectors`
  for SPAs to `body[data-loaded]`", "add `--ignore-selectors` for known
  decorative icons", "auto-detect modal open buttons">
- **Verdict:** [Smooth / Minor friction / Major friction]

Anti-patterns this skill refuses

• Running on a static screenshot — the whole point is to use a real layout

engine. If Playwright is not installable, stop and say so.

• Auto-applying fixes without showing the report first. The user wants

to see what's broken.

• Suggesting transform: scale() or zoom as fixes. Both break touch

hit-areas and accessibility — fix the layout instead.

• Flagging hover-only elements that legitimately don't need a touch target

(decorative pseudo-elements, etc.). The scanner already filters by visibility + isInteractive; if a class of false positives slips through, surface them in LEARNINGS so /evolve can add an ignore.

• Suggesting overflow: hidden on a parent as a fix for clipped text. It

hides the bug, doesn't fix it.