impeccable detect catches real contrast bugs. It passed a Pay button that does nothing

impeccable's detect command is worth a place in CI, for about half of what it reports. On our own homepage it found four text colours that fail WCAG contrast and five labels set below a readable size, all real, in eight seconds and with the same answer twice. The other half of its findings are taste rules, which make a poor merge gate. And a clean exit means less than it looks: a checkout page whose Pay button is wired to nothing scanned with zero findings and exit code 0.

impeccable is #4 on GitHub Trending today, with a fresh engine release on 23 October. The detect command is the part that runs without a model. The docs describe it as scanning for "known design problems without starting an AI conversation": contrast, overflowing text, "common AI design patterns", and drift from a design system. Its author has been pointing people at it since July:

What we ran

impeccable 4.7.0 via npx impeccable@latest detect --json, on macOS arm64 with Node 22. Targets:

1. Five live pages on this site: the homepage, a blog post, /alternatives, /signal, and the homepage again at a 390×844 mobile viewport.
2. Two local pages we wrote to probe the edges. One is deliberately "AI-styled": a purple-to-violet gradient, gradient heading text, glassy nested cards, emoji labels, 10px text and the word "Supercharge". The other is a plain, well-typeset checkout whose Pay button has an empty onclick.

What it got right

Contrast, measured, with the colours named. Four homepage findings were real WCAG 1.4.3 failures. Our status red #ff5470 sat on an off-white card at 2.8:1, a green label measured 3.4:1 on white, and the brand orange #e8652d measured 3.0:1 on a light panel. WCAG asks for 4.5:1 for body-size text. We had a rule for exactly this case, a darker orange for light surfaces, and one component still missed it. A human review had not caught it. The linter did, with a CSS selector pointing at the element.

Text that is too small to use. Five findings flagged 10px and 11px text, most of it labels inside a diagram. That is a judgement call on a diagram, but the measurement is correct, and on a phone it is the right flag.

Deterministic. Two desktop runs of the homepage returned the same 31 findings. That makes it usable as a gate: a new finding in a pull request is a real change, not noise.

The "AI look" rules work on the AI look. The planted page got eight findings: gradient text twice, an "AI color palette", low contrast three times, an overused font and a marketing buzzword. It did not flag the card nested inside a card on that page, though it flagged three on our homepage.

What it flags that is taste

Fourteen of the homepage's 24 non-advisory findings are style opinions rather than defects:

• Nine: a small label above a section heading ("kicker above heading").
• Three: cards inside cards.
• One: a marquee of logos.
• One: that 81% of the text is set in a single typeface.

Each of those is a defensible design note. Not one of them breaks the page for a user. If you run detect as a required check, you will either redesign to its taste or start writing ignore comments, and the second path ends with nobody reading the output. Its own --scope flag helps, but the cleaner split is by rule: gate on contrast, size, overflow and line length, and read the rest as review comments.

What it cannot see

The dead-button checkout is the sharp result. It is a tidy page: one serif font, black text on white, a clear button. Its Pay handler is empty, and its script fires a charge request on page load. detect returned no findings and exit code 0.

That is not a bug in impeccable. It looks at the rendered page and its styles, and it never clicks anything. Its docs are careful about this: "A clean scan means these checks found no reportable problem". The risk is the shortcut a team takes when the check goes green in CI. An agent that builds a page, runs detect and sees exit 0 has proved the page looks deliberate. Whether Pay charges the card once is a different question, and nothing in a static scan answers it.

Verdict

Install it, and gate on its measurable rules. Real contrast failures on a production page, found in seconds and the same twice, are worth the npx. Keep the taste rules advisory. Do not let exit 0 stand in for "the UI works": pair it with a check that drives the flow and reads what happened, because the most expensive bug on our test pages passed with the cleanest score.