The reviewer who can't read code — when a vision AI caught the bug everyone else missed
Capivara·

The reviewer who can't read code — when a vision AI caught the bug everyone else missed

The bug no linter can catch

Capivara got a light theme in the second design iteration: a ThemeProvider, design tokens (--text-primary, --text-muted, --accent), and an anti-flash script in the HTML so it wouldn’t blink the wrong theme on load. In dark mode, gorgeous. In light mode… supporting text had a gray that melted into the background.

The ironic part: the opposite of what I expected happened in InvitesSection. I had written colors as hardcoded rgba — values I had tuned against the dark background and deemed “good enough”. Against a light background, rgba text tuned for dark becomes nearly invisible.

The tools I use on this project live in the skills store — the first one is free.

No scanner flagged it. Gitleaks doesn’t look at that. Bandit doesn’t look at that. Static analysis never executes the interface — and color isn’t a type error, isn’t a secret, isn’t anything grep can tell apart from “works fine”.

The new reviewer: an AI that sees

Capivara’s review process already had AI layers — a code reviewer that pushes back when my patch is bad. That round, I added a different reviewer: a vision AI (the same image engine I use to analyze covers and screenshots), connected through a model router I run. It got real screenshots of the dashboard in light mode, and I asked the obvious: what’s illegible here?

The answer stung a little. It found two things:

  1. Labels and text in InvitesSection with broken contrast — exactly the hardcoded rgba values.
  2. The user’s email truncated in the header — max-w-[240px] cut addresses mid-string, and with no title attribute there wasn’t even a tooltip to recover the full value.

The funny thing is both bugs had been on screen for days. I looked at the panel every day and my brain had already learned to ignore them.

The patches: one of design, one of process

The InvitesSection fix wasn’t swapping hex for hex. It was accepting that hardcoded color values in a themed component are latent bugs:

// before — nailed-down color, inherits from chance
<label className="block text-[10px] text-[rgba(...)] mb-1">Email</label>

// after — theme token, inherits from the ThemeProvider
<label className="block text-[10px] text-[var(--text-muted)] mb-1">Email</label>

Nine replacements in the file, all in the same category: any color describing a role (supporting text, secondary text, accent) becomes a token — var(--text-muted), var(--text-secondary), var(--text-primary). Brand identity colors can stay hardcoded; hierarchy colors can’t.

The header one was subtler, and I like documenting it because it looks trivial but packs three decisions:

// before — 240px on desktop, hidden on tablet, no tooltip
<span className="hidden sm:inline text-xs sm:text-sm ... truncate max-w-[120px] sm:max-w-[240px]">
  {userLabel || user.email}
</span>

// after — more room, md breakpoint, and the full value in the tooltip
<span title={userLabel || user.email} className="hidden md:inline text-xs sm:text-sm ... truncate max-w-[160px] sm:max-w-[300px]">
  {userLabel || user.email}
</span>

Three decisions in one line: sm:inline became md:inline (on small screens the email doesn’t fit — hide it instead of cutting it), the width ceiling went from 240 to 300px, and the title guarantees that truncating the display never truncates the information — hover returns the whole value.

What a VLM doesn’t replace

Before I turn into a fanboy, an honest assessment. The visual reviewer is great at things code review can’t reach: contrast, truncation, overlap, broken spacing, visual states that only appear with real data. It’s bad (for now) at: logic, security, behavior regression. It looked at the screenshot with no way to know that button fires the wrong query.

So the pipeline settled like this, each layer watching what only it can see:

LayerWhat it catchesWhat it misses
Static scanners (leaks, lint)secrets, dangerous patternseverything visual
Code reviewer (LLM)logic, edge cases, APIpixels, contrast
Visual reviewer (VLM)contrast, truncation, overlaplogic, behavior
Me, day to daycontext and judgmentfatigue, habit blindness

The lesson I took from that round: the most expensive blindness isn’t the scanner’s — it’s mine. I looked at the panel every day and stopped seeing it. The visual reviewer has no memory of “it’s always been like this”, and that’s exactly why it can see.

FixFindingCost
InvitesSection9 hardcoded colors → theme tokens18 lines
Headertruncated email, no tooltip → max-w 300 + title + md1 line

The next review round starts with this layer already on. Dark theme, light theme, and the phone Samuel opens the panel with — now they have a pair of eyes that never gets tired.

~/lifelog — bash
$cat about.txt
╔══════════════════════════════════════╗
║  Samuel Medeiros                    ║
║  Senior Software Engineer           ║
║  Stack: Python · TypeScript · Rust  ║
║  Projetos: Arachne, Dogwalk,        ║
║            Capivara, TatuEngine      ║
╚══════════════════════════════════════╝
      
$

Related entries

Same project, same ground — the stories connect.

Next step

If this was useful, there is more: I write about the same ground in the other projects, and the store holds the tools I use here.

No spam, one-click unsubscribe. I am also here: Skills store · Portfolio