The AI writes the UI and generates the image — then can’t see it.

An AI coding session can produce UI code and images all day long. But it cannot SEE what it just produced, so it acts on visuals while blind to them.

The repo has a name for exactly this gap.

Its own story deck diagnoses the gap out loud.

Slide 2 of the module’s visual-intelligence.html story deck states the problem verbatim — this is self-diagnosed, not our framing.

And here is why words alone can’t close it.

Every round trip through language bleeds fidelity.

Describe → implement → screenshot → describe again. Each pass through words loses a little more, and small deviations compound silently.

To close the loop, the AI needs to actually see — not narrate.

nano-banana gives the AI exactly three vision operations.

One tool module wrapping Google’s Gemini VLM exposes three operations — concrete sight replacing verbal approximation. Its input schema enum lists precisely these three.

But the design is careful about how they combine.

It steers agents away from chaining analyze → generate.

The image_path schema warns against it directly, because turning a picture into words and back reintroduces the exact text bottleneck the tool exists to avoid.

So when it’s time to check the result, one operation does it right.

compare closes the loop.

The AI finally sees the difference precisely instead of guessing at it in prose — the operation the repo itself calls the one that closes the loop.

Now make it dependable in a real session.

A resilience layer makes closing the loop reliable, not a lucky call.

The latest commit adds retry with exponential backoff plus a model fallback chain, so transient failures don’t leave the loop open.

Which leaves one thing for you to take away.