Fixing Coding Agents' Visual Blind Spots: Introducing SceneProof

ReyJ94 · reddit · 2026-08-03

Coding agents (like Claude Code) have a visual blind spot when working with 3D scenes like Three.js—the code compiles successfully and reports it "looks great," but the actual render is often a black screen. Screenshots alone only reveal that a problem exists; they cannot diagnose the root cause (e.g., missing mesh, wrong camera position, or material issues).

SceneProof is a CLI tool designed to bridge this gap:

Core Functionality: Loads real React or Three.js source code, renders it in actual Chrome, and returns the structural scene data behind the pixels.
Key Commands:
tree: Provides the scene graph with bounds, materials, lights, and cameras, turning "why is it invisible" into a direct lookup.
scout: Tries a set of cameras on a target and scores each by actual visibility.
render-region: Re-renders a specific region from the source code at any scale, providing fresh high-definition details rather than just enlarged pixels.
Design Philosophy: Clearly separates "the command ran successfully" from "the output can support a valid judgment." If a render is unjudgeable (e.g., target out of frame), it returns unjudgeable, preventing agents from having the false confidence that a successful build means a perfect design.
Scope: Currently supports TypeScript/JavaScript, React DOM (Tailwind v4), and Three.js (WebGL/WebGPU). Ships with a Skill that can be integrated into agent frameworks like Claude Code.

Original post →

Fixing Coding Agents' Visual Blind Spots: Introducing SceneProof

More from coding & agent