IKANDY AI Shader Error Sandbox:
Self-Healing GPU Shader Generation
via Closed-Loop Compile Feedback

Abstract

This document discloses a novel method implemented in IKANDY for automatically detecting GPU shader compile errors produced by AI-generated GLSL and WGSL code, silently feeding those exact compiler error strings back to the originating large language model, and retrying compilation — up to a configurable maximum — before surfacing any failure to the user. This method is referred to internally as the AI Shader Error Sandbox.

Background & Problem Statement

GPU shader languages (GLSL, WGSL) are strictly typed and structurally rigid. Large language models produce syntactically valid code for common patterns but frequently generate shaders that fail at the GPU driver's compile stage due to:

In existing LLM-integrated tools, shader compile errors are surfaced directly to the user as raw driver error strings, requiring manual code intervention. No prior system in the music visualizer or real-time graphics domain closes the feedback loop by automatically returning compiler output to the model for self-correction.

The Novel Method

// Pseudocode — error sandbox retry loop
const AUTO_RETRY_MAX = 2;
let retryCount = 0;

async function compileWithSandbox(source, userPrompt, mode) {
  const result = await compile(source, mode);

  if (result.ok) {
    retryCount = 0;
    return result;
  }

  if (retryCount >= AUTO_RETRY_MAX) {
    showManualFixButton();
    return result;
  }

  retryCount++;
  setStatus(`compile failed — silent AI auto-fix ${retryCount}/${AUTO_RETRY_MAX}…`);

  const fixed = await requestAiFix({
    originalPrompt: userPrompt,
    failedSource: source,
    compilerError: result.error,  // raw GPU driver string
    mode
  });

  await saveScene(fixed.code);     // overwrite on disk
  return compileWithSandbox(fixed.code, userPrompt, mode);
}

What Makes This Novel

The combination of these elements constitutes the novel method:

1 AI fragment generation with structural boilerplate isolation
2 Raw GPU driver error string capture (not parsed or summarized)
3 Closed-loop LLM re-submission with error as context
4 Silent on-disk overwrite of the corrected fragment
5 Configurable retry ceiling with user-visible fallback
6 Zero user interaction required during the fix loop

Shader Type Coverage

The loop applies at two distinct points: during the AI generation flow (immediately after first compile attempt) and during browser-mount (when a user clicks a saved user scene card). This means a shader saved from an older AI output that later fails due to a driver update will also self-correct on next load.

Implementation Context

The AI Shader Error Sandbox is implemented in IKANDY, an Electron 30-based music visualizer running on Windows. AI generation calls target the Anthropic Claude API (claude-sonnet-4-20250514). GLSL shaders compile into inline WebGL contexts; WGSL shaders compile into WebGPU pipelines via _wgpuUserScene() in scenes.js. IPC handlers in main.js manage file writes with strict filename regex validation and size caps.

This technical disclosure was published at ikandy.app and constitutes prior art as of its effective date.

Prior Art Search Summary

A search of npm, GitHub, and general web sources conducted May 2026 found no implementation combining GPU compiler error capture with closed-loop LLM re-submission for automatic shader self-correction in any real-time graphics, music visualization, or creative coding context. Shader editors with AI-assist features (Cursor, GitHub Copilot in VSCode) offer suggestion and inline diff but require manual user acceptance of corrections and do not execute a silent compile-verify-overwrite loop.