@tryhamster/gerbil 1.0.0-rc.0 → 1.0.0-rc.2

package/docs/vision.md

@@ -0,0 +1,396 @@
+# Vision Models in Gerbil
+
+Gerbil supports **Vision Language Models (VLMs)** like Ministral 3B that can understand and describe images. This guide covers how to use vision capabilities across all Gerbil interfaces.
+
+## Quick Start
+
+```typescript
+import { Gerbil } from "@tryhamster/gerbil";
+
+const g = new Gerbil();
+await g.loadModel("ministral-3b"); // Vision-capable model
+
+const result = await g.generate("What's in this image?", {
+  images: [{ source: "https://example.com/photo.jpg" }]
+});
+
+console.log(result.text);
+```
+
+## Supported Models
+
+| Model ID | Vision | Reasoning | Context | Size |
+|----------|--------|-----------|---------|------|
+| `ministral-3b` | ✅ | ✅ | 256K | ~2.5GB |
+
+More vision models coming soon as they become available in ONNX format.
+
+## Image Input Types
+
+Gerbil accepts images in several formats:
+
+```typescript
+// URL (recommended for web images)
+images: [{ source: "https://example.com/image.jpg" }]
+
+// Data URI (base64 encoded)
+images: [{ source: "data:image/png;base64,iVBORw0KGgo..." }]
+
+// Local file path (Node.js only, auto-converted to data URI)
+images: [{ source: "/path/to/image.png" }]
+
+// With alt text (optional, provides context)
+images: [{ source: "...", alt: "A photo of a sunset" }]
+```
+
+## Multiple Images
+
+You can pass multiple images for comparison or multi-image understanding:
+
+```typescript
+const result = await g.generate("What's the difference between these two images?", {
+  images: [
+    { source: "https://example.com/before.jpg" },
+    { source: "https://example.com/after.jpg" }
+  ]
+});
+```
+
+## Model Capability Detection
+
+Check if the loaded model supports vision:
+
+```typescript
+await g.loadModel("ministral-3b");
+
+if (g.supportsVision()) {
+  // Use vision features
+} else {
+  // Text-only mode
+}
+```
+
+## Graceful Fallback
+
+If you pass images to a non-vision model, Gerbil will:
+1. Log a warning to console
+2. Ignore the images
+3. Process the text prompt normally
+
+This allows you to write code that works with any model:
+
+```typescript
+// This works with any model - images are used if supported
+const result = await g.generate("Describe this", {
+  images: [{ source: imageUrl }]
+});
+```
+
+---
+
+## AI SDK Integration
+
+Use vision models with Vercel AI SDK v5+:
+
+```typescript
+import { generateText } from "ai";
+import { gerbil } from "@tryhamster/gerbil/ai";
+
+const { text } = await generateText({
+  model: gerbil("ministral-3b"),
+  messages: [
+    {
+      role: "user",
+      content: [
+        { type: "image", image: new URL("https://example.com/photo.jpg") },
+        { type: "text", text: "Describe this image in detail" },
+      ],
+    },
+  ],
+});
+```
+
+### Image Part Formats
+
+The AI SDK integration accepts images in these formats:
+
+```typescript
+// URL object
+{ type: "image", image: new URL("https://...") }
+
+// URL string
+{ type: "image", image: "https://..." }
+
+// Base64 string
+{ type: "image", image: "data:image/png;base64,..." }
+
+// Uint8Array with mime type
+{ type: "image", image: imageBytes, mimeType: "image/png" }
+```
+
+---
+
+## Express & Next.js Integration
+
+### Express
+
+```typescript
+import express from "express";
+import { gerbil } from "@tryhamster/gerbil/express";
+
+const app = express();
+app.use("/ai", gerbil({ model: "ministral-3b" })());
+
+// POST /ai/generate
+// Body: { prompt: "Describe this", images: [{ source: "https://..." }] }
+```
+
+### Next.js App Router
+
+```typescript
+// app/api/chat/route.ts
+import { gerbil } from "@tryhamster/gerbil/next";
+
+export const POST = gerbil.handler({ model: "ministral-3b" });
+
+// Fetch from client:
+// fetch("/api/chat", {
+//   method: "POST",
+//   body: JSON.stringify({
+//     prompt: "What's in this image?",
+//     images: [{ source: dataUri }]
+//   })
+// })
+```
+
+---
+
+## React Hooks (Browser)
+
+### useChat with Images
+
+```tsx
+import { useChat } from "@tryhamster/gerbil/browser";
+
+function VisionChat() {
+  const { 
+    messages, 
+    input, 
+    setInput, 
+    handleSubmit,
+    attachImage,
+    attachedImages,
+    clearImages,
+    sendWithImages,
+  } = useChat({ model: "ministral-3b" });
+
+  const handleFileSelect = (e: React.ChangeEvent<HTMLInputElement>) => {
+    const file = e.target.files?.[0];
+    if (file) {
+      const reader = new FileReader();
+      reader.onload = () => attachImage(reader.result as string);
+      reader.readAsDataURL(file);
+    }
+  };
+
+  return (
+    <div>
+      {/* Messages */}
+      {messages.map(m => (
+        <div key={m.id}>
+          {m.images?.map((img, i) => (
+            <img key={i} src={img} alt="" className="max-w-xs" />
+          ))}
+          <p>{m.content}</p>
+        </div>
+      ))}
+
+      {/* Image attachment */}
+      <input type="file" accept="image/*" onChange={handleFileSelect} />
+      
+      {attachedImages.length > 0 && (
+        <div>
+          📎 {attachedImages.length} image(s) attached
+          <button onClick={clearImages}>Clear</button>
+        </div>
+      )}
+
+      {/* Input */}
+      <form onSubmit={handleSubmit}>
+        <input 
+          value={input} 
+          onChange={e => setInput(e.target.value)}
+          placeholder="Describe the image..."
+        />
+        <button type="submit">Send</button>
+      </form>
+    </div>
+  );
+}
+```
+
+### Direct Image Send
+
+```tsx
+// Send a message with specific images
+sendWithImages("Compare these two photos", [image1DataUri, image2DataUri]);
+```
+
+---
+
+## Built-in Vision Skills
+
+Gerbil includes pre-built skills for common vision tasks:
+
+### Describe Image
+
+```typescript
+import { describeImage } from "@tryhamster/gerbil/skills";
+
+const description = await describeImage({
+  image: "https://example.com/photo.jpg",
+  focus: "details", // "general" | "details" | "text" | "objects" | "scene"
+  format: "bullets", // "paragraph" | "bullets" | "structured"
+});
+```
+
+### Analyze Screenshot
+
+```typescript
+import { analyzeScreenshot } from "@tryhamster/gerbil/skills";
+
+const analysis = await analyzeScreenshot({
+  image: screenshotDataUri,
+  type: "accessibility", // "ui-review" | "accessibility" | "suggestions" | "qa"
+});
+```
+
+### Extract from Image
+
+```typescript
+import { extractFromImage } from "@tryhamster/gerbil/skills";
+
+const extracted = await extractFromImage({
+  image: documentPhoto,
+  extract: "text", // "text" | "data" | "code" | "table" | "diagram"
+  outputFormat: "markdown", // "raw" | "json" | "markdown"
+});
+```
+
+### Compare Images
+
+```typescript
+import { compareImages } from "@tryhamster/gerbil/skills";
+
+const comparison = await compareImages({
+  image1: beforeScreenshot,
+  image2: afterScreenshot,
+  focus: "differences", // "differences" | "similarities" | "detailed"
+});
+```
+
+### Caption Image
+
+```typescript
+import { captionImage } from "@tryhamster/gerbil/skills";
+
+const caption = await captionImage({
+  image: photo,
+  style: "descriptive", // "concise" | "descriptive" | "creative" | "funny"
+});
+```
+
+---
+
+## Performance Tips
+
+### WebGPU Acceleration
+
+Vision models benefit significantly from GPU acceleration:
+
+```typescript
+// Node.js: Uses Chrome backend for WebGPU
+await g.loadModel("ministral-3b"); // Auto-detects WebGPU
+
+// Browser: Native WebGPU
+await g.loadModel("ministral-3b", { device: "webgpu" });
+```
+
+### Image Size
+
+- Larger images take longer to process
+- Consider resizing before sending to the model
+- 512x512 to 1024x1024 is generally optimal
+
+### Caching
+
+The model caches in the browser's IndexedDB (via Chrome backend in Node.js), so subsequent loads are fast.
+
+---
+
+## Troubleshooting
+
+### "Model doesn't support vision"
+
+Make sure you're using a vision-capable model like `ministral-3b`.
+
+### Slow image processing
+
+- Ensure WebGPU is being used (check `g.getDeviceMode()`)
+- Resize large images before sending
+- In Node.js, the Chrome backend provides GPU acceleration
+
+### Image not loading
+
+- Check the URL is accessible
+- For local files, ensure the path is absolute
+- Base64 data URIs must include the mime type prefix
+
+---
+
+## API Reference
+
+### ImageInput
+
+```typescript
+interface ImageInput {
+  /** Image source: URL, base64 data URI, or local file path */
+  source: string;
+  /** Optional alt text for context */
+  alt?: string;
+}
+```
+
+### GenerateOptions (with images)
+
+```typescript
+interface GenerateOptions {
+  // ... standard options ...
+  
+  /** Images to include (only used if model supports vision) */
+  images?: ImageInput[];
+}
+```
+
+### supportsVision()
+
+```typescript
+g.supportsVision(): boolean
+```
+
+Returns `true` if the loaded model supports vision input.
+
+### ModelConfig
+
+```typescript
+interface ModelConfig {
+  // ... standard properties ...
+  
+  /** Whether model supports vision/image input */
+  supportsVision?: boolean;
+  
+  /** Size of vision encoder (if applicable) */
+  visionEncoderSize?: string;
+}
+```
+

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tryhamster/gerbil",
-  "version": "1.0.0-rc.0",
+  "version": "1.0.0-rc.2",
   "description": "Local LLM inference for Node.js. GPU-accelerated. Zero config. Works standalone or with Vercel AI SDK.",
   "type": "module",
   "main": "dist/index.mjs",
@@ -66,28 +66,17 @@
       "types": "./dist/browser/index.d.mts"
     }
   },
-  "scripts": {
-    "build": "tsdown",
-    "dev": "tsx src/cli/index.ts",
-    "typecheck": "tsc --noEmit",
-    "check": "ultracite check",
-    "fix": "ultracite fix",
-    "test": "vitest run",
-    "test:watch": "vitest",
-    "prepublishOnly": "pnpm build",
-    "changeset": "changeset",
-    "release": "changeset publish",
-    "prepare": "lefthook install"
-  },
   "dependencies": {
     "@huggingface/hub": "^2.7.1",
     "@huggingface/transformers": "^3.8.0",
     "chalk": "^5.3.0",
     "cli-progress": "^3.12.0",
     "commander": "^12.1.0",
+    "kokoro-js": "^1.2.1",
     "onnxruntime-web": "^1.21.0-dev.20250114-228dd16893",
     "ora": "^8.0.1",
     "puppeteer-core": "^24.31.0",
+    "react": "^19.0.0",
     "webgpu": "^0.3.8",
     "zod": "^3.23.0"
   },
@@ -101,8 +90,7 @@
     "hono": ">=4.0.0",
     "langchain": ">=0.1.0",
     "llamaindex": ">=0.1.0",
-    "next": ">=14.0.0",
-    "react": ">=18.0.0"
+    "next": ">=14.0.0"
   },
   "peerDependenciesMeta": {
     "@ai-sdk/provider": {
@@ -158,7 +146,6 @@
     "ink-spinner": "^5.0.0",
     "ink-text-input": "^6.0.0",
     "lefthook": "^2.0.5",
-    "react": "^19.0.0",
     "tsdown": "^0.17.0-beta.3",
     "tsx": "^4.15.0",
     "typescript": "^5.4.5",
@@ -203,5 +190,17 @@
   ],
   "publishConfig": {
     "access": "public"
+  },
+  "scripts": {
+    "build": "tsdown",
+    "dev": "tsx src/cli/index.ts",
+    "typecheck": "tsc --noEmit",
+    "check": "ultracite check",
+    "fix": "ultracite fix",
+    "test": "vitest run",
+    "test:watch": "vitest",
+    "changeset": "changeset",
+    "version": "changeset version",
+    "release": "pnpm publish && changeset tag"
   }
-}
+}

package/dist/auto-update-BbNHbSU1.mjs

@@ -1,3 +0,0 @@
-import { i as installUpdate, n as checkForUpdate, r as compareVersions, t as CURRENT_VERSION } from "./cli.mjs";
-
-export { CURRENT_VERSION, checkForUpdate, installUpdate };

package/dist/chrome-backend-C5Un08O4.mjs.map

@@ -1 +0,0 @@
-{"version":3,"file":"chrome-backend-C5Un08O4.mjs","names":["files: { path: string; size?: number }[]","globalBrowser: Browser | null","globalBrowserPromise: Promise<Browser> | null","globalServer: Server | null","pid: number | null","err: any"],"sources":["../src/core/chrome-backend.ts"],"sourcesContent":["/**\n * Chrome DevTools Protocol Backend for WebGPU Inference\n *\n * Uses headless Chrome as a WebGPU accelerator for Node.js environments.\n * Provides the same performance as browser inference (~100+ tok/s with q4f16).\n */\n\nimport { execSync } from \"child_process\";\nimport { existsSync, mkdirSync, readFileSync, rmSync, unlinkSync, writeFileSync } from \"fs\";\nimport { createServer, type Server } from \"http\";\nimport { homedir } from \"os\";\nimport { join } from \"path\";\nimport puppeteer, { type Browser, type CDPSession, type Page } from \"puppeteer-core\";\n\n// Persistent cache directory for Chrome profile (keeps model cache between runs)\nconst GERBIL_CACHE_DIR = join(homedir(), \".gerbil\", \"chrome-cache\");\nconst WS_ENDPOINT_FILE = join(GERBIL_CACHE_DIR, \"ws-endpoint.txt\");\nconst CACHED_MODELS_FILE = join(homedir(), \".gerbil\", \"cached-models.json\");\n\n// ============================================\n// Cached Models Tracking\n// ============================================\n\ntype CachedModelEntry = {\n  modelId: string;\n  downloadedAt: string;\n  lastUsed: string;\n  sizeBytes?: number;\n};\n\n/** Get list of models cached in Chrome's IndexedDB */\nexport function getChromeCachedModels(): CachedModelEntry[] {\n  try {\n    if (!existsSync(CACHED_MODELS_FILE)) {\n      return [];\n    }\n    const data = JSON.parse(readFileSync(CACHED_MODELS_FILE, \"utf-8\"));\n    return data.models || [];\n  } catch {\n    return [];\n  }\n}\n\n/** Fetch model size from HuggingFace API */\nasync function fetchModelSize(modelId: string): Promise<number | undefined> {\n  try {\n    // Try to get ONNX file size from tree API\n    const treeRes = await fetch(`https://huggingface.co/api/models/${modelId}/tree/main/onnx`);\n    if (treeRes.ok) {\n      const files: { path: string; size?: number }[] = await treeRes.json();\n      // Prefer q4f16 > q4 > fp16 > any .onnx file\n      const q4f16 = files.find((f) => f.path.includes(\"q4f16\") && f.path.endsWith(\".onnx\"));\n      const q4 = files.find(\n        (f) => f.path.includes(\"q4\") && !f.path.includes(\"f16\") && f.path.endsWith(\".onnx\"),\n      );\n      const fp16 = files.find((f) => f.path.includes(\"fp16\") && f.path.endsWith(\".onnx\"));\n      const anyOnnx = files.find((f) => f.path.endsWith(\".onnx\"));\n      const bestFile = q4f16 || q4 || fp16 || anyOnnx;\n      if (bestFile?.size) {\n        return bestFile.size;\n      }\n    }\n\n    // Fallback to model info API\n    const res = await fetch(`https://huggingface.co/api/models/${modelId}`);\n    if (res.ok) {\n      const info = await res.json();\n      return info.usedStorage;\n    }\n  } catch {\n    // Ignore fetch errors\n  }\n  return;\n}\n\n/** Track a model as cached */\nexport function trackCachedModel(modelId: string, sizeBytes?: number): void {\n  try {\n    const dir = join(homedir(), \".gerbil\");\n    if (!existsSync(dir)) {\n      mkdirSync(dir, { recursive: true });\n    }\n\n    const models = getChromeCachedModels();\n    const existing = models.find((m) => m.modelId === modelId);\n    const now = new Date().toISOString();\n\n    if (existing) {\n      existing.lastUsed = now;\n      if (sizeBytes) {\n        existing.sizeBytes = sizeBytes;\n      }\n    } else {\n      models.push({\n        modelId,\n        downloadedAt: now,\n        lastUsed: now,\n        sizeBytes,\n      });\n    }\n\n    writeFileSync(CACHED_MODELS_FILE, JSON.stringify({ models }, null, 2));\n\n    // Fetch size in background if not provided\n    if (!sizeBytes) {\n      fetchModelSize(modelId)\n        .then((size) => {\n          if (size) {\n            const updatedModels = getChromeCachedModels();\n            const model = updatedModels.find((m) => m.modelId === modelId);\n            if (model) {\n              model.sizeBytes = size;\n              writeFileSync(CACHED_MODELS_FILE, JSON.stringify({ models: updatedModels }, null, 2));\n            }\n          }\n        })\n        .catch(() => {});\n    }\n  } catch {\n    // Ignore tracking errors\n  }\n}\n\n/** Remove a model from cache tracking */\nexport function untrackCachedModel(modelId: string): void {\n  try {\n    const models = getChromeCachedModels().filter((m) => m.modelId !== modelId);\n    writeFileSync(CACHED_MODELS_FILE, JSON.stringify({ models }, null, 2));\n  } catch {\n    // Ignore tracking errors\n  }\n}\n\n/** Refresh sizes for cached models that don't have them */\nexport async function refreshCachedModelSizes(): Promise<void> {\n  try {\n    const models = getChromeCachedModels();\n    const needsSize = models.filter((m) => !m.sizeBytes);\n    if (needsSize.length === 0) {\n      return;\n    }\n\n    // Fetch sizes in parallel (max 3 at a time)\n    const batchSize = 3;\n    for (let i = 0; i < needsSize.length; i += batchSize) {\n      const batch = needsSize.slice(i, i + batchSize);\n      await Promise.all(\n        batch.map(async (model) => {\n          const size = await fetchModelSize(model.modelId);\n          if (size) {\n            model.sizeBytes = size;\n          }\n        }),\n      );\n    }\n\n    // Save updated models\n    writeFileSync(CACHED_MODELS_FILE, JSON.stringify({ models }, null, 2));\n  } catch {\n    // Ignore refresh errors\n  }\n}\n\n// Fixed port for local server - IndexedDB cache is origin-scoped, so using a\n// consistent port ensures the model cache persists between runs\n// Port 43724 = \"GERBI\" on phone keypad (GERBIL=437245 is too big for port range)\nconst GERBIL_LOCAL_PORT = 43_724;\n\n// Global singletons - multiple Gerbil instances share browser and server\nlet globalBrowser: Browser | null = null;\nlet globalBrowserPromise: Promise<Browser> | null = null;\nlet globalServer: Server | null = null;\nlet globalServerPort = 0;\n\n// Page tracking for memory management\nlet activePagesCount = 0;\nconst MAX_CONCURRENT_PAGES = 5; // Limit to prevent dev mistakes\n\n// ============================================\n// Types\n// ============================================\n\nexport type ChromeBackendOptions = {\n  /** Custom Chrome executable path */\n  chromePath?: string;\n  /** Model ID to load */\n  modelId?: string;\n  /** Progress callback */\n  onProgress?: (info: { status: string; progress?: number; file?: string }) => void;\n  /** Token callback for streaming */\n  onToken?: (token: { text: string; state: string; numTokens: number; tps: number }) => void;\n};\n\nexport type GenerateOptions = {\n  maxTokens?: number;\n  temperature?: number;\n  topP?: number;\n  topK?: number;\n  thinking?: boolean;\n  system?: string;\n  onToken?: (token: { text: string; state: string; numTokens: number; tps: number }) => void;\n};\n\n// ============================================\n// Chrome Path Detection\n// ============================================\n\nconst CHROME_PATHS = {\n  darwin: [\n    \"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome\",\n    \"/Applications/Google Chrome Canary.app/Contents/MacOS/Google Chrome Canary\",\n    \"/Applications/Chromium.app/Contents/MacOS/Chromium\",\n    \"/Applications/Microsoft Edge.app/Contents/MacOS/Microsoft Edge\",\n    \"/Applications/Brave Browser.app/Contents/MacOS/Brave Browser\",\n  ],\n  linux: [\n    \"google-chrome-stable\",\n    \"google-chrome\",\n    \"chromium-browser\",\n    \"chromium\",\n    \"microsoft-edge\",\n    \"brave-browser\",\n  ],\n  win32: [\n    \"C:\\\\Program Files\\\\Google\\\\Chrome\\\\Application\\\\chrome.exe\",\n    \"C:\\\\Program Files (x86)\\\\Google\\\\Chrome\\\\Application\\\\chrome.exe\",\n    `${process.env.LOCALAPPDATA}\\\\Google\\\\Chrome\\\\Application\\\\chrome.exe`,\n    \"C:\\\\Program Files\\\\Microsoft\\\\Edge\\\\Application\\\\msedge.exe\",\n    \"C:\\\\Program Files\\\\BraveSoftware\\\\Brave-Browser\\\\Application\\\\brave.exe\",\n  ],\n};\n\nfunction findChrome(): string {\n  // Check env override first\n  if (process.env.CHROME_PATH) {\n    return process.env.CHROME_PATH;\n  }\n\n  const platform = process.platform as \"darwin\" | \"linux\" | \"win32\";\n  const paths = CHROME_PATHS[platform] || [];\n\n  for (const p of paths) {\n    try {\n      if (platform === \"linux\") {\n        // For Linux, check if command exists in PATH\n        execSync(`which ${p}`, { stdio: \"ignore\" });\n        return p;\n      }\n      // For macOS/Windows, use fs.existsSync (portable)\n      if (existsSync(p)) {\n        return p;\n      }\n    } catch {}\n  }\n\n  throw new Error(\"Chrome not found. Install Chrome or set CHROME_PATH environment variable.\");\n}\n\n// ============================================\n// Chrome Launch Flags\n// ============================================\n\nfunction getChromeFlags(userDataDir: string, _debuggingPort: number): string[] {\n  const flags = [\"--no-sandbox\", `--user-data-dir=${userDataDir}`];\n\n  // Platform-specific WebGPU flags\n  if (process.platform === \"linux\") {\n    // Linux: use Vulkan backend for WebGPU\n    flags.push(\n      \"--enable-unsafe-webgpu\",\n      \"--enable-features=Vulkan\",\n      \"--use-angle=vulkan\",\n      \"--disable-vulkan-surface\",\n    );\n  } else if (process.platform === \"darwin\") {\n    // macOS: WebGPU uses Metal by default, minimal flags needed\n    // Only add --enable-unsafe-webgpu if WebGPU is disabled (rare)\n    // For now, try without it to avoid triggering GPU bugs\n  } else {\n    // Windows: use default DirectX/D3D12 backend\n    flags.push(\"--enable-unsafe-webgpu\");\n  }\n\n  return flags;\n}\n\n// ============================================\n// Worker Page HTML\n// ============================================\n\nfunction getWorkerPageHTML(modelPath: string): string {\n  return `\n<!DOCTYPE html>\n<html>\n<head>\n  <title>Gerbil WebGPU Backend</title>\n  <script type=\"module\">\n    import {\n      AutoTokenizer,\n      AutoModelForCausalLM,\n      TextStreamer,\n      InterruptableStoppingCriteria,\n    } from \"https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.8.0\";\n\n    class ModelPipeline {\n      static tokenizer = null;\n      static model = null;\n      static modelId = \"${modelPath}\";\n\n      static async getInstance(progressCallback) {\n        if (!this.tokenizer) {\n          this.tokenizer = await AutoTokenizer.from_pretrained(this.modelId, {\n            progress_callback: progressCallback,\n          });\n        }\n\n        if (!this.model) {\n          this.model = await AutoModelForCausalLM.from_pretrained(this.modelId, {\n            dtype: \"q4f16\",\n            device: \"webgpu\",\n            progress_callback: progressCallback,\n          });\n        }\n\n        return { tokenizer: this.tokenizer, model: this.model };\n      }\n    }\n\n    const stoppingCriteria = new InterruptableStoppingCriteria();\n    let pastKeyValuesCache = null;\n    let totalTokensInCache = 0;\n    \n    // Context length for auto-reset (Qwen3 default: 2048)\n    // Cache beyond this provides no benefit and wastes memory\n    const CONTEXT_LENGTH = 2048;\n\n    // Auto-load model on page init\n    (async function() {\n      console.log(JSON.stringify({ type: \"progress\", status: \"Loading model...\" }));\n      \n      try {\n        const { tokenizer, model } = await ModelPipeline.getInstance((progress) => {\n          if (progress.status === \"progress\" && progress.file) {\n            console.log(JSON.stringify({\n              type: \"progress\",\n              status: \"progress\",\n              file: progress.file,\n              progress: Math.round(progress.progress || 0),\n            }));\n          }\n        });\n\n        console.log(JSON.stringify({ type: \"progress\", status: \"Compiling shaders...\" }));\n        const warmupInputs = tokenizer(\"a\");\n        await model.generate({ ...warmupInputs, max_new_tokens: 1 });\n\n        console.log(JSON.stringify({ type: \"ready\" }));\n      } catch (error) {\n        console.log(JSON.stringify({ type: \"error\", error: error.message || String(error) }));\n      }\n    })();\n\n    window.gerbilGenerate = async function(messages, options = {}) {\n      const { maxTokens = 256, temperature = 0.7, topP = 0.9, topK = 20, thinking = false } = options;\n      \n      // Auto-reset KV cache if it exceeds context length\n      // This prevents unbounded memory growth while preserving performance\n      if (totalTokensInCache > CONTEXT_LENGTH) {\n        console.log(JSON.stringify({ \n          type: \"cache_reset\", \n          reason: \"context_exceeded\",\n          tokensInCache: totalTokensInCache,\n          contextLength: CONTEXT_LENGTH\n        }));\n        pastKeyValuesCache = null;\n        totalTokensInCache = 0;\n      }\n\n      try {\n        const { tokenizer, model } = await ModelPipeline.getInstance();\n\n        const inputs = tokenizer.apply_chat_template(messages, {\n          add_generation_prompt: true,\n          return_dict: true,\n          enable_thinking: thinking,\n        });\n\n        let state = \"answering\";\n        let prevState = \"answering\";\n        const [START_THINKING_TOKEN_ID, END_THINKING_TOKEN_ID] = tokenizer.encode(\n          \"<think></think>\",\n          { add_special_tokens: false }\n        );\n\n        let startTime = null;\n        let numTokens = 0;\n\n        const tokenCallback = (tokens) => {\n          startTime ??= performance.now();\n          numTokens++;\n          \n          const tokenId = Number(tokens[0]);\n          if (tokenId === START_THINKING_TOKEN_ID) {\n            state = \"thinking\";\n          } else if (tokenId === END_THINKING_TOKEN_ID) {\n            state = \"answering\";\n          }\n        };\n\n        const streamCallback = (text) => {\n          const tps = startTime ? (numTokens / (performance.now() - startTime)) * 1000 : 0;\n          \n          // Inject <think> markers when state changes (since skip_special_tokens removes them)\n          let outputText = text;\n          if (thinking) {\n            if (state === \"thinking\" && prevState !== \"thinking\") {\n              outputText = \"<think>\" + text;\n            } else if (state === \"answering\" && prevState === \"thinking\") {\n              outputText = \"</think>\" + text;\n            }\n          }\n          prevState = state;\n          \n          console.log(JSON.stringify({ type: \"token\", text: outputText, state, numTokens, tps }));\n        };\n\n        const streamer = new TextStreamer(tokenizer, {\n          skip_prompt: true,\n          skip_special_tokens: true,\n          callback_function: streamCallback,\n          token_callback_function: tokenCallback,\n        });\n\n        console.log(JSON.stringify({ type: \"start\" }));\n\n        const { past_key_values, sequences } = await model.generate({\n          ...inputs,\n          past_key_values: pastKeyValuesCache,\n          do_sample: temperature > 0,\n          temperature: temperature > 0 ? temperature : undefined,\n          top_p: topP,\n          top_k: topK,\n          max_new_tokens: maxTokens,\n          streamer,\n          stopping_criteria: stoppingCriteria,\n          return_dict_in_generate: true,\n        });\n\n        pastKeyValuesCache = past_key_values;\n        \n        // Track total tokens in cache (input + generated)\n        const inputLength = inputs.input_ids.dims[1];\n        totalTokensInCache += inputLength + numTokens;\n\n        const endTime = performance.now();\n        const totalTime = startTime ? endTime - startTime : 0;\n        \n        // Extract only the generated tokens (exclude input prompt)\n        const generatedTokens = sequences.slice(null, [inputLength, null]);\n        const decoded = tokenizer.batch_decode(generatedTokens, { skip_special_tokens: true });\n\n        console.log(JSON.stringify({\n          type: \"complete\",\n          text: decoded[0] || \"\",\n          numTokens,\n          totalTime,\n          tps: totalTime > 0 ? (numTokens / totalTime) * 1000 : 0,\n          tokensInCache: totalTokensInCache,\n        }));\n\n        return decoded[0] || \"\";\n      } catch (error) {\n        console.log(JSON.stringify({ type: \"error\", error: error.message || String(error) }));\n        throw error;\n      }\n    };\n\n    window.gerbilInterrupt = function() {\n      stoppingCriteria.interrupt();\n    };\n\n    window.gerbilReset = function() {\n      pastKeyValuesCache = null;\n      totalTokensInCache = 0;\n      stoppingCriteria.reset();\n      console.log(JSON.stringify({ type: \"cache_reset\", reason: \"manual\" }));\n    };\n\n    // Signal that the page is ready for commands\n    console.log(JSON.stringify({ type: \"init\" }));\n  </script>\n</head>\n<body>\n  <h1>Gerbil WebGPU Backend</h1>\n  <p>This page provides WebGPU inference for the Gerbil CLI.</p>\n</body>\n</html>\n`;\n}\n\n// ============================================\n// Chrome GPU Backend\n// ============================================\n\nexport class ChromeGPUBackend {\n  private browser: Browser | null = null;\n  private page: Page | null = null;\n  private cdp: CDPSession | null = null;\n  private serverPort = 0;\n  private userDataDir: string = GERBIL_CACHE_DIR; // Always use shared cache\n  private readonly modelId: string;\n  private isReady = false;\n  private readonly messageHandlers: Map<string, (data: any) => void> = new Map();\n  private pendingRejects: Array<(err: Error) => void> = []; // Track pending waits for cleanup\n\n  public server: Server | null = null;\n\n  private constructor(modelId: string) {\n    this.modelId = modelId;\n  }\n\n  /**\n   * Create and initialize a Chrome GPU backend\n   */\n  static async create(options: ChromeBackendOptions = {}): Promise<ChromeGPUBackend> {\n    const modelId = options.modelId || \"onnx-community/Qwen3-0.6B-ONNX\";\n    const backend = new ChromeGPUBackend(modelId);\n    await backend.launch(options);\n    return backend;\n  }\n\n  /**\n   * Get existing browser or launch a new one (singleton pattern)\n   * Multiple Gerbil instances share the same browser process\n   */\n  private async getOrCreateBrowser(\n    chromePath: string,\n    options: ChromeBackendOptions,\n  ): Promise<Browser> {\n    // If we already have a global browser, reuse it\n    if (globalBrowser?.connected) {\n      options.onProgress?.({ status: \"Reusing existing Chrome...\" });\n      return globalBrowser;\n    }\n\n    // If another caller is launching, wait for them\n    if (globalBrowserPromise) {\n      options.onProgress?.({ status: \"Waiting for Chrome startup...\" });\n      return globalBrowserPromise;\n    }\n\n    // Try to connect to existing browser via saved WebSocket endpoint\n    if (existsSync(WS_ENDPOINT_FILE)) {\n      try {\n        const wsEndpoint = readFileSync(WS_ENDPOINT_FILE, \"utf-8\").trim();\n        options.onProgress?.({ status: \"Connecting to existing Chrome...\" });\n        globalBrowser = await puppeteer.connect({\n          browserWSEndpoint: wsEndpoint,\n        });\n        return globalBrowser;\n      } catch {\n        // Stale endpoint, remove it and launch fresh\n        try {\n          unlinkSync(WS_ENDPOINT_FILE);\n        } catch {}\n      }\n    }\n\n    // Launch new browser\n    globalBrowserPromise = this.launchBrowser(chromePath, options);\n    try {\n      globalBrowser = await globalBrowserPromise;\n      return globalBrowser;\n    } finally {\n      globalBrowserPromise = null;\n    }\n  }\n\n  /**\n   * Launch a new Chrome browser instance\n   */\n  private async launchBrowser(\n    chromePath: string,\n    _options: ChromeBackendOptions,\n  ): Promise<Browser> {\n    const debuggingPort = 9222 + Math.floor(Math.random() * 1000);\n\n    // Clean up stale lock file if Chrome crashed\n    const lockFile = join(this.userDataDir, \"SingletonLock\");\n    if (existsSync(lockFile)) {\n      try {\n        unlinkSync(lockFile);\n        await new Promise((r) => setTimeout(r, 200));\n      } catch {}\n    }\n\n    // Use new headless mode - more compatible with WebGPU than old headless\n    // Previous crashes were caused by killing our own server, not headless mode\n    const browser = await puppeteer.launch({\n      executablePath: chromePath,\n      headless: true, // Standard headless mode - crashes before were from killing our own server\n      args: [\n        ...getChromeFlags(this.userDataDir, debuggingPort),\n        \"--enable-gpu\",\n        \"--no-first-run\",\n        \"--no-default-browser-check\",\n        \"--disable-background-timer-throttling\",\n        \"--disable-renderer-backgrounding\",\n        \"--disable-dev-shm-usage\",\n      ],\n      handleSIGINT: false,\n      handleSIGTERM: false,\n      handleSIGHUP: false,\n    });\n\n    // Save WebSocket endpoint for reconnection\n    writeFileSync(WS_ENDPOINT_FILE, browser.wsEndpoint());\n\n    // Clean up endpoint file when browser closes\n    browser.on(\"disconnected\", () => {\n      globalBrowser = null;\n      try {\n        unlinkSync(WS_ENDPOINT_FILE);\n      } catch {}\n    });\n\n    return browser;\n  }\n\n  /**\n   * Launch Chrome and initialize the worker page\n   */\n  private async launch(options: ChromeBackendOptions): Promise<void> {\n    // Check page limit to prevent memory issues during development\n    if (activePagesCount >= MAX_CONCURRENT_PAGES) {\n      throw new Error(\n        `Maximum concurrent pages (${MAX_CONCURRENT_PAGES}) reached. ` +\n          \"Call dispose() on old Gerbil instances to free resources. \" +\n          `Currently active: ${activePagesCount}`,\n      );\n    }\n\n    const chromePath = options.chromePath || findChrome();\n\n    // Use persistent cache directory (keeps model downloads between runs)\n    this.userDataDir = GERBIL_CACHE_DIR;\n    if (!existsSync(this.userDataDir)) {\n      mkdirSync(this.userDataDir, { recursive: true });\n    }\n\n    // Start tiny HTTP server to serve the worker page\n    // (Required because file:// + ES modules + CDN imports doesn't work due to CORS)\n    const html = getWorkerPageHTML(this.modelId);\n    await this.startServer(html);\n\n    options.onProgress?.({ status: \"Starting Chrome...\" });\n\n    // Get or create the shared browser instance\n    this.browser = await this.getOrCreateBrowser(chromePath, options);\n\n    // Create page and set up CDP session\n    this.page = await this.browser.newPage();\n    this.cdp = await this.page.createCDPSession();\n\n    // Increment active page counter\n    activePagesCount++;\n    options.onProgress?.({\n      status: `Active pages: ${activePagesCount}/${MAX_CONCURRENT_PAGES}`,\n    });\n\n    // Listen for browser disconnect (OOM kill, crash, etc.)\n    this.browser.on(\"disconnected\", () => {\n      console.error(\"[Chrome] Browser disconnected unexpectedly\");\n      this.isReady = false;\n      this.browser = null;\n      this.page = null;\n      this.cdp = null;\n      // Fail fast: reject any pending waits so callers don't hang\n      this.rejectPendingWaits(new Error(\"CHROME_DISCONNECTED\"));\n    });\n\n    // Enable console API events and exceptions\n    await this.cdp.send(\"Runtime.enable\");\n    await this.cdp.send(\"Runtime.setAsyncCallStackDepth\", { maxDepth: 32 });\n\n    // Set up console message handler\n    this.cdp.on(\"Runtime.consoleAPICalled\", (event) => {\n      const text = event.args.map((a: any) => a.value || a.description || \"\").join(\" \");\n\n      if (event.type === \"log\" && event.args[0]?.value) {\n        try {\n          const data = JSON.parse(event.args[0].value);\n          this.handleMessage(data, options);\n        } catch {\n          // Not JSON - only log short messages, skip large data dumps (like KV cache)\n          if (\n            text.length < 500 &&\n            !text.includes(\"Float32Array\") &&\n            !text.includes(\"past_key_values\")\n          ) {\n            // Uncomment for debugging: console.log(\"[Chrome Log]\", text);\n          }\n        }\n      } else if (event.type === \"error\" || event.type === \"warning\") {\n        // Filter out noisy messages\n        if (\n          !(\n            text.includes(\"onnxruntime\") ||\n            text.includes(\"content-length\") ||\n            text.includes(\"Float32Array\") ||\n            text.includes(\"past_key_values\")\n          ) &&\n          text.length < 1000\n        ) {\n          console.error(`[Chrome ${event.type}]`, text);\n        }\n      }\n    });\n\n    // Listen for exceptions\n    this.cdp.on(\"Runtime.exceptionThrown\", (event) => {\n      const errText =\n        event.exceptionDetails?.text || event.exceptionDetails?.exception?.description || \"\";\n      // Skip noisy tensor/KV cache dumps\n      if (\n        errText.includes(\"Float32Array\") ||\n        errText.includes(\"past_key_values\") ||\n        errText.length > 1000\n      ) {\n        return;\n      }\n      console.error(\"[Chrome Exception]\", errText);\n    });\n\n    // Navigate to our HTTP server - model loads automatically\n    await this.page.goto(`http://127.0.0.1:${this.serverPort}/`, {\n      waitUntil: \"domcontentloaded\",\n      timeout: 30_000,\n    });\n\n    // Wait for model to be ready (loads automatically on page init)\n    await this.waitForMessage(\"ready\", 300_000); // 5 min timeout for model download\n\n    this.isReady = true;\n    options.onProgress?.({ status: \"Ready (WebGPU)!\" });\n\n    // Track this model as cached\n    trackCachedModel(this.modelId);\n  }\n\n  /**\n   * Handle incoming messages from the page\n   */\n  private handleMessage(data: any, options: ChromeBackendOptions): void {\n    const { type, ...rest } = data;\n\n    // Call registered handler\n    const handler = this.messageHandlers.get(type);\n    if (handler) {\n      handler(rest);\n    }\n\n    // Also call option callbacks\n    if (type === \"progress\") {\n      options.onProgress?.(rest);\n    } else if (type === \"token\") {\n      options.onToken?.(rest);\n    }\n  }\n\n  /**\n   * Wait for a specific message type\n   */\n  private waitForMessage(type: string, timeout = 30_000): Promise<any> {\n    return new Promise((resolve, reject) => {\n      // Track this reject for cleanup on browser disconnect\n      this.pendingRejects.push(reject);\n\n      const cleanup = () => {\n        clearTimeout(timer);\n        this.messageHandlers.delete(type);\n        const idx = this.pendingRejects.indexOf(reject);\n        if (idx >= 0) {\n          this.pendingRejects.splice(idx, 1);\n        }\n      };\n\n      const timer = setTimeout(() => {\n        cleanup();\n        reject(new Error(`Timeout waiting for ${type} message`));\n      }, timeout);\n\n      this.messageHandlers.set(type, (data) => {\n        cleanup();\n        resolve(data);\n      });\n    });\n  }\n\n  /**\n   * Check if Chrome backend is still alive\n   */\n  isAlive(): boolean {\n    return this.isReady && this.browser !== null && this.page !== null;\n  }\n\n  /**\n   * Get Chrome backend status information\n   */\n  getStatus(): {\n    pid: number | null;\n    port: number;\n    modelId: string;\n    startedAt: Date | null;\n  } {\n    // Try instance browser first, then global browser\n    let pid: number | null = null;\n    const browserProcess = this.browser?.process?.() || globalBrowser?.process?.();\n    if (browserProcess?.pid) {\n      pid = browserProcess.pid;\n    }\n    return {\n      pid,\n      port: this.serverPort || globalServerPort,\n      modelId: this.modelId,\n      startedAt: this.isReady ? new Date() : null,\n    };\n  }\n\n  /**\n   * Get Chrome memory usage via CDP Performance metrics\n   * Returns memory in bytes or null if unavailable\n   */\n  async getMemoryUsage(): Promise<{\n    jsHeapUsed: number;\n    jsHeapTotal: number;\n  } | null> {\n    if (!(this.cdp && this.isReady)) {\n      return null;\n    }\n\n    try {\n      // Enable Performance domain if needed\n      await this.cdp.send(\"Performance.enable\");\n\n      const { metrics } = (await this.cdp.send(\"Performance.getMetrics\")) as {\n        metrics: Array<{ name: string; value: number }>;\n      };\n\n      const jsHeapUsed = metrics.find((m) => m.name === \"JSHeapUsedSize\")?.value ?? 0;\n      const jsHeapTotal = metrics.find((m) => m.name === \"JSHeapTotalSize\")?.value ?? 0;\n\n      return { jsHeapUsed, jsHeapTotal };\n    } catch {\n      return null;\n    }\n  }\n\n  /**\n   * Check memory usage and auto-cleanup if threshold exceeded\n   * @param thresholdGB Memory threshold in GB (default: 8)\n   * @returns true if cleanup was performed\n   */\n  async checkMemoryAndCleanup(thresholdGB = 8): Promise<boolean> {\n    const mem = await this.getMemoryUsage();\n    if (!mem) {\n      return false;\n    }\n\n    const usedGB = mem.jsHeapUsed / 1024 ** 3;\n\n    if (usedGB > thresholdGB) {\n      console.warn(\n        `[Gerbil] Memory usage high (${usedGB.toFixed(\n          1,\n        )}GB > ${thresholdGB}GB), clearing KV cache...`,\n      );\n      await this.reset();\n      return true;\n    }\n\n    return false;\n  }\n\n  /**\n   * Get memory usage in a human-readable format\n   */\n  async getMemoryStats(): Promise<{\n    usedGB: number;\n    totalGB: number;\n    usedPercent: number;\n  } | null> {\n    const mem = await this.getMemoryUsage();\n    if (!mem) {\n      return null;\n    }\n\n    const usedGB = mem.jsHeapUsed / 1024 ** 3;\n    const totalGB = mem.jsHeapTotal / 1024 ** 3;\n    const usedPercent = (mem.jsHeapUsed / mem.jsHeapTotal) * 100;\n\n    return { usedGB, totalGB, usedPercent };\n  }\n\n  /**\n   * Generate text with streaming\n   */\n  async generate(prompt: string, options: GenerateOptions = {}): Promise<string> {\n    if (!this.isAlive()) {\n      throw new Error(\"CHROME_BACKEND_DEAD\");\n    }\n\n    const system = options.system || \"You are a helpful assistant.\";\n    const messages = [\n      { role: \"system\", content: system },\n      { role: \"user\", content: prompt },\n    ];\n\n    const genOptions = {\n      maxTokens: options.maxTokens ?? 256,\n      temperature: options.temperature ?? 0.7,\n      topP: options.topP ?? 0.9,\n      topK: options.topK ?? 20,\n      thinking: options.thinking ?? false,\n    };\n\n    // Set up token handler if callback provided\n    if (options.onToken) {\n      this.messageHandlers.set(\"token\", options.onToken);\n    }\n\n    try {\n      // Start generation\n      const resultPromise = this.page!.evaluate(\n        (msgs, opts) => (window as any).gerbilGenerate(msgs, opts),\n        messages,\n        genOptions,\n      );\n\n      // Wait for completion\n      const completeData = await this.waitForMessage(\"complete\", 600_000); // 10 min timeout\n\n      // Clean up token handler\n      this.messageHandlers.delete(\"token\");\n\n      await resultPromise; // Ensure evaluate completes\n\n      return completeData.text || \"\";\n    } catch (err: any) {\n      // Check if Chrome died during generation\n      if (!this.isAlive()) {\n        throw new Error(\"CHROME_BACKEND_DEAD\");\n      }\n      throw err;\n    }\n  }\n\n  /**\n   * Interrupt current generation\n   */\n  async interrupt(): Promise<void> {\n    if (this.page) {\n      await this.page.evaluate(\"window.gerbilInterrupt()\");\n    }\n  }\n\n  /**\n   * Reset conversation cache\n   */\n  async reset(): Promise<void> {\n    if (this.page) {\n      await this.page.evaluate(\"window.gerbilReset()\");\n    }\n  }\n\n  /**\n   * Check if backend is ready\n   */\n  ready(): boolean {\n    return this.isReady;\n  }\n\n  /**\n   * Start or reuse the global HTTP server\n   * Uses singleton pattern to prevent killing our own server\n   */\n  private async startServer(html: string): Promise<void> {\n    // If global server is already running, reuse it\n    if (globalServer && globalServerPort) {\n      this.server = globalServer;\n      this.serverPort = globalServerPort;\n      return;\n    }\n\n    return new Promise((resolve, reject) => {\n      const server = createServer((_req, res) => {\n        res.writeHead(200, { \"Content-Type\": \"text/html\" });\n        res.end(html);\n      });\n\n      server.on(\"error\", (err: NodeJS.ErrnoException) => {\n        if (err.code === \"EADDRINUSE\") {\n          // Port in use - assume it's our server from a previous run\n          // Just use that port (Chrome will connect to existing server)\n          this.serverPort = GERBIL_LOCAL_PORT;\n          globalServerPort = GERBIL_LOCAL_PORT;\n          resolve();\n        } else {\n          reject(err);\n        }\n      });\n\n      // Listen on fixed port for consistent IndexedDB origin (cache persistence)\n      server.listen(GERBIL_LOCAL_PORT, \"127.0.0.1\", () => {\n        this.server = server;\n        this.serverPort = GERBIL_LOCAL_PORT;\n        globalServer = server;\n        globalServerPort = GERBIL_LOCAL_PORT;\n        resolve();\n      });\n    });\n  }\n\n  /**\n   * Dispose of the backend and clean up\n   * Note: We keep the shared browser running for other backends\n   */\n  async dispose(): Promise<void> {\n    // Mark as not ready first to prevent new operations\n    this.isReady = false;\n\n    // Clear pending waits silently (don't reject - just clear them)\n    this.pendingRejects = [];\n    this.messageHandlers.clear();\n\n    // Close our page (but NOT the shared browser - other backends may use it)\n    if (this.page) {\n      try {\n        await this.page.close();\n        // Decrement active page counter\n        activePagesCount = Math.max(0, activePagesCount - 1);\n      } catch {\n        // Page may already be closed\n      }\n      this.page = null;\n    }\n\n    // Clear CDP session reference (closed with page)\n    this.cdp = null;\n\n    // We don't close the shared browser or server - they persist for reuse\n    // Just clear our references\n    this.browser = null;\n    this.server = null;\n  }\n\n  /**\n   * Reject all pending waits (called on browser disconnect or dispose)\n   */\n  private rejectPendingWaits(error: Error): void {\n    for (const reject of this.pendingRejects) {\n      reject(error);\n    }\n    this.pendingRejects = [];\n    this.messageHandlers.clear();\n  }\n\n  /**\n   * Clear the model cache (forces re-download on next start)\n   */\n  static clearCache(): void {\n    if (existsSync(GERBIL_CACHE_DIR)) {\n      rmSync(GERBIL_CACHE_DIR, { recursive: true, force: true });\n    }\n  }\n\n  /**\n   * Get the number of active Chrome pages\n   */\n  static getActivePageCount(): number {\n    return activePagesCount;\n  }\n\n  /**\n   * Get memory usage info for all active pages\n   */\n  static getMemoryInfo(): { activePagesCount: number; maxPages: number } {\n    return {\n      activePagesCount,\n      maxPages: MAX_CONCURRENT_PAGES,\n    };\n  }\n\n  /**\n   * Gracefully close the shared browser (call on process exit)\n   */\n  static async closeSharedBrowser(): Promise<void> {\n    if (globalBrowser) {\n      try {\n        await globalBrowser.close();\n      } catch {\n        // Browser may already be closed\n      }\n      globalBrowser = null;\n      globalBrowserPromise = null;\n    }\n\n    if (globalServer) {\n      globalServer.close();\n      globalServer = null;\n      globalServerPort = 0;\n    }\n\n    // Reset page counter\n    activePagesCount = 0;\n\n    // Clean up endpoint file\n    try {\n      unlinkSync(WS_ENDPOINT_FILE);\n    } catch {}\n  }\n}\n\n// Register cleanup on process exit to prevent mutex errors\nlet cleanupRegistered = false;\nfunction registerCleanup() {\n  if (cleanupRegistered) {\n    return;\n  }\n  cleanupRegistered = true;\n\n  const cleanup = () => {\n    // Synchronous close - can't await in exit handlers\n    if (globalBrowser) {\n      try {\n        // Force kill the browser process\n        const browserProcess = globalBrowser.process();\n        if (browserProcess) {\n          browserProcess.kill(\"SIGTERM\");\n        }\n      } catch {}\n      globalBrowser = null;\n    }\n    if (globalServer) {\n      globalServer.close();\n      globalServer = null;\n    }\n  };\n\n  process.on(\"exit\", cleanup);\n  process.on(\"SIGINT\", () => {\n    cleanup();\n    process.exit(0);\n  });\n  process.on(\"SIGTERM\", () => {\n    cleanup();\n    process.exit(0);\n  });\n}\n\n// Auto-register when module loads\nregisterCleanup();\n\nexport default ChromeGPUBackend;\n"],"mappings":";;;;;;;;;;;;;;AAeA,MAAM,mBAAmB,KAAK,SAAS,EAAE,WAAW,eAAe;AACnE,MAAM,mBAAmB,KAAK,kBAAkB,kBAAkB;AAClE,MAAM,qBAAqB,KAAK,SAAS,EAAE,WAAW,qBAAqB;;AAc3E,SAAgB,wBAA4C;AAC1D,KAAI;AACF,MAAI,CAAC,WAAW,mBAAmB,CACjC,QAAO,EAAE;AAGX,SADa,KAAK,MAAM,aAAa,oBAAoB,QAAQ,CAAC,CACtD,UAAU,EAAE;SAClB;AACN,SAAO,EAAE;;;;AAKb,eAAe,eAAe,SAA8C;AAC1E,KAAI;EAEF,MAAM,UAAU,MAAM,MAAM,qCAAqC,QAAQ,iBAAiB;AAC1F,MAAI,QAAQ,IAAI;GACd,MAAMA,QAA2C,MAAM,QAAQ,MAAM;GAErE,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,KAAK,SAAS,QAAQ,IAAI,EAAE,KAAK,SAAS,QAAQ,CAAC;GACrF,MAAM,KAAK,MAAM,MACd,MAAM,EAAE,KAAK,SAAS,KAAK,IAAI,CAAC,EAAE,KAAK,SAAS,MAAM,IAAI,EAAE,KAAK,SAAS,QAAQ,CACpF;GACD,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,KAAK,SAAS,OAAO,IAAI,EAAE,KAAK,SAAS,QAAQ,CAAC;GACnF,MAAM,UAAU,MAAM,MAAM,MAAM,EAAE,KAAK,SAAS,QAAQ,CAAC;GAC3D,MAAM,WAAW,SAAS,MAAM,QAAQ;AACxC,OAAI,UAAU,KACZ,QAAO,SAAS;;EAKpB,MAAM,MAAM,MAAM,MAAM,qCAAqC,UAAU;AACvE,MAAI,IAAI,GAEN,SADa,MAAM,IAAI,MAAM,EACjB;SAER;;;AAOV,SAAgB,iBAAiB,SAAiB,WAA0B;AAC1E,KAAI;EACF,MAAM,MAAM,KAAK,SAAS,EAAE,UAAU;AACtC,MAAI,CAAC,WAAW,IAAI,CAClB,WAAU,KAAK,EAAE,WAAW,MAAM,CAAC;EAGrC,MAAM,SAAS,uBAAuB;EACtC,MAAM,WAAW,OAAO,MAAM,MAAM,EAAE,YAAY,QAAQ;EAC1D,MAAM,uBAAM,IAAI,MAAM,EAAC,aAAa;AAEpC,MAAI,UAAU;AACZ,YAAS,WAAW;AACpB,OAAI,UACF,UAAS,YAAY;QAGvB,QAAO,KAAK;GACV;GACA,cAAc;GACd,UAAU;GACV;GACD,CAAC;AAGJ,gBAAc,oBAAoB,KAAK,UAAU,EAAE,QAAQ,EAAE,MAAM,EAAE,CAAC;AAGtE,MAAI,CAAC,UACH,gBAAe,QAAQ,CACpB,MAAM,SAAS;AACd,OAAI,MAAM;IACR,MAAM,gBAAgB,uBAAuB;IAC7C,MAAM,QAAQ,cAAc,MAAM,MAAM,EAAE,YAAY,QAAQ;AAC9D,QAAI,OAAO;AACT,WAAM,YAAY;AAClB,mBAAc,oBAAoB,KAAK,UAAU,EAAE,QAAQ,eAAe,EAAE,MAAM,EAAE,CAAC;;;IAGzF,CACD,YAAY,GAAG;SAEd;;;AAgBV,eAAsB,0BAAyC;AAC7D,KAAI;EACF,MAAM,SAAS,uBAAuB;EACtC,MAAM,YAAY,OAAO,QAAQ,MAAM,CAAC,EAAE,UAAU;AACpD,MAAI,UAAU,WAAW,EACvB;EAIF,MAAM,YAAY;AAClB,OAAK,IAAI,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK,WAAW;GACpD,MAAM,QAAQ,UAAU,MAAM,GAAG,IAAI,UAAU;AAC/C,SAAM,QAAQ,IACZ,MAAM,IAAI,OAAO,UAAU;IACzB,MAAM,OAAO,MAAM,eAAe,MAAM,QAAQ;AAChD,QAAI,KACF,OAAM,YAAY;KAEpB,CACH;;AAIH,gBAAc,oBAAoB,KAAK,UAAU,EAAE,QAAQ,EAAE,MAAM,EAAE,CAAC;SAChE;;AAQV,MAAM,oBAAoB;AAG1B,IAAIC,gBAAgC;AACpC,IAAIC,uBAAgD;AACpD,IAAIC,eAA8B;AAClC,IAAI,mBAAmB;AAGvB,IAAI,mBAAmB;AACvB,MAAM,uBAAuB;AA+B7B,MAAM,eAAe;CACnB,QAAQ;EACN;EACA;EACA;EACA;EACA;EACD;CACD,OAAO;EACL;EACA;EACA;EACA;EACA;EACA;EACD;CACD,OAAO;EACL;EACA;EACA,GAAG,QAAQ,IAAI,aAAa;EAC5B;EACA;EACD;CACF;AAED,SAAS,aAAqB;AAE5B,KAAI,QAAQ,IAAI,YACd,QAAO,QAAQ,IAAI;CAGrB,MAAM,WAAW,QAAQ;CACzB,MAAM,QAAQ,aAAa,aAAa,EAAE;AAE1C,MAAK,MAAM,KAAK,MACd,KAAI;AACF,MAAI,aAAa,SAAS;AAExB,YAAS,SAAS,KAAK,EAAE,OAAO,UAAU,CAAC;AAC3C,UAAO;;AAGT,MAAI,WAAW,EAAE,CACf,QAAO;SAEH;AAGV,OAAM,IAAI,MAAM,4EAA4E;;AAO9F,SAAS,eAAe,aAAqB,gBAAkC;CAC7E,MAAM,QAAQ,CAAC,gBAAgB,mBAAmB,cAAc;AAGhE,KAAI,QAAQ,aAAa,QAEvB,OAAM,KACJ,0BACA,4BACA,sBACA,2BACD;UACQ,QAAQ,aAAa,UAAU,OAMxC,OAAM,KAAK,yBAAyB;AAGtC,QAAO;;AAOT,SAAS,kBAAkB,WAA2B;AACpD,QAAO;;;;;;;;;;;;;;;;0BAgBiB,UAAU;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqMpC,IAAa,mBAAb,MAAa,iBAAiB;CAC5B,AAAQ,UAA0B;CAClC,AAAQ,OAAoB;CAC5B,AAAQ,MAAyB;CACjC,AAAQ,aAAa;CACrB,AAAQ,cAAsB;CAC9B,AAAiB;CACjB,AAAQ,UAAU;CAClB,AAAiB,kCAAoD,IAAI,KAAK;CAC9E,AAAQ,iBAA8C,EAAE;CAExD,AAAO,SAAwB;CAE/B,AAAQ,YAAY,SAAiB;AACnC,OAAK,UAAU;;;;;CAMjB,aAAa,OAAO,UAAgC,EAAE,EAA6B;EAEjF,MAAM,UAAU,IAAI,iBADJ,QAAQ,WAAW,iCACU;AAC7C,QAAM,QAAQ,OAAO,QAAQ;AAC7B,SAAO;;;;;;CAOT,MAAc,mBACZ,YACA,SACkB;AAElB,MAAI,eAAe,WAAW;AAC5B,WAAQ,aAAa,EAAE,QAAQ,8BAA8B,CAAC;AAC9D,UAAO;;AAIT,MAAI,sBAAsB;AACxB,WAAQ,aAAa,EAAE,QAAQ,iCAAiC,CAAC;AACjE,UAAO;;AAIT,MAAI,WAAW,iBAAiB,CAC9B,KAAI;GACF,MAAM,aAAa,aAAa,kBAAkB,QAAQ,CAAC,MAAM;AACjE,WAAQ,aAAa,EAAE,QAAQ,oCAAoC,CAAC;AACpE,mBAAgB,MAAM,UAAU,QAAQ,EACtC,mBAAmB,YACpB,CAAC;AACF,UAAO;UACD;AAEN,OAAI;AACF,eAAW,iBAAiB;WACtB;;AAKZ,yBAAuB,KAAK,cAAc,YAAY,QAAQ;AAC9D,MAAI;AACF,mBAAgB,MAAM;AACtB,UAAO;YACC;AACR,0BAAuB;;;;;;CAO3B,MAAc,cACZ,YACA,UACkB;EAClB,MAAM,gBAAgB,OAAO,KAAK,MAAM,KAAK,QAAQ,GAAG,IAAK;EAG7D,MAAM,WAAW,KAAK,KAAK,aAAa,gBAAgB;AACxD,MAAI,WAAW,SAAS,CACtB,KAAI;AACF,cAAW,SAAS;AACpB,SAAM,IAAI,SAAS,MAAM,WAAW,GAAG,IAAI,CAAC;UACtC;EAKV,MAAM,UAAU,MAAM,UAAU,OAAO;GACrC,gBAAgB;GAChB,UAAU;GACV,MAAM;IACJ,GAAG,eAAe,KAAK,aAAa,cAAc;IAClD;IACA;IACA;IACA;IACA;IACA;IACD;GACD,cAAc;GACd,eAAe;GACf,cAAc;GACf,CAAC;AAGF,gBAAc,kBAAkB,QAAQ,YAAY,CAAC;AAGrD,UAAQ,GAAG,sBAAsB;AAC/B,mBAAgB;AAChB,OAAI;AACF,eAAW,iBAAiB;WACtB;IACR;AAEF,SAAO;;;;;CAMT,MAAc,OAAO,SAA8C;AAEjE,MAAI,oBAAoB,qBACtB,OAAM,IAAI,MACR,6BAA6B,qBAAqB,yFAE3B,mBACxB;EAGH,MAAM,aAAa,QAAQ,cAAc,YAAY;AAGrD,OAAK,cAAc;AACnB,MAAI,CAAC,WAAW,KAAK,YAAY,CAC/B,WAAU,KAAK,aAAa,EAAE,WAAW,MAAM,CAAC;EAKlD,MAAM,OAAO,kBAAkB,KAAK,QAAQ;AAC5C,QAAM,KAAK,YAAY,KAAK;AAE5B,UAAQ,aAAa,EAAE,QAAQ,sBAAsB,CAAC;AAGtD,OAAK,UAAU,MAAM,KAAK,mBAAmB,YAAY,QAAQ;AAGjE,OAAK,OAAO,MAAM,KAAK,QAAQ,SAAS;AACxC,OAAK,MAAM,MAAM,KAAK,KAAK,kBAAkB;AAG7C;AACA,UAAQ,aAAa,EACnB,QAAQ,iBAAiB,iBAAiB,GAAG,wBAC9C,CAAC;AAGF,OAAK,QAAQ,GAAG,sBAAsB;AACpC,WAAQ,MAAM,6CAA6C;AAC3D,QAAK,UAAU;AACf,QAAK,UAAU;AACf,QAAK,OAAO;AACZ,QAAK,MAAM;AAEX,QAAK,mCAAmB,IAAI,MAAM,sBAAsB,CAAC;IACzD;AAGF,QAAM,KAAK,IAAI,KAAK,iBAAiB;AACrC,QAAM,KAAK,IAAI,KAAK,kCAAkC,EAAE,UAAU,IAAI,CAAC;AAGvE,OAAK,IAAI,GAAG,6BAA6B,UAAU;GACjD,MAAM,OAAO,MAAM,KAAK,KAAK,MAAW,EAAE,SAAS,EAAE,eAAe,GAAG,CAAC,KAAK,IAAI;AAEjF,OAAI,MAAM,SAAS,SAAS,MAAM,KAAK,IAAI,MACzC,KAAI;IACF,MAAM,OAAO,KAAK,MAAM,MAAM,KAAK,GAAG,MAAM;AAC5C,SAAK,cAAc,MAAM,QAAQ;WAC3B;AAEN,QACE,KAAK,SAAS,OACd,CAAC,KAAK,SAAS,eAAe,IAC9B,CAAC,KAAK,SAAS,kBAAkB,EACjC;;YAIK,MAAM,SAAS,WAAW,MAAM,SAAS,WAElD;QACE,EACE,KAAK,SAAS,cAAc,IAC5B,KAAK,SAAS,iBAAiB,IAC/B,KAAK,SAAS,eAAe,IAC7B,KAAK,SAAS,kBAAkB,KAElC,KAAK,SAAS,IAEd,SAAQ,MAAM,WAAW,MAAM,KAAK,IAAI,KAAK;;IAGjD;AAGF,OAAK,IAAI,GAAG,4BAA4B,UAAU;GAChD,MAAM,UACJ,MAAM,kBAAkB,QAAQ,MAAM,kBAAkB,WAAW,eAAe;AAEpF,OACE,QAAQ,SAAS,eAAe,IAChC,QAAQ,SAAS,kBAAkB,IACnC,QAAQ,SAAS,IAEjB;AAEF,WAAQ,MAAM,sBAAsB,QAAQ;IAC5C;AAGF,QAAM,KAAK,KAAK,KAAK,oBAAoB,KAAK,WAAW,IAAI;GAC3D,WAAW;GACX,SAAS;GACV,CAAC;AAGF,QAAM,KAAK,eAAe,SAAS,IAAQ;AAE3C,OAAK,UAAU;AACf,UAAQ,aAAa,EAAE,QAAQ,mBAAmB,CAAC;AAGnD,mBAAiB,KAAK,QAAQ;;;;;CAMhC,AAAQ,cAAc,MAAW,SAAqC;EACpE,MAAM,EAAE,MAAM,GAAG,SAAS;EAG1B,MAAM,UAAU,KAAK,gBAAgB,IAAI,KAAK;AAC9C,MAAI,QACF,SAAQ,KAAK;AAIf,MAAI,SAAS,WACX,SAAQ,aAAa,KAAK;WACjB,SAAS,QAClB,SAAQ,UAAU,KAAK;;;;;CAO3B,AAAQ,eAAe,MAAc,UAAU,KAAsB;AACnE,SAAO,IAAI,SAAS,SAAS,WAAW;AAEtC,QAAK,eAAe,KAAK,OAAO;GAEhC,MAAM,gBAAgB;AACpB,iBAAa,MAAM;AACnB,SAAK,gBAAgB,OAAO,KAAK;IACjC,MAAM,MAAM,KAAK,eAAe,QAAQ,OAAO;AAC/C,QAAI,OAAO,EACT,MAAK,eAAe,OAAO,KAAK,EAAE;;GAItC,MAAM,QAAQ,iBAAiB;AAC7B,aAAS;AACT,2BAAO,IAAI,MAAM,uBAAuB,KAAK,UAAU,CAAC;MACvD,QAAQ;AAEX,QAAK,gBAAgB,IAAI,OAAO,SAAS;AACvC,aAAS;AACT,YAAQ,KAAK;KACb;IACF;;;;;CAMJ,UAAmB;AACjB,SAAO,KAAK,WAAW,KAAK,YAAY,QAAQ,KAAK,SAAS;;;;;CAMhE,YAKE;EAEA,IAAIC,MAAqB;EACzB,MAAM,iBAAiB,KAAK,SAAS,WAAW,IAAI,eAAe,WAAW;AAC9E,MAAI,gBAAgB,IAClB,OAAM,eAAe;AAEvB,SAAO;GACL;GACA,MAAM,KAAK,cAAc;GACzB,SAAS,KAAK;GACd,WAAW,KAAK,0BAAU,IAAI,MAAM,GAAG;GACxC;;;;;;CAOH,MAAM,iBAGI;AACR,MAAI,EAAE,KAAK,OAAO,KAAK,SACrB,QAAO;AAGT,MAAI;AAEF,SAAM,KAAK,IAAI,KAAK,qBAAqB;GAEzC,MAAM,EAAE,YAAa,MAAM,KAAK,IAAI,KAAK,yBAAyB;AAOlE,UAAO;IAAE,YAHU,QAAQ,MAAM,MAAM,EAAE,SAAS,iBAAiB,EAAE,SAAS;IAGzD,aAFD,QAAQ,MAAM,MAAM,EAAE,SAAS,kBAAkB,EAAE,SAAS;IAE9C;UAC5B;AACN,UAAO;;;;;;;;CASX,MAAM,sBAAsB,cAAc,GAAqB;EAC7D,MAAM,MAAM,MAAM,KAAK,gBAAgB;AACvC,MAAI,CAAC,IACH,QAAO;EAGT,MAAM,SAAS,IAAI,aAAa,QAAQ;AAExC,MAAI,SAAS,aAAa;AACxB,WAAQ,KACN,+BAA+B,OAAO,QACpC,EACD,CAAC,OAAO,YAAY,2BACtB;AACD,SAAM,KAAK,OAAO;AAClB,UAAO;;AAGT,SAAO;;;;;CAMT,MAAM,iBAII;EACR,MAAM,MAAM,MAAM,KAAK,gBAAgB;AACvC,MAAI,CAAC,IACH,QAAO;AAOT,SAAO;GAAE,QAJM,IAAI,aAAa,QAAQ;GAIvB,SAHD,IAAI,cAAc,QAAQ;GAGhB,aAFL,IAAI,aAAa,IAAI,cAAe;GAElB;;;;;CAMzC,MAAM,SAAS,QAAgB,UAA2B,EAAE,EAAmB;AAC7E,MAAI,CAAC,KAAK,SAAS,CACjB,OAAM,IAAI,MAAM,sBAAsB;EAIxC,MAAM,WAAW,CACf;GAAE,MAAM;GAAU,SAFL,QAAQ,UAAU;GAEI,EACnC;GAAE,MAAM;GAAQ,SAAS;GAAQ,CAClC;EAED,MAAM,aAAa;GACjB,WAAW,QAAQ,aAAa;GAChC,aAAa,QAAQ,eAAe;GACpC,MAAM,QAAQ,QAAQ;GACtB,MAAM,QAAQ,QAAQ;GACtB,UAAU,QAAQ,YAAY;GAC/B;AAGD,MAAI,QAAQ,QACV,MAAK,gBAAgB,IAAI,SAAS,QAAQ,QAAQ;AAGpD,MAAI;GAEF,MAAM,gBAAgB,KAAK,KAAM,UAC9B,MAAM,SAAU,OAAe,eAAe,MAAM,KAAK,EAC1D,UACA,WACD;GAGD,MAAM,eAAe,MAAM,KAAK,eAAe,YAAY,IAAQ;AAGnE,QAAK,gBAAgB,OAAO,QAAQ;AAEpC,SAAM;AAEN,UAAO,aAAa,QAAQ;WACrBC,KAAU;AAEjB,OAAI,CAAC,KAAK,SAAS,CACjB,OAAM,IAAI,MAAM,sBAAsB;AAExC,SAAM;;;;;;CAOV,MAAM,YAA2B;AAC/B,MAAI,KAAK,KACP,OAAM,KAAK,KAAK,SAAS,2BAA2B;;;;;CAOxD,MAAM,QAAuB;AAC3B,MAAI,KAAK,KACP,OAAM,KAAK,KAAK,SAAS,uBAAuB;;;;;CAOpD,QAAiB;AACf,SAAO,KAAK;;;;;;CAOd,MAAc,YAAY,MAA6B;AAErD,MAAI,gBAAgB,kBAAkB;AACpC,QAAK,SAAS;AACd,QAAK,aAAa;AAClB;;AAGF,SAAO,IAAI,SAAS,SAAS,WAAW;GACtC,MAAM,SAAS,cAAc,MAAM,QAAQ;AACzC,QAAI,UAAU,KAAK,EAAE,gBAAgB,aAAa,CAAC;AACnD,QAAI,IAAI,KAAK;KACb;AAEF,UAAO,GAAG,UAAU,QAA+B;AACjD,QAAI,IAAI,SAAS,cAAc;AAG7B,UAAK,aAAa;AAClB,wBAAmB;AACnB,cAAS;UAET,QAAO,IAAI;KAEb;AAGF,UAAO,OAAO,mBAAmB,mBAAmB;AAClD,SAAK,SAAS;AACd,SAAK,aAAa;AAClB,mBAAe;AACf,uBAAmB;AACnB,aAAS;KACT;IACF;;;;;;CAOJ,MAAM,UAAyB;AAE7B,OAAK,UAAU;AAGf,OAAK,iBAAiB,EAAE;AACxB,OAAK,gBAAgB,OAAO;AAG5B,MAAI,KAAK,MAAM;AACb,OAAI;AACF,UAAM,KAAK,KAAK,OAAO;AAEvB,uBAAmB,KAAK,IAAI,GAAG,mBAAmB,EAAE;WAC9C;AAGR,QAAK,OAAO;;AAId,OAAK,MAAM;AAIX,OAAK,UAAU;AACf,OAAK,SAAS;;;;;CAMhB,AAAQ,mBAAmB,OAAoB;AAC7C,OAAK,MAAM,UAAU,KAAK,eACxB,QAAO,MAAM;AAEf,OAAK,iBAAiB,EAAE;AACxB,OAAK,gBAAgB,OAAO;;;;;CAM9B,OAAO,aAAmB;AACxB,MAAI,WAAW,iBAAiB,CAC9B,QAAO,kBAAkB;GAAE,WAAW;GAAM,OAAO;GAAM,CAAC;;;;;CAO9D,OAAO,qBAA6B;AAClC,SAAO;;;;;CAMT,OAAO,gBAAgE;AACrE,SAAO;GACL;GACA,UAAU;GACX;;;;;CAMH,aAAa,qBAAoC;AAC/C,MAAI,eAAe;AACjB,OAAI;AACF,UAAM,cAAc,OAAO;WACrB;AAGR,mBAAgB;AAChB,0BAAuB;;AAGzB,MAAI,cAAc;AAChB,gBAAa,OAAO;AACpB,kBAAe;AACf,sBAAmB;;AAIrB,qBAAmB;AAGnB,MAAI;AACF,cAAW,iBAAiB;UACtB;;;AAKZ,IAAI,oBAAoB;AACxB,SAAS,kBAAkB;AACzB,KAAI,kBACF;AAEF,qBAAoB;CAEpB,MAAM,gBAAgB;AAEpB,MAAI,eAAe;AACjB,OAAI;IAEF,MAAM,iBAAiB,cAAc,SAAS;AAC9C,QAAI,eACF,gBAAe,KAAK,UAAU;WAE1B;AACR,mBAAgB;;AAElB,MAAI,cAAc;AAChB,gBAAa,OAAO;AACpB,kBAAe;;;AAInB,SAAQ,GAAG,QAAQ,QAAQ;AAC3B,SAAQ,GAAG,gBAAgB;AACzB,WAAS;AACT,UAAQ,KAAK,EAAE;GACf;AACF,SAAQ,GAAG,iBAAiB;AAC1B,WAAS;AACT,UAAQ,KAAK,EAAE;GACf;;AAIJ,iBAAiB"}