@tryhamster/gerbil 1.0.0-rc.1 → 1.0.0-rc.3

package/dist/{chrome-backend-Y9F7W5VQ.mjs.map → chrome-backend-CORwaIyC.mjs.map}

@@ -1 +1 @@
-{"version":3,"file":"chrome-backend-Y9F7W5VQ.mjs","names":["files: { path: string; size?: number; lfs?: { size?: number } }[]","globalBrowser: Browser | null","globalBrowserPromise: Promise<Browser> | null","globalServer: Server | null","activeBackends: Set<ChromeGPUBackend>","pid: number | null","err: any","wsEndpoint: string | null","modelId: string | null","memory: { usedGB: number; totalGB: number } | null"],"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 \"node:child_process\";\nimport { existsSync, mkdirSync, readFileSync, rmSync, unlinkSync, writeFileSync } from \"node:fs\";\nimport { createServer, type Server } from \"node:http\";\nimport { homedir } from \"node:os\";\nimport { join } from \"node: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  contextLength?: 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 context length from HuggingFace (config.json preferred for actual limit) */\nasync function fetchContextLength(modelId: string): Promise<number | undefined> {\n  // Priority 1: config.json → max_position_embeddings (actual architectural limit)\n  try {\n    const res = await fetch(`https://huggingface.co/${modelId}/raw/main/config.json`);\n    if (res.ok) {\n      const config = await res.json();\n      // Check root level first, then nested text_config (for multimodal models like Ministral)\n      const textConfig = config.text_config || {};\n      const ctxLen =\n        config.max_position_embeddings ||\n        textConfig.max_position_embeddings ||\n        config.sliding_window ||\n        textConfig.sliding_window ||\n        config.max_seq_len ||\n        config.max_sequence_length ||\n        config.n_ctx ||\n        config.n_positions;\n      if (ctxLen) {\n        return ctxLen;\n      }\n    }\n  } catch {\n    // Continue to fallback\n  }\n\n  // Priority 2: tokenizer_config.json → model_max_length (fallback only)\n  try {\n    const tokRes = await fetch(`https://huggingface.co/${modelId}/raw/main/tokenizer_config.json`);\n    if (tokRes.ok) {\n      const tokConfig = await tokRes.json();\n      // Only use if reasonable (< 1M tokens - anything larger is likely placeholder)\n      if (tokConfig.model_max_length && tokConfig.model_max_length < 1e6) {\n        return tokConfig.model_max_length;\n      }\n    }\n  } catch {\n    // Ignore\n  }\n\n  return;\n}\n\n/** Get file size from HuggingFace tree entry (handles both regular and LFS files) */\nfunction getFileSize(file: { size?: number; lfs?: { size?: number } }): number {\n  // LFS files have size in lfs.size, regular files in size\n  return file.lfs?.size || file.size || 0;\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/${modelId}/tree/main/onnx`);\n    if (treeRes.ok) {\n      const files: { path: string; size?: number; lfs?: { size?: number } }[] =\n        await treeRes.json();\n\n      // Prefer q4f16 > q4 > fp16 > any .onnx file\n      const q4f16 = files.find((f) => f.path.includes(\"q4f16\") && f.path.endsWith(\".onnx\"));\n      const q4 = files.find(\n        (f) => f.path.includes(\"q4\") && !f.path.includes(\"f16\") && f.path.endsWith(\".onnx\"),\n      );\n      const fp16 = files.find((f) => f.path.includes(\"fp16\") && f.path.endsWith(\".onnx\"));\n      const anyOnnx = files.find((f) => f.path.endsWith(\".onnx\"));\n      const bestFile = q4f16 || q4 || fp16 || anyOnnx;\n\n      if (bestFile) {\n        // Sum up the .onnx file AND all associated .onnx_data* files\n        const baseName = bestFile.path.replace(\".onnx\", \"\");\n        const relatedFiles = files.filter(\n          (f) => f.path === bestFile.path || f.path.startsWith(`${baseName}.onnx_data`),\n        );\n        const totalSize = relatedFiles.reduce((sum, f) => sum + getFileSize(f), 0);\n        if (totalSize > 0) {\n          return totalSize;\n        }\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(\n  modelId: string,\n  sizeBytes?: number,\n  contextLength?: number,\n): 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      if (contextLength) {\n        existing.contextLength = contextLength;\n      }\n    } else {\n      models.push({\n        modelId,\n        downloadedAt: now,\n        lastUsed: now,\n        sizeBytes,\n        contextLength,\n      });\n    }\n\n    writeFileSync(CACHED_MODELS_FILE, JSON.stringify({ models }, null, 2));\n\n    // Fetch missing metadata in background\n    const needsSize = !(sizeBytes || existing?.sizeBytes);\n    const needsContext = !(contextLength || existing?.contextLength);\n\n    if (needsSize || needsContext) {\n      Promise.all([\n        needsSize ? fetchModelSize(modelId) : Promise.resolve(undefined),\n        needsContext ? fetchContextLength(modelId) : Promise.resolve(undefined),\n      ])\n        .then(([size, context]) => {\n          const updatedModels = getChromeCachedModels();\n          const model = updatedModels.find((m) => m.modelId === modelId);\n          if (model) {\n            if (size) {\n              model.sizeBytes = size;\n            }\n            if (context) {\n              model.contextLength = context;\n            }\n            writeFileSync(CACHED_MODELS_FILE, JSON.stringify({ models: updatedModels }, null, 2));\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 metadata (size, context length) for cached models that need it */\nexport async function refreshCachedModelSizes(): Promise<void> {\n  try {\n    const models = getChromeCachedModels();\n    // Refresh if: no size, size < 1MB (wrong), or missing context length\n    const MIN_EXPECTED_SIZE = 1_000_000; // 1MB - real models are always bigger\n    const needsRefresh = models.filter(\n      (m) => !m.sizeBytes || m.sizeBytes < MIN_EXPECTED_SIZE || !m.contextLength,\n    );\n    if (needsRefresh.length === 0) {\n      return;\n    }\n\n    // Fetch metadata in parallel (max 3 at a time)\n    const batchSize = 3;\n    for (let i = 0; i < needsRefresh.length; i += batchSize) {\n      const batch = needsRefresh.slice(i, i + batchSize);\n      await Promise.all(\n        batch.map(async (model) => {\n          const [size, context] = await Promise.all([\n            !model.sizeBytes || model.sizeBytes < MIN_EXPECTED_SIZE\n              ? fetchModelSize(model.modelId)\n              : Promise.resolve(undefined),\n            model.contextLength ? Promise.resolve(undefined) : fetchContextLength(model.modelId),\n          ]);\n          if (size) {\n            model.sizeBytes = size;\n          }\n          if (context) {\n            model.contextLength = context;\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;\nlet globalServerHtml = \"\"; // Current HTML content to serve\n\n// Page tracking for memory management\nlet activePagesCount = 0;\nconst MAX_CONCURRENT_PAGES = 5; // Limit to prevent dev mistakes\n\n// Track active backend instances for memory monitoring and cleanup\nconst activeBackends: Set<ChromeGPUBackend> = new Set();\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  /** Model context length (for KV cache management) */\n  contextLength?: number;\n  /** Whether this is a vision model (auto-detected if not specified) */\n  isVision?: boolean;\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  /** Images for vision models (URLs or data URIs) */\n  images?: 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, contextLength = 32_768, isVision = false): 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      AutoProcessor,\n      AutoModelForImageTextToText,\n      RawImage,\n      TextStreamer,\n      InterruptableStoppingCriteria,\n      env,\n    } from \"https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.8.1\";\n\n    // Enable IndexedDB caching (prevents re-downloading models)\n    env.useBrowserCache = true;\n    env.allowLocalModels = false;\n\n    const IS_VISION = ${isVision};\n    \n    class ModelPipeline {\n      static tokenizer = null;\n      static processor = null;\n      static model = null;\n      static modelId = \"${modelPath}\";\n      static isVision = IS_VISION;\n\n      static async getInstance(progressCallback) {\n        if (this.isVision) {\n          // Vision model: use AutoProcessor + AutoModelForImageTextToText\n          if (!this.processor) {\n            this.processor = await AutoProcessor.from_pretrained(this.modelId, {\n              progress_callback: progressCallback,\n            });\n          }\n          if (!this.model) {\n            this.model = await AutoModelForImageTextToText.from_pretrained(this.modelId, {\n              device: \"webgpu\",\n              progress_callback: progressCallback,\n            });\n          }\n          return { \n            processor: this.processor, \n            tokenizer: this.processor.tokenizer, \n            model: this.model,\n            isVision: true \n          };\n        } else {\n          // Text model: use AutoTokenizer + AutoModelForCausalLM\n          if (!this.tokenizer) {\n            this.tokenizer = await AutoTokenizer.from_pretrained(this.modelId, {\n              progress_callback: progressCallback,\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          return { \n            tokenizer: this.tokenizer, \n            model: this.model,\n            isVision: false \n          };\n        }\n      }\n    }\n\n    const stoppingCriteria = new InterruptableStoppingCriteria();\n    let pastKeyValuesCache = null;\n    let totalTokensInCache = 0;\n    \n    // Context length for auto-reset (passed from model config)\n    const CONTEXT_LENGTH = ${contextLength};\n\n    // Auto-load model on page init\n    (async function() {\n      console.log(JSON.stringify({ type: \"progress\", status: IS_VISION ? \"Loading vision model...\" : \"Loading model...\" }));\n      \n      try {\n        const result = 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        \n        // Warmup generation to compile shaders and initialize model\n        // Always do text warmup first\n        const textWarmupInputs = result.tokenizer(\"hello\");\n        await result.model.generate({ ...textWarmupInputs, max_new_tokens: 1 });\n        \n        // Vision models also need vision warmup\n        if (result.isVision) {\n          console.log(JSON.stringify({ type: \"progress\", status: \"Warming up vision encoder...\" }));\n          try {\n            // Create a tiny 8x8 red test image\n            const canvas = new OffscreenCanvas(8, 8);\n            const ctx = canvas.getContext('2d');\n            ctx.fillStyle = 'red';\n            ctx.fillRect(0, 0, 8, 8);\n            const blob = await canvas.convertToBlob({ type: 'image/png' });\n            const warmupImage = await RawImage.fromBlob(blob);\n            \n            // Process with vision pipeline\n            const warmupContent = [{ type: \"image\" }, { type: \"text\", text: \"hi\" }];\n            const warmupMessages = [{ role: \"user\", content: warmupContent }];\n            const warmupPrompt = result.processor.apply_chat_template(warmupMessages);\n            const warmupInputs = await result.processor(warmupImage, warmupPrompt, { add_special_tokens: false });\n            \n            // Run vision warmup generation\n            await result.model.generate({\n              ...warmupInputs,\n              max_new_tokens: 1,\n              do_sample: false,\n            });\n          } catch {\n            // Vision warmup failed, text warmup was done so continue\n          }\n        }\n        \n        // Set page title to model ID for cross-process identification\n        document.title = \"Gerbil: \" + ModelPipeline.modelId;\n        \n        console.log(JSON.stringify({ type: \"ready\", isVision: result.isVision }));\n      } catch (error) {\n        console.log(JSON.stringify({ type: \"error\", error: error.message || String(error) }));\n      }\n    })();\n\n    // Text generation (for non-vision models or vision without images)\n    window.gerbilGenerate = async function(messages, options = {}) {\n      const { maxTokens = 256, temperature = 0.7, topP = 0.9, topK = 20, thinking = false, images = [] } = options;\n      \n      const result = await ModelPipeline.getInstance();\n      \n      // Route to vision generation if we have images and this is a vision model\n      if (images.length > 0 && result.isVision) {\n        return window.gerbilGenerateVision(messages, images, options);\n      }\n      \n      // Auto-reset KV cache if it exceeds context length\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 } = result;\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 += 1;\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          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        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        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    // Vision generation (for vision models with images)\n    window.gerbilGenerateVision = async function(messages, imageUrls, options = {}) {\n      const { maxTokens = 2048, temperature = 0.7, topP = 0.9, topK = 20 } = options;\n\n      try {\n        const { processor, tokenizer, model } = await ModelPipeline.getInstance();\n\n        // Build message content with image placeholders for the user prompt\n        const lastMessage = messages[messages.length - 1];\n        const content = [];\n        for (let i = 0; i < imageUrls.length; i += 1) {\n          content.push({ type: \"image\" });\n        }\n        content.push({ type: \"text\", text: lastMessage.content });\n\n        // For vision models, include a brief system instruction for concise responses\n        const visionMessages = [\n          { role: \"system\", content: \"You are a helpful assistant. Be concise and direct in your responses.\" },\n          { role: \"user\", content }\n        ];\n\n        // Apply chat template with generation prompt\n        const chatPrompt = processor.apply_chat_template(visionMessages, { \n          add_generation_prompt: true \n        });\n\n        // Load images\n        console.log(JSON.stringify({ type: \"progress\", status: \"Loading images...\" }));\n        const loadedImages = await Promise.all(\n          imageUrls.map(url => RawImage.fromURL(url))\n        );\n\n        // Process inputs\n        const inputs = await processor(\n          loadedImages.length === 1 ? loadedImages[0] : loadedImages,\n          chatPrompt,\n          { add_special_tokens: false }\n        );\n\n        let startTime = null;\n        let numTokens = 0;\n\n        const streamCallback = (text) => {\n          startTime ??= performance.now();\n          numTokens += 1;\n          const tps = (numTokens / (performance.now() - startTime)) * 1000;\n          console.log(JSON.stringify({ type: \"token\", text, state: \"answering\", numTokens, tps }));\n        };\n\n        const streamer = new TextStreamer(tokenizer, {\n          skip_prompt: true,\n          skip_special_tokens: true,\n          callback_function: streamCallback,\n        });\n\n        console.log(JSON.stringify({ type: \"start\" }));\n\n        const outputs = await model.generate({\n          ...inputs,\n          max_new_tokens: maxTokens,\n          do_sample: temperature > 0,\n          temperature: temperature > 0 ? temperature : undefined,\n          top_p: topP,\n          top_k: topK,\n          streamer,\n          stopping_criteria: stoppingCriteria,\n        });\n\n        // Decode output (skip prompt)\n        const inputLength = inputs.input_ids.dims?.at(-1) || 0;\n        const decoded = processor.batch_decode(\n          outputs.slice(null, [inputLength, null]),\n          { skip_special_tokens: true }\n        );\n\n        const endTime = performance.now();\n        const totalTime = startTime ? endTime - startTime : 0;\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        }));\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 server: Server | 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 isVisionModel: boolean = false; // Whether this is a vision model\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  private constructor(modelId: string, isVision = false) {\n    this.modelId = modelId;\n    this.isVisionModel = isVision;\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\n    // Detect if this is a vision model (can be overridden via options)\n    const isVision = options.isVision ?? ChromeGPUBackend.detectVisionModel(modelId);\n\n    const backend = new ChromeGPUBackend(modelId, isVision);\n    await backend.launch(options);\n    return backend;\n  }\n\n  /**\n   * Detect if a model is a vision model based on its ID\n   */\n  private static detectVisionModel(modelId: string): boolean {\n    const visionModelPatterns = [\n      /ministral/i,\n      /pixtral/i,\n      /llava/i,\n      /vision/i,\n      /vl/i,\n      /image-text/i,\n      /multimodal/i,\n    ];\n    return visionModelPatterns.some((pattern) => pattern.test(modelId));\n  }\n\n  /**\n   * Check if this backend is for a vision model\n   */\n  isVision(): boolean {\n    return this.isVisionModel;\n  }\n\n  /**\n   * Clean up orphan Gerbil pages from previous sessions\n   * These are pages that were left behind when process exited without proper cleanup\n   */\n  private async cleanupOrphanPages(\n    browser: Browser,\n    options: ChromeBackendOptions,\n  ): Promise<number> {\n    try {\n      const pages = await browser.pages();\n      const gerbilPages = pages.filter((p) => {\n        const url = p.url();\n        // Match our worker pages by URL pattern (127.0.0.1:PORT where PORT > 40000)\n        return /127\\.0\\.0\\.1:4\\d{4}/.test(url);\n      });\n\n      // Only clean up if there are orphan pages not tracked by this process\n      const orphanCount = gerbilPages.length - activeBackends.size;\n      if (orphanCount > 0) {\n        options.onProgress?.({ status: `Cleaning up ${orphanCount} orphan page(s)...` });\n\n        for (const page of gerbilPages) {\n          // Check if this page is owned by a current backend\n          let isOwned = false;\n          for (const backend of activeBackends) {\n            if (backend.page === page) {\n              isOwned = true;\n              break;\n            }\n          }\n\n          // Close unowned pages\n          if (!isOwned) {\n            try {\n              await page.close();\n            } catch {\n              // Page may already be closed\n            }\n          }\n        }\n      }\n\n      return orphanCount;\n    } catch {\n      return 0;\n    }\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      // Clean up orphan pages from previous sessions\n      await this.cleanupOrphanPages(globalBrowser, options);\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        // Clean up orphan pages from previous sessions\n        await this.cleanupOrphanPages(globalBrowser, options);\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 contextLength = options.contextLength || 32_768; // Default to 32K if not specified\n    const html = getWorkerPageHTML(this.modelId, contextLength, this.isVisionModel);\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 and register this backend\n    activePagesCount += 1;\n    activeBackends.add(this);\n    options.onProgress?.({ status: `Active pages: ${activePagesCount}/${MAX_CONCURRENT_PAGES}` });\n\n    // Listen for browser disconnect (OOM kill, crash, etc.)\n    this.browser.on(\"disconnected\", () => {\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        }\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    });\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(): { pid: number | null; port: number; modelId: string; startedAt: Date | null } {\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<{ jsHeapUsed: number; jsHeapTotal: number } | 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      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<{ usedGB: number; totalGB: number; usedPercent: number } | 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 ?? (this.isVisionModel ? 2048 : 256),\n      temperature: options.temperature ?? 0.7,\n      topP: options.topP ?? 0.9,\n      topK: options.topK ?? 20,\n      thinking: options.thinking ?? false,\n      images: options.images ?? [], // Pass images for vision models\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   * Updates HTML content for new model loads\n   */\n  private async startServer(html: string): Promise<void> {\n    // Always update the HTML content for new model loads\n    globalServerHtml = html;\n\n    // If global server is already running, reuse it (with updated HTML)\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(globalServerHtml); // Serve current HTML (not captured at creation time)\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   * @param disconnect If true, also disconnect from shared browser (for clean script exit)\n   */\n  async dispose(disconnect = false): 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    // Detach CDP session first\n    if (this.cdp) {\n      try {\n        await this.cdp.detach();\n      } catch {\n        // CDP may already be detached\n      }\n      this.cdp = null;\n    }\n\n    // Close our page (but NOT the shared browser - other backends may use it)\n    if (this.page) {\n      try {\n        // Navigate away first to ensure page is \"dead\"\n        await this.page.goto(\"about:blank\").catch(() => {});\n\n        // Small delay for navigation\n        await new Promise((r) => setTimeout(r, 50));\n\n        // Now close\n        await this.page.close({ runBeforeUnload: false });\n\n        // Decrement active page counter\n        activePagesCount = Math.max(0, activePagesCount - 1);\n      } catch {\n        // Ignore close errors\n      }\n      this.page = null;\n    }\n\n    // Unregister from active backends\n    activeBackends.delete(this);\n\n    // Clear our references first\n    this.browser = null;\n    this.server = null;\n\n    // Small delay to ensure page close completes before disconnect\n    if (disconnect) {\n      await new Promise((r) => setTimeout(r, 100));\n    }\n\n    // If requested and we're the last backend, disconnect from browser\n    // This allows clean script exit without killing browser for other processes\n    if (disconnect && activeBackends.size === 0 && globalBrowser) {\n      try {\n        globalBrowser.disconnect();\n        globalBrowser = null;\n        globalBrowserPromise = null;\n      } catch {\n        // Browser may already be disconnected\n      }\n    }\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   * Get global browser status (even if no active backends)\n   */\n  static getGlobalBrowserStatus(): {\n    running: boolean;\n    pid: number | null;\n    port: number;\n    activePagesCount: number;\n    maxPages: number;\n    wsEndpoint: string | null;\n  } {\n    let pid: number | null = null;\n    let wsEndpoint: string | null = null;\n\n    if (globalBrowser?.connected) {\n      const browserProcess = globalBrowser.process?.();\n      if (browserProcess?.pid) {\n        pid = browserProcess.pid;\n      }\n      wsEndpoint = globalBrowser.wsEndpoint();\n    }\n\n    return {\n      running: globalBrowser?.connected ?? false,\n      pid,\n      port: globalServerPort,\n      activePagesCount, // This process only\n      maxPages: MAX_CONCURRENT_PAGES,\n      wsEndpoint,\n    };\n  }\n\n  /**\n   * Get total page count from Chrome (all processes)\n   */\n  static async getTotalPageCount(): Promise<number> {\n    if (!globalBrowser?.connected) {\n      return 0;\n    }\n\n    try {\n      const pages = await globalBrowser.pages();\n      // Count only Gerbil pages\n      return pages.filter((p) => {\n        const url = p.url();\n        return url.includes(`127.0.0.1:${globalServerPort}`);\n      }).length;\n    } catch {\n      return 0;\n    }\n  }\n\n  /**\n   * Get all active backends with their memory usage (this process only)\n   */\n  static async getAllBackendsInfo(): Promise<\n    Array<{\n      modelId: string;\n      isVision: boolean;\n      isReady: boolean;\n      memory: { usedGB: number; totalGB: number; usedPercent: number } | null;\n    }>\n  > {\n    const results = [];\n\n    for (const backend of activeBackends) {\n      const mem = await backend.getMemoryStats();\n      results.push({\n        modelId: backend.modelId,\n        isVision: backend.isVisionModel,\n        isReady: backend.isReady,\n        memory: mem,\n      });\n    }\n\n    return results;\n  }\n\n  /**\n   * Get ALL pages in Chrome browser (cross-process visibility)\n   * This shows pages from ALL Gerbil processes sharing the browser\n   */\n  static async getAllChromePages(): Promise<\n    Array<{\n      url: string;\n      title: string;\n      isOurs: boolean; // true if this process owns it\n      modelId: string | null; // extracted from URL/title if possible\n      memory: { usedGB: number; totalGB: number } | null;\n    }>\n  > {\n    if (!globalBrowser?.connected) {\n      return [];\n    }\n\n    try {\n      const pages = await globalBrowser.pages();\n\n      const results = [];\n      for (const page of pages) {\n        const url = page.url();\n        const title = await page.title().catch(() => \"\");\n\n        // Skip about:blank and non-gerbil pages\n        if (url === \"about:blank\" || !url.includes(`127.0.0.1:${globalServerPort}`)) {\n          continue;\n        }\n\n        // Check if we own this page\n        let modelId: string | null = null;\n        let isOurs = false;\n        let memory: { usedGB: number; totalGB: number } | null = null;\n\n        for (const backend of activeBackends) {\n          if (backend.page === page) {\n            isOurs = true;\n            modelId = backend.modelId;\n            // Get memory from our backend\n            const mem = await backend.getMemoryStats();\n            if (mem) {\n              memory = { usedGB: mem.usedGB, totalGB: mem.totalGB };\n            }\n            break;\n          }\n        }\n\n        // For pages we don't own, extract model ID from title and try to get memory\n        if (!isOurs) {\n          if (title.startsWith(\"Gerbil: \")) {\n            modelId = title.replace(\"Gerbil: \", \"\");\n          }\n\n          // Try to get memory for other sessions via CDP\n          try {\n            const cdp = await page.createCDPSession();\n            await cdp.send(\"Performance.enable\");\n            const { metrics } = (await cdp.send(\"Performance.getMetrics\")) as {\n              metrics: Array<{ name: string; value: number }>;\n            };\n            const jsHeapUsed = metrics.find((m) => m.name === \"JSHeapUsedSize\")?.value ?? 0;\n            const jsHeapTotal = metrics.find((m) => m.name === \"JSHeapTotalSize\")?.value ?? 0;\n            memory = {\n              usedGB: jsHeapUsed / 1024 ** 3,\n              totalGB: jsHeapTotal / 1024 ** 3,\n            };\n            await cdp.detach();\n          } catch {\n            // Can't get memory for this page\n          }\n        }\n\n        results.push({\n          url,\n          title: title || \"Gerbil WebGPU Backend\",\n          isOurs,\n          modelId,\n          memory,\n        });\n      }\n\n      return results;\n    } catch {\n      return [];\n    }\n  }\n\n  /**\n   * Kill a Chrome page by index (works cross-process)\n   */\n  static async killPageByIndex(index: number): Promise<boolean> {\n    if (!globalBrowser?.connected) {\n      return false;\n    }\n\n    try {\n      const pages = await globalBrowser.pages();\n      // Filter to only Gerbil pages\n      const gerbilPages = pages.filter((p) => {\n        const url = p.url();\n        return url.includes(`127.0.0.1:${globalServerPort}`);\n      });\n\n      if (index < 0 || index >= gerbilPages.length) {\n        return false;\n      }\n\n      const page = gerbilPages[index];\n\n      // If it's our page, use dispose() for proper cleanup\n      for (const backend of activeBackends) {\n        if (backend.page === page) {\n          await backend.dispose();\n          return true;\n        }\n      }\n\n      // Otherwise just close the page directly\n      await page.close();\n      return true;\n    } catch {\n      return false;\n    }\n  }\n\n  /**\n   * Kill a specific backend by index (this process only)\n   */\n  static async killBackendByIndex(index: number): Promise<boolean> {\n    const backends = [...activeBackends];\n    if (index < 0 || index >= backends.length) {\n      return false;\n    }\n\n    const backend = backends[index];\n    try {\n      await backend.dispose();\n      return true;\n    } catch {\n      return false;\n    }\n  }\n\n  /**\n   * Force kill all backends (for zombie cleanup)\n   */\n  static async killAllBackends(): Promise<{ pagesKilled: number; browserKilled: boolean }> {\n    const count = activeBackends.size;\n\n    // Dispose all active backends\n    for (const backend of [...activeBackends]) {\n      try {\n        await backend.dispose();\n      } catch {\n        // Ignore errors during cleanup\n      }\n    }\n    activeBackends.clear();\n\n    // Close shared browser\n    let browserKilled = false;\n    if (globalBrowser) {\n      try {\n        await globalBrowser.close();\n        browserKilled = true;\n      } catch {\n        // Browser may already be closed\n      }\n      globalBrowser = null;\n      globalBrowserPromise = null;\n    }\n\n    // Close server\n    if (globalServer) {\n      globalServer.close();\n      globalServer = null;\n      globalServerPort = 0;\n    }\n\n    activePagesCount = 0;\n\n    // Clean up endpoint file\n    try {\n      unlinkSync(WS_ENDPOINT_FILE);\n    } catch {}\n\n    return { pagesKilled: count, browserKilled };\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;;AAe3E,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,mBAAmB,SAA8C;AAE9E,KAAI;EACF,MAAM,MAAM,MAAM,MAAM,0BAA0B,QAAQ,uBAAuB;AACjF,MAAI,IAAI,IAAI;GACV,MAAM,SAAS,MAAM,IAAI,MAAM;GAE/B,MAAM,aAAa,OAAO,eAAe,EAAE;GAC3C,MAAM,SACJ,OAAO,2BACP,WAAW,2BACX,OAAO,kBACP,WAAW,kBACX,OAAO,eACP,OAAO,uBACP,OAAO,SACP,OAAO;AACT,OAAI,OACF,QAAO;;SAGL;AAKR,KAAI;EACF,MAAM,SAAS,MAAM,MAAM,0BAA0B,QAAQ,iCAAiC;AAC9F,MAAI,OAAO,IAAI;GACb,MAAM,YAAY,MAAM,OAAO,MAAM;AAErC,OAAI,UAAU,oBAAoB,UAAU,mBAAmB,IAC7D,QAAO,UAAU;;SAGf;;;AAQV,SAAS,YAAY,MAA0D;AAE7E,QAAO,KAAK,KAAK,QAAQ,KAAK,QAAQ;;;AAIxC,eAAe,eAAe,SAA8C;AAC1E,KAAI;EAEF,MAAM,UAAU,MAAM,MAAM,qCAAqC,QAAQ,iBAAiB;AAC1F,MAAI,QAAQ,IAAI;GACd,MAAMA,QACJ,MAAM,QAAQ,MAAM;GAGtB,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;AAExC,OAAI,UAAU;IAEZ,MAAM,WAAW,SAAS,KAAK,QAAQ,SAAS,GAAG;IAInD,MAAM,YAHe,MAAM,QACxB,MAAM,EAAE,SAAS,SAAS,QAAQ,EAAE,KAAK,WAAW,GAAG,SAAS,YAAY,CAC9E,CAC8B,QAAQ,KAAK,MAAM,MAAM,YAAY,EAAE,EAAE,EAAE;AAC1E,QAAI,YAAY,EACd,QAAO;;;EAMb,MAAM,MAAM,MAAM,MAAM,qCAAqC,UAAU;AACvE,MAAI,IAAI,GAEN,SADa,MAAM,IAAI,MAAM,EACjB;SAER;;;AAOV,SAAgB,iBACd,SACA,WACA,eACM;AACN,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;AAEvB,OAAI,cACF,UAAS,gBAAgB;QAG3B,QAAO,KAAK;GACV;GACA,cAAc;GACd,UAAU;GACV;GACA;GACD,CAAC;AAGJ,gBAAc,oBAAoB,KAAK,UAAU,EAAE,QAAQ,EAAE,MAAM,EAAE,CAAC;EAGtE,MAAM,YAAY,EAAE,aAAa,UAAU;EAC3C,MAAM,eAAe,EAAE,iBAAiB,UAAU;AAElD,MAAI,aAAa,aACf,SAAQ,IAAI,CACV,YAAY,eAAe,QAAQ,GAAG,QAAQ,QAAQ,OAAU,EAChE,eAAe,mBAAmB,QAAQ,GAAG,QAAQ,QAAQ,OAAU,CACxE,CAAC,CACC,MAAM,CAAC,MAAM,aAAa;GACzB,MAAM,gBAAgB,uBAAuB;GAC7C,MAAM,QAAQ,cAAc,MAAM,MAAM,EAAE,YAAY,QAAQ;AAC9D,OAAI,OAAO;AACT,QAAI,KACF,OAAM,YAAY;AAEpB,QAAI,QACF,OAAM,gBAAgB;AAExB,kBAAc,oBAAoB,KAAK,UAAU,EAAE,QAAQ,eAAe,EAAE,MAAM,EAAE,CAAC;;IAEvF,CACD,YAAY,GAAG;SAEd;;;AAgBV,eAAsB,0BAAyC;AAC7D,KAAI;EACF,MAAM,SAAS,uBAAuB;EAEtC,MAAM,oBAAoB;EAC1B,MAAM,eAAe,OAAO,QACzB,MAAM,CAAC,EAAE,aAAa,EAAE,YAAY,qBAAqB,CAAC,EAAE,cAC9D;AACD,MAAI,aAAa,WAAW,EAC1B;EAIF,MAAM,YAAY;AAClB,OAAK,IAAI,IAAI,GAAG,IAAI,aAAa,QAAQ,KAAK,WAAW;GACvD,MAAM,QAAQ,aAAa,MAAM,GAAG,IAAI,UAAU;AAClD,SAAM,QAAQ,IACZ,MAAM,IAAI,OAAO,UAAU;IACzB,MAAM,CAAC,MAAM,WAAW,MAAM,QAAQ,IAAI,CACxC,CAAC,MAAM,aAAa,MAAM,YAAY,oBAClC,eAAe,MAAM,QAAQ,GAC7B,QAAQ,QAAQ,OAAU,EAC9B,MAAM,gBAAgB,QAAQ,QAAQ,OAAU,GAAG,mBAAmB,MAAM,QAAQ,CACrF,CAAC;AACF,QAAI,KACF,OAAM,YAAY;AAEpB,QAAI,QACF,OAAM,gBAAgB;KAExB,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;AACvB,IAAI,mBAAmB;AAGvB,IAAI,mBAAmB;AACvB,MAAM,uBAAuB;AAG7B,MAAMC,iCAAwC,IAAI,KAAK;AAqCvD,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,WAAmB,gBAAgB,OAAQ,WAAW,OAAe;AAC9F,QAAO;;;;;;;;;;;;;;;;;;;;;wBAqBe,SAAS;;;;;;0BAMP,UAAU;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;6BAmDP,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8S3C,IAAa,mBAAb,MAAa,iBAAiB;CAC5B,AAAQ,UAA0B;CAClC,AAAQ,OAAoB;CAC5B,AAAQ,MAAyB;CACjC,AAAQ,SAAwB;CAChC,AAAQ,aAAa;CACrB,AAAQ,cAAsB;CAC9B,AAAiB;CACjB,AAAQ,UAAU;CAClB,AAAiB,gBAAyB;CAC1C,AAAiB,kCAAoD,IAAI,KAAK;CAC9E,AAAQ,iBAA8C,EAAE;CAExD,AAAQ,YAAY,SAAiB,WAAW,OAAO;AACrD,OAAK,UAAU;AACf,OAAK,gBAAgB;;;;;CAMvB,aAAa,OAAO,UAAgC,EAAE,EAA6B;EACjF,MAAM,UAAU,QAAQ,WAAW;EAKnC,MAAM,UAAU,IAAI,iBAAiB,SAFpB,QAAQ,YAAY,iBAAiB,kBAAkB,QAAQ,CAEzB;AACvD,QAAM,QAAQ,OAAO,QAAQ;AAC7B,SAAO;;;;;CAMT,OAAe,kBAAkB,SAA0B;AAUzD,SAT4B;GAC1B;GACA;GACA;GACA;GACA;GACA;GACA;GACD,CAC0B,MAAM,YAAY,QAAQ,KAAK,QAAQ,CAAC;;;;;CAMrE,WAAoB;AAClB,SAAO,KAAK;;;;;;CAOd,MAAc,mBACZ,SACA,SACiB;AACjB,MAAI;GAEF,MAAM,eADQ,MAAM,QAAQ,OAAO,EACT,QAAQ,MAAM;IACtC,MAAM,MAAM,EAAE,KAAK;AAEnB,WAAO,sBAAsB,KAAK,IAAI;KACtC;GAGF,MAAM,cAAc,YAAY,SAAS,eAAe;AACxD,OAAI,cAAc,GAAG;AACnB,YAAQ,aAAa,EAAE,QAAQ,eAAe,YAAY,qBAAqB,CAAC;AAEhF,SAAK,MAAM,QAAQ,aAAa;KAE9B,IAAI,UAAU;AACd,UAAK,MAAM,WAAW,eACpB,KAAI,QAAQ,SAAS,MAAM;AACzB,gBAAU;AACV;;AAKJ,SAAI,CAAC,QACH,KAAI;AACF,YAAM,KAAK,OAAO;aACZ;;;AAOd,UAAO;UACD;AACN,UAAO;;;;;;;CAQX,MAAc,mBACZ,YACA,SACkB;AAElB,MAAI,eAAe,WAAW;AAC5B,WAAQ,aAAa,EAAE,QAAQ,8BAA8B,CAAC;AAE9D,SAAM,KAAK,mBAAmB,eAAe,QAAQ;AACrD,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;AAEF,SAAM,KAAK,mBAAmB,eAAe,QAAQ;AACrD,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,gBAAgB,QAAQ,iBAAiB;EAC/C,MAAM,OAAO,kBAAkB,KAAK,SAAS,eAAe,KAAK,cAAc;AAC/E,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,sBAAoB;AACpB,iBAAe,IAAI,KAAK;AACxB,UAAQ,aAAa,EAAE,QAAQ,iBAAiB,iBAAiB,GAAG,wBAAwB,CAAC;AAG7F,OAAK,QAAQ,GAAG,sBAAsB;AACpC,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,KACd;;IAGJ;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;IAEF;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,YAA2F;EAEzF,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,iBAA8E;AAClF,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;AAKT,MAFe,IAAI,aAAa,QAAQ,IAE3B,aAAa;AACxB,SAAM,KAAK,OAAO;AAClB,UAAO;;AAGT,SAAO;;;;;CAMT,MAAM,iBAA2F;EAC/F,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,cAAc,KAAK,gBAAgB,OAAO;GAC7D,aAAa,QAAQ,eAAe;GACpC,MAAM,QAAQ,QAAQ;GACtB,MAAM,QAAQ,QAAQ;GACtB,UAAU,QAAQ,YAAY;GAC9B,QAAQ,QAAQ,UAAU,EAAE;GAC7B;AAGD,MAAI,QAAQ,QACV,MAAK,gBAAgB,IAAI,SAAS,QAAQ,QAAQ;AAGpD,MAAI;GAEF,MAAM,gBAAgB,KAAK,MAAM,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;;;;;;;CAQd,MAAc,YAAY,MAA6B;AAErD,qBAAmB;AAGnB,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,iBAAiB;KACzB;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;;;;;;;CAQJ,MAAM,QAAQ,aAAa,OAAsB;AAE/C,OAAK,UAAU;AAGf,OAAK,iBAAiB,EAAE;AACxB,OAAK,gBAAgB,OAAO;AAG5B,MAAI,KAAK,KAAK;AACZ,OAAI;AACF,UAAM,KAAK,IAAI,QAAQ;WACjB;AAGR,QAAK,MAAM;;AAIb,MAAI,KAAK,MAAM;AACb,OAAI;AAEF,UAAM,KAAK,KAAK,KAAK,cAAc,CAAC,YAAY,GAAG;AAGnD,UAAM,IAAI,SAAS,MAAM,WAAW,GAAG,GAAG,CAAC;AAG3C,UAAM,KAAK,KAAK,MAAM,EAAE,iBAAiB,OAAO,CAAC;AAGjD,uBAAmB,KAAK,IAAI,GAAG,mBAAmB,EAAE;WAC9C;AAGR,QAAK,OAAO;;AAId,iBAAe,OAAO,KAAK;AAG3B,OAAK,UAAU;AACf,OAAK,SAAS;AAGd,MAAI,WACF,OAAM,IAAI,SAAS,MAAM,WAAW,GAAG,IAAI,CAAC;AAK9C,MAAI,cAAc,eAAe,SAAS,KAAK,cAC7C,KAAI;AACF,iBAAc,YAAY;AAC1B,mBAAgB;AAChB,0BAAuB;UACjB;;;;;CASZ,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,OAAO,yBAOL;EACA,IAAID,MAAqB;EACzB,IAAIE,aAA4B;AAEhC,MAAI,eAAe,WAAW;GAC5B,MAAM,iBAAiB,cAAc,WAAW;AAChD,OAAI,gBAAgB,IAClB,OAAM,eAAe;AAEvB,gBAAa,cAAc,YAAY;;AAGzC,SAAO;GACL,SAAS,eAAe,aAAa;GACrC;GACA,MAAM;GACN;GACA,UAAU;GACV;GACD;;;;;CAMH,aAAa,oBAAqC;AAChD,MAAI,CAAC,eAAe,UAClB,QAAO;AAGT,MAAI;AAGF,WAFc,MAAM,cAAc,OAAO,EAE5B,QAAQ,MAAM;AAEzB,WADY,EAAE,KAAK,CACR,SAAS,aAAa,mBAAmB;KACpD,CAAC;UACG;AACN,UAAO;;;;;;CAOX,aAAa,qBAOX;EACA,MAAM,UAAU,EAAE;AAElB,OAAK,MAAM,WAAW,gBAAgB;GACpC,MAAM,MAAM,MAAM,QAAQ,gBAAgB;AAC1C,WAAQ,KAAK;IACX,SAAS,QAAQ;IACjB,UAAU,QAAQ;IAClB,SAAS,QAAQ;IACjB,QAAQ;IACT,CAAC;;AAGJ,SAAO;;;;;;CAOT,aAAa,oBAQX;AACA,MAAI,CAAC,eAAe,UAClB,QAAO,EAAE;AAGX,MAAI;GACF,MAAM,QAAQ,MAAM,cAAc,OAAO;GAEzC,MAAM,UAAU,EAAE;AAClB,QAAK,MAAM,QAAQ,OAAO;IACxB,MAAM,MAAM,KAAK,KAAK;IACtB,MAAM,QAAQ,MAAM,KAAK,OAAO,CAAC,YAAY,GAAG;AAGhD,QAAI,QAAQ,iBAAiB,CAAC,IAAI,SAAS,aAAa,mBAAmB,CACzE;IAIF,IAAIC,UAAyB;IAC7B,IAAI,SAAS;IACb,IAAIC,SAAqD;AAEzD,SAAK,MAAM,WAAW,eACpB,KAAI,QAAQ,SAAS,MAAM;AACzB,cAAS;AACT,eAAU,QAAQ;KAElB,MAAM,MAAM,MAAM,QAAQ,gBAAgB;AAC1C,SAAI,IACF,UAAS;MAAE,QAAQ,IAAI;MAAQ,SAAS,IAAI;MAAS;AAEvD;;AAKJ,QAAI,CAAC,QAAQ;AACX,SAAI,MAAM,WAAW,WAAW,CAC9B,WAAU,MAAM,QAAQ,YAAY,GAAG;AAIzC,SAAI;MACF,MAAM,MAAM,MAAM,KAAK,kBAAkB;AACzC,YAAM,IAAI,KAAK,qBAAqB;MACpC,MAAM,EAAE,YAAa,MAAM,IAAI,KAAK,yBAAyB;MAG7D,MAAM,aAAa,QAAQ,MAAM,MAAM,EAAE,SAAS,iBAAiB,EAAE,SAAS;MAC9E,MAAM,cAAc,QAAQ,MAAM,MAAM,EAAE,SAAS,kBAAkB,EAAE,SAAS;AAChF,eAAS;OACP,QAAQ,aAAa,QAAQ;OAC7B,SAAS,cAAc,QAAQ;OAChC;AACD,YAAM,IAAI,QAAQ;aACZ;;AAKV,YAAQ,KAAK;KACX;KACA,OAAO,SAAS;KAChB;KACA;KACA;KACD,CAAC;;AAGJ,UAAO;UACD;AACN,UAAO,EAAE;;;;;;CAOb,aAAa,gBAAgB,OAAiC;AAC5D,MAAI,CAAC,eAAe,UAClB,QAAO;AAGT,MAAI;GAGF,MAAM,eAFQ,MAAM,cAAc,OAAO,EAEf,QAAQ,MAAM;AAEtC,WADY,EAAE,KAAK,CACR,SAAS,aAAa,mBAAmB;KACpD;AAEF,OAAI,QAAQ,KAAK,SAAS,YAAY,OACpC,QAAO;GAGT,MAAM,OAAO,YAAY;AAGzB,QAAK,MAAM,WAAW,eACpB,KAAI,QAAQ,SAAS,MAAM;AACzB,UAAM,QAAQ,SAAS;AACvB,WAAO;;AAKX,SAAM,KAAK,OAAO;AAClB,UAAO;UACD;AACN,UAAO;;;;;;CAOX,aAAa,mBAAmB,OAAiC;EAC/D,MAAM,WAAW,CAAC,GAAG,eAAe;AACpC,MAAI,QAAQ,KAAK,SAAS,SAAS,OACjC,QAAO;EAGT,MAAM,UAAU,SAAS;AACzB,MAAI;AACF,SAAM,QAAQ,SAAS;AACvB,UAAO;UACD;AACN,UAAO;;;;;;CAOX,aAAa,kBAA4E;EACvF,MAAM,QAAQ,eAAe;AAG7B,OAAK,MAAM,WAAW,CAAC,GAAG,eAAe,CACvC,KAAI;AACF,SAAM,QAAQ,SAAS;UACjB;AAIV,iBAAe,OAAO;EAGtB,IAAI,gBAAgB;AACpB,MAAI,eAAe;AACjB,OAAI;AACF,UAAM,cAAc,OAAO;AAC3B,oBAAgB;WACV;AAGR,mBAAgB;AAChB,0BAAuB;;AAIzB,MAAI,cAAc;AAChB,gBAAa,OAAO;AACpB,kBAAe;AACf,sBAAmB;;AAGrB,qBAAmB;AAGnB,MAAI;AACF,cAAW,iBAAiB;UACtB;AAER,SAAO;GAAE,aAAa;GAAO;GAAe;;;;;CAM9C,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"}
+{"version":3,"file":"chrome-backend-CORwaIyC.mjs","names":["files: { path: string; size?: number; lfs?: { size?: number } }[]","globalBrowser: Browser | null","globalBrowserPromise: Promise<Browser> | null","globalServer: Server | null","activeBackends: Set<ChromeGPUBackend>","pid: number | null","err: any","wsEndpoint: string | null","modelId: string | null","memory: { usedGB: number; totalGB: number } | null"],"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 \"node:child_process\";\nimport { existsSync, mkdirSync, readFileSync, rmSync, unlinkSync, writeFileSync } from \"node:fs\";\nimport { createServer, type Server } from \"node:http\";\nimport { homedir } from \"node:os\";\nimport { join } from \"node: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  contextLength?: 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 context length from HuggingFace (config.json preferred for actual limit) */\nasync function fetchContextLength(modelId: string): Promise<number | undefined> {\n  // Priority 1: config.json → max_position_embeddings (actual architectural limit)\n  try {\n    const res = await fetch(`https://huggingface.co/${modelId}/raw/main/config.json`);\n    if (res.ok) {\n      const config = await res.json();\n      // Check root level first, then nested text_config (for multimodal models like Ministral)\n      const textConfig = config.text_config || {};\n      const ctxLen =\n        config.max_position_embeddings ||\n        textConfig.max_position_embeddings ||\n        config.sliding_window ||\n        textConfig.sliding_window ||\n        config.max_seq_len ||\n        config.max_sequence_length ||\n        config.n_ctx ||\n        config.n_positions;\n      if (ctxLen) {\n        return ctxLen;\n      }\n    }\n  } catch {\n    // Continue to fallback\n  }\n\n  // Priority 2: tokenizer_config.json → model_max_length (fallback only)\n  try {\n    const tokRes = await fetch(`https://huggingface.co/${modelId}/raw/main/tokenizer_config.json`);\n    if (tokRes.ok) {\n      const tokConfig = await tokRes.json();\n      // Only use if reasonable (< 1M tokens - anything larger is likely placeholder)\n      if (tokConfig.model_max_length && tokConfig.model_max_length < 1e6) {\n        return tokConfig.model_max_length;\n      }\n    }\n  } catch {\n    // Ignore\n  }\n\n  return;\n}\n\n/** Get file size from HuggingFace tree entry (handles both regular and LFS files) */\nfunction getFileSize(file: { size?: number; lfs?: { size?: number } }): number {\n  // LFS files have size in lfs.size, regular files in size\n  return file.lfs?.size || file.size || 0;\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/${modelId}/tree/main/onnx`);\n    if (treeRes.ok) {\n      const files: { path: string; size?: number; lfs?: { size?: number } }[] =\n        await treeRes.json();\n\n      // Prefer q4f16 > q4 > fp16 > any .onnx file\n      const q4f16 = files.find((f) => f.path.includes(\"q4f16\") && f.path.endsWith(\".onnx\"));\n      const q4 = files.find(\n        (f) => f.path.includes(\"q4\") && !f.path.includes(\"f16\") && f.path.endsWith(\".onnx\"),\n      );\n      const fp16 = files.find((f) => f.path.includes(\"fp16\") && f.path.endsWith(\".onnx\"));\n      const anyOnnx = files.find((f) => f.path.endsWith(\".onnx\"));\n      const bestFile = q4f16 || q4 || fp16 || anyOnnx;\n\n      if (bestFile) {\n        // Sum up the .onnx file AND all associated .onnx_data* files\n        const baseName = bestFile.path.replace(\".onnx\", \"\");\n        const relatedFiles = files.filter(\n          (f) => f.path === bestFile.path || f.path.startsWith(`${baseName}.onnx_data`),\n        );\n        const totalSize = relatedFiles.reduce((sum, f) => sum + getFileSize(f), 0);\n        if (totalSize > 0) {\n          return totalSize;\n        }\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(\n  modelId: string,\n  sizeBytes?: number,\n  contextLength?: number,\n): 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      if (contextLength) {\n        existing.contextLength = contextLength;\n      }\n    } else {\n      models.push({\n        modelId,\n        downloadedAt: now,\n        lastUsed: now,\n        sizeBytes,\n        contextLength,\n      });\n    }\n\n    writeFileSync(CACHED_MODELS_FILE, JSON.stringify({ models }, null, 2));\n\n    // Fetch missing metadata in background\n    const needsSize = !(sizeBytes || existing?.sizeBytes);\n    const needsContext = !(contextLength || existing?.contextLength);\n\n    if (needsSize || needsContext) {\n      Promise.all([\n        needsSize ? fetchModelSize(modelId) : Promise.resolve(undefined),\n        needsContext ? fetchContextLength(modelId) : Promise.resolve(undefined),\n      ])\n        .then(([size, context]) => {\n          const updatedModels = getChromeCachedModels();\n          const model = updatedModels.find((m) => m.modelId === modelId);\n          if (model) {\n            if (size) {\n              model.sizeBytes = size;\n            }\n            if (context) {\n              model.contextLength = context;\n            }\n            writeFileSync(CACHED_MODELS_FILE, JSON.stringify({ models: updatedModels }, null, 2));\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 metadata (size, context length) for cached models that need it */\nexport async function refreshCachedModelSizes(): Promise<void> {\n  try {\n    const models = getChromeCachedModels();\n    // Refresh if: no size, size < 1MB (wrong), or missing context length\n    const MIN_EXPECTED_SIZE = 1_000_000; // 1MB - real models are always bigger\n    const needsRefresh = models.filter(\n      (m) => !m.sizeBytes || m.sizeBytes < MIN_EXPECTED_SIZE || !m.contextLength,\n    );\n    if (needsRefresh.length === 0) {\n      return;\n    }\n\n    // Fetch metadata in parallel (max 3 at a time)\n    const batchSize = 3;\n    for (let i = 0; i < needsRefresh.length; i += batchSize) {\n      const batch = needsRefresh.slice(i, i + batchSize);\n      await Promise.all(\n        batch.map(async (model) => {\n          const [size, context] = await Promise.all([\n            !model.sizeBytes || model.sizeBytes < MIN_EXPECTED_SIZE\n              ? fetchModelSize(model.modelId)\n              : Promise.resolve(undefined),\n            model.contextLength ? Promise.resolve(undefined) : fetchContextLength(model.modelId),\n          ]);\n          if (size) {\n            model.sizeBytes = size;\n          }\n          if (context) {\n            model.contextLength = context;\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;\nlet globalServerHtml = \"\"; // Current HTML content to serve\n\n// Page tracking for memory management\nlet activePagesCount = 0;\nconst MAX_CONCURRENT_PAGES = 5; // Limit to prevent dev mistakes\n\n// Track active backend instances for memory monitoring and cleanup\nconst activeBackends: Set<ChromeGPUBackend> = new Set();\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  /** Model context length (for KV cache management) */\n  contextLength?: number;\n  /** Whether this is a vision model (auto-detected if not specified) */\n  isVision?: boolean;\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  /** Images for vision models (URLs or data URIs) */\n  images?: 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, contextLength = 32_768, isVision = false): 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      AutoProcessor,\n      AutoModelForImageTextToText,\n      RawImage,\n      TextStreamer,\n      InterruptableStoppingCriteria,\n      env,\n    } from \"https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.8.1\";\n\n    // Enable IndexedDB caching (prevents re-downloading models)\n    env.useBrowserCache = true;\n    env.allowLocalModels = false;\n\n    const IS_VISION = ${isVision};\n    \n    class ModelPipeline {\n      static tokenizer = null;\n      static processor = null;\n      static model = null;\n      static modelId = \"${modelPath}\";\n      static isVision = IS_VISION;\n\n      static async getInstance(progressCallback) {\n        if (this.isVision) {\n          // Vision model: use AutoProcessor + AutoModelForImageTextToText\n          if (!this.processor) {\n            this.processor = await AutoProcessor.from_pretrained(this.modelId, {\n              progress_callback: progressCallback,\n            });\n          }\n          if (!this.model) {\n            this.model = await AutoModelForImageTextToText.from_pretrained(this.modelId, {\n              device: \"webgpu\",\n              progress_callback: progressCallback,\n            });\n          }\n          return { \n            processor: this.processor, \n            tokenizer: this.processor.tokenizer, \n            model: this.model,\n            isVision: true \n          };\n        } else {\n          // Text model: use AutoTokenizer + AutoModelForCausalLM\n          if (!this.tokenizer) {\n            this.tokenizer = await AutoTokenizer.from_pretrained(this.modelId, {\n              progress_callback: progressCallback,\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          return { \n            tokenizer: this.tokenizer, \n            model: this.model,\n            isVision: false \n          };\n        }\n      }\n    }\n\n    const stoppingCriteria = new InterruptableStoppingCriteria();\n    let pastKeyValuesCache = null;\n    let totalTokensInCache = 0;\n    \n    // Context length for auto-reset (passed from model config)\n    const CONTEXT_LENGTH = ${contextLength};\n\n    // Auto-load model on page init\n    (async function() {\n      console.log(JSON.stringify({ type: \"progress\", status: IS_VISION ? \"Loading vision model...\" : \"Loading model...\" }));\n      \n      try {\n        const result = 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        \n        // Warmup generation to compile shaders and initialize model\n        // Always do text warmup first\n        const textWarmupInputs = result.tokenizer(\"hello\");\n        await result.model.generate({ ...textWarmupInputs, max_new_tokens: 1 });\n        \n        // Vision models also need vision warmup\n        if (result.isVision) {\n          console.log(JSON.stringify({ type: \"progress\", status: \"Warming up vision encoder...\" }));\n          try {\n            // Create a tiny 8x8 red test image\n            const canvas = new OffscreenCanvas(8, 8);\n            const ctx = canvas.getContext('2d');\n            ctx.fillStyle = 'red';\n            ctx.fillRect(0, 0, 8, 8);\n            const blob = await canvas.convertToBlob({ type: 'image/png' });\n            const warmupImage = await RawImage.fromBlob(blob);\n            \n            // Process with vision pipeline\n            const warmupContent = [{ type: \"image\" }, { type: \"text\", text: \"hi\" }];\n            const warmupMessages = [{ role: \"user\", content: warmupContent }];\n            const warmupPrompt = result.processor.apply_chat_template(warmupMessages);\n            const warmupInputs = await result.processor(warmupImage, warmupPrompt, { add_special_tokens: false });\n            \n            // Run vision warmup generation\n            await result.model.generate({\n              ...warmupInputs,\n              max_new_tokens: 1,\n              do_sample: false,\n            });\n          } catch {\n            // Vision warmup failed, text warmup was done so continue\n          }\n        }\n        \n        // Set page title to model ID for cross-process identification\n        document.title = \"Gerbil: \" + ModelPipeline.modelId;\n        \n        console.log(JSON.stringify({ type: \"ready\", isVision: result.isVision }));\n      } catch (error) {\n        console.log(JSON.stringify({ type: \"error\", error: error.message || String(error) }));\n      }\n    })();\n\n    // Text generation (for non-vision models or vision without images)\n    window.gerbilGenerate = async function(messages, options = {}) {\n      const { maxTokens = 256, temperature = 0.7, topP = 0.9, topK = 20, thinking = false, images = [] } = options;\n      \n      const result = await ModelPipeline.getInstance();\n      \n      // Route to vision generation if we have images and this is a vision model\n      if (images.length > 0 && result.isVision) {\n        return window.gerbilGenerateVision(messages, images, options);\n      }\n      \n      // Auto-reset KV cache if it exceeds context length\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 } = result;\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 += 1;\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          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        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        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    // Vision generation (for vision models with images)\n    window.gerbilGenerateVision = async function(messages, imageUrls, options = {}) {\n      const { maxTokens = 2048, temperature = 0.7, topP = 0.9, topK = 20 } = options;\n\n      try {\n        const { processor, tokenizer, model } = await ModelPipeline.getInstance();\n\n        // Build message content with image placeholders for the user prompt\n        const lastMessage = messages[messages.length - 1];\n        const content = [];\n        for (let i = 0; i < imageUrls.length; i += 1) {\n          content.push({ type: \"image\" });\n        }\n        content.push({ type: \"text\", text: lastMessage.content });\n\n        // For vision models, include a brief system instruction for concise responses\n        const visionMessages = [\n          { role: \"system\", content: \"You are a helpful assistant. Be concise and direct in your responses.\" },\n          { role: \"user\", content }\n        ];\n\n        // Apply chat template with generation prompt\n        const chatPrompt = processor.apply_chat_template(visionMessages, { \n          add_generation_prompt: true \n        });\n\n        // Load images\n        console.log(JSON.stringify({ type: \"progress\", status: \"Loading images...\" }));\n        const loadedImages = await Promise.all(\n          imageUrls.map(url => RawImage.fromURL(url))\n        );\n\n        // Process inputs\n        const inputs = await processor(\n          loadedImages.length === 1 ? loadedImages[0] : loadedImages,\n          chatPrompt,\n          { add_special_tokens: false }\n        );\n\n        let startTime = null;\n        let numTokens = 0;\n\n        const streamCallback = (text) => {\n          startTime ??= performance.now();\n          numTokens += 1;\n          const tps = (numTokens / (performance.now() - startTime)) * 1000;\n          console.log(JSON.stringify({ type: \"token\", text, state: \"answering\", numTokens, tps }));\n        };\n\n        const streamer = new TextStreamer(tokenizer, {\n          skip_prompt: true,\n          skip_special_tokens: true,\n          callback_function: streamCallback,\n        });\n\n        console.log(JSON.stringify({ type: \"start\" }));\n\n        const outputs = await model.generate({\n          ...inputs,\n          max_new_tokens: maxTokens,\n          do_sample: temperature > 0,\n          temperature: temperature > 0 ? temperature : undefined,\n          top_p: topP,\n          top_k: topK,\n          streamer,\n          stopping_criteria: stoppingCriteria,\n        });\n\n        // Decode output (skip prompt)\n        const inputLength = inputs.input_ids.dims?.at(-1) || 0;\n        const decoded = processor.batch_decode(\n          outputs.slice(null, [inputLength, null]),\n          { skip_special_tokens: true }\n        );\n\n        const endTime = performance.now();\n        const totalTime = startTime ? endTime - startTime : 0;\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        }));\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 server: Server | 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 isVisionModel: boolean = false; // Whether this is a vision model\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  private constructor(modelId: string, isVision = false) {\n    this.modelId = modelId;\n    this.isVisionModel = isVision;\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\n    // Detect if this is a vision model (can be overridden via options)\n    const isVision = options.isVision ?? ChromeGPUBackend.detectVisionModel(modelId);\n\n    const backend = new ChromeGPUBackend(modelId, isVision);\n    await backend.launch(options);\n    return backend;\n  }\n\n  /**\n   * Detect if a model is a vision model based on its ID\n   */\n  private static detectVisionModel(modelId: string): boolean {\n    const visionModelPatterns = [\n      /ministral/i,\n      /pixtral/i,\n      /llava/i,\n      /vision/i,\n      /vl/i,\n      /image-text/i,\n      /multimodal/i,\n    ];\n    return visionModelPatterns.some((pattern) => pattern.test(modelId));\n  }\n\n  /**\n   * Check if this backend is for a vision model\n   */\n  isVision(): boolean {\n    return this.isVisionModel;\n  }\n\n  /**\n   * Clean up orphan Gerbil pages from previous sessions\n   * These are pages that were left behind when process exited without proper cleanup\n   */\n  private async cleanupOrphanPages(\n    browser: Browser,\n    options: ChromeBackendOptions,\n  ): Promise<number> {\n    try {\n      const pages = await browser.pages();\n      const gerbilPages = pages.filter((p) => {\n        const url = p.url();\n        // Match our worker pages by URL pattern (127.0.0.1:PORT where PORT > 40000)\n        return /127\\.0\\.0\\.1:4\\d{4}/.test(url);\n      });\n\n      // Only clean up if there are orphan pages not tracked by this process\n      const orphanCount = gerbilPages.length - activeBackends.size;\n      if (orphanCount > 0) {\n        options.onProgress?.({ status: `Cleaning up ${orphanCount} orphan page(s)...` });\n\n        for (const page of gerbilPages) {\n          // Check if this page is owned by a current backend\n          let isOwned = false;\n          for (const backend of activeBackends) {\n            if (backend.page === page) {\n              isOwned = true;\n              break;\n            }\n          }\n\n          // Close unowned pages\n          if (!isOwned) {\n            try {\n              await page.close();\n            } catch {\n              // Page may already be closed\n            }\n          }\n        }\n      }\n\n      return orphanCount;\n    } catch {\n      return 0;\n    }\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      // Clean up orphan pages from previous sessions\n      await this.cleanupOrphanPages(globalBrowser, options);\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        // Clean up orphan pages from previous sessions\n        await this.cleanupOrphanPages(globalBrowser, options);\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 contextLength = options.contextLength || 32_768; // Default to 32K if not specified\n    const html = getWorkerPageHTML(this.modelId, contextLength, this.isVisionModel);\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 and register this backend\n    activePagesCount += 1;\n    activeBackends.add(this);\n    options.onProgress?.({ status: `Active pages: ${activePagesCount}/${MAX_CONCURRENT_PAGES}` });\n\n    // Listen for browser disconnect (OOM kill, crash, etc.)\n    this.browser.on(\"disconnected\", () => {\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        }\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    });\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(): { pid: number | null; port: number; modelId: string; startedAt: Date | null } {\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<{ jsHeapUsed: number; jsHeapTotal: number } | 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      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<{ usedGB: number; totalGB: number; usedPercent: number } | 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 ?? (this.isVisionModel ? 2048 : 256),\n      temperature: options.temperature ?? 0.7,\n      topP: options.topP ?? 0.9,\n      topK: options.topK ?? 20,\n      thinking: options.thinking ?? false,\n      images: options.images ?? [], // Pass images for vision models\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   * Updates HTML content for new model loads\n   */\n  private async startServer(html: string): Promise<void> {\n    // Always update the HTML content for new model loads\n    globalServerHtml = html;\n\n    // If global server is already running, reuse it (with updated HTML)\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(globalServerHtml); // Serve current HTML (not captured at creation time)\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   * @param disconnect If true, also disconnect from shared browser (for clean script exit)\n   */\n  async dispose(disconnect = false): 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    // Detach CDP session first\n    if (this.cdp) {\n      try {\n        await this.cdp.detach();\n      } catch {\n        // CDP may already be detached\n      }\n      this.cdp = null;\n    }\n\n    // Close our page (but NOT the shared browser - other backends may use it)\n    if (this.page) {\n      try {\n        // Navigate away first to ensure page is \"dead\"\n        await this.page.goto(\"about:blank\").catch(() => {});\n\n        // Small delay for navigation\n        await new Promise((r) => setTimeout(r, 50));\n\n        // Now close\n        await this.page.close({ runBeforeUnload: false });\n\n        // Decrement active page counter\n        activePagesCount = Math.max(0, activePagesCount - 1);\n      } catch {\n        // Ignore close errors\n      }\n      this.page = null;\n    }\n\n    // Unregister from active backends\n    activeBackends.delete(this);\n\n    // Clear our references first\n    this.browser = null;\n    this.server = null;\n\n    // Small delay to ensure page close completes before disconnect\n    if (disconnect) {\n      await new Promise((r) => setTimeout(r, 100));\n    }\n\n    // If requested and we're the last backend, disconnect from browser\n    // This allows clean script exit without killing browser for other processes\n    if (disconnect && activeBackends.size === 0 && globalBrowser) {\n      try {\n        globalBrowser.disconnect();\n        globalBrowser = null;\n        globalBrowserPromise = null;\n      } catch {\n        // Browser may already be disconnected\n      }\n    }\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   * Get global browser status (even if no active backends)\n   */\n  static getGlobalBrowserStatus(): {\n    running: boolean;\n    pid: number | null;\n    port: number;\n    activePagesCount: number;\n    maxPages: number;\n    wsEndpoint: string | null;\n  } {\n    let pid: number | null = null;\n    let wsEndpoint: string | null = null;\n\n    if (globalBrowser?.connected) {\n      const browserProcess = globalBrowser.process?.();\n      if (browserProcess?.pid) {\n        pid = browserProcess.pid;\n      }\n      wsEndpoint = globalBrowser.wsEndpoint();\n    }\n\n    return {\n      running: globalBrowser?.connected ?? false,\n      pid,\n      port: globalServerPort,\n      activePagesCount, // This process only\n      maxPages: MAX_CONCURRENT_PAGES,\n      wsEndpoint,\n    };\n  }\n\n  /**\n   * Get total page count from Chrome (all processes)\n   */\n  static async getTotalPageCount(): Promise<number> {\n    if (!globalBrowser?.connected) {\n      return 0;\n    }\n\n    try {\n      const pages = await globalBrowser.pages();\n      // Count only Gerbil pages\n      return pages.filter((p) => {\n        const url = p.url();\n        return url.includes(`127.0.0.1:${globalServerPort}`);\n      }).length;\n    } catch {\n      return 0;\n    }\n  }\n\n  /**\n   * Get all active backends with their memory usage (this process only)\n   */\n  static async getAllBackendsInfo(): Promise<\n    Array<{\n      modelId: string;\n      isVision: boolean;\n      isReady: boolean;\n      memory: { usedGB: number; totalGB: number; usedPercent: number } | null;\n    }>\n  > {\n    const results = [];\n\n    for (const backend of activeBackends) {\n      const mem = await backend.getMemoryStats();\n      results.push({\n        modelId: backend.modelId,\n        isVision: backend.isVisionModel,\n        isReady: backend.isReady,\n        memory: mem,\n      });\n    }\n\n    return results;\n  }\n\n  /**\n   * Get ALL pages in Chrome browser (cross-process visibility)\n   * This shows pages from ALL Gerbil processes sharing the browser\n   */\n  static async getAllChromePages(): Promise<\n    Array<{\n      url: string;\n      title: string;\n      isOurs: boolean; // true if this process owns it\n      modelId: string | null; // extracted from URL/title if possible\n      memory: { usedGB: number; totalGB: number } | null;\n    }>\n  > {\n    if (!globalBrowser?.connected) {\n      return [];\n    }\n\n    try {\n      const pages = await globalBrowser.pages();\n\n      const results = [];\n      for (const page of pages) {\n        const url = page.url();\n        const title = await page.title().catch(() => \"\");\n\n        // Skip about:blank and non-gerbil pages\n        if (url === \"about:blank\" || !url.includes(`127.0.0.1:${globalServerPort}`)) {\n          continue;\n        }\n\n        // Check if we own this page\n        let modelId: string | null = null;\n        let isOurs = false;\n        let memory: { usedGB: number; totalGB: number } | null = null;\n\n        for (const backend of activeBackends) {\n          if (backend.page === page) {\n            isOurs = true;\n            modelId = backend.modelId;\n            // Get memory from our backend\n            const mem = await backend.getMemoryStats();\n            if (mem) {\n              memory = { usedGB: mem.usedGB, totalGB: mem.totalGB };\n            }\n            break;\n          }\n        }\n\n        // For pages we don't own, extract model ID from title and try to get memory\n        if (!isOurs) {\n          if (title.startsWith(\"Gerbil: \")) {\n            modelId = title.replace(\"Gerbil: \", \"\");\n          }\n\n          // Try to get memory for other sessions via CDP\n          try {\n            const cdp = await page.createCDPSession();\n            await cdp.send(\"Performance.enable\");\n            const { metrics } = (await cdp.send(\"Performance.getMetrics\")) as {\n              metrics: Array<{ name: string; value: number }>;\n            };\n            const jsHeapUsed = metrics.find((m) => m.name === \"JSHeapUsedSize\")?.value ?? 0;\n            const jsHeapTotal = metrics.find((m) => m.name === \"JSHeapTotalSize\")?.value ?? 0;\n            memory = {\n              usedGB: jsHeapUsed / 1024 ** 3,\n              totalGB: jsHeapTotal / 1024 ** 3,\n            };\n            await cdp.detach();\n          } catch {\n            // Can't get memory for this page\n          }\n        }\n\n        results.push({\n          url,\n          title: title || \"Gerbil WebGPU Backend\",\n          isOurs,\n          modelId,\n          memory,\n        });\n      }\n\n      return results;\n    } catch {\n      return [];\n    }\n  }\n\n  /**\n   * Kill a Chrome page by index (works cross-process)\n   */\n  static async killPageByIndex(index: number): Promise<boolean> {\n    if (!globalBrowser?.connected) {\n      return false;\n    }\n\n    try {\n      const pages = await globalBrowser.pages();\n      // Filter to only Gerbil pages\n      const gerbilPages = pages.filter((p) => {\n        const url = p.url();\n        return url.includes(`127.0.0.1:${globalServerPort}`);\n      });\n\n      if (index < 0 || index >= gerbilPages.length) {\n        return false;\n      }\n\n      const page = gerbilPages[index];\n\n      // If it's our page, use dispose() for proper cleanup\n      for (const backend of activeBackends) {\n        if (backend.page === page) {\n          await backend.dispose();\n          return true;\n        }\n      }\n\n      // Otherwise just close the page directly\n      await page.close();\n      return true;\n    } catch {\n      return false;\n    }\n  }\n\n  /**\n   * Kill a specific backend by index (this process only)\n   */\n  static async killBackendByIndex(index: number): Promise<boolean> {\n    const backends = [...activeBackends];\n    if (index < 0 || index >= backends.length) {\n      return false;\n    }\n\n    const backend = backends[index];\n    try {\n      await backend.dispose();\n      return true;\n    } catch {\n      return false;\n    }\n  }\n\n  /**\n   * Force kill all backends (for zombie cleanup)\n   */\n  static async killAllBackends(): Promise<{ pagesKilled: number; browserKilled: boolean }> {\n    const count = activeBackends.size;\n\n    // Dispose all active backends\n    for (const backend of [...activeBackends]) {\n      try {\n        await backend.dispose();\n      } catch {\n        // Ignore errors during cleanup\n      }\n    }\n    activeBackends.clear();\n\n    // Close shared browser\n    let browserKilled = false;\n    if (globalBrowser) {\n      try {\n        await globalBrowser.close();\n        browserKilled = true;\n      } catch {\n        // Browser may already be closed\n      }\n      globalBrowser = null;\n      globalBrowserPromise = null;\n    }\n\n    // Close server\n    if (globalServer) {\n      globalServer.close();\n      globalServer = null;\n      globalServerPort = 0;\n    }\n\n    activePagesCount = 0;\n\n    // Clean up endpoint file\n    try {\n      unlinkSync(WS_ENDPOINT_FILE);\n    } catch {}\n\n    return { pagesKilled: count, browserKilled };\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;;AAe3E,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,mBAAmB,SAA8C;AAE9E,KAAI;EACF,MAAM,MAAM,MAAM,MAAM,0BAA0B,QAAQ,uBAAuB;AACjF,MAAI,IAAI,IAAI;GACV,MAAM,SAAS,MAAM,IAAI,MAAM;GAE/B,MAAM,aAAa,OAAO,eAAe,EAAE;GAC3C,MAAM,SACJ,OAAO,2BACP,WAAW,2BACX,OAAO,kBACP,WAAW,kBACX,OAAO,eACP,OAAO,uBACP,OAAO,SACP,OAAO;AACT,OAAI,OACF,QAAO;;SAGL;AAKR,KAAI;EACF,MAAM,SAAS,MAAM,MAAM,0BAA0B,QAAQ,iCAAiC;AAC9F,MAAI,OAAO,IAAI;GACb,MAAM,YAAY,MAAM,OAAO,MAAM;AAErC,OAAI,UAAU,oBAAoB,UAAU,mBAAmB,IAC7D,QAAO,UAAU;;SAGf;;;AAQV,SAAS,YAAY,MAA0D;AAE7E,QAAO,KAAK,KAAK,QAAQ,KAAK,QAAQ;;;AAIxC,eAAe,eAAe,SAA8C;AAC1E,KAAI;EAEF,MAAM,UAAU,MAAM,MAAM,qCAAqC,QAAQ,iBAAiB;AAC1F,MAAI,QAAQ,IAAI;GACd,MAAMA,QACJ,MAAM,QAAQ,MAAM;GAGtB,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;AAExC,OAAI,UAAU;IAEZ,MAAM,WAAW,SAAS,KAAK,QAAQ,SAAS,GAAG;IAInD,MAAM,YAHe,MAAM,QACxB,MAAM,EAAE,SAAS,SAAS,QAAQ,EAAE,KAAK,WAAW,GAAG,SAAS,YAAY,CAC9E,CAC8B,QAAQ,KAAK,MAAM,MAAM,YAAY,EAAE,EAAE,EAAE;AAC1E,QAAI,YAAY,EACd,QAAO;;;EAMb,MAAM,MAAM,MAAM,MAAM,qCAAqC,UAAU;AACvE,MAAI,IAAI,GAEN,SADa,MAAM,IAAI,MAAM,EACjB;SAER;;;AAOV,SAAgB,iBACd,SACA,WACA,eACM;AACN,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;AAEvB,OAAI,cACF,UAAS,gBAAgB;QAG3B,QAAO,KAAK;GACV;GACA,cAAc;GACd,UAAU;GACV;GACA;GACD,CAAC;AAGJ,gBAAc,oBAAoB,KAAK,UAAU,EAAE,QAAQ,EAAE,MAAM,EAAE,CAAC;EAGtE,MAAM,YAAY,EAAE,aAAa,UAAU;EAC3C,MAAM,eAAe,EAAE,iBAAiB,UAAU;AAElD,MAAI,aAAa,aACf,SAAQ,IAAI,CACV,YAAY,eAAe,QAAQ,GAAG,QAAQ,QAAQ,OAAU,EAChE,eAAe,mBAAmB,QAAQ,GAAG,QAAQ,QAAQ,OAAU,CACxE,CAAC,CACC,MAAM,CAAC,MAAM,aAAa;GACzB,MAAM,gBAAgB,uBAAuB;GAC7C,MAAM,QAAQ,cAAc,MAAM,MAAM,EAAE,YAAY,QAAQ;AAC9D,OAAI,OAAO;AACT,QAAI,KACF,OAAM,YAAY;AAEpB,QAAI,QACF,OAAM,gBAAgB;AAExB,kBAAc,oBAAoB,KAAK,UAAU,EAAE,QAAQ,eAAe,EAAE,MAAM,EAAE,CAAC;;IAEvF,CACD,YAAY,GAAG;SAEd;;;AAgBV,eAAsB,0BAAyC;AAC7D,KAAI;EACF,MAAM,SAAS,uBAAuB;EAEtC,MAAM,oBAAoB;EAC1B,MAAM,eAAe,OAAO,QACzB,MAAM,CAAC,EAAE,aAAa,EAAE,YAAY,qBAAqB,CAAC,EAAE,cAC9D;AACD,MAAI,aAAa,WAAW,EAC1B;EAIF,MAAM,YAAY;AAClB,OAAK,IAAI,IAAI,GAAG,IAAI,aAAa,QAAQ,KAAK,WAAW;GACvD,MAAM,QAAQ,aAAa,MAAM,GAAG,IAAI,UAAU;AAClD,SAAM,QAAQ,IACZ,MAAM,IAAI,OAAO,UAAU;IACzB,MAAM,CAAC,MAAM,WAAW,MAAM,QAAQ,IAAI,CACxC,CAAC,MAAM,aAAa,MAAM,YAAY,oBAClC,eAAe,MAAM,QAAQ,GAC7B,QAAQ,QAAQ,OAAU,EAC9B,MAAM,gBAAgB,QAAQ,QAAQ,OAAU,GAAG,mBAAmB,MAAM,QAAQ,CACrF,CAAC;AACF,QAAI,KACF,OAAM,YAAY;AAEpB,QAAI,QACF,OAAM,gBAAgB;KAExB,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;AACvB,IAAI,mBAAmB;AAGvB,IAAI,mBAAmB;AACvB,MAAM,uBAAuB;AAG7B,MAAMC,iCAAwC,IAAI,KAAK;AAqCvD,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,WAAmB,gBAAgB,OAAQ,WAAW,OAAe;AAC9F,QAAO;;;;;;;;;;;;;;;;;;;;;wBAqBe,SAAS;;;;;;0BAMP,UAAU;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;6BAmDP,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8S3C,IAAa,mBAAb,MAAa,iBAAiB;CAC5B,AAAQ,UAA0B;CAClC,AAAQ,OAAoB;CAC5B,AAAQ,MAAyB;CACjC,AAAQ,SAAwB;CAChC,AAAQ,aAAa;CACrB,AAAQ,cAAsB;CAC9B,AAAiB;CACjB,AAAQ,UAAU;CAClB,AAAiB,gBAAyB;CAC1C,AAAiB,kCAAoD,IAAI,KAAK;CAC9E,AAAQ,iBAA8C,EAAE;CAExD,AAAQ,YAAY,SAAiB,WAAW,OAAO;AACrD,OAAK,UAAU;AACf,OAAK,gBAAgB;;;;;CAMvB,aAAa,OAAO,UAAgC,EAAE,EAA6B;EACjF,MAAM,UAAU,QAAQ,WAAW;EAKnC,MAAM,UAAU,IAAI,iBAAiB,SAFpB,QAAQ,YAAY,iBAAiB,kBAAkB,QAAQ,CAEzB;AACvD,QAAM,QAAQ,OAAO,QAAQ;AAC7B,SAAO;;;;;CAMT,OAAe,kBAAkB,SAA0B;AAUzD,SAT4B;GAC1B;GACA;GACA;GACA;GACA;GACA;GACA;GACD,CAC0B,MAAM,YAAY,QAAQ,KAAK,QAAQ,CAAC;;;;;CAMrE,WAAoB;AAClB,SAAO,KAAK;;;;;;CAOd,MAAc,mBACZ,SACA,SACiB;AACjB,MAAI;GAEF,MAAM,eADQ,MAAM,QAAQ,OAAO,EACT,QAAQ,MAAM;IACtC,MAAM,MAAM,EAAE,KAAK;AAEnB,WAAO,sBAAsB,KAAK,IAAI;KACtC;GAGF,MAAM,cAAc,YAAY,SAAS,eAAe;AACxD,OAAI,cAAc,GAAG;AACnB,YAAQ,aAAa,EAAE,QAAQ,eAAe,YAAY,qBAAqB,CAAC;AAEhF,SAAK,MAAM,QAAQ,aAAa;KAE9B,IAAI,UAAU;AACd,UAAK,MAAM,WAAW,eACpB,KAAI,QAAQ,SAAS,MAAM;AACzB,gBAAU;AACV;;AAKJ,SAAI,CAAC,QACH,KAAI;AACF,YAAM,KAAK,OAAO;aACZ;;;AAOd,UAAO;UACD;AACN,UAAO;;;;;;;CAQX,MAAc,mBACZ,YACA,SACkB;AAElB,MAAI,eAAe,WAAW;AAC5B,WAAQ,aAAa,EAAE,QAAQ,8BAA8B,CAAC;AAE9D,SAAM,KAAK,mBAAmB,eAAe,QAAQ;AACrD,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;AAEF,SAAM,KAAK,mBAAmB,eAAe,QAAQ;AACrD,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,gBAAgB,QAAQ,iBAAiB;EAC/C,MAAM,OAAO,kBAAkB,KAAK,SAAS,eAAe,KAAK,cAAc;AAC/E,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,sBAAoB;AACpB,iBAAe,IAAI,KAAK;AACxB,UAAQ,aAAa,EAAE,QAAQ,iBAAiB,iBAAiB,GAAG,wBAAwB,CAAC;AAG7F,OAAK,QAAQ,GAAG,sBAAsB;AACpC,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,KACd;;IAGJ;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;IAEF;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,YAA2F;EAEzF,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,iBAA8E;AAClF,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;AAKT,MAFe,IAAI,aAAa,QAAQ,IAE3B,aAAa;AACxB,SAAM,KAAK,OAAO;AAClB,UAAO;;AAGT,SAAO;;;;;CAMT,MAAM,iBAA2F;EAC/F,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,cAAc,KAAK,gBAAgB,OAAO;GAC7D,aAAa,QAAQ,eAAe;GACpC,MAAM,QAAQ,QAAQ;GACtB,MAAM,QAAQ,QAAQ;GACtB,UAAU,QAAQ,YAAY;GAC9B,QAAQ,QAAQ,UAAU,EAAE;GAC7B;AAGD,MAAI,QAAQ,QACV,MAAK,gBAAgB,IAAI,SAAS,QAAQ,QAAQ;AAGpD,MAAI;GAEF,MAAM,gBAAgB,KAAK,MAAM,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;;;;;;;CAQd,MAAc,YAAY,MAA6B;AAErD,qBAAmB;AAGnB,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,iBAAiB;KACzB;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;;;;;;;CAQJ,MAAM,QAAQ,aAAa,OAAsB;AAE/C,OAAK,UAAU;AAGf,OAAK,iBAAiB,EAAE;AACxB,OAAK,gBAAgB,OAAO;AAG5B,MAAI,KAAK,KAAK;AACZ,OAAI;AACF,UAAM,KAAK,IAAI,QAAQ;WACjB;AAGR,QAAK,MAAM;;AAIb,MAAI,KAAK,MAAM;AACb,OAAI;AAEF,UAAM,KAAK,KAAK,KAAK,cAAc,CAAC,YAAY,GAAG;AAGnD,UAAM,IAAI,SAAS,MAAM,WAAW,GAAG,GAAG,CAAC;AAG3C,UAAM,KAAK,KAAK,MAAM,EAAE,iBAAiB,OAAO,CAAC;AAGjD,uBAAmB,KAAK,IAAI,GAAG,mBAAmB,EAAE;WAC9C;AAGR,QAAK,OAAO;;AAId,iBAAe,OAAO,KAAK;AAG3B,OAAK,UAAU;AACf,OAAK,SAAS;AAGd,MAAI,WACF,OAAM,IAAI,SAAS,MAAM,WAAW,GAAG,IAAI,CAAC;AAK9C,MAAI,cAAc,eAAe,SAAS,KAAK,cAC7C,KAAI;AACF,iBAAc,YAAY;AAC1B,mBAAgB;AAChB,0BAAuB;UACjB;;;;;CASZ,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,OAAO,yBAOL;EACA,IAAID,MAAqB;EACzB,IAAIE,aAA4B;AAEhC,MAAI,eAAe,WAAW;GAC5B,MAAM,iBAAiB,cAAc,WAAW;AAChD,OAAI,gBAAgB,IAClB,OAAM,eAAe;AAEvB,gBAAa,cAAc,YAAY;;AAGzC,SAAO;GACL,SAAS,eAAe,aAAa;GACrC;GACA,MAAM;GACN;GACA,UAAU;GACV;GACD;;;;;CAMH,aAAa,oBAAqC;AAChD,MAAI,CAAC,eAAe,UAClB,QAAO;AAGT,MAAI;AAGF,WAFc,MAAM,cAAc,OAAO,EAE5B,QAAQ,MAAM;AAEzB,WADY,EAAE,KAAK,CACR,SAAS,aAAa,mBAAmB;KACpD,CAAC;UACG;AACN,UAAO;;;;;;CAOX,aAAa,qBAOX;EACA,MAAM,UAAU,EAAE;AAElB,OAAK,MAAM,WAAW,gBAAgB;GACpC,MAAM,MAAM,MAAM,QAAQ,gBAAgB;AAC1C,WAAQ,KAAK;IACX,SAAS,QAAQ;IACjB,UAAU,QAAQ;IAClB,SAAS,QAAQ;IACjB,QAAQ;IACT,CAAC;;AAGJ,SAAO;;;;;;CAOT,aAAa,oBAQX;AACA,MAAI,CAAC,eAAe,UAClB,QAAO,EAAE;AAGX,MAAI;GACF,MAAM,QAAQ,MAAM,cAAc,OAAO;GAEzC,MAAM,UAAU,EAAE;AAClB,QAAK,MAAM,QAAQ,OAAO;IACxB,MAAM,MAAM,KAAK,KAAK;IACtB,MAAM,QAAQ,MAAM,KAAK,OAAO,CAAC,YAAY,GAAG;AAGhD,QAAI,QAAQ,iBAAiB,CAAC,IAAI,SAAS,aAAa,mBAAmB,CACzE;IAIF,IAAIC,UAAyB;IAC7B,IAAI,SAAS;IACb,IAAIC,SAAqD;AAEzD,SAAK,MAAM,WAAW,eACpB,KAAI,QAAQ,SAAS,MAAM;AACzB,cAAS;AACT,eAAU,QAAQ;KAElB,MAAM,MAAM,MAAM,QAAQ,gBAAgB;AAC1C,SAAI,IACF,UAAS;MAAE,QAAQ,IAAI;MAAQ,SAAS,IAAI;MAAS;AAEvD;;AAKJ,QAAI,CAAC,QAAQ;AACX,SAAI,MAAM,WAAW,WAAW,CAC9B,WAAU,MAAM,QAAQ,YAAY,GAAG;AAIzC,SAAI;MACF,MAAM,MAAM,MAAM,KAAK,kBAAkB;AACzC,YAAM,IAAI,KAAK,qBAAqB;MACpC,MAAM,EAAE,YAAa,MAAM,IAAI,KAAK,yBAAyB;MAG7D,MAAM,aAAa,QAAQ,MAAM,MAAM,EAAE,SAAS,iBAAiB,EAAE,SAAS;MAC9E,MAAM,cAAc,QAAQ,MAAM,MAAM,EAAE,SAAS,kBAAkB,EAAE,SAAS;AAChF,eAAS;OACP,QAAQ,aAAa,QAAQ;OAC7B,SAAS,cAAc,QAAQ;OAChC;AACD,YAAM,IAAI,QAAQ;aACZ;;AAKV,YAAQ,KAAK;KACX;KACA,OAAO,SAAS;KAChB;KACA;KACA;KACD,CAAC;;AAGJ,UAAO;UACD;AACN,UAAO,EAAE;;;;;;CAOb,aAAa,gBAAgB,OAAiC;AAC5D,MAAI,CAAC,eAAe,UAClB,QAAO;AAGT,MAAI;GAGF,MAAM,eAFQ,MAAM,cAAc,OAAO,EAEf,QAAQ,MAAM;AAEtC,WADY,EAAE,KAAK,CACR,SAAS,aAAa,mBAAmB;KACpD;AAEF,OAAI,QAAQ,KAAK,SAAS,YAAY,OACpC,QAAO;GAGT,MAAM,OAAO,YAAY;AAGzB,QAAK,MAAM,WAAW,eACpB,KAAI,QAAQ,SAAS,MAAM;AACzB,UAAM,QAAQ,SAAS;AACvB,WAAO;;AAKX,SAAM,KAAK,OAAO;AAClB,UAAO;UACD;AACN,UAAO;;;;;;CAOX,aAAa,mBAAmB,OAAiC;EAC/D,MAAM,WAAW,CAAC,GAAG,eAAe;AACpC,MAAI,QAAQ,KAAK,SAAS,SAAS,OACjC,QAAO;EAGT,MAAM,UAAU,SAAS;AACzB,MAAI;AACF,SAAM,QAAQ,SAAS;AACvB,UAAO;UACD;AACN,UAAO;;;;;;CAOX,aAAa,kBAA4E;EACvF,MAAM,QAAQ,eAAe;AAG7B,OAAK,MAAM,WAAW,CAAC,GAAG,eAAe,CACvC,KAAI;AACF,SAAM,QAAQ,SAAS;UACjB;AAIV,iBAAe,OAAO;EAGtB,IAAI,gBAAgB;AACpB,MAAI,eAAe;AACjB,OAAI;AACF,UAAM,cAAc,OAAO;AAC3B,oBAAgB;WACV;AAGR,mBAAgB;AAChB,0BAAuB;;AAIzB,MAAI,cAAc;AAChB,gBAAa,OAAO;AACpB,kBAAe;AACf,sBAAmB;;AAGrB,qBAAmB;AAGnB,MAAI;AACF,cAAW,iBAAiB;UACtB;AAER,SAAO;GAAE,aAAa;GAAO;GAAe;;;;;CAM9C,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"}

package/dist/{chrome-backend-JEPeM2YE.mjs → chrome-backend-DIKYoWj-.mjs}

@@ -1,3 +1,3 @@
-import { i as trackCachedModel, n as getChromeCachedModels, r as refreshCachedModelSizes, t as ChromeGPUBackend } from "./chrome-backend-Y9F7W5VQ.mjs";
+import { i as trackCachedModel, n as getChromeCachedModels, r as refreshCachedModelSizes, t as ChromeGPUBackend } from "./chrome-backend-CORwaIyC.mjs";
 
 export { ChromeGPUBackend };

package/dist/cli.mjs

@@ -1,13 +1,12 @@
 #!/usr/bin/env node
-import { t as __require } from "./chunk-Ct1HF2bE.mjs";
-import { t as Gerbil } from "./gerbil-yoSpRHgv.mjs";
-import { n as getChromeCachedModels, r as refreshCachedModelSizes } from "./chrome-backend-Y9F7W5VQ.mjs";
-import { t as BUILTIN_MODELS } from "./models-BAtL8qsA.mjs";
-import "./utils-CkB4Roi6.mjs";
-import "./one-liner-B1rmFto6.mjs";
-import { D as listSkills, T as getSkillInfo, a as summarize, d as explain, k as useSkill, m as commit, s as review, v as loadProjectSkills } from "./skills-5DxAV-rn.mjs";
-import { r as startMCPServer } from "./mcp-Bitg4sjX.mjs";
-import { a as getToolDefinitions, c as setToolContext, i as getTool, n as executeToolCall, o as loadProjectTools, r as formatToolsForPrompt, s as parseToolCall } from "./tools-IYPrqoek.mjs";
+import { t as __require } from "./chunk-CkXuGtQK.mjs";
+import { n as BUILTIN_MODELS, t as Gerbil } from "./gerbil-Dq039a6V.mjs";
+import { n as getChromeCachedModels, r as refreshCachedModelSizes } from "./chrome-backend-CORwaIyC.mjs";
+import "./utils-CZBZ8dgR.mjs";
+import "./one-liner-CgRVfe5K.mjs";
+import { D as listSkills, T as getSkillInfo, a as summarize, d as explain, k as useSkill, m as commit, s as review, v as loadProjectSkills } from "./skills-BGS20rGK.mjs";
+import { r as startMCPServer } from "./mcp-DY57Whwj.mjs";
+import { a as getToolDefinitions, c as setToolContext, i as getTool, n as executeToolCall, o as loadProjectTools, r as formatToolsForPrompt, s as parseToolCall } from "./tools-Bi1P7Xoy.mjs";
 import { exec, spawn, spawnSync } from "node:child_process";
 import fs, { existsSync, readFileSync, unlinkSync } from "node:fs";
 import http from "node:http";
@@ -25,7 +24,7 @@ import SelectInput from "ink-select-input";
 import TextInput from "ink-text-input";
 
 //#region package.json
-var version = "1.0.0-rc.1";
+var version = "1.0.0-rc.3";
 
 //#endregion
 //#region src/cli/repl/auto-update.ts
@@ -8182,7 +8181,7 @@ function App({ initialView = "menu", forcedDevice } = {}) {
 		return () => {
 			mounted = false;
 			if (gerbilRef.current) {
-				import("./repl-D20JO260.mjs").then(({ setCleanupPromise: setCleanupPromise$1 }) => {
+				import("./repl-BEusmMZs.mjs").then(({ setCleanupPromise: setCleanupPromise$1 }) => {
 					setCleanupPromise$1(gerbilRef.current?.dispose(true) ?? Promise.resolve());
 				});
 				gerbilRef.current = null;
@@ -9509,7 +9508,7 @@ program.command("serve").description("Start Gerbil server (use --mcp or --http f
 	}
 	if (opts.mcp) await startMCPServer({ model: opts.model });
 	else {
-		const { Gerbil: Gerbil$1 } = await import("./gerbil-DeQlX_Mt.mjs");
+		const { Gerbil: Gerbil$1 } = await import("./gerbil-DyTEWXLy.mjs");
 		const g = new Gerbil$1();
 		const spinner = ora("Loading model...").start();
 		await g.loadModel(opts.model);
@@ -9653,7 +9652,7 @@ program.command("cache").description("Show or manage model cache").option("--cle
 });
 const formatSize = formatBytes;
 program.command("cleanup").description("Clean up zombie Chrome pages and free memory").option("--kill-browser", "Also kill the shared Chrome browser (forces restart on next use)").option("--force", "Force kill all Gerbil Chrome processes (use if --kill-browser fails)").action(async (opts) => {
-	const { ChromeGPUBackend } = await import("./chrome-backend-JEPeM2YE.mjs");
+	const { ChromeGPUBackend } = await import("./chrome-backend-DIKYoWj-.mjs");
 	const { execSync: execSync$1 } = await import("node:child_process");
 	console.log(chalk.cyan("\nChecking Chrome backend status...\n"));
 	let orphanPids = [];
@@ -9753,7 +9752,7 @@ program.command("bench").description("Benchmark model performance").option("-m,
 	}
 });
 program.command("update").description("Update Gerbil to the latest version").action(async () => {
-	const { checkForUpdate: checkForUpdate$1, installUpdate: installUpdate$1, CURRENT_VERSION: CURRENT_VERSION$1 } = await import("./auto-update-DsWBBnEk.mjs");
+	const { checkForUpdate: checkForUpdate$1, installUpdate: installUpdate$1, CURRENT_VERSION: CURRENT_VERSION$1 } = await import("./auto-update-S9s5-g0C.mjs");
 	const spinner = ora("Checking for updates...").start();
 	try {
 		const check = await checkForUpdate$1();
@@ -9781,7 +9780,7 @@ program.parse = (...args) => {
 	return result;
 };
 async function checkForUpdateCLI() {
-	const { checkForUpdate: checkForUpdate$1, CURRENT_VERSION: CURRENT_VERSION$1 } = await import("./auto-update-DsWBBnEk.mjs");
+	const { checkForUpdate: checkForUpdate$1, CURRENT_VERSION: CURRENT_VERSION$1 } = await import("./auto-update-S9s5-g0C.mjs");
 	if ((await checkForUpdate$1()).updateAvailable) {}
 }
 program.parse();