@tryhamster/gerbil 1.0.0-rc.0 → 1.0.0-rc.10

package/dist/gerbil-DoDGHe6Z.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"gerbil-DoDGHe6Z.mjs","names":["globalCache: ResponseCache | null","BUILTIN_MODELS: Record<string, ModelConfig>","FAMILY_CONTEXT_DEFAULTS: Record<string, number>","family: ModelConfig[\"family\"]","pipeline","rawPipeline","KOKORO_VOICES_DEFAULT: VoiceInfo[]","isBrowser","tfDevice: \"webgpu\" | \"wasm\" | \"cpu\"","chromeErr: any","result","result: GenerateResult","tokenQueue: string[]","resolveNext: ((value: string | null) => void) | null","content: Array<{ type: string; text?: string }>","text","results: EmbedResult[]","messages: Array<{ role: string; content: string }>"],"sources":["../src/core/cache.ts","../src/core/models.ts","../src/core/gerbil.ts"],"sourcesContent":["/**\n * Response Cache for Gerbil\n *\n * LRU cache with TTL expiration for caching inference responses.\n * Enables instant responses for repeated prompts.\n */\n\nimport type { GenerateResult } from \"./types.js\";\n\n// ============================================\n// Types\n// ============================================\n\ntype CacheEntry = {\n  result: GenerateResult;\n  createdAt: number;\n  ttl: number;\n};\n\ntype CacheStats = {\n  hits: number;\n  misses: number;\n  size: number;\n  maxSize: number;\n};\n\n// ============================================\n// Cache Key Generation\n// ============================================\n\n/**\n * Generate a deterministic cache key from prompt and options.\n * Key includes all parameters that affect the output.\n */\nexport function generateCacheKey(\n  prompt: string,\n  modelId: string,\n  options: {\n    maxTokens?: number;\n    temperature?: number;\n    topP?: number;\n    topK?: number;\n    system?: string;\n    thinking?: boolean;\n  },\n): string {\n  const keyParts = [\n    prompt,\n    modelId,\n    options.maxTokens ?? 256,\n    options.temperature ?? 0.7,\n    options.topP ?? 0.9,\n    options.topK ?? 50,\n    options.system ?? \"\",\n    options.thinking ?? false,\n  ];\n\n  // Simple hash function for cache key\n  const str = JSON.stringify(keyParts);\n  let hash = 0;\n  for (let i = 0; i < str.length; i++) {\n    const char = str.charCodeAt(i);\n    hash = (hash << 5) - hash + char;\n    hash = hash & hash; // Convert to 32bit integer\n  }\n  return `gerbil:${hash.toString(16)}`;\n}\n\n// ============================================\n// Response Cache\n// ============================================\n\n/**\n * LRU cache with TTL expiration for inference responses.\n */\nexport class ResponseCache {\n  private cache: Map<string, CacheEntry> = new Map();\n  private maxSize: number;\n  private defaultTtl: number;\n  private hits = 0;\n  private misses = 0;\n\n  /**\n   * Create a new response cache.\n   * @param maxSize Maximum number of entries (default: 100)\n   * @param defaultTtl Default TTL in ms (default: 5 minutes)\n   */\n  constructor(maxSize = 100, defaultTtl = 5 * 60 * 1000) {\n    this.maxSize = maxSize;\n    this.defaultTtl = defaultTtl;\n  }\n\n  /**\n   * Get a cached response if it exists and hasn't expired.\n   */\n  get(key: string): GenerateResult | null {\n    const entry = this.cache.get(key);\n\n    if (!entry) {\n      this.misses++;\n      return null;\n    }\n\n    // Check if expired\n    if (Date.now() - entry.createdAt > entry.ttl) {\n      this.cache.delete(key);\n      this.misses++;\n      return null;\n    }\n\n    // Move to end for LRU (delete and re-add)\n    this.cache.delete(key);\n    this.cache.set(key, entry);\n\n    this.hits++;\n    return { ...entry.result, cached: true };\n  }\n\n  /**\n   * Store a response in the cache.\n   */\n  set(key: string, result: GenerateResult, ttl?: number): void {\n    // Evict oldest entries if at capacity\n    while (this.cache.size >= this.maxSize) {\n      const firstKey = this.cache.keys().next().value;\n      if (firstKey) {\n        this.cache.delete(firstKey);\n      }\n    }\n\n    this.cache.set(key, {\n      result,\n      createdAt: Date.now(),\n      ttl: ttl ?? this.defaultTtl,\n    });\n  }\n\n  /**\n   * Check if a key exists and is not expired.\n   */\n  has(key: string): boolean {\n    const entry = this.cache.get(key);\n    if (!entry) return false;\n\n    if (Date.now() - entry.createdAt > entry.ttl) {\n      this.cache.delete(key);\n      return false;\n    }\n\n    return true;\n  }\n\n  /**\n   * Remove a specific key from the cache.\n   */\n  delete(key: string): boolean {\n    return this.cache.delete(key);\n  }\n\n  /**\n   * Clear all entries from the cache.\n   */\n  clear(): void {\n    this.cache.clear();\n    this.hits = 0;\n    this.misses = 0;\n  }\n\n  /**\n   * Remove all expired entries.\n   */\n  prune(): number {\n    const now = Date.now();\n    let pruned = 0;\n\n    for (const [key, entry] of this.cache) {\n      if (now - entry.createdAt > entry.ttl) {\n        this.cache.delete(key);\n        pruned++;\n      }\n    }\n\n    return pruned;\n  }\n\n  /**\n   * Get cache statistics.\n   */\n  getStats(): CacheStats {\n    return {\n      hits: this.hits,\n      misses: this.misses,\n      size: this.cache.size,\n      maxSize: this.maxSize,\n    };\n  }\n\n  /**\n   * Get hit rate as a percentage.\n   */\n  getHitRate(): number {\n    const total = this.hits + this.misses;\n    if (total === 0) return 0;\n    return (this.hits / total) * 100;\n  }\n}\n\n// ============================================\n// Global Cache Instance\n// ============================================\n\nlet globalCache: ResponseCache | null = null;\n\n/**\n * Get the global response cache instance.\n * Creates one if it doesn't exist.\n */\nexport function getGlobalCache(): ResponseCache {\n  if (!globalCache) {\n    globalCache = new ResponseCache();\n  }\n  return globalCache;\n}\n\n/**\n * Configure the global cache with custom settings.\n */\nexport function configureGlobalCache(maxSize?: number, defaultTtl?: number): ResponseCache {\n  globalCache = new ResponseCache(maxSize, defaultTtl);\n  return globalCache;\n}\n\n/**\n * Clear and reset the global cache.\n */\nexport function clearGlobalCache(): void {\n  if (globalCache) {\n    globalCache.clear();\n  }\n}\n","/**\n * Model Registry\n *\n * Supports built-in models and any HuggingFace model via hf:org/model syntax\n */\n\nimport type { ModelConfig, ModelSource } from \"./types.js\";\n\n// ============================================\n// Built-in Models (curated & tested)\n// ============================================\n\nexport const BUILTIN_MODELS: Record<string, ModelConfig> = {\n  \"qwen3-0.6b\": {\n    id: \"qwen3-0.6b\",\n    repo: \"onnx-community/Qwen3-0.6B-ONNX\",\n    description: \"Qwen3 0.6B - Best balance of speed and quality, supports thinking\",\n    size: \"~400MB\",\n    contextLength: 32_768,\n    supportsThinking: true,\n    supportsJson: true,\n    family: \"qwen\",\n  },\n  \"qwen2.5-0.5b\": {\n    id: \"qwen2.5-0.5b\",\n    repo: \"onnx-community/Qwen2.5-0.5B-Instruct\",\n    description: \"Qwen2.5 0.5B - Fast and capable\",\n    size: \"~350MB\",\n    contextLength: 32_768,\n    supportsThinking: false,\n    supportsJson: true,\n    family: \"qwen\",\n  },\n  \"qwen2.5-coder-0.5b\": {\n    id: \"qwen2.5-coder-0.5b\",\n    repo: \"onnx-community/Qwen2.5-Coder-0.5B-Instruct\",\n    description: \"Qwen2.5 Coder 0.5B - Optimized for code\",\n    size: \"~400MB\",\n    contextLength: 32_768,\n    supportsThinking: false,\n    supportsJson: true,\n    family: \"qwen\",\n  },\n  \"smollm2-360m\": {\n    id: \"smollm2-360m\",\n    repo: \"HuggingFaceTB/SmolLM2-360M-Instruct\",\n    description: \"SmolLM2 360M - Fast, good for simple tasks\",\n    size: \"~250MB\",\n    contextLength: 8192,\n    supportsThinking: false,\n    supportsJson: false,\n    family: \"smollm\",\n  },\n  \"smollm2-135m\": {\n    id: \"smollm2-135m\",\n    repo: \"HuggingFaceTB/SmolLM2-135M-Instruct\",\n    description: \"SmolLM2 135M - Fastest, basic generation\",\n    size: \"~100MB\",\n    contextLength: 8192,\n    supportsThinking: false,\n    supportsJson: false,\n    family: \"smollm\",\n  },\n  \"phi-3-mini\": {\n    id: \"phi-3-mini\",\n    repo: \"microsoft/Phi-3-mini-4k-instruct-onnx\",\n    description: \"Phi-3 Mini - High quality, larger model\",\n    size: \"~2.1GB\",\n    contextLength: 4096,\n    supportsThinking: false,\n    supportsJson: true,\n    family: \"phi\",\n  },\n  \"ministral-3b\": {\n    id: \"ministral-3b\",\n    repo: \"mistralai/Ministral-3-3B-Instruct-2512-ONNX\",\n    description: \"Ministral 3 3B - Vision + Reasoning, 256k context\",\n    size: \"~2.5GB\",\n    contextLength: 262_144,\n    supportsThinking: true,\n    supportsJson: true,\n    supportsVision: true,\n    visionEncoderSize: \"0.4B\",\n    family: \"mistral\",\n  },\n};\n\n// ============================================\n// Model Resolution\n// ============================================\n\n/**\n * Parse model identifier and resolve to source\n *\n * Supported formats:\n * - \"qwen3-0.6b\" (built-in)\n * - \"hf:org/model\" (HuggingFace shorthand)\n * - \"https://huggingface.co/org/model\" (full URL)\n * - \"file:./path/to/model\" (local path)\n */\nexport function resolveModel(modelId: string): ModelSource {\n  // Built-in model\n  if (BUILTIN_MODELS[modelId]) {\n    return {\n      type: \"builtin\",\n      path: BUILTIN_MODELS[modelId].repo,\n    };\n  }\n\n  // HuggingFace shorthand: hf:org/model\n  if (modelId.startsWith(\"hf:\")) {\n    const repo = modelId.slice(3);\n    return {\n      type: \"huggingface\",\n      path: repo,\n    };\n  }\n\n  // HuggingFace URL\n  if (modelId.startsWith(\"https://huggingface.co/\")) {\n    const repo = modelId.replace(\"https://huggingface.co/\", \"\");\n    return {\n      type: \"huggingface\",\n      path: repo,\n    };\n  }\n\n  // Local file\n  if (modelId.startsWith(\"file:\")) {\n    const path = modelId.slice(5);\n    return {\n      type: \"local\",\n      path,\n    };\n  }\n\n  // Assume it's a HuggingFace repo if it contains a slash\n  if (modelId.includes(\"/\")) {\n    return {\n      type: \"huggingface\",\n      path: modelId,\n    };\n  }\n\n  // Unknown - treat as HuggingFace\n  return {\n    type: \"huggingface\",\n    path: modelId,\n  };\n}\n\n/**\n * Get model config (built-in only)\n */\nexport function getModelConfig(modelId: string): ModelConfig | null {\n  return BUILTIN_MODELS[modelId] || null;\n}\n\n// Default context lengths by model family (when config.json is unavailable)\nconst FAMILY_CONTEXT_DEFAULTS: Record<string, number> = {\n  qwen: 32_768,\n  mistral: 262_144, // Ministral models support up to 256K\n  llama: 8192,\n  phi: 4096,\n  smollm: 8192,\n  other: 4096,\n};\n\n/**\n * Create model config for external model\n */\nexport function createExternalModelConfig(\n  modelId: string,\n  repo: string,\n  contextLength?: number,\n): ModelConfig {\n  // Try to infer family from repo name\n  let family: ModelConfig[\"family\"] = \"other\";\n  const repoLower = repo.toLowerCase();\n\n  if (repoLower.includes(\"qwen\")) {\n    family = \"qwen\";\n  } else if (repoLower.includes(\"smollm\")) {\n    family = \"smollm\";\n  } else if (repoLower.includes(\"phi\")) {\n    family = \"phi\";\n  } else if (repoLower.includes(\"mistral\") || repoLower.includes(\"ministral\")) {\n    family = \"mistral\";\n  } else if (repoLower.includes(\"llama\")) {\n    family = \"llama\";\n  }\n\n  // Detect vision models from common patterns\n  const supportsVision =\n    repoLower.includes(\"vision\") ||\n    repoLower.includes(\"vlm\") ||\n    repoLower.includes(\"image-text\") ||\n    repoLower.includes(\"ministral\");\n\n  return {\n    id: modelId,\n    repo,\n    description: `External model: ${repo}`,\n    size: \"Unknown\",\n    contextLength: contextLength || FAMILY_CONTEXT_DEFAULTS[family] || 4096,\n    supportsThinking: family === \"qwen\" || family === \"mistral\",\n    supportsJson: family === \"qwen\" || family === \"phi\" || family === \"mistral\",\n    supportsVision,\n    family,\n  };\n}\n\n/**\n * Fetch context length from HuggingFace model config\n */\nexport async function fetchModelContextLength(repo: string): Promise<number | null> {\n  try {\n    const res = await fetch(`https://huggingface.co/${repo}/raw/main/config.json`);\n    if (!res.ok) {\n      return null;\n    }\n\n    const config = await res.json();\n\n    // Different models use different field names\n    return (\n      config.max_position_embeddings ||\n      config.n_positions ||\n      config.max_seq_len ||\n      config.sliding_window || // Some models use this\n      config.context_length ||\n      null\n    );\n  } catch {\n    return null;\n  }\n}\n\n/**\n * List all built-in models\n */\nexport function listBuiltinModels(): ModelConfig[] {\n  return Object.values(BUILTIN_MODELS);\n}\n\n/**\n * Search HuggingFace models (placeholder - would need HF API)\n */\nexport async function searchModels(query: string): Promise<ModelConfig[]> {\n  // TODO: Implement HuggingFace API search\n  // For now, filter built-in models\n  const q = query.toLowerCase();\n  return listBuiltinModels().filter(\n    (m) =>\n      m.id.toLowerCase().includes(q) ||\n      m.description.toLowerCase().includes(q) ||\n      m.family.toLowerCase().includes(q),\n  );\n}\n","/**\n * Gerbil - Local GPU-accelerated LLM inference\n */\n\nimport {\n  AutoModelForCausalLM,\n  AutoModelForImageTextToText,\n  AutoProcessor,\n  AutoTokenizer,\n  env,\n  type FeatureExtractionPipeline,\n  type PreTrainedTokenizer,\n  RawImage,\n  pipeline as rawPipeline,\n  type TextGenerationPipeline,\n  TextStreamer,\n} from \"@huggingface/transformers\";\n\n// Wrapper to avoid TypeScript complexity issues with transformers.js types\nconst pipeline = rawPipeline as (task: string, model: string, options?: any) => Promise<any>;\n\n// Suppress noisy transformers.js warnings during model loading\nfunction suppressNoisyWarnings<T>(fn: () => Promise<T>): Promise<T> {\n  const originalWarn = console.warn;\n  console.warn = (...args: any[]) => {\n    const msg = args[0]?.toString?.() || \"\";\n    // Suppress \"Unable to determine content-length\" warnings from transformers.js\n    if (msg.includes(\"content-length\") || msg.includes(\"Unable to determine\")) {\n      return;\n    }\n    originalWarn.apply(console, args);\n  };\n\n  return fn().finally(() => {\n    console.warn = originalWarn;\n  });\n}\n\nimport { generateCacheKey, getGlobalCache } from \"./cache.js\";\nimport {\n  BUILTIN_MODELS,\n  createExternalModelConfig,\n  fetchModelContextLength,\n  getModelConfig,\n  resolveModel,\n} from \"./models.js\";\nimport type {\n  AudioChunk,\n  EmbedOptions,\n  EmbedResult,\n  GenerateOptions,\n  GenerateResult,\n  GerbilConfig,\n  JsonOptions,\n  LoadOptions,\n  LoadSTTOptions,\n  LoadTTSOptions,\n  ModelConfig,\n  SessionStats,\n  SpeakOptions,\n  SpeakResult,\n  STTModelConfig,\n  StreamingTranscriptionOptions,\n  StreamingTranscriptionSession,\n  SystemInfo,\n  TranscribeOptions,\n  TranscribeResult,\n  VoiceInfo,\n} from \"./types.js\";\n\n// TTS types for lazy loading\ntype KokoroTTSType = import(\"./tts.js\").KokoroTTS;\ntype SupertonicTTSType = import(\"./tts.js\").SupertonicTTS;\ntype TTSBackendType = KokoroTTSType | SupertonicTTSType;\n\n// STT type for lazy loading\ntype WhisperSTTType = import(\"./stt.js\").WhisperSTT;\n\n// Default voices for listVoices() when TTS not loaded\nconst KOKORO_VOICES_DEFAULT: VoiceInfo[] = [\n  {\n    id: \"af_bella\",\n    name: \"Bella\",\n    gender: \"female\",\n    language: \"en-us\",\n    description: \"American female, warm and friendly\",\n  },\n  {\n    id: \"af_sarah\",\n    name: \"Sarah\",\n    gender: \"female\",\n    language: \"en-us\",\n    description: \"American female, clear and professional\",\n  },\n  {\n    id: \"af_nicole\",\n    name: \"Nicole\",\n    gender: \"female\",\n    language: \"en-us\",\n    description: \"American female, soft and gentle\",\n  },\n  {\n    id: \"af_sky\",\n    name: \"Sky\",\n    gender: \"female\",\n    language: \"en-us\",\n    description: \"American female, young and energetic\",\n  },\n  {\n    id: \"am_adam\",\n    name: \"Adam\",\n    gender: \"male\",\n    language: \"en-us\",\n    description: \"American male, deep and confident\",\n  },\n  {\n    id: \"am_michael\",\n    name: \"Michael\",\n    gender: \"male\",\n    language: \"en-us\",\n    description: \"American male, warm and friendly\",\n  },\n  {\n    id: \"bf_emma\",\n    name: \"Emma\",\n    gender: \"female\",\n    language: \"en-gb\",\n    description: \"British female, elegant and clear\",\n  },\n  {\n    id: \"bf_isabella\",\n    name: \"Isabella\",\n    gender: \"female\",\n    language: \"en-gb\",\n    description: \"British female, sophisticated\",\n  },\n  {\n    id: \"bm_george\",\n    name: \"George\",\n    gender: \"male\",\n    language: \"en-gb\",\n    description: \"British male, distinguished\",\n  },\n  {\n    id: \"bm_lewis\",\n    name: \"Lewis\",\n    gender: \"male\",\n    language: \"en-gb\",\n    description: \"British male, friendly and warm\",\n  },\n];\n\nimport { extractJson, zodToJsonSchema } from \"./utils.js\";\n\n// Configure transformers.js based on environment\nconst isBrowser = typeof window !== \"undefined\";\nenv.allowLocalModels = !isBrowser; // false in browser (fetch from HuggingFace)\nenv.useBrowserCache = isBrowser; // true in browser (cache in IndexedDB)\n\n// ============================================\n// Gerbil Class\n// ============================================\n\n// WebGPU initialization state for Node.js\nlet webgpuInitialized = false;\nlet webgpuAvailable = false;\n\n/**\n * Initialize WebGPU for Node.js environments\n * Called automatically before model loading\n */\nasync function initNodeWebGPU(): Promise<boolean> {\n  if (webgpuInitialized) {\n    return webgpuAvailable;\n  }\n  webgpuInitialized = true;\n\n  // Skip if in browser (already has WebGPU)\n  if (typeof window !== \"undefined\") {\n    webgpuAvailable = \"gpu\" in navigator;\n    return webgpuAvailable;\n  }\n\n  // Try to initialize WebGPU in Node.js via Dawn\n  // Use Function constructor to hide import from bundlers\n  try {\n    const dynamicImport = new Function(\"specifier\", \"return import(specifier)\");\n    const webgpuModule = await dynamicImport(\"webgpu\");\n    const { create, globals } = webgpuModule;\n\n    // Extend globalThis with WebGPU globals\n    Object.assign(globalThis, globals);\n\n    // Create navigator.gpu\n    if (!(globalThis as any).navigator) {\n      (globalThis as any).navigator = {};\n    }\n    (globalThis as any).navigator.gpu = create([]);\n\n    webgpuAvailable = true;\n  } catch {\n    // WebGPU not available, will fall back to CPU\n    webgpuAvailable = false;\n  }\n\n  return webgpuAvailable;\n}\n\n// ChromeGPUBackend is dynamically imported only in Node.js to avoid bundling puppeteer in browser\ntype ChromeGPUBackendType = import(\"./chrome-backend.js\").ChromeGPUBackend;\n\nexport class Gerbil {\n  private generator: TextGenerationPipeline | null = null;\n  private tokenizer: PreTrainedTokenizer | null = null;\n  private model: any = null; // AutoModelForCausalLM instance\n  private embedder: FeatureExtractionPipeline | null = null;\n  private currentModel: string | null = null;\n  private modelConfig: ModelConfig | null = null;\n  private readonly config: GerbilConfig;\n  private stats: SessionStats;\n  private useDirect = false; // Use direct model loading (for WebGPU)\n  private chromeBackend: ChromeGPUBackendType | null = null; // Chrome backend for Node.js WebGPU\n  private _deviceMode: \"webgpu\" | \"cpu\" | \"wasm\" = \"cpu\"; // Track which backend is active\n\n  // Vision model components\n  private processor: any = null; // AutoProcessor for vision models\n  private visionModel: any = null; // AutoModelForImageTextToText instance\n  private isVisionModel = false; // Whether current model supports vision\n\n  constructor(config: GerbilConfig = {}) {\n    this.config = config;\n    this.stats = {\n      prompts: 0,\n      tokensIn: 0,\n      tokensOut: 0,\n      avgSpeed: 0,\n      totalTime: 0,\n      cacheHits: 0,\n      cacheMisses: 0,\n    };\n  }\n\n  // ============================================\n  // Static Methods\n  // ============================================\n\n  static listModels(): ModelConfig[] {\n    return Object.values(BUILTIN_MODELS);\n  }\n\n  static getModel(modelId: string): ModelConfig | undefined {\n    return BUILTIN_MODELS[modelId];\n  }\n\n  // ============================================\n  // Model Loading\n  // ============================================\n\n  /**\n   * Load a model\n   *\n   * @example\n   * ```ts\n   * // Built-in model\n   * await g.loadModel(\"qwen3-0.6b\");\n   *\n   * // HuggingFace model\n   * await g.loadModel(\"hf:microsoft/Phi-3-mini\");\n   *\n   * // Local model\n   * await g.loadModel(\"file:./models/my-model\");\n   *\n   * // Vision model\n   * await g.loadModel(\"ministral-3b\");\n   * ```\n   */\n  async loadModel(modelId = \"qwen3-0.6b\", options: LoadOptions = {}): Promise<void> {\n    // Dispose any existing model/backend before loading a new one\n    // This prevents zombie Chrome pages when switching models\n    if (this.isLoaded()) {\n      await this.dispose();\n    }\n\n    // Initialize WebGPU for Node.js if needed\n    await initNodeWebGPU();\n\n    const source = resolveModel(modelId);\n    const { onProgress, device = \"auto\", dtype: userDtype } = options;\n\n    // Get or create model config\n    let config = getModelConfig(modelId);\n    if (!config) {\n      // Try to fetch actual context length from HuggingFace config.json\n      const contextLength = await fetchModelContextLength(source.path).catch(() => null);\n      config = createExternalModelConfig(modelId, source.path, contextLength || undefined);\n    }\n\n    // Route to vision model loading if needed\n    if (config.supportsVision) {\n      return this.loadVisionModel(modelId, source.path, config, options);\n    }\n\n    onProgress?.({ status: `Loading ${modelId}...` });\n\n    // Map device to transformers.js device\n    // Browser supports: webgpu, wasm (no cpu)\n    // Node supports: webgpu, cpu\n    const isBrowser = typeof window !== \"undefined\";\n    const fallbackDevice = isBrowser ? \"wasm\" : \"cpu\";\n    let tfDevice: \"webgpu\" | \"wasm\" | \"cpu\" = fallbackDevice;\n    if (device === \"webgpu\" || device === \"gpu\" || device === \"auto\") {\n      tfDevice = \"webgpu\";\n    }\n\n    // Use q4f16 for WebGPU (required for Qwen3), q4 for CPU/WASM\n    const dtype = userDtype ?? (tfDevice === \"webgpu\" ? \"q4f16\" : \"q4\");\n\n    // Track if we're still in loading phase (to suppress progress during inference)\n    let isLoading = true;\n    let lastFile = \"\";\n    let lastPct = -1;\n\n    const progressCallback = (progress: any) => {\n      if (!isLoading) {\n        return; // Suppress progress after initial load\n      }\n\n      if (progress.status === \"progress\" && progress.file) {\n        const pct = Math.round(progress.progress || 0);\n        // Only report if file changed or progress increased significantly\n        if (progress.file !== lastFile || pct >= lastPct + 5) {\n          lastFile = progress.file;\n          lastPct = pct;\n          onProgress?.({\n            status: `Downloading ${progress.file}`,\n            progress: pct,\n            file: progress.file,\n          });\n        }\n      }\n    };\n\n    try {\n      // Use direct model loading for browser WebGPU (like qwen-web does)\n      // This bypasses pipeline() which may have different ONNX session config\n      if (isBrowser && tfDevice === \"webgpu\") {\n        onProgress?.({ status: \"Loading tokenizer...\" });\n        this.tokenizer = (await suppressNoisyWarnings(() =>\n          AutoTokenizer.from_pretrained(source.path, {\n            progress_callback: progressCallback,\n          }),\n        )) as PreTrainedTokenizer;\n\n        onProgress?.({ status: \"Loading model...\" });\n        this.model = await suppressNoisyWarnings(() =>\n          AutoModelForCausalLM.from_pretrained(source.path, {\n            dtype,\n            device: tfDevice,\n            progress_callback: progressCallback,\n          }),\n        );\n\n        this.useDirect = true;\n        this._deviceMode = \"webgpu\";\n        this.isVisionModel = false;\n        isLoading = false;\n        this.currentModel = modelId;\n        this.modelConfig = config;\n        onProgress?.({ status: \"Ready (WebGPU)!\" });\n      } else if (!isBrowser && tfDevice === \"webgpu\") {\n        // Node.js + WebGPU: Use Chrome backend for real GPU acceleration\n        onProgress?.({ status: \"Starting Chrome WebGPU backend...\" });\n\n        // Dynamic import to avoid bundling puppeteer in browser builds\n        const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n        this.chromeBackend = await ChromeGPUBackend.create({\n          modelId: source.path,\n          contextLength: config.contextLength,\n          onProgress,\n        });\n\n        this.useDirect = false;\n        this._deviceMode = \"webgpu\";\n        this.isVisionModel = false;\n        isLoading = false;\n        this.currentModel = modelId;\n        this.modelConfig = config;\n        // Ready status is set by ChromeGPUBackend\n      } else {\n        // Use pipeline for CPU / WASM\n        const pipelineOptions = {\n          dtype,\n          device: tfDevice,\n          progress_callback: progressCallback,\n        };\n        this.generator = (await suppressNoisyWarnings(() =>\n          pipeline(\"text-generation\", source.path, pipelineOptions as any),\n        )) as TextGenerationPipeline;\n\n        this.useDirect = false;\n        this._deviceMode = tfDevice as \"cpu\" | \"wasm\";\n        this.isVisionModel = false;\n        isLoading = false;\n        this.currentModel = modelId;\n        this.modelConfig = config;\n        onProgress?.({ status: `Ready (${tfDevice.toUpperCase()})!` });\n      }\n    } catch (err) {\n      // Fallback to CPU/WASM if GPU fails (silently)\n      if (tfDevice !== fallbackDevice) {\n        onProgress?.({ status: `Using ${fallbackDevice.toUpperCase()}...` });\n\n        // Clean up Chrome backend if it was partially initialized\n        if (this.chromeBackend) {\n          await this.chromeBackend.dispose();\n          this.chromeBackend = null;\n        }\n\n        // Fallback always uses pipeline (WASM/CPU don't need direct loading)\n        this.generator = (await suppressNoisyWarnings(() =>\n          pipeline(\"text-generation\", source.path, {\n            dtype: \"q4\",\n            device: fallbackDevice,\n            progress_callback: progressCallback,\n          } as any),\n        )) as TextGenerationPipeline;\n\n        this.useDirect = false;\n        this._deviceMode = fallbackDevice as \"cpu\" | \"wasm\";\n        this.isVisionModel = false;\n        isLoading = false;\n        this.currentModel = modelId;\n        this.modelConfig = config;\n        onProgress?.({ status: `Ready (${fallbackDevice.toUpperCase()})!` });\n      } else {\n        throw err;\n      }\n    }\n  }\n\n  /**\n   * Load a vision model (VLM)\n   * Uses AutoProcessor + AutoModelForImageTextToText instead of tokenizer + causal LM\n   */\n  private async loadVisionModel(\n    modelId: string,\n    repoPath: string,\n    config: ModelConfig,\n    options: LoadOptions = {},\n  ): Promise<void> {\n    const { onProgress, device = \"auto\" } = options;\n\n    onProgress?.({ status: `Loading ${modelId} (vision model)...` });\n\n    const isBrowser = typeof window !== \"undefined\";\n    const fallbackDevice = isBrowser ? \"wasm\" : \"cpu\";\n    let tfDevice: \"webgpu\" | \"wasm\" | \"cpu\" = fallbackDevice;\n    if (device === \"webgpu\" || device === \"gpu\" || device === \"auto\") {\n      tfDevice = \"webgpu\";\n    }\n\n    // Node.js + WebGPU: Use Chrome backend for GPU acceleration\n    if (!isBrowser && tfDevice === \"webgpu\") {\n      onProgress?.({ status: \"Starting Chrome WebGPU backend (vision)...\" });\n\n      // Dynamic import to avoid bundling puppeteer in browser builds\n      const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n      this.chromeBackend = await ChromeGPUBackend.create({\n        modelId: repoPath,\n        contextLength: config.contextLength,\n        isVision: true, // Enable vision mode in Chrome backend\n        onProgress,\n      });\n\n      this.useDirect = false;\n      this._deviceMode = \"webgpu\";\n      this.isVisionModel = true;\n      this.currentModel = modelId;\n      this.modelConfig = config;\n      // Ready status is set by ChromeGPUBackend\n      return;\n    }\n\n    // Browser or CPU/WASM: Load directly\n    let lastFile = \"\";\n    let lastPct = -1;\n\n    const progressCallback = (progress: any) => {\n      if (progress.status === \"progress\" && progress.file) {\n        const pct = Math.round(progress.progress || 0);\n        if (progress.file !== lastFile || pct >= lastPct + 5) {\n          lastFile = progress.file;\n          lastPct = pct;\n          onProgress?.({\n            status: `Downloading ${progress.file}`,\n            progress: pct,\n            file: progress.file,\n          });\n        }\n      }\n    };\n\n    try {\n      // Load processor (handles both tokenization and image preprocessing)\n      onProgress?.({ status: \"Loading processor...\" });\n      this.processor = await suppressNoisyWarnings(() =>\n        AutoProcessor.from_pretrained(repoPath, {\n          progress_callback: progressCallback,\n        }),\n      );\n\n      // Load vision model\n      onProgress?.({ status: \"Loading vision model...\" });\n      this.visionModel = await suppressNoisyWarnings(() =>\n        AutoModelForImageTextToText.from_pretrained(repoPath, {\n          device: tfDevice,\n          progress_callback: progressCallback,\n        }),\n      );\n\n      this.isVisionModel = true;\n      this.useDirect = true;\n      this._deviceMode = tfDevice === \"webgpu\" ? \"webgpu\" : (tfDevice as \"cpu\" | \"wasm\");\n      this.currentModel = modelId;\n      this.modelConfig = config;\n      onProgress?.({ status: `Ready (Vision, ${tfDevice.toUpperCase()})!` });\n    } catch (err) {\n      // Fallback to CPU/WASM if GPU fails\n      if (tfDevice !== fallbackDevice) {\n        onProgress?.({ status: `Vision model: Using ${fallbackDevice.toUpperCase()}...` });\n\n        this.processor = await suppressNoisyWarnings(() =>\n          AutoProcessor.from_pretrained(repoPath, {\n            progress_callback: progressCallback,\n          }),\n        );\n\n        this.visionModel = await suppressNoisyWarnings(() =>\n          AutoModelForImageTextToText.from_pretrained(repoPath, {\n            device: fallbackDevice,\n            progress_callback: progressCallback,\n          }),\n        );\n\n        this.isVisionModel = true;\n        this.useDirect = true;\n        this._deviceMode = fallbackDevice as \"cpu\" | \"wasm\";\n        this.currentModel = modelId;\n        this.modelConfig = config;\n        onProgress?.({ status: `Ready (Vision, ${fallbackDevice.toUpperCase()})!` });\n      } else {\n        throw err;\n      }\n    }\n  }\n\n  /**\n   * Check if a model is loaded\n   */\n  isLoaded(): boolean {\n    return (\n      this.generator !== null ||\n      (this.useDirect && this.model !== null) ||\n      this.chromeBackend !== null ||\n      (this.isVisionModel && this.visionModel !== null)\n    );\n  }\n\n  /**\n   * Check if current model supports vision\n   */\n  supportsVision(): boolean {\n    return this.isVisionModel && this.modelConfig?.supportsVision === true;\n  }\n\n  /**\n   * Get current model info\n   */\n  getModelInfo(): ModelConfig | null {\n    return this.modelConfig;\n  }\n\n  /**\n   * Get current device mode (webgpu, cpu, or wasm)\n   */\n  getDeviceMode(): \"webgpu\" | \"cpu\" | \"wasm\" {\n    return this._deviceMode;\n  }\n\n  /**\n   * Get dtype used for current model\n   */\n  getDtype(): string {\n    // WebGPU uses q4f16, CPU/WASM use q4\n    return this._deviceMode === \"webgpu\" ? \"q4f16\" : \"q4\";\n  }\n\n  /**\n   * Get response cache statistics\n   */\n  getResponseCacheStats(): { hits: number; misses: number; size: number; hitRate: number } {\n    const cache = getGlobalCache();\n    const stats = cache.getStats();\n    return {\n      hits: stats.hits,\n      misses: stats.misses,\n      size: stats.size,\n      hitRate: cache.getHitRate(),\n    };\n  }\n\n  /**\n   * Clear the response cache (for cached generate() results)\n   */\n  clearResponseCache(): void {\n    getGlobalCache().clear();\n  }\n\n  /**\n   * Get Chrome backend status (if using WebGPU via Chrome)\n   */\n  getChromeStatus(): {\n    pid: number | null;\n    port: number;\n    modelId: string;\n    startedAt: Date | null;\n  } | null {\n    if (!this.chromeBackend) {\n      return null;\n    }\n    return this.chromeBackend.getStatus();\n  }\n\n  /**\n   * Get Chrome memory usage (if using WebGPU via Chrome)\n   * Returns JS heap memory in bytes\n   */\n  async getChromeMemory(): Promise<{ jsHeapUsed: number; jsHeapTotal: number } | null> {\n    if (!this.chromeBackend) {\n      return null;\n    }\n    return this.chromeBackend.getMemoryUsage();\n  }\n\n  /**\n   * Get memory usage in GB (if using WebGPU via Chrome)\n   */\n  async getMemoryUsage(): Promise<{ usedGB: number; totalGB: number; usedPercent: number } | null> {\n    if (!this.chromeBackend) {\n      return null;\n    }\n    return this.chromeBackend.getMemoryStats();\n  }\n\n  /**\n   * Clear KV cache to free memory\n   * This will reset the conversation context but free up memory\n   */\n  async clearCache(): Promise<void> {\n    if (this.chromeBackend) {\n      await this.chromeBackend.reset();\n    }\n  }\n\n  /**\n   * Check memory usage and cleanup if needed\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    if (!this.chromeBackend) {\n      return false;\n    }\n    return this.chromeBackend.checkMemoryAndCleanup(thresholdGB);\n  }\n\n  // ============================================\n  // Text Generation\n  // ============================================\n\n  /**\n   * Generate text (automatically routes to vision generation if images provided)\n   *\n   * @example\n   * ```ts\n   * // Text generation\n   * const result = await g.generate(\"Hello!\");\n   *\n   * // Vision generation (with vision model)\n   * const result = await g.generate(\"What's in this image?\", {\n   *   images: [{ source: \"https://example.com/cat.jpg\" }]\n   * });\n   * ```\n   */\n  async generate(prompt: string, options: GenerateOptions = {}): Promise<GenerateResult> {\n    if (!this.isLoaded()) {\n      // Auto-load default model\n      await this.loadModel(this.config.model || \"qwen3-0.6b\");\n    }\n\n    const { images } = options;\n\n    // Route to local vision generation if:\n    // 1. Images provided\n    // 2. Model supports vision\n    // 3. NOT using Chrome backend (Chrome backend handles vision internally)\n    if (images?.length && this.isVisionModel && !this.chromeBackend) {\n      return this.generateWithVision(prompt, options);\n    }\n\n    // Warn if images provided but model doesn't support vision\n    if (images?.length && !this.isVisionModel) {\n    }\n\n    const {\n      maxTokens = 256,\n      temperature = 0.7,\n      topP = 0.9,\n      topK = 50,\n      thinking = false,\n      system,\n      cache = false,\n      cacheTtl,\n    } = options;\n\n    // Check cache if enabled (skip for streaming/vision)\n    if (cache && !options.onToken && !images?.length) {\n      const cacheKey = generateCacheKey(prompt, this.currentModel || \"\", {\n        maxTokens,\n        temperature,\n        topP,\n        topK,\n        system,\n        thinking,\n      });\n      const cached = getGlobalCache().get(cacheKey);\n      if (cached) {\n        return cached;\n      }\n    }\n\n    const startTime = performance.now();\n\n    try {\n      let rawText = \"\";\n\n      if (this.chromeBackend) {\n        // Chrome backend approach (for Node.js WebGPU via Chrome)\n        try {\n          rawText = await this.chromeBackend.generate(prompt, {\n            maxTokens,\n            temperature,\n            topP,\n            topK,\n            thinking,\n            system,\n            // Pass images for vision models\n            images: images?.map((img) => img.source),\n            // Wrap onToken to match Gerbil's simpler signature\n            onToken: options.onToken ? (t) => options.onToken?.(t.text) : undefined,\n          });\n        } catch (chromeErr: any) {\n          // If Chrome died (OOM, crash), fall back to CPU silently\n          if (chromeErr?.message === \"CHROME_BACKEND_DEAD\" || !this.chromeBackend?.isAlive()) {\n            await this.chromeBackend?.dispose().catch(() => {});\n            this.chromeBackend = null;\n            this._deviceMode = \"cpu\";\n            // Load CPU fallback and retry\n            const modelPath = this.currentModel || \"qwen3-0.6b\";\n            this.generator = (await pipeline(\"text-generation\", modelPath, {\n              dtype: \"q4\",\n              device: \"cpu\",\n            } as any)) as TextGenerationPipeline;\n            // Retry with CPU\n            return this.generate(prompt, options);\n          }\n          throw chromeErr;\n        }\n      } else if (this.useDirect && this.model && this.tokenizer) {\n        // Direct model approach (for browser WebGPU)\n        const messages = this.buildMessages(prompt, { ...options, thinking });\n\n        const inputs = (this.tokenizer as any).apply_chat_template(messages, {\n          add_generation_prompt: true,\n          return_dict: true,\n          enable_thinking: thinking, // Qwen3 thinking mode\n        });\n\n        const output = await this.model.generate({\n          ...inputs,\n          max_new_tokens: maxTokens,\n          temperature: temperature > 0 ? temperature : undefined,\n          top_p: topP,\n          top_k: topK,\n          do_sample: temperature > 0,\n        });\n\n        // Get input length to extract only generated tokens\n        const inputLength = inputs.input_ids.dims?.[1] || inputs.input_ids.data?.length || 0;\n\n        // Slice output tensor to get only new tokens (skip prompt)\n        const outputTokens = output.slice(null, [inputLength, null]);\n        const decoded = this.tokenizer.batch_decode(outputTokens, {\n          skip_special_tokens: true,\n        });\n\n        rawText = decoded[0] || \"\";\n\n        // If we still have prompt artifacts, extract assistant response\n        if (rawText.toLowerCase().includes(\"assistant\")) {\n          const match = rawText.match(/assistant[:\\s]*([\\s\\S]*)/i);\n          if (match) {\n            rawText = match[1].trim();\n          }\n        }\n      } else if (this.generator) {\n        // Pipeline approach (for Node.js / CPU / WASM)\n        const formattedPrompt = this.formatPrompt(prompt, { ...options, thinking });\n\n        const output = await this.generator(formattedPrompt, {\n          max_new_tokens: maxTokens,\n          temperature,\n          top_p: topP,\n          top_k: topK,\n          do_sample: temperature > 0,\n          return_full_text: false,\n        });\n\n        // Extract text from pipeline output\n        if (Array.isArray(output) && output[0]) {\n          const result = output[0] as any;\n          if (Array.isArray(result.generated_text)) {\n            const last = result.generated_text.at(-1);\n            rawText = last?.content || \"\";\n          } else {\n            rawText = result.generated_text || \"\";\n          }\n        }\n      } else {\n        throw new Error(\"No model loaded\");\n      }\n\n      const endTime = performance.now();\n      const totalTime = endTime - startTime;\n\n      rawText = this.cleanOutput(rawText);\n\n      // Always parse thinking to strip <think> tags from output\n      // (model may generate them even without thinking mode enabled)\n      const { thinking: thinkingText, response } = this.parseThinking(rawText);\n\n      // Only include thinking in result if mode was enabled\n      const finalThinking = thinking ? thinkingText : undefined;\n\n      const tokensGenerated = Math.ceil(response.length / 4);\n\n      // Update stats\n      this.stats.prompts += 1;\n      this.stats.tokensOut += tokensGenerated;\n      this.stats.totalTime += totalTime;\n      this.stats.avgSpeed = (this.stats.tokensOut / this.stats.totalTime) * 1000;\n\n      const result: GenerateResult = {\n        text: response,\n        thinking: finalThinking,\n        tokensGenerated,\n        tokensPerSecond: (tokensGenerated / totalTime) * 1000,\n        totalTime,\n        finishReason: \"stop\",\n        provider: \"local\",\n        cached: false,\n      };\n\n      // Store in cache if enabled\n      if (cache && !options.onToken && !images?.length) {\n        const cacheKey = generateCacheKey(prompt, this.currentModel || \"\", {\n          maxTokens,\n          temperature,\n          topP,\n          topK,\n          system,\n          thinking,\n        });\n        getGlobalCache().set(cacheKey, result, cacheTtl);\n      }\n\n      return result;\n    } catch (_error) {\n      return {\n        text: \"\",\n        tokensGenerated: 0,\n        tokensPerSecond: 0,\n        totalTime: performance.now() - startTime,\n        finishReason: \"error\",\n        provider: \"local\",\n        cached: false,\n      };\n    }\n  }\n\n  /**\n   * Stream text generation (simulated token-by-token)\n   *\n   * Note: Yields the raw output including <think> tags if thinking mode is enabled.\n   * The final result has parsed thinking separated out.\n   */\n  async *stream(\n    prompt: string,\n    options: GenerateOptions = {},\n  ): AsyncGenerator<string, GenerateResult, unknown> {\n    if (!this.isLoaded()) {\n      await this.loadModel(this.config.model || \"qwen3-0.6b\");\n    }\n\n    const startTime = performance.now();\n\n    // For Chrome backend, use real streaming via onToken callback\n    if (this.chromeBackend) {\n      let fullText = \"\";\n      const tokenQueue: string[] = [];\n      let resolveNext: ((value: string | null) => void) | null = null;\n      let done = false;\n\n      // Start generation with streaming callback\n      const generatePromise = this.chromeBackend\n        .generate(prompt, {\n          ...options,\n          // Convert ImageInput[] to string[] for Chrome backend\n          images: options.images?.map((img) => img.source),\n          onToken: (token) => {\n            fullText += token.text;\n            if (resolveNext) {\n              resolveNext(token.text);\n              resolveNext = null;\n            } else {\n              tokenQueue.push(token.text);\n            }\n          },\n        })\n        .then(() => {\n          done = true;\n          if (resolveNext) {\n            resolveNext(null);\n          }\n        })\n        .catch((err) => {\n          done = true;\n          if (resolveNext) {\n            resolveNext(null);\n          }\n          throw err;\n        });\n\n      // Yield tokens as they arrive\n      while (!done || tokenQueue.length > 0) {\n        if (tokenQueue.length > 0) {\n          const token = tokenQueue.shift()!;\n          yield token;\n          options.onToken?.(token);\n        } else if (!done) {\n          const token = await new Promise<string | null>((resolve) => {\n            resolveNext = resolve;\n          });\n          if (token) {\n            yield token;\n            options.onToken?.(token);\n          }\n        }\n      }\n\n      await generatePromise;\n\n      const { thinking: thinkingText, response } = this.parseThinking(fullText);\n      const tokensGenerated = Math.ceil(response.length / 4);\n      const totalTime = performance.now() - startTime;\n\n      return {\n        text: response,\n        thinking: options.thinking ? thinkingText : undefined,\n        tokensGenerated,\n        totalTime,\n        tokensPerSecond: (tokensGenerated / totalTime) * 1000,\n        finishReason: \"stop\" as const,\n      };\n    }\n\n    // For pipeline/direct model, use fake streaming (generate then yield)\n    const result = await this.generateRaw(prompt, options);\n\n    // Yield word by word for more accurate token simulation\n    // (actual tokens average ~4 chars, words are a reasonable approximation)\n    const words = result.rawText.split(/(\\s+)/);\n    for (const word of words) {\n      if (word) {\n        yield word;\n        options.onToken?.(word);\n      }\n    }\n\n    return result.result;\n  }\n\n  /**\n   * Internal: Generate with raw text access for streaming\n   */\n  private async generateRaw(\n    prompt: string,\n    options: GenerateOptions = {},\n  ): Promise<{ rawText: string; result: GenerateResult }> {\n    const { maxTokens = 256, temperature = 0.7, topP = 0.9, topK = 50, thinking = false } = options;\n\n    const startTime = performance.now();\n    const formattedPrompt = this.formatPrompt(prompt, { ...options, thinking });\n\n    try {\n      const output = await this.generator?.(formattedPrompt, {\n        max_new_tokens: maxTokens,\n        temperature,\n        top_p: topP,\n        top_k: topK,\n        do_sample: temperature > 0,\n        return_full_text: false,\n      });\n\n      const endTime = performance.now();\n      const totalTime = endTime - startTime;\n\n      // Extract text from output\n      let rawText = \"\";\n      if (Array.isArray(output) && output[0]) {\n        const result = output[0] as any;\n        if (Array.isArray(result.generated_text)) {\n          const last = result.generated_text.at(-1);\n          rawText = last?.content || \"\";\n        } else {\n          rawText = result.generated_text || \"\";\n        }\n      }\n\n      rawText = this.cleanOutput(rawText);\n      const { thinking: thinkingText, response } = this.parseThinking(rawText);\n      const finalThinking = thinking ? thinkingText : undefined;\n      const tokensGenerated = Math.ceil(response.length / 4);\n\n      // Update stats\n      this.stats.prompts += 1;\n      this.stats.tokensOut += tokensGenerated;\n      this.stats.totalTime += totalTime;\n      this.stats.avgSpeed = (this.stats.tokensOut / this.stats.totalTime) * 1000;\n\n      return {\n        rawText,\n        result: {\n          text: response,\n          thinking: finalThinking,\n          tokensGenerated,\n          tokensPerSecond: (tokensGenerated / totalTime) * 1000,\n          totalTime,\n          finishReason: \"stop\",\n          provider: \"local\",\n          cached: false,\n        },\n      };\n    } catch (_error) {\n      return {\n        rawText: \"\",\n        result: {\n          text: \"\",\n          tokensGenerated: 0,\n          tokensPerSecond: 0,\n          totalTime: performance.now() - startTime,\n          finishReason: \"error\",\n          provider: \"local\",\n          cached: false,\n        },\n      };\n    }\n  }\n\n  // ============================================\n  // Vision Generation\n  // ============================================\n\n  /**\n   * Generate text from images using a vision model\n   * Called automatically by generate() when images are provided\n   */\n  private async generateWithVision(\n    prompt: string,\n    options: GenerateOptions,\n  ): Promise<GenerateResult> {\n    if (!(this.processor && this.visionModel)) {\n      throw new Error(\"Vision model not loaded. Load a vision model first.\");\n    }\n\n    const {\n      images = [],\n      maxTokens = 2048,\n      temperature = 0.7,\n      topP = 0.9,\n      topK = 20,\n      system,\n    } = options;\n\n    const startTime = performance.now();\n\n    try {\n      // Build message content with images and text\n      const content: Array<{ type: string; text?: string }> = [];\n\n      // Add image placeholders (the actual images are passed separately)\n      for (let i = 0; i < images.length; i += 1) {\n        content.push({ type: \"image\" });\n      }\n\n      // Add text prompt\n      content.push({ type: \"text\", text: prompt });\n\n      const messages = [\n        ...(system ? [{ role: \"system\", content: system }] : []),\n        { role: \"user\", content },\n      ];\n\n      // Apply chat template\n      const chatPrompt = this.processor.apply_chat_template(messages);\n\n      // Load images using RawImage\n      const loadedImages = await Promise.all(\n        images.map(async (img) => await RawImage.fromURL(img.source)),\n      );\n\n      // Process inputs (image + text)\n      const inputs = await this.processor(\n        loadedImages.length === 1 ? loadedImages[0] : loadedImages,\n        chatPrompt,\n        { add_special_tokens: false },\n      );\n\n      // Set up streaming if callback provided\n      let fullText = \"\";\n      const streamer = options.onToken\n        ? new TextStreamer(this.processor.tokenizer, {\n            skip_prompt: true,\n            skip_special_tokens: true,\n            callback_function: (text: string) => {\n              fullText += text;\n              options.onToken?.(text);\n            },\n          })\n        : undefined;\n\n      // Generate\n      const outputs = await this.visionModel.generate({\n        ...inputs,\n        max_new_tokens: maxTokens,\n        temperature: temperature > 0 ? temperature : undefined,\n        top_p: topP,\n        top_k: topK,\n        do_sample: temperature > 0,\n        ...(streamer ? { streamer } : {}),\n      });\n\n      // Decode output (skip the prompt tokens)\n      const inputLength = inputs.input_ids.dims?.at(-1) || 0;\n      const decoded = this.processor.batch_decode(outputs.slice(null, [inputLength, null]), {\n        skip_special_tokens: true,\n      });\n\n      const text = decoded[0] || fullText || \"\";\n      const totalTime = performance.now() - startTime;\n      const tokensGenerated = Math.ceil(text.length / 4);\n\n      // Update stats\n      this.stats.prompts += 1;\n      this.stats.tokensOut += tokensGenerated;\n      this.stats.totalTime += totalTime;\n      this.stats.avgSpeed = (this.stats.tokensOut / this.stats.totalTime) * 1000;\n\n      return {\n        text: this.cleanOutput(text),\n        tokensGenerated,\n        tokensPerSecond: (tokensGenerated / totalTime) * 1000,\n        totalTime,\n        finishReason: \"stop\",\n        provider: \"local\",\n        cached: false,\n      };\n    } catch (_error) {\n      return {\n        text: \"\",\n        tokensGenerated: 0,\n        tokensPerSecond: 0,\n        totalTime: performance.now() - startTime,\n        finishReason: \"error\",\n        provider: \"local\",\n        cached: false,\n      };\n    }\n  }\n\n  // ============================================\n  // Structured Output (JSON)\n  // ============================================\n\n  /**\n   * Generate structured JSON output\n   */\n  async json<T>(prompt: string, options: JsonOptions<T>): Promise<T> {\n    const { schema, retries = 3, temperature = 0.3 } = options;\n\n    const systemPrompt = `You are a JSON generator. You MUST respond with valid JSON only.\nNo explanations, no markdown, no code blocks. Just pure JSON.\nThe JSON must conform to this schema: ${JSON.stringify(zodToJsonSchema(schema))}`;\n\n    for (let attempt = 0; attempt < retries; attempt += 1) {\n      const result = await this.generate(prompt, {\n        system: options.system || systemPrompt,\n        temperature,\n        maxTokens: 1000,\n      });\n\n      try {\n        // Try to extract JSON from response\n        const jsonStr = extractJson(result.text);\n        const parsed = JSON.parse(jsonStr);\n        const validated = schema.parse(parsed);\n        return validated;\n      } catch (error) {\n        if (attempt === retries - 1) {\n          throw new Error(`Failed to generate valid JSON after ${retries} attempts: ${error}`);\n        }\n      }\n    }\n\n    throw new Error(\"Failed to generate valid JSON\");\n  }\n\n  // ============================================\n  // Embeddings\n  // ============================================\n\n  /**\n   * Generate embeddings\n   */\n  async embed(text: string, options: EmbedOptions = {}): Promise<EmbedResult> {\n    if (!this.embedder) {\n      // Load embedding model\n      const model = options.model || \"Xenova/all-MiniLM-L6-v2\";\n      this.embedder = (await pipeline(\"feature-extraction\", model)) as FeatureExtractionPipeline;\n    }\n\n    const startTime = performance.now();\n    const output = await this.embedder(text, {\n      pooling: \"mean\",\n      normalize: options.normalize !== false,\n    });\n\n    const vector = Array.from(output.data as Float32Array);\n\n    return {\n      vector,\n      text,\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /**\n   * Generate embeddings for multiple texts\n   */\n  async embedBatch(texts: string[], options: EmbedOptions = {}): Promise<EmbedResult[]> {\n    const results: EmbedResult[] = [];\n    for (const text of texts) {\n      results.push(await this.embed(text, options));\n    }\n    return results;\n  }\n\n  // ============================================\n  // Stats & Info\n  // ============================================\n\n  /**\n   * Get session stats\n   */\n  getStats(): SessionStats {\n    return { ...this.stats };\n  }\n\n  /**\n   * Get system info\n   */\n  getInfo(): SystemInfo {\n    return {\n      version: \"1.0.0\",\n      model: this.modelConfig,\n      device: {\n        backend: \"transformers.js\",\n        gpu: null, // TODO: detect GPU\n        vram: null,\n        status: this.isLoaded() ? \"ready\" : \"loading\",\n      },\n      context: {\n        max: this.modelConfig?.contextLength || 0,\n        used: 0,\n        available: this.modelConfig?.contextLength || 0,\n      },\n      cache: {\n        location: \"~/.gerbil/models\",\n        size: \"0 MB\",\n        modelCount: 0,\n      },\n    };\n  }\n\n  /**\n   * Reset stats\n   */\n  resetStats(): void {\n    this.stats = {\n      prompts: 0,\n      tokensIn: 0,\n      tokensOut: 0,\n      avgSpeed: 0,\n      totalTime: 0,\n      cacheHits: 0,\n      cacheMisses: 0,\n    };\n  }\n\n  // ============================================\n  // Text-to-Speech (TTS)\n  // ============================================\n\n  private tts: TTSBackendType | null = null;\n  private ttsModelId: string = \"kokoro-82m\";\n\n  /**\n   * Load TTS model for text-to-speech synthesis\n   *\n   * @example\n   * ```ts\n   * // Load default (Kokoro)\n   * await g.loadTTS({ onProgress: (p) => console.log(p.status) });\n   *\n   * // Load Supertonic (faster, 44kHz output)\n   * await g.loadTTS({ model: \"supertonic-66m\" });\n   *\n   * const result = await g.speak(\"Hello world\");\n   * // result.audio = Float32Array, result.sampleRate = 24000 or 44100\n   * ```\n   */\n  async loadTTS(options: LoadTTSOptions & { model?: string } = {}): Promise<void> {\n    const modelId = options.model || \"kokoro-82m\";\n\n    // If switching models, dispose the old one\n    if (this.tts && this.ttsModelId !== modelId) {\n      await this.tts.dispose();\n      this.tts = null;\n    }\n\n    if (this.tts?.isLoaded()) {\n      return;\n    }\n\n    this.ttsModelId = modelId;\n\n    // Dynamic import to avoid bundling TTS code when not used\n    const { createTTS } = await import(\"./tts.js\");\n\n    if (!this.tts) {\n      this.tts = createTTS(modelId);\n    }\n\n    await this.tts.load(options);\n  }\n\n  /**\n   * Ensure TTS model is loaded (lazy loading)\n   */\n  async ensureTTSLoaded(options?: LoadTTSOptions): Promise<void> {\n    if (!this.tts?.isLoaded()) {\n      await this.loadTTS(options);\n    }\n  }\n\n  /**\n   * Generate speech from text\n   *\n   * @example\n   * ```ts\n   * const result = await g.speak(\"Hello world\", { voice: \"af_bella\" });\n   * // result.audio = Float32Array PCM\n   * // result.sampleRate = 24000\n   * // result.duration = seconds\n   * ```\n   */\n  async speak(text: string, options: SpeakOptions = {}): Promise<SpeakResult> {\n    await this.ensureTTSLoaded({ onProgress: options.onProgress });\n    return this.tts!.speak(text, options);\n  }\n\n  /**\n   * Stream speech generation (yields audio chunks as they're generated)\n   *\n   * @example\n   * ```ts\n   * for await (const chunk of g.speakStream(\"Long text...\")) {\n   *   // chunk.samples = Float32Array\n   *   // chunk.isFinal = boolean\n   *   playChunk(chunk);\n   * }\n   * ```\n   */\n  async *speakStream(\n    text: string,\n    options: SpeakOptions = {},\n  ): AsyncGenerator<AudioChunk, SpeakResult, unknown> {\n    await this.ensureTTSLoaded({ onProgress: options.onProgress });\n    return yield* this.tts!.speakStream(text, options);\n  }\n\n  /**\n   * Get list of available TTS voices\n   */\n  listVoices(): VoiceInfo[] {\n    if (!this.tts) {\n      // Return default voices from static import\n      return KOKORO_VOICES_DEFAULT;\n    }\n    return this.tts.listVoices();\n  }\n\n  /**\n   * Check if TTS model is loaded\n   */\n  isTTSLoaded(): boolean {\n    return this.tts?.isLoaded() ?? false;\n  }\n\n  /**\n   * Get current TTS model info\n   */\n  getTTSModelInfo(): { id: string; loaded: boolean; device?: \"webgpu\" | \"cpu\" } | null {\n    if (!this.tts) {\n      return null;\n    }\n    return {\n      id: this.ttsModelId,\n      loaded: this.tts.isLoaded(),\n      device: this.tts.isLoaded() ? this.tts.getDeviceMode() : undefined,\n    };\n  }\n\n  /**\n   * List available TTS models\n   */\n  async listTTSModels(): Promise<\n    Array<{ id: string; description: string; sampleRate: number; voiceCount: number }>\n  > {\n    const { TTS_MODELS } = await import(\"./tts.js\");\n    return Object.values(TTS_MODELS).map((m) => ({\n      id: m.id,\n      description: m.description,\n      sampleRate: m.sampleRate,\n      voiceCount: m.voices.length,\n    }));\n  }\n\n  // ============================================\n  // Speech-to-Text (STT)\n  // ============================================\n\n  private stt: WhisperSTTType | null = null;\n\n  /**\n   * Load STT model for speech-to-text transcription\n   *\n   * @example\n   * ```ts\n   * await g.loadSTT({\n   *   onProgress: (p) => console.log(p.status)\n   * });\n   *\n   * const result = await g.transcribe(audioData);\n   * console.log(result.text);\n   * ```\n   */\n  async loadSTT(modelId?: string, options: LoadSTTOptions = {}): Promise<void> {\n    if (this.stt?.isLoaded()) {\n      return;\n    }\n\n    // Dynamic import to avoid bundling STT code when not used\n    const { WhisperSTT } = await import(\"./stt.js\");\n\n    if (!this.stt) {\n      this.stt = new WhisperSTT(modelId);\n    }\n\n    await this.stt.load(options);\n  }\n\n  /**\n   * Ensure STT model is loaded (lazy loading)\n   */\n  public async ensureSTTLoaded(modelId?: string, options?: LoadSTTOptions): Promise<void> {\n    if (!this.stt?.isLoaded()) {\n      await this.loadSTT(modelId, options);\n    }\n  }\n\n  /**\n   * Transcribe audio to text\n   *\n   * @param audio - Audio data as Float32Array (16kHz mono) or Uint8Array (WAV file)\n   * @param options - Transcription options\n   *\n   * @example\n   * ```ts\n   * // From Float32Array (16kHz mono)\n   * const result = await g.transcribe(audioData);\n   * console.log(result.text);\n   *\n   * // With timestamps\n   * const result = await g.transcribe(audioData, { timestamps: true });\n   * for (const seg of result.segments) {\n   *   console.log(`[${seg.start}s] ${seg.text}`);\n   * }\n   *\n   * // From WAV file\n   * const wavData = fs.readFileSync(\"audio.wav\");\n   * const result = await g.transcribe(new Uint8Array(wavData));\n   * ```\n   */\n  async transcribe(\n    audio: Float32Array | Uint8Array,\n    options: TranscribeOptions = {},\n  ): Promise<TranscribeResult> {\n    await this.ensureSTTLoaded(undefined, { onProgress: options.onProgress });\n    return this.stt!.transcribe(audio, options);\n  }\n\n  /**\n   * Create a streaming transcription session\n   *\n   * Transcribes audio in real-time by processing chunks at regular intervals.\n   * Perfect for live captioning, call transcription, or real-time subtitles.\n   *\n   * @param options - Streaming options\n   * @returns Streaming session controller\n   *\n   * @example\n   * ```ts\n   * const session = await g.createStreamingTranscription({\n   *   chunkDuration: 3000,  // Transcribe every 3 seconds\n   *   onChunk: (text, idx) => console.log(`Chunk ${idx}: ${text}`),\n   *   onTranscript: (fullText) => console.log(\"Full:\", fullText),\n   * });\n   *\n   * // Feed audio data as it comes in\n   * session.feedAudio(audioChunk);\n   *\n   * // Start automatic interval-based transcription\n   * session.start();\n   *\n   * // Later, stop and get final transcript\n   * const finalText = await session.stop();\n   * ```\n   */\n  async createStreamingTranscription(\n    options: StreamingTranscriptionOptions = {},\n  ): Promise<StreamingTranscriptionSession> {\n    await this.ensureSTTLoaded();\n    return this.stt!.createStreamingSession(options);\n  }\n\n  /**\n   * Get list of available STT models\n   */\n  async listSTTModels(): Promise<STTModelConfig[]> {\n    // Dynamic import to avoid bundling STT code when not used\n    const { WhisperSTT } = await import(\"./stt.js\");\n    return WhisperSTT.listModels();\n  }\n\n  /**\n   * Check if STT model is loaded\n   */\n  isSTTLoaded(): boolean {\n    return this.stt?.isLoaded() ?? false;\n  }\n\n  /**\n   * Get current STT model info\n   */\n  getSTTModelInfo(): { id: string; loaded: boolean; device?: \"webgpu\" | \"cpu\" } | null {\n    if (!this.stt) {\n      return null;\n    }\n    return {\n      id: this.stt.getModelInfo().id,\n      loaded: this.stt.isLoaded(),\n      device: this.stt.isLoaded() ? this.stt.getDeviceMode() : undefined,\n    };\n  }\n\n  // ============================================\n  // Microphone Input\n  // ============================================\n\n  /**\n   * Record audio from microphone and transcribe\n   *\n   * @example\n   * ```ts\n   * // Record for 5 seconds and transcribe\n   * const result = await g.listen(5000);\n   * console.log(result.text);\n   *\n   * // Use with voice chat\n   * const userInput = await g.listen(10000);\n   * const response = await g.generate(userInput.text);\n   * await g.speak(response.text);\n   * ```\n   */\n  async listen(\n    durationMs: number = 5000,\n    options: { onProgress?: (status: string) => void } = {},\n  ): Promise<TranscribeResult> {\n    // Dynamic import for microphone (avoids bundling when not used)\n    const { Microphone, isSoxAvailable } = await import(\"./microphone.js\");\n\n    if (!isSoxAvailable()) {\n      throw new Error(\n        \"Microphone recording requires SoX. Install with:\\n\" +\n          \"  macOS: brew install sox\\n\" +\n          \"  Ubuntu: sudo apt install sox\\n\" +\n          \"  Windows: https://sox.sourceforge.net/\",\n      );\n    }\n\n    options.onProgress?.(\"Starting microphone...\");\n\n    const mic = new Microphone({ sampleRate: 16000 });\n    await mic.start();\n\n    options.onProgress?.(`Recording for ${(durationMs / 1000).toFixed(1)}s...`);\n\n    // Wait for the specified duration\n    await new Promise((r) => setTimeout(r, durationMs));\n\n    options.onProgress?.(\"Processing audio...\");\n    const { audio } = await mic.stop();\n\n    options.onProgress?.(\"Transcribing...\");\n    return this.transcribe(audio, {\n      onProgress: (p) => options.onProgress?.(p.status || \"Transcribing...\"),\n    });\n  }\n\n  /**\n   * Check if microphone recording is available\n   */\n  async isMicrophoneAvailable(): Promise<boolean> {\n    try {\n      const { isSoxAvailable } = await import(\"./microphone.js\");\n      return isSoxAvailable();\n    } catch {\n      return false;\n    }\n  }\n\n  // ============================================\n  // Cleanup\n  // ============================================\n\n  /**\n   * Dispose of resources\n   * @param disconnect If true, also disconnect from shared browser (for clean script exit)\n   */\n  async dispose(disconnect = false): Promise<void> {\n    // Clean up Chrome backend first (most important to release resources)\n    if (this.chromeBackend) {\n      try {\n        await this.chromeBackend.dispose(disconnect);\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.chromeBackend = null;\n    }\n\n    if (this.generator) {\n      if (typeof (this.generator as any).dispose === \"function\") {\n        try {\n          await (this.generator as any).dispose();\n        } catch {\n          // Ignore errors during cleanup\n        }\n      }\n      this.generator = null;\n    }\n    if (this.embedder) {\n      if (typeof (this.embedder as any).dispose === \"function\") {\n        try {\n          await (this.embedder as any).dispose();\n        } catch {\n          // Ignore errors during cleanup\n        }\n      }\n      this.embedder = null;\n    }\n\n    // Clean up vision model resources\n    if (this.visionModel) {\n      if (typeof this.visionModel.dispose === \"function\") {\n        try {\n          await this.visionModel.dispose();\n        } catch {\n          // Ignore errors during cleanup\n        }\n      }\n      this.visionModel = null;\n    }\n    if (this.processor) {\n      this.processor = null;\n    }\n\n    // Clean up TTS resources\n    if (this.tts) {\n      try {\n        await this.tts.dispose();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.tts = null;\n    }\n\n    // Clean up STT resources\n    if (this.stt) {\n      try {\n        this.stt.dispose();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.stt = null;\n    }\n\n    this.currentModel = null;\n    this.modelConfig = null;\n    this.isVisionModel = false;\n  }\n\n  /**\n   * Shutdown the shared Chrome backend completely.\n   * Call this when your script/process is done to ensure proper cleanup.\n   * This closes the shared browser used for WebGPU acceleration.\n   */\n  static async shutdown(): Promise<void> {\n    // Dynamic import to match how ChromeGPUBackend is loaded\n    const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n    await ChromeGPUBackend.closeSharedBrowser();\n  }\n\n  /**\n   * Get global WebGPU process info (all active backends)\n   * Useful for monitoring and debugging memory leaks\n   */\n  static async getWebGPUProcesses(): Promise<{\n    browser: {\n      running: boolean;\n      pid: number | null;\n      port: number;\n      activePagesCount: number;\n      maxPages: number;\n    };\n    backends: Array<{\n      modelId: string;\n      isVision: boolean;\n      isReady: boolean;\n      memory: { usedGB: number; totalGB: number; usedPercent: number } | null;\n    }>;\n  } | null> {\n    // Not available in browser\n    if (typeof window !== \"undefined\") {\n      return null;\n    }\n\n    try {\n      const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n      const browser = ChromeGPUBackend.getGlobalBrowserStatus();\n      const backends = await ChromeGPUBackend.getAllBackendsInfo();\n\n      return { browser, backends };\n    } catch {\n      return null;\n    }\n  }\n\n  /**\n   * Kill all WebGPU processes (for zombie cleanup)\n   * Use this if you suspect memory leaks from undisposed Gerbil instances\n   */\n  static async killAllWebGPU(): Promise<{ pagesKilled: number; browserKilled: boolean } | null> {\n    // Not available in browser\n    if (typeof window !== \"undefined\") {\n      return null;\n    }\n\n    try {\n      const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n      return await ChromeGPUBackend.killAllBackends();\n    } catch {\n      return null;\n    }\n  }\n\n  /**\n   * Kill a specific WebGPU backend by index\n   * @param index Index of the backend to kill (0-based)\n   */\n  static async killWebGPUBackend(index: number): Promise<boolean> {\n    // Not available in browser\n    if (typeof window !== \"undefined\") {\n      return false;\n    }\n\n    try {\n      const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n      return await ChromeGPUBackend.killBackendByIndex(index);\n    } catch {\n      return false;\n    }\n  }\n\n  /**\n   * Get all Chrome pages across ALL Gerbil processes\n   * This provides cross-process visibility into WebGPU backends\n   */\n  static async getAllChromePagesInfo(): Promise<Array<{\n    url: string;\n    title: string;\n    isOurs: boolean;\n    modelId: string | null;\n    memory: { usedGB: number; totalGB: number } | null;\n  }> | null> {\n    if (typeof window !== \"undefined\") {\n      return null;\n    }\n\n    try {\n      const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n      return await ChromeGPUBackend.getAllChromePages();\n    } catch {\n      return null;\n    }\n  }\n\n  /**\n   * Kill a Chrome page by index (works across processes)\n   * @param index Index of the page to kill (0-based)\n   */\n  static async killChromePage(index: number): Promise<boolean> {\n    if (typeof window !== \"undefined\") {\n      return false;\n    }\n\n    try {\n      const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n      return await ChromeGPUBackend.killPageByIndex(index);\n    } catch {\n      return false;\n    }\n  }\n\n  /**\n   * Get total Chrome page count (all processes)\n   */\n  static async getTotalChromePageCount(): Promise<number> {\n    if (typeof window !== \"undefined\") {\n      return 0;\n    }\n\n    try {\n      const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n      return await ChromeGPUBackend.getTotalPageCount();\n    } catch {\n      return 0;\n    }\n  }\n\n  // ============================================\n  // Private Methods\n  // ============================================\n\n  private formatPrompt(prompt: string, options: GenerateOptions): string {\n    const system = options.system || \"You are a helpful assistant.\";\n    const isQwen = this.currentModel?.includes(\"qwen\");\n\n    if (options.thinking && this.modelConfig?.supportsThinking) {\n      const thinkSystem = `${system}\\n\\nThink step-by-step before answering. Wrap your reasoning in <think></think> tags, then provide your answer.`;\n      return `<|im_start|>system\\n${thinkSystem}<|im_end|>\\n<|im_start|>user\\n${prompt}<|im_end|>\\n<|im_start|>assistant\\n`;\n    }\n\n    if (isQwen) {\n      return `<|im_start|>system\\n${system}<|im_end|>\\n<|im_start|>user\\n${prompt} /no_think<|im_end|>\\n<|im_start|>assistant\\n`;\n    }\n\n    return `<|im_start|>system\\n${system}<|im_end|>\\n<|im_start|>user\\n${prompt}<|im_end|>\\n<|im_start|>assistant\\n`;\n  }\n\n  private buildMessages(\n    prompt: string,\n    options: GenerateOptions,\n  ): Array<{ role: string; content: string }> {\n    const system = options.system || \"You are a helpful assistant.\";\n    const messages: Array<{ role: string; content: string }> = [];\n\n    // For direct model (WebGPU), enable_thinking is passed to apply_chat_template\n    // so we don't need to add /no_think or modify the system prompt\n    messages.push({ role: \"system\", content: system });\n    messages.push({ role: \"user\", content: prompt });\n\n    return messages;\n  }\n\n  private parseThinking(text: string): {\n    thinking?: string;\n    response: string;\n  } {\n    // Handle complete <think>...</think> blocks\n    const match = text.match(/<think>([\\s\\S]*?)<\\/think>/);\n    if (match) {\n      const thinking = match[1].trim();\n      const response = text.replace(/<think>[\\s\\S]*?<\\/think>/, \"\").trim();\n      return { thinking, response };\n    }\n\n    // Handle unclosed <think> tags (model stopped mid-thought)\n    const unclosedMatch = text.match(/<think>([\\s\\S]*)$/);\n    if (unclosedMatch) {\n      const thinking = unclosedMatch[1].trim();\n      const response = text.replace(/<think>[\\s\\S]*$/, \"\").trim();\n      return { thinking: thinking || undefined, response };\n    }\n\n    // Handle any remaining think tags\n    const response = text.replace(/<\\/?think>/g, \"\").trim();\n    return { response };\n  }\n\n  private cleanOutput(text: string): string {\n    return (\n      text\n        .replace(/<\\|im_end\\|>/g, \"\")\n        .replace(/<\\|im_start\\|>/g, \"\")\n        .replace(/<\\|endoftext\\|>/g, \"\")\n        .replace(/<\\/s>/g, \"\")\n        // Clean up artifacts from direct model output\n        .replace(/^\\/no_think\\s*/i, \"\")\n        .replace(/^assistant\\s*/i, \"\")\n        .replace(/^\\s*\\/no_think\\s*/gim, \"\")\n        .replace(/^\\s*assistant\\s*/gim, \"\")\n        // Clean up role markers that might appear\n        .replace(/^(system|user|assistant):\\s*/gim, \"\")\n        .trim()\n    );\n  }\n}\n\nexport default Gerbil;\n"],"mappings":";;;;;;;;AAkCA,SAAgB,iBACd,QACA,SACA,SAQQ;CACR,MAAM,WAAW;EACf;EACA;EACA,QAAQ,aAAa;EACrB,QAAQ,eAAe;EACvB,QAAQ,QAAQ;EAChB,QAAQ,QAAQ;EAChB,QAAQ,UAAU;EAClB,QAAQ,YAAY;EACrB;CAGD,MAAM,MAAM,KAAK,UAAU,SAAS;CACpC,IAAI,OAAO;AACX,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,MAAM,OAAO,IAAI,WAAW,EAAE;AAC9B,UAAQ,QAAQ,KAAK,OAAO;AAC5B,SAAO,OAAO;;AAEhB,QAAO,UAAU,KAAK,SAAS,GAAG;;;;;AAUpC,IAAa,gBAAb,MAA2B;CACzB,AAAQ,wBAAiC,IAAI,KAAK;CAClD,AAAQ;CACR,AAAQ;CACR,AAAQ,OAAO;CACf,AAAQ,SAAS;;;;;;CAOjB,YAAY,UAAU,KAAK,aAAa,MAAS,KAAM;AACrD,OAAK,UAAU;AACf,OAAK,aAAa;;;;;CAMpB,IAAI,KAAoC;EACtC,MAAM,QAAQ,KAAK,MAAM,IAAI,IAAI;AAEjC,MAAI,CAAC,OAAO;AACV,QAAK;AACL,UAAO;;AAIT,MAAI,KAAK,KAAK,GAAG,MAAM,YAAY,MAAM,KAAK;AAC5C,QAAK,MAAM,OAAO,IAAI;AACtB,QAAK;AACL,UAAO;;AAIT,OAAK,MAAM,OAAO,IAAI;AACtB,OAAK,MAAM,IAAI,KAAK,MAAM;AAE1B,OAAK;AACL,SAAO;GAAE,GAAG,MAAM;GAAQ,QAAQ;GAAM;;;;;CAM1C,IAAI,KAAa,QAAwB,KAAoB;AAE3D,SAAO,KAAK,MAAM,QAAQ,KAAK,SAAS;GACtC,MAAM,WAAW,KAAK,MAAM,MAAM,CAAC,MAAM,CAAC;AAC1C,OAAI,SACF,MAAK,MAAM,OAAO,SAAS;;AAI/B,OAAK,MAAM,IAAI,KAAK;GAClB;GACA,WAAW,KAAK,KAAK;GACrB,KAAK,OAAO,KAAK;GAClB,CAAC;;;;;CAMJ,IAAI,KAAsB;EACxB,MAAM,QAAQ,KAAK,MAAM,IAAI,IAAI;AACjC,MAAI,CAAC,MAAO,QAAO;AAEnB,MAAI,KAAK,KAAK,GAAG,MAAM,YAAY,MAAM,KAAK;AAC5C,QAAK,MAAM,OAAO,IAAI;AACtB,UAAO;;AAGT,SAAO;;;;;CAMT,OAAO,KAAsB;AAC3B,SAAO,KAAK,MAAM,OAAO,IAAI;;;;;CAM/B,QAAc;AACZ,OAAK,MAAM,OAAO;AAClB,OAAK,OAAO;AACZ,OAAK,SAAS;;;;;CAMhB,QAAgB;EACd,MAAM,MAAM,KAAK,KAAK;EACtB,IAAI,SAAS;AAEb,OAAK,MAAM,CAAC,KAAK,UAAU,KAAK,MAC9B,KAAI,MAAM,MAAM,YAAY,MAAM,KAAK;AACrC,QAAK,MAAM,OAAO,IAAI;AACtB;;AAIJ,SAAO;;;;;CAMT,WAAuB;AACrB,SAAO;GACL,MAAM,KAAK;GACX,QAAQ,KAAK;GACb,MAAM,KAAK,MAAM;GACjB,SAAS,KAAK;GACf;;;;;CAMH,aAAqB;EACnB,MAAM,QAAQ,KAAK,OAAO,KAAK;AAC/B,MAAI,UAAU,EAAG,QAAO;AACxB,SAAQ,KAAK,OAAO,QAAS;;;AAQjC,IAAIA,cAAoC;;;;;AAMxC,SAAgB,iBAAgC;AAC9C,KAAI,CAAC,YACH,eAAc,IAAI,eAAe;AAEnC,QAAO;;;;;AAMT,SAAgB,qBAAqB,SAAkB,YAAoC;AACzF,eAAc,IAAI,cAAc,SAAS,WAAW;AACpD,QAAO;;;;;AAMT,SAAgB,mBAAyB;AACvC,KAAI,YACF,aAAY,OAAO;;;;;ACjOvB,MAAaC,iBAA8C;CACzD,cAAc;EACZ,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,gBAAgB;EACd,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,sBAAsB;EACpB,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,gBAAgB;EACd,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,gBAAgB;EACd,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,cAAc;EACZ,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,gBAAgB;EACd,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,gBAAgB;EAChB,mBAAmB;EACnB,QAAQ;EACT;CACF;;;;;;;;;;AAeD,SAAgB,aAAa,SAA8B;AAEzD,KAAI,eAAe,SACjB,QAAO;EACL,MAAM;EACN,MAAM,eAAe,SAAS;EAC/B;AAIH,KAAI,QAAQ,WAAW,MAAM,CAE3B,QAAO;EACL,MAAM;EACN,MAHW,QAAQ,MAAM,EAAE;EAI5B;AAIH,KAAI,QAAQ,WAAW,0BAA0B,CAE/C,QAAO;EACL,MAAM;EACN,MAHW,QAAQ,QAAQ,2BAA2B,GAAG;EAI1D;AAIH,KAAI,QAAQ,WAAW,QAAQ,CAE7B,QAAO;EACL,MAAM;EACN,MAHW,QAAQ,MAAM,EAAE;EAI5B;AAIH,KAAI,QAAQ,SAAS,IAAI,CACvB,QAAO;EACL,MAAM;EACN,MAAM;EACP;AAIH,QAAO;EACL,MAAM;EACN,MAAM;EACP;;;;;AAMH,SAAgB,eAAe,SAAqC;AAClE,QAAO,eAAe,YAAY;;AAIpC,MAAMC,0BAAkD;CACtD,MAAM;CACN,SAAS;CACT,OAAO;CACP,KAAK;CACL,QAAQ;CACR,OAAO;CACR;;;;AAKD,SAAgB,0BACd,SACA,MACA,eACa;CAEb,IAAIC,SAAgC;CACpC,MAAM,YAAY,KAAK,aAAa;AAEpC,KAAI,UAAU,SAAS,OAAO,CAC5B,UAAS;UACA,UAAU,SAAS,SAAS,CACrC,UAAS;UACA,UAAU,SAAS,MAAM,CAClC,UAAS;UACA,UAAU,SAAS,UAAU,IAAI,UAAU,SAAS,YAAY,CACzE,UAAS;UACA,UAAU,SAAS,QAAQ,CACpC,UAAS;CAIX,MAAM,iBACJ,UAAU,SAAS,SAAS,IAC5B,UAAU,SAAS,MAAM,IACzB,UAAU,SAAS,aAAa,IAChC,UAAU,SAAS,YAAY;AAEjC,QAAO;EACL,IAAI;EACJ;EACA,aAAa,mBAAmB;EAChC,MAAM;EACN,eAAe,iBAAiB,wBAAwB,WAAW;EACnE,kBAAkB,WAAW,UAAU,WAAW;EAClD,cAAc,WAAW,UAAU,WAAW,SAAS,WAAW;EAClE;EACA;EACD;;;;;AAMH,eAAsB,wBAAwB,MAAsC;AAClF,KAAI;EACF,MAAM,MAAM,MAAM,MAAM,0BAA0B,KAAK,uBAAuB;AAC9E,MAAI,CAAC,IAAI,GACP,QAAO;EAGT,MAAM,SAAS,MAAM,IAAI,MAAM;AAG/B,SACE,OAAO,2BACP,OAAO,eACP,OAAO,eACP,OAAO,kBACP,OAAO,kBACP;SAEI;AACN,SAAO;;;;;;AAOX,SAAgB,oBAAmC;AACjD,QAAO,OAAO,OAAO,eAAe;;;;;;;;AC/NtC,MAAMC,aAAWC;AAGjB,SAAS,sBAAyB,IAAkC;CAClE,MAAM,eAAe,QAAQ;AAC7B,SAAQ,QAAQ,GAAG,SAAgB;EACjC,MAAM,MAAM,KAAK,IAAI,YAAY,IAAI;AAErC,MAAI,IAAI,SAAS,iBAAiB,IAAI,IAAI,SAAS,sBAAsB,CACvE;AAEF,eAAa,MAAM,SAAS,KAAK;;AAGnC,QAAO,IAAI,CAAC,cAAc;AACxB,UAAQ,OAAO;GACf;;AA4CJ,MAAMC,wBAAqC;CACzC;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACD;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACD;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACD;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACD;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACD;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACD;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACD;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACD;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACD;EACE,IAAI;EACJ,MAAM;EACN,QAAQ;EACR,UAAU;EACV,aAAa;EACd;CACF;AAKD,MAAM,YAAY,OAAO,WAAW;AACpC,IAAI,mBAAmB,CAAC;AACxB,IAAI,kBAAkB;AAOtB,IAAI,oBAAoB;AACxB,IAAI,kBAAkB;;;;;AAMtB,eAAe,iBAAmC;AAChD,KAAI,kBACF,QAAO;AAET,qBAAoB;AAGpB,KAAI,OAAO,WAAW,aAAa;AACjC,oBAAkB,SAAS;AAC3B,SAAO;;AAKT,KAAI;EAGF,MAAM,EAAE,QAAQ,YADK,MADC,IAAI,SAAS,aAAa,2BAA2B,CAClC,SAAS;AAIlD,SAAO,OAAO,YAAY,QAAQ;AAGlC,MAAI,CAAE,WAAmB,UACvB,CAAC,WAAmB,YAAY,EAAE;AAEpC,EAAC,WAAmB,UAAU,MAAM,OAAO,EAAE,CAAC;AAE9C,oBAAkB;SACZ;AAEN,oBAAkB;;AAGpB,QAAO;;AAMT,IAAa,SAAb,MAAoB;CAClB,AAAQ,YAA2C;CACnD,AAAQ,YAAwC;CAChD,AAAQ,QAAa;CACrB,AAAQ,WAA6C;CACrD,AAAQ,eAA8B;CACtC,AAAQ,cAAkC;CAC1C,AAAiB;CACjB,AAAQ;CACR,AAAQ,YAAY;CACpB,AAAQ,gBAA6C;CACrD,AAAQ,cAAyC;CAGjD,AAAQ,YAAiB;CACzB,AAAQ,cAAmB;CAC3B,AAAQ,gBAAgB;CAExB,YAAY,SAAuB,EAAE,EAAE;AACrC,OAAK,SAAS;AACd,OAAK,QAAQ;GACX,SAAS;GACT,UAAU;GACV,WAAW;GACX,UAAU;GACV,WAAW;GACX,WAAW;GACX,aAAa;GACd;;CAOH,OAAO,aAA4B;AACjC,SAAO,OAAO,OAAO,eAAe;;CAGtC,OAAO,SAAS,SAA0C;AACxD,SAAO,eAAe;;;;;;;;;;;;;;;;;;;;CAyBxB,MAAM,UAAU,UAAU,cAAc,UAAuB,EAAE,EAAiB;AAGhF,MAAI,KAAK,UAAU,CACjB,OAAM,KAAK,SAAS;AAItB,QAAM,gBAAgB;EAEtB,MAAM,SAAS,aAAa,QAAQ;EACpC,MAAM,EAAE,YAAY,SAAS,QAAQ,OAAO,cAAc;EAG1D,IAAI,SAAS,eAAe,QAAQ;AACpC,MAAI,CAAC,QAAQ;GAEX,MAAM,gBAAgB,MAAM,wBAAwB,OAAO,KAAK,CAAC,YAAY,KAAK;AAClF,YAAS,0BAA0B,SAAS,OAAO,MAAM,iBAAiB,OAAU;;AAItF,MAAI,OAAO,eACT,QAAO,KAAK,gBAAgB,SAAS,OAAO,MAAM,QAAQ,QAAQ;AAGpE,eAAa,EAAE,QAAQ,WAAW,QAAQ,MAAM,CAAC;EAKjD,MAAMC,cAAY,OAAO,WAAW;EACpC,MAAM,iBAAiBA,cAAY,SAAS;EAC5C,IAAIC,WAAsC;AAC1C,MAAI,WAAW,YAAY,WAAW,SAAS,WAAW,OACxD,YAAW;EAIb,MAAM,QAAQ,cAAc,aAAa,WAAW,UAAU;EAG9D,IAAI,YAAY;EAChB,IAAI,WAAW;EACf,IAAI,UAAU;EAEd,MAAM,oBAAoB,aAAkB;AAC1C,OAAI,CAAC,UACH;AAGF,OAAI,SAAS,WAAW,cAAc,SAAS,MAAM;IACnD,MAAM,MAAM,KAAK,MAAM,SAAS,YAAY,EAAE;AAE9C,QAAI,SAAS,SAAS,YAAY,OAAO,UAAU,GAAG;AACpD,gBAAW,SAAS;AACpB,eAAU;AACV,kBAAa;MACX,QAAQ,eAAe,SAAS;MAChC,UAAU;MACV,MAAM,SAAS;MAChB,CAAC;;;;AAKR,MAAI;AAGF,OAAID,eAAa,aAAa,UAAU;AACtC,iBAAa,EAAE,QAAQ,wBAAwB,CAAC;AAChD,SAAK,YAAa,MAAM,4BACtB,cAAc,gBAAgB,OAAO,MAAM,EACzC,mBAAmB,kBACpB,CAAC,CACH;AAED,iBAAa,EAAE,QAAQ,oBAAoB,CAAC;AAC5C,SAAK,QAAQ,MAAM,4BACjB,qBAAqB,gBAAgB,OAAO,MAAM;KAChD;KACA,QAAQ;KACR,mBAAmB;KACpB,CAAC,CACH;AAED,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,SAAK,gBAAgB;AACrB,gBAAY;AACZ,SAAK,eAAe;AACpB,SAAK,cAAc;AACnB,iBAAa,EAAE,QAAQ,mBAAmB,CAAC;cAClC,CAACA,eAAa,aAAa,UAAU;AAE9C,iBAAa,EAAE,QAAQ,qCAAqC,CAAC;IAG7D,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAC1C,SAAK,gBAAgB,MAAM,iBAAiB,OAAO;KACjD,SAAS,OAAO;KAChB,eAAe,OAAO;KACtB;KACD,CAAC;AAEF,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,SAAK,gBAAgB;AACrB,gBAAY;AACZ,SAAK,eAAe;AACpB,SAAK,cAAc;UAEd;IAEL,MAAM,kBAAkB;KACtB;KACA,QAAQ;KACR,mBAAmB;KACpB;AACD,SAAK,YAAa,MAAM,4BACtBH,WAAS,mBAAmB,OAAO,MAAM,gBAAuB,CACjE;AAED,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,SAAK,gBAAgB;AACrB,gBAAY;AACZ,SAAK,eAAe;AACpB,SAAK,cAAc;AACnB,iBAAa,EAAE,QAAQ,UAAU,SAAS,aAAa,CAAC,KAAK,CAAC;;WAEzD,KAAK;AAEZ,OAAI,aAAa,gBAAgB;AAC/B,iBAAa,EAAE,QAAQ,SAAS,eAAe,aAAa,CAAC,MAAM,CAAC;AAGpE,QAAI,KAAK,eAAe;AACtB,WAAM,KAAK,cAAc,SAAS;AAClC,UAAK,gBAAgB;;AAIvB,SAAK,YAAa,MAAM,4BACtBA,WAAS,mBAAmB,OAAO,MAAM;KACvC,OAAO;KACP,QAAQ;KACR,mBAAmB;KACpB,CAAQ,CACV;AAED,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,SAAK,gBAAgB;AACrB,gBAAY;AACZ,SAAK,eAAe;AACpB,SAAK,cAAc;AACnB,iBAAa,EAAE,QAAQ,UAAU,eAAe,aAAa,CAAC,KAAK,CAAC;SAEpE,OAAM;;;;;;;CASZ,MAAc,gBACZ,SACA,UACA,QACA,UAAuB,EAAE,EACV;EACf,MAAM,EAAE,YAAY,SAAS,WAAW;AAExC,eAAa,EAAE,QAAQ,WAAW,QAAQ,qBAAqB,CAAC;EAEhE,MAAMG,cAAY,OAAO,WAAW;EACpC,MAAM,iBAAiBA,cAAY,SAAS;EAC5C,IAAIC,WAAsC;AAC1C,MAAI,WAAW,YAAY,WAAW,SAAS,WAAW,OACxD,YAAW;AAIb,MAAI,CAACD,eAAa,aAAa,UAAU;AACvC,gBAAa,EAAE,QAAQ,8CAA8C,CAAC;GAGtE,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAC1C,QAAK,gBAAgB,MAAM,iBAAiB,OAAO;IACjD,SAAS;IACT,eAAe,OAAO;IACtB,UAAU;IACV;IACD,CAAC;AAEF,QAAK,YAAY;AACjB,QAAK,cAAc;AACnB,QAAK,gBAAgB;AACrB,QAAK,eAAe;AACpB,QAAK,cAAc;AAEnB;;EAIF,IAAI,WAAW;EACf,IAAI,UAAU;EAEd,MAAM,oBAAoB,aAAkB;AAC1C,OAAI,SAAS,WAAW,cAAc,SAAS,MAAM;IACnD,MAAM,MAAM,KAAK,MAAM,SAAS,YAAY,EAAE;AAC9C,QAAI,SAAS,SAAS,YAAY,OAAO,UAAU,GAAG;AACpD,gBAAW,SAAS;AACpB,eAAU;AACV,kBAAa;MACX,QAAQ,eAAe,SAAS;MAChC,UAAU;MACV,MAAM,SAAS;MAChB,CAAC;;;;AAKR,MAAI;AAEF,gBAAa,EAAE,QAAQ,wBAAwB,CAAC;AAChD,QAAK,YAAY,MAAM,4BACrB,cAAc,gBAAgB,UAAU,EACtC,mBAAmB,kBACpB,CAAC,CACH;AAGD,gBAAa,EAAE,QAAQ,2BAA2B,CAAC;AACnD,QAAK,cAAc,MAAM,4BACvB,4BAA4B,gBAAgB,UAAU;IACpD,QAAQ;IACR,mBAAmB;IACpB,CAAC,CACH;AAED,QAAK,gBAAgB;AACrB,QAAK,YAAY;AACjB,QAAK,cAAc,aAAa,WAAW,WAAY;AACvD,QAAK,eAAe;AACpB,QAAK,cAAc;AACnB,gBAAa,EAAE,QAAQ,kBAAkB,SAAS,aAAa,CAAC,KAAK,CAAC;WAC/D,KAAK;AAEZ,OAAI,aAAa,gBAAgB;AAC/B,iBAAa,EAAE,QAAQ,uBAAuB,eAAe,aAAa,CAAC,MAAM,CAAC;AAElF,SAAK,YAAY,MAAM,4BACrB,cAAc,gBAAgB,UAAU,EACtC,mBAAmB,kBACpB,CAAC,CACH;AAED,SAAK,cAAc,MAAM,4BACvB,4BAA4B,gBAAgB,UAAU;KACpD,QAAQ;KACR,mBAAmB;KACpB,CAAC,CACH;AAED,SAAK,gBAAgB;AACrB,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,SAAK,eAAe;AACpB,SAAK,cAAc;AACnB,iBAAa,EAAE,QAAQ,kBAAkB,eAAe,aAAa,CAAC,KAAK,CAAC;SAE5E,OAAM;;;;;;CAQZ,WAAoB;AAClB,SACE,KAAK,cAAc,QAClB,KAAK,aAAa,KAAK,UAAU,QAClC,KAAK,kBAAkB,QACtB,KAAK,iBAAiB,KAAK,gBAAgB;;;;;CAOhD,iBAA0B;AACxB,SAAO,KAAK,iBAAiB,KAAK,aAAa,mBAAmB;;;;;CAMpE,eAAmC;AACjC,SAAO,KAAK;;;;;CAMd,gBAA2C;AACzC,SAAO,KAAK;;;;;CAMd,WAAmB;AAEjB,SAAO,KAAK,gBAAgB,WAAW,UAAU;;;;;CAMnD,wBAAyF;EACvF,MAAM,QAAQ,gBAAgB;EAC9B,MAAM,QAAQ,MAAM,UAAU;AAC9B,SAAO;GACL,MAAM,MAAM;GACZ,QAAQ,MAAM;GACd,MAAM,MAAM;GACZ,SAAS,MAAM,YAAY;GAC5B;;;;;CAMH,qBAA2B;AACzB,kBAAgB,CAAC,OAAO;;;;;CAM1B,kBAKS;AACP,MAAI,CAAC,KAAK,cACR,QAAO;AAET,SAAO,KAAK,cAAc,WAAW;;;;;;CAOvC,MAAM,kBAA+E;AACnF,MAAI,CAAC,KAAK,cACR,QAAO;AAET,SAAO,KAAK,cAAc,gBAAgB;;;;;CAM5C,MAAM,iBAA2F;AAC/F,MAAI,CAAC,KAAK,cACR,QAAO;AAET,SAAO,KAAK,cAAc,gBAAgB;;;;;;CAO5C,MAAM,aAA4B;AAChC,MAAI,KAAK,cACP,OAAM,KAAK,cAAc,OAAO;;;;;;;CASpC,MAAM,sBAAsB,cAAc,GAAqB;AAC7D,MAAI,CAAC,KAAK,cACR,QAAO;AAET,SAAO,KAAK,cAAc,sBAAsB,YAAY;;;;;;;;;;;;;;;;CAqB9D,MAAM,SAAS,QAAgB,UAA2B,EAAE,EAA2B;AACrF,MAAI,CAAC,KAAK,UAAU,CAElB,OAAM,KAAK,UAAU,KAAK,OAAO,SAAS,aAAa;EAGzD,MAAM,EAAE,WAAW;AAMnB,MAAI,QAAQ,UAAU,KAAK,iBAAiB,CAAC,KAAK,cAChD,QAAO,KAAK,mBAAmB,QAAQ,QAAQ;AAIjD,MAAI,QAAQ,UAAU,CAAC,KAAK,eAAe;EAG3C,MAAM,EACJ,YAAY,KACZ,cAAc,IACd,OAAO,IACP,OAAO,IACP,WAAW,OACX,QACA,QAAQ,OACR,aACE;AAGJ,MAAI,SAAS,CAAC,QAAQ,WAAW,CAAC,QAAQ,QAAQ;GAChD,MAAM,WAAW,iBAAiB,QAAQ,KAAK,gBAAgB,IAAI;IACjE;IACA;IACA;IACA;IACA;IACA;IACD,CAAC;GACF,MAAM,SAAS,gBAAgB,CAAC,IAAI,SAAS;AAC7C,OAAI,OACF,QAAO;;EAIX,MAAM,YAAY,YAAY,KAAK;AAEnC,MAAI;GACF,IAAI,UAAU;AAEd,OAAI,KAAK,cAEP,KAAI;AACF,cAAU,MAAM,KAAK,cAAc,SAAS,QAAQ;KAClD;KACA;KACA;KACA;KACA;KACA;KAEA,QAAQ,QAAQ,KAAK,QAAQ,IAAI,OAAO;KAExC,SAAS,QAAQ,WAAW,MAAM,QAAQ,UAAU,EAAE,KAAK,GAAG;KAC/D,CAAC;YACKE,WAAgB;AAEvB,QAAI,WAAW,YAAY,yBAAyB,CAAC,KAAK,eAAe,SAAS,EAAE;AAClF,WAAM,KAAK,eAAe,SAAS,CAAC,YAAY,GAAG;AACnD,UAAK,gBAAgB;AACrB,UAAK,cAAc;AAGnB,UAAK,YAAa,MAAML,WAAS,mBADf,KAAK,gBAAgB,cACwB;MAC7D,OAAO;MACP,QAAQ;MACT,CAAQ;AAET,YAAO,KAAK,SAAS,QAAQ,QAAQ;;AAEvC,UAAM;;YAEC,KAAK,aAAa,KAAK,SAAS,KAAK,WAAW;IAEzD,MAAM,WAAW,KAAK,cAAc,QAAQ;KAAE,GAAG;KAAS;KAAU,CAAC;IAErE,MAAM,SAAU,KAAK,UAAkB,oBAAoB,UAAU;KACnE,uBAAuB;KACvB,aAAa;KACb,iBAAiB;KAClB,CAAC;IAEF,MAAM,SAAS,MAAM,KAAK,MAAM,SAAS;KACvC,GAAG;KACH,gBAAgB;KAChB,aAAa,cAAc,IAAI,cAAc;KAC7C,OAAO;KACP,OAAO;KACP,WAAW,cAAc;KAC1B,CAAC;IAGF,MAAM,cAAc,OAAO,UAAU,OAAO,MAAM,OAAO,UAAU,MAAM,UAAU;IAGnF,MAAM,eAAe,OAAO,MAAM,MAAM,CAAC,aAAa,KAAK,CAAC;AAK5D,cAJgB,KAAK,UAAU,aAAa,cAAc,EACxD,qBAAqB,MACtB,CAAC,CAEgB,MAAM;AAGxB,QAAI,QAAQ,aAAa,CAAC,SAAS,YAAY,EAAE;KAC/C,MAAM,QAAQ,QAAQ,MAAM,4BAA4B;AACxD,SAAI,MACF,WAAU,MAAM,GAAG,MAAM;;cAGpB,KAAK,WAAW;IAEzB,MAAM,kBAAkB,KAAK,aAAa,QAAQ;KAAE,GAAG;KAAS;KAAU,CAAC;IAE3E,MAAM,SAAS,MAAM,KAAK,UAAU,iBAAiB;KACnD,gBAAgB;KAChB;KACA,OAAO;KACP,OAAO;KACP,WAAW,cAAc;KACzB,kBAAkB;KACnB,CAAC;AAGF,QAAI,MAAM,QAAQ,OAAO,IAAI,OAAO,IAAI;KACtC,MAAMM,WAAS,OAAO;AACtB,SAAI,MAAM,QAAQA,SAAO,eAAe,CAEtC,WADaA,SAAO,eAAe,GAAG,GAAG,EACzB,WAAW;SAE3B,WAAUA,SAAO,kBAAkB;;SAIvC,OAAM,IAAI,MAAM,kBAAkB;GAIpC,MAAM,YADU,YAAY,KAAK,GACL;AAE5B,aAAU,KAAK,YAAY,QAAQ;GAInC,MAAM,EAAE,UAAU,cAAc,aAAa,KAAK,cAAc,QAAQ;GAGxE,MAAM,gBAAgB,WAAW,eAAe;GAEhD,MAAM,kBAAkB,KAAK,KAAK,SAAS,SAAS,EAAE;AAGtD,QAAK,MAAM,WAAW;AACtB,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,WAAY,KAAK,MAAM,YAAY,KAAK,MAAM,YAAa;GAEtE,MAAMC,SAAyB;IAC7B,MAAM;IACN,UAAU;IACV;IACA,iBAAkB,kBAAkB,YAAa;IACjD;IACA,cAAc;IACd,UAAU;IACV,QAAQ;IACT;AAGD,OAAI,SAAS,CAAC,QAAQ,WAAW,CAAC,QAAQ,QAAQ;IAChD,MAAM,WAAW,iBAAiB,QAAQ,KAAK,gBAAgB,IAAI;KACjE;KACA;KACA;KACA;KACA;KACA;KACD,CAAC;AACF,oBAAgB,CAAC,IAAI,UAAU,QAAQ,SAAS;;AAGlD,UAAO;WACA,QAAQ;AACf,UAAO;IACL,MAAM;IACN,iBAAiB;IACjB,iBAAiB;IACjB,WAAW,YAAY,KAAK,GAAG;IAC/B,cAAc;IACd,UAAU;IACV,QAAQ;IACT;;;;;;;;;CAUL,OAAO,OACL,QACA,UAA2B,EAAE,EACoB;AACjD,MAAI,CAAC,KAAK,UAAU,CAClB,OAAM,KAAK,UAAU,KAAK,OAAO,SAAS,aAAa;EAGzD,MAAM,YAAY,YAAY,KAAK;AAGnC,MAAI,KAAK,eAAe;GACtB,IAAI,WAAW;GACf,MAAMC,aAAuB,EAAE;GAC/B,IAAIC,cAAuD;GAC3D,IAAI,OAAO;GAGX,MAAM,kBAAkB,KAAK,cAC1B,SAAS,QAAQ;IAChB,GAAG;IAEH,QAAQ,QAAQ,QAAQ,KAAK,QAAQ,IAAI,OAAO;IAChD,UAAU,UAAU;AAClB,iBAAY,MAAM;AAClB,SAAI,aAAa;AACf,kBAAY,MAAM,KAAK;AACvB,oBAAc;WAEd,YAAW,KAAK,MAAM,KAAK;;IAGhC,CAAC,CACD,WAAW;AACV,WAAO;AACP,QAAI,YACF,aAAY,KAAK;KAEnB,CACD,OAAO,QAAQ;AACd,WAAO;AACP,QAAI,YACF,aAAY,KAAK;AAEnB,UAAM;KACN;AAGJ,UAAO,CAAC,QAAQ,WAAW,SAAS,EAClC,KAAI,WAAW,SAAS,GAAG;IACzB,MAAM,QAAQ,WAAW,OAAO;AAChC,UAAM;AACN,YAAQ,UAAU,MAAM;cACf,CAAC,MAAM;IAChB,MAAM,QAAQ,MAAM,IAAI,SAAwB,YAAY;AAC1D,mBAAc;MACd;AACF,QAAI,OAAO;AACT,WAAM;AACN,aAAQ,UAAU,MAAM;;;AAK9B,SAAM;GAEN,MAAM,EAAE,UAAU,cAAc,aAAa,KAAK,cAAc,SAAS;GACzE,MAAM,kBAAkB,KAAK,KAAK,SAAS,SAAS,EAAE;GACtD,MAAM,YAAY,YAAY,KAAK,GAAG;AAEtC,UAAO;IACL,MAAM;IACN,UAAU,QAAQ,WAAW,eAAe;IAC5C;IACA;IACA,iBAAkB,kBAAkB,YAAa;IACjD,cAAc;IACf;;EAIH,MAAM,SAAS,MAAM,KAAK,YAAY,QAAQ,QAAQ;EAItD,MAAM,QAAQ,OAAO,QAAQ,MAAM,QAAQ;AAC3C,OAAK,MAAM,QAAQ,MACjB,KAAI,MAAM;AACR,SAAM;AACN,WAAQ,UAAU,KAAK;;AAI3B,SAAO,OAAO;;;;;CAMhB,MAAc,YACZ,QACA,UAA2B,EAAE,EACyB;EACtD,MAAM,EAAE,YAAY,KAAK,cAAc,IAAK,OAAO,IAAK,OAAO,IAAI,WAAW,UAAU;EAExF,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,kBAAkB,KAAK,aAAa,QAAQ;GAAE,GAAG;GAAS;GAAU,CAAC;AAE3E,MAAI;GACF,MAAM,SAAS,MAAM,KAAK,YAAY,iBAAiB;IACrD,gBAAgB;IAChB;IACA,OAAO;IACP,OAAO;IACP,WAAW,cAAc;IACzB,kBAAkB;IACnB,CAAC;GAGF,MAAM,YADU,YAAY,KAAK,GACL;GAG5B,IAAI,UAAU;AACd,OAAI,MAAM,QAAQ,OAAO,IAAI,OAAO,IAAI;IACtC,MAAM,SAAS,OAAO;AACtB,QAAI,MAAM,QAAQ,OAAO,eAAe,CAEtC,WADa,OAAO,eAAe,GAAG,GAAG,EACzB,WAAW;QAE3B,WAAU,OAAO,kBAAkB;;AAIvC,aAAU,KAAK,YAAY,QAAQ;GACnC,MAAM,EAAE,UAAU,cAAc,aAAa,KAAK,cAAc,QAAQ;GACxE,MAAM,gBAAgB,WAAW,eAAe;GAChD,MAAM,kBAAkB,KAAK,KAAK,SAAS,SAAS,EAAE;AAGtD,QAAK,MAAM,WAAW;AACtB,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,WAAY,KAAK,MAAM,YAAY,KAAK,MAAM,YAAa;AAEtE,UAAO;IACL;IACA,QAAQ;KACN,MAAM;KACN,UAAU;KACV;KACA,iBAAkB,kBAAkB,YAAa;KACjD;KACA,cAAc;KACd,UAAU;KACV,QAAQ;KACT;IACF;WACM,QAAQ;AACf,UAAO;IACL,SAAS;IACT,QAAQ;KACN,MAAM;KACN,iBAAiB;KACjB,iBAAiB;KACjB,WAAW,YAAY,KAAK,GAAG;KAC/B,cAAc;KACd,UAAU;KACV,QAAQ;KACT;IACF;;;;;;;CAYL,MAAc,mBACZ,QACA,SACyB;AACzB,MAAI,EAAE,KAAK,aAAa,KAAK,aAC3B,OAAM,IAAI,MAAM,sDAAsD;EAGxE,MAAM,EACJ,SAAS,EAAE,EACX,YAAY,MACZ,cAAc,IACd,OAAO,IACP,OAAO,IACP,WACE;EAEJ,MAAM,YAAY,YAAY,KAAK;AAEnC,MAAI;GAEF,MAAMC,UAAkD,EAAE;AAG1D,QAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK,EACtC,SAAQ,KAAK,EAAE,MAAM,SAAS,CAAC;AAIjC,WAAQ,KAAK;IAAE,MAAM;IAAQ,MAAM;IAAQ,CAAC;GAE5C,MAAM,WAAW,CACf,GAAI,SAAS,CAAC;IAAE,MAAM;IAAU,SAAS;IAAQ,CAAC,GAAG,EAAE,EACvD;IAAE,MAAM;IAAQ;IAAS,CAC1B;GAGD,MAAM,aAAa,KAAK,UAAU,oBAAoB,SAAS;GAG/D,MAAM,eAAe,MAAM,QAAQ,IACjC,OAAO,IAAI,OAAO,QAAQ,MAAM,SAAS,QAAQ,IAAI,OAAO,CAAC,CAC9D;GAGD,MAAM,SAAS,MAAM,KAAK,UACxB,aAAa,WAAW,IAAI,aAAa,KAAK,cAC9C,YACA,EAAE,oBAAoB,OAAO,CAC9B;GAGD,IAAI,WAAW;GACf,MAAM,WAAW,QAAQ,UACrB,IAAI,aAAa,KAAK,UAAU,WAAW;IACzC,aAAa;IACb,qBAAqB;IACrB,oBAAoB,WAAiB;AACnC,iBAAYC;AACZ,aAAQ,UAAUA,OAAK;;IAE1B,CAAC,GACF;GAGJ,MAAM,UAAU,MAAM,KAAK,YAAY,SAAS;IAC9C,GAAG;IACH,gBAAgB;IAChB,aAAa,cAAc,IAAI,cAAc;IAC7C,OAAO;IACP,OAAO;IACP,WAAW,cAAc;IACzB,GAAI,WAAW,EAAE,UAAU,GAAG,EAAE;IACjC,CAAC;GAGF,MAAM,cAAc,OAAO,UAAU,MAAM,GAAG,GAAG,IAAI;GAKrD,MAAM,OAJU,KAAK,UAAU,aAAa,QAAQ,MAAM,MAAM,CAAC,aAAa,KAAK,CAAC,EAAE,EACpF,qBAAqB,MACtB,CAAC,CAEmB,MAAM,YAAY;GACvC,MAAM,YAAY,YAAY,KAAK,GAAG;GACtC,MAAM,kBAAkB,KAAK,KAAK,KAAK,SAAS,EAAE;AAGlD,QAAK,MAAM,WAAW;AACtB,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,WAAY,KAAK,MAAM,YAAY,KAAK,MAAM,YAAa;AAEtE,UAAO;IACL,MAAM,KAAK,YAAY,KAAK;IAC5B;IACA,iBAAkB,kBAAkB,YAAa;IACjD;IACA,cAAc;IACd,UAAU;IACV,QAAQ;IACT;WACM,QAAQ;AACf,UAAO;IACL,MAAM;IACN,iBAAiB;IACjB,iBAAiB;IACjB,WAAW,YAAY,KAAK,GAAG;IAC/B,cAAc;IACd,UAAU;IACV,QAAQ;IACT;;;;;;CAWL,MAAM,KAAQ,QAAgB,SAAqC;EACjE,MAAM,EAAE,QAAQ,UAAU,GAAG,cAAc,OAAQ;EAEnD,MAAM,eAAe;;wCAEe,KAAK,UAAU,gBAAgB,OAAO,CAAC;AAE3E,OAAK,IAAI,UAAU,GAAG,UAAU,SAAS,WAAW,GAAG;GACrD,MAAM,SAAS,MAAM,KAAK,SAAS,QAAQ;IACzC,QAAQ,QAAQ,UAAU;IAC1B;IACA,WAAW;IACZ,CAAC;AAEF,OAAI;IAEF,MAAM,UAAU,YAAY,OAAO,KAAK;IACxC,MAAM,SAAS,KAAK,MAAM,QAAQ;AAElC,WADkB,OAAO,MAAM,OAAO;YAE/B,OAAO;AACd,QAAI,YAAY,UAAU,EACxB,OAAM,IAAI,MAAM,uCAAuC,QAAQ,aAAa,QAAQ;;;AAK1F,QAAM,IAAI,MAAM,gCAAgC;;;;;CAUlD,MAAM,MAAM,MAAc,UAAwB,EAAE,EAAwB;AAC1E,MAAI,CAAC,KAAK,SAGR,MAAK,WAAY,MAAMX,WAAS,sBADlB,QAAQ,SAAS,0BAC6B;EAG9D,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,SAAS,MAAM,KAAK,SAAS,MAAM;GACvC,SAAS;GACT,WAAW,QAAQ,cAAc;GAClC,CAAC;AAIF,SAAO;GACL,QAHa,MAAM,KAAK,OAAO,KAAqB;GAIpD;GACA,WAAW,YAAY,KAAK,GAAG;GAChC;;;;;CAMH,MAAM,WAAW,OAAiB,UAAwB,EAAE,EAA0B;EACpF,MAAMY,UAAyB,EAAE;AACjC,OAAK,MAAM,QAAQ,MACjB,SAAQ,KAAK,MAAM,KAAK,MAAM,MAAM,QAAQ,CAAC;AAE/C,SAAO;;;;;CAUT,WAAyB;AACvB,SAAO,EAAE,GAAG,KAAK,OAAO;;;;;CAM1B,UAAsB;AACpB,SAAO;GACL,SAAS;GACT,OAAO,KAAK;GACZ,QAAQ;IACN,SAAS;IACT,KAAK;IACL,MAAM;IACN,QAAQ,KAAK,UAAU,GAAG,UAAU;IACrC;GACD,SAAS;IACP,KAAK,KAAK,aAAa,iBAAiB;IACxC,MAAM;IACN,WAAW,KAAK,aAAa,iBAAiB;IAC/C;GACD,OAAO;IACL,UAAU;IACV,MAAM;IACN,YAAY;IACb;GACF;;;;;CAMH,aAAmB;AACjB,OAAK,QAAQ;GACX,SAAS;GACT,UAAU;GACV,WAAW;GACX,UAAU;GACV,WAAW;GACX,WAAW;GACX,aAAa;GACd;;CAOH,AAAQ,MAA6B;CACrC,AAAQ,aAAqB;;;;;;;;;;;;;;;;CAiB7B,MAAM,QAAQ,UAA+C,EAAE,EAAiB;EAC9E,MAAM,UAAU,QAAQ,SAAS;AAGjC,MAAI,KAAK,OAAO,KAAK,eAAe,SAAS;AAC3C,SAAM,KAAK,IAAI,SAAS;AACxB,QAAK,MAAM;;AAGb,MAAI,KAAK,KAAK,UAAU,CACtB;AAGF,OAAK,aAAa;EAGlB,MAAM,EAAE,cAAc,MAAM,OAAO;AAEnC,MAAI,CAAC,KAAK,IACR,MAAK,MAAM,UAAU,QAAQ;AAG/B,QAAM,KAAK,IAAI,KAAK,QAAQ;;;;;CAM9B,MAAM,gBAAgB,SAAyC;AAC7D,MAAI,CAAC,KAAK,KAAK,UAAU,CACvB,OAAM,KAAK,QAAQ,QAAQ;;;;;;;;;;;;;CAe/B,MAAM,MAAM,MAAc,UAAwB,EAAE,EAAwB;AAC1E,QAAM,KAAK,gBAAgB,EAAE,YAAY,QAAQ,YAAY,CAAC;AAC9D,SAAO,KAAK,IAAK,MAAM,MAAM,QAAQ;;;;;;;;;;;;;;CAevC,OAAO,YACL,MACA,UAAwB,EAAE,EACwB;AAClD,QAAM,KAAK,gBAAgB,EAAE,YAAY,QAAQ,YAAY,CAAC;AAC9D,SAAO,OAAO,KAAK,IAAK,YAAY,MAAM,QAAQ;;;;;CAMpD,aAA0B;AACxB,MAAI,CAAC,KAAK,IAER,QAAO;AAET,SAAO,KAAK,IAAI,YAAY;;;;;CAM9B,cAAuB;AACrB,SAAO,KAAK,KAAK,UAAU,IAAI;;;;;CAMjC,kBAAqF;AACnF,MAAI,CAAC,KAAK,IACR,QAAO;AAET,SAAO;GACL,IAAI,KAAK;GACT,QAAQ,KAAK,IAAI,UAAU;GAC3B,QAAQ,KAAK,IAAI,UAAU,GAAG,KAAK,IAAI,eAAe,GAAG;GAC1D;;;;;CAMH,MAAM,gBAEJ;EACA,MAAM,EAAE,eAAe,MAAM,OAAO;AACpC,SAAO,OAAO,OAAO,WAAW,CAAC,KAAK,OAAO;GAC3C,IAAI,EAAE;GACN,aAAa,EAAE;GACf,YAAY,EAAE;GACd,YAAY,EAAE,OAAO;GACtB,EAAE;;CAOL,AAAQ,MAA6B;;;;;;;;;;;;;;CAerC,MAAM,QAAQ,SAAkB,UAA0B,EAAE,EAAiB;AAC3E,MAAI,KAAK,KAAK,UAAU,CACtB;EAIF,MAAM,EAAE,eAAe,MAAM,OAAO;AAEpC,MAAI,CAAC,KAAK,IACR,MAAK,MAAM,IAAI,WAAW,QAAQ;AAGpC,QAAM,KAAK,IAAI,KAAK,QAAQ;;;;;CAM9B,MAAa,gBAAgB,SAAkB,SAAyC;AACtF,MAAI,CAAC,KAAK,KAAK,UAAU,CACvB,OAAM,KAAK,QAAQ,SAAS,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;CA2BxC,MAAM,WACJ,OACA,UAA6B,EAAE,EACJ;AAC3B,QAAM,KAAK,gBAAgB,QAAW,EAAE,YAAY,QAAQ,YAAY,CAAC;AACzE,SAAO,KAAK,IAAK,WAAW,OAAO,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8B7C,MAAM,6BACJ,UAAyC,EAAE,EACH;AACxC,QAAM,KAAK,iBAAiB;AAC5B,SAAO,KAAK,IAAK,uBAAuB,QAAQ;;;;;CAMlD,MAAM,gBAA2C;EAE/C,MAAM,EAAE,eAAe,MAAM,OAAO;AACpC,SAAO,WAAW,YAAY;;;;;CAMhC,cAAuB;AACrB,SAAO,KAAK,KAAK,UAAU,IAAI;;;;;CAMjC,kBAAqF;AACnF,MAAI,CAAC,KAAK,IACR,QAAO;AAET,SAAO;GACL,IAAI,KAAK,IAAI,cAAc,CAAC;GAC5B,QAAQ,KAAK,IAAI,UAAU;GAC3B,QAAQ,KAAK,IAAI,UAAU,GAAG,KAAK,IAAI,eAAe,GAAG;GAC1D;;;;;;;;;;;;;;;;;CAsBH,MAAM,OACJ,aAAqB,KACrB,UAAqD,EAAE,EAC5B;EAE3B,MAAM,EAAE,YAAY,mBAAmB,MAAM,OAAO;AAEpD,MAAI,CAAC,gBAAgB,CACnB,OAAM,IAAI,MACR,uJAID;AAGH,UAAQ,aAAa,yBAAyB;EAE9C,MAAM,MAAM,IAAI,WAAW,EAAE,YAAY,MAAO,CAAC;AACjD,QAAM,IAAI,OAAO;AAEjB,UAAQ,aAAa,kBAAkB,aAAa,KAAM,QAAQ,EAAE,CAAC,MAAM;AAG3E,QAAM,IAAI,SAAS,MAAM,WAAW,GAAG,WAAW,CAAC;AAEnD,UAAQ,aAAa,sBAAsB;EAC3C,MAAM,EAAE,UAAU,MAAM,IAAI,MAAM;AAElC,UAAQ,aAAa,kBAAkB;AACvC,SAAO,KAAK,WAAW,OAAO,EAC5B,aAAa,MAAM,QAAQ,aAAa,EAAE,UAAU,kBAAkB,EACvE,CAAC;;;;;CAMJ,MAAM,wBAA0C;AAC9C,MAAI;GACF,MAAM,EAAE,mBAAmB,MAAM,OAAO;AACxC,UAAO,gBAAgB;UACjB;AACN,UAAO;;;;;;;CAYX,MAAM,QAAQ,aAAa,OAAsB;AAE/C,MAAI,KAAK,eAAe;AACtB,OAAI;AACF,UAAM,KAAK,cAAc,QAAQ,WAAW;WACtC;AAGR,QAAK,gBAAgB;;AAGvB,MAAI,KAAK,WAAW;AAClB,OAAI,OAAQ,KAAK,UAAkB,YAAY,WAC7C,KAAI;AACF,UAAO,KAAK,UAAkB,SAAS;WACjC;AAIV,QAAK,YAAY;;AAEnB,MAAI,KAAK,UAAU;AACjB,OAAI,OAAQ,KAAK,SAAiB,YAAY,WAC5C,KAAI;AACF,UAAO,KAAK,SAAiB,SAAS;WAChC;AAIV,QAAK,WAAW;;AAIlB,MAAI,KAAK,aAAa;AACpB,OAAI,OAAO,KAAK,YAAY,YAAY,WACtC,KAAI;AACF,UAAM,KAAK,YAAY,SAAS;WAC1B;AAIV,QAAK,cAAc;;AAErB,MAAI,KAAK,UACP,MAAK,YAAY;AAInB,MAAI,KAAK,KAAK;AACZ,OAAI;AACF,UAAM,KAAK,IAAI,SAAS;WAClB;AAGR,QAAK,MAAM;;AAIb,MAAI,KAAK,KAAK;AACZ,OAAI;AACF,SAAK,IAAI,SAAS;WACZ;AAGR,QAAK,MAAM;;AAGb,OAAK,eAAe;AACpB,OAAK,cAAc;AACnB,OAAK,gBAAgB;;;;;;;CAQvB,aAAa,WAA0B;EAErC,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAC1C,QAAM,iBAAiB,oBAAoB;;;;;;CAO7C,aAAa,qBAcH;AAER,MAAI,OAAO,WAAW,YACpB,QAAO;AAGT,MAAI;GACF,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAI1C,UAAO;IAAE,SAHO,iBAAiB,wBAAwB;IAGvC,UAFD,MAAM,iBAAiB,oBAAoB;IAEhC;UACtB;AACN,UAAO;;;;;;;CAQX,aAAa,gBAAiF;AAE5F,MAAI,OAAO,WAAW,YACpB,QAAO;AAGT,MAAI;GACF,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAC1C,UAAO,MAAM,iBAAiB,iBAAiB;UACzC;AACN,UAAO;;;;;;;CAQX,aAAa,kBAAkB,OAAiC;AAE9D,MAAI,OAAO,WAAW,YACpB,QAAO;AAGT,MAAI;GACF,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAC1C,UAAO,MAAM,iBAAiB,mBAAmB,MAAM;UACjD;AACN,UAAO;;;;;;;CAQX,aAAa,wBAMF;AACT,MAAI,OAAO,WAAW,YACpB,QAAO;AAGT,MAAI;GACF,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAC1C,UAAO,MAAM,iBAAiB,mBAAmB;UAC3C;AACN,UAAO;;;;;;;CAQX,aAAa,eAAe,OAAiC;AAC3D,MAAI,OAAO,WAAW,YACpB,QAAO;AAGT,MAAI;GACF,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAC1C,UAAO,MAAM,iBAAiB,gBAAgB,MAAM;UAC9C;AACN,UAAO;;;;;;CAOX,aAAa,0BAA2C;AACtD,MAAI,OAAO,WAAW,YACpB,QAAO;AAGT,MAAI;GACF,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAC1C,UAAO,MAAM,iBAAiB,mBAAmB;UAC3C;AACN,UAAO;;;CAQX,AAAQ,aAAa,QAAgB,SAAkC;EACrE,MAAM,SAAS,QAAQ,UAAU;EACjC,MAAM,SAAS,KAAK,cAAc,SAAS,OAAO;AAElD,MAAI,QAAQ,YAAY,KAAK,aAAa,iBAExC,QAAO,uBADa,GAAG,OAAO,iHACY,gCAAgC,OAAO;AAGnF,MAAI,OACF,QAAO,uBAAuB,OAAO,gCAAgC,OAAO;AAG9E,SAAO,uBAAuB,OAAO,gCAAgC,OAAO;;CAG9E,AAAQ,cACN,QACA,SAC0C;EAC1C,MAAM,SAAS,QAAQ,UAAU;EACjC,MAAMC,WAAqD,EAAE;AAI7D,WAAS,KAAK;GAAE,MAAM;GAAU,SAAS;GAAQ,CAAC;AAClD,WAAS,KAAK;GAAE,MAAM;GAAQ,SAAS;GAAQ,CAAC;AAEhD,SAAO;;CAGT,AAAQ,cAAc,MAGpB;EAEA,MAAM,QAAQ,KAAK,MAAM,6BAA6B;AACtD,MAAI,MAGF,QAAO;GAAE,UAFQ,MAAM,GAAG,MAAM;GAEb,UADF,KAAK,QAAQ,4BAA4B,GAAG,CAAC,MAAM;GACvC;EAI/B,MAAM,gBAAgB,KAAK,MAAM,oBAAoB;AACrD,MAAI,eAAe;GACjB,MAAM,WAAW,cAAc,GAAG,MAAM;GACxC,MAAM,WAAW,KAAK,QAAQ,mBAAmB,GAAG,CAAC,MAAM;AAC3D,UAAO;IAAE,UAAU,YAAY;IAAW;IAAU;;AAKtD,SAAO,EAAE,UADQ,KAAK,QAAQ,eAAe,GAAG,CAAC,MAAM,EACpC;;CAGrB,AAAQ,YAAY,MAAsB;AACxC,SACE,KACG,QAAQ,iBAAiB,GAAG,CAC5B,QAAQ,mBAAmB,GAAG,CAC9B,QAAQ,oBAAoB,GAAG,CAC/B,QAAQ,UAAU,GAAG,CAErB,QAAQ,mBAAmB,GAAG,CAC9B,QAAQ,kBAAkB,GAAG,CAC7B,QAAQ,wBAAwB,GAAG,CACnC,QAAQ,uBAAuB,GAAG,CAElC,QAAQ,mCAAmC,GAAG,CAC9C,MAAM"}

package/dist/gerbil-qOTe1nl2.d.mts

@@ -0,0 +1,431 @@
+import { C as SpeakResult, E as SystemInfo, O as TranscribeOptions, S as SpeakOptions, T as StreamingTranscriptionSession, b as STTModelConfig, c as GerbilConfig, f as JsonOptions, g as ModelConfig, h as LoadTTSOptions, i as EmbedResult, j as VoiceInfo, k as TranscribeResult, m as LoadSTTOptions, o as GenerateOptions, p as LoadOptions, r as EmbedOptions, s as GenerateResult, t as AudioChunk, w as StreamingTranscriptionOptions, x as SessionStats } from "./types-CiTc7ez3.mjs";
+
+//#region src/core/gerbil.d.ts
+
+declare class Gerbil {
+  private generator;
+  private tokenizer;
+  private model;
+  private embedder;
+  private currentModel;
+  private modelConfig;
+  private readonly config;
+  private stats;
+  private useDirect;
+  private chromeBackend;
+  private _deviceMode;
+  private processor;
+  private visionModel;
+  private isVisionModel;
+  constructor(config?: GerbilConfig);
+  static listModels(): ModelConfig[];
+  static getModel(modelId: string): ModelConfig | undefined;
+  /**
+   * Load a model
+   *
+   * @example
+   * ```ts
+   * // Built-in model
+   * await g.loadModel("qwen3-0.6b");
+   *
+   * // HuggingFace model
+   * await g.loadModel("hf:microsoft/Phi-3-mini");
+   *
+   * // Local model
+   * await g.loadModel("file:./models/my-model");
+   *
+   * // Vision model
+   * await g.loadModel("ministral-3b");
+   * ```
+   */
+  loadModel(modelId?: string, options?: LoadOptions): Promise<void>;
+  /**
+   * Load a vision model (VLM)
+   * Uses AutoProcessor + AutoModelForImageTextToText instead of tokenizer + causal LM
+   */
+  private loadVisionModel;
+  /**
+   * Check if a model is loaded
+   */
+  isLoaded(): boolean;
+  /**
+   * Check if current model supports vision
+   */
+  supportsVision(): boolean;
+  /**
+   * Get current model info
+   */
+  getModelInfo(): ModelConfig | null;
+  /**
+   * Get current device mode (webgpu, cpu, or wasm)
+   */
+  getDeviceMode(): "webgpu" | "cpu" | "wasm";
+  /**
+   * Get dtype used for current model
+   */
+  getDtype(): string;
+  /**
+   * Get response cache statistics
+   */
+  getResponseCacheStats(): {
+    hits: number;
+    misses: number;
+    size: number;
+    hitRate: number;
+  };
+  /**
+   * Clear the response cache (for cached generate() results)
+   */
+  clearResponseCache(): void;
+  /**
+   * Get Chrome backend status (if using WebGPU via Chrome)
+   */
+  getChromeStatus(): {
+    pid: number | null;
+    port: number;
+    modelId: string;
+    startedAt: Date | null;
+  } | null;
+  /**
+   * Get Chrome memory usage (if using WebGPU via Chrome)
+   * Returns JS heap memory in bytes
+   */
+  getChromeMemory(): Promise<{
+    jsHeapUsed: number;
+    jsHeapTotal: number;
+  } | null>;
+  /**
+   * Get memory usage in GB (if using WebGPU via Chrome)
+   */
+  getMemoryUsage(): Promise<{
+    usedGB: number;
+    totalGB: number;
+    usedPercent: number;
+  } | null>;
+  /**
+   * Clear KV cache to free memory
+   * This will reset the conversation context but free up memory
+   */
+  clearCache(): Promise<void>;
+  /**
+   * Check memory usage and cleanup if needed
+   * @param thresholdGB Memory threshold in GB (default: 8)
+   * @returns true if cleanup was performed
+   */
+  checkMemoryAndCleanup(thresholdGB?: number): Promise<boolean>;
+  /**
+   * Generate text (automatically routes to vision generation if images provided)
+   *
+   * @example
+   * ```ts
+   * // Text generation
+   * const result = await g.generate("Hello!");
+   *
+   * // Vision generation (with vision model)
+   * const result = await g.generate("What's in this image?", {
+   *   images: [{ source: "https://example.com/cat.jpg" }]
+   * });
+   * ```
+   */
+  generate(prompt: string, options?: GenerateOptions): Promise<GenerateResult>;
+  /**
+   * Stream text generation (simulated token-by-token)
+   *
+   * Note: Yields the raw output including <think> tags if thinking mode is enabled.
+   * The final result has parsed thinking separated out.
+   */
+  stream(prompt: string, options?: GenerateOptions): AsyncGenerator<string, GenerateResult, unknown>;
+  /**
+   * Internal: Generate with raw text access for streaming
+   */
+  private generateRaw;
+  /**
+   * Generate text from images using a vision model
+   * Called automatically by generate() when images are provided
+   */
+  private generateWithVision;
+  /**
+   * Generate structured JSON output
+   */
+  json<T>(prompt: string, options: JsonOptions<T>): Promise<T>;
+  /**
+   * Generate embeddings
+   */
+  embed(text: string, options?: EmbedOptions): Promise<EmbedResult>;
+  /**
+   * Generate embeddings for multiple texts
+   */
+  embedBatch(texts: string[], options?: EmbedOptions): Promise<EmbedResult[]>;
+  /**
+   * Get session stats
+   */
+  getStats(): SessionStats;
+  /**
+   * Get system info
+   */
+  getInfo(): SystemInfo;
+  /**
+   * Reset stats
+   */
+  resetStats(): void;
+  private tts;
+  private ttsModelId;
+  /**
+   * Load TTS model for text-to-speech synthesis
+   *
+   * @example
+   * ```ts
+   * // Load default (Kokoro)
+   * await g.loadTTS({ onProgress: (p) => console.log(p.status) });
+   *
+   * // Load Supertonic (faster, 44kHz output)
+   * await g.loadTTS({ model: "supertonic-66m" });
+   *
+   * const result = await g.speak("Hello world");
+   * // result.audio = Float32Array, result.sampleRate = 24000 or 44100
+   * ```
+   */
+  loadTTS(options?: LoadTTSOptions & {
+    model?: string;
+  }): Promise<void>;
+  /**
+   * Ensure TTS model is loaded (lazy loading)
+   */
+  ensureTTSLoaded(options?: LoadTTSOptions): Promise<void>;
+  /**
+   * Generate speech from text
+   *
+   * @example
+   * ```ts
+   * const result = await g.speak("Hello world", { voice: "af_bella" });
+   * // result.audio = Float32Array PCM
+   * // result.sampleRate = 24000
+   * // result.duration = seconds
+   * ```
+   */
+  speak(text: string, options?: SpeakOptions): Promise<SpeakResult>;
+  /**
+   * Stream speech generation (yields audio chunks as they're generated)
+   *
+   * @example
+   * ```ts
+   * for await (const chunk of g.speakStream("Long text...")) {
+   *   // chunk.samples = Float32Array
+   *   // chunk.isFinal = boolean
+   *   playChunk(chunk);
+   * }
+   * ```
+   */
+  speakStream(text: string, options?: SpeakOptions): AsyncGenerator<AudioChunk, SpeakResult, unknown>;
+  /**
+   * Get list of available TTS voices
+   */
+  listVoices(): VoiceInfo[];
+  /**
+   * Check if TTS model is loaded
+   */
+  isTTSLoaded(): boolean;
+  /**
+   * Get current TTS model info
+   */
+  getTTSModelInfo(): {
+    id: string;
+    loaded: boolean;
+    device?: "webgpu" | "cpu";
+  } | null;
+  /**
+   * List available TTS models
+   */
+  listTTSModels(): Promise<Array<{
+    id: string;
+    description: string;
+    sampleRate: number;
+    voiceCount: number;
+  }>>;
+  private stt;
+  /**
+   * Load STT model for speech-to-text transcription
+   *
+   * @example
+   * ```ts
+   * await g.loadSTT({
+   *   onProgress: (p) => console.log(p.status)
+   * });
+   *
+   * const result = await g.transcribe(audioData);
+   * console.log(result.text);
+   * ```
+   */
+  loadSTT(modelId?: string, options?: LoadSTTOptions): Promise<void>;
+  /**
+   * Ensure STT model is loaded (lazy loading)
+   */
+  ensureSTTLoaded(modelId?: string, options?: LoadSTTOptions): Promise<void>;
+  /**
+   * Transcribe audio to text
+   *
+   * @param audio - Audio data as Float32Array (16kHz mono) or Uint8Array (WAV file)
+   * @param options - Transcription options
+   *
+   * @example
+   * ```ts
+   * // From Float32Array (16kHz mono)
+   * const result = await g.transcribe(audioData);
+   * console.log(result.text);
+   *
+   * // With timestamps
+   * const result = await g.transcribe(audioData, { timestamps: true });
+   * for (const seg of result.segments) {
+   *   console.log(`[${seg.start}s] ${seg.text}`);
+   * }
+   *
+   * // From WAV file
+   * const wavData = fs.readFileSync("audio.wav");
+   * const result = await g.transcribe(new Uint8Array(wavData));
+   * ```
+   */
+  transcribe(audio: Float32Array | Uint8Array, options?: TranscribeOptions): Promise<TranscribeResult>;
+  /**
+   * Create a streaming transcription session
+   *
+   * Transcribes audio in real-time by processing chunks at regular intervals.
+   * Perfect for live captioning, call transcription, or real-time subtitles.
+   *
+   * @param options - Streaming options
+   * @returns Streaming session controller
+   *
+   * @example
+   * ```ts
+   * const session = await g.createStreamingTranscription({
+   *   chunkDuration: 3000,  // Transcribe every 3 seconds
+   *   onChunk: (text, idx) => console.log(`Chunk ${idx}: ${text}`),
+   *   onTranscript: (fullText) => console.log("Full:", fullText),
+   * });
+   *
+   * // Feed audio data as it comes in
+   * session.feedAudio(audioChunk);
+   *
+   * // Start automatic interval-based transcription
+   * session.start();
+   *
+   * // Later, stop and get final transcript
+   * const finalText = await session.stop();
+   * ```
+   */
+  createStreamingTranscription(options?: StreamingTranscriptionOptions): Promise<StreamingTranscriptionSession>;
+  /**
+   * Get list of available STT models
+   */
+  listSTTModels(): Promise<STTModelConfig[]>;
+  /**
+   * Check if STT model is loaded
+   */
+  isSTTLoaded(): boolean;
+  /**
+   * Get current STT model info
+   */
+  getSTTModelInfo(): {
+    id: string;
+    loaded: boolean;
+    device?: "webgpu" | "cpu";
+  } | null;
+  /**
+   * Record audio from microphone and transcribe
+   *
+   * @example
+   * ```ts
+   * // Record for 5 seconds and transcribe
+   * const result = await g.listen(5000);
+   * console.log(result.text);
+   *
+   * // Use with voice chat
+   * const userInput = await g.listen(10000);
+   * const response = await g.generate(userInput.text);
+   * await g.speak(response.text);
+   * ```
+   */
+  listen(durationMs?: number, options?: {
+    onProgress?: (status: string) => void;
+  }): Promise<TranscribeResult>;
+  /**
+   * Check if microphone recording is available
+   */
+  isMicrophoneAvailable(): Promise<boolean>;
+  /**
+   * Dispose of resources
+   * @param disconnect If true, also disconnect from shared browser (for clean script exit)
+   */
+  dispose(disconnect?: boolean): Promise<void>;
+  /**
+   * Shutdown the shared Chrome backend completely.
+   * Call this when your script/process is done to ensure proper cleanup.
+   * This closes the shared browser used for WebGPU acceleration.
+   */
+  static shutdown(): Promise<void>;
+  /**
+   * Get global WebGPU process info (all active backends)
+   * Useful for monitoring and debugging memory leaks
+   */
+  static getWebGPUProcesses(): Promise<{
+    browser: {
+      running: boolean;
+      pid: number | null;
+      port: number;
+      activePagesCount: number;
+      maxPages: number;
+    };
+    backends: Array<{
+      modelId: string;
+      isVision: boolean;
+      isReady: boolean;
+      memory: {
+        usedGB: number;
+        totalGB: number;
+        usedPercent: number;
+      } | null;
+    }>;
+  } | null>;
+  /**
+   * Kill all WebGPU processes (for zombie cleanup)
+   * Use this if you suspect memory leaks from undisposed Gerbil instances
+   */
+  static killAllWebGPU(): Promise<{
+    pagesKilled: number;
+    browserKilled: boolean;
+  } | null>;
+  /**
+   * Kill a specific WebGPU backend by index
+   * @param index Index of the backend to kill (0-based)
+   */
+  static killWebGPUBackend(index: number): Promise<boolean>;
+  /**
+   * Get all Chrome pages across ALL Gerbil processes
+   * This provides cross-process visibility into WebGPU backends
+   */
+  static getAllChromePagesInfo(): Promise<Array<{
+    url: string;
+    title: string;
+    isOurs: boolean;
+    modelId: string | null;
+    memory: {
+      usedGB: number;
+      totalGB: number;
+    } | null;
+  }> | null>;
+  /**
+   * Kill a Chrome page by index (works across processes)
+   * @param index Index of the page to kill (0-based)
+   */
+  static killChromePage(index: number): Promise<boolean>;
+  /**
+   * Get total Chrome page count (all processes)
+   */
+  static getTotalChromePageCount(): Promise<number>;
+  private formatPrompt;
+  private buildMessages;
+  private parseThinking;
+  private cleanOutput;
+}
+//#endregion
+export { Gerbil as t };
+//# sourceMappingURL=gerbil-qOTe1nl2.d.mts.map