@tryhamster/gerbil 1.0.3 → 1.0.5

package/dist/{gerbil-D4eIu8hx.mjs.map → gerbil-CZYYq4Sq.mjs.map}

@@ -1 +1 @@
-{"version":3,"file":"gerbil-D4eIu8hx.mjs","names":["globalCache: ResponseCache | null","BUILTIN_MODELS: Record<string, ModelConfig>","FAMILY_CONTEXT_DEFAULTS: Record<string, number>","family: ModelConfig[\"family\"]","idat: Uint8Array[]","v: number","KOKORO_VOICES_DEFAULT: VoiceInfo[]","concurrency: ConcurrencyConfig","result","result: GenerateResult","tokenQueue: string[]","resolveNext: ((value: string | null) => void) | null","results: EmbedResult[]"],"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// Canonical default model\n// ============================================\n\n/**\n * The default model used everywhere a model id is not explicitly provided\n * (CLI flags, REPL, framework adapters, integrations, one-liner). This is the\n * e2e-validated model; reference this constant instead of hard-coding the id.\n */\nexport const DEFAULT_MODEL = \"qwen3.5-0.8b\";\n\n// ============================================\n// Built-in Models (curated & tested)\n// ============================================\n\n// Every entry is a standard HuggingFace safetensors repo whose architecture the\n// native WebGPU engine supports (Qwen2/Qwen3/Qwen3.5, LFM2 — see\n// src/gpu/architectures/index.ts). The engine quantizes weights to INT4 on load;\n// `size` is the bf16/fp16 download size. Only add a repo here whose architecture\n// has a graph generator in the registry.\nexport const BUILTIN_MODELS: Record<string, ModelConfig> = {\n  \"qwen3.5-0.8b\": {\n    id: \"qwen3.5-0.8b\",\n    repo: \"Qwen/Qwen3.5-0.8B\",\n    description:\n      \"Qwen3.5 0.8B - Fast, multimodal (vision), 262k context, supports thinking (default)\",\n    size: \"~1.6GB\",\n    contextLength: 262_144,\n    supportsThinking: true,\n    supportsJson: true,\n    supportsVision: true,\n    family: \"qwen\",\n  },\n  \"qwen3.5-2b\": {\n    id: \"qwen3.5-2b\",\n    repo: \"Qwen/Qwen3.5-2B\",\n    description:\n      \"Qwen3.5 2B - Higher quality, multimodal (vision), 262k context, supports thinking\",\n    size: \"~4GB\",\n    contextLength: 262_144,\n    supportsThinking: true,\n    supportsJson: true,\n    supportsVision: true,\n    family: \"qwen\",\n  },\n  \"lfm2.5-1.2b-thinking\": {\n    id: \"lfm2.5-1.2b-thinking\",\n    repo: \"LiquidAI/LFM2.5-1.2B-Thinking\",\n    description: \"LFM2.5 1.2B Thinking - Efficient reasoning model, 128k context\",\n    size: \"~2.4GB\",\n    contextLength: 128_000,\n    supportsThinking: true,\n    supportsJson: false,\n    family: \"other\",\n  },\n};\n\n// ============================================\n// Model Resolution\n// ============================================\n\n/**\n * Parse model identifier and resolve to source\n *\n * Supported formats:\n * - \"qwen3.5-0.8b\" (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 for the families the native engine actually supports\n// (a graph generator exists in src/gpu/architectures). Other families fall back\n// to a conservative default.\nconst FAMILY_CONTEXT_DEFAULTS: Record<string, number> = {\n  qwen: 32_768,\n  other: 32_768, // LFM2 supports 128k but config.json is the real source of truth\n};\n\n/**\n * Create model config for an external HuggingFace model.\n *\n * Inference is restricted to families the engine can actually run — Qwen\n * (Qwen2/Qwen3/Qwen3.5) and LFM2 (Liquid). Everything else is left as \"other\"\n * with conservative capability flags so the REPL doesn't advertise features the\n * engine can't deliver.\n */\nexport function createExternalModelConfig(\n  modelId: string,\n  repo: string,\n  contextLength?: number,\n): ModelConfig {\n  const repoLower = repo.toLowerCase();\n\n  // Only infer families that have a graph generator in the registry.\n  let family: ModelConfig[\"family\"] = \"other\";\n  if (repoLower.includes(\"qwen\")) {\n    family = \"qwen\";\n  }\n\n  const isLiquid = repoLower.includes(\"lfm\") || repoLower.includes(\"liquid\");\n  const isQwen = family === \"qwen\";\n\n  return {\n    id: modelId,\n    repo,\n    description: `External model: ${repo}`,\n    size: \"Unknown\",\n    contextLength: contextLength || FAMILY_CONTEXT_DEFAULTS[family] || 32_768,\n    // Qwen3/Qwen3.5 and LFM2.5-Thinking expose thinking; nothing here is vision.\n    supportsThinking: isQwen || isLiquid,\n    supportsJson: isQwen,\n    family,\n  };\n}\n\n/**\n * Fetch context length from HuggingFace model config\n */\nexport async function fetchModelContextLength(repo: string): Promise<number | null> {\n  try {\n    const res = await fetch(`https://huggingface.co/${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 { existsSync } from \"node:fs\";\nimport os from \"node:os\";\nimport path from \"node:path\";\nimport zlib from \"node:zlib\";\nimport PQueue from \"p-queue\";\nimport { generateCacheKey, getGlobalCache } from \"./cache.js\";\nimport {\n  BUILTIN_MODELS,\n  createExternalModelConfig,\n  DEFAULT_MODEL,\n  fetchModelContextLength,\n  getModelConfig,\n  resolveModel,\n} from \"./models.js\";\nimport type {\n  AudioChunk,\n  ConcurrencyConfig,\n  EmbedOptions,\n  EmbedResult,\n  ErrorContext,\n  GenerateOptions,\n  GenerateResult,\n  GerbilConfig,\n  JsonOptions,\n  LoadOptions,\n  LoadSTTOptions,\n  LoadTTSOptions,\n  ModelConfig,\n  PreloadOptions,\n  SessionStats,\n  SpeakOptions,\n  SpeakResult,\n  STTModelConfig,\n  StreamingTranscriptionOptions,\n  StreamingTranscriptionSession,\n  SystemInfo,\n  TelemetryConfig,\n  TranscribeOptions,\n  TranscribeResult,\n  VoiceInfo,\n} from \"./types.js\";\n\n/**\n * Minimal PNG decoder: 8-bit, non-interlaced, color type 2 (RGB) or 6 (RGBA).\n * Returns packed RGB pixels for the native vision encoder. Replaces the\n * transformers.js RawImage decoder for the common PNG case.\n */\nfunction decodePng(buf: Uint8Array): { pixels: Uint8Array; width: number; height: number } {\n  const view = new DataView(buf.buffer, buf.byteOffset, buf.byteLength);\n  const w = view.getUint32(16);\n  const h = view.getUint32(20);\n  const bitDepth = buf[24];\n  const colorType = buf[25];\n  if (bitDepth !== 8 || (colorType !== 2 && colorType !== 6)) {\n    throw new Error(\n      `Unsupported PNG (bitDepth=${bitDepth} colorType=${colorType}); need 8-bit RGB/RGBA`,\n    );\n  }\n  const channels = colorType === 6 ? 4 : 3;\n  let off = 8;\n  const idat: Uint8Array[] = [];\n  while (off < buf.length) {\n    const len = view.getUint32(off);\n    const type = String.fromCharCode(buf[off + 4], buf[off + 5], buf[off + 6], buf[off + 7]);\n    if (type === \"IDAT\") {\n      idat.push(buf.subarray(off + 8, off + 8 + len));\n    }\n    off += 12 + len;\n    if (type === \"IEND\") {\n      break;\n    }\n  }\n  const raw = zlib.inflateSync(Buffer.concat(idat));\n  const stride = w * channels;\n  const out = new Uint8Array(w * h * 3);\n  const line = new Uint8Array(stride);\n  const prev = new Uint8Array(stride);\n  let p = 0;\n  for (let y = 0; y < h; y += 1) {\n    const filter = raw[p];\n    p += 1;\n    for (let i = 0; i < stride; i += 1) {\n      const x = raw[p];\n      p += 1;\n      const a = i >= channels ? line[i - channels] : 0;\n      const b = prev[i];\n      const c = i >= channels ? prev[i - channels] : 0;\n      let v: number;\n      switch (filter) {\n        case 0:\n          v = x;\n          break;\n        case 1:\n          v = x + a;\n          break;\n        case 2:\n          v = x + b;\n          break;\n        case 3:\n          v = x + ((a + b) >> 1);\n          break;\n        case 4: {\n          const pp = a + b - c;\n          const pa = Math.abs(pp - a);\n          const pb = Math.abs(pp - b);\n          const pc = Math.abs(pp - c);\n          let pred = c;\n          if (pa <= pb && pa <= pc) {\n            pred = a;\n          } else if (pb <= pc) {\n            pred = b;\n          }\n          v = x + pred;\n          break;\n        }\n        default:\n          throw new Error(`bad PNG filter ${filter}`);\n      }\n      line[i] = v & 0xff;\n    }\n    for (let x = 0; x < w; x += 1) {\n      out[(y * w + x) * 3 + 0] = line[x * channels + 0];\n      out[(y * w + x) * 3 + 1] = line[x * channels + 1];\n      out[(y * w + x) * 3 + 2] = line[x * channels + 2];\n    }\n    prev.set(line);\n  }\n  return { pixels: out, width: w, height: h };\n}\n\n// Default voices for listVoices() when native 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 type { GenerateObjectOptions, GenerateObjectResult } from \"../gpu/index.js\";\nimport { extractJson, zodToJsonSchema } from \"./utils.js\";\n\n// ============================================\n// Gerbil Class\n// ============================================\n\nexport class Gerbil {\n  private currentModel: string | null = null;\n  private modelConfig: ModelConfig | null = null;\n  private readonly config: GerbilConfig;\n  private stats: SessionStats;\n  private _deviceMode: \"webgpu\" | \"cpu\" | \"wasm\" = \"cpu\"; // Track which backend is active\n  private webgpuEngine: any = null; // Native WebGPU engine (src/gpu)\n\n  // ── Native (src/gpu) capability engines, lazily created and independent of the\n  // loaded text model. These let embed / transcribe / speak run on the native\n  // WebGPU engine even when the active model isn't that capability. They are torn\n  // down in dispose().\n  private nativeEmbedEngine: any = null; // Native WebGPUEngine loaded as an embedder\n  private nativeEmbedRepo: string | null = null; // Repo backing nativeEmbedEngine (for cache-busting on model switch)\n  private nativeSTT: any = null; // Native MoonshineSTT (src/gpu/moonshine-stt)\n  private nativeTTSEngine: any = null; // Native WebGPUEngine (Kani-TTS) backing speak()\n\n  private isVisionModel = false; // Whether current model supports vision\n\n  // Request queue for concurrency control (prevents GPU OOM under load)\n  private readonly queue: PQueue;\n  private readonly telemetry: TelemetryConfig;\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    // Initialize request queue (default: 1 concurrent for LLM, 5 min timeout)\n    const concurrency: ConcurrencyConfig = config.concurrency || {};\n    this.queue = new PQueue({\n      concurrency: concurrency.maxConcurrent ?? 1,\n      timeout: concurrency.timeout ?? 300_000, // 5 minutes (auto-throws on timeout)\n    });\n\n    // Store telemetry hooks\n    this.telemetry = config.telemetry || {};\n  }\n\n  // ============================================\n  // Telemetry Helpers\n  // ============================================\n\n  private reportError(error: Error, context: ErrorContext): void {\n    try {\n      this.telemetry.onError?.(error, context);\n    } catch {\n      // Never throw from telemetry\n    }\n  }\n\n  /**\n   * Whether the native (src/gpu) WebGPU engine should be used for a capability\n   * (embed / transcribe / speak / vision). The native WebGPU engine is the only\n   * inference backend, so this is always true; kept as a seam for callers.\n   */\n  private preferNative(): boolean {\n    return true;\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(\"qwen3.5-0.8b\");\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 = DEFAULT_MODEL, options: LoadOptions = {}): Promise<void> {\n    const loadStartTime = performance.now();\n\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    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    // The native WebGPU engine is the only inference backend. CPU/WASM and the\n    // legacy ONNX path have been removed — guard against legacy runtime values.\n    if ((device as string) === \"cpu\" || (device as string) === \"gpu\") {\n      throw new Error(\n        'Gerbil requires WebGPU. CPU/WASM and the legacy ONNX backend have been removed; use device \"webgpu\" or \"auto\".',\n      );\n    }\n\n    onProgress?.({ status: `Loading ${modelId}...` });\n\n    try {\n      onProgress?.({ status: \"Initializing WebGPU engine...\" });\n      const { WebGPUEngine } = await import(\"../gpu/index.js\");\n\n      // Resolve the HF repo — for built-in models, use the standard HF repo\n      // (not the ONNX repo) since the native engine loads safetensors directly\n      let hfRepo = source.path;\n      // If it points to an ONNX repo, map to the standard repo\n      if (hfRepo.includes(\"onnx-community/\") || hfRepo.includes(\"-ONNX\")) {\n        const nativeRepoMap: Record<string, string> = {\n          \"onnx-community/Qwen3.5-0.8B-ONNX\": \"Qwen/Qwen3.5-0.8B\",\n          \"onnx-community/Qwen3-0.6B-ONNX\": \"Qwen/Qwen3-0.6B\",\n          \"onnx-community/Qwen3-1.7B-ONNX\": \"Qwen/Qwen3-1.7B\",\n          \"onnx-community/Qwen3.5-2B-ONNX\": \"Qwen/Qwen3.5-2B\",\n        };\n        hfRepo = nativeRepoMap[hfRepo] || hfRepo;\n      }\n\n      // Map user dtype to GPU engine dtype (\"q4\" for INT4, \"f32\" default)\n      const gpuDtype = userDtype === \"q4\" ? (\"q4\" as const) : undefined;\n\n      this.webgpuEngine = await WebGPUEngine.create({\n        repo: hfRepo,\n        maxSeqLen: options.contextLength ?? config.contextLength ?? 4096,\n        dtype: gpuDtype,\n        onProgress: (loaded, total, message) => {\n          onProgress?.({\n            status: message,\n            progress: total > 0 ? Math.round((loaded / total) * 100) : undefined,\n          });\n        },\n      });\n\n      this._deviceMode = \"webgpu\";\n      this.isVisionModel = false;\n      this.currentModel = modelId;\n      this.modelConfig = config;\n      onProgress?.({ status: \"Ready (WebGPU Native)!\" });\n\n      // Report successful model load\n      if (this.telemetry.onModelLoad) {\n        try {\n          this.telemetry.onModelLoad({\n            modelId,\n            loadTimeMs: performance.now() - loadStartTime,\n            fromCache: false, // TODO: detect if loaded from cache\n            device: this._deviceMode,\n            success: true,\n          });\n        } catch {\n          // Never throw from telemetry\n        }\n      }\n    } catch (err) {\n      // No CPU/ONNX fallback — surface the load failure.\n      this.reportError(err instanceof Error ? err : new Error(String(err)), {\n        operation: \"load\",\n        modelId,\n      });\n      if (this.telemetry.onModelLoad) {\n        try {\n          this.telemetry.onModelLoad({\n            modelId,\n            loadTimeMs: performance.now() - loadStartTime,\n            fromCache: false,\n            device: this._deviceMode,\n            success: false,\n            error: err instanceof Error ? err.message : String(err),\n          });\n        } catch {\n          // Never throw from telemetry\n        }\n      }\n      throw err;\n    }\n  }\n\n  /**\n   * Load a vision model (VLM) on the native WebGPU engine.\n   * The native engine loads the vision-capable safetensors checkpoint directly\n   * and builds its ViT tower on demand (enableVision: true). describeImage() then\n   * runs encode → splice → decode entirely in WebGPU compute.\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    // The native WebGPU engine is the only backend. CPU/WASM/ONNX are unsupported.\n    if ((device as string) === \"cpu\" || (device as string) === \"gpu\") {\n      throw new Error(\n        'Gerbil vision models require WebGPU. CPU/WASM and the legacy ONNX backend have been removed; use device \"webgpu\" or \"auto\".',\n      );\n    }\n\n    onProgress?.({ status: \"Initializing WebGPU vision engine...\" });\n    const { WebGPUEngine } = await import(\"../gpu/index.js\");\n    // Map an ONNX vision repo to its standard HF checkpoint when possible;\n    // otherwise honor the configured path (defaults resolve inside create()).\n    let visRepo = repoPath;\n    if (visRepo.includes(\"onnx-community/\") || visRepo.includes(\"-ONNX\")) {\n      const map: Record<string, string> = {\n        \"onnx-community/Qwen3.5-0.8B-ONNX\": \"Qwen/Qwen3.5-0.8B\",\n      };\n      visRepo = map[visRepo] || visRepo;\n    }\n    this.webgpuEngine = await WebGPUEngine.create({\n      repo: visRepo,\n      enableVision: true,\n      maxSeqLen: options.contextLength ?? config.contextLength ?? 4096,\n      onProgress: (loaded, total, message) =>\n        onProgress?.({\n          status: message,\n          progress: total > 0 ? Math.round((loaded / total) * 100) : undefined,\n        }),\n    });\n    this._deviceMode = \"webgpu\";\n    this.isVisionModel = true;\n    this.currentModel = modelId;\n    this.modelConfig = config;\n    onProgress?.({ status: \"Ready (Vision, WebGPU Native)!\" });\n  }\n\n  /**\n   * Check if a model is loaded\n   */\n  isLoaded(): boolean {\n    return this.webgpuEngine !== null;\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 the in-memory weight quantization the native engine uses for the loaded\n   * model. The WebGPU engine quantizes weights to INT4 (\"q4\") on load; the KV\n   * cache precision (f16/f32) is separate and device-detected.\n   */\n  getDtype(): string {\n    return \"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  // Model Preloading & Cache Checking\n  // ============================================\n\n  /**\n   * Check if a model is cached (downloaded) without loading it\n   *\n   * @example\n   * ```ts\n   * if (await g.isModelCached(\"qwen3.5-0.8b\")) {\n   *   console.log(\"Model ready, will load instantly\");\n   * } else {\n   *   console.log(\"Model needs to download (~400MB)\");\n   * }\n   * ```\n   */\n  async isModelCached(modelId: string): Promise<boolean> {\n    const source = resolveModel(modelId);\n    return this.isNativeRepoCached(source.path);\n  }\n\n  /**\n   * Check whether the native WebGPU engine has a repo cached on disk.\n   * The native loader stores files under ~/.cache/gerbil/<repo>/<revision>/.\n   */\n  private isNativeRepoCached(repo: string, revision = \"main\"): boolean {\n    try {\n      const home = process.env.HOME || process.env.USERPROFILE || os.homedir();\n      if (!home) {\n        return false;\n      }\n      const modelDir = path.join(home, \".cache\", \"gerbil\", repo.replace(/\\//g, \"_\"), revision);\n      // config.json is the first file the native loader writes for any repo.\n      return existsSync(path.join(modelDir, \"config.json\".replace(/\\//g, \"_\")));\n    } catch {\n      return false;\n    }\n  }\n\n  /**\n   * Preload a model (download without initializing for inference)\n   *\n   * Use this to download models ahead of time, e.g., during app startup,\n   * so users don't wait when they first use AI.\n   *\n   * @example\n   * ```ts\n   * // Preload for later (download only, free memory)\n   * await g.preloadModel(\"qwen3.5-0.8b\", {\n   *   onProgress: (p) => console.log(p.status, p.progress),\n   * });\n   *\n   * // Preload and keep in memory for instant use\n   * await g.preloadModel(\"qwen3.5-0.8b\", { keepLoaded: true });\n   * await g.generate(\"Hello\"); // Instant, no loading needed\n   * ```\n   */\n  async preloadModel(modelId: string, options: PreloadOptions = {}): Promise<void> {\n    const source = resolveModel(modelId);\n    const { onProgress, keepLoaded = false } = options;\n\n    // Already loaded? Nothing to do\n    if (keepLoaded && this.isLoaded() && this.currentModel === modelId) {\n      onProgress?.({ status: \"Model already loaded\" });\n      return;\n    }\n\n    // Already cached and not keeping loaded? Nothing to do\n    if (!keepLoaded && (await this.isModelCached(modelId))) {\n      onProgress?.({ status: \"Model already cached\" });\n      return;\n    }\n\n    // If keepLoaded, use the regular loadModel path\n    if (keepLoaded) {\n      await this.loadModel(modelId, { onProgress });\n      return;\n    }\n\n    onProgress?.({ status: `Preloading ${modelId}...` });\n\n    // The native engine has no \"download-only\" mode, so load it (which fetches\n    // and caches all weight files to ~/.cache/gerbil) then dispose to free GPU\n    // memory. The cached files remain on disk for an instant subsequent load.\n    await this.loadModel(modelId, { onProgress });\n    await this.dispose();\n\n    onProgress?.({ status: \"Preload complete\" });\n  }\n\n  /**\n   * Check if the native TTS model is cached. The native engine always uses the\n   * Kani-TTS-2 checkpoint, so `modelId` is accepted for API compatibility only.\n   */\n  async isTTSCached(_modelId?: string): Promise<boolean> {\n    const { DEFAULT_MODELS } = await import(\"../gpu/index.js\");\n    return this.isNativeRepoCached(DEFAULT_MODELS.tts);\n  }\n\n  /**\n   * Preload the native TTS model (downloads Kani-TTS-2 weights to disk cache).\n   *\n   * @param modelId - Accepted for API compatibility; native TTS uses Kani-TTS-2.\n   * @param options.keepLoaded - Keep the engine in memory for instant use.\n   */\n  async preloadTTS(modelId?: string, options: PreloadOptions = {}): Promise<void> {\n    const { onProgress, keepLoaded = false } = options;\n\n    if (keepLoaded && this.isTTSLoaded()) {\n      onProgress?.({ status: \"TTS model already loaded\" });\n      return;\n    }\n\n    if (!keepLoaded && (await this.isTTSCached(modelId))) {\n      onProgress?.({ status: \"TTS model already cached\" });\n      return;\n    }\n\n    onProgress?.({ status: \"Preloading TTS model...\" });\n    // Build the native TTS engine (downloads + caches weights).\n    await this.ensureNativeTTSEngine();\n\n    if (!keepLoaded && this.nativeTTSEngine) {\n      try {\n        this.nativeTTSEngine.destroy();\n      } catch {\n        // best-effort\n      }\n      this.nativeTTSEngine = null;\n    }\n    onProgress?.({ status: \"Preload complete\" });\n  }\n\n  /**\n   * Check if the native STT model is cached. The native engine always uses the\n   * Moonshine checkpoint, so `modelId` is accepted for API compatibility only.\n   */\n  async isSTTCached(_modelId?: string): Promise<boolean> {\n    const { DEFAULT_MODELS } = await import(\"../gpu/index.js\");\n    return this.isNativeRepoCached(DEFAULT_MODELS.stt);\n  }\n\n  /**\n   * Preload the native STT model (downloads Moonshine weights to disk cache).\n   *\n   * @param modelId - Accepted for API compatibility; native STT uses Moonshine.\n   * @param options.keepLoaded - Keep the engine in memory for instant use.\n   */\n  async preloadSTT(modelId?: string, options: PreloadOptions = {}): Promise<void> {\n    const { onProgress, keepLoaded = false } = options;\n\n    if (keepLoaded && this.isSTTLoaded()) {\n      onProgress?.({ status: \"STT model already loaded\" });\n      return;\n    }\n\n    if (!keepLoaded && (await this.isSTTCached(modelId))) {\n      onProgress?.({ status: \"STT model already cached\" });\n      return;\n    }\n\n    onProgress?.({ status: \"Preloading STT model...\" });\n    // Build the native STT engine (downloads + caches weights).\n    await this.ensureNativeSTT();\n\n    if (!keepLoaded && this.nativeSTT) {\n      try {\n        this.nativeSTT.destroy?.();\n      } catch {\n        // best-effort\n      }\n      this.nativeSTT = null;\n    }\n    onProgress?.({ status: \"Preload complete\" });\n  }\n\n  /**\n   * Check if a native embedding model is cached. Defaults to the native\n   * EmbeddingGemma checkpoint when no repo is provided.\n   */\n  async isEmbeddingCached(modelId?: string): Promise<boolean> {\n    const { DEFAULT_MODELS } = await import(\"../gpu/index.js\");\n    return this.isNativeRepoCached(modelId || DEFAULT_MODELS.embedding);\n  }\n\n  /**\n   * Preload a native embedding model (downloads weights to disk cache).\n   *\n   * @param modelId - Embedding repo (default: native EmbeddingGemma).\n   * @param options.keepLoaded - Keep the engine in memory for instant use.\n   */\n  async preloadEmbedding(modelId?: string, options: PreloadOptions = {}): Promise<void> {\n    const { onProgress, keepLoaded = false } = options;\n\n    if (keepLoaded && this.nativeEmbedEngine) {\n      onProgress?.({ status: \"Embedding model already loaded\" });\n      return;\n    }\n\n    if (!keepLoaded && (await this.isEmbeddingCached(modelId))) {\n      onProgress?.({ status: \"Embedding model already cached\" });\n      return;\n    }\n\n    onProgress?.({ status: \"Preloading embedding model...\" });\n    // Build the native embedding engine (downloads + caches weights).\n    await this.ensureNativeEmbedEngine(modelId);\n\n    if (!keepLoaded && this.nativeEmbedEngine) {\n      try {\n        this.nativeEmbedEngine.destroy();\n      } catch {\n        // best-effort\n      }\n      this.nativeEmbedEngine = null;\n      this.nativeEmbedRepo = null;\n    }\n    onProgress?.({ status: \"Preload complete\" });\n  }\n\n  /**\n   * Clear KV cache to free memory.\n   * The native engine manages its own KV cache; this is a no-op kept for API\n   * compatibility.\n   */\n  // biome-ignore lint/suspicious/useAwait: kept async for API compatibility\n  async clearCache(): Promise<void> {\n    // no-op\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    const queueStartTime = performance.now();\n\n    try {\n      // Queue the generation request (prevents GPU OOM under concurrent load)\n      const result = await this.queue.add(async () => {\n        const queueWaitTime = performance.now() - queueStartTime;\n\n        // Report queue wait time if significant (>100ms)\n        if (queueWaitTime > 100 && this.telemetry.onQueueWait) {\n          try {\n            this.telemetry.onQueueWait(queueWaitTime);\n          } catch {\n            // Never throw from telemetry\n          }\n        }\n\n        const generatedResult = await this.generateInternal(prompt, options);\n\n        // Report successful generation\n        if (this.telemetry.onGenerate) {\n          try {\n            this.telemetry.onGenerate({\n              modelId: this.currentModel || \"unknown\",\n              result: generatedResult,\n              cached: generatedResult.cached ?? false,\n              queueTimeMs: queueWaitTime > 100 ? queueWaitTime : undefined,\n            });\n          } catch {\n            // Never throw from telemetry\n          }\n        }\n\n        return generatedResult;\n      });\n\n      return result as GenerateResult;\n    } catch (error) {\n      // Report errors to telemetry (including queue timeouts)\n      if (this.telemetry.onError) {\n        try {\n          this.telemetry.onError(error instanceof Error ? error : new Error(String(error)), {\n            method: \"generate\",\n            modelId: this.currentModel || \"unknown\",\n            prompt: prompt.slice(0, 100), // Truncate for privacy\n            queueWaitTime: performance.now() - queueStartTime,\n          });\n        } catch {\n          // Never throw from telemetry\n        }\n      }\n      throw error;\n    }\n  }\n\n  /**\n   * Internal generate implementation (called within queue)\n   */\n  private async generateInternal(\n    prompt: string,\n    options: GenerateOptions = {},\n  ): Promise<GenerateResult> {\n    if (!this.isLoaded()) {\n      // Auto-load default model\n      await this.loadModel(this.config.model || DEFAULT_MODEL);\n    }\n\n    const { images } = options;\n\n    // Route to native vision generation when images are provided and the loaded\n    // model supports vision.\n    if (images?.length && this.isVisionModel) {\n      return this.generateWithVision(prompt, options);\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      // Real token counts come from the engine, not a length heuristic.\n      let engineTokensGenerated = 0;\n      let engineTokensPerSecond = 0;\n\n      if (this.webgpuEngine) {\n        // Native WebGPU engine — pure compute shaders, no ONNX\n        const result = await this.webgpuEngine.generate(prompt, {\n          maxTokens,\n          sampling: {\n            temperature,\n            topP,\n            topK,\n          },\n          systemPrompt: system,\n          onToken: options.onToken ? (t: string) => options.onToken?.(t) : undefined,\n        });\n        rawText = result.text;\n        engineTokensGenerated = result.tokensGenerated;\n        engineTokensPerSecond = result.tokensPerSecond;\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      // Use the engine's true decoded-token count and throughput.\n      const tokensGenerated = engineTokensGenerated;\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: engineTokensPerSecond,\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      // Report error to telemetry\n      this.reportError(error instanceof Error ? error : new Error(String(error)), {\n        operation: \"generate\",\n        modelId: this.currentModel || undefined,\n      });\n\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 || DEFAULT_MODEL);\n    }\n\n    const startTime = performance.now();\n\n    // For native WebGPU engine, use real streaming via onToken callback\n    if (this.webgpuEngine) {\n      let fullText = \"\";\n      const tokenQueue: string[] = [];\n      let resolveNext: ((value: string | null) => void) | null = null;\n      let done = false;\n      // Capture the engine's true token count / throughput for the final result.\n      let engineTokensGenerated = 0;\n      let engineTokensPerSecond = 0;\n\n      const generatePromise = this.webgpuEngine\n        .generate(prompt, {\n          ...options,\n          sampling: {\n            temperature: options.temperature,\n            topP: options.topP,\n            topK: options.topK,\n          },\n          systemPrompt: options.system,\n          onToken: (token: string) => {\n            fullText += token;\n            if (resolveNext) {\n              resolveNext(token);\n              resolveNext = null;\n            } else {\n              tokenQueue.push(token);\n            }\n          },\n        })\n        .then((result: { tokensGenerated: number; tokensPerSecond: number }) => {\n          engineTokensGenerated = result.tokensGenerated;\n          engineTokensPerSecond = result.tokensPerSecond;\n          done = true;\n          if (resolveNext) {\n            resolveNext(null);\n          }\n        })\n        .catch((err: Error) => {\n          done = true;\n          if (resolveNext) {\n            resolveNext(null);\n          }\n          throw err;\n        });\n\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 totalTime = performance.now() - startTime;\n\n      return {\n        text: response,\n        thinking: options.thinking ? thinkingText : undefined,\n        tokensGenerated: engineTokensGenerated,\n        totalTime,\n        tokensPerSecond: engineTokensPerSecond,\n        finishReason: \"stop\" as const,\n      };\n    }\n\n    // The native WebGPU engine is the only backend. If we reach here, no model\n    // is loaded (auto-load happens at the top of stream()).\n    throw new Error(\"No model loaded\");\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    // ── Native WebGPU vision ──\n    // loadModel() brought up the native vision engine (this.webgpuEngine has its\n    // ViT tower built via enableVision: true). Decode the image to RGB pixels and\n    // run the native encode → splice → decode path (engine.describeImage). The\n    // native engine currently supports a single image per request.\n    if (!(this.webgpuEngine && typeof this.webgpuEngine.describeImage === \"function\")) {\n      throw new Error(\n        \"Vision model not loaded. Load a vision-capable model with device 'webgpu' first.\",\n      );\n    }\n\n    const imgs = options.images ?? [];\n    if (imgs.length !== 1) {\n      throw new Error(\n        `Native WebGPU vision supports exactly one image per request (got ${imgs.length}).`,\n      );\n    }\n\n    const startTime = performance.now();\n    const { pixels, width, height } = await this.decodeImageToPixels(imgs[0].source);\n    const result = await this.webgpuEngine.describeImage({ pixels, width, height }, prompt, {\n      maxTokens: options.maxTokens ?? 512,\n      sampling: {\n        temperature: options.temperature ?? 0.7,\n        topP: options.topP ?? 0.9,\n        topK: options.topK ?? 20,\n      },\n      onToken: options.onToken ? (t: string) => options.onToken?.(t) : undefined,\n    });\n    const totalTime = performance.now() - startTime;\n\n    // Update stats\n    this.stats.prompts += 1;\n    this.stats.tokensOut += result.tokensGenerated;\n    this.stats.totalTime += totalTime;\n    this.stats.avgSpeed = (this.stats.tokensOut / this.stats.totalTime) * 1000;\n\n    return {\n      text: this.cleanOutput(result.text),\n      tokensGenerated: result.tokensGenerated,\n      tokensPerSecond: result.tokensPerSecond,\n      totalTime,\n      finishReason: \"stop\",\n      provider: \"local\",\n      cached: false,\n    };\n  }\n\n  /**\n   * Decode an image source (http(s) URL, file path, or data URI) to raw RGB\n   * pixels for the native vision encoder. Supports 8-bit non-interlaced PNG\n   * (color types 2/RGB and 6/RGBA). Other formats throw a clear error — callers\n   * can pre-decode and use the lower-level WebGPUEngine.describeImage() with\n   * pixels directly.\n   */\n  private async decodeImageToPixels(\n    source: string,\n  ): Promise<{ pixels: Uint8Array; width: number; height: number }> {\n    const bytes = await this.fetchImageBytes(source);\n    // PNG magic number: 89 50 4E 47 0D 0A 1A 0A\n    const isPng =\n      bytes.length > 8 &&\n      bytes[0] === 0x89 &&\n      bytes[1] === 0x50 &&\n      bytes[2] === 0x4e &&\n      bytes[3] === 0x47;\n    if (!isPng) {\n      throw new Error(\n        \"Native vision currently decodes PNG images only. For other formats, pre-decode \" +\n          \"to RGB pixels and call the GPU engine's describeImage() directly.\",\n      );\n    }\n    return decodePng(bytes);\n  }\n\n  /** Fetch an image source to raw bytes (URL, data URI, or local file path). */\n  private async fetchImageBytes(source: string): Promise<Uint8Array> {\n    if (source.startsWith(\"data:\")) {\n      const comma = source.indexOf(\",\");\n      const meta = source.slice(5, comma);\n      const data = source.slice(comma + 1);\n      if (meta.includes(\"base64\")) {\n        return Uint8Array.from(Buffer.from(data, \"base64\"));\n      }\n      return Uint8Array.from(Buffer.from(decodeURIComponent(data), \"binary\"));\n    }\n    if (source.startsWith(\"http://\") || source.startsWith(\"https://\")) {\n      const res = await fetch(source);\n      if (!res.ok) {\n        throw new Error(`Failed to fetch image (${res.status}): ${source}`);\n      }\n      return new Uint8Array(await res.arrayBuffer());\n    }\n    // Local file path (Node.js only)\n    const { readFile } = await import(\"node:fs/promises\");\n    return new Uint8Array(await readFile(source));\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   * Generate a structured object via the native engine's retrying\n   * `generateObject` (extract JSON → validate → retry with a nudge).\n   *\n   * Unlike {@link json} (which is Zod-driven), this passes through to the engine\n   * and accepts either a predicate validator `(o) => boolean` or a minimal\n   * `{ required: [...] }` schema; omit `schema` to accept any valid JSON.\n   *\n   * @example\n   * ```ts\n   * const { object } = await g.generateObject<{ name: string; age: number }>(\n   *   'Extract {name, age} from: \"I am Sarah, 28\"',\n   *   { schema: { required: [\"name\", \"age\"] } },\n   * );\n   * ```\n   */\n  async generateObject<T = unknown>(\n    prompt: string,\n    options: GenerateObjectOptions = {},\n  ): Promise<GenerateObjectResult<T>> {\n    if (!this.isLoaded()) {\n      await this.loadModel(this.config.model || DEFAULT_MODEL);\n    }\n    if (!this.webgpuEngine) {\n      throw new Error(\"No model loaded\");\n    }\n    return this.webgpuEngine.generateObject(prompt, options) as Promise<GenerateObjectResult<T>>;\n  }\n\n  // ============================================\n  // Embeddings\n  // ============================================\n\n  /**\n   * Generate embeddings\n   */\n  async embed(text: string, options: EmbedOptions = {}): Promise<EmbedResult> {\n    // ── Native WebGPU embedding ──\n    // The active text model may not be an embedder, so embeddings run on a\n    // dedicated native engine loaded with { embedding: true }. It is created\n    // lazily on first embed(), keyed by repo: if the caller asks for a different\n    // embedding model than the cached one, the old engine is torn down and a new\n    // one is built. The engine is owned by this Gerbil instance and released in\n    // dispose(). Native engine.embed() returns an L2-normalized Float32Array.\n    if (!this.preferNative()) {\n      throw new Error(\n        \"Embeddings require WebGPU. CPU/WASM and the legacy ONNX backend have been removed.\",\n      );\n    }\n\n    const native = await this.ensureNativeEmbedEngine(options.model);\n    const startTime = performance.now();\n    const vec = await native.embed(text);\n    return {\n      vector: Array.from(vec as Float32Array),\n      text,\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /**\n   * Lazily build (or reuse) the native embedding engine. Re-creates it when the\n   * requested repo differs from the cached one. The default native embedding\n   * model is resolved by the engine itself (EmbeddingGemma) when no repo given.\n   */\n  private async ensureNativeEmbedEngine(repo?: string): Promise<any> {\n    if (this.nativeEmbedEngine && (!repo || repo === this.nativeEmbedRepo)) {\n      return this.nativeEmbedEngine;\n    }\n    if (this.nativeEmbedEngine) {\n      // Switching embedding models — release the previous native engine.\n      try {\n        this.nativeEmbedEngine.destroy();\n      } catch {\n        // best-effort cleanup\n      }\n      this.nativeEmbedEngine = null;\n      this.nativeEmbedRepo = null;\n    }\n    const { WebGPUEngine } = await import(\"../gpu/index.js\");\n    this.nativeEmbedEngine = await WebGPUEngine.create({ repo, embedding: true });\n    this.nativeEmbedRepo = repo ?? null;\n    return this.nativeEmbedEngine;\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   * Compute cosine similarity between two vectors\n   *\n   * @example\n   * ```ts\n   * const sim = g.cosineSimilarity([1, 0, 0], [1, 0, 0]); // 1.0\n   * const sim2 = g.cosineSimilarity([1, 0, 0], [0, 1, 0]); // 0.0\n   * ```\n   */\n  cosineSimilarity(a: number[], b: number[]): number {\n    if (a.length !== b.length) {\n      throw new Error(`Vector dimensions must match: ${a.length} vs ${b.length}`);\n    }\n\n    let dotProduct = 0;\n    let normA = 0;\n    let normB = 0;\n\n    for (let i = 0; i < a.length; i++) {\n      dotProduct += a[i] * b[i];\n      normA += a[i] * a[i];\n      normB += b[i] * b[i];\n    }\n\n    const magnitude = Math.sqrt(normA) * Math.sqrt(normB);\n    if (magnitude === 0) return 0;\n\n    return dotProduct / magnitude;\n  }\n\n  /**\n   * Compare similarity between two texts\n   *\n   * @example\n   * ```ts\n   * const result = await g.similarity(\"Hello world\", \"Hi there\");\n   * console.log(result.score); // 0.85\n   * ```\n   */\n  async similarity(\n    textA: string,\n    textB: string,\n    options: EmbedOptions = {},\n  ): Promise<import(\"./types.js\").SimilarityResult> {\n    const startTime = performance.now();\n\n    const [embedA, embedB] = await Promise.all([\n      this.embed(textA, options),\n      this.embed(textB, options),\n    ]);\n\n    const score = this.cosineSimilarity(embedA.vector, embedB.vector);\n\n    return {\n      score,\n      textA,\n      textB,\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /**\n   * Semantic search - find most similar texts from a corpus\n   *\n   * @example\n   * ```ts\n   * const results = await g.search(\"capital of France\", [\n   *   \"Paris is beautiful\",\n   *   \"London is in England\",\n   *   \"Dogs are pets\"\n   * ]);\n   * // [{ text: \"Paris is beautiful\", score: 0.89, index: 0 }, ...]\n   * ```\n   */\n  async search(\n    query: string,\n    corpus: string[],\n    options: EmbedOptions & { topK?: number } = {},\n  ): Promise<import(\"./types.js\").SearchResult[]> {\n    const { topK = corpus.length, ...embedOptions } = options;\n\n    // Embed query and all documents\n    const queryEmbedding = await this.embed(query, embedOptions);\n    const corpusEmbeddings = await this.embedBatch(corpus, embedOptions);\n\n    // Calculate similarities\n    const results: import(\"./types.js\").SearchResult[] = corpusEmbeddings.map((doc, index) => ({\n      text: doc.text,\n      score: this.cosineSimilarity(queryEmbedding.vector, doc.vector),\n      index,\n    }));\n\n    // Sort by score descending and take top K\n    return results.sort((a, b) => b.score - a.score).slice(0, topK);\n  }\n\n  /**\n   * Find the nearest text to an embedding vector\n   *\n   * @example\n   * ```ts\n   * const embedding = (await g.embed(\"dog\")).vector;\n   * const match = await g.findNearest(embedding, [\"cat\", \"car\", \"tree\"]);\n   * // { text: \"cat\", score: 0.85, index: 0 }\n   * ```\n   */\n  async findNearest(\n    embedding: number[],\n    candidates: string[],\n    options: EmbedOptions & { topK?: number } = {},\n  ): Promise<import(\"./types.js\").SearchResult[]> {\n    const { topK = candidates.length, ...embedOptions } = options;\n\n    // Embed all candidates\n    const candidateEmbeddings = await this.embedBatch(candidates, embedOptions);\n\n    // Calculate similarities\n    const results: import(\"./types.js\").SearchResult[] = candidateEmbeddings.map((doc, index) => ({\n      text: doc.text,\n      score: this.cosineSimilarity(embedding, doc.vector),\n      index,\n    }));\n\n    // Sort by score descending and take top K\n    return results.sort((a, b) => b.score - a.score).slice(0, topK);\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: \"webgpu-native\",\n        gpu: null, // TODO: surface Dawn adapter info\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: \"~/.cache/gerbil\",\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 readonly ttsModelId = \"kani-tts-2\";\n\n  /**\n   * Load the native TTS model (Kani-TTS-2) for text-to-speech synthesis.\n   *\n   * @example\n   * ```ts\n   * await g.loadTTS({ onProgress: (p) => console.log(p.status) });\n   * const result = await g.speak(\"Hello world\");\n   * // result.audio = Float32Array PCM, result.sampleRate = 22050\n   * ```\n   */\n  // biome-ignore lint/suspicious/useAwait: native TTS is built lazily by ensureNativeTTSEngine\n  async loadTTS(_options: LoadTTSOptions & { model?: string } = {}): Promise<void> {\n    await this.ensureNativeTTSEngine();\n  }\n\n  /**\n   * Ensure TTS model is loaded (lazy loading)\n   */\n  async ensureTTSLoaded(_options?: LoadTTSOptions): Promise<void> {\n    await this.ensureNativeTTSEngine();\n  }\n\n  /**\n   * Generate speech from text using the native Kani-TTS-2 WebGPU engine.\n   *\n   * @example\n   * ```ts\n   * const result = await g.speak(\"Hello world\");\n   * // result.audio = Float32Array PCM, result.sampleRate = 22050\n   * ```\n   */\n  async speak(text: string, options: SpeakOptions = {}): Promise<SpeakResult> {\n    // ── Native WebGPU TTS ──\n    // Native speech runs through a dedicated WebGPUEngine loaded with the\n    // Kani-TTS-2 checkpoint; engine.speak() returns { pcm, sampleRate,\n    // audioSeconds }. The engine is lazily built on first speak() and owned by\n    // this instance (released in dispose()).\n    if (!this.preferNative()) {\n      throw new Error(\n        \"Speech synthesis requires WebGPU. CPU/WASM and the legacy ONNX backend have been removed.\",\n      );\n    }\n\n    const native = await this.ensureNativeTTSEngine();\n    const startTime = performance.now();\n    const out = await native.speak(text, {});\n    return {\n      audio: out.pcm,\n      sampleRate: out.sampleRate,\n      duration: out.audioSeconds,\n      voice: options.voice ?? \"default\",\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /** Lazily build (or reuse) the native Kani-TTS WebGPUEngine (default repo). */\n  private async ensureNativeTTSEngine(): Promise<any> {\n    if (!this.nativeTTSEngine) {\n      const { WebGPUEngine, DEFAULT_MODELS } = await import(\"../gpu/index.js\");\n      this.nativeTTSEngine = await WebGPUEngine.create({ repo: DEFAULT_MODELS.tts });\n    }\n    return this.nativeTTSEngine;\n  }\n\n  /**\n   * Stream speech generation. The native engine synthesizes the full clip, so a\n   * single final audio chunk is yielded.\n   */\n  async *speakStream(\n    text: string,\n    options: SpeakOptions = {},\n  ): AsyncGenerator<AudioChunk, SpeakResult, unknown> {\n    const result = await this.speak(text, options);\n    yield {\n      samples: result.audio,\n      sampleRate: result.sampleRate,\n      index: 0,\n      isFinal: true,\n    };\n    return result;\n  }\n\n  /**\n   * Get list of available TTS voices (native Kani-TTS-2 default voice).\n   */\n  listVoices(): VoiceInfo[] {\n    return KOKORO_VOICES_DEFAULT;\n  }\n\n  /**\n   * Check if TTS model is loaded\n   */\n  isTTSLoaded(): boolean {\n    return this.nativeTTSEngine !== null;\n  }\n\n  /**\n   * Get current TTS model info\n   */\n  getTTSModelInfo(): { id: string; loaded: boolean; device?: \"webgpu\" | \"cpu\" } | null {\n    if (!this.nativeTTSEngine) {\n      return null;\n    }\n    return {\n      id: this.ttsModelId,\n      loaded: true,\n      device: \"webgpu\",\n    };\n  }\n\n  /**\n   * List available TTS models (native Kani-TTS-2).\n   */\n  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  async listTTSModels(): Promise<\n    Array<{ id: string; description: string; sampleRate: number; voiceCount: number }>\n  > {\n    return [\n      {\n        id: this.ttsModelId,\n        description: \"Kani-TTS-2 native WebGPU TTS\",\n        sampleRate: 22_050,\n        voiceCount: 1,\n      },\n    ];\n  }\n\n  // ============================================\n  // Speech-to-Text (STT)\n  // ============================================\n\n  /**\n   * Load the native STT model (Moonshine) for speech-to-text transcription.\n   *\n   * @example\n   * ```ts\n   * await g.loadSTT();\n   * const result = await g.transcribe(audioData);\n   * console.log(result.text);\n   * ```\n   */\n  // biome-ignore lint/suspicious/useAwait: native STT is built lazily by ensureNativeSTT\n  async loadSTT(_modelId?: string, _options: LoadSTTOptions = {}): Promise<void> {\n    // Native STT uses the Moonshine checkpoint, built lazily on first transcribe().\n    await this.ensureNativeSTT();\n  }\n\n  /**\n   * Ensure STT model is loaded (lazy loading)\n   */\n  public async ensureSTTLoaded(_modelId?: string, _options?: LoadSTTOptions): Promise<void> {\n    await this.ensureNativeSTT();\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    // ── Native WebGPU STT ──\n    // MoonshineSTT consumes raw 16 kHz mono Float32 PCM. A Uint8Array WAV\n    // container cannot be decoded natively, and timestamps are not produced by\n    // the native path — both are unsupported now that the ONNX backend is gone.\n    if (!this.preferNative()) {\n      throw new Error(\n        \"Transcription requires WebGPU. CPU/WASM and the legacy ONNX backend have been removed.\",\n      );\n    }\n    if (!(audio instanceof Float32Array)) {\n      throw new Error(\n        \"Native transcription requires 16 kHz mono Float32Array PCM. Decode WAV bytes to PCM first.\",\n      );\n    }\n    if (options.timestamps) {\n      throw new Error(\"Native transcription does not produce timestamps.\");\n    }\n\n    const native = await this.ensureNativeSTT();\n    const startTime = performance.now();\n    const out = await native.transcribe(audio);\n    return {\n      text: out.text,\n      language: options.language ?? \"en\",\n      duration: out.audioSeconds,\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /** Lazily build (or reuse) the native MoonshineSTT engine (default repo). */\n  private async ensureNativeSTT(): Promise<any> {\n    if (!this.nativeSTT) {\n      const { MoonshineSTT } = await import(\"../gpu/moonshine-stt.js\");\n      this.nativeSTT = await MoonshineSTT.create();\n    }\n    return this.nativeSTT;\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  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  async createStreamingTranscription(\n    _options: StreamingTranscriptionOptions = {},\n  ): Promise<StreamingTranscriptionSession> {\n    throw new Error(\n      \"Streaming transcription is not supported by the native WebGPU STT engine. \" +\n        \"Use transcribe() on buffered 16 kHz Float32Array PCM instead.\",\n    );\n  }\n\n  /**\n   * Get list of available STT models (native Moonshine).\n   */\n  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  async listSTTModels(): Promise<STTModelConfig[]> {\n    return [\n      {\n        id: \"moonshine-base\",\n        repo: \"UsefulSensors/moonshine-base\",\n        description: \"Moonshine native WebGPU STT\",\n        size: \"61M\",\n        multilingual: false,\n        languages: [\"en\"],\n        sampleRate: 16000,\n      },\n    ];\n  }\n\n  /**\n   * Check if STT model is loaded\n   */\n  isSTTLoaded(): boolean {\n    return this.nativeSTT !== null;\n  }\n\n  /**\n   * Get current STT model info\n   */\n  getSTTModelInfo(): { id: string; loaded: boolean; device?: \"webgpu\" | \"cpu\" } | null {\n    if (!this.nativeSTT) {\n      return null;\n    }\n    return {\n      id: \"moonshine-base\",\n      loaded: true,\n      device: \"webgpu\",\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 (releases all native WebGPU engines and their devices).\n   * @param _disconnect Accepted for API compatibility; no longer used.\n   */\n  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  async dispose(_disconnect = false): Promise<void> {\n    // Clean up native WebGPU engine\n    if (this.webgpuEngine) {\n      try {\n        this.webgpuEngine.destroy();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.webgpuEngine = null;\n    }\n\n    // Clean up the lazily-created native capability engines (embed / STT / TTS).\n    // These are independent of this.webgpuEngine (they back embed()/transcribe()/\n    // speak() even when the active model isn't that capability), so they need\n    // their own teardown. MoonshineSTT has no GPU device of its own to destroy\n    // beyond its weight maps; WebGPUEngine.destroy() releases its device.\n    if (this.nativeEmbedEngine) {\n      try {\n        this.nativeEmbedEngine.destroy();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.nativeEmbedEngine = null;\n      this.nativeEmbedRepo = null;\n    }\n    if (this.nativeSTT) {\n      try {\n        this.nativeSTT.destroy?.();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.nativeSTT = null;\n    }\n    if (this.nativeTTSEngine) {\n      try {\n        this.nativeTTSEngine.destroy();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.nativeTTSEngine = null;\n    }\n\n    this.currentModel = null;\n    this.modelConfig = null;\n    this.isVisionModel = false;\n  }\n\n  /**\n   * @deprecated The shared Chrome backend was removed; this is now a no-op.\n   */\n  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  static async shutdown(): Promise<void> {\n    // no-op\n  }\n\n  // ============================================\n  // Private Methods\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;;;;;;;;;;AC5NvB,MAAa,gBAAgB;AAW7B,MAAaC,iBAA8C;CACzD,gBAAgB;EACd,IAAI;EACJ,MAAM;EACN,aACE;EACF,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,gBAAgB;EAChB,QAAQ;EACT;CACD,cAAc;EACZ,IAAI;EACJ,MAAM;EACN,aACE;EACF,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,gBAAgB;EAChB,QAAQ;EACT;CACD,wBAAwB;EACtB,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,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;;AAMpC,MAAMC,0BAAkD;CACtD,MAAM;CACN,OAAO;CACR;;;;;;;;;AAUD,SAAgB,0BACd,SACA,MACA,eACa;CACb,MAAM,YAAY,KAAK,aAAa;CAGpC,IAAIC,SAAgC;AACpC,KAAI,UAAU,SAAS,OAAO,CAC5B,UAAS;CAGX,MAAM,WAAW,UAAU,SAAS,MAAM,IAAI,UAAU,SAAS,SAAS;CAC1E,MAAM,SAAS,WAAW;AAE1B,QAAO;EACL,IAAI;EACJ;EACA,aAAa,mBAAmB;EAChC,MAAM;EACN,eAAe,iBAAiB,wBAAwB,WAAW;EAEnE,kBAAkB,UAAU;EAC5B,cAAc;EACd;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;;;;;;;;;;;;;AChKtC,SAAS,UAAU,KAAwE;CACzF,MAAM,OAAO,IAAI,SAAS,IAAI,QAAQ,IAAI,YAAY,IAAI,WAAW;CACrE,MAAM,IAAI,KAAK,UAAU,GAAG;CAC5B,MAAM,IAAI,KAAK,UAAU,GAAG;CAC5B,MAAM,WAAW,IAAI;CACrB,MAAM,YAAY,IAAI;AACtB,KAAI,aAAa,KAAM,cAAc,KAAK,cAAc,EACtD,OAAM,IAAI,MACR,6BAA6B,SAAS,aAAa,UAAU,wBAC9D;CAEH,MAAM,WAAW,cAAc,IAAI,IAAI;CACvC,IAAI,MAAM;CACV,MAAMC,OAAqB,EAAE;AAC7B,QAAO,MAAM,IAAI,QAAQ;EACvB,MAAM,MAAM,KAAK,UAAU,IAAI;EAC/B,MAAM,OAAO,OAAO,aAAa,IAAI,MAAM,IAAI,IAAI,MAAM,IAAI,IAAI,MAAM,IAAI,IAAI,MAAM,GAAG;AACxF,MAAI,SAAS,OACX,MAAK,KAAK,IAAI,SAAS,MAAM,GAAG,MAAM,IAAI,IAAI,CAAC;AAEjD,SAAO,KAAK;AACZ,MAAI,SAAS,OACX;;CAGJ,MAAM,MAAM,KAAK,YAAY,OAAO,OAAO,KAAK,CAAC;CACjD,MAAM,SAAS,IAAI;CACnB,MAAM,MAAM,IAAI,WAAW,IAAI,IAAI,EAAE;CACrC,MAAM,OAAO,IAAI,WAAW,OAAO;CACnC,MAAM,OAAO,IAAI,WAAW,OAAO;CACnC,IAAI,IAAI;AACR,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,GAAG;EAC7B,MAAM,SAAS,IAAI;AACnB,OAAK;AACL,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK,GAAG;GAClC,MAAM,IAAI,IAAI;AACd,QAAK;GACL,MAAM,IAAI,KAAK,WAAW,KAAK,IAAI,YAAY;GAC/C,MAAM,IAAI,KAAK;GACf,MAAM,IAAI,KAAK,WAAW,KAAK,IAAI,YAAY;GAC/C,IAAIC;AACJ,WAAQ,QAAR;IACE,KAAK;AACH,SAAI;AACJ;IACF,KAAK;AACH,SAAI,IAAI;AACR;IACF,KAAK;AACH,SAAI,IAAI;AACR;IACF,KAAK;AACH,SAAI,KAAM,IAAI,KAAM;AACpB;IACF,KAAK,GAAG;KACN,MAAM,KAAK,IAAI,IAAI;KACnB,MAAM,KAAK,KAAK,IAAI,KAAK,EAAE;KAC3B,MAAM,KAAK,KAAK,IAAI,KAAK,EAAE;KAC3B,MAAM,KAAK,KAAK,IAAI,KAAK,EAAE;KAC3B,IAAI,OAAO;AACX,SAAI,MAAM,MAAM,MAAM,GACpB,QAAO;cACE,MAAM,GACf,QAAO;AAET,SAAI,IAAI;AACR;;IAEF,QACE,OAAM,IAAI,MAAM,kBAAkB,SAAS;;AAE/C,QAAK,KAAK,IAAI;;AAEhB,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,GAAG;AAC7B,QAAK,IAAI,IAAI,KAAK,IAAI,KAAK,KAAK,IAAI,WAAW;AAC/C,QAAK,IAAI,IAAI,KAAK,IAAI,KAAK,KAAK,IAAI,WAAW;AAC/C,QAAK,IAAI,IAAI,KAAK,IAAI,KAAK,KAAK,IAAI,WAAW;;AAEjD,OAAK,IAAI,KAAK;;AAEhB,QAAO;EAAE,QAAQ;EAAK,OAAO;EAAG,QAAQ;EAAG;;AAI7C,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;AASD,IAAa,SAAb,MAAoB;CAClB,AAAQ,eAA8B;CACtC,AAAQ,cAAkC;CAC1C,AAAiB;CACjB,AAAQ;CACR,AAAQ,cAAyC;CACjD,AAAQ,eAAoB;CAM5B,AAAQ,oBAAyB;CACjC,AAAQ,kBAAiC;CACzC,AAAQ,YAAiB;CACzB,AAAQ,kBAAuB;CAE/B,AAAQ,gBAAgB;CAGxB,AAAiB;CACjB,AAAiB;CAEjB,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;EAGD,MAAMC,cAAiC,OAAO,eAAe,EAAE;AAC/D,OAAK,QAAQ,IAAI,OAAO;GACtB,aAAa,YAAY,iBAAiB;GAC1C,SAAS,YAAY,WAAW;GACjC,CAAC;AAGF,OAAK,YAAY,OAAO,aAAa,EAAE;;CAOzC,AAAQ,YAAY,OAAc,SAA6B;AAC7D,MAAI;AACF,QAAK,UAAU,UAAU,OAAO,QAAQ;UAClC;;;;;;;CAUV,AAAQ,eAAwB;AAC9B,SAAO;;CAOT,OAAO,aAA4B;AACjC,SAAO,OAAO,OAAO,eAAe;;CAGtC,OAAO,SAAS,SAA0C;AACxD,SAAO,eAAe;;;;;;;;;;;;;;;;;;;;CAyBxB,MAAM,UAAU,UAAU,eAAe,UAAuB,EAAE,EAAiB;EACjF,MAAM,gBAAgB,YAAY,KAAK;AAIvC,MAAI,KAAK,UAAU,CACjB,OAAM,KAAK,SAAS;EAGtB,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;AAKpE,MAAK,WAAsB,SAAU,WAAsB,MACzD,OAAM,IAAI,MACR,qHACD;AAGH,eAAa,EAAE,QAAQ,WAAW,QAAQ,MAAM,CAAC;AAEjD,MAAI;AACF,gBAAa,EAAE,QAAQ,iCAAiC,CAAC;GACzD,MAAM,EAAE,iBAAiB,MAAM,OAAO;GAItC,IAAI,SAAS,OAAO;AAEpB,OAAI,OAAO,SAAS,kBAAkB,IAAI,OAAO,SAAS,QAAQ,CAOhE,UAN8C;IAC5C,oCAAoC;IACpC,kCAAkC;IAClC,kCAAkC;IAClC,kCAAkC;IACnC,CACsB,WAAW;GAIpC,MAAM,WAAW,cAAc,OAAQ,OAAiB;AAExD,QAAK,eAAe,MAAM,aAAa,OAAO;IAC5C,MAAM;IACN,WAAW,QAAQ,iBAAiB,OAAO,iBAAiB;IAC5D,OAAO;IACP,aAAa,QAAQ,OAAO,YAAY;AACtC,kBAAa;MACX,QAAQ;MACR,UAAU,QAAQ,IAAI,KAAK,MAAO,SAAS,QAAS,IAAI,GAAG;MAC5D,CAAC;;IAEL,CAAC;AAEF,QAAK,cAAc;AACnB,QAAK,gBAAgB;AACrB,QAAK,eAAe;AACpB,QAAK,cAAc;AACnB,gBAAa,EAAE,QAAQ,0BAA0B,CAAC;AAGlD,OAAI,KAAK,UAAU,YACjB,KAAI;AACF,SAAK,UAAU,YAAY;KACzB;KACA,YAAY,YAAY,KAAK,GAAG;KAChC,WAAW;KACX,QAAQ,KAAK;KACb,SAAS;KACV,CAAC;WACI;WAIH,KAAK;AAEZ,QAAK,YAAY,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,IAAI,CAAC,EAAE;IACpE,WAAW;IACX;IACD,CAAC;AACF,OAAI,KAAK,UAAU,YACjB,KAAI;AACF,SAAK,UAAU,YAAY;KACzB;KACA,YAAY,YAAY,KAAK,GAAG;KAChC,WAAW;KACX,QAAQ,KAAK;KACb,SAAS;KACT,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,IAAI;KACxD,CAAC;WACI;AAIV,SAAM;;;;;;;;;CAUV,MAAc,gBACZ,SACA,UACA,QACA,UAAuB,EAAE,EACV;EACf,MAAM,EAAE,YAAY,SAAS,WAAW;AAExC,eAAa,EAAE,QAAQ,WAAW,QAAQ,qBAAqB,CAAC;AAGhE,MAAK,WAAsB,SAAU,WAAsB,MACzD,OAAM,IAAI,MACR,kIACD;AAGH,eAAa,EAAE,QAAQ,wCAAwC,CAAC;EAChE,MAAM,EAAE,iBAAiB,MAAM,OAAO;EAGtC,IAAI,UAAU;AACd,MAAI,QAAQ,SAAS,kBAAkB,IAAI,QAAQ,SAAS,QAAQ,CAIlE,WAHoC,EAClC,oCAAoC,qBACrC,CACa,YAAY;AAE5B,OAAK,eAAe,MAAM,aAAa,OAAO;GAC5C,MAAM;GACN,cAAc;GACd,WAAW,QAAQ,iBAAiB,OAAO,iBAAiB;GAC5D,aAAa,QAAQ,OAAO,YAC1B,aAAa;IACX,QAAQ;IACR,UAAU,QAAQ,IAAI,KAAK,MAAO,SAAS,QAAS,IAAI,GAAG;IAC5D,CAAC;GACL,CAAC;AACF,OAAK,cAAc;AACnB,OAAK,gBAAgB;AACrB,OAAK,eAAe;AACpB,OAAK,cAAc;AACnB,eAAa,EAAE,QAAQ,kCAAkC,CAAC;;;;;CAM5D,WAAoB;AAClB,SAAO,KAAK,iBAAiB;;;;;CAM/B,iBAA0B;AACxB,SAAO,KAAK,iBAAiB,KAAK,aAAa,mBAAmB;;;;;CAMpE,eAAmC;AACjC,SAAO,KAAK;;;;;CAMd,gBAA2C;AACzC,SAAO,KAAK;;;;;;;CAQd,WAAmB;AACjB,SAAO;;;;;CAMT,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;;;;;;;;;;;;;;CAmB1B,MAAM,cAAc,SAAmC;EACrD,MAAM,SAAS,aAAa,QAAQ;AACpC,SAAO,KAAK,mBAAmB,OAAO,KAAK;;;;;;CAO7C,AAAQ,mBAAmB,MAAc,WAAW,QAAiB;AACnE,MAAI;GACF,MAAM,OAAO,QAAQ,IAAI,QAAQ,QAAQ,IAAI,eAAe,GAAG,SAAS;AACxE,OAAI,CAAC,KACH,QAAO;GAET,MAAM,WAAW,KAAK,KAAK,MAAM,UAAU,UAAU,KAAK,QAAQ,OAAO,IAAI,EAAE,SAAS;AAExF,UAAO,WAAW,KAAK,KAAK,UAAU,cAAc,QAAQ,OAAO,IAAI,CAAC,CAAC;UACnE;AACN,UAAO;;;;;;;;;;;;;;;;;;;;;CAsBX,MAAM,aAAa,SAAiB,UAA0B,EAAE,EAAiB;AAChE,eAAa,QAAQ;EACpC,MAAM,EAAE,YAAY,aAAa,UAAU;AAG3C,MAAI,cAAc,KAAK,UAAU,IAAI,KAAK,iBAAiB,SAAS;AAClE,gBAAa,EAAE,QAAQ,wBAAwB,CAAC;AAChD;;AAIF,MAAI,CAAC,cAAe,MAAM,KAAK,cAAc,QAAQ,EAAG;AACtD,gBAAa,EAAE,QAAQ,wBAAwB,CAAC;AAChD;;AAIF,MAAI,YAAY;AACd,SAAM,KAAK,UAAU,SAAS,EAAE,YAAY,CAAC;AAC7C;;AAGF,eAAa,EAAE,QAAQ,cAAc,QAAQ,MAAM,CAAC;AAKpD,QAAM,KAAK,UAAU,SAAS,EAAE,YAAY,CAAC;AAC7C,QAAM,KAAK,SAAS;AAEpB,eAAa,EAAE,QAAQ,oBAAoB,CAAC;;;;;;CAO9C,MAAM,YAAY,UAAqC;EACrD,MAAM,EAAE,mBAAmB,MAAM,OAAO;AACxC,SAAO,KAAK,mBAAmB,eAAe,IAAI;;;;;;;;CASpD,MAAM,WAAW,SAAkB,UAA0B,EAAE,EAAiB;EAC9E,MAAM,EAAE,YAAY,aAAa,UAAU;AAE3C,MAAI,cAAc,KAAK,aAAa,EAAE;AACpC,gBAAa,EAAE,QAAQ,4BAA4B,CAAC;AACpD;;AAGF,MAAI,CAAC,cAAe,MAAM,KAAK,YAAY,QAAQ,EAAG;AACpD,gBAAa,EAAE,QAAQ,4BAA4B,CAAC;AACpD;;AAGF,eAAa,EAAE,QAAQ,2BAA2B,CAAC;AAEnD,QAAM,KAAK,uBAAuB;AAElC,MAAI,CAAC,cAAc,KAAK,iBAAiB;AACvC,OAAI;AACF,SAAK,gBAAgB,SAAS;WACxB;AAGR,QAAK,kBAAkB;;AAEzB,eAAa,EAAE,QAAQ,oBAAoB,CAAC;;;;;;CAO9C,MAAM,YAAY,UAAqC;EACrD,MAAM,EAAE,mBAAmB,MAAM,OAAO;AACxC,SAAO,KAAK,mBAAmB,eAAe,IAAI;;;;;;;;CASpD,MAAM,WAAW,SAAkB,UAA0B,EAAE,EAAiB;EAC9E,MAAM,EAAE,YAAY,aAAa,UAAU;AAE3C,MAAI,cAAc,KAAK,aAAa,EAAE;AACpC,gBAAa,EAAE,QAAQ,4BAA4B,CAAC;AACpD;;AAGF,MAAI,CAAC,cAAe,MAAM,KAAK,YAAY,QAAQ,EAAG;AACpD,gBAAa,EAAE,QAAQ,4BAA4B,CAAC;AACpD;;AAGF,eAAa,EAAE,QAAQ,2BAA2B,CAAC;AAEnD,QAAM,KAAK,iBAAiB;AAE5B,MAAI,CAAC,cAAc,KAAK,WAAW;AACjC,OAAI;AACF,SAAK,UAAU,WAAW;WACpB;AAGR,QAAK,YAAY;;AAEnB,eAAa,EAAE,QAAQ,oBAAoB,CAAC;;;;;;CAO9C,MAAM,kBAAkB,SAAoC;EAC1D,MAAM,EAAE,mBAAmB,MAAM,OAAO;AACxC,SAAO,KAAK,mBAAmB,WAAW,eAAe,UAAU;;;;;;;;CASrE,MAAM,iBAAiB,SAAkB,UAA0B,EAAE,EAAiB;EACpF,MAAM,EAAE,YAAY,aAAa,UAAU;AAE3C,MAAI,cAAc,KAAK,mBAAmB;AACxC,gBAAa,EAAE,QAAQ,kCAAkC,CAAC;AAC1D;;AAGF,MAAI,CAAC,cAAe,MAAM,KAAK,kBAAkB,QAAQ,EAAG;AAC1D,gBAAa,EAAE,QAAQ,kCAAkC,CAAC;AAC1D;;AAGF,eAAa,EAAE,QAAQ,iCAAiC,CAAC;AAEzD,QAAM,KAAK,wBAAwB,QAAQ;AAE3C,MAAI,CAAC,cAAc,KAAK,mBAAmB;AACzC,OAAI;AACF,SAAK,kBAAkB,SAAS;WAC1B;AAGR,QAAK,oBAAoB;AACzB,QAAK,kBAAkB;;AAEzB,eAAa,EAAE,QAAQ,oBAAoB,CAAC;;;;;;;CAS9C,MAAM,aAA4B;;;;;;;;;;;;;;;CAsBlC,MAAM,SAAS,QAAgB,UAA2B,EAAE,EAA2B;EACrF,MAAM,iBAAiB,YAAY,KAAK;AAExC,MAAI;AAiCF,UA/Be,MAAM,KAAK,MAAM,IAAI,YAAY;IAC9C,MAAM,gBAAgB,YAAY,KAAK,GAAG;AAG1C,QAAI,gBAAgB,OAAO,KAAK,UAAU,YACxC,KAAI;AACF,UAAK,UAAU,YAAY,cAAc;YACnC;IAKV,MAAM,kBAAkB,MAAM,KAAK,iBAAiB,QAAQ,QAAQ;AAGpE,QAAI,KAAK,UAAU,WACjB,KAAI;AACF,UAAK,UAAU,WAAW;MACxB,SAAS,KAAK,gBAAgB;MAC9B,QAAQ;MACR,QAAQ,gBAAgB,UAAU;MAClC,aAAa,gBAAgB,MAAM,gBAAgB;MACpD,CAAC;YACI;AAKV,WAAO;KACP;WAGK,OAAO;AAEd,OAAI,KAAK,UAAU,QACjB,KAAI;AACF,SAAK,UAAU,QAAQ,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,MAAM,CAAC,EAAE;KAChF,QAAQ;KACR,SAAS,KAAK,gBAAgB;KAC9B,QAAQ,OAAO,MAAM,GAAG,IAAI;KAC5B,eAAe,YAAY,KAAK,GAAG;KACpC,CAAC;WACI;AAIV,SAAM;;;;;;CAOV,MAAc,iBACZ,QACA,UAA2B,EAAE,EACJ;AACzB,MAAI,CAAC,KAAK,UAAU,CAElB,OAAM,KAAK,UAAU,KAAK,OAAO,SAAS,cAAc;EAG1D,MAAM,EAAE,WAAW;AAInB,MAAI,QAAQ,UAAU,KAAK,cACzB,QAAO,KAAK,mBAAmB,QAAQ,QAAQ;EAGjD,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;GAEd,IAAI,wBAAwB;GAC5B,IAAI,wBAAwB;AAE5B,OAAI,KAAK,cAAc;IAErB,MAAMC,WAAS,MAAM,KAAK,aAAa,SAAS,QAAQ;KACtD;KACA,UAAU;MACR;MACA;MACA;MACD;KACD,cAAc;KACd,SAAS,QAAQ,WAAW,MAAc,QAAQ,UAAU,EAAE,GAAG;KAClE,CAAC;AACF,cAAUA,SAAO;AACjB,4BAAwBA,SAAO;AAC/B,4BAAwBA,SAAO;SAE/B,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;GAGhD,MAAM,kBAAkB;AAGxB,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,iBAAiB;IACjB;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,OAAO;AAEd,QAAK,YAAY,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,MAAM,CAAC,EAAE;IAC1E,WAAW;IACX,SAAS,KAAK,gBAAgB;IAC/B,CAAC;AAEF,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,cAAc;EAG1D,MAAM,YAAY,YAAY,KAAK;AAGnC,MAAI,KAAK,cAAc;GACrB,IAAI,WAAW;GACf,MAAMC,aAAuB,EAAE;GAC/B,IAAIC,cAAuD;GAC3D,IAAI,OAAO;GAEX,IAAI,wBAAwB;GAC5B,IAAI,wBAAwB;GAE5B,MAAM,kBAAkB,KAAK,aAC1B,SAAS,QAAQ;IAChB,GAAG;IACH,UAAU;KACR,aAAa,QAAQ;KACrB,MAAM,QAAQ;KACd,MAAM,QAAQ;KACf;IACD,cAAc,QAAQ;IACtB,UAAU,UAAkB;AAC1B,iBAAY;AACZ,SAAI,aAAa;AACf,kBAAY,MAAM;AAClB,oBAAc;WAEd,YAAW,KAAK,MAAM;;IAG3B,CAAC,CACD,MAAM,WAAiE;AACtE,4BAAwB,OAAO;AAC/B,4BAAwB,OAAO;AAC/B,WAAO;AACP,QAAI,YACF,aAAY,KAAK;KAEnB,CACD,OAAO,QAAe;AACrB,WAAO;AACP,QAAI,YACF,aAAY,KAAK;AAEnB,UAAM;KACN;AAEJ,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,YAAY,YAAY,KAAK,GAAG;AAEtC,UAAO;IACL,MAAM;IACN,UAAU,QAAQ,WAAW,eAAe;IAC5C,iBAAiB;IACjB;IACA,iBAAiB;IACjB,cAAc;IACf;;AAKH,QAAM,IAAI,MAAM,kBAAkB;;;;;;CAWpC,MAAc,mBACZ,QACA,SACyB;AAMzB,MAAI,EAAE,KAAK,gBAAgB,OAAO,KAAK,aAAa,kBAAkB,YACpE,OAAM,IAAI,MACR,mFACD;EAGH,MAAM,OAAO,QAAQ,UAAU,EAAE;AACjC,MAAI,KAAK,WAAW,EAClB,OAAM,IAAI,MACR,oEAAoE,KAAK,OAAO,IACjF;EAGH,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,EAAE,QAAQ,OAAO,WAAW,MAAM,KAAK,oBAAoB,KAAK,GAAG,OAAO;EAChF,MAAM,SAAS,MAAM,KAAK,aAAa,cAAc;GAAE;GAAQ;GAAO;GAAQ,EAAE,QAAQ;GACtF,WAAW,QAAQ,aAAa;GAChC,UAAU;IACR,aAAa,QAAQ,eAAe;IACpC,MAAM,QAAQ,QAAQ;IACtB,MAAM,QAAQ,QAAQ;IACvB;GACD,SAAS,QAAQ,WAAW,MAAc,QAAQ,UAAU,EAAE,GAAG;GAClE,CAAC;EACF,MAAM,YAAY,YAAY,KAAK,GAAG;AAGtC,OAAK,MAAM,WAAW;AACtB,OAAK,MAAM,aAAa,OAAO;AAC/B,OAAK,MAAM,aAAa;AACxB,OAAK,MAAM,WAAY,KAAK,MAAM,YAAY,KAAK,MAAM,YAAa;AAEtE,SAAO;GACL,MAAM,KAAK,YAAY,OAAO,KAAK;GACnC,iBAAiB,OAAO;GACxB,iBAAiB,OAAO;GACxB;GACA,cAAc;GACd,UAAU;GACV,QAAQ;GACT;;;;;;;;;CAUH,MAAc,oBACZ,QACgE;EAChE,MAAM,QAAQ,MAAM,KAAK,gBAAgB,OAAO;AAQhD,MAAI,EALF,MAAM,SAAS,KACf,MAAM,OAAO,OACb,MAAM,OAAO,MACb,MAAM,OAAO,MACb,MAAM,OAAO,IAEb,OAAM,IAAI,MACR,mJAED;AAEH,SAAO,UAAU,MAAM;;;CAIzB,MAAc,gBAAgB,QAAqC;AACjE,MAAI,OAAO,WAAW,QAAQ,EAAE;GAC9B,MAAM,QAAQ,OAAO,QAAQ,IAAI;GACjC,MAAM,OAAO,OAAO,MAAM,GAAG,MAAM;GACnC,MAAM,OAAO,OAAO,MAAM,QAAQ,EAAE;AACpC,OAAI,KAAK,SAAS,SAAS,CACzB,QAAO,WAAW,KAAK,OAAO,KAAK,MAAM,SAAS,CAAC;AAErD,UAAO,WAAW,KAAK,OAAO,KAAK,mBAAmB,KAAK,EAAE,SAAS,CAAC;;AAEzE,MAAI,OAAO,WAAW,UAAU,IAAI,OAAO,WAAW,WAAW,EAAE;GACjE,MAAM,MAAM,MAAM,MAAM,OAAO;AAC/B,OAAI,CAAC,IAAI,GACP,OAAM,IAAI,MAAM,0BAA0B,IAAI,OAAO,KAAK,SAAS;AAErE,UAAO,IAAI,WAAW,MAAM,IAAI,aAAa,CAAC;;EAGhD,MAAM,EAAE,aAAa,MAAM,OAAO;AAClC,SAAO,IAAI,WAAW,MAAM,SAAS,OAAO,CAAC;;;;;CAU/C,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;;;;;;;;;;;;;;;;;;CAmBlD,MAAM,eACJ,QACA,UAAiC,EAAE,EACD;AAClC,MAAI,CAAC,KAAK,UAAU,CAClB,OAAM,KAAK,UAAU,KAAK,OAAO,SAAS,cAAc;AAE1D,MAAI,CAAC,KAAK,aACR,OAAM,IAAI,MAAM,kBAAkB;AAEpC,SAAO,KAAK,aAAa,eAAe,QAAQ,QAAQ;;;;;CAU1D,MAAM,MAAM,MAAc,UAAwB,EAAE,EAAwB;AAQ1E,MAAI,CAAC,KAAK,cAAc,CACtB,OAAM,IAAI,MACR,qFACD;EAGH,MAAM,SAAS,MAAM,KAAK,wBAAwB,QAAQ,MAAM;EAChE,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,MAAM,MAAM,OAAO,MAAM,KAAK;AACpC,SAAO;GACL,QAAQ,MAAM,KAAK,IAAoB;GACvC;GACA,WAAW,YAAY,KAAK,GAAG;GAChC;;;;;;;CAQH,MAAc,wBAAwB,MAA6B;AACjE,MAAI,KAAK,sBAAsB,CAAC,QAAQ,SAAS,KAAK,iBACpD,QAAO,KAAK;AAEd,MAAI,KAAK,mBAAmB;AAE1B,OAAI;AACF,SAAK,kBAAkB,SAAS;WAC1B;AAGR,QAAK,oBAAoB;AACzB,QAAK,kBAAkB;;EAEzB,MAAM,EAAE,iBAAiB,MAAM,OAAO;AACtC,OAAK,oBAAoB,MAAM,aAAa,OAAO;GAAE;GAAM,WAAW;GAAM,CAAC;AAC7E,OAAK,kBAAkB,QAAQ;AAC/B,SAAO,KAAK;;;;;CAMd,MAAM,WAAW,OAAiB,UAAwB,EAAE,EAA0B;EACpF,MAAMC,UAAyB,EAAE;AACjC,OAAK,MAAM,QAAQ,MACjB,SAAQ,KAAK,MAAM,KAAK,MAAM,MAAM,QAAQ,CAAC;AAE/C,SAAO;;;;;;;;;;;CAYT,iBAAiB,GAAa,GAAqB;AACjD,MAAI,EAAE,WAAW,EAAE,OACjB,OAAM,IAAI,MAAM,iCAAiC,EAAE,OAAO,MAAM,EAAE,SAAS;EAG7E,IAAI,aAAa;EACjB,IAAI,QAAQ;EACZ,IAAI,QAAQ;AAEZ,OAAK,IAAI,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,iBAAc,EAAE,KAAK,EAAE;AACvB,YAAS,EAAE,KAAK,EAAE;AAClB,YAAS,EAAE,KAAK,EAAE;;EAGpB,MAAM,YAAY,KAAK,KAAK,MAAM,GAAG,KAAK,KAAK,MAAM;AACrD,MAAI,cAAc,EAAG,QAAO;AAE5B,SAAO,aAAa;;;;;;;;;;;CAYtB,MAAM,WACJ,OACA,OACA,UAAwB,EAAE,EACsB;EAChD,MAAM,YAAY,YAAY,KAAK;EAEnC,MAAM,CAAC,QAAQ,UAAU,MAAM,QAAQ,IAAI,CACzC,KAAK,MAAM,OAAO,QAAQ,EAC1B,KAAK,MAAM,OAAO,QAAQ,CAC3B,CAAC;AAIF,SAAO;GACL,OAHY,KAAK,iBAAiB,OAAO,QAAQ,OAAO,OAAO;GAI/D;GACA;GACA,WAAW,YAAY,KAAK,GAAG;GAChC;;;;;;;;;;;;;;;CAgBH,MAAM,OACJ,OACA,QACA,UAA4C,EAAE,EACA;EAC9C,MAAM,EAAE,OAAO,OAAO,QAAQ,GAAG,iBAAiB;EAGlD,MAAM,iBAAiB,MAAM,KAAK,MAAM,OAAO,aAAa;AAW5D,UAVyB,MAAM,KAAK,WAAW,QAAQ,aAAa,EAGE,KAAK,KAAK,WAAW;GACzF,MAAM,IAAI;GACV,OAAO,KAAK,iBAAiB,eAAe,QAAQ,IAAI,OAAO;GAC/D;GACD,EAAE,CAGY,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;;;;;;;;;;;;CAajE,MAAM,YACJ,WACA,YACA,UAA4C,EAAE,EACA;EAC9C,MAAM,EAAE,OAAO,WAAW,QAAQ,GAAG,iBAAiB;AAatD,UAV4B,MAAM,KAAK,WAAW,YAAY,aAAa,EAGF,KAAK,KAAK,WAAW;GAC5F,MAAM,IAAI;GACV,OAAO,KAAK,iBAAiB,WAAW,IAAI,OAAO;GACnD;GACD,EAAE,CAGY,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;;;;;CAUjE,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,AAAiB,aAAa;;;;;;;;;;;CAa9B,MAAM,QAAQ,WAAgD,EAAE,EAAiB;AAC/E,QAAM,KAAK,uBAAuB;;;;;CAMpC,MAAM,gBAAgB,UAA0C;AAC9D,QAAM,KAAK,uBAAuB;;;;;;;;;;;CAYpC,MAAM,MAAM,MAAc,UAAwB,EAAE,EAAwB;AAM1E,MAAI,CAAC,KAAK,cAAc,CACtB,OAAM,IAAI,MACR,4FACD;EAGH,MAAM,SAAS,MAAM,KAAK,uBAAuB;EACjD,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,MAAM,MAAM,OAAO,MAAM,MAAM,EAAE,CAAC;AACxC,SAAO;GACL,OAAO,IAAI;GACX,YAAY,IAAI;GAChB,UAAU,IAAI;GACd,OAAO,QAAQ,SAAS;GACxB,WAAW,YAAY,KAAK,GAAG;GAChC;;;CAIH,MAAc,wBAAsC;AAClD,MAAI,CAAC,KAAK,iBAAiB;GACzB,MAAM,EAAE,cAAc,mBAAmB,MAAM,OAAO;AACtD,QAAK,kBAAkB,MAAM,aAAa,OAAO,EAAE,MAAM,eAAe,KAAK,CAAC;;AAEhF,SAAO,KAAK;;;;;;CAOd,OAAO,YACL,MACA,UAAwB,EAAE,EACwB;EAClD,MAAM,SAAS,MAAM,KAAK,MAAM,MAAM,QAAQ;AAC9C,QAAM;GACJ,SAAS,OAAO;GAChB,YAAY,OAAO;GACnB,OAAO;GACP,SAAS;GACV;AACD,SAAO;;;;;CAMT,aAA0B;AACxB,SAAO;;;;;CAMT,cAAuB;AACrB,SAAO,KAAK,oBAAoB;;;;;CAMlC,kBAAqF;AACnF,MAAI,CAAC,KAAK,gBACR,QAAO;AAET,SAAO;GACL,IAAI,KAAK;GACT,QAAQ;GACR,QAAQ;GACT;;;;;CAOH,MAAM,gBAEJ;AACA,SAAO,CACL;GACE,IAAI,KAAK;GACT,aAAa;GACb,YAAY;GACZ,YAAY;GACb,CACF;;;;;;;;;;;;CAkBH,MAAM,QAAQ,UAAmB,WAA2B,EAAE,EAAiB;AAE7E,QAAM,KAAK,iBAAiB;;;;;CAM9B,MAAa,gBAAgB,UAAmB,UAA0C;AACxF,QAAM,KAAK,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;CA0B9B,MAAM,WACJ,OACA,UAA6B,EAAE,EACJ;AAK3B,MAAI,CAAC,KAAK,cAAc,CACtB,OAAM,IAAI,MACR,yFACD;AAEH,MAAI,EAAE,iBAAiB,cACrB,OAAM,IAAI,MACR,6FACD;AAEH,MAAI,QAAQ,WACV,OAAM,IAAI,MAAM,oDAAoD;EAGtE,MAAM,SAAS,MAAM,KAAK,iBAAiB;EAC3C,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,MAAM,MAAM,OAAO,WAAW,MAAM;AAC1C,SAAO;GACL,MAAM,IAAI;GACV,UAAU,QAAQ,YAAY;GAC9B,UAAU,IAAI;GACd,WAAW,YAAY,KAAK,GAAG;GAChC;;;CAIH,MAAc,kBAAgC;AAC5C,MAAI,CAAC,KAAK,WAAW;GACnB,MAAM,EAAE,iBAAiB,MAAM,OAAO;AACtC,QAAK,YAAY,MAAM,aAAa,QAAQ;;AAE9C,SAAO,KAAK;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA+Bd,MAAM,6BACJ,WAA0C,EAAE,EACJ;AACxC,QAAM,IAAI,MACR,0IAED;;;;;CAOH,MAAM,gBAA2C;AAC/C,SAAO,CACL;GACE,IAAI;GACJ,MAAM;GACN,aAAa;GACb,MAAM;GACN,cAAc;GACd,WAAW,CAAC,KAAK;GACjB,YAAY;GACb,CACF;;;;;CAMH,cAAuB;AACrB,SAAO,KAAK,cAAc;;;;;CAM5B,kBAAqF;AACnF,MAAI,CAAC,KAAK,UACR,QAAO;AAET,SAAO;GACL,IAAI;GACJ,QAAQ;GACR,QAAQ;GACT;;;;;;;;;;;;;;;;;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;;;;;;;CAaX,MAAM,QAAQ,cAAc,OAAsB;AAEhD,MAAI,KAAK,cAAc;AACrB,OAAI;AACF,SAAK,aAAa,SAAS;WACrB;AAGR,QAAK,eAAe;;AAQtB,MAAI,KAAK,mBAAmB;AAC1B,OAAI;AACF,SAAK,kBAAkB,SAAS;WAC1B;AAGR,QAAK,oBAAoB;AACzB,QAAK,kBAAkB;;AAEzB,MAAI,KAAK,WAAW;AAClB,OAAI;AACF,SAAK,UAAU,WAAW;WACpB;AAGR,QAAK,YAAY;;AAEnB,MAAI,KAAK,iBAAiB;AACxB,OAAI;AACF,SAAK,gBAAgB,SAAS;WACxB;AAGR,QAAK,kBAAkB;;AAGzB,OAAK,eAAe;AACpB,OAAK,cAAc;AACnB,OAAK,gBAAgB;;;;;CAOvB,aAAa,WAA0B;CAQvC,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"}
+{"version":3,"file":"gerbil-CZYYq4Sq.mjs","names":["globalCache: ResponseCache | null","BUILTIN_MODELS: Record<string, ModelConfig>","FAMILY_CONTEXT_DEFAULTS: Record<string, number>","family: ModelConfig[\"family\"]","idat: Uint8Array[]","v: number","KOKORO_VOICES_DEFAULT: VoiceInfo[]","concurrency: ConcurrencyConfig","result","result: GenerateResult","tokenQueue: string[]","resolveNext: ((value: string | null) => void) | null","results: EmbedResult[]"],"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// Canonical default model\n// ============================================\n\n/**\n * The default model used everywhere a model id is not explicitly provided\n * (CLI flags, REPL, framework adapters, integrations, one-liner). This is the\n * e2e-validated model; reference this constant instead of hard-coding the id.\n */\nexport const DEFAULT_MODEL = \"qwen3.5-0.8b\";\n\n// ============================================\n// Built-in Models (curated & tested)\n// ============================================\n\n// Every entry is a standard HuggingFace safetensors repo whose architecture the\n// native WebGPU engine supports (Qwen2/Qwen3/Qwen3.5, LFM2 — see\n// src/gpu/architectures/index.ts). The engine quantizes weights to INT4 on load;\n// `size` is the bf16/fp16 download size. Only add a repo here whose architecture\n// has a graph generator in the registry.\nexport const BUILTIN_MODELS: Record<string, ModelConfig> = {\n  \"qwen3.5-0.8b\": {\n    id: \"qwen3.5-0.8b\",\n    repo: \"Qwen/Qwen3.5-0.8B\",\n    description:\n      \"Qwen3.5 0.8B - Fast, multimodal (vision), 262k context, supports thinking (default)\",\n    size: \"~1.6GB\",\n    contextLength: 262_144,\n    supportsThinking: true,\n    supportsJson: true,\n    supportsVision: true,\n    family: \"qwen\",\n  },\n  \"qwen3.5-2b\": {\n    id: \"qwen3.5-2b\",\n    repo: \"Qwen/Qwen3.5-2B\",\n    description:\n      \"Qwen3.5 2B - Higher quality, multimodal (vision), 262k context, supports thinking\",\n    size: \"~4GB\",\n    contextLength: 262_144,\n    supportsThinking: true,\n    supportsJson: true,\n    supportsVision: true,\n    family: \"qwen\",\n  },\n  \"lfm2.5-1.2b-thinking\": {\n    id: \"lfm2.5-1.2b-thinking\",\n    repo: \"LiquidAI/LFM2.5-1.2B-Thinking\",\n    description: \"LFM2.5 1.2B Thinking - Efficient reasoning model, 128k context\",\n    size: \"~2.4GB\",\n    contextLength: 128_000,\n    supportsThinking: true,\n    supportsJson: false,\n    family: \"other\",\n  },\n};\n\n// ============================================\n// Model Resolution\n// ============================================\n\n/**\n * Parse model identifier and resolve to source\n *\n * Supported formats:\n * - \"qwen3.5-0.8b\" (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 for the families the native engine actually supports\n// (a graph generator exists in src/gpu/architectures). Other families fall back\n// to a conservative default.\nconst FAMILY_CONTEXT_DEFAULTS: Record<string, number> = {\n  qwen: 32_768,\n  other: 32_768, // LFM2 supports 128k but config.json is the real source of truth\n};\n\n/**\n * Create model config for an external HuggingFace model.\n *\n * Inference is restricted to families the engine can actually run — Qwen\n * (Qwen2/Qwen3/Qwen3.5) and LFM2 (Liquid). Everything else is left as \"other\"\n * with conservative capability flags so the REPL doesn't advertise features the\n * engine can't deliver.\n */\nexport function createExternalModelConfig(\n  modelId: string,\n  repo: string,\n  contextLength?: number,\n): ModelConfig {\n  const repoLower = repo.toLowerCase();\n\n  // Only infer families that have a graph generator in the registry.\n  let family: ModelConfig[\"family\"] = \"other\";\n  if (repoLower.includes(\"qwen\")) {\n    family = \"qwen\";\n  }\n\n  const isLiquid = repoLower.includes(\"lfm\") || repoLower.includes(\"liquid\");\n  const isQwen = family === \"qwen\";\n\n  return {\n    id: modelId,\n    repo,\n    description: `External model: ${repo}`,\n    size: \"Unknown\",\n    contextLength: contextLength || FAMILY_CONTEXT_DEFAULTS[family] || 32_768,\n    // Qwen3/Qwen3.5 and LFM2.5-Thinking expose thinking; nothing here is vision.\n    supportsThinking: isQwen || isLiquid,\n    supportsJson: isQwen,\n    family,\n  };\n}\n\n/**\n * Fetch context length from HuggingFace model config\n */\nexport async function fetchModelContextLength(repo: string): Promise<number | null> {\n  try {\n    const res = await fetch(`https://huggingface.co/${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 { existsSync } from \"node:fs\";\nimport os from \"node:os\";\nimport path from \"node:path\";\nimport zlib from \"node:zlib\";\nimport PQueue from \"p-queue\";\nimport { generateCacheKey, getGlobalCache } from \"./cache.js\";\nimport {\n  BUILTIN_MODELS,\n  createExternalModelConfig,\n  DEFAULT_MODEL,\n  fetchModelContextLength,\n  getModelConfig,\n  resolveModel,\n} from \"./models.js\";\nimport type {\n  AudioChunk,\n  ConcurrencyConfig,\n  EmbedOptions,\n  EmbedResult,\n  ErrorContext,\n  GenerateOptions,\n  GenerateResult,\n  GerbilConfig,\n  JsonOptions,\n  LoadOptions,\n  LoadSTTOptions,\n  LoadTTSOptions,\n  ModelConfig,\n  PreloadOptions,\n  SessionStats,\n  SpeakOptions,\n  SpeakResult,\n  STTModelConfig,\n  StreamingTranscriptionOptions,\n  StreamingTranscriptionSession,\n  SystemInfo,\n  TelemetryConfig,\n  TranscribeOptions,\n  TranscribeResult,\n  VoiceInfo,\n} from \"./types.js\";\n\n/**\n * Minimal PNG decoder: 8-bit, non-interlaced, color type 2 (RGB) or 6 (RGBA).\n * Returns packed RGB pixels for the native vision encoder. Replaces the\n * transformers.js RawImage decoder for the common PNG case.\n */\nfunction decodePng(buf: Uint8Array): { pixels: Uint8Array; width: number; height: number } {\n  const view = new DataView(buf.buffer, buf.byteOffset, buf.byteLength);\n  const w = view.getUint32(16);\n  const h = view.getUint32(20);\n  const bitDepth = buf[24];\n  const colorType = buf[25];\n  if (bitDepth !== 8 || (colorType !== 2 && colorType !== 6)) {\n    throw new Error(\n      `Unsupported PNG (bitDepth=${bitDepth} colorType=${colorType}); need 8-bit RGB/RGBA`,\n    );\n  }\n  const channels = colorType === 6 ? 4 : 3;\n  let off = 8;\n  const idat: Uint8Array[] = [];\n  while (off < buf.length) {\n    const len = view.getUint32(off);\n    const type = String.fromCharCode(buf[off + 4], buf[off + 5], buf[off + 6], buf[off + 7]);\n    if (type === \"IDAT\") {\n      idat.push(buf.subarray(off + 8, off + 8 + len));\n    }\n    off += 12 + len;\n    if (type === \"IEND\") {\n      break;\n    }\n  }\n  const raw = zlib.inflateSync(Buffer.concat(idat));\n  const stride = w * channels;\n  const out = new Uint8Array(w * h * 3);\n  const line = new Uint8Array(stride);\n  const prev = new Uint8Array(stride);\n  let p = 0;\n  for (let y = 0; y < h; y += 1) {\n    const filter = raw[p];\n    p += 1;\n    for (let i = 0; i < stride; i += 1) {\n      const x = raw[p];\n      p += 1;\n      const a = i >= channels ? line[i - channels] : 0;\n      const b = prev[i];\n      const c = i >= channels ? prev[i - channels] : 0;\n      let v: number;\n      switch (filter) {\n        case 0:\n          v = x;\n          break;\n        case 1:\n          v = x + a;\n          break;\n        case 2:\n          v = x + b;\n          break;\n        case 3:\n          v = x + ((a + b) >> 1);\n          break;\n        case 4: {\n          const pp = a + b - c;\n          const pa = Math.abs(pp - a);\n          const pb = Math.abs(pp - b);\n          const pc = Math.abs(pp - c);\n          let pred = c;\n          if (pa <= pb && pa <= pc) {\n            pred = a;\n          } else if (pb <= pc) {\n            pred = b;\n          }\n          v = x + pred;\n          break;\n        }\n        default:\n          throw new Error(`bad PNG filter ${filter}`);\n      }\n      line[i] = v & 0xff;\n    }\n    for (let x = 0; x < w; x += 1) {\n      out[(y * w + x) * 3 + 0] = line[x * channels + 0];\n      out[(y * w + x) * 3 + 1] = line[x * channels + 1];\n      out[(y * w + x) * 3 + 2] = line[x * channels + 2];\n    }\n    prev.set(line);\n  }\n  return { pixels: out, width: w, height: h };\n}\n\n// Default voices for listVoices() when native 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 type { GenerateObjectOptions, GenerateObjectResult } from \"../gpu/index.js\";\nimport { extractJson, zodToJsonSchema } from \"./utils.js\";\n\n// ============================================\n// Gerbil Class\n// ============================================\n\nexport class Gerbil {\n  private currentModel: string | null = null;\n  private modelConfig: ModelConfig | null = null;\n  private readonly config: GerbilConfig;\n  private stats: SessionStats;\n  private _deviceMode: \"webgpu\" | \"cpu\" | \"wasm\" = \"cpu\"; // Track which backend is active\n  private webgpuEngine: any = null; // Native WebGPU engine (src/gpu)\n\n  // ── Native (src/gpu) capability engines, lazily created and independent of the\n  // loaded text model. These let embed / transcribe / speak run on the native\n  // WebGPU engine even when the active model isn't that capability. They are torn\n  // down in dispose().\n  private nativeEmbedEngine: any = null; // Native WebGPUEngine loaded as an embedder\n  private nativeEmbedRepo: string | null = null; // Repo backing nativeEmbedEngine (for cache-busting on model switch)\n  private nativeSTT: any = null; // Native MoonshineSTT (src/gpu/moonshine-stt)\n  private nativeTTSEngine: any = null; // Native WebGPUEngine (Kani-TTS) backing speak()\n\n  private isVisionModel = false; // Whether current model supports vision\n\n  // Request queue for concurrency control (prevents GPU OOM under load)\n  private readonly queue: PQueue;\n  private readonly telemetry: TelemetryConfig;\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    // Initialize request queue (default: 1 concurrent for LLM, 5 min timeout)\n    const concurrency: ConcurrencyConfig = config.concurrency || {};\n    this.queue = new PQueue({\n      concurrency: concurrency.maxConcurrent ?? 1,\n      timeout: concurrency.timeout ?? 300_000, // 5 minutes (auto-throws on timeout)\n    });\n\n    // Store telemetry hooks\n    this.telemetry = config.telemetry || {};\n  }\n\n  // ============================================\n  // Telemetry Helpers\n  // ============================================\n\n  private reportError(error: Error, context: ErrorContext): void {\n    try {\n      this.telemetry.onError?.(error, context);\n    } catch {\n      // Never throw from telemetry\n    }\n  }\n\n  /**\n   * Whether the native (src/gpu) WebGPU engine should be used for a capability\n   * (embed / transcribe / speak / vision). The native WebGPU engine is the only\n   * inference backend, so this is always true; kept as a seam for callers.\n   */\n  private preferNative(): boolean {\n    return true;\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(\"qwen3.5-0.8b\");\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 = DEFAULT_MODEL, options: LoadOptions = {}): Promise<void> {\n    const loadStartTime = performance.now();\n\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    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    // The native WebGPU engine is the only inference backend. CPU/WASM and the\n    // legacy ONNX path have been removed — guard against legacy runtime values.\n    if ((device as string) === \"cpu\" || (device as string) === \"gpu\") {\n      throw new Error(\n        'Gerbil requires WebGPU. CPU/WASM and the legacy ONNX backend have been removed; use device \"webgpu\" or \"auto\".',\n      );\n    }\n\n    onProgress?.({ status: `Loading ${modelId}...` });\n\n    try {\n      onProgress?.({ status: \"Initializing WebGPU engine...\" });\n      const { WebGPUEngine } = await import(\"../gpu/index.js\");\n\n      // Resolve the HF repo — for built-in models, use the standard HF repo\n      // (not the ONNX repo) since the native engine loads safetensors directly\n      let hfRepo = source.path;\n      // If it points to an ONNX repo, map to the standard repo\n      if (hfRepo.includes(\"onnx-community/\") || hfRepo.includes(\"-ONNX\")) {\n        const nativeRepoMap: Record<string, string> = {\n          \"onnx-community/Qwen3.5-0.8B-ONNX\": \"Qwen/Qwen3.5-0.8B\",\n          \"onnx-community/Qwen3-0.6B-ONNX\": \"Qwen/Qwen3-0.6B\",\n          \"onnx-community/Qwen3-1.7B-ONNX\": \"Qwen/Qwen3-1.7B\",\n          \"onnx-community/Qwen3.5-2B-ONNX\": \"Qwen/Qwen3.5-2B\",\n        };\n        hfRepo = nativeRepoMap[hfRepo] || hfRepo;\n      }\n\n      // Map user dtype to GPU engine dtype (\"q4\" for INT4, \"f32\" default)\n      const gpuDtype = userDtype === \"q4\" ? (\"q4\" as const) : undefined;\n\n      this.webgpuEngine = await WebGPUEngine.create({\n        repo: hfRepo,\n        maxSeqLen: options.contextLength ?? config.contextLength ?? 4096,\n        dtype: gpuDtype,\n        onProgress: (loaded, total, message) => {\n          onProgress?.({\n            status: message,\n            progress: total > 0 ? Math.round((loaded / total) * 100) : undefined,\n          });\n        },\n      });\n\n      this._deviceMode = \"webgpu\";\n      this.isVisionModel = false;\n      this.currentModel = modelId;\n      this.modelConfig = config;\n      onProgress?.({ status: \"Ready (WebGPU Native)!\" });\n\n      // Report successful model load\n      if (this.telemetry.onModelLoad) {\n        try {\n          this.telemetry.onModelLoad({\n            modelId,\n            loadTimeMs: performance.now() - loadStartTime,\n            fromCache: false, // TODO: detect if loaded from cache\n            device: this._deviceMode,\n            success: true,\n          });\n        } catch {\n          // Never throw from telemetry\n        }\n      }\n    } catch (err) {\n      // No CPU/ONNX fallback — surface the load failure.\n      this.reportError(err instanceof Error ? err : new Error(String(err)), {\n        operation: \"load\",\n        modelId,\n      });\n      if (this.telemetry.onModelLoad) {\n        try {\n          this.telemetry.onModelLoad({\n            modelId,\n            loadTimeMs: performance.now() - loadStartTime,\n            fromCache: false,\n            device: this._deviceMode,\n            success: false,\n            error: err instanceof Error ? err.message : String(err),\n          });\n        } catch {\n          // Never throw from telemetry\n        }\n      }\n      throw err;\n    }\n  }\n\n  /**\n   * Load a vision model (VLM) on the native WebGPU engine.\n   * The native engine loads the vision-capable safetensors checkpoint directly\n   * and builds its ViT tower on demand (enableVision: true). describeImage() then\n   * runs encode → splice → decode entirely in WebGPU compute.\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    // The native WebGPU engine is the only backend. CPU/WASM/ONNX are unsupported.\n    if ((device as string) === \"cpu\" || (device as string) === \"gpu\") {\n      throw new Error(\n        'Gerbil vision models require WebGPU. CPU/WASM and the legacy ONNX backend have been removed; use device \"webgpu\" or \"auto\".',\n      );\n    }\n\n    onProgress?.({ status: \"Initializing WebGPU vision engine...\" });\n    const { WebGPUEngine } = await import(\"../gpu/index.js\");\n    // Map an ONNX vision repo to its standard HF checkpoint when possible;\n    // otherwise honor the configured path (defaults resolve inside create()).\n    let visRepo = repoPath;\n    if (visRepo.includes(\"onnx-community/\") || visRepo.includes(\"-ONNX\")) {\n      const map: Record<string, string> = {\n        \"onnx-community/Qwen3.5-0.8B-ONNX\": \"Qwen/Qwen3.5-0.8B\",\n      };\n      visRepo = map[visRepo] || visRepo;\n    }\n    this.webgpuEngine = await WebGPUEngine.create({\n      repo: visRepo,\n      enableVision: true,\n      maxSeqLen: options.contextLength ?? config.contextLength ?? 4096,\n      onProgress: (loaded, total, message) =>\n        onProgress?.({\n          status: message,\n          progress: total > 0 ? Math.round((loaded / total) * 100) : undefined,\n        }),\n    });\n    this._deviceMode = \"webgpu\";\n    this.isVisionModel = true;\n    this.currentModel = modelId;\n    this.modelConfig = config;\n    onProgress?.({ status: \"Ready (Vision, WebGPU Native)!\" });\n  }\n\n  /**\n   * Check if a model is loaded\n   */\n  isLoaded(): boolean {\n    return this.webgpuEngine !== null;\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 the in-memory weight quantization the native engine uses for the loaded\n   * model. The WebGPU engine quantizes weights to INT4 (\"q4\") on load; the KV\n   * cache precision (f16/f32) is separate and device-detected.\n   */\n  getDtype(): string {\n    return \"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  // Model Preloading & Cache Checking\n  // ============================================\n\n  /**\n   * Check if a model is cached (downloaded) without loading it\n   *\n   * @example\n   * ```ts\n   * if (await g.isModelCached(\"qwen3.5-0.8b\")) {\n   *   console.log(\"Model ready, will load instantly\");\n   * } else {\n   *   console.log(\"Model needs to download (~400MB)\");\n   * }\n   * ```\n   */\n  async isModelCached(modelId: string): Promise<boolean> {\n    const source = resolveModel(modelId);\n    return this.isNativeRepoCached(source.path);\n  }\n\n  /**\n   * Check whether the native WebGPU engine has a repo cached on disk.\n   * The native loader stores files under ~/.cache/gerbil/<repo>/<revision>/.\n   */\n  private isNativeRepoCached(repo: string, revision = \"main\"): boolean {\n    try {\n      const home = process.env.HOME || process.env.USERPROFILE || os.homedir();\n      if (!home) {\n        return false;\n      }\n      const modelDir = path.join(home, \".cache\", \"gerbil\", repo.replace(/\\//g, \"_\"), revision);\n      // config.json is the first file the native loader writes for any repo.\n      return existsSync(path.join(modelDir, \"config.json\".replace(/\\//g, \"_\")));\n    } catch {\n      return false;\n    }\n  }\n\n  /**\n   * Preload a model (download without initializing for inference)\n   *\n   * Use this to download models ahead of time, e.g., during app startup,\n   * so users don't wait when they first use AI.\n   *\n   * @example\n   * ```ts\n   * // Preload for later (download only, free memory)\n   * await g.preloadModel(\"qwen3.5-0.8b\", {\n   *   onProgress: (p) => console.log(p.status, p.progress),\n   * });\n   *\n   * // Preload and keep in memory for instant use\n   * await g.preloadModel(\"qwen3.5-0.8b\", { keepLoaded: true });\n   * await g.generate(\"Hello\"); // Instant, no loading needed\n   * ```\n   */\n  async preloadModel(modelId: string, options: PreloadOptions = {}): Promise<void> {\n    const source = resolveModel(modelId);\n    const { onProgress, keepLoaded = false } = options;\n\n    // Already loaded? Nothing to do\n    if (keepLoaded && this.isLoaded() && this.currentModel === modelId) {\n      onProgress?.({ status: \"Model already loaded\" });\n      return;\n    }\n\n    // Already cached and not keeping loaded? Nothing to do\n    if (!keepLoaded && (await this.isModelCached(modelId))) {\n      onProgress?.({ status: \"Model already cached\" });\n      return;\n    }\n\n    // If keepLoaded, use the regular loadModel path\n    if (keepLoaded) {\n      await this.loadModel(modelId, { onProgress });\n      return;\n    }\n\n    onProgress?.({ status: `Preloading ${modelId}...` });\n\n    // The native engine has no \"download-only\" mode, so load it (which fetches\n    // and caches all weight files to ~/.cache/gerbil) then dispose to free GPU\n    // memory. The cached files remain on disk for an instant subsequent load.\n    await this.loadModel(modelId, { onProgress });\n    await this.dispose();\n\n    onProgress?.({ status: \"Preload complete\" });\n  }\n\n  /**\n   * Check if the native TTS model is cached. The native engine always uses the\n   * Kani-TTS-2 checkpoint, so `modelId` is accepted for API compatibility only.\n   */\n  async isTTSCached(_modelId?: string): Promise<boolean> {\n    const { DEFAULT_MODELS } = await import(\"../gpu/index.js\");\n    return this.isNativeRepoCached(DEFAULT_MODELS.tts);\n  }\n\n  /**\n   * Preload the native TTS model (downloads Kani-TTS-2 weights to disk cache).\n   *\n   * @param modelId - Accepted for API compatibility; native TTS uses Kani-TTS-2.\n   * @param options.keepLoaded - Keep the engine in memory for instant use.\n   */\n  async preloadTTS(modelId?: string, options: PreloadOptions = {}): Promise<void> {\n    const { onProgress, keepLoaded = false } = options;\n\n    if (keepLoaded && this.isTTSLoaded()) {\n      onProgress?.({ status: \"TTS model already loaded\" });\n      return;\n    }\n\n    if (!keepLoaded && (await this.isTTSCached(modelId))) {\n      onProgress?.({ status: \"TTS model already cached\" });\n      return;\n    }\n\n    onProgress?.({ status: \"Preloading TTS model...\" });\n    // Build the native TTS engine (downloads + caches weights).\n    await this.ensureNativeTTSEngine();\n\n    if (!keepLoaded && this.nativeTTSEngine) {\n      try {\n        this.nativeTTSEngine.destroy();\n      } catch {\n        // best-effort\n      }\n      this.nativeTTSEngine = null;\n    }\n    onProgress?.({ status: \"Preload complete\" });\n  }\n\n  /**\n   * Check if the native STT model is cached. The native engine always uses the\n   * Moonshine checkpoint, so `modelId` is accepted for API compatibility only.\n   */\n  async isSTTCached(_modelId?: string): Promise<boolean> {\n    const { DEFAULT_MODELS } = await import(\"../gpu/index.js\");\n    return this.isNativeRepoCached(DEFAULT_MODELS.stt);\n  }\n\n  /**\n   * Preload the native STT model (downloads Moonshine weights to disk cache).\n   *\n   * @param modelId - Accepted for API compatibility; native STT uses Moonshine.\n   * @param options.keepLoaded - Keep the engine in memory for instant use.\n   */\n  async preloadSTT(modelId?: string, options: PreloadOptions = {}): Promise<void> {\n    const { onProgress, keepLoaded = false } = options;\n\n    if (keepLoaded && this.isSTTLoaded()) {\n      onProgress?.({ status: \"STT model already loaded\" });\n      return;\n    }\n\n    if (!keepLoaded && (await this.isSTTCached(modelId))) {\n      onProgress?.({ status: \"STT model already cached\" });\n      return;\n    }\n\n    onProgress?.({ status: \"Preloading STT model...\" });\n    // Build the native STT engine (downloads + caches weights).\n    await this.ensureNativeSTT();\n\n    if (!keepLoaded && this.nativeSTT) {\n      try {\n        this.nativeSTT.destroy?.();\n      } catch {\n        // best-effort\n      }\n      this.nativeSTT = null;\n    }\n    onProgress?.({ status: \"Preload complete\" });\n  }\n\n  /**\n   * Check if a native embedding model is cached. Defaults to the native\n   * EmbeddingGemma checkpoint when no repo is provided.\n   */\n  async isEmbeddingCached(modelId?: string): Promise<boolean> {\n    const { DEFAULT_MODELS } = await import(\"../gpu/index.js\");\n    return this.isNativeRepoCached(modelId || DEFAULT_MODELS.embedding);\n  }\n\n  /**\n   * Preload a native embedding model (downloads weights to disk cache).\n   *\n   * @param modelId - Embedding repo (default: native EmbeddingGemma).\n   * @param options.keepLoaded - Keep the engine in memory for instant use.\n   */\n  async preloadEmbedding(modelId?: string, options: PreloadOptions = {}): Promise<void> {\n    const { onProgress, keepLoaded = false } = options;\n\n    if (keepLoaded && this.nativeEmbedEngine) {\n      onProgress?.({ status: \"Embedding model already loaded\" });\n      return;\n    }\n\n    if (!keepLoaded && (await this.isEmbeddingCached(modelId))) {\n      onProgress?.({ status: \"Embedding model already cached\" });\n      return;\n    }\n\n    onProgress?.({ status: \"Preloading embedding model...\" });\n    // Build the native embedding engine (downloads + caches weights).\n    await this.ensureNativeEmbedEngine(modelId);\n\n    if (!keepLoaded && this.nativeEmbedEngine) {\n      try {\n        this.nativeEmbedEngine.destroy();\n      } catch {\n        // best-effort\n      }\n      this.nativeEmbedEngine = null;\n      this.nativeEmbedRepo = null;\n    }\n    onProgress?.({ status: \"Preload complete\" });\n  }\n\n  /**\n   * Clear KV cache to free memory.\n   * The native engine manages its own KV cache; this is a no-op kept for API\n   * compatibility.\n   */\n  // biome-ignore lint/suspicious/useAwait: kept async for API compatibility\n  async clearCache(): Promise<void> {\n    // no-op\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    const queueStartTime = performance.now();\n\n    try {\n      // Queue the generation request (prevents GPU OOM under concurrent load)\n      const result = await this.queue.add(async () => {\n        const queueWaitTime = performance.now() - queueStartTime;\n\n        // Report queue wait time if significant (>100ms)\n        if (queueWaitTime > 100 && this.telemetry.onQueueWait) {\n          try {\n            this.telemetry.onQueueWait(queueWaitTime);\n          } catch {\n            // Never throw from telemetry\n          }\n        }\n\n        const generatedResult = await this.generateInternal(prompt, options);\n\n        // Report successful generation\n        if (this.telemetry.onGenerate) {\n          try {\n            this.telemetry.onGenerate({\n              modelId: this.currentModel || \"unknown\",\n              result: generatedResult,\n              cached: generatedResult.cached ?? false,\n              queueTimeMs: queueWaitTime > 100 ? queueWaitTime : undefined,\n            });\n          } catch {\n            // Never throw from telemetry\n          }\n        }\n\n        return generatedResult;\n      });\n\n      return result as GenerateResult;\n    } catch (error) {\n      // Report errors to telemetry (including queue timeouts)\n      if (this.telemetry.onError) {\n        try {\n          this.telemetry.onError(error instanceof Error ? error : new Error(String(error)), {\n            method: \"generate\",\n            modelId: this.currentModel || \"unknown\",\n            prompt: prompt.slice(0, 100), // Truncate for privacy\n            queueWaitTime: performance.now() - queueStartTime,\n          });\n        } catch {\n          // Never throw from telemetry\n        }\n      }\n      throw error;\n    }\n  }\n\n  /**\n   * Internal generate implementation (called within queue)\n   */\n  private async generateInternal(\n    prompt: string,\n    options: GenerateOptions = {},\n  ): Promise<GenerateResult> {\n    if (!this.isLoaded()) {\n      // Auto-load default model\n      await this.loadModel(this.config.model || DEFAULT_MODEL);\n    }\n\n    const { images } = options;\n\n    // Route to native vision generation when images are provided and the loaded\n    // model supports vision.\n    if (images?.length && this.isVisionModel) {\n      return this.generateWithVision(prompt, options);\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      // Real token counts come from the engine, not a length heuristic.\n      let engineTokensGenerated = 0;\n      let engineTokensPerSecond = 0;\n\n      if (this.webgpuEngine) {\n        // Native WebGPU engine — pure compute shaders, no ONNX\n        const result = await this.webgpuEngine.generate(prompt, {\n          maxTokens,\n          sampling: {\n            temperature,\n            topP,\n            topK,\n          },\n          systemPrompt: system,\n          onToken: options.onToken ? (t: string) => options.onToken?.(t) : undefined,\n        });\n        rawText = result.text;\n        engineTokensGenerated = result.tokensGenerated;\n        engineTokensPerSecond = result.tokensPerSecond;\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      // Use the engine's true decoded-token count and throughput.\n      const tokensGenerated = engineTokensGenerated;\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: engineTokensPerSecond,\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      // Report error to telemetry\n      this.reportError(error instanceof Error ? error : new Error(String(error)), {\n        operation: \"generate\",\n        modelId: this.currentModel || undefined,\n      });\n\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 || DEFAULT_MODEL);\n    }\n\n    const startTime = performance.now();\n\n    // For native WebGPU engine, use real streaming via onToken callback\n    if (this.webgpuEngine) {\n      let fullText = \"\";\n      const tokenQueue: string[] = [];\n      let resolveNext: ((value: string | null) => void) | null = null;\n      let done = false;\n      // Capture the engine's true token count / throughput for the final result.\n      let engineTokensGenerated = 0;\n      let engineTokensPerSecond = 0;\n\n      const generatePromise = this.webgpuEngine\n        .generate(prompt, {\n          ...options,\n          sampling: {\n            temperature: options.temperature,\n            topP: options.topP,\n            topK: options.topK,\n          },\n          systemPrompt: options.system,\n          onToken: (token: string) => {\n            fullText += token;\n            if (resolveNext) {\n              resolveNext(token);\n              resolveNext = null;\n            } else {\n              tokenQueue.push(token);\n            }\n          },\n        })\n        .then((result: { tokensGenerated: number; tokensPerSecond: number }) => {\n          engineTokensGenerated = result.tokensGenerated;\n          engineTokensPerSecond = result.tokensPerSecond;\n          done = true;\n          if (resolveNext) {\n            resolveNext(null);\n          }\n        })\n        .catch((err: Error) => {\n          done = true;\n          if (resolveNext) {\n            resolveNext(null);\n          }\n          throw err;\n        });\n\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 totalTime = performance.now() - startTime;\n\n      return {\n        text: response,\n        thinking: options.thinking ? thinkingText : undefined,\n        tokensGenerated: engineTokensGenerated,\n        totalTime,\n        tokensPerSecond: engineTokensPerSecond,\n        finishReason: \"stop\" as const,\n      };\n    }\n\n    // The native WebGPU engine is the only backend. If we reach here, no model\n    // is loaded (auto-load happens at the top of stream()).\n    throw new Error(\"No model loaded\");\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    // ── Native WebGPU vision ──\n    // loadModel() brought up the native vision engine (this.webgpuEngine has its\n    // ViT tower built via enableVision: true). Decode the image to RGB pixels and\n    // run the native encode → splice → decode path (engine.describeImage). The\n    // native engine currently supports a single image per request.\n    if (!(this.webgpuEngine && typeof this.webgpuEngine.describeImage === \"function\")) {\n      throw new Error(\n        \"Vision model not loaded. Load a vision-capable model with device 'webgpu' first.\",\n      );\n    }\n\n    const imgs = options.images ?? [];\n    if (imgs.length !== 1) {\n      throw new Error(\n        `Native WebGPU vision supports exactly one image per request (got ${imgs.length}).`,\n      );\n    }\n\n    const startTime = performance.now();\n    const { pixels, width, height } = await this.decodeImageToPixels(imgs[0].source);\n    const result = await this.webgpuEngine.describeImage({ pixels, width, height }, prompt, {\n      maxTokens: options.maxTokens ?? 512,\n      sampling: {\n        temperature: options.temperature ?? 0.7,\n        topP: options.topP ?? 0.9,\n        topK: options.topK ?? 20,\n      },\n      onToken: options.onToken ? (t: string) => options.onToken?.(t) : undefined,\n    });\n    const totalTime = performance.now() - startTime;\n\n    // Update stats\n    this.stats.prompts += 1;\n    this.stats.tokensOut += result.tokensGenerated;\n    this.stats.totalTime += totalTime;\n    this.stats.avgSpeed = (this.stats.tokensOut / this.stats.totalTime) * 1000;\n\n    return {\n      text: this.cleanOutput(result.text),\n      tokensGenerated: result.tokensGenerated,\n      tokensPerSecond: result.tokensPerSecond,\n      totalTime,\n      finishReason: \"stop\",\n      provider: \"local\",\n      cached: false,\n    };\n  }\n\n  /**\n   * Decode an image source (http(s) URL, file path, or data URI) to raw RGB\n   * pixels for the native vision encoder. Supports 8-bit non-interlaced PNG\n   * (color types 2/RGB and 6/RGBA). Other formats throw a clear error — callers\n   * can pre-decode and use the lower-level WebGPUEngine.describeImage() with\n   * pixels directly.\n   */\n  private async decodeImageToPixels(\n    source: string,\n  ): Promise<{ pixels: Uint8Array; width: number; height: number }> {\n    const bytes = await this.fetchImageBytes(source);\n    // PNG magic number: 89 50 4E 47 0D 0A 1A 0A\n    const isPng =\n      bytes.length > 8 &&\n      bytes[0] === 0x89 &&\n      bytes[1] === 0x50 &&\n      bytes[2] === 0x4e &&\n      bytes[3] === 0x47;\n    if (!isPng) {\n      throw new Error(\n        \"Native vision currently decodes PNG images only. For other formats, pre-decode \" +\n          \"to RGB pixels and call the GPU engine's describeImage() directly.\",\n      );\n    }\n    return decodePng(bytes);\n  }\n\n  /** Fetch an image source to raw bytes (URL, data URI, or local file path). */\n  private async fetchImageBytes(source: string): Promise<Uint8Array> {\n    if (source.startsWith(\"data:\")) {\n      const comma = source.indexOf(\",\");\n      const meta = source.slice(5, comma);\n      const data = source.slice(comma + 1);\n      if (meta.includes(\"base64\")) {\n        return Uint8Array.from(Buffer.from(data, \"base64\"));\n      }\n      return Uint8Array.from(Buffer.from(decodeURIComponent(data), \"binary\"));\n    }\n    if (source.startsWith(\"http://\") || source.startsWith(\"https://\")) {\n      const res = await fetch(source);\n      if (!res.ok) {\n        throw new Error(`Failed to fetch image (${res.status}): ${source}`);\n      }\n      return new Uint8Array(await res.arrayBuffer());\n    }\n    // Local file path (Node.js only)\n    const { readFile } = await import(\"node:fs/promises\");\n    return new Uint8Array(await readFile(source));\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   * Generate a structured object via the native engine's retrying\n   * `generateObject` (extract JSON → validate → retry with a nudge).\n   *\n   * Unlike {@link json} (which is Zod-driven), this passes through to the engine\n   * and accepts either a predicate validator `(o) => boolean` or a minimal\n   * `{ required: [...] }` schema; omit `schema` to accept any valid JSON.\n   *\n   * @example\n   * ```ts\n   * const { object } = await g.generateObject<{ name: string; age: number }>(\n   *   'Extract {name, age} from: \"I am Sarah, 28\"',\n   *   { schema: { required: [\"name\", \"age\"] } },\n   * );\n   * ```\n   */\n  async generateObject<T = unknown>(\n    prompt: string,\n    options: GenerateObjectOptions = {},\n  ): Promise<GenerateObjectResult<T>> {\n    if (!this.isLoaded()) {\n      await this.loadModel(this.config.model || DEFAULT_MODEL);\n    }\n    if (!this.webgpuEngine) {\n      throw new Error(\"No model loaded\");\n    }\n    return this.webgpuEngine.generateObject(prompt, options) as Promise<GenerateObjectResult<T>>;\n  }\n\n  // ============================================\n  // Embeddings\n  // ============================================\n\n  /**\n   * Generate embeddings\n   */\n  async embed(text: string, options: EmbedOptions = {}): Promise<EmbedResult> {\n    // ── Native WebGPU embedding ──\n    // The active text model may not be an embedder, so embeddings run on a\n    // dedicated native engine loaded with { embedding: true }. It is created\n    // lazily on first embed(), keyed by repo: if the caller asks for a different\n    // embedding model than the cached one, the old engine is torn down and a new\n    // one is built. The engine is owned by this Gerbil instance and released in\n    // dispose(). Native engine.embed() returns an L2-normalized Float32Array.\n    if (!this.preferNative()) {\n      throw new Error(\n        \"Embeddings require WebGPU. CPU/WASM and the legacy ONNX backend have been removed.\",\n      );\n    }\n\n    const native = await this.ensureNativeEmbedEngine(options.model);\n    const startTime = performance.now();\n    const vec = await native.embed(text);\n    return {\n      vector: Array.from(vec as Float32Array),\n      text,\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /**\n   * Lazily build (or reuse) the native embedding engine. Re-creates it when the\n   * requested repo differs from the cached one. The default native embedding\n   * model is resolved by the engine itself (EmbeddingGemma) when no repo given.\n   */\n  private async ensureNativeEmbedEngine(repo?: string): Promise<any> {\n    if (this.nativeEmbedEngine && (!repo || repo === this.nativeEmbedRepo)) {\n      return this.nativeEmbedEngine;\n    }\n    if (this.nativeEmbedEngine) {\n      // Switching embedding models — release the previous native engine.\n      try {\n        this.nativeEmbedEngine.destroy();\n      } catch {\n        // best-effort cleanup\n      }\n      this.nativeEmbedEngine = null;\n      this.nativeEmbedRepo = null;\n    }\n    const { WebGPUEngine } = await import(\"../gpu/index.js\");\n    this.nativeEmbedEngine = await WebGPUEngine.create({ repo, embedding: true });\n    this.nativeEmbedRepo = repo ?? null;\n    return this.nativeEmbedEngine;\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   * Compute cosine similarity between two vectors\n   *\n   * @example\n   * ```ts\n   * const sim = g.cosineSimilarity([1, 0, 0], [1, 0, 0]); // 1.0\n   * const sim2 = g.cosineSimilarity([1, 0, 0], [0, 1, 0]); // 0.0\n   * ```\n   */\n  cosineSimilarity(a: number[], b: number[]): number {\n    if (a.length !== b.length) {\n      throw new Error(`Vector dimensions must match: ${a.length} vs ${b.length}`);\n    }\n\n    let dotProduct = 0;\n    let normA = 0;\n    let normB = 0;\n\n    for (let i = 0; i < a.length; i++) {\n      dotProduct += a[i] * b[i];\n      normA += a[i] * a[i];\n      normB += b[i] * b[i];\n    }\n\n    const magnitude = Math.sqrt(normA) * Math.sqrt(normB);\n    if (magnitude === 0) return 0;\n\n    return dotProduct / magnitude;\n  }\n\n  /**\n   * Compare similarity between two texts\n   *\n   * @example\n   * ```ts\n   * const result = await g.similarity(\"Hello world\", \"Hi there\");\n   * console.log(result.score); // 0.85\n   * ```\n   */\n  async similarity(\n    textA: string,\n    textB: string,\n    options: EmbedOptions = {},\n  ): Promise<import(\"./types.js\").SimilarityResult> {\n    const startTime = performance.now();\n\n    const [embedA, embedB] = await Promise.all([\n      this.embed(textA, options),\n      this.embed(textB, options),\n    ]);\n\n    const score = this.cosineSimilarity(embedA.vector, embedB.vector);\n\n    return {\n      score,\n      textA,\n      textB,\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /**\n   * Semantic search - find most similar texts from a corpus\n   *\n   * @example\n   * ```ts\n   * const results = await g.search(\"capital of France\", [\n   *   \"Paris is beautiful\",\n   *   \"London is in England\",\n   *   \"Dogs are pets\"\n   * ]);\n   * // [{ text: \"Paris is beautiful\", score: 0.89, index: 0 }, ...]\n   * ```\n   */\n  async search(\n    query: string,\n    corpus: string[],\n    options: EmbedOptions & { topK?: number } = {},\n  ): Promise<import(\"./types.js\").SearchResult[]> {\n    const { topK = corpus.length, ...embedOptions } = options;\n\n    // Embed query and all documents\n    const queryEmbedding = await this.embed(query, embedOptions);\n    const corpusEmbeddings = await this.embedBatch(corpus, embedOptions);\n\n    // Calculate similarities\n    const results: import(\"./types.js\").SearchResult[] = corpusEmbeddings.map((doc, index) => ({\n      text: doc.text,\n      score: this.cosineSimilarity(queryEmbedding.vector, doc.vector),\n      index,\n    }));\n\n    // Sort by score descending and take top K\n    return results.sort((a, b) => b.score - a.score).slice(0, topK);\n  }\n\n  /**\n   * Find the nearest text to an embedding vector\n   *\n   * @example\n   * ```ts\n   * const embedding = (await g.embed(\"dog\")).vector;\n   * const match = await g.findNearest(embedding, [\"cat\", \"car\", \"tree\"]);\n   * // { text: \"cat\", score: 0.85, index: 0 }\n   * ```\n   */\n  async findNearest(\n    embedding: number[],\n    candidates: string[],\n    options: EmbedOptions & { topK?: number } = {},\n  ): Promise<import(\"./types.js\").SearchResult[]> {\n    const { topK = candidates.length, ...embedOptions } = options;\n\n    // Embed all candidates\n    const candidateEmbeddings = await this.embedBatch(candidates, embedOptions);\n\n    // Calculate similarities\n    const results: import(\"./types.js\").SearchResult[] = candidateEmbeddings.map((doc, index) => ({\n      text: doc.text,\n      score: this.cosineSimilarity(embedding, doc.vector),\n      index,\n    }));\n\n    // Sort by score descending and take top K\n    return results.sort((a, b) => b.score - a.score).slice(0, topK);\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: \"webgpu-native\",\n        gpu: null, // TODO: surface Dawn adapter info\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: \"~/.cache/gerbil\",\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 readonly ttsModelId = \"kani-tts-2\";\n\n  /**\n   * Load the native TTS model (Kani-TTS-2) for text-to-speech synthesis.\n   *\n   * @example\n   * ```ts\n   * await g.loadTTS({ onProgress: (p) => console.log(p.status) });\n   * const result = await g.speak(\"Hello world\");\n   * // result.audio = Float32Array PCM, result.sampleRate = 22050\n   * ```\n   */\n  // biome-ignore lint/suspicious/useAwait: native TTS is built lazily by ensureNativeTTSEngine\n  async loadTTS(_options: LoadTTSOptions & { model?: string } = {}): Promise<void> {\n    await this.ensureNativeTTSEngine();\n  }\n\n  /**\n   * Ensure TTS model is loaded (lazy loading)\n   */\n  async ensureTTSLoaded(_options?: LoadTTSOptions): Promise<void> {\n    await this.ensureNativeTTSEngine();\n  }\n\n  /**\n   * Generate speech from text using the native Kani-TTS-2 WebGPU engine.\n   *\n   * @example\n   * ```ts\n   * const result = await g.speak(\"Hello world\");\n   * // result.audio = Float32Array PCM, result.sampleRate = 22050\n   * ```\n   */\n  async speak(text: string, options: SpeakOptions = {}): Promise<SpeakResult> {\n    // ── Native WebGPU TTS ──\n    // Native speech runs through a dedicated WebGPUEngine loaded with the\n    // Kani-TTS-2 checkpoint; engine.speak() returns { pcm, sampleRate,\n    // audioSeconds }. The engine is lazily built on first speak() and owned by\n    // this instance (released in dispose()).\n    if (!this.preferNative()) {\n      throw new Error(\n        \"Speech synthesis requires WebGPU. CPU/WASM and the legacy ONNX backend have been removed.\",\n      );\n    }\n\n    const native = await this.ensureNativeTTSEngine();\n    const startTime = performance.now();\n    const out = await native.speak(text, {});\n    return {\n      audio: out.pcm,\n      sampleRate: out.sampleRate,\n      duration: out.audioSeconds,\n      voice: options.voice ?? \"default\",\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /** Lazily build (or reuse) the native Kani-TTS WebGPUEngine (default repo). */\n  private async ensureNativeTTSEngine(): Promise<any> {\n    if (!this.nativeTTSEngine) {\n      const { WebGPUEngine, DEFAULT_MODELS } = await import(\"../gpu/index.js\");\n      this.nativeTTSEngine = await WebGPUEngine.create({ repo: DEFAULT_MODELS.tts });\n    }\n    return this.nativeTTSEngine;\n  }\n\n  /**\n   * Stream speech generation. The native engine synthesizes the full clip, so a\n   * single final audio chunk is yielded.\n   */\n  async *speakStream(\n    text: string,\n    options: SpeakOptions = {},\n  ): AsyncGenerator<AudioChunk, SpeakResult, unknown> {\n    const result = await this.speak(text, options);\n    yield {\n      samples: result.audio,\n      sampleRate: result.sampleRate,\n      index: 0,\n      isFinal: true,\n    };\n    return result;\n  }\n\n  /**\n   * Get list of available TTS voices (native Kani-TTS-2 default voice).\n   */\n  listVoices(): VoiceInfo[] {\n    return KOKORO_VOICES_DEFAULT;\n  }\n\n  /**\n   * Check if TTS model is loaded\n   */\n  isTTSLoaded(): boolean {\n    return this.nativeTTSEngine !== null;\n  }\n\n  /**\n   * Get current TTS model info\n   */\n  getTTSModelInfo(): { id: string; loaded: boolean; device?: \"webgpu\" | \"cpu\" } | null {\n    if (!this.nativeTTSEngine) {\n      return null;\n    }\n    return {\n      id: this.ttsModelId,\n      loaded: true,\n      device: \"webgpu\",\n    };\n  }\n\n  /**\n   * List available TTS models (native Kani-TTS-2).\n   */\n  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  async listTTSModels(): Promise<\n    Array<{ id: string; description: string; sampleRate: number; voiceCount: number }>\n  > {\n    return [\n      {\n        id: this.ttsModelId,\n        description: \"Kani-TTS-2 native WebGPU TTS\",\n        sampleRate: 22_050,\n        voiceCount: 1,\n      },\n    ];\n  }\n\n  // ============================================\n  // Speech-to-Text (STT)\n  // ============================================\n\n  /**\n   * Load the native STT model (Moonshine) for speech-to-text transcription.\n   *\n   * @example\n   * ```ts\n   * await g.loadSTT();\n   * const result = await g.transcribe(audioData);\n   * console.log(result.text);\n   * ```\n   */\n  // biome-ignore lint/suspicious/useAwait: native STT is built lazily by ensureNativeSTT\n  async loadSTT(_modelId?: string, _options: LoadSTTOptions = {}): Promise<void> {\n    // Native STT uses the Moonshine checkpoint, built lazily on first transcribe().\n    await this.ensureNativeSTT();\n  }\n\n  /**\n   * Ensure STT model is loaded (lazy loading)\n   */\n  public async ensureSTTLoaded(_modelId?: string, _options?: LoadSTTOptions): Promise<void> {\n    await this.ensureNativeSTT();\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    // ── Native WebGPU STT ──\n    // MoonshineSTT consumes raw 16 kHz mono Float32 PCM. A Uint8Array WAV\n    // container cannot be decoded natively, and timestamps are not produced by\n    // the native path — both are unsupported now that the ONNX backend is gone.\n    if (!this.preferNative()) {\n      throw new Error(\n        \"Transcription requires WebGPU. CPU/WASM and the legacy ONNX backend have been removed.\",\n      );\n    }\n    if (!(audio instanceof Float32Array)) {\n      throw new Error(\n        \"Native transcription requires 16 kHz mono Float32Array PCM. Decode WAV bytes to PCM first.\",\n      );\n    }\n    if (options.timestamps) {\n      throw new Error(\"Native transcription does not produce timestamps.\");\n    }\n\n    const native = await this.ensureNativeSTT();\n    const startTime = performance.now();\n    const out = await native.transcribe(audio);\n    return {\n      text: out.text,\n      language: options.language ?? \"en\",\n      duration: out.audioSeconds,\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /** Lazily build (or reuse) the native MoonshineSTT engine (default repo). */\n  private async ensureNativeSTT(): Promise<any> {\n    if (!this.nativeSTT) {\n      const { MoonshineSTT } = await import(\"../gpu/moonshine-stt.js\");\n      this.nativeSTT = await MoonshineSTT.create();\n    }\n    return this.nativeSTT;\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  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  async createStreamingTranscription(\n    _options: StreamingTranscriptionOptions = {},\n  ): Promise<StreamingTranscriptionSession> {\n    throw new Error(\n      \"Streaming transcription is not supported by the native WebGPU STT engine. \" +\n        \"Use transcribe() on buffered 16 kHz Float32Array PCM instead.\",\n    );\n  }\n\n  /**\n   * Get list of available STT models (native Moonshine).\n   */\n  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  async listSTTModels(): Promise<STTModelConfig[]> {\n    return [\n      {\n        id: \"moonshine-base\",\n        repo: \"UsefulSensors/moonshine-base\",\n        description: \"Moonshine native WebGPU STT\",\n        size: \"61M\",\n        multilingual: false,\n        languages: [\"en\"],\n        sampleRate: 16000,\n      },\n    ];\n  }\n\n  /**\n   * Check if STT model is loaded\n   */\n  isSTTLoaded(): boolean {\n    return this.nativeSTT !== null;\n  }\n\n  /**\n   * Get current STT model info\n   */\n  getSTTModelInfo(): { id: string; loaded: boolean; device?: \"webgpu\" | \"cpu\" } | null {\n    if (!this.nativeSTT) {\n      return null;\n    }\n    return {\n      id: \"moonshine-base\",\n      loaded: true,\n      device: \"webgpu\",\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 (releases all native WebGPU engines and their devices).\n   * @param _disconnect Accepted for API compatibility; no longer used.\n   */\n  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  async dispose(_disconnect = false): Promise<void> {\n    // Clean up native WebGPU engine\n    if (this.webgpuEngine) {\n      try {\n        this.webgpuEngine.destroy();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.webgpuEngine = null;\n    }\n\n    // Clean up the lazily-created native capability engines (embed / STT / TTS).\n    // These are independent of this.webgpuEngine (they back embed()/transcribe()/\n    // speak() even when the active model isn't that capability), so they need\n    // their own teardown. MoonshineSTT has no GPU device of its own to destroy\n    // beyond its weight maps; WebGPUEngine.destroy() releases its device.\n    if (this.nativeEmbedEngine) {\n      try {\n        this.nativeEmbedEngine.destroy();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.nativeEmbedEngine = null;\n      this.nativeEmbedRepo = null;\n    }\n    if (this.nativeSTT) {\n      try {\n        this.nativeSTT.destroy?.();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.nativeSTT = null;\n    }\n    if (this.nativeTTSEngine) {\n      try {\n        this.nativeTTSEngine.destroy();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.nativeTTSEngine = null;\n    }\n\n    this.currentModel = null;\n    this.modelConfig = null;\n    this.isVisionModel = false;\n  }\n\n  /**\n   * @deprecated The shared Chrome backend was removed; this is now a no-op.\n   */\n  // biome-ignore lint/suspicious/useAwait: async kept for API compatibility\n  static async shutdown(): Promise<void> {\n    // no-op\n  }\n\n  // ============================================\n  // Private Methods\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;;;;;;;;;;AC5NvB,MAAa,gBAAgB;AAW7B,MAAaC,iBAA8C;CACzD,gBAAgB;EACd,IAAI;EACJ,MAAM;EACN,aACE;EACF,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,gBAAgB;EAChB,QAAQ;EACT;CACD,cAAc;EACZ,IAAI;EACJ,MAAM;EACN,aACE;EACF,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,gBAAgB;EAChB,QAAQ;EACT;CACD,wBAAwB;EACtB,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,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;;AAMpC,MAAMC,0BAAkD;CACtD,MAAM;CACN,OAAO;CACR;;;;;;;;;AAUD,SAAgB,0BACd,SACA,MACA,eACa;CACb,MAAM,YAAY,KAAK,aAAa;CAGpC,IAAIC,SAAgC;AACpC,KAAI,UAAU,SAAS,OAAO,CAC5B,UAAS;CAGX,MAAM,WAAW,UAAU,SAAS,MAAM,IAAI,UAAU,SAAS,SAAS;CAC1E,MAAM,SAAS,WAAW;AAE1B,QAAO;EACL,IAAI;EACJ;EACA,aAAa,mBAAmB;EAChC,MAAM;EACN,eAAe,iBAAiB,wBAAwB,WAAW;EAEnE,kBAAkB,UAAU;EAC5B,cAAc;EACd;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;;;;;;;;;;;;;AChKtC,SAAS,UAAU,KAAwE;CACzF,MAAM,OAAO,IAAI,SAAS,IAAI,QAAQ,IAAI,YAAY,IAAI,WAAW;CACrE,MAAM,IAAI,KAAK,UAAU,GAAG;CAC5B,MAAM,IAAI,KAAK,UAAU,GAAG;CAC5B,MAAM,WAAW,IAAI;CACrB,MAAM,YAAY,IAAI;AACtB,KAAI,aAAa,KAAM,cAAc,KAAK,cAAc,EACtD,OAAM,IAAI,MACR,6BAA6B,SAAS,aAAa,UAAU,wBAC9D;CAEH,MAAM,WAAW,cAAc,IAAI,IAAI;CACvC,IAAI,MAAM;CACV,MAAMC,OAAqB,EAAE;AAC7B,QAAO,MAAM,IAAI,QAAQ;EACvB,MAAM,MAAM,KAAK,UAAU,IAAI;EAC/B,MAAM,OAAO,OAAO,aAAa,IAAI,MAAM,IAAI,IAAI,MAAM,IAAI,IAAI,MAAM,IAAI,IAAI,MAAM,GAAG;AACxF,MAAI,SAAS,OACX,MAAK,KAAK,IAAI,SAAS,MAAM,GAAG,MAAM,IAAI,IAAI,CAAC;AAEjD,SAAO,KAAK;AACZ,MAAI,SAAS,OACX;;CAGJ,MAAM,MAAM,KAAK,YAAY,OAAO,OAAO,KAAK,CAAC;CACjD,MAAM,SAAS,IAAI;CACnB,MAAM,MAAM,IAAI,WAAW,IAAI,IAAI,EAAE;CACrC,MAAM,OAAO,IAAI,WAAW,OAAO;CACnC,MAAM,OAAO,IAAI,WAAW,OAAO;CACnC,IAAI,IAAI;AACR,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,GAAG;EAC7B,MAAM,SAAS,IAAI;AACnB,OAAK;AACL,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,KAAK,GAAG;GAClC,MAAM,IAAI,IAAI;AACd,QAAK;GACL,MAAM,IAAI,KAAK,WAAW,KAAK,IAAI,YAAY;GAC/C,MAAM,IAAI,KAAK;GACf,MAAM,IAAI,KAAK,WAAW,KAAK,IAAI,YAAY;GAC/C,IAAIC;AACJ,WAAQ,QAAR;IACE,KAAK;AACH,SAAI;AACJ;IACF,KAAK;AACH,SAAI,IAAI;AACR;IACF,KAAK;AACH,SAAI,IAAI;AACR;IACF,KAAK;AACH,SAAI,KAAM,IAAI,KAAM;AACpB;IACF,KAAK,GAAG;KACN,MAAM,KAAK,IAAI,IAAI;KACnB,MAAM,KAAK,KAAK,IAAI,KAAK,EAAE;KAC3B,MAAM,KAAK,KAAK,IAAI,KAAK,EAAE;KAC3B,MAAM,KAAK,KAAK,IAAI,KAAK,EAAE;KAC3B,IAAI,OAAO;AACX,SAAI,MAAM,MAAM,MAAM,GACpB,QAAO;cACE,MAAM,GACf,QAAO;AAET,SAAI,IAAI;AACR;;IAEF,QACE,OAAM,IAAI,MAAM,kBAAkB,SAAS;;AAE/C,QAAK,KAAK,IAAI;;AAEhB,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK,GAAG;AAC7B,QAAK,IAAI,IAAI,KAAK,IAAI,KAAK,KAAK,IAAI,WAAW;AAC/C,QAAK,IAAI,IAAI,KAAK,IAAI,KAAK,KAAK,IAAI,WAAW;AAC/C,QAAK,IAAI,IAAI,KAAK,IAAI,KAAK,KAAK,IAAI,WAAW;;AAEjD,OAAK,IAAI,KAAK;;AAEhB,QAAO;EAAE,QAAQ;EAAK,OAAO;EAAG,QAAQ;EAAG;;AAI7C,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;AASD,IAAa,SAAb,MAAoB;CAClB,AAAQ,eAA8B;CACtC,AAAQ,cAAkC;CAC1C,AAAiB;CACjB,AAAQ;CACR,AAAQ,cAAyC;CACjD,AAAQ,eAAoB;CAM5B,AAAQ,oBAAyB;CACjC,AAAQ,kBAAiC;CACzC,AAAQ,YAAiB;CACzB,AAAQ,kBAAuB;CAE/B,AAAQ,gBAAgB;CAGxB,AAAiB;CACjB,AAAiB;CAEjB,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;EAGD,MAAMC,cAAiC,OAAO,eAAe,EAAE;AAC/D,OAAK,QAAQ,IAAI,OAAO;GACtB,aAAa,YAAY,iBAAiB;GAC1C,SAAS,YAAY,WAAW;GACjC,CAAC;AAGF,OAAK,YAAY,OAAO,aAAa,EAAE;;CAOzC,AAAQ,YAAY,OAAc,SAA6B;AAC7D,MAAI;AACF,QAAK,UAAU,UAAU,OAAO,QAAQ;UAClC;;;;;;;CAUV,AAAQ,eAAwB;AAC9B,SAAO;;CAOT,OAAO,aAA4B;AACjC,SAAO,OAAO,OAAO,eAAe;;CAGtC,OAAO,SAAS,SAA0C;AACxD,SAAO,eAAe;;;;;;;;;;;;;;;;;;;;CAyBxB,MAAM,UAAU,UAAU,eAAe,UAAuB,EAAE,EAAiB;EACjF,MAAM,gBAAgB,YAAY,KAAK;AAIvC,MAAI,KAAK,UAAU,CACjB,OAAM,KAAK,SAAS;EAGtB,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;AAKpE,MAAK,WAAsB,SAAU,WAAsB,MACzD,OAAM,IAAI,MACR,qHACD;AAGH,eAAa,EAAE,QAAQ,WAAW,QAAQ,MAAM,CAAC;AAEjD,MAAI;AACF,gBAAa,EAAE,QAAQ,iCAAiC,CAAC;GACzD,MAAM,EAAE,iBAAiB,MAAM,OAAO;GAItC,IAAI,SAAS,OAAO;AAEpB,OAAI,OAAO,SAAS,kBAAkB,IAAI,OAAO,SAAS,QAAQ,CAOhE,UAN8C;IAC5C,oCAAoC;IACpC,kCAAkC;IAClC,kCAAkC;IAClC,kCAAkC;IACnC,CACsB,WAAW;GAIpC,MAAM,WAAW,cAAc,OAAQ,OAAiB;AAExD,QAAK,eAAe,MAAM,aAAa,OAAO;IAC5C,MAAM;IACN,WAAW,QAAQ,iBAAiB,OAAO,iBAAiB;IAC5D,OAAO;IACP,aAAa,QAAQ,OAAO,YAAY;AACtC,kBAAa;MACX,QAAQ;MACR,UAAU,QAAQ,IAAI,KAAK,MAAO,SAAS,QAAS,IAAI,GAAG;MAC5D,CAAC;;IAEL,CAAC;AAEF,QAAK,cAAc;AACnB,QAAK,gBAAgB;AACrB,QAAK,eAAe;AACpB,QAAK,cAAc;AACnB,gBAAa,EAAE,QAAQ,0BAA0B,CAAC;AAGlD,OAAI,KAAK,UAAU,YACjB,KAAI;AACF,SAAK,UAAU,YAAY;KACzB;KACA,YAAY,YAAY,KAAK,GAAG;KAChC,WAAW;KACX,QAAQ,KAAK;KACb,SAAS;KACV,CAAC;WACI;WAIH,KAAK;AAEZ,QAAK,YAAY,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,IAAI,CAAC,EAAE;IACpE,WAAW;IACX;IACD,CAAC;AACF,OAAI,KAAK,UAAU,YACjB,KAAI;AACF,SAAK,UAAU,YAAY;KACzB;KACA,YAAY,YAAY,KAAK,GAAG;KAChC,WAAW;KACX,QAAQ,KAAK;KACb,SAAS;KACT,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,IAAI;KACxD,CAAC;WACI;AAIV,SAAM;;;;;;;;;CAUV,MAAc,gBACZ,SACA,UACA,QACA,UAAuB,EAAE,EACV;EACf,MAAM,EAAE,YAAY,SAAS,WAAW;AAExC,eAAa,EAAE,QAAQ,WAAW,QAAQ,qBAAqB,CAAC;AAGhE,MAAK,WAAsB,SAAU,WAAsB,MACzD,OAAM,IAAI,MACR,kIACD;AAGH,eAAa,EAAE,QAAQ,wCAAwC,CAAC;EAChE,MAAM,EAAE,iBAAiB,MAAM,OAAO;EAGtC,IAAI,UAAU;AACd,MAAI,QAAQ,SAAS,kBAAkB,IAAI,QAAQ,SAAS,QAAQ,CAIlE,WAHoC,EAClC,oCAAoC,qBACrC,CACa,YAAY;AAE5B,OAAK,eAAe,MAAM,aAAa,OAAO;GAC5C,MAAM;GACN,cAAc;GACd,WAAW,QAAQ,iBAAiB,OAAO,iBAAiB;GAC5D,aAAa,QAAQ,OAAO,YAC1B,aAAa;IACX,QAAQ;IACR,UAAU,QAAQ,IAAI,KAAK,MAAO,SAAS,QAAS,IAAI,GAAG;IAC5D,CAAC;GACL,CAAC;AACF,OAAK,cAAc;AACnB,OAAK,gBAAgB;AACrB,OAAK,eAAe;AACpB,OAAK,cAAc;AACnB,eAAa,EAAE,QAAQ,kCAAkC,CAAC;;;;;CAM5D,WAAoB;AAClB,SAAO,KAAK,iBAAiB;;;;;CAM/B,iBAA0B;AACxB,SAAO,KAAK,iBAAiB,KAAK,aAAa,mBAAmB;;;;;CAMpE,eAAmC;AACjC,SAAO,KAAK;;;;;CAMd,gBAA2C;AACzC,SAAO,KAAK;;;;;;;CAQd,WAAmB;AACjB,SAAO;;;;;CAMT,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;;;;;;;;;;;;;;CAmB1B,MAAM,cAAc,SAAmC;EACrD,MAAM,SAAS,aAAa,QAAQ;AACpC,SAAO,KAAK,mBAAmB,OAAO,KAAK;;;;;;CAO7C,AAAQ,mBAAmB,MAAc,WAAW,QAAiB;AACnE,MAAI;GACF,MAAM,OAAO,QAAQ,IAAI,QAAQ,QAAQ,IAAI,eAAe,GAAG,SAAS;AACxE,OAAI,CAAC,KACH,QAAO;GAET,MAAM,WAAW,KAAK,KAAK,MAAM,UAAU,UAAU,KAAK,QAAQ,OAAO,IAAI,EAAE,SAAS;AAExF,UAAO,WAAW,KAAK,KAAK,UAAU,cAAc,QAAQ,OAAO,IAAI,CAAC,CAAC;UACnE;AACN,UAAO;;;;;;;;;;;;;;;;;;;;;CAsBX,MAAM,aAAa,SAAiB,UAA0B,EAAE,EAAiB;AAChE,eAAa,QAAQ;EACpC,MAAM,EAAE,YAAY,aAAa,UAAU;AAG3C,MAAI,cAAc,KAAK,UAAU,IAAI,KAAK,iBAAiB,SAAS;AAClE,gBAAa,EAAE,QAAQ,wBAAwB,CAAC;AAChD;;AAIF,MAAI,CAAC,cAAe,MAAM,KAAK,cAAc,QAAQ,EAAG;AACtD,gBAAa,EAAE,QAAQ,wBAAwB,CAAC;AAChD;;AAIF,MAAI,YAAY;AACd,SAAM,KAAK,UAAU,SAAS,EAAE,YAAY,CAAC;AAC7C;;AAGF,eAAa,EAAE,QAAQ,cAAc,QAAQ,MAAM,CAAC;AAKpD,QAAM,KAAK,UAAU,SAAS,EAAE,YAAY,CAAC;AAC7C,QAAM,KAAK,SAAS;AAEpB,eAAa,EAAE,QAAQ,oBAAoB,CAAC;;;;;;CAO9C,MAAM,YAAY,UAAqC;EACrD,MAAM,EAAE,mBAAmB,MAAM,OAAO;AACxC,SAAO,KAAK,mBAAmB,eAAe,IAAI;;;;;;;;CASpD,MAAM,WAAW,SAAkB,UAA0B,EAAE,EAAiB;EAC9E,MAAM,EAAE,YAAY,aAAa,UAAU;AAE3C,MAAI,cAAc,KAAK,aAAa,EAAE;AACpC,gBAAa,EAAE,QAAQ,4BAA4B,CAAC;AACpD;;AAGF,MAAI,CAAC,cAAe,MAAM,KAAK,YAAY,QAAQ,EAAG;AACpD,gBAAa,EAAE,QAAQ,4BAA4B,CAAC;AACpD;;AAGF,eAAa,EAAE,QAAQ,2BAA2B,CAAC;AAEnD,QAAM,KAAK,uBAAuB;AAElC,MAAI,CAAC,cAAc,KAAK,iBAAiB;AACvC,OAAI;AACF,SAAK,gBAAgB,SAAS;WACxB;AAGR,QAAK,kBAAkB;;AAEzB,eAAa,EAAE,QAAQ,oBAAoB,CAAC;;;;;;CAO9C,MAAM,YAAY,UAAqC;EACrD,MAAM,EAAE,mBAAmB,MAAM,OAAO;AACxC,SAAO,KAAK,mBAAmB,eAAe,IAAI;;;;;;;;CASpD,MAAM,WAAW,SAAkB,UAA0B,EAAE,EAAiB;EAC9E,MAAM,EAAE,YAAY,aAAa,UAAU;AAE3C,MAAI,cAAc,KAAK,aAAa,EAAE;AACpC,gBAAa,EAAE,QAAQ,4BAA4B,CAAC;AACpD;;AAGF,MAAI,CAAC,cAAe,MAAM,KAAK,YAAY,QAAQ,EAAG;AACpD,gBAAa,EAAE,QAAQ,4BAA4B,CAAC;AACpD;;AAGF,eAAa,EAAE,QAAQ,2BAA2B,CAAC;AAEnD,QAAM,KAAK,iBAAiB;AAE5B,MAAI,CAAC,cAAc,KAAK,WAAW;AACjC,OAAI;AACF,SAAK,UAAU,WAAW;WACpB;AAGR,QAAK,YAAY;;AAEnB,eAAa,EAAE,QAAQ,oBAAoB,CAAC;;;;;;CAO9C,MAAM,kBAAkB,SAAoC;EAC1D,MAAM,EAAE,mBAAmB,MAAM,OAAO;AACxC,SAAO,KAAK,mBAAmB,WAAW,eAAe,UAAU;;;;;;;;CASrE,MAAM,iBAAiB,SAAkB,UAA0B,EAAE,EAAiB;EACpF,MAAM,EAAE,YAAY,aAAa,UAAU;AAE3C,MAAI,cAAc,KAAK,mBAAmB;AACxC,gBAAa,EAAE,QAAQ,kCAAkC,CAAC;AAC1D;;AAGF,MAAI,CAAC,cAAe,MAAM,KAAK,kBAAkB,QAAQ,EAAG;AAC1D,gBAAa,EAAE,QAAQ,kCAAkC,CAAC;AAC1D;;AAGF,eAAa,EAAE,QAAQ,iCAAiC,CAAC;AAEzD,QAAM,KAAK,wBAAwB,QAAQ;AAE3C,MAAI,CAAC,cAAc,KAAK,mBAAmB;AACzC,OAAI;AACF,SAAK,kBAAkB,SAAS;WAC1B;AAGR,QAAK,oBAAoB;AACzB,QAAK,kBAAkB;;AAEzB,eAAa,EAAE,QAAQ,oBAAoB,CAAC;;;;;;;CAS9C,MAAM,aAA4B;;;;;;;;;;;;;;;CAsBlC,MAAM,SAAS,QAAgB,UAA2B,EAAE,EAA2B;EACrF,MAAM,iBAAiB,YAAY,KAAK;AAExC,MAAI;AAiCF,UA/Be,MAAM,KAAK,MAAM,IAAI,YAAY;IAC9C,MAAM,gBAAgB,YAAY,KAAK,GAAG;AAG1C,QAAI,gBAAgB,OAAO,KAAK,UAAU,YACxC,KAAI;AACF,UAAK,UAAU,YAAY,cAAc;YACnC;IAKV,MAAM,kBAAkB,MAAM,KAAK,iBAAiB,QAAQ,QAAQ;AAGpE,QAAI,KAAK,UAAU,WACjB,KAAI;AACF,UAAK,UAAU,WAAW;MACxB,SAAS,KAAK,gBAAgB;MAC9B,QAAQ;MACR,QAAQ,gBAAgB,UAAU;MAClC,aAAa,gBAAgB,MAAM,gBAAgB;MACpD,CAAC;YACI;AAKV,WAAO;KACP;WAGK,OAAO;AAEd,OAAI,KAAK,UAAU,QACjB,KAAI;AACF,SAAK,UAAU,QAAQ,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,MAAM,CAAC,EAAE;KAChF,QAAQ;KACR,SAAS,KAAK,gBAAgB;KAC9B,QAAQ,OAAO,MAAM,GAAG,IAAI;KAC5B,eAAe,YAAY,KAAK,GAAG;KACpC,CAAC;WACI;AAIV,SAAM;;;;;;CAOV,MAAc,iBACZ,QACA,UAA2B,EAAE,EACJ;AACzB,MAAI,CAAC,KAAK,UAAU,CAElB,OAAM,KAAK,UAAU,KAAK,OAAO,SAAS,cAAc;EAG1D,MAAM,EAAE,WAAW;AAInB,MAAI,QAAQ,UAAU,KAAK,cACzB,QAAO,KAAK,mBAAmB,QAAQ,QAAQ;EAGjD,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;GAEd,IAAI,wBAAwB;GAC5B,IAAI,wBAAwB;AAE5B,OAAI,KAAK,cAAc;IAErB,MAAMC,WAAS,MAAM,KAAK,aAAa,SAAS,QAAQ;KACtD;KACA,UAAU;MACR;MACA;MACA;MACD;KACD,cAAc;KACd,SAAS,QAAQ,WAAW,MAAc,QAAQ,UAAU,EAAE,GAAG;KAClE,CAAC;AACF,cAAUA,SAAO;AACjB,4BAAwBA,SAAO;AAC/B,4BAAwBA,SAAO;SAE/B,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;GAGhD,MAAM,kBAAkB;AAGxB,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,iBAAiB;IACjB;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,OAAO;AAEd,QAAK,YAAY,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,MAAM,CAAC,EAAE;IAC1E,WAAW;IACX,SAAS,KAAK,gBAAgB;IAC/B,CAAC;AAEF,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,cAAc;EAG1D,MAAM,YAAY,YAAY,KAAK;AAGnC,MAAI,KAAK,cAAc;GACrB,IAAI,WAAW;GACf,MAAMC,aAAuB,EAAE;GAC/B,IAAIC,cAAuD;GAC3D,IAAI,OAAO;GAEX,IAAI,wBAAwB;GAC5B,IAAI,wBAAwB;GAE5B,MAAM,kBAAkB,KAAK,aAC1B,SAAS,QAAQ;IAChB,GAAG;IACH,UAAU;KACR,aAAa,QAAQ;KACrB,MAAM,QAAQ;KACd,MAAM,QAAQ;KACf;IACD,cAAc,QAAQ;IACtB,UAAU,UAAkB;AAC1B,iBAAY;AACZ,SAAI,aAAa;AACf,kBAAY,MAAM;AAClB,oBAAc;WAEd,YAAW,KAAK,MAAM;;IAG3B,CAAC,CACD,MAAM,WAAiE;AACtE,4BAAwB,OAAO;AAC/B,4BAAwB,OAAO;AAC/B,WAAO;AACP,QAAI,YACF,aAAY,KAAK;KAEnB,CACD,OAAO,QAAe;AACrB,WAAO;AACP,QAAI,YACF,aAAY,KAAK;AAEnB,UAAM;KACN;AAEJ,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,YAAY,YAAY,KAAK,GAAG;AAEtC,UAAO;IACL,MAAM;IACN,UAAU,QAAQ,WAAW,eAAe;IAC5C,iBAAiB;IACjB;IACA,iBAAiB;IACjB,cAAc;IACf;;AAKH,QAAM,IAAI,MAAM,kBAAkB;;;;;;CAWpC,MAAc,mBACZ,QACA,SACyB;AAMzB,MAAI,EAAE,KAAK,gBAAgB,OAAO,KAAK,aAAa,kBAAkB,YACpE,OAAM,IAAI,MACR,mFACD;EAGH,MAAM,OAAO,QAAQ,UAAU,EAAE;AACjC,MAAI,KAAK,WAAW,EAClB,OAAM,IAAI,MACR,oEAAoE,KAAK,OAAO,IACjF;EAGH,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,EAAE,QAAQ,OAAO,WAAW,MAAM,KAAK,oBAAoB,KAAK,GAAG,OAAO;EAChF,MAAM,SAAS,MAAM,KAAK,aAAa,cAAc;GAAE;GAAQ;GAAO;GAAQ,EAAE,QAAQ;GACtF,WAAW,QAAQ,aAAa;GAChC,UAAU;IACR,aAAa,QAAQ,eAAe;IACpC,MAAM,QAAQ,QAAQ;IACtB,MAAM,QAAQ,QAAQ;IACvB;GACD,SAAS,QAAQ,WAAW,MAAc,QAAQ,UAAU,EAAE,GAAG;GAClE,CAAC;EACF,MAAM,YAAY,YAAY,KAAK,GAAG;AAGtC,OAAK,MAAM,WAAW;AACtB,OAAK,MAAM,aAAa,OAAO;AAC/B,OAAK,MAAM,aAAa;AACxB,OAAK,MAAM,WAAY,KAAK,MAAM,YAAY,KAAK,MAAM,YAAa;AAEtE,SAAO;GACL,MAAM,KAAK,YAAY,OAAO,KAAK;GACnC,iBAAiB,OAAO;GACxB,iBAAiB,OAAO;GACxB;GACA,cAAc;GACd,UAAU;GACV,QAAQ;GACT;;;;;;;;;CAUH,MAAc,oBACZ,QACgE;EAChE,MAAM,QAAQ,MAAM,KAAK,gBAAgB,OAAO;AAQhD,MAAI,EALF,MAAM,SAAS,KACf,MAAM,OAAO,OACb,MAAM,OAAO,MACb,MAAM,OAAO,MACb,MAAM,OAAO,IAEb,OAAM,IAAI,MACR,mJAED;AAEH,SAAO,UAAU,MAAM;;;CAIzB,MAAc,gBAAgB,QAAqC;AACjE,MAAI,OAAO,WAAW,QAAQ,EAAE;GAC9B,MAAM,QAAQ,OAAO,QAAQ,IAAI;GACjC,MAAM,OAAO,OAAO,MAAM,GAAG,MAAM;GACnC,MAAM,OAAO,OAAO,MAAM,QAAQ,EAAE;AACpC,OAAI,KAAK,SAAS,SAAS,CACzB,QAAO,WAAW,KAAK,OAAO,KAAK,MAAM,SAAS,CAAC;AAErD,UAAO,WAAW,KAAK,OAAO,KAAK,mBAAmB,KAAK,EAAE,SAAS,CAAC;;AAEzE,MAAI,OAAO,WAAW,UAAU,IAAI,OAAO,WAAW,WAAW,EAAE;GACjE,MAAM,MAAM,MAAM,MAAM,OAAO;AAC/B,OAAI,CAAC,IAAI,GACP,OAAM,IAAI,MAAM,0BAA0B,IAAI,OAAO,KAAK,SAAS;AAErE,UAAO,IAAI,WAAW,MAAM,IAAI,aAAa,CAAC;;EAGhD,MAAM,EAAE,aAAa,MAAM,OAAO;AAClC,SAAO,IAAI,WAAW,MAAM,SAAS,OAAO,CAAC;;;;;CAU/C,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;;;;;;;;;;;;;;;;;;CAmBlD,MAAM,eACJ,QACA,UAAiC,EAAE,EACD;AAClC,MAAI,CAAC,KAAK,UAAU,CAClB,OAAM,KAAK,UAAU,KAAK,OAAO,SAAS,cAAc;AAE1D,MAAI,CAAC,KAAK,aACR,OAAM,IAAI,MAAM,kBAAkB;AAEpC,SAAO,KAAK,aAAa,eAAe,QAAQ,QAAQ;;;;;CAU1D,MAAM,MAAM,MAAc,UAAwB,EAAE,EAAwB;AAQ1E,MAAI,CAAC,KAAK,cAAc,CACtB,OAAM,IAAI,MACR,qFACD;EAGH,MAAM,SAAS,MAAM,KAAK,wBAAwB,QAAQ,MAAM;EAChE,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,MAAM,MAAM,OAAO,MAAM,KAAK;AACpC,SAAO;GACL,QAAQ,MAAM,KAAK,IAAoB;GACvC;GACA,WAAW,YAAY,KAAK,GAAG;GAChC;;;;;;;CAQH,MAAc,wBAAwB,MAA6B;AACjE,MAAI,KAAK,sBAAsB,CAAC,QAAQ,SAAS,KAAK,iBACpD,QAAO,KAAK;AAEd,MAAI,KAAK,mBAAmB;AAE1B,OAAI;AACF,SAAK,kBAAkB,SAAS;WAC1B;AAGR,QAAK,oBAAoB;AACzB,QAAK,kBAAkB;;EAEzB,MAAM,EAAE,iBAAiB,MAAM,OAAO;AACtC,OAAK,oBAAoB,MAAM,aAAa,OAAO;GAAE;GAAM,WAAW;GAAM,CAAC;AAC7E,OAAK,kBAAkB,QAAQ;AAC/B,SAAO,KAAK;;;;;CAMd,MAAM,WAAW,OAAiB,UAAwB,EAAE,EAA0B;EACpF,MAAMC,UAAyB,EAAE;AACjC,OAAK,MAAM,QAAQ,MACjB,SAAQ,KAAK,MAAM,KAAK,MAAM,MAAM,QAAQ,CAAC;AAE/C,SAAO;;;;;;;;;;;CAYT,iBAAiB,GAAa,GAAqB;AACjD,MAAI,EAAE,WAAW,EAAE,OACjB,OAAM,IAAI,MAAM,iCAAiC,EAAE,OAAO,MAAM,EAAE,SAAS;EAG7E,IAAI,aAAa;EACjB,IAAI,QAAQ;EACZ,IAAI,QAAQ;AAEZ,OAAK,IAAI,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,iBAAc,EAAE,KAAK,EAAE;AACvB,YAAS,EAAE,KAAK,EAAE;AAClB,YAAS,EAAE,KAAK,EAAE;;EAGpB,MAAM,YAAY,KAAK,KAAK,MAAM,GAAG,KAAK,KAAK,MAAM;AACrD,MAAI,cAAc,EAAG,QAAO;AAE5B,SAAO,aAAa;;;;;;;;;;;CAYtB,MAAM,WACJ,OACA,OACA,UAAwB,EAAE,EACsB;EAChD,MAAM,YAAY,YAAY,KAAK;EAEnC,MAAM,CAAC,QAAQ,UAAU,MAAM,QAAQ,IAAI,CACzC,KAAK,MAAM,OAAO,QAAQ,EAC1B,KAAK,MAAM,OAAO,QAAQ,CAC3B,CAAC;AAIF,SAAO;GACL,OAHY,KAAK,iBAAiB,OAAO,QAAQ,OAAO,OAAO;GAI/D;GACA;GACA,WAAW,YAAY,KAAK,GAAG;GAChC;;;;;;;;;;;;;;;CAgBH,MAAM,OACJ,OACA,QACA,UAA4C,EAAE,EACA;EAC9C,MAAM,EAAE,OAAO,OAAO,QAAQ,GAAG,iBAAiB;EAGlD,MAAM,iBAAiB,MAAM,KAAK,MAAM,OAAO,aAAa;AAW5D,UAVyB,MAAM,KAAK,WAAW,QAAQ,aAAa,EAGE,KAAK,KAAK,WAAW;GACzF,MAAM,IAAI;GACV,OAAO,KAAK,iBAAiB,eAAe,QAAQ,IAAI,OAAO;GAC/D;GACD,EAAE,CAGY,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;;;;;;;;;;;;CAajE,MAAM,YACJ,WACA,YACA,UAA4C,EAAE,EACA;EAC9C,MAAM,EAAE,OAAO,WAAW,QAAQ,GAAG,iBAAiB;AAatD,UAV4B,MAAM,KAAK,WAAW,YAAY,aAAa,EAGF,KAAK,KAAK,WAAW;GAC5F,MAAM,IAAI;GACV,OAAO,KAAK,iBAAiB,WAAW,IAAI,OAAO;GACnD;GACD,EAAE,CAGY,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;;;;;CAUjE,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,AAAiB,aAAa;;;;;;;;;;;CAa9B,MAAM,QAAQ,WAAgD,EAAE,EAAiB;AAC/E,QAAM,KAAK,uBAAuB;;;;;CAMpC,MAAM,gBAAgB,UAA0C;AAC9D,QAAM,KAAK,uBAAuB;;;;;;;;;;;CAYpC,MAAM,MAAM,MAAc,UAAwB,EAAE,EAAwB;AAM1E,MAAI,CAAC,KAAK,cAAc,CACtB,OAAM,IAAI,MACR,4FACD;EAGH,MAAM,SAAS,MAAM,KAAK,uBAAuB;EACjD,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,MAAM,MAAM,OAAO,MAAM,MAAM,EAAE,CAAC;AACxC,SAAO;GACL,OAAO,IAAI;GACX,YAAY,IAAI;GAChB,UAAU,IAAI;GACd,OAAO,QAAQ,SAAS;GACxB,WAAW,YAAY,KAAK,GAAG;GAChC;;;CAIH,MAAc,wBAAsC;AAClD,MAAI,CAAC,KAAK,iBAAiB;GACzB,MAAM,EAAE,cAAc,mBAAmB,MAAM,OAAO;AACtD,QAAK,kBAAkB,MAAM,aAAa,OAAO,EAAE,MAAM,eAAe,KAAK,CAAC;;AAEhF,SAAO,KAAK;;;;;;CAOd,OAAO,YACL,MACA,UAAwB,EAAE,EACwB;EAClD,MAAM,SAAS,MAAM,KAAK,MAAM,MAAM,QAAQ;AAC9C,QAAM;GACJ,SAAS,OAAO;GAChB,YAAY,OAAO;GACnB,OAAO;GACP,SAAS;GACV;AACD,SAAO;;;;;CAMT,aAA0B;AACxB,SAAO;;;;;CAMT,cAAuB;AACrB,SAAO,KAAK,oBAAoB;;;;;CAMlC,kBAAqF;AACnF,MAAI,CAAC,KAAK,gBACR,QAAO;AAET,SAAO;GACL,IAAI,KAAK;GACT,QAAQ;GACR,QAAQ;GACT;;;;;CAOH,MAAM,gBAEJ;AACA,SAAO,CACL;GACE,IAAI,KAAK;GACT,aAAa;GACb,YAAY;GACZ,YAAY;GACb,CACF;;;;;;;;;;;;CAkBH,MAAM,QAAQ,UAAmB,WAA2B,EAAE,EAAiB;AAE7E,QAAM,KAAK,iBAAiB;;;;;CAM9B,MAAa,gBAAgB,UAAmB,UAA0C;AACxF,QAAM,KAAK,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;CA0B9B,MAAM,WACJ,OACA,UAA6B,EAAE,EACJ;AAK3B,MAAI,CAAC,KAAK,cAAc,CACtB,OAAM,IAAI,MACR,yFACD;AAEH,MAAI,EAAE,iBAAiB,cACrB,OAAM,IAAI,MACR,6FACD;AAEH,MAAI,QAAQ,WACV,OAAM,IAAI,MAAM,oDAAoD;EAGtE,MAAM,SAAS,MAAM,KAAK,iBAAiB;EAC3C,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,MAAM,MAAM,OAAO,WAAW,MAAM;AAC1C,SAAO;GACL,MAAM,IAAI;GACV,UAAU,QAAQ,YAAY;GAC9B,UAAU,IAAI;GACd,WAAW,YAAY,KAAK,GAAG;GAChC;;;CAIH,MAAc,kBAAgC;AAC5C,MAAI,CAAC,KAAK,WAAW;GACnB,MAAM,EAAE,iBAAiB,MAAM,OAAO;AACtC,QAAK,YAAY,MAAM,aAAa,QAAQ;;AAE9C,SAAO,KAAK;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA+Bd,MAAM,6BACJ,WAA0C,EAAE,EACJ;AACxC,QAAM,IAAI,MACR,0IAED;;;;;CAOH,MAAM,gBAA2C;AAC/C,SAAO,CACL;GACE,IAAI;GACJ,MAAM;GACN,aAAa;GACb,MAAM;GACN,cAAc;GACd,WAAW,CAAC,KAAK;GACjB,YAAY;GACb,CACF;;;;;CAMH,cAAuB;AACrB,SAAO,KAAK,cAAc;;;;;CAM5B,kBAAqF;AACnF,MAAI,CAAC,KAAK,UACR,QAAO;AAET,SAAO;GACL,IAAI;GACJ,QAAQ;GACR,QAAQ;GACT;;;;;;;;;;;;;;;;;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;;;;;;;CAaX,MAAM,QAAQ,cAAc,OAAsB;AAEhD,MAAI,KAAK,cAAc;AACrB,OAAI;AACF,SAAK,aAAa,SAAS;WACrB;AAGR,QAAK,eAAe;;AAQtB,MAAI,KAAK,mBAAmB;AAC1B,OAAI;AACF,SAAK,kBAAkB,SAAS;WAC1B;AAGR,QAAK,oBAAoB;AACzB,QAAK,kBAAkB;;AAEzB,MAAI,KAAK,WAAW;AAClB,OAAI;AACF,SAAK,UAAU,WAAW;WACpB;AAGR,QAAK,YAAY;;AAEnB,MAAI,KAAK,iBAAiB;AACxB,OAAI;AACF,SAAK,gBAAgB,SAAS;WACxB;AAGR,QAAK,kBAAkB;;AAGzB,OAAK,eAAe;AACpB,OAAK,cAAc;AACnB,OAAK,gBAAgB;;;;;CAOvB,aAAa,WAA0B;CAQvC,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-BetB5xb0.d.mts → gerbil-CvNqpVg0.d.mts}

@@ -1,5 +1,5 @@
-import { A as SpeakResult, C as PreloadOptions, D as SessionStats, E as SearchResult, I as TranscribeOptions, L as TranscribeResult, M as StreamingTranscriptionSession, N as SystemInfo, O as SimilarityResult, T as STTModelConfig, _ as LoadSTTOptions, a as EmbedResult, d as GerbilConfig, g as LoadOptions, h as JsonOptions, i as EmbedOptions, j as StreamingTranscriptionOptions, k as SpeakOptions, l as GenerateOptions, t as AudioChunk, u as GenerateResult, v as LoadTTSOptions, y as ModelConfig, z as VoiceInfo } from "./types-LlyYILII.mjs";
-import { a as GenerateObjectOptions, o as GenerateObjectResult } from "./index-DukkJRMj.mjs";
+import { A as SpeakResult, C as PreloadOptions, D as SessionStats, E as SearchResult, I as TranscribeOptions, L as TranscribeResult, M as StreamingTranscriptionSession, N as SystemInfo, O as SimilarityResult, T as STTModelConfig, _ as LoadSTTOptions, a as EmbedResult, d as GerbilConfig, g as LoadOptions, h as JsonOptions, i as EmbedOptions, j as StreamingTranscriptionOptions, k as SpeakOptions, l as GenerateOptions, t as AudioChunk, u as GenerateResult, v as LoadTTSOptions, y as ModelConfig, z as VoiceInfo } from "./types-Bxwe_uS7.mjs";
+import { a as GenerateObjectOptions, o as GenerateObjectResult } from "./index-Nl40RdSs.mjs";
 
 //#region src/core/gerbil.d.ts
 
@@ -485,4 +485,4 @@ declare class Gerbil {
 }
 //#endregion
 export { Gerbil as t };
-//# sourceMappingURL=gerbil-BetB5xb0.d.mts.map
+//# sourceMappingURL=gerbil-CvNqpVg0.d.mts.map

package/dist/{gerbil-BetB5xb0.d.mts.map → gerbil-CvNqpVg0.d.mts.map}

@@ -1 +1 @@
-{"version":3,"file":"gerbil-BetB5xb0.d.mts","names":[],"sources":["../src/core/gerbil.ts"],"sourcesContent":[],"mappings":";;;;;AA8OsB,cAvBT,MAAA,CAuBS;EAgDC,QAAA,YAAA;EAIa,QAAA,WAAA;EA0BgB,iBAAA,MAAA;EAAmB,QAAA,KAAA;EAoLrD,QAAA,WAAA;EAyDsB,QAAA,YAAA;EAyCO,QAAA,iBAAA;EAAsB,QAAA,eAAA;EAqC7B,QAAA,SAAA;EAWM,QAAA,eAAA;EAAsB,QAAA,aAAA;EAgC5B,iBAAA,KAAA;EAWM,iBAAA,SAAA;EAAsB,WAAA,CAAA,MAAA,CAAA,EA/b9C,YA+b8C;EAgCvB,QAAA,WAAA;EAWO;;;;;EAyDa,QAAA,YAAA;EA0MpD,OAAA,UAAA,CAAA,CAAA,EA7rBU,WA6rBV,EAAA;EACe,OAAA,QAAA,CAAA,OAAA,EAAA,MAAA,CAAA,EA1rBQ,WA0rBR,GAAA,SAAA;EAAvB;;;;;;;;;;;;;;;;;;EAwZQ,SAAA,CAAA,OAAA,CAAA,EAAA,MAAA,EAAA,OAAA,CAAA,EAxjCuC,WAwjCvC,CAAA,EAxjC0D,OAwjC1D,CAAA,IAAA,CAAA;EAAY;;;;;;EAgEZ,QAAA,eAAA;EAuDa;;;EAO0B,QAAA,CAAA,CAAA,EAAA,OAAA;EAaf;;;EAuCxB,cAAA,CAAA,CAAA,EAAA,OAAA;EACO;;;EAcJ,YAAA,CAAA,CAAA,EArkCE,WAqkCF,GAAA,IAAA;EA8BZ;;;EA2B+D,aAAA,CAAA,CAAA,EAAA,QAAA,GAAA,KAAA,GAAA,MAAA;EAQN;;;;;EA8BhD,QAAA,CAAA,CAAA,EAAA,MAAA;EAAR;;;EAqEA,qBAAA,CAAA,CAAA,EAAA;IAW4B,IAAA,EAAA,MAAA;IAAR,MAAA,EAAA,MAAA;IAyDZ,IAAA,EAAA,MAAA;IAAR,OAAA,EAAA,MAAA;EAmC4B,CAAA;EAkBK;;;;;;;;;;;;;;;;kCAzyCE;;;;;;;;;;;;;;;;;;;;;;;;0CAyCO,iBAAsB;;;;;kCAqC7B;;;;;;;yCAWM,iBAAsB;;;;;kCAgC5B;;;;;;;yCAWM,iBAAsB;;;;;uCAgCvB;;;;;;;+CAWO,iBAAsB;;;;;;gBAmCpD;;;;;;;;;;;;;;;qCAsBoB,kBAAuB,QAAQ;;;;;;;;;;;mCA0M5D,kBACR,uBAAuB;;;;;;;;;;;;;;;;;;;mCA6Ma,YAAY,KAAK,QAAQ;;;;;;;;;;;;;;;;;wDAgDrD,wBACR,QAAQ,qBAAqB;;;;gCAiBG,eAAoB,QAAQ;;;;;;;;;;wCAoDpB,eAAoB,QAAQ;;;;;;;;;;;;;;;;;;;;qDAkD5D,eACR,QADyB,gBAAA;;;;;;;;;;;;;;oDAmCjB;;MACR,QADoB,YAAA;;;;;;;;;;;mEAgCZ;;MACR,QADoB,YAAA;;;;cAyBX;;;;aAOD;;;;;;;;;;;;;;;;qBAuDa;;MAA2C;;;;6BAOlC,iBAAiB;;;;;;;;;;gCAaf,eAAoB,QAAQ;;;;;;;sCAuCpD,eACR,eAAe,YAAY;;;;gBAchB;;;;;;;;;;;;;;;;mBA6BS,QACrB;;;;;;;;;;;;;;;;wCA2ByC,iBAAsB;;;;gDAQN,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;oBA4BnE,eAAe,sBACb,oBACR,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;0CAoEC,gCACT,QAAQ;;;;mBAWY,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;MAyD5B,QAAQ;;;;2BAmCoB;;;;;kCAkBK;;;;qBAmDX"}
+{"version":3,"file":"gerbil-CvNqpVg0.d.mts","names":[],"sources":["../src/core/gerbil.ts"],"sourcesContent":[],"mappings":";;;;;AA8OsB,cAvBT,MAAA,CAuBS;EAgDC,QAAA,YAAA;EAIa,QAAA,WAAA;EA0BgB,iBAAA,MAAA;EAAmB,QAAA,KAAA;EAoLrD,QAAA,WAAA;EAyDsB,QAAA,YAAA;EAyCO,QAAA,iBAAA;EAAsB,QAAA,eAAA;EAqC7B,QAAA,SAAA;EAWM,QAAA,eAAA;EAAsB,QAAA,aAAA;EAgC5B,iBAAA,KAAA;EAWM,iBAAA,SAAA;EAAsB,WAAA,CAAA,MAAA,CAAA,EA/b9C,YA+b8C;EAgCvB,QAAA,WAAA;EAWO;;;;;EAyDa,QAAA,YAAA;EA0MpD,OAAA,UAAA,CAAA,CAAA,EA7rBU,WA6rBV,EAAA;EACe,OAAA,QAAA,CAAA,OAAA,EAAA,MAAA,CAAA,EA1rBQ,WA0rBR,GAAA,SAAA;EAAvB;;;;;;;;;;;;;;;;;;EAwZQ,SAAA,CAAA,OAAA,CAAA,EAAA,MAAA,EAAA,OAAA,CAAA,EAxjCuC,WAwjCvC,CAAA,EAxjC0D,OAwjC1D,CAAA,IAAA,CAAA;EAAY;;;;;;EAgEZ,QAAA,eAAA;EAuDa;;;EAO0B,QAAA,CAAA,CAAA,EAAA,OAAA;EAaf;;;EAuCxB,cAAA,CAAA,CAAA,EAAA,OAAA;EACO;;;EAcJ,YAAA,CAAA,CAAA,EArkCE,WAqkCF,GAAA,IAAA;EA8BZ;;;EA2B+D,aAAA,CAAA,CAAA,EAAA,QAAA,GAAA,KAAA,GAAA,MAAA;EAQN;;;;;EA8BhD,QAAA,CAAA,CAAA,EAAA,MAAA;EAAR;;;EAqEA,qBAAA,CAAA,CAAA,EAAA;IAW4B,IAAA,EAAA,MAAA;IAAR,MAAA,EAAA,MAAA;IAyDZ,IAAA,EAAA,MAAA;IAAR,OAAA,EAAA,MAAA;EAmC4B,CAAA;EAkBK;;;;;;;;;;;;;;;;kCAzyCE;;;;;;;;;;;;;;;;;;;;;;;;0CAyCO,iBAAsB;;;;;kCAqC7B;;;;;;;yCAWM,iBAAsB;;;;;kCAgC5B;;;;;;;yCAWM,iBAAsB;;;;;uCAgCvB;;;;;;;+CAWO,iBAAsB;;;;;;gBAmCpD;;;;;;;;;;;;;;;qCAsBoB,kBAAuB,QAAQ;;;;;;;;;;;mCA0M5D,kBACR,uBAAuB;;;;;;;;;;;;;;;;;;;mCA6Ma,YAAY,KAAK,QAAQ;;;;;;;;;;;;;;;;;wDAgDrD,wBACR,QAAQ,qBAAqB;;;;gCAiBG,eAAoB,QAAQ;;;;;;;;;;wCAoDpB,eAAoB,QAAQ;;;;;;;;;;;;;;;;;;;;qDAkD5D,eACR,QADyB,gBAAA;;;;;;;;;;;;;;oDAmCjB;;MACR,QADoB,YAAA;;;;;;;;;;;mEAgCZ;;MACR,QADoB,YAAA;;;;cAyBX;;;;aAOD;;;;;;;;;;;;;;;;qBAuDa;;MAA2C;;;;6BAOlC,iBAAiB;;;;;;;;;;gCAaf,eAAoB,QAAQ;;;;;;;sCAuCpD,eACR,eAAe,YAAY;;;;gBAchB;;;;;;;;;;;;;;;;mBA6BS,QACrB;;;;;;;;;;;;;;;;wCA2ByC,iBAAsB;;;;gDAQN,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;oBA4BnE,eAAe,sBACb,oBACR,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;0CAoEC,gCACT,QAAQ;;;;mBAWY,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;MAyD5B,QAAQ;;;;2BAmCoB;;;;;kCAkBK;;;;qBAmDX"}

package/dist/gpu/hooks.d.mts

@@ -1,5 +1,5 @@
-import { n as AgentTool, t as AgentStep } from "../index-DukkJRMj.mjs";
-import { c as MemorySearchResult, d as RecallOptions, f as RecallResult, m as SearchOptions, s as MemoryRecord, t as AddOptions } from "../types-DQBe2lFo.mjs";
+import { n as AgentTool, t as AgentStep } from "../index-Nl40RdSs.mjs";
+import { c as MemorySearchResult, d as RecallOptions, f as RecallResult, m as SearchOptions, s as MemoryRecord, t as AddOptions } from "../types-D7kn-0i2.mjs";
 import { ReactNode } from "react";
 import * as react_jsx_runtime0 from "react/jsx-runtime";