@tryhamster/gerbil 1.0.0 → 1.0.1

package/dist/browser/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","names":["BUILTIN_MODELS: Record<string, ModelConfig>","audioContext: AudioContext | null","MODEL_SIZES: Record<string, number>","chat: string","reason: string","state: SessionState","bytesDownloaded","e: any"],"sources":["../../src/core/models.ts","../../src/browser/pwa.ts","../../src/browser/audio.ts","../../src/browser/device-guards.ts","../../src/browser/download.ts"],"sourcesContent":["/**\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 * Mobile / PWA storage helpers.\n *\n * On-device models are large (a 4-bit 0.8B is ~400 MB; vision/larger models are\n * GBs). Mobile browsers — iOS Safari especially — wall a web origin off from the\n * real disk with TWO independent ceilings:\n *\n *   1. **Storage quota** (disk for the model cache). An *uninstalled* Safari tab\n *      gets only ~1 GB, best-effort and evictable, regardless of how much free\n *      disk the device has. Exceed it and every cache write fails → the model\n *      re-downloads on every visit.\n *   2. **Tab memory** (RAM during load/inference) — a separate, smaller ceiling.\n *\n * The unlock for the storage ceiling is **persistent storage**, which iOS Safari\n * grants when the site is **installed to the Home Screen** (a PWA). Installed, the\n * quota jumps to a large fraction of actual disk and is never evicted — so models\n * cache once and stay. These helpers let an app surface that to its users and\n * request it, so on-device AI is actually practical on mobile.\n *\n * All functions are SSR/Node-safe (guarded; return conservative defaults).\n */\n\n/** True when the page is running as an installed/standalone PWA (Home Screen). */\nexport function isStandalone(): boolean {\n  if (typeof window === \"undefined\") return false;\n  // iOS Safari exposes navigator.standalone; everyone else uses display-mode.\n  const iosStandalone = (navigator as { standalone?: boolean }).standalone === true;\n  const displayStandalone =\n    typeof window.matchMedia === \"function\" &&\n    window.matchMedia(\"(display-mode: standalone)\").matches;\n  return iosStandalone || displayStandalone;\n}\n\n/** True when running on iOS/iPadOS (where install is the quota unlock and the\n * install flow is manual: Share → Add to Home Screen). iPadOS masquerades as\n * macOS, so we also treat touch-capable WebKit-on-Mac as iOS. */\nexport function isIOS(): boolean {\n  if (typeof navigator === \"undefined\") return false;\n  const ua = navigator.userAgent || \"\";\n  if (/iPhone|iPad|iPod/.test(ua)) return true;\n  // iPadOS 13+ reports a Mac UA — detect via touch points + WebKit.\n  const isMacWebKit = /Macintosh/.test(ua) && /AppleWebKit/.test(ua) && !/Chrome/.test(ua);\n  return isMacWebKit && ((navigator as { maxTouchPoints?: number }).maxTouchPoints ?? 0) > 1;\n}\n\nexport type StorageStatus = {\n  /** Total quota granted to this origin, in MB (best-effort estimate). */\n  quotaMB: number;\n  /** Bytes currently used by this origin, in MB. */\n  usageMB: number;\n  /** quota − usage, in MB. */\n  availableMB: number;\n  /** Storage is persistent (exempt from eviction). On iOS this is effectively\n   * only true once the site is installed to the Home Screen. */\n  persisted: boolean;\n  /** Running as an installed/standalone PWA. */\n  installed: boolean;\n  /** Platform is iOS/iPadOS (install is the quota unlock here). */\n  ios: boolean;\n};\n\n/** Snapshot of the origin's storage situation — quota, usage, persistence, and\n * whether the app is installed. Use it to decide whether to recommend install\n * before downloading a large model. */\nexport async function getStorageStatus(): Promise<StorageStatus> {\n  const installed = isStandalone();\n  const ios = isIOS();\n  let quotaMB = 0;\n  let usageMB = 0;\n  let persisted = false;\n  try {\n    const est = await navigator.storage?.estimate?.();\n    quotaMB = Math.round((est?.quota || 0) / 1e6);\n    usageMB = Math.round((est?.usage || 0) / 1e6);\n  } catch {\n    /* estimate unsupported */\n  }\n  try {\n    persisted = (await navigator.storage?.persisted?.()) ?? false;\n  } catch {\n    /* persisted unsupported */\n  }\n  return {\n    quotaMB,\n    usageMB,\n    availableMB: Math.max(0, quotaMB - usageMB),\n    persisted,\n    installed,\n    ios,\n  };\n}\n\n/**\n * Request persistent storage (exempt from eviction). Returns whether the origin\n * is persistent afterwards. Browsers grant this based on engagement/installation;\n * on iOS Safari it is effectively granted only to an installed (Home Screen) PWA,\n * so call this AND guide users to install when it returns false on iOS.\n */\nexport async function requestPersistentStorage(): Promise<boolean> {\n  try {\n    if (await navigator.storage?.persisted?.()) return true;\n    return (await navigator.storage?.persist?.()) ?? false;\n  } catch {\n    return false;\n  }\n}\n\nexport type ModelFit = {\n  /** The model likely fits in the currently-available quota. */\n  fits: boolean;\n  availableMB: number;\n  /** Caching durably would benefit from installing to the Home Screen — true when\n   * not installed on iOS, or when the model doesn't fit the current quota. */\n  recommendInstall: boolean;\n};\n\n/**\n * Estimate whether a model of `sizeMB` will cache in the current quota, and\n * whether you should recommend installing to the Home Screen first. Pair with a\n * one-time \"Install for offline use\" prompt before a large download on mobile.\n */\nexport async function canCacheModel(sizeMB: number): Promise<ModelFit> {\n  const s = await getStorageStatus();\n  // Headroom: leave ~10% slack so we don't recommend a download that just barely\n  // fits and then fails mid-write.\n  const fits = s.availableMB >= sizeMB * 1.1;\n  const recommendInstall = (!fits || (s.ios && !s.installed)) && !s.persisted;\n  return { fits, availableMB: s.availableMB, recommendInstall };\n}\n\n/**\n * Platform-appropriate install guidance. iOS Safari has NO programmatic install\n * prompt — installation is manual (Share → Add to Home Screen), so apps should\n * show these instructions. Other platforms (Android/Chrome) fire\n * `beforeinstallprompt`, which apps can capture for a one-tap button.\n */\nexport function getInstallGuidance(): { installed: boolean; manual: boolean; steps: string } {\n  const installed = isStandalone();\n  if (installed) return { installed: true, manual: false, steps: \"Already installed.\" };\n  if (isIOS()) {\n    return {\n      installed: false,\n      manual: true,\n      steps:\n        \"Tap the Share button, then 'Add to Home Screen'. Installing unlocks durable storage so models download once instead of every visit.\",\n    };\n  }\n  return {\n    installed: false,\n    manual: false,\n    steps:\n      \"Use your browser's Install option (or the install icon in the address bar) to add this app for offline use and durable model storage.\",\n  };\n}\n","// ============================================\n// Audio Playback Utilities\n// ============================================\n\n/**\n * Play audio from Float32Array using Web Audio API\n *\n * @example\n * ```ts\n * import { playAudio } from \"@tryhamster/gerbil/browser\";\n *\n * const audio = new Float32Array([...]); // TTS output\n * const controller = await playAudio(audio, 24000);\n *\n * // Stop playback\n * controller.stop();\n * ```\n */\nexport async function playAudio(\n  audio: Float32Array,\n  sampleRate: number = 24000,\n): Promise<{ stop: () => void; onEnded: Promise<void> }> {\n  const audioContext = new AudioContext();\n\n  // Resume if suspended\n  if (audioContext.state === \"suspended\") {\n    await audioContext.resume();\n  }\n\n  const audioBuffer = audioContext.createBuffer(1, audio.length, sampleRate);\n  const channelData = new Float32Array(audio);\n  audioBuffer.copyToChannel(channelData, 0);\n\n  const sourceNode = audioContext.createBufferSource();\n  sourceNode.buffer = audioBuffer;\n  sourceNode.connect(audioContext.destination);\n\n  const onEnded = new Promise<void>((resolve) => {\n    sourceNode.onended = () => {\n      audioContext.close();\n      resolve();\n    };\n  });\n\n  sourceNode.start();\n\n  return {\n    stop: () => {\n      sourceNode.stop();\n      audioContext.close();\n    },\n    onEnded,\n  };\n}\n\n/**\n * Create a reusable audio player for streaming TTS\n *\n * @example\n * ```ts\n * import { createAudioPlayer } from \"@tryhamster/gerbil/browser\";\n *\n * const player = createAudioPlayer(24000);\n *\n * // Queue audio chunks as they arrive\n * player.queue(chunk1);\n * player.queue(chunk2);\n *\n * // Stop and clear\n * player.stop();\n * ```\n */\nexport function createAudioPlayer(sampleRate: number = 24000): {\n  queue: (audio: Float32Array) => void;\n  stop: () => void;\n  isPlaying: () => boolean;\n} {\n  let audioContext: AudioContext | null = null;\n  let nextStartTime = 0;\n  let isActive = false;\n\n  const ensureContext = async () => {\n    if (!audioContext) {\n      audioContext = new AudioContext();\n    }\n    if (audioContext.state === \"suspended\") {\n      await audioContext.resume();\n    }\n    return audioContext;\n  };\n\n  return {\n    queue: async (audio: Float32Array) => {\n      const ctx = await ensureContext();\n      isActive = true;\n\n      const buffer = ctx.createBuffer(1, audio.length, sampleRate);\n      const channelData = new Float32Array(audio);\n      buffer.copyToChannel(channelData, 0);\n\n      const source = ctx.createBufferSource();\n      source.buffer = buffer;\n      source.connect(ctx.destination);\n\n      // Schedule seamlessly after previous chunk\n      const startTime = Math.max(ctx.currentTime, nextStartTime);\n      source.start(startTime);\n      nextStartTime = startTime + buffer.duration;\n\n      source.onended = () => {\n        if (ctx.currentTime >= nextStartTime - 0.1) {\n          isActive = false;\n        }\n      };\n    },\n\n    stop: () => {\n      isActive = false;\n      nextStartTime = 0;\n      if (audioContext) {\n        audioContext.close();\n        audioContext = null;\n      }\n    },\n\n    isPlaying: () => isActive,\n  };\n}\n","// ============================================\n// iOS Model Guards & Device Capability Detection\n// ============================================\n\n// ============================================\n// Real native-engine model ids\n// ============================================\n//\n// The website loads models by their actual repo id (MLX 4-bit on-device builds,\n// or the upstream Qwen/Liquid repos on desktop). The guard matches on those ids\n// — NOT the old ONNX-era shorthands — so it must stay in sync with the real\n// checkpoints the engine can run. Matching is substring-based on a normalized\n// id so both `mlx-community/Qwen3.5-0.8B-4bit` and `Qwen/Qwen3.5-0.8B` resolve.\n\n/** Recommended safe model ids per modality (used as fallbacks on mobile). */\nconst SAFE_MOBILE_CHAT = \"mlx-community/Qwen3.5-0.8B-4bit\";\n\n/**\n * Approximate on-device (INT4) memory footprint in MB for the models the native\n * engine actually ships. Used for memory-aware selection and messaging.\n */\nexport const MODEL_SIZES: Record<string, number> = {\n  // Chat models (INT4, on-device)\n  \"qwen3.5-0.8b\": 650, // ~0.65GB\n  \"qwen3.5-2b\": 1700, // ~1.7GB\n  \"gemma-4-e2b\": 3600, // ~3.6GB\n  \"lfm2.5-350m\": 300, // ~0.3GB\n  // TTS models\n  \"kokoro-82m\": 350,\n  \"supertonic-66m\": 300,\n  // STT models\n  \"whisper-tiny\": 150,\n  \"whisper-tiny.en\": 150,\n  \"whisper-small\": 500,\n  // Embedding models\n  \"all-minilm-l6-v2\": 100,\n};\n\n/**\n * Normalize a repo/model id to a lowercase token stream for substring matching\n * (strips org prefixes' punctuation while preserving the model name tokens).\n */\nfunction normalizeId(modelId: string): string {\n  return modelId.toLowerCase().replace(/[^a-z0-9]/g, \"-\");\n}\n\n/**\n * iOS (WKWebView) model classification keyed off the REAL native-engine ids.\n *\n * - blocked: too large for the WKWebView memory ceiling on iPhone — will crash.\n *   gemma-4-e2b (~3.6GB) plus any vision checkpoint (the vision encoder pushes\n *   the working set well past what an iPhone can hold).\n * - risky: borderline on iPhone — Qwen3.5-2B (~1.7GB) fits on newer devices but\n *   warns and can OOM on older ones.\n * - everything else (Qwen3.5-0.8B ~0.65GB, LFM2.5-350M) is allowed everywhere.\n */\nconst IOS_MODEL_LIMITS = {\n  /** Substrings (normalized) of ids that HARD-BLOCK on iPhone. */\n  blocked: [\"gemma-4-e2b\", \"gemma-4-e4b\"],\n  /** Substrings that mark a vision checkpoint (blocked on iPhone). */\n  visionMarkers: [\"vision\", \"-vl-\", \"-vl\", \"vlm\", \"image-text\", \"-it-vision\"],\n  /** Substrings of ids that WARN on iPhone (Qwen3.5-2B class). */\n  risky: [\"qwen3.5-2b\", \"qwen3-5-2b\"],\n  /** Maximum total memory budget in MB for iOS WKWebView. */\n  maxBudgetMB: 1800,\n} as const;\n\n/**\n * Check if a model is safe to load on the current device.\n * Returns guidance specific to iOS memory constraints. Matches on the real\n * native-engine repo ids (MLX 4-bit / upstream Qwen / Liquid).\n */\nexport function isModelSafeForDevice(modelId: string): {\n  safe: boolean;\n  reason: string;\n  recommendation?: string;\n  maxSafeModel?: string;\n} {\n  const ua = typeof navigator !== \"undefined\" ? navigator.userAgent : \"\";\n  const isIPhone = /iPhone|iPod/.test(ua);\n  const isIPad = /iPad/.test(ua);\n  const isIOS = isIPhone || isIPad;\n  const isIOSChrome = isIOS && /CriOS/.test(ua);\n  const normalizedId = normalizeId(modelId);\n\n  const isVision = IOS_MODEL_LIMITS.visionMarkers.some((m) =>\n    normalizedId.includes(normalizeId(m)),\n  );\n  const isBlocked =\n    IOS_MODEL_LIMITS.blocked.some((m) => normalizedId.includes(normalizeId(m))) || isVision;\n  const isRisky = IOS_MODEL_LIMITS.risky.some((m) => normalizedId.includes(normalizeId(m)));\n\n  // iPhone is the hard constraint (smallest WKWebView budget). iPad has more\n  // headroom, so only the genuinely huge / vision checkpoints are blocked there.\n  if (isIPhone) {\n    if (isBlocked) {\n      const browserNote = isIOSChrome ? \" (iOS Chrome uses WKWebView, same limits as Safari)\" : \"\";\n      const why = isVision\n        ? \"Vision checkpoints need a separate image encoder in memory\"\n        : \"It is too large (~3.6GB)\";\n      return {\n        safe: false,\n        reason: `Model ${modelId} will crash on iPhone${browserNote}. ${why}, which exceeds the WKWebView memory ceiling.`,\n        recommendation: `Use ${SAFE_MOBILE_CHAT} (Qwen3.5-0.8B) on iPhone, or run larger models on desktop.`,\n        maxSafeModel: SAFE_MOBILE_CHAT,\n      };\n    }\n    if (isRisky) {\n      return {\n        safe: true,\n        reason: `Model ${modelId} (~1.7GB) is borderline on iPhone and may run out of memory on older devices.`,\n        recommendation: `If it crashes, fall back to ${SAFE_MOBILE_CHAT} (Qwen3.5-0.8B).`,\n        maxSafeModel: SAFE_MOBILE_CHAT,\n      };\n    }\n    return { safe: true, reason: \"Model is within iPhone memory limits.\" };\n  }\n\n  if (isIPad) {\n    // iPad tolerates the 2B class; only block the huge / vision checkpoints.\n    if (isBlocked) {\n      const why = isVision\n        ? \"Vision checkpoints need a separate image encoder in memory\"\n        : \"It is too large (~3.6GB)\";\n      return {\n        safe: false,\n        reason: `Model ${modelId} may crash on iPad. ${why}, which can exceed the WKWebView memory ceiling.`,\n        recommendation: `Use ${SAFE_MOBILE_CHAT} (Qwen3.5-0.8B) or Qwen3.5-2B on iPad.`,\n        maxSafeModel: \"mlx-community/Qwen3.5-2B-4bit\",\n      };\n    }\n    return { safe: true, reason: \"Model is within iPad memory limits.\" };\n  }\n\n  // Android - block the huge / vision checkpoints, allow the rest.\n  const isAndroid = /Android/.test(ua);\n  if (isAndroid && isBlocked) {\n    return {\n      safe: false,\n      reason: `Model ${modelId} is very large and may crash on Android devices.`,\n      recommendation: `Use ${SAFE_MOBILE_CHAT} (Qwen3.5-0.8B) or Qwen3.5-2B on Android.`,\n      maxSafeModel: \"mlx-community/Qwen3.5-2B-4bit\",\n    };\n  }\n\n  // Desktop - all models are safe.\n  return { safe: true, reason: \"Desktop browser has sufficient memory.\" };\n}\n\n/**\n * Get recommended models based on device memory and capabilities.\n * Helps prevent OOM crashes on low-memory mobile devices.\n */\nexport function getRecommendedModels(): {\n  chat: string;\n  tts: string;\n  stt: string;\n  embedding: string;\n  reason: string;\n  deviceMemory: number | null;\n  isMobile: boolean;\n} {\n  const ua = typeof navigator !== \"undefined\" ? navigator.userAgent : \"\";\n  const deviceMemory = typeof navigator !== \"undefined\" ? (navigator as any).deviceMemory : null;\n  const isMobile = /iPhone|iPad|iPod|Android|Mobile/.test(ua);\n\n  // Estimate available memory (deviceMemory reports total GB, not available)\n  // Mobile devices typically have less free memory due to OS overhead\n  const effectiveMemory = deviceMemory ? (isMobile ? deviceMemory * 0.4 : deviceMemory * 0.6) : 4;\n  const availableMB = effectiveMemory * 1024;\n\n  let chat: string;\n  let reason: string;\n\n  if (availableMB < 600) {\n    chat = \"LiquidAI/LFM2.5-350M\";\n    reason = \"Very low memory device - using smallest model (LFM2.5-350M)\";\n  } else if (isMobile && availableMB < 2200) {\n    chat = SAFE_MOBILE_CHAT;\n    reason = \"Mobile device - using Qwen3.5-0.8B to stay within the WKWebView memory limit\";\n  } else if (availableMB < 2200) {\n    chat = SAFE_MOBILE_CHAT;\n    reason = \"Standard model for moderate memory (Qwen3.5-0.8B)\";\n  } else {\n    chat = \"mlx-community/Qwen3.5-2B-4bit\";\n    reason = \"High memory available - using Qwen3.5-2B for better quality\";\n  }\n\n  return {\n    chat,\n    tts: \"kokoro-82m\",\n    stt: \"whisper-tiny.en\",\n    embedding: \"all-MiniLM-L6-v2\",\n    reason,\n    deviceMemory,\n    isMobile,\n  };\n}\n\n// ============================================\n// Session Phase Tracking (Reload Detection)\n// ============================================\n\ntype DownloadPhase = \"idle\" | \"downloading\" | \"caching\" | \"initializing\" | \"ready\" | \"error\";\n\nexport const SESSION_STORAGE_KEY = \"gerbil_session_phase\";\n\nexport type SessionState = {\n  phase: DownloadPhase;\n  modelId: string | null;\n  sessionId: string;\n  timestamp: number;\n  bytesDownloaded?: number;\n  totalBytes?: number;\n};\n\n/**\n * Generate a unique session ID for tracking across reloads.\n */\nfunction generateSessionId(): string {\n  return `${Date.now()}-${Math.random().toString(36).slice(2, 9)}`;\n}\n\n/**\n * Get or create the current session ID.\n */\nfunction getSessionId(): string {\n  if (typeof localStorage === \"undefined\") return generateSessionId();\n\n  let sessionId = sessionStorage.getItem(\"gerbil_session_id\");\n  if (!sessionId) {\n    sessionId = generateSessionId();\n    sessionStorage.setItem(\"gerbil_session_id\", sessionId);\n  }\n  return sessionId;\n}\n\n/**\n * Set the current download/initialization phase.\n * Used to detect if a reload happened during a critical operation.\n */\nexport function setDownloadPhase(\n  phase: DownloadPhase,\n  modelId?: string,\n  progress?: { bytesDownloaded: number; totalBytes: number },\n): void {\n  if (typeof localStorage === \"undefined\") return;\n\n  const state: SessionState = {\n    phase,\n    modelId: modelId || null,\n    sessionId: getSessionId(),\n    timestamp: Date.now(),\n    bytesDownloaded: progress?.bytesDownloaded,\n    totalBytes: progress?.totalBytes,\n  };\n\n  localStorage.setItem(SESSION_STORAGE_KEY, JSON.stringify(state));\n}\n\n/**\n * Get the last known download phase from storage.\n */\nexport function getDownloadPhase(): SessionState | null {\n  if (typeof localStorage === \"undefined\") return null;\n\n  try {\n    const raw = localStorage.getItem(SESSION_STORAGE_KEY);\n    if (!raw) return null;\n    return JSON.parse(raw) as SessionState;\n  } catch {\n    return null;\n  }\n}\n\n/**\n * Detect if the page reloaded during a model download/initialization.\n * This typically indicates an iOS memory crash.\n *\n * @returns Detection result with recommended action\n */\nexport function detectMemoryCrash(): {\n  crashed: boolean;\n  phase?: DownloadPhase;\n  modelId?: string;\n  timeSinceCrash?: number;\n  recommendation?: string;\n} {\n  const lastState = getDownloadPhase();\n  const currentSessionId = getSessionId();\n\n  if (!lastState) {\n    return { crashed: false };\n  }\n\n  // If session ID changed and we were in a critical phase, it's likely a crash/reload\n  const criticalPhases: DownloadPhase[] = [\"downloading\", \"caching\", \"initializing\"];\n  const wasInCriticalPhase = criticalPhases.includes(lastState.phase);\n  const sessionChanged = lastState.sessionId !== currentSessionId;\n  const timeSinceCrash = Date.now() - lastState.timestamp;\n\n  // Only count as crash if it happened recently (within 5 minutes) and session changed\n  const recentCrash = timeSinceCrash < 5 * 60 * 1000;\n\n  if (wasInCriticalPhase && sessionChanged && recentCrash) {\n    // Clear the state to prevent repeated detection\n    localStorage.removeItem(SESSION_STORAGE_KEY);\n\n    return {\n      crashed: true,\n      phase: lastState.phase,\n      modelId: lastState.modelId || undefined,\n      timeSinceCrash,\n      recommendation:\n        lastState.modelId && /2b|gemma|vision/i.test(lastState.modelId)\n          ? \"The model was too large for your device. Try Qwen3.5-0.8B instead.\"\n          : \"Your device ran out of memory. Try a smaller model or use a desktop browser.\",\n    };\n  }\n\n  return { crashed: false };\n}\n\n/**\n * Clear session phase (call when model loads successfully).\n */\nexport function clearDownloadPhase(): void {\n  if (typeof localStorage === \"undefined\") return;\n  localStorage.removeItem(SESSION_STORAGE_KEY);\n}\n\n// ============================================\n// WebKit Submit-Granularity (group size) Probe\n// ============================================\n//\n// On WebKit/iOS WebGPU, decode speed is bound by GPU round-trips: the executor\n// groups `webkitGroupSize` dispatches per command buffer, then submits + drains\n// (queue.onSubmittedWorkDone) per group. Larger groups amortize the drains and\n// can be ~5x faster on iPad. BUT the safe ceiling is device-dependent in three\n// classes:\n//   1. Newer WebKit (iPad 26.5): high group is correct + fast.\n//   2. Older WebKit: high group yields zero/garbage logits (wrong output).\n//   3. iPhone (iOS 18.7): high group HARD-CRASHES the GPU process (page dies).\n//\n// The crash class is the hard one: a bad group kills the page before any code\n// can record the result. So we use a localStorage BREADCRUMB that OUTLIVES the\n// page kill: we persist the candidate we're about to try BEFORE running any\n// GPU-heavy work. If the page survives, a later promotion clears `trying` and\n// records it as known-good. If the page crashed, `trying` is still set on the\n// next load, so we KNOW that rung crashed and cap below it.\n//\n// The probe escalates UP from the safe floor (never starts optimistic — a\n// first-visit crash is a terrible UX) and persists per-device in localStorage.\n\nexport const WEBKIT_GROUP_PROBE_KEY = \"gerbil-webkit-group-v1\";\n\n/**\n * Candidate group-size ladder. We escalate UP one rung per page load. 128 is\n * treated as effectively \"batch-all\" for this model — kept simple on purpose.\n */\nexport const WEBKIT_GROUP_LADDER = [1, 8, 32, 64, 128] as const;\n\nexport type WebkitGroupProbe = {\n  /** Largest group size proven crash-free AND correct on this device. Starts at 1. */\n  knownGood: number;\n  /** Candidate being attempted this page-load (null when not probing). */\n  trying: number | null;\n  /** True once we hit a crash/incorrect ceiling and should stop escalating. */\n  capped: boolean;\n};\n\nconst DEFAULT_PROBE: WebkitGroupProbe = { knownGood: 1, trying: null, capped: false };\n\n/** Read the persisted WebKit group probe record (guarded; safe on node). */\nexport function readGroupProbe(): WebkitGroupProbe {\n  if (typeof localStorage === \"undefined\") return { ...DEFAULT_PROBE };\n  try {\n    const raw = localStorage.getItem(WEBKIT_GROUP_PROBE_KEY);\n    if (!raw) return { ...DEFAULT_PROBE };\n    const parsed = JSON.parse(raw) as Partial<WebkitGroupProbe>;\n    return {\n      knownGood:\n        typeof parsed.knownGood === \"number\" && parsed.knownGood >= 1 ? parsed.knownGood : 1,\n      trying: typeof parsed.trying === \"number\" ? parsed.trying : null,\n      capped: parsed.capped === true,\n    };\n  } catch {\n    return { ...DEFAULT_PROBE };\n  }\n}\n\n/** Persist the WebKit group probe record (guarded; no-op on node). */\nexport function writeGroupProbe(rec: WebkitGroupProbe): void {\n  if (typeof localStorage === \"undefined\") return;\n  try {\n    localStorage.setItem(WEBKIT_GROUP_PROBE_KEY, JSON.stringify(rec));\n  } catch {\n    // Best-effort; private mode / quota can throw.\n  }\n}\n\n/**\n * The validated non-phone sweet spot. iPad swept 1→7.9, 8→19, 32→24.8, 64→26.6,\n * 128→26.9 (peak), 256→26.2 tok/s — a plateau from ~64 up, so 128 is the best\n * stable target (more batching just costs memory). Non-phone WebKit jumps here\n * directly; the crash breadcrumb caps it down if a device can't sustain it.\n */\nconst NONPHONE_TARGET_GROUP = 128;\n\n/**\n * Resolve the WebKit group size to use this session, recording `trying` as a\n * side effect so a crash this load is detectable on the next load.\n *\n * Algorithm (only meaningful on WebKit; inert otherwise):\n *  1. Read the record (default {knownGood:1, trying:null, capped:false}).\n *  2. If `trying !== null` on entry → the previous load set it but never cleared\n *     it → that load CRASHED at `trying`. Cap there, keep `knownGood`, clear\n *     `trying`. Use `knownGood` this session.\n *  3. Else if !capped and there is a rung above `knownGood` → set `trying = next`,\n *     persist BEFORE any GPU work, and use it (we're escalating).\n *  4. Else → use `knownGood`.\n *\n * @returns the group size to use this session.\n */\nexport function resolveWebkitGroupSize(args: {\n  override?: number;\n  isWebKit: boolean;\n  /**\n   * Memory-constrained devices (phones) are the \"crash class\": batching crashes\n   * the GPU process, and discovering that costs one user-visible crash before the\n   * breadcrumb caps it. When `conservative` is set we never auto-escalate — the\n   * device stays at its proven floor (group=1 on first run) and never risks that\n   * calibration crash. `?group=N` still lets such a device opt in explicitly.\n   */\n  conservative?: boolean;\n}): number {\n  // Explicit ?group=N override wins and skips the probe entirely.\n  if (args.override && args.override > 0) return args.override;\n  // Non-WebKit (desktop / Dawn / node) never groups — floor of 1, no bookkeeping.\n  if (!args.isWebKit) return 1;\n\n  const rec = readGroupProbe();\n\n  // Step 2: a stale `trying` means last load crashed at that rung.\n  if (rec.trying !== null) {\n    const crashedAt = rec.trying;\n    const next: WebkitGroupProbe = {\n      knownGood: rec.knownGood,\n      trying: null,\n      capped: true,\n    };\n    writeGroupProbe(next);\n    console.log(\n      `[engine] webkit group probe: previous load crashed at group=${crashedAt} → capping at knownGood=${next.knownGood}`,\n    );\n    return next.knownGood;\n  }\n\n  // Step 3: non-phone WebKit (iPad / Mac Safari) jumps straight to the validated\n  // sweet spot (group=128) instead of climbing rung-by-rung — these devices\n  // tolerate batching (iPad ran every rung to 256 coherently), so there's no\n  // reason to crawl. The breadcrumb (Step 2) caps it down if a particular device\n  // can't sustain it. Phones (conservative) NEVER escalate — they stay at the\n  // proven floor (group=1) to avoid the one calibration crash batching costs on\n  // the crash class.\n  if (!rec.capped && !args.conservative && rec.knownGood < NONPHONE_TARGET_GROUP) {\n    const next: WebkitGroupProbe = {\n      knownGood: rec.knownGood,\n      trying: NONPHONE_TARGET_GROUP,\n      capped: false,\n    };\n    writeGroupProbe(next); // PERSIST BEFORE any GPU-heavy work runs\n    console.log(\n      `[engine] webkit group probe: trying target group=${NONPHONE_TARGET_GROUP} (knownGood=${rec.knownGood})`,\n    );\n    return NONPHONE_TARGET_GROUP;\n  }\n\n  // Step 4: nothing to escalate (capped or at top rung) — use the proven floor.\n  console.log(\n    `[engine] webkit group probe: knownGood=${rec.knownGood} trying=null capped=${rec.capped} → using ${rec.knownGood}`,\n  );\n  return rec.knownGood;\n}\n\n/**\n * Promote (or cap) the WebKit group probe after the first successful forward.\n *\n * Call this once per page-load, after the model has loaded AND a first forward\n * completed without the page dying. The breadcrumb already handles the crash\n * class (the page death leaves `trying` set for the next load); this handles the\n * wrong-output class and records success.\n *\n * @param correct  true if the first forward produced non-corrupt output.\n *   - correct  → promote: knownGood = trying, trying = null.\n *   - incorrect → cap: keep knownGood at the prior rung, trying = null, capped.\n */\nexport function promoteGroupProbe(correct: boolean): void {\n  if (typeof localStorage === \"undefined\") return;\n  const rec = readGroupProbe();\n  // Nothing was being attempted this load — nothing to promote.\n  if (rec.trying === null) return;\n\n  if (correct) {\n    writeGroupProbe({ knownGood: rec.trying, trying: null, capped: rec.capped });\n    console.log(`[engine] webkit group probe: PROMOTED group=${rec.trying} to known-good`);\n  } else {\n    // Wrong-output class: this rung is incorrect (but did not crash). Keep the\n    // prior knownGood and stop escalating.\n    writeGroupProbe({ knownGood: rec.knownGood, trying: null, capped: true });\n    console.log(\n      `[engine] webkit group probe: group=${rec.trying} produced INCORRECT output → capping at knownGood=${rec.knownGood}`,\n    );\n  }\n}\n","import { setDownloadPhase } from \"./device-guards.js\";\n\n// ============================================\n// Chunked Resumable Downloader\n// ============================================\n\n/** Chunk size for downloads: 1.5MB (safe for iOS IndexedDB transactions) */\nexport const CHUNK_SIZE_BYTES = 1.5 * 1024 * 1024;\n\n/** IndexedDB database name for chunked downloads */\nexport const DOWNLOAD_DB_NAME = \"gerbil-model-chunks\";\nconst DOWNLOAD_DB_VERSION = 1;\n\n/**\n * Manifest stored in IndexedDB to track download progress.\n */\ntype DownloadManifest = {\n  modelId: string;\n  url: string;\n  etag: string | null;\n  totalBytes: number;\n  chunkSize: number;\n  completedChunks: number[];\n  createdAt: number;\n  updatedAt: number;\n};\n\n/**\n * Open (or create) the IndexedDB for chunked downloads.\n */\nasync function openDownloadDB(): Promise<IDBDatabase> {\n  return new Promise((resolve, reject) => {\n    const request = indexedDB.open(DOWNLOAD_DB_NAME, DOWNLOAD_DB_VERSION);\n\n    request.onerror = () =>\n      reject(new Error(`Failed to open download DB: ${request.error?.message}`));\n\n    request.onsuccess = () => resolve(request.result);\n\n    request.onupgradeneeded = (event) => {\n      const db = (event.target as IDBOpenDBRequest).result;\n\n      // Store for download manifests\n      if (!db.objectStoreNames.contains(\"manifests\")) {\n        db.createObjectStore(\"manifests\", { keyPath: \"modelId\" });\n      }\n\n      // Store for actual chunks (key: modelId-chunkIndex)\n      if (!db.objectStoreNames.contains(\"chunks\")) {\n        db.createObjectStore(\"chunks\");\n      }\n    };\n  });\n}\n\n/**\n * Get download manifest for a model.\n */\nasync function getManifest(db: IDBDatabase, modelId: string): Promise<DownloadManifest | null> {\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction(\"manifests\", \"readonly\");\n    const store = tx.objectStore(\"manifests\");\n    const request = store.get(modelId);\n\n    request.onerror = () => reject(new Error(`Failed to get manifest: ${request.error?.message}`));\n    request.onsuccess = () => resolve(request.result || null);\n  });\n}\n\n/**\n * Save download manifest.\n */\nasync function saveManifest(db: IDBDatabase, manifest: DownloadManifest): Promise<void> {\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction(\"manifests\", \"readwrite\");\n    const store = tx.objectStore(\"manifests\");\n    const request = store.put(manifest);\n\n    request.onerror = () => reject(new Error(`Failed to save manifest: ${request.error?.message}`));\n    request.onsuccess = () => resolve();\n  });\n}\n\n/**\n * Save a single chunk.\n */\nasync function saveChunk(\n  db: IDBDatabase,\n  modelId: string,\n  chunkIndex: number,\n  data: ArrayBuffer,\n): Promise<void> {\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction(\"chunks\", \"readwrite\");\n    const store = tx.objectStore(\"chunks\");\n    const key = `${modelId}-${chunkIndex}`;\n    const request = store.put(data, key);\n\n    request.onerror = () =>\n      reject(new Error(`Failed to save chunk ${chunkIndex}: ${request.error?.message}`));\n    request.onsuccess = () => resolve();\n  });\n}\n\n/**\n * Get a single chunk.\n */\nasync function getChunk(\n  db: IDBDatabase,\n  modelId: string,\n  chunkIndex: number,\n): Promise<ArrayBuffer | null> {\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction(\"chunks\", \"readonly\");\n    const store = tx.objectStore(\"chunks\");\n    const key = `${modelId}-${chunkIndex}`;\n    const request = store.get(key);\n\n    request.onerror = () =>\n      reject(new Error(`Failed to get chunk ${chunkIndex}: ${request.error?.message}`));\n    request.onsuccess = () => resolve(request.result || null);\n  });\n}\n\n/**\n * Delete all chunks and manifest for a model.\n */\nasync function clearModelData(db: IDBDatabase, modelId: string): Promise<void> {\n  // Get manifest to know how many chunks to delete\n  const manifest = await getManifest(db, modelId);\n\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction([\"manifests\", \"chunks\"], \"readwrite\");\n\n    // Delete manifest\n    tx.objectStore(\"manifests\").delete(modelId);\n\n    // Delete all chunks\n    if (manifest) {\n      const totalChunks = Math.ceil(manifest.totalBytes / manifest.chunkSize);\n      const chunkStore = tx.objectStore(\"chunks\");\n      for (let i = 0; i < totalChunks; i++) {\n        chunkStore.delete(`${modelId}-${i}`);\n      }\n    }\n\n    tx.oncomplete = () => resolve();\n    tx.onerror = () => reject(new Error(`Failed to clear model data: ${tx.error?.message}`));\n  });\n}\n\n/**\n * Chunked resumable downloader for large model files.\n * Downloads in 1.5MB chunks to avoid iOS memory pressure.\n */\nexport async function downloadModelChunked(\n  url: string,\n  modelId: string,\n  options: {\n    onProgress?: (info: {\n      phase: string;\n      bytesDownloaded: number;\n      totalBytes: number;\n      percent: number;\n    }) => void;\n    signal?: AbortSignal;\n  } = {},\n): Promise<ArrayBuffer> {\n  const { onProgress, signal } = options;\n\n  // Update session phase\n  setDownloadPhase(\"downloading\", modelId);\n\n  const db = await openDownloadDB();\n\n  try {\n    // Check for existing manifest\n    let manifest = await getManifest(db, modelId);\n\n    // Fetch headers to get content-length and etag\n    const headResponse = await fetch(url, { method: \"HEAD\", signal });\n    if (!headResponse.ok) {\n      throw new Error(`HEAD request failed: ${headResponse.status} ${headResponse.statusText}`);\n    }\n\n    const contentLength = parseInt(headResponse.headers.get(\"content-length\") || \"0\", 10);\n    const etag = headResponse.headers.get(\"etag\");\n    const acceptRanges = headResponse.headers.get(\"accept-ranges\");\n\n    if (!contentLength) {\n      throw new Error(\"Server did not provide content-length\");\n    }\n\n    // Check if we need to restart (etag mismatch means model updated)\n    if (manifest && manifest.etag !== etag) {\n      console.warn(`Model ${modelId} has been updated (etag mismatch). Clearing cached chunks.`);\n      await clearModelData(db, modelId);\n      manifest = null;\n    }\n\n    // Check if server supports range requests\n    const supportsRange = acceptRanges === \"bytes\";\n\n    if (!supportsRange) {\n      // Fall back to regular download\n      console.warn(`Server doesn't support range requests for ${modelId}. Using regular download.`);\n      db.close();\n\n      const response = await fetch(url, { signal });\n      if (!response.ok) throw new Error(`Download failed: ${response.status}`);\n\n      setDownloadPhase(\"caching\", modelId);\n      const buffer = await response.arrayBuffer();\n      setDownloadPhase(\"ready\", modelId);\n      return buffer;\n    }\n\n    // Create or update manifest\n    const totalChunks = Math.ceil(contentLength / CHUNK_SIZE_BYTES);\n\n    if (!manifest) {\n      manifest = {\n        modelId,\n        url,\n        etag,\n        totalBytes: contentLength,\n        chunkSize: CHUNK_SIZE_BYTES,\n        completedChunks: [],\n        createdAt: Date.now(),\n        updatedAt: Date.now(),\n      };\n      await saveManifest(db, manifest);\n    }\n\n    // Download missing chunks\n    for (let i = 0; i < totalChunks; i++) {\n      if (signal?.aborted) {\n        throw new Error(\"Download aborted\");\n      }\n\n      // Skip already downloaded chunks\n      if (manifest.completedChunks.includes(i)) {\n        const bytesDownloaded = (manifest.completedChunks.length / totalChunks) * contentLength;\n        onProgress?.({\n          phase: \"resuming\",\n          bytesDownloaded,\n          totalBytes: contentLength,\n          percent: Math.round((bytesDownloaded / contentLength) * 100),\n        });\n        continue;\n      }\n\n      const start = i * CHUNK_SIZE_BYTES;\n      const end = Math.min(start + CHUNK_SIZE_BYTES - 1, contentLength - 1);\n\n      // Download chunk with Range header\n      const response = await fetch(url, {\n        headers: { Range: `bytes=${start}-${end}` },\n        signal,\n      });\n\n      if (response.status !== 206) {\n        throw new Error(`Range request failed: ${response.status} (expected 206)`);\n      }\n\n      const chunkData = await response.arrayBuffer();\n\n      // Save chunk to IndexedDB\n      await saveChunk(db, modelId, i, chunkData);\n\n      // Update manifest\n      manifest.completedChunks.push(i);\n      manifest.updatedAt = Date.now();\n      await saveManifest(db, manifest);\n\n      // Update session phase with progress\n      const bytesDownloaded = manifest.completedChunks.length * CHUNK_SIZE_BYTES;\n      setDownloadPhase(\"downloading\", modelId, { bytesDownloaded, totalBytes: contentLength });\n\n      onProgress?.({\n        phase: \"downloading\",\n        bytesDownloaded: Math.min(bytesDownloaded, contentLength),\n        totalBytes: contentLength,\n        percent: Math.round((manifest.completedChunks.length / totalChunks) * 100),\n      });\n\n      // Null out reference to allow GC before next chunk\n      // @ts-expect-error - intentional null for GC\n      response.body = null;\n    }\n\n    // All chunks downloaded - reassemble\n    setDownloadPhase(\"caching\", modelId);\n    onProgress?.({\n      phase: \"assembling\",\n      bytesDownloaded: contentLength,\n      totalBytes: contentLength,\n      percent: 100,\n    });\n\n    // Assemble chunks into final buffer\n    // We do this incrementally to avoid holding all chunks in memory at once\n    const finalBuffer = new ArrayBuffer(contentLength);\n    const finalView = new Uint8Array(finalBuffer);\n\n    for (let i = 0; i < totalChunks; i++) {\n      const chunk = await getChunk(db, modelId, i);\n      if (!chunk) {\n        throw new Error(`Missing chunk ${i} during assembly`);\n      }\n\n      const offset = i * CHUNK_SIZE_BYTES;\n      finalView.set(new Uint8Array(chunk), offset);\n\n      // Null out chunk reference immediately after copying\n      // This helps GC reclaim memory on iOS\n    }\n\n    // Clean up - delete chunks now that we have the full file\n    await clearModelData(db, modelId);\n    db.close();\n\n    setDownloadPhase(\"ready\", modelId);\n    return finalBuffer;\n  } catch (error) {\n    setDownloadPhase(\"error\", modelId);\n    db.close();\n    throw error;\n  }\n}\n\n/**\n * Check if a model has an incomplete download.\n */\nexport async function hasIncompleteDownload(modelId: string): Promise<{\n  incomplete: boolean;\n  bytesDownloaded?: number;\n  totalBytes?: number;\n  percent?: number;\n}> {\n  try {\n    const db = await openDownloadDB();\n    const manifest = await getManifest(db, modelId);\n    db.close();\n\n    if (!manifest) {\n      return { incomplete: false };\n    }\n\n    const totalChunks = Math.ceil(manifest.totalBytes / manifest.chunkSize);\n    const completedChunks = manifest.completedChunks.length;\n\n    if (completedChunks < totalChunks) {\n      return {\n        incomplete: true,\n        bytesDownloaded: completedChunks * manifest.chunkSize,\n        totalBytes: manifest.totalBytes,\n        percent: Math.round((completedChunks / totalChunks) * 100),\n      };\n    }\n\n    return { incomplete: false };\n  } catch {\n    return { incomplete: false };\n  }\n}\n\n/**\n * Clear incomplete download data for a model.\n */\nexport async function clearIncompleteDownload(modelId: string): Promise<void> {\n  try {\n    const db = await openDownloadDB();\n    await clearModelData(db, modelId);\n    db.close();\n  } catch {\n    // Ignore errors\n  }\n}\n\n/**\n * Check if there's enough storage quota for a model download.\n * Returns estimated available space and whether download should proceed.\n */\nexport async function checkStorageQuota(requiredMB: number = 500): Promise<{\n  ok: boolean;\n  availableMB: number;\n  usedMB: number;\n  quotaMB: number;\n  message?: string;\n}> {\n  if (typeof navigator === \"undefined\" || !navigator.storage?.estimate) {\n    return {\n      ok: true,\n      availableMB: -1,\n      usedMB: -1,\n      quotaMB: -1,\n      message: \"Storage API not available\",\n    };\n  }\n\n  try {\n    const { quota, usage } = await navigator.storage.estimate();\n    const quotaMB = Math.round((quota || 0) / 1_000_000);\n    const usedMB = Math.round((usage || 0) / 1_000_000);\n    const availableMB = quotaMB - usedMB;\n\n    if (availableMB < requiredMB) {\n      return {\n        ok: false,\n        availableMB,\n        usedMB,\n        quotaMB,\n        message: `Need ${requiredMB}MB but only ${availableMB}MB available. Clear browser data or free up space.`,\n      };\n    }\n\n    return { ok: true, availableMB, usedMB, quotaMB };\n  } catch (e: any) {\n    return {\n      ok: true, // Proceed optimistically if we can't check\n      availableMB: -1,\n      usedMB: -1,\n      quotaMB: -1,\n      message: `Storage check failed: ${e.message}`,\n    };\n  }\n}\n"],"mappings":";AA4BA,MAAaA,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;;;;;;;;;;;;;;;;;;;;;;;;;;ACxCD,SAAgB,eAAwB;AACtC,KAAI,OAAO,WAAW,YAAa,QAAO;CAE1C,MAAM,gBAAiB,UAAuC,eAAe;CAC7E,MAAM,oBACJ,OAAO,OAAO,eAAe,cAC7B,OAAO,WAAW,6BAA6B,CAAC;AAClD,QAAO,iBAAiB;;;;;AAM1B,SAAgB,QAAiB;AAC/B,KAAI,OAAO,cAAc,YAAa,QAAO;CAC7C,MAAM,KAAK,UAAU,aAAa;AAClC,KAAI,mBAAmB,KAAK,GAAG,CAAE,QAAO;AAGxC,QADoB,YAAY,KAAK,GAAG,IAAI,cAAc,KAAK,GAAG,IAAI,CAAC,SAAS,KAAK,GAAG,KAChE,UAA0C,kBAAkB,KAAK;;;;;AAsB3F,eAAsB,mBAA2C;CAC/D,MAAM,YAAY,cAAc;CAChC,MAAM,MAAM,OAAO;CACnB,IAAI,UAAU;CACd,IAAI,UAAU;CACd,IAAI,YAAY;AAChB,KAAI;EACF,MAAM,MAAM,MAAM,UAAU,SAAS,YAAY;AACjD,YAAU,KAAK,OAAO,KAAK,SAAS,KAAK,IAAI;AAC7C,YAAU,KAAK,OAAO,KAAK,SAAS,KAAK,IAAI;SACvC;AAGR,KAAI;AACF,cAAa,MAAM,UAAU,SAAS,aAAa,IAAK;SAClD;AAGR,QAAO;EACL;EACA;EACA,aAAa,KAAK,IAAI,GAAG,UAAU,QAAQ;EAC3C;EACA;EACA;EACD;;;;;;;;AASH,eAAsB,2BAA6C;AACjE,KAAI;AACF,MAAI,MAAM,UAAU,SAAS,aAAa,CAAE,QAAO;AACnD,SAAQ,MAAM,UAAU,SAAS,WAAW,IAAK;SAC3C;AACN,SAAO;;;;;;;;AAkBX,eAAsB,cAAc,QAAmC;CACrE,MAAM,IAAI,MAAM,kBAAkB;CAGlC,MAAM,OAAO,EAAE,eAAe,SAAS;CACvC,MAAM,oBAAoB,CAAC,QAAS,EAAE,OAAO,CAAC,EAAE,cAAe,CAAC,EAAE;AAClE,QAAO;EAAE;EAAM,aAAa,EAAE;EAAa;EAAkB;;;;;;;;AAS/D,SAAgB,qBAA6E;AAE3F,KADkB,cAAc,CACjB,QAAO;EAAE,WAAW;EAAM,QAAQ;EAAO,OAAO;EAAsB;AACrF,KAAI,OAAO,CACT,QAAO;EACL,WAAW;EACX,QAAQ;EACR,OACE;EACH;AAEH,QAAO;EACL,WAAW;EACX,QAAQ;EACR,OACE;EACH;;;;;;;;;;;;;;;;;;;ACtIH,eAAsB,UACpB,OACA,aAAqB,MACkC;CACvD,MAAM,eAAe,IAAI,cAAc;AAGvC,KAAI,aAAa,UAAU,YACzB,OAAM,aAAa,QAAQ;CAG7B,MAAM,cAAc,aAAa,aAAa,GAAG,MAAM,QAAQ,WAAW;CAC1E,MAAM,cAAc,IAAI,aAAa,MAAM;AAC3C,aAAY,cAAc,aAAa,EAAE;CAEzC,MAAM,aAAa,aAAa,oBAAoB;AACpD,YAAW,SAAS;AACpB,YAAW,QAAQ,aAAa,YAAY;CAE5C,MAAM,UAAU,IAAI,SAAe,YAAY;AAC7C,aAAW,gBAAgB;AACzB,gBAAa,OAAO;AACpB,YAAS;;GAEX;AAEF,YAAW,OAAO;AAElB,QAAO;EACL,YAAY;AACV,cAAW,MAAM;AACjB,gBAAa,OAAO;;EAEtB;EACD;;;;;;;;;;;;;;;;;;;AAoBH,SAAgB,kBAAkB,aAAqB,MAIrD;CACA,IAAIC,eAAoC;CACxC,IAAI,gBAAgB;CACpB,IAAI,WAAW;CAEf,MAAM,gBAAgB,YAAY;AAChC,MAAI,CAAC,aACH,gBAAe,IAAI,cAAc;AAEnC,MAAI,aAAa,UAAU,YACzB,OAAM,aAAa,QAAQ;AAE7B,SAAO;;AAGT,QAAO;EACL,OAAO,OAAO,UAAwB;GACpC,MAAM,MAAM,MAAM,eAAe;AACjC,cAAW;GAEX,MAAM,SAAS,IAAI,aAAa,GAAG,MAAM,QAAQ,WAAW;GAC5D,MAAM,cAAc,IAAI,aAAa,MAAM;AAC3C,UAAO,cAAc,aAAa,EAAE;GAEpC,MAAM,SAAS,IAAI,oBAAoB;AACvC,UAAO,SAAS;AAChB,UAAO,QAAQ,IAAI,YAAY;GAG/B,MAAM,YAAY,KAAK,IAAI,IAAI,aAAa,cAAc;AAC1D,UAAO,MAAM,UAAU;AACvB,mBAAgB,YAAY,OAAO;AAEnC,UAAO,gBAAgB;AACrB,QAAI,IAAI,eAAe,gBAAgB,GACrC,YAAW;;;EAKjB,YAAY;AACV,cAAW;AACX,mBAAgB;AAChB,OAAI,cAAc;AAChB,iBAAa,OAAO;AACpB,mBAAe;;;EAInB,iBAAiB;EAClB;;;;;;AC/GH,MAAM,mBAAmB;;;;;AAMzB,MAAaC,cAAsC;CAEjD,gBAAgB;CAChB,cAAc;CACd,eAAe;CACf,eAAe;CAEf,cAAc;CACd,kBAAkB;CAElB,gBAAgB;CAChB,mBAAmB;CACnB,iBAAiB;CAEjB,oBAAoB;CACrB;;;;;AAMD,SAAS,YAAY,SAAyB;AAC5C,QAAO,QAAQ,aAAa,CAAC,QAAQ,cAAc,IAAI;;;;;;;;;;;;AAazD,MAAM,mBAAmB;CAEvB,SAAS,CAAC,eAAe,cAAc;CAEvC,eAAe;EAAC;EAAU;EAAQ;EAAO;EAAO;EAAc;EAAa;CAE3E,OAAO,CAAC,cAAc,aAAa;CAEnC,aAAa;CACd;;;;;;AAOD,SAAgB,qBAAqB,SAKnC;CACA,MAAM,KAAK,OAAO,cAAc,cAAc,UAAU,YAAY;CACpE,MAAM,WAAW,cAAc,KAAK,GAAG;CACvC,MAAM,SAAS,OAAO,KAAK,GAAG;CAE9B,MAAM,eADQ,YAAY,WACG,QAAQ,KAAK,GAAG;CAC7C,MAAM,eAAe,YAAY,QAAQ;CAEzC,MAAM,WAAW,iBAAiB,cAAc,MAAM,MACpD,aAAa,SAAS,YAAY,EAAE,CAAC,CACtC;CACD,MAAM,YACJ,iBAAiB,QAAQ,MAAM,MAAM,aAAa,SAAS,YAAY,EAAE,CAAC,CAAC,IAAI;CACjF,MAAM,UAAU,iBAAiB,MAAM,MAAM,MAAM,aAAa,SAAS,YAAY,EAAE,CAAC,CAAC;AAIzF,KAAI,UAAU;AACZ,MAAI,UAKF,QAAO;GACL,MAAM;GACN,QAAQ,SAAS,QAAQ,uBANP,cAAc,wDAAwD,GAM5B,IALlD,WACR,+DACA,2BAGkE;GACpE,gBAAgB,OAAO,iBAAiB;GACxC,cAAc;GACf;AAEH,MAAI,QACF,QAAO;GACL,MAAM;GACN,QAAQ,SAAS,QAAQ;GACzB,gBAAgB,+BAA+B,iBAAiB;GAChE,cAAc;GACf;AAEH,SAAO;GAAE,MAAM;GAAM,QAAQ;GAAyC;;AAGxE,KAAI,QAAQ;AAEV,MAAI,UAIF,QAAO;GACL,MAAM;GACN,QAAQ,SAAS,QAAQ,sBALf,WACR,+DACA,2BAGiD;GACnD,gBAAgB,OAAO,iBAAiB;GACxC,cAAc;GACf;AAEH,SAAO;GAAE,MAAM;GAAM,QAAQ;GAAuC;;AAKtE,KADkB,UAAU,KAAK,GAAG,IACnB,UACf,QAAO;EACL,MAAM;EACN,QAAQ,SAAS,QAAQ;EACzB,gBAAgB,OAAO,iBAAiB;EACxC,cAAc;EACf;AAIH,QAAO;EAAE,MAAM;EAAM,QAAQ;EAA0C;;;;;;AAOzE,SAAgB,uBAQd;CACA,MAAM,KAAK,OAAO,cAAc,cAAc,UAAU,YAAY;CACpE,MAAM,eAAe,OAAO,cAAc,cAAe,UAAkB,eAAe;CAC1F,MAAM,WAAW,kCAAkC,KAAK,GAAG;CAK3D,MAAM,eADkB,eAAgB,WAAW,eAAe,KAAM,eAAe,KAAO,KACxD;CAEtC,IAAIC;CACJ,IAAIC;AAEJ,KAAI,cAAc,KAAK;AACrB,SAAO;AACP,WAAS;YACA,YAAY,cAAc,MAAM;AACzC,SAAO;AACP,WAAS;YACA,cAAc,MAAM;AAC7B,SAAO;AACP,WAAS;QACJ;AACL,SAAO;AACP,WAAS;;AAGX,QAAO;EACL;EACA,KAAK;EACL,KAAK;EACL,WAAW;EACX;EACA;EACA;EACD;;AASH,MAAa,sBAAsB;;;;AAcnC,SAAS,oBAA4B;AACnC,QAAO,GAAG,KAAK,KAAK,CAAC,GAAG,KAAK,QAAQ,CAAC,SAAS,GAAG,CAAC,MAAM,GAAG,EAAE;;;;;AAMhE,SAAS,eAAuB;AAC9B,KAAI,OAAO,iBAAiB,YAAa,QAAO,mBAAmB;CAEnE,IAAI,YAAY,eAAe,QAAQ,oBAAoB;AAC3D,KAAI,CAAC,WAAW;AACd,cAAY,mBAAmB;AAC/B,iBAAe,QAAQ,qBAAqB,UAAU;;AAExD,QAAO;;;;;;AAOT,SAAgB,iBACd,OACA,SACA,UACM;AACN,KAAI,OAAO,iBAAiB,YAAa;CAEzC,MAAMC,QAAsB;EAC1B;EACA,SAAS,WAAW;EACpB,WAAW,cAAc;EACzB,WAAW,KAAK,KAAK;EACrB,iBAAiB,UAAU;EAC3B,YAAY,UAAU;EACvB;AAED,cAAa,QAAQ,qBAAqB,KAAK,UAAU,MAAM,CAAC;;;;;AAMlE,SAAgB,mBAAwC;AACtD,KAAI,OAAO,iBAAiB,YAAa,QAAO;AAEhD,KAAI;EACF,MAAM,MAAM,aAAa,QAAQ,oBAAoB;AACrD,MAAI,CAAC,IAAK,QAAO;AACjB,SAAO,KAAK,MAAM,IAAI;SAChB;AACN,SAAO;;;;;;;;;AAUX,SAAgB,oBAMd;CACA,MAAM,YAAY,kBAAkB;CACpC,MAAM,mBAAmB,cAAc;AAEvC,KAAI,CAAC,UACH,QAAO,EAAE,SAAS,OAAO;CAK3B,MAAM,qBADkC;EAAC;EAAe;EAAW;EAAe,CACxC,SAAS,UAAU,MAAM;CACnE,MAAM,iBAAiB,UAAU,cAAc;CAC/C,MAAM,iBAAiB,KAAK,KAAK,GAAG,UAAU;AAK9C,KAAI,sBAAsB,kBAFN,iBAAiB,MAAS,KAEW;AAEvD,eAAa,WAAW,oBAAoB;AAE5C,SAAO;GACL,SAAS;GACT,OAAO,UAAU;GACjB,SAAS,UAAU,WAAW;GAC9B;GACA,gBACE,UAAU,WAAW,mBAAmB,KAAK,UAAU,QAAQ,GAC3D,uEACA;GACP;;AAGH,QAAO,EAAE,SAAS,OAAO;;;;;AAM3B,SAAgB,qBAA2B;AACzC,KAAI,OAAO,iBAAiB,YAAa;AACzC,cAAa,WAAW,oBAAoB;;;;;;ACjU9C,MAAa,mBAAmB,MAAM,OAAO;;AAG7C,MAAa,mBAAmB;AAChC,MAAM,sBAAsB;;;;AAmB5B,eAAe,iBAAuC;AACpD,QAAO,IAAI,SAAS,SAAS,WAAW;EACtC,MAAM,UAAU,UAAU,KAAK,kBAAkB,oBAAoB;AAErE,UAAQ,gBACN,uBAAO,IAAI,MAAM,+BAA+B,QAAQ,OAAO,UAAU,CAAC;AAE5E,UAAQ,kBAAkB,QAAQ,QAAQ,OAAO;AAEjD,UAAQ,mBAAmB,UAAU;GACnC,MAAM,KAAM,MAAM,OAA4B;AAG9C,OAAI,CAAC,GAAG,iBAAiB,SAAS,YAAY,CAC5C,IAAG,kBAAkB,aAAa,EAAE,SAAS,WAAW,CAAC;AAI3D,OAAI,CAAC,GAAG,iBAAiB,SAAS,SAAS,CACzC,IAAG,kBAAkB,SAAS;;GAGlC;;;;;AAMJ,eAAe,YAAY,IAAiB,SAAmD;AAC7F,QAAO,IAAI,SAAS,SAAS,WAAW;EAGtC,MAAM,UAFK,GAAG,YAAY,aAAa,WAAW,CACjC,YAAY,YAAY,CACnB,IAAI,QAAQ;AAElC,UAAQ,gBAAgB,uBAAO,IAAI,MAAM,2BAA2B,QAAQ,OAAO,UAAU,CAAC;AAC9F,UAAQ,kBAAkB,QAAQ,QAAQ,UAAU,KAAK;GACzD;;;;;AAMJ,eAAe,aAAa,IAAiB,UAA2C;AACtF,QAAO,IAAI,SAAS,SAAS,WAAW;EAGtC,MAAM,UAFK,GAAG,YAAY,aAAa,YAAY,CAClC,YAAY,YAAY,CACnB,IAAI,SAAS;AAEnC,UAAQ,gBAAgB,uBAAO,IAAI,MAAM,4BAA4B,QAAQ,OAAO,UAAU,CAAC;AAC/F,UAAQ,kBAAkB,SAAS;GACnC;;;;;AAMJ,eAAe,UACb,IACA,SACA,YACA,MACe;AACf,QAAO,IAAI,SAAS,SAAS,WAAW;EAEtC,MAAM,QADK,GAAG,YAAY,UAAU,YAAY,CAC/B,YAAY,SAAS;EACtC,MAAM,MAAM,GAAG,QAAQ,GAAG;EAC1B,MAAM,UAAU,MAAM,IAAI,MAAM,IAAI;AAEpC,UAAQ,gBACN,uBAAO,IAAI,MAAM,wBAAwB,WAAW,IAAI,QAAQ,OAAO,UAAU,CAAC;AACpF,UAAQ,kBAAkB,SAAS;GACnC;;;;;AAMJ,eAAe,SACb,IACA,SACA,YAC6B;AAC7B,QAAO,IAAI,SAAS,SAAS,WAAW;EAEtC,MAAM,QADK,GAAG,YAAY,UAAU,WAAW,CAC9B,YAAY,SAAS;EACtC,MAAM,MAAM,GAAG,QAAQ,GAAG;EAC1B,MAAM,UAAU,MAAM,IAAI,IAAI;AAE9B,UAAQ,gBACN,uBAAO,IAAI,MAAM,uBAAuB,WAAW,IAAI,QAAQ,OAAO,UAAU,CAAC;AACnF,UAAQ,kBAAkB,QAAQ,QAAQ,UAAU,KAAK;GACzD;;;;;AAMJ,eAAe,eAAe,IAAiB,SAAgC;CAE7E,MAAM,WAAW,MAAM,YAAY,IAAI,QAAQ;AAE/C,QAAO,IAAI,SAAS,SAAS,WAAW;EACtC,MAAM,KAAK,GAAG,YAAY,CAAC,aAAa,SAAS,EAAE,YAAY;AAG/D,KAAG,YAAY,YAAY,CAAC,OAAO,QAAQ;AAG3C,MAAI,UAAU;GACZ,MAAM,cAAc,KAAK,KAAK,SAAS,aAAa,SAAS,UAAU;GACvE,MAAM,aAAa,GAAG,YAAY,SAAS;AAC3C,QAAK,IAAI,IAAI,GAAG,IAAI,aAAa,IAC/B,YAAW,OAAO,GAAG,QAAQ,GAAG,IAAI;;AAIxC,KAAG,mBAAmB,SAAS;AAC/B,KAAG,gBAAgB,uBAAO,IAAI,MAAM,+BAA+B,GAAG,OAAO,UAAU,CAAC;GACxF;;;;;;AAOJ,eAAsB,qBACpB,KACA,SACA,UAQI,EAAE,EACgB;CACtB,MAAM,EAAE,YAAY,WAAW;AAG/B,kBAAiB,eAAe,QAAQ;CAExC,MAAM,KAAK,MAAM,gBAAgB;AAEjC,KAAI;EAEF,IAAI,WAAW,MAAM,YAAY,IAAI,QAAQ;EAG7C,MAAM,eAAe,MAAM,MAAM,KAAK;GAAE,QAAQ;GAAQ;GAAQ,CAAC;AACjE,MAAI,CAAC,aAAa,GAChB,OAAM,IAAI,MAAM,wBAAwB,aAAa,OAAO,GAAG,aAAa,aAAa;EAG3F,MAAM,gBAAgB,SAAS,aAAa,QAAQ,IAAI,iBAAiB,IAAI,KAAK,GAAG;EACrF,MAAM,OAAO,aAAa,QAAQ,IAAI,OAAO;EAC7C,MAAM,eAAe,aAAa,QAAQ,IAAI,gBAAgB;AAE9D,MAAI,CAAC,cACH,OAAM,IAAI,MAAM,wCAAwC;AAI1D,MAAI,YAAY,SAAS,SAAS,MAAM;AACtC,WAAQ,KAAK,SAAS,QAAQ,4DAA4D;AAC1F,SAAM,eAAe,IAAI,QAAQ;AACjC,cAAW;;AAMb,MAAI,EAFkB,iBAAiB,UAEnB;AAElB,WAAQ,KAAK,6CAA6C,QAAQ,2BAA2B;AAC7F,MAAG,OAAO;GAEV,MAAM,WAAW,MAAM,MAAM,KAAK,EAAE,QAAQ,CAAC;AAC7C,OAAI,CAAC,SAAS,GAAI,OAAM,IAAI,MAAM,oBAAoB,SAAS,SAAS;AAExE,oBAAiB,WAAW,QAAQ;GACpC,MAAM,SAAS,MAAM,SAAS,aAAa;AAC3C,oBAAiB,SAAS,QAAQ;AAClC,UAAO;;EAIT,MAAM,cAAc,KAAK,KAAK,gBAAgB,iBAAiB;AAE/D,MAAI,CAAC,UAAU;AACb,cAAW;IACT;IACA;IACA;IACA,YAAY;IACZ,WAAW;IACX,iBAAiB,EAAE;IACnB,WAAW,KAAK,KAAK;IACrB,WAAW,KAAK,KAAK;IACtB;AACD,SAAM,aAAa,IAAI,SAAS;;AAIlC,OAAK,IAAI,IAAI,GAAG,IAAI,aAAa,KAAK;AACpC,OAAI,QAAQ,QACV,OAAM,IAAI,MAAM,mBAAmB;AAIrC,OAAI,SAAS,gBAAgB,SAAS,EAAE,EAAE;IACxC,MAAMC,oBAAmB,SAAS,gBAAgB,SAAS,cAAe;AAC1E,iBAAa;KACX,OAAO;KACP;KACA,YAAY;KACZ,SAAS,KAAK,MAAOA,oBAAkB,gBAAiB,IAAI;KAC7D,CAAC;AACF;;GAGF,MAAM,QAAQ,IAAI;GAClB,MAAM,MAAM,KAAK,IAAI,QAAQ,mBAAmB,GAAG,gBAAgB,EAAE;GAGrE,MAAM,WAAW,MAAM,MAAM,KAAK;IAChC,SAAS,EAAE,OAAO,SAAS,MAAM,GAAG,OAAO;IAC3C;IACD,CAAC;AAEF,OAAI,SAAS,WAAW,IACtB,OAAM,IAAI,MAAM,yBAAyB,SAAS,OAAO,iBAAiB;GAG5E,MAAM,YAAY,MAAM,SAAS,aAAa;AAG9C,SAAM,UAAU,IAAI,SAAS,GAAG,UAAU;AAG1C,YAAS,gBAAgB,KAAK,EAAE;AAChC,YAAS,YAAY,KAAK,KAAK;AAC/B,SAAM,aAAa,IAAI,SAAS;GAGhC,MAAM,kBAAkB,SAAS,gBAAgB,SAAS;AAC1D,oBAAiB,eAAe,SAAS;IAAE;IAAiB,YAAY;IAAe,CAAC;AAExF,gBAAa;IACX,OAAO;IACP,iBAAiB,KAAK,IAAI,iBAAiB,cAAc;IACzD,YAAY;IACZ,SAAS,KAAK,MAAO,SAAS,gBAAgB,SAAS,cAAe,IAAI;IAC3E,CAAC;AAIF,YAAS,OAAO;;AAIlB,mBAAiB,WAAW,QAAQ;AACpC,eAAa;GACX,OAAO;GACP,iBAAiB;GACjB,YAAY;GACZ,SAAS;GACV,CAAC;EAIF,MAAM,cAAc,IAAI,YAAY,cAAc;EAClD,MAAM,YAAY,IAAI,WAAW,YAAY;AAE7C,OAAK,IAAI,IAAI,GAAG,IAAI,aAAa,KAAK;GACpC,MAAM,QAAQ,MAAM,SAAS,IAAI,SAAS,EAAE;AAC5C,OAAI,CAAC,MACH,OAAM,IAAI,MAAM,iBAAiB,EAAE,kBAAkB;GAGvD,MAAM,SAAS,IAAI;AACnB,aAAU,IAAI,IAAI,WAAW,MAAM,EAAE,OAAO;;AAO9C,QAAM,eAAe,IAAI,QAAQ;AACjC,KAAG,OAAO;AAEV,mBAAiB,SAAS,QAAQ;AAClC,SAAO;UACA,OAAO;AACd,mBAAiB,SAAS,QAAQ;AAClC,KAAG,OAAO;AACV,QAAM;;;;;;AAOV,eAAsB,sBAAsB,SAKzC;AACD,KAAI;EACF,MAAM,KAAK,MAAM,gBAAgB;EACjC,MAAM,WAAW,MAAM,YAAY,IAAI,QAAQ;AAC/C,KAAG,OAAO;AAEV,MAAI,CAAC,SACH,QAAO,EAAE,YAAY,OAAO;EAG9B,MAAM,cAAc,KAAK,KAAK,SAAS,aAAa,SAAS,UAAU;EACvE,MAAM,kBAAkB,SAAS,gBAAgB;AAEjD,MAAI,kBAAkB,YACpB,QAAO;GACL,YAAY;GACZ,iBAAiB,kBAAkB,SAAS;GAC5C,YAAY,SAAS;GACrB,SAAS,KAAK,MAAO,kBAAkB,cAAe,IAAI;GAC3D;AAGH,SAAO,EAAE,YAAY,OAAO;SACtB;AACN,SAAO,EAAE,YAAY,OAAO;;;;;;AAOhC,eAAsB,wBAAwB,SAAgC;AAC5E,KAAI;EACF,MAAM,KAAK,MAAM,gBAAgB;AACjC,QAAM,eAAe,IAAI,QAAQ;AACjC,KAAG,OAAO;SACJ;;;;;;AASV,eAAsB,kBAAkB,aAAqB,KAM1D;AACD,KAAI,OAAO,cAAc,eAAe,CAAC,UAAU,SAAS,SAC1D,QAAO;EACL,IAAI;EACJ,aAAa;EACb,QAAQ;EACR,SAAS;EACT,SAAS;EACV;AAGH,KAAI;EACF,MAAM,EAAE,OAAO,UAAU,MAAM,UAAU,QAAQ,UAAU;EAC3D,MAAM,UAAU,KAAK,OAAO,SAAS,KAAK,IAAU;EACpD,MAAM,SAAS,KAAK,OAAO,SAAS,KAAK,IAAU;EACnD,MAAM,cAAc,UAAU;AAE9B,MAAI,cAAc,WAChB,QAAO;GACL,IAAI;GACJ;GACA;GACA;GACA,SAAS,QAAQ,WAAW,cAAc,YAAY;GACvD;AAGH,SAAO;GAAE,IAAI;GAAM;GAAa;GAAQ;GAAS;UAC1CC,GAAQ;AACf,SAAO;GACL,IAAI;GACJ,aAAa;GACb,QAAQ;GACR,SAAS;GACT,SAAS,yBAAyB,EAAE;GACrC"}
+{"version":3,"file":"index.js","names":["BUILTIN_MODELS: Record<string, ModelConfig>","audioContext: AudioContext | null","MODEL_SIZES: Record<string, number>","chat: string","reason: string","state: SessionState","bytesDownloaded","e: any"],"sources":["../../src/core/models.ts","../../src/browser/pwa.ts","../../src/browser/audio.ts","../../src/browser/device-guards.ts","../../src/browser/download.ts"],"sourcesContent":["/**\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 * Mobile / PWA storage helpers.\n *\n * On-device models are large (a 4-bit 0.8B is ~400 MB; vision/larger models are\n * GBs). Mobile browsers — iOS Safari especially — wall a web origin off from the\n * real disk with TWO independent ceilings:\n *\n *   1. **Storage quota** (disk for the model cache). An *uninstalled* Safari tab\n *      gets only ~1 GB, best-effort and evictable, regardless of how much free\n *      disk the device has. Exceed it and every cache write fails → the model\n *      re-downloads on every visit.\n *   2. **Tab memory** (RAM during load/inference) — a separate, smaller ceiling.\n *\n * The unlock for the storage ceiling is **persistent storage**, which iOS Safari\n * grants when the site is **installed to the Home Screen** (a PWA). Installed, the\n * quota jumps to a large fraction of actual disk and is never evicted — so models\n * cache once and stay. These helpers let an app surface that to its users and\n * request it, so on-device AI is actually practical on mobile.\n *\n * All functions are SSR/Node-safe (guarded; return conservative defaults).\n */\n\n/** True when the page is running as an installed/standalone PWA (Home Screen). */\nexport function isStandalone(): boolean {\n  if (typeof window === \"undefined\") return false;\n  // iOS Safari exposes navigator.standalone; everyone else uses display-mode.\n  const iosStandalone = (navigator as { standalone?: boolean }).standalone === true;\n  const displayStandalone =\n    typeof window.matchMedia === \"function\" &&\n    window.matchMedia(\"(display-mode: standalone)\").matches;\n  return iosStandalone || displayStandalone;\n}\n\n/** True when running on iOS/iPadOS (where install is the quota unlock and the\n * install flow is manual: Share → Add to Home Screen). iPadOS masquerades as\n * macOS, so we also treat touch-capable WebKit-on-Mac as iOS. */\nexport function isIOS(): boolean {\n  if (typeof navigator === \"undefined\") return false;\n  const ua = navigator.userAgent || \"\";\n  if (/iPhone|iPad|iPod/.test(ua)) return true;\n  // iPadOS 13+ reports a Mac UA — detect via touch points + WebKit.\n  const isMacWebKit = /Macintosh/.test(ua) && /AppleWebKit/.test(ua) && !/Chrome/.test(ua);\n  return isMacWebKit && ((navigator as { maxTouchPoints?: number }).maxTouchPoints ?? 0) > 1;\n}\n\nexport type StorageStatus = {\n  /** Total quota granted to this origin, in MB (best-effort estimate). */\n  quotaMB: number;\n  /** Bytes currently used by this origin, in MB. */\n  usageMB: number;\n  /** quota − usage, in MB. */\n  availableMB: number;\n  /** Storage is persistent (exempt from eviction). On iOS this is effectively\n   * only true once the site is installed to the Home Screen. */\n  persisted: boolean;\n  /** Running as an installed/standalone PWA. */\n  installed: boolean;\n  /** Platform is iOS/iPadOS (install is the quota unlock here). */\n  ios: boolean;\n};\n\n/** Snapshot of the origin's storage situation — quota, usage, persistence, and\n * whether the app is installed. Use it to decide whether to recommend install\n * before downloading a large model. */\nexport async function getStorageStatus(): Promise<StorageStatus> {\n  const installed = isStandalone();\n  const ios = isIOS();\n  let quotaMB = 0;\n  let usageMB = 0;\n  let persisted = false;\n  try {\n    const est = await navigator.storage?.estimate?.();\n    quotaMB = Math.round((est?.quota || 0) / 1e6);\n    usageMB = Math.round((est?.usage || 0) / 1e6);\n  } catch {\n    /* estimate unsupported */\n  }\n  try {\n    persisted = (await navigator.storage?.persisted?.()) ?? false;\n  } catch {\n    /* persisted unsupported */\n  }\n  return {\n    quotaMB,\n    usageMB,\n    availableMB: Math.max(0, quotaMB - usageMB),\n    persisted,\n    installed,\n    ios,\n  };\n}\n\n/**\n * Request persistent storage (exempt from eviction). Returns whether the origin\n * is persistent afterwards. Browsers grant this based on engagement/installation;\n * on iOS Safari it is effectively granted only to an installed (Home Screen) PWA,\n * so call this AND guide users to install when it returns false on iOS.\n */\nexport async function requestPersistentStorage(): Promise<boolean> {\n  try {\n    if (await navigator.storage?.persisted?.()) return true;\n    return (await navigator.storage?.persist?.()) ?? false;\n  } catch {\n    return false;\n  }\n}\n\nexport type ModelFit = {\n  /** The model likely fits in the currently-available quota. */\n  fits: boolean;\n  availableMB: number;\n  /** Caching durably would benefit from installing to the Home Screen — true when\n   * not installed on iOS, or when the model doesn't fit the current quota. */\n  recommendInstall: boolean;\n};\n\n/**\n * Estimate whether a model of `sizeMB` will cache in the current quota, and\n * whether you should recommend installing to the Home Screen first. Pair with a\n * one-time \"Install for offline use\" prompt before a large download on mobile.\n */\nexport async function canCacheModel(sizeMB: number): Promise<ModelFit> {\n  const s = await getStorageStatus();\n  // Headroom: leave ~10% slack so we don't recommend a download that just barely\n  // fits and then fails mid-write.\n  const fits = s.availableMB >= sizeMB * 1.1;\n  const recommendInstall = (!fits || (s.ios && !s.installed)) && !s.persisted;\n  return { fits, availableMB: s.availableMB, recommendInstall };\n}\n\n/**\n * Platform-appropriate install guidance. iOS Safari has NO programmatic install\n * prompt — installation is manual (Share → Add to Home Screen), so apps should\n * show these instructions. Other platforms (Android/Chrome) fire\n * `beforeinstallprompt`, which apps can capture for a one-tap button.\n */\nexport function getInstallGuidance(): { installed: boolean; manual: boolean; steps: string } {\n  const installed = isStandalone();\n  if (installed) return { installed: true, manual: false, steps: \"Already installed.\" };\n  if (isIOS()) {\n    return {\n      installed: false,\n      manual: true,\n      steps:\n        \"Tap the Share button, then 'Add to Home Screen'. Installing unlocks durable storage so models download once instead of every visit.\",\n    };\n  }\n  return {\n    installed: false,\n    manual: false,\n    steps:\n      \"Use your browser's Install option (or the install icon in the address bar) to add this app for offline use and durable model storage.\",\n  };\n}\n","// ============================================\n// Audio Playback Utilities\n// ============================================\n\n/**\n * Play audio from Float32Array using Web Audio API\n *\n * @example\n * ```ts\n * import { playAudio } from \"@tryhamster/gerbil/browser\";\n *\n * const audio = new Float32Array([...]); // TTS output\n * const controller = await playAudio(audio, 24000);\n *\n * // Stop playback\n * controller.stop();\n * ```\n */\nexport async function playAudio(\n  audio: Float32Array,\n  sampleRate: number = 24000,\n): Promise<{ stop: () => void; onEnded: Promise<void> }> {\n  const audioContext = new AudioContext();\n\n  // Resume if suspended\n  if (audioContext.state === \"suspended\") {\n    await audioContext.resume();\n  }\n\n  const audioBuffer = audioContext.createBuffer(1, audio.length, sampleRate);\n  const channelData = new Float32Array(audio);\n  audioBuffer.copyToChannel(channelData, 0);\n\n  const sourceNode = audioContext.createBufferSource();\n  sourceNode.buffer = audioBuffer;\n  sourceNode.connect(audioContext.destination);\n\n  const onEnded = new Promise<void>((resolve) => {\n    sourceNode.onended = () => {\n      audioContext.close();\n      resolve();\n    };\n  });\n\n  sourceNode.start();\n\n  return {\n    stop: () => {\n      sourceNode.stop();\n      audioContext.close();\n    },\n    onEnded,\n  };\n}\n\n/**\n * Create a reusable audio player for streaming TTS\n *\n * @example\n * ```ts\n * import { createAudioPlayer } from \"@tryhamster/gerbil/browser\";\n *\n * const player = createAudioPlayer(24000);\n *\n * // Queue audio chunks as they arrive\n * player.queue(chunk1);\n * player.queue(chunk2);\n *\n * // Stop and clear\n * player.stop();\n * ```\n */\nexport function createAudioPlayer(sampleRate: number = 24000): {\n  queue: (audio: Float32Array) => void;\n  stop: () => void;\n  isPlaying: () => boolean;\n} {\n  let audioContext: AudioContext | null = null;\n  let nextStartTime = 0;\n  let isActive = false;\n\n  const ensureContext = async () => {\n    if (!audioContext) {\n      audioContext = new AudioContext();\n    }\n    if (audioContext.state === \"suspended\") {\n      await audioContext.resume();\n    }\n    return audioContext;\n  };\n\n  return {\n    queue: async (audio: Float32Array) => {\n      const ctx = await ensureContext();\n      isActive = true;\n\n      const buffer = ctx.createBuffer(1, audio.length, sampleRate);\n      const channelData = new Float32Array(audio);\n      buffer.copyToChannel(channelData, 0);\n\n      const source = ctx.createBufferSource();\n      source.buffer = buffer;\n      source.connect(ctx.destination);\n\n      // Schedule seamlessly after previous chunk\n      const startTime = Math.max(ctx.currentTime, nextStartTime);\n      source.start(startTime);\n      nextStartTime = startTime + buffer.duration;\n\n      source.onended = () => {\n        if (ctx.currentTime >= nextStartTime - 0.1) {\n          isActive = false;\n        }\n      };\n    },\n\n    stop: () => {\n      isActive = false;\n      nextStartTime = 0;\n      if (audioContext) {\n        audioContext.close();\n        audioContext = null;\n      }\n    },\n\n    isPlaying: () => isActive,\n  };\n}\n","// ============================================\n// iOS Model Guards & Device Capability Detection\n// ============================================\n\n// ============================================\n// Real native-engine model ids\n// ============================================\n//\n// The website loads models by their actual repo id (MLX 4-bit on-device builds,\n// or the upstream Qwen/Liquid repos on desktop). The guard matches on those ids\n// — NOT the old ONNX-era shorthands — so it must stay in sync with the real\n// checkpoints the engine can run. Matching is substring-based on a normalized\n// id so both `mlx-community/Qwen3.5-0.8B-4bit` and `Qwen/Qwen3.5-0.8B` resolve.\n\n/** Recommended safe model ids per modality (used as fallbacks on mobile). */\nconst SAFE_MOBILE_CHAT = \"mlx-community/Qwen3.5-0.8B-4bit\";\n\n/**\n * Approximate on-device (INT4) memory footprint in MB for the models the native\n * engine actually ships. Used for memory-aware selection and messaging.\n */\nexport const MODEL_SIZES: Record<string, number> = {\n  // Chat models (INT4, on-device)\n  \"qwen3.5-0.8b\": 650, // ~0.65GB\n  \"qwen3.5-2b\": 1700, // ~1.7GB\n  \"gemma-4-e2b\": 3600, // ~3.6GB\n  \"lfm2.5-350m\": 300, // ~0.3GB\n  // TTS models\n  \"kokoro-82m\": 350,\n  \"supertonic-66m\": 300,\n  // STT models\n  \"whisper-tiny\": 150,\n  \"whisper-tiny.en\": 150,\n  \"whisper-small\": 500,\n  // Embedding models\n  \"all-minilm-l6-v2\": 100,\n};\n\n/**\n * Normalize a repo/model id to a lowercase token stream for substring matching\n * (strips org prefixes' punctuation while preserving the model name tokens).\n */\nfunction normalizeId(modelId: string): string {\n  return modelId.toLowerCase().replace(/[^a-z0-9]/g, \"-\");\n}\n\n/**\n * iOS (WKWebView) model classification keyed off the REAL native-engine ids.\n *\n * - blocked: too large for the WKWebView memory ceiling on iPhone — will crash.\n *   gemma-4-e2b (~3.6GB) plus any vision checkpoint (the vision encoder pushes\n *   the working set well past what an iPhone can hold).\n * - risky: Qwen3.5-2B (~1.7GB) OOM-crashes iPhone WKWebView in practice — now\n *   treated as unsafe to auto-load (flagged `risky` so a UI can offer to force it).\n * - everything else (Qwen3.5-0.8B ~0.65GB, LFM2.5-350M) is allowed everywhere.\n */\nconst IOS_MODEL_LIMITS = {\n  /** Substrings (normalized) of ids that HARD-BLOCK on iPhone. */\n  blocked: [\"gemma-4-e2b\", \"gemma-4-e4b\"],\n  /** Substrings that mark a vision checkpoint (blocked on iPhone). */\n  visionMarkers: [\"vision\", \"-vl-\", \"-vl\", \"vlm\", \"image-text\", \"-it-vision\"],\n  /** Substrings of ids that WARN on iPhone (Qwen3.5-2B class). */\n  risky: [\"qwen3.5-2b\", \"qwen3-5-2b\"],\n  /** Maximum total memory budget in MB for iOS WKWebView. */\n  maxBudgetMB: 1800,\n} as const;\n\n/**\n * Check if a model is safe to load on the current device.\n * Returns guidance specific to iOS memory constraints. Matches on the real\n * native-engine repo ids (MLX 4-bit / upstream Qwen / Liquid).\n */\nexport function isModelSafeForDevice(modelId: string): {\n  safe: boolean;\n  /**\n   * Borderline: may run on the newest hardware but is prone to OOM. Reported\n   * unsafe (`safe: false`) so callers block by default; a UI can use `risky` to\n   * offer an explicit \"load anyway\" on capable devices.\n   */\n  risky: boolean;\n  reason: string;\n  recommendation?: string;\n  maxSafeModel?: string;\n} {\n  const ua = typeof navigator !== \"undefined\" ? navigator.userAgent : \"\";\n  const isIPhone = /iPhone|iPod/.test(ua);\n  const isIPad = /iPad/.test(ua);\n  const isIOS = isIPhone || isIPad;\n  const isIOSChrome = isIOS && /CriOS/.test(ua);\n  const normalizedId = normalizeId(modelId);\n\n  const isVision = IOS_MODEL_LIMITS.visionMarkers.some((m) =>\n    normalizedId.includes(normalizeId(m)),\n  );\n  const isBlocked =\n    IOS_MODEL_LIMITS.blocked.some((m) => normalizedId.includes(normalizeId(m))) || isVision;\n  const isRisky = IOS_MODEL_LIMITS.risky.some((m) => normalizedId.includes(normalizeId(m)));\n\n  // iPhone is the hard constraint (smallest WKWebView budget). iPad has more\n  // headroom, so only the genuinely huge / vision checkpoints are blocked there.\n  if (isIPhone) {\n    if (isBlocked) {\n      const browserNote = isIOSChrome ? \" (iOS Chrome uses WKWebView, same limits as Safari)\" : \"\";\n      const why = isVision\n        ? \"Vision checkpoints need a separate image encoder in memory\"\n        : \"It is too large (~3.6GB)\";\n      return {\n        safe: false,\n        risky: false,\n        reason: `Model ${modelId} will crash on iPhone${browserNote}. ${why}, which exceeds the WKWebView memory ceiling.`,\n        recommendation: `Use ${SAFE_MOBILE_CHAT} (Qwen3.5-0.8B) on iPhone, or run larger models on desktop.`,\n        maxSafeModel: SAFE_MOBILE_CHAT,\n      };\n    }\n    if (isRisky) {\n      return {\n        safe: false,\n        risky: true,\n        reason: `Model ${modelId} (~1.7GB) exceeds the iPhone WKWebView memory budget and is likely to crash.`,\n        recommendation: `Use ${SAFE_MOBILE_CHAT} (Qwen3.5-0.8B) on iPhone; run Qwen3.5-2B on iPad or desktop.`,\n        maxSafeModel: SAFE_MOBILE_CHAT,\n      };\n    }\n    return { safe: true, risky: false, reason: \"Model is within iPhone memory limits.\" };\n  }\n\n  if (isIPad) {\n    // iPad tolerates the 2B class; only block the huge / vision checkpoints.\n    if (isBlocked) {\n      const why = isVision\n        ? \"Vision checkpoints need a separate image encoder in memory\"\n        : \"It is too large (~3.6GB)\";\n      return {\n        safe: false,\n        risky: false,\n        reason: `Model ${modelId} may crash on iPad. ${why}, which can exceed the WKWebView memory ceiling.`,\n        recommendation: `Use ${SAFE_MOBILE_CHAT} (Qwen3.5-0.8B) or Qwen3.5-2B on iPad.`,\n        maxSafeModel: \"mlx-community/Qwen3.5-2B-4bit\",\n      };\n    }\n    return { safe: true, risky: false, reason: \"Model is within iPad memory limits.\" };\n  }\n\n  // Android - block the huge / vision checkpoints, allow the rest.\n  const isAndroid = /Android/.test(ua);\n  if (isAndroid && isBlocked) {\n    return {\n      safe: false,\n      risky: false,\n      reason: `Model ${modelId} is very large and may crash on Android devices.`,\n      recommendation: `Use ${SAFE_MOBILE_CHAT} (Qwen3.5-0.8B) or Qwen3.5-2B on Android.`,\n      maxSafeModel: \"mlx-community/Qwen3.5-2B-4bit\",\n    };\n  }\n\n  // Desktop - all models are safe.\n  return { safe: true, risky: false, reason: \"Desktop browser has sufficient memory.\" };\n}\n\n/**\n * Get recommended models based on device memory and capabilities.\n * Helps prevent OOM crashes on low-memory mobile devices.\n */\nexport function getRecommendedModels(): {\n  chat: string;\n  tts: string;\n  stt: string;\n  embedding: string;\n  reason: string;\n  deviceMemory: number | null;\n  isMobile: boolean;\n} {\n  const ua = typeof navigator !== \"undefined\" ? navigator.userAgent : \"\";\n  const deviceMemory = typeof navigator !== \"undefined\" ? (navigator as any).deviceMemory : null;\n  const isMobile = /iPhone|iPad|iPod|Android|Mobile/.test(ua);\n\n  // Estimate available memory (deviceMemory reports total GB, not available)\n  // Mobile devices typically have less free memory due to OS overhead\n  const effectiveMemory = deviceMemory ? (isMobile ? deviceMemory * 0.4 : deviceMemory * 0.6) : 4;\n  const availableMB = effectiveMemory * 1024;\n\n  let chat: string;\n  let reason: string;\n\n  if (availableMB < 600) {\n    chat = \"LiquidAI/LFM2.5-350M\";\n    reason = \"Very low memory device - using smallest model (LFM2.5-350M)\";\n  } else if (isMobile && availableMB < 2200) {\n    chat = SAFE_MOBILE_CHAT;\n    reason = \"Mobile device - using Qwen3.5-0.8B to stay within the WKWebView memory limit\";\n  } else if (availableMB < 2200) {\n    chat = SAFE_MOBILE_CHAT;\n    reason = \"Standard model for moderate memory (Qwen3.5-0.8B)\";\n  } else {\n    chat = \"mlx-community/Qwen3.5-2B-4bit\";\n    reason = \"High memory available - using Qwen3.5-2B for better quality\";\n  }\n\n  return {\n    chat,\n    tts: \"kokoro-82m\",\n    stt: \"whisper-tiny.en\",\n    embedding: \"all-MiniLM-L6-v2\",\n    reason,\n    deviceMemory,\n    isMobile,\n  };\n}\n\n// ============================================\n// Session Phase Tracking (Reload Detection)\n// ============================================\n\ntype DownloadPhase = \"idle\" | \"downloading\" | \"caching\" | \"initializing\" | \"ready\" | \"error\";\n\nexport const SESSION_STORAGE_KEY = \"gerbil_session_phase\";\n\nexport type SessionState = {\n  phase: DownloadPhase;\n  modelId: string | null;\n  sessionId: string;\n  timestamp: number;\n  bytesDownloaded?: number;\n  totalBytes?: number;\n};\n\n/**\n * Generate a unique session ID for tracking across reloads.\n */\nfunction generateSessionId(): string {\n  return `${Date.now()}-${Math.random().toString(36).slice(2, 9)}`;\n}\n\n/**\n * Get or create the current session ID.\n */\nfunction getSessionId(): string {\n  if (typeof localStorage === \"undefined\") return generateSessionId();\n\n  let sessionId = sessionStorage.getItem(\"gerbil_session_id\");\n  if (!sessionId) {\n    sessionId = generateSessionId();\n    sessionStorage.setItem(\"gerbil_session_id\", sessionId);\n  }\n  return sessionId;\n}\n\n/**\n * Set the current download/initialization phase.\n * Used to detect if a reload happened during a critical operation.\n */\nexport function setDownloadPhase(\n  phase: DownloadPhase,\n  modelId?: string,\n  progress?: { bytesDownloaded: number; totalBytes: number },\n): void {\n  if (typeof localStorage === \"undefined\") return;\n\n  const state: SessionState = {\n    phase,\n    modelId: modelId || null,\n    sessionId: getSessionId(),\n    timestamp: Date.now(),\n    bytesDownloaded: progress?.bytesDownloaded,\n    totalBytes: progress?.totalBytes,\n  };\n\n  localStorage.setItem(SESSION_STORAGE_KEY, JSON.stringify(state));\n}\n\n/**\n * Get the last known download phase from storage.\n */\nexport function getDownloadPhase(): SessionState | null {\n  if (typeof localStorage === \"undefined\") return null;\n\n  try {\n    const raw = localStorage.getItem(SESSION_STORAGE_KEY);\n    if (!raw) return null;\n    return JSON.parse(raw) as SessionState;\n  } catch {\n    return null;\n  }\n}\n\n/**\n * Detect if the page reloaded during a model download/initialization.\n * This typically indicates an iOS memory crash.\n *\n * @returns Detection result with recommended action\n */\nexport function detectMemoryCrash(): {\n  crashed: boolean;\n  phase?: DownloadPhase;\n  modelId?: string;\n  timeSinceCrash?: number;\n  recommendation?: string;\n} {\n  const lastState = getDownloadPhase();\n  const currentSessionId = getSessionId();\n\n  if (!lastState) {\n    return { crashed: false };\n  }\n\n  // If session ID changed and we were in a critical phase, it's likely a crash/reload\n  const criticalPhases: DownloadPhase[] = [\"downloading\", \"caching\", \"initializing\"];\n  const wasInCriticalPhase = criticalPhases.includes(lastState.phase);\n  const sessionChanged = lastState.sessionId !== currentSessionId;\n  const timeSinceCrash = Date.now() - lastState.timestamp;\n\n  // Only count as crash if it happened recently (within 5 minutes) and session changed\n  const recentCrash = timeSinceCrash < 5 * 60 * 1000;\n\n  if (wasInCriticalPhase && sessionChanged && recentCrash) {\n    // Clear the state to prevent repeated detection\n    localStorage.removeItem(SESSION_STORAGE_KEY);\n\n    return {\n      crashed: true,\n      phase: lastState.phase,\n      modelId: lastState.modelId || undefined,\n      timeSinceCrash,\n      recommendation:\n        lastState.modelId && /2b|gemma|vision/i.test(lastState.modelId)\n          ? \"The model was too large for your device. Try Qwen3.5-0.8B instead.\"\n          : \"Your device ran out of memory. Try a smaller model or use a desktop browser.\",\n    };\n  }\n\n  return { crashed: false };\n}\n\n/**\n * Clear session phase (call when model loads successfully).\n */\nexport function clearDownloadPhase(): void {\n  if (typeof localStorage === \"undefined\") return;\n  localStorage.removeItem(SESSION_STORAGE_KEY);\n}\n\n// ============================================\n// WebKit Submit-Granularity (group size) Probe\n// ============================================\n//\n// On WebKit/iOS WebGPU, decode speed is bound by GPU round-trips: the executor\n// groups `webkitGroupSize` dispatches per command buffer, then submits + drains\n// (queue.onSubmittedWorkDone) per group. Larger groups amortize the drains and\n// can be ~5x faster on iPad. BUT the safe ceiling is device-dependent in three\n// classes:\n//   1. Newer WebKit (iPad 26.5): high group is correct + fast.\n//   2. Older WebKit: high group yields zero/garbage logits (wrong output).\n//   3. iPhone (iOS 18.7): high group HARD-CRASHES the GPU process (page dies).\n//\n// The crash class is the hard one: a bad group kills the page before any code\n// can record the result. So we use a localStorage BREADCRUMB that OUTLIVES the\n// page kill: we persist the candidate we're about to try BEFORE running any\n// GPU-heavy work. If the page survives, a later promotion clears `trying` and\n// records it as known-good. If the page crashed, `trying` is still set on the\n// next load, so we KNOW that rung crashed and cap below it.\n//\n// The probe escalates UP from the safe floor (never starts optimistic — a\n// first-visit crash is a terrible UX) and persists per-device in localStorage.\n\nexport const WEBKIT_GROUP_PROBE_KEY = \"gerbil-webkit-group-v1\";\n\n/**\n * Candidate group-size ladder. We escalate UP one rung per page load. 128 is\n * treated as effectively \"batch-all\" for this model — kept simple on purpose.\n */\nexport const WEBKIT_GROUP_LADDER = [1, 8, 32, 64, 128] as const;\n\nexport type WebkitGroupProbe = {\n  /** Largest group size proven crash-free AND correct on this device. Starts at 1. */\n  knownGood: number;\n  /** Candidate being attempted this page-load (null when not probing). */\n  trying: number | null;\n  /** True once we hit a crash/incorrect ceiling and should stop escalating. */\n  capped: boolean;\n};\n\nconst DEFAULT_PROBE: WebkitGroupProbe = { knownGood: 1, trying: null, capped: false };\n\n/** Read the persisted WebKit group probe record (guarded; safe on node). */\nexport function readGroupProbe(): WebkitGroupProbe {\n  if (typeof localStorage === \"undefined\") return { ...DEFAULT_PROBE };\n  try {\n    const raw = localStorage.getItem(WEBKIT_GROUP_PROBE_KEY);\n    if (!raw) return { ...DEFAULT_PROBE };\n    const parsed = JSON.parse(raw) as Partial<WebkitGroupProbe>;\n    return {\n      knownGood:\n        typeof parsed.knownGood === \"number\" && parsed.knownGood >= 1 ? parsed.knownGood : 1,\n      trying: typeof parsed.trying === \"number\" ? parsed.trying : null,\n      capped: parsed.capped === true,\n    };\n  } catch {\n    return { ...DEFAULT_PROBE };\n  }\n}\n\n/** Persist the WebKit group probe record (guarded; no-op on node). */\nexport function writeGroupProbe(rec: WebkitGroupProbe): void {\n  if (typeof localStorage === \"undefined\") return;\n  try {\n    localStorage.setItem(WEBKIT_GROUP_PROBE_KEY, JSON.stringify(rec));\n  } catch {\n    // Best-effort; private mode / quota can throw.\n  }\n}\n\n/**\n * The validated non-phone sweet spot. iPad swept 1→7.9, 8→19, 32→24.8, 64→26.6,\n * 128→26.9 (peak), 256→26.2 tok/s — a plateau from ~64 up, so 128 is the best\n * stable target (more batching just costs memory). Non-phone WebKit jumps here\n * directly; the crash breadcrumb caps it down if a device can't sustain it.\n */\nconst NONPHONE_TARGET_GROUP = 128;\n\n/**\n * Resolve the WebKit group size to use this session, recording `trying` as a\n * side effect so a crash this load is detectable on the next load.\n *\n * Algorithm (only meaningful on WebKit; inert otherwise):\n *  1. Read the record (default {knownGood:1, trying:null, capped:false}).\n *  2. If `trying !== null` on entry → the previous load set it but never cleared\n *     it → that load CRASHED at `trying`. Cap there, keep `knownGood`, clear\n *     `trying`. Use `knownGood` this session.\n *  3. Else if !capped and there is a rung above `knownGood` → set `trying = next`,\n *     persist BEFORE any GPU work, and use it (we're escalating).\n *  4. Else → use `knownGood`.\n *\n * @returns the group size to use this session.\n */\nexport function resolveWebkitGroupSize(args: {\n  override?: number;\n  isWebKit: boolean;\n  /**\n   * Memory-constrained devices (phones) are the \"crash class\": batching crashes\n   * the GPU process, and discovering that costs one user-visible crash before the\n   * breadcrumb caps it. When `conservative` is set we never auto-escalate — the\n   * device stays at its proven floor (group=1 on first run) and never risks that\n   * calibration crash. `?group=N` still lets such a device opt in explicitly.\n   */\n  conservative?: boolean;\n}): number {\n  // Explicit ?group=N override wins and skips the probe entirely.\n  if (args.override && args.override > 0) return args.override;\n  // Non-WebKit (desktop / Dawn / node) never groups — floor of 1, no bookkeeping.\n  if (!args.isWebKit) return 1;\n\n  const rec = readGroupProbe();\n\n  // Step 2: a stale `trying` means last load crashed at that rung.\n  if (rec.trying !== null) {\n    const crashedAt = rec.trying;\n    const next: WebkitGroupProbe = {\n      knownGood: rec.knownGood,\n      trying: null,\n      capped: true,\n    };\n    writeGroupProbe(next);\n    console.log(\n      `[engine] webkit group probe: previous load crashed at group=${crashedAt} → capping at knownGood=${next.knownGood}`,\n    );\n    return next.knownGood;\n  }\n\n  // Step 3: non-phone WebKit (iPad / Mac Safari) jumps straight to the validated\n  // sweet spot (group=128) instead of climbing rung-by-rung — these devices\n  // tolerate batching (iPad ran every rung to 256 coherently), so there's no\n  // reason to crawl. The breadcrumb (Step 2) caps it down if a particular device\n  // can't sustain it. Phones (conservative) NEVER escalate — they stay at the\n  // proven floor (group=1) to avoid the one calibration crash batching costs on\n  // the crash class.\n  if (!rec.capped && !args.conservative && rec.knownGood < NONPHONE_TARGET_GROUP) {\n    const next: WebkitGroupProbe = {\n      knownGood: rec.knownGood,\n      trying: NONPHONE_TARGET_GROUP,\n      capped: false,\n    };\n    writeGroupProbe(next); // PERSIST BEFORE any GPU-heavy work runs\n    console.log(\n      `[engine] webkit group probe: trying target group=${NONPHONE_TARGET_GROUP} (knownGood=${rec.knownGood})`,\n    );\n    return NONPHONE_TARGET_GROUP;\n  }\n\n  // Step 4: nothing to escalate (capped or at top rung) — use the proven floor.\n  console.log(\n    `[engine] webkit group probe: knownGood=${rec.knownGood} trying=null capped=${rec.capped} → using ${rec.knownGood}`,\n  );\n  return rec.knownGood;\n}\n\n/**\n * Promote (or cap) the WebKit group probe after the first successful forward.\n *\n * Call this once per page-load, after the model has loaded AND a first forward\n * completed without the page dying. The breadcrumb already handles the crash\n * class (the page death leaves `trying` set for the next load); this handles the\n * wrong-output class and records success.\n *\n * @param correct  true if the first forward produced non-corrupt output.\n *   - correct  → promote: knownGood = trying, trying = null.\n *   - incorrect → cap: keep knownGood at the prior rung, trying = null, capped.\n */\nexport function promoteGroupProbe(correct: boolean): void {\n  if (typeof localStorage === \"undefined\") return;\n  const rec = readGroupProbe();\n  // Nothing was being attempted this load — nothing to promote.\n  if (rec.trying === null) return;\n\n  if (correct) {\n    writeGroupProbe({ knownGood: rec.trying, trying: null, capped: rec.capped });\n    console.log(`[engine] webkit group probe: PROMOTED group=${rec.trying} to known-good`);\n  } else {\n    // Wrong-output class: this rung is incorrect (but did not crash). Keep the\n    // prior knownGood and stop escalating.\n    writeGroupProbe({ knownGood: rec.knownGood, trying: null, capped: true });\n    console.log(\n      `[engine] webkit group probe: group=${rec.trying} produced INCORRECT output → capping at knownGood=${rec.knownGood}`,\n    );\n  }\n}\n","import { setDownloadPhase } from \"./device-guards.js\";\n\n// ============================================\n// Chunked Resumable Downloader\n// ============================================\n\n/** Chunk size for downloads: 1.5MB (safe for iOS IndexedDB transactions) */\nexport const CHUNK_SIZE_BYTES = 1.5 * 1024 * 1024;\n\n/** IndexedDB database name for chunked downloads */\nexport const DOWNLOAD_DB_NAME = \"gerbil-model-chunks\";\nconst DOWNLOAD_DB_VERSION = 1;\n\n/**\n * Manifest stored in IndexedDB to track download progress.\n */\ntype DownloadManifest = {\n  modelId: string;\n  url: string;\n  etag: string | null;\n  totalBytes: number;\n  chunkSize: number;\n  completedChunks: number[];\n  createdAt: number;\n  updatedAt: number;\n};\n\n/**\n * Open (or create) the IndexedDB for chunked downloads.\n */\nasync function openDownloadDB(): Promise<IDBDatabase> {\n  return new Promise((resolve, reject) => {\n    const request = indexedDB.open(DOWNLOAD_DB_NAME, DOWNLOAD_DB_VERSION);\n\n    request.onerror = () =>\n      reject(new Error(`Failed to open download DB: ${request.error?.message}`));\n\n    request.onsuccess = () => resolve(request.result);\n\n    request.onupgradeneeded = (event) => {\n      const db = (event.target as IDBOpenDBRequest).result;\n\n      // Store for download manifests\n      if (!db.objectStoreNames.contains(\"manifests\")) {\n        db.createObjectStore(\"manifests\", { keyPath: \"modelId\" });\n      }\n\n      // Store for actual chunks (key: modelId-chunkIndex)\n      if (!db.objectStoreNames.contains(\"chunks\")) {\n        db.createObjectStore(\"chunks\");\n      }\n    };\n  });\n}\n\n/**\n * Get download manifest for a model.\n */\nasync function getManifest(db: IDBDatabase, modelId: string): Promise<DownloadManifest | null> {\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction(\"manifests\", \"readonly\");\n    const store = tx.objectStore(\"manifests\");\n    const request = store.get(modelId);\n\n    request.onerror = () => reject(new Error(`Failed to get manifest: ${request.error?.message}`));\n    request.onsuccess = () => resolve(request.result || null);\n  });\n}\n\n/**\n * Save download manifest.\n */\nasync function saveManifest(db: IDBDatabase, manifest: DownloadManifest): Promise<void> {\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction(\"manifests\", \"readwrite\");\n    const store = tx.objectStore(\"manifests\");\n    const request = store.put(manifest);\n\n    request.onerror = () => reject(new Error(`Failed to save manifest: ${request.error?.message}`));\n    request.onsuccess = () => resolve();\n  });\n}\n\n/**\n * Save a single chunk.\n */\nasync function saveChunk(\n  db: IDBDatabase,\n  modelId: string,\n  chunkIndex: number,\n  data: ArrayBuffer,\n): Promise<void> {\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction(\"chunks\", \"readwrite\");\n    const store = tx.objectStore(\"chunks\");\n    const key = `${modelId}-${chunkIndex}`;\n    const request = store.put(data, key);\n\n    request.onerror = () =>\n      reject(new Error(`Failed to save chunk ${chunkIndex}: ${request.error?.message}`));\n    request.onsuccess = () => resolve();\n  });\n}\n\n/**\n * Get a single chunk.\n */\nasync function getChunk(\n  db: IDBDatabase,\n  modelId: string,\n  chunkIndex: number,\n): Promise<ArrayBuffer | null> {\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction(\"chunks\", \"readonly\");\n    const store = tx.objectStore(\"chunks\");\n    const key = `${modelId}-${chunkIndex}`;\n    const request = store.get(key);\n\n    request.onerror = () =>\n      reject(new Error(`Failed to get chunk ${chunkIndex}: ${request.error?.message}`));\n    request.onsuccess = () => resolve(request.result || null);\n  });\n}\n\n/**\n * Delete all chunks and manifest for a model.\n */\nasync function clearModelData(db: IDBDatabase, modelId: string): Promise<void> {\n  // Get manifest to know how many chunks to delete\n  const manifest = await getManifest(db, modelId);\n\n  return new Promise((resolve, reject) => {\n    const tx = db.transaction([\"manifests\", \"chunks\"], \"readwrite\");\n\n    // Delete manifest\n    tx.objectStore(\"manifests\").delete(modelId);\n\n    // Delete all chunks\n    if (manifest) {\n      const totalChunks = Math.ceil(manifest.totalBytes / manifest.chunkSize);\n      const chunkStore = tx.objectStore(\"chunks\");\n      for (let i = 0; i < totalChunks; i++) {\n        chunkStore.delete(`${modelId}-${i}`);\n      }\n    }\n\n    tx.oncomplete = () => resolve();\n    tx.onerror = () => reject(new Error(`Failed to clear model data: ${tx.error?.message}`));\n  });\n}\n\n/**\n * Chunked resumable downloader for large model files.\n * Downloads in 1.5MB chunks to avoid iOS memory pressure.\n */\nexport async function downloadModelChunked(\n  url: string,\n  modelId: string,\n  options: {\n    onProgress?: (info: {\n      phase: string;\n      bytesDownloaded: number;\n      totalBytes: number;\n      percent: number;\n    }) => void;\n    signal?: AbortSignal;\n  } = {},\n): Promise<ArrayBuffer> {\n  const { onProgress, signal } = options;\n\n  // Update session phase\n  setDownloadPhase(\"downloading\", modelId);\n\n  const db = await openDownloadDB();\n\n  try {\n    // Check for existing manifest\n    let manifest = await getManifest(db, modelId);\n\n    // Fetch headers to get content-length and etag\n    const headResponse = await fetch(url, { method: \"HEAD\", signal });\n    if (!headResponse.ok) {\n      throw new Error(`HEAD request failed: ${headResponse.status} ${headResponse.statusText}`);\n    }\n\n    const contentLength = parseInt(headResponse.headers.get(\"content-length\") || \"0\", 10);\n    const etag = headResponse.headers.get(\"etag\");\n    const acceptRanges = headResponse.headers.get(\"accept-ranges\");\n\n    if (!contentLength) {\n      throw new Error(\"Server did not provide content-length\");\n    }\n\n    // Check if we need to restart (etag mismatch means model updated)\n    if (manifest && manifest.etag !== etag) {\n      console.warn(`Model ${modelId} has been updated (etag mismatch). Clearing cached chunks.`);\n      await clearModelData(db, modelId);\n      manifest = null;\n    }\n\n    // Check if server supports range requests\n    const supportsRange = acceptRanges === \"bytes\";\n\n    if (!supportsRange) {\n      // Fall back to regular download\n      console.warn(`Server doesn't support range requests for ${modelId}. Using regular download.`);\n      db.close();\n\n      const response = await fetch(url, { signal });\n      if (!response.ok) throw new Error(`Download failed: ${response.status}`);\n\n      setDownloadPhase(\"caching\", modelId);\n      const buffer = await response.arrayBuffer();\n      setDownloadPhase(\"ready\", modelId);\n      return buffer;\n    }\n\n    // Create or update manifest\n    const totalChunks = Math.ceil(contentLength / CHUNK_SIZE_BYTES);\n\n    if (!manifest) {\n      manifest = {\n        modelId,\n        url,\n        etag,\n        totalBytes: contentLength,\n        chunkSize: CHUNK_SIZE_BYTES,\n        completedChunks: [],\n        createdAt: Date.now(),\n        updatedAt: Date.now(),\n      };\n      await saveManifest(db, manifest);\n    }\n\n    // Download missing chunks\n    for (let i = 0; i < totalChunks; i++) {\n      if (signal?.aborted) {\n        throw new Error(\"Download aborted\");\n      }\n\n      // Skip already downloaded chunks\n      if (manifest.completedChunks.includes(i)) {\n        const bytesDownloaded = (manifest.completedChunks.length / totalChunks) * contentLength;\n        onProgress?.({\n          phase: \"resuming\",\n          bytesDownloaded,\n          totalBytes: contentLength,\n          percent: Math.round((bytesDownloaded / contentLength) * 100),\n        });\n        continue;\n      }\n\n      const start = i * CHUNK_SIZE_BYTES;\n      const end = Math.min(start + CHUNK_SIZE_BYTES - 1, contentLength - 1);\n\n      // Download chunk with Range header\n      const response = await fetch(url, {\n        headers: { Range: `bytes=${start}-${end}` },\n        signal,\n      });\n\n      if (response.status !== 206) {\n        throw new Error(`Range request failed: ${response.status} (expected 206)`);\n      }\n\n      const chunkData = await response.arrayBuffer();\n\n      // Save chunk to IndexedDB\n      await saveChunk(db, modelId, i, chunkData);\n\n      // Update manifest\n      manifest.completedChunks.push(i);\n      manifest.updatedAt = Date.now();\n      await saveManifest(db, manifest);\n\n      // Update session phase with progress\n      const bytesDownloaded = manifest.completedChunks.length * CHUNK_SIZE_BYTES;\n      setDownloadPhase(\"downloading\", modelId, { bytesDownloaded, totalBytes: contentLength });\n\n      onProgress?.({\n        phase: \"downloading\",\n        bytesDownloaded: Math.min(bytesDownloaded, contentLength),\n        totalBytes: contentLength,\n        percent: Math.round((manifest.completedChunks.length / totalChunks) * 100),\n      });\n\n      // Null out reference to allow GC before next chunk\n      // @ts-expect-error - intentional null for GC\n      response.body = null;\n    }\n\n    // All chunks downloaded - reassemble\n    setDownloadPhase(\"caching\", modelId);\n    onProgress?.({\n      phase: \"assembling\",\n      bytesDownloaded: contentLength,\n      totalBytes: contentLength,\n      percent: 100,\n    });\n\n    // Assemble chunks into final buffer\n    // We do this incrementally to avoid holding all chunks in memory at once\n    const finalBuffer = new ArrayBuffer(contentLength);\n    const finalView = new Uint8Array(finalBuffer);\n\n    for (let i = 0; i < totalChunks; i++) {\n      const chunk = await getChunk(db, modelId, i);\n      if (!chunk) {\n        throw new Error(`Missing chunk ${i} during assembly`);\n      }\n\n      const offset = i * CHUNK_SIZE_BYTES;\n      finalView.set(new Uint8Array(chunk), offset);\n\n      // Null out chunk reference immediately after copying\n      // This helps GC reclaim memory on iOS\n    }\n\n    // Clean up - delete chunks now that we have the full file\n    await clearModelData(db, modelId);\n    db.close();\n\n    setDownloadPhase(\"ready\", modelId);\n    return finalBuffer;\n  } catch (error) {\n    setDownloadPhase(\"error\", modelId);\n    db.close();\n    throw error;\n  }\n}\n\n/**\n * Check if a model has an incomplete download.\n */\nexport async function hasIncompleteDownload(modelId: string): Promise<{\n  incomplete: boolean;\n  bytesDownloaded?: number;\n  totalBytes?: number;\n  percent?: number;\n}> {\n  try {\n    const db = await openDownloadDB();\n    const manifest = await getManifest(db, modelId);\n    db.close();\n\n    if (!manifest) {\n      return { incomplete: false };\n    }\n\n    const totalChunks = Math.ceil(manifest.totalBytes / manifest.chunkSize);\n    const completedChunks = manifest.completedChunks.length;\n\n    if (completedChunks < totalChunks) {\n      return {\n        incomplete: true,\n        bytesDownloaded: completedChunks * manifest.chunkSize,\n        totalBytes: manifest.totalBytes,\n        percent: Math.round((completedChunks / totalChunks) * 100),\n      };\n    }\n\n    return { incomplete: false };\n  } catch {\n    return { incomplete: false };\n  }\n}\n\n/**\n * Clear incomplete download data for a model.\n */\nexport async function clearIncompleteDownload(modelId: string): Promise<void> {\n  try {\n    const db = await openDownloadDB();\n    await clearModelData(db, modelId);\n    db.close();\n  } catch {\n    // Ignore errors\n  }\n}\n\n/**\n * Check if there's enough storage quota for a model download.\n * Returns estimated available space and whether download should proceed.\n */\nexport async function checkStorageQuota(requiredMB: number = 500): Promise<{\n  ok: boolean;\n  availableMB: number;\n  usedMB: number;\n  quotaMB: number;\n  message?: string;\n}> {\n  if (typeof navigator === \"undefined\" || !navigator.storage?.estimate) {\n    return {\n      ok: true,\n      availableMB: -1,\n      usedMB: -1,\n      quotaMB: -1,\n      message: \"Storage API not available\",\n    };\n  }\n\n  try {\n    const { quota, usage } = await navigator.storage.estimate();\n    const quotaMB = Math.round((quota || 0) / 1_000_000);\n    const usedMB = Math.round((usage || 0) / 1_000_000);\n    const availableMB = quotaMB - usedMB;\n\n    if (availableMB < requiredMB) {\n      return {\n        ok: false,\n        availableMB,\n        usedMB,\n        quotaMB,\n        message: `Need ${requiredMB}MB but only ${availableMB}MB available. Clear browser data or free up space.`,\n      };\n    }\n\n    return { ok: true, availableMB, usedMB, quotaMB };\n  } catch (e: any) {\n    return {\n      ok: true, // Proceed optimistically if we can't check\n      availableMB: -1,\n      usedMB: -1,\n      quotaMB: -1,\n      message: `Storage check failed: ${e.message}`,\n    };\n  }\n}\n"],"mappings":";AA4BA,MAAaA,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;;;;;;;;;;;;;;;;;;;;;;;;;;ACxCD,SAAgB,eAAwB;AACtC,KAAI,OAAO,WAAW,YAAa,QAAO;CAE1C,MAAM,gBAAiB,UAAuC,eAAe;CAC7E,MAAM,oBACJ,OAAO,OAAO,eAAe,cAC7B,OAAO,WAAW,6BAA6B,CAAC;AAClD,QAAO,iBAAiB;;;;;AAM1B,SAAgB,QAAiB;AAC/B,KAAI,OAAO,cAAc,YAAa,QAAO;CAC7C,MAAM,KAAK,UAAU,aAAa;AAClC,KAAI,mBAAmB,KAAK,GAAG,CAAE,QAAO;AAGxC,QADoB,YAAY,KAAK,GAAG,IAAI,cAAc,KAAK,GAAG,IAAI,CAAC,SAAS,KAAK,GAAG,KAChE,UAA0C,kBAAkB,KAAK;;;;;AAsB3F,eAAsB,mBAA2C;CAC/D,MAAM,YAAY,cAAc;CAChC,MAAM,MAAM,OAAO;CACnB,IAAI,UAAU;CACd,IAAI,UAAU;CACd,IAAI,YAAY;AAChB,KAAI;EACF,MAAM,MAAM,MAAM,UAAU,SAAS,YAAY;AACjD,YAAU,KAAK,OAAO,KAAK,SAAS,KAAK,IAAI;AAC7C,YAAU,KAAK,OAAO,KAAK,SAAS,KAAK,IAAI;SACvC;AAGR,KAAI;AACF,cAAa,MAAM,UAAU,SAAS,aAAa,IAAK;SAClD;AAGR,QAAO;EACL;EACA;EACA,aAAa,KAAK,IAAI,GAAG,UAAU,QAAQ;EAC3C;EACA;EACA;EACD;;;;;;;;AASH,eAAsB,2BAA6C;AACjE,KAAI;AACF,MAAI,MAAM,UAAU,SAAS,aAAa,CAAE,QAAO;AACnD,SAAQ,MAAM,UAAU,SAAS,WAAW,IAAK;SAC3C;AACN,SAAO;;;;;;;;AAkBX,eAAsB,cAAc,QAAmC;CACrE,MAAM,IAAI,MAAM,kBAAkB;CAGlC,MAAM,OAAO,EAAE,eAAe,SAAS;CACvC,MAAM,oBAAoB,CAAC,QAAS,EAAE,OAAO,CAAC,EAAE,cAAe,CAAC,EAAE;AAClE,QAAO;EAAE;EAAM,aAAa,EAAE;EAAa;EAAkB;;;;;;;;AAS/D,SAAgB,qBAA6E;AAE3F,KADkB,cAAc,CACjB,QAAO;EAAE,WAAW;EAAM,QAAQ;EAAO,OAAO;EAAsB;AACrF,KAAI,OAAO,CACT,QAAO;EACL,WAAW;EACX,QAAQ;EACR,OACE;EACH;AAEH,QAAO;EACL,WAAW;EACX,QAAQ;EACR,OACE;EACH;;;;;;;;;;;;;;;;;;;ACtIH,eAAsB,UACpB,OACA,aAAqB,MACkC;CACvD,MAAM,eAAe,IAAI,cAAc;AAGvC,KAAI,aAAa,UAAU,YACzB,OAAM,aAAa,QAAQ;CAG7B,MAAM,cAAc,aAAa,aAAa,GAAG,MAAM,QAAQ,WAAW;CAC1E,MAAM,cAAc,IAAI,aAAa,MAAM;AAC3C,aAAY,cAAc,aAAa,EAAE;CAEzC,MAAM,aAAa,aAAa,oBAAoB;AACpD,YAAW,SAAS;AACpB,YAAW,QAAQ,aAAa,YAAY;CAE5C,MAAM,UAAU,IAAI,SAAe,YAAY;AAC7C,aAAW,gBAAgB;AACzB,gBAAa,OAAO;AACpB,YAAS;;GAEX;AAEF,YAAW,OAAO;AAElB,QAAO;EACL,YAAY;AACV,cAAW,MAAM;AACjB,gBAAa,OAAO;;EAEtB;EACD;;;;;;;;;;;;;;;;;;;AAoBH,SAAgB,kBAAkB,aAAqB,MAIrD;CACA,IAAIC,eAAoC;CACxC,IAAI,gBAAgB;CACpB,IAAI,WAAW;CAEf,MAAM,gBAAgB,YAAY;AAChC,MAAI,CAAC,aACH,gBAAe,IAAI,cAAc;AAEnC,MAAI,aAAa,UAAU,YACzB,OAAM,aAAa,QAAQ;AAE7B,SAAO;;AAGT,QAAO;EACL,OAAO,OAAO,UAAwB;GACpC,MAAM,MAAM,MAAM,eAAe;AACjC,cAAW;GAEX,MAAM,SAAS,IAAI,aAAa,GAAG,MAAM,QAAQ,WAAW;GAC5D,MAAM,cAAc,IAAI,aAAa,MAAM;AAC3C,UAAO,cAAc,aAAa,EAAE;GAEpC,MAAM,SAAS,IAAI,oBAAoB;AACvC,UAAO,SAAS;AAChB,UAAO,QAAQ,IAAI,YAAY;GAG/B,MAAM,YAAY,KAAK,IAAI,IAAI,aAAa,cAAc;AAC1D,UAAO,MAAM,UAAU;AACvB,mBAAgB,YAAY,OAAO;AAEnC,UAAO,gBAAgB;AACrB,QAAI,IAAI,eAAe,gBAAgB,GACrC,YAAW;;;EAKjB,YAAY;AACV,cAAW;AACX,mBAAgB;AAChB,OAAI,cAAc;AAChB,iBAAa,OAAO;AACpB,mBAAe;;;EAInB,iBAAiB;EAClB;;;;;;AC/GH,MAAM,mBAAmB;;;;;AAMzB,MAAaC,cAAsC;CAEjD,gBAAgB;CAChB,cAAc;CACd,eAAe;CACf,eAAe;CAEf,cAAc;CACd,kBAAkB;CAElB,gBAAgB;CAChB,mBAAmB;CACnB,iBAAiB;CAEjB,oBAAoB;CACrB;;;;;AAMD,SAAS,YAAY,SAAyB;AAC5C,QAAO,QAAQ,aAAa,CAAC,QAAQ,cAAc,IAAI;;;;;;;;;;;;AAazD,MAAM,mBAAmB;CAEvB,SAAS,CAAC,eAAe,cAAc;CAEvC,eAAe;EAAC;EAAU;EAAQ;EAAO;EAAO;EAAc;EAAa;CAE3E,OAAO,CAAC,cAAc,aAAa;CAEnC,aAAa;CACd;;;;;;AAOD,SAAgB,qBAAqB,SAWnC;CACA,MAAM,KAAK,OAAO,cAAc,cAAc,UAAU,YAAY;CACpE,MAAM,WAAW,cAAc,KAAK,GAAG;CACvC,MAAM,SAAS,OAAO,KAAK,GAAG;CAE9B,MAAM,eADQ,YAAY,WACG,QAAQ,KAAK,GAAG;CAC7C,MAAM,eAAe,YAAY,QAAQ;CAEzC,MAAM,WAAW,iBAAiB,cAAc,MAAM,MACpD,aAAa,SAAS,YAAY,EAAE,CAAC,CACtC;CACD,MAAM,YACJ,iBAAiB,QAAQ,MAAM,MAAM,aAAa,SAAS,YAAY,EAAE,CAAC,CAAC,IAAI;CACjF,MAAM,UAAU,iBAAiB,MAAM,MAAM,MAAM,aAAa,SAAS,YAAY,EAAE,CAAC,CAAC;AAIzF,KAAI,UAAU;AACZ,MAAI,UAKF,QAAO;GACL,MAAM;GACN,OAAO;GACP,QAAQ,SAAS,QAAQ,uBAPP,cAAc,wDAAwD,GAO5B,IANlD,WACR,+DACA,2BAIkE;GACpE,gBAAgB,OAAO,iBAAiB;GACxC,cAAc;GACf;AAEH,MAAI,QACF,QAAO;GACL,MAAM;GACN,OAAO;GACP,QAAQ,SAAS,QAAQ;GACzB,gBAAgB,OAAO,iBAAiB;GACxC,cAAc;GACf;AAEH,SAAO;GAAE,MAAM;GAAM,OAAO;GAAO,QAAQ;GAAyC;;AAGtF,KAAI,QAAQ;AAEV,MAAI,UAIF,QAAO;GACL,MAAM;GACN,OAAO;GACP,QAAQ,SAAS,QAAQ,sBANf,WACR,+DACA,2BAIiD;GACnD,gBAAgB,OAAO,iBAAiB;GACxC,cAAc;GACf;AAEH,SAAO;GAAE,MAAM;GAAM,OAAO;GAAO,QAAQ;GAAuC;;AAKpF,KADkB,UAAU,KAAK,GAAG,IACnB,UACf,QAAO;EACL,MAAM;EACN,OAAO;EACP,QAAQ,SAAS,QAAQ;EACzB,gBAAgB,OAAO,iBAAiB;EACxC,cAAc;EACf;AAIH,QAAO;EAAE,MAAM;EAAM,OAAO;EAAO,QAAQ;EAA0C;;;;;;AAOvF,SAAgB,uBAQd;CACA,MAAM,KAAK,OAAO,cAAc,cAAc,UAAU,YAAY;CACpE,MAAM,eAAe,OAAO,cAAc,cAAe,UAAkB,eAAe;CAC1F,MAAM,WAAW,kCAAkC,KAAK,GAAG;CAK3D,MAAM,eADkB,eAAgB,WAAW,eAAe,KAAM,eAAe,KAAO,KACxD;CAEtC,IAAIC;CACJ,IAAIC;AAEJ,KAAI,cAAc,KAAK;AACrB,SAAO;AACP,WAAS;YACA,YAAY,cAAc,MAAM;AACzC,SAAO;AACP,WAAS;YACA,cAAc,MAAM;AAC7B,SAAO;AACP,WAAS;QACJ;AACL,SAAO;AACP,WAAS;;AAGX,QAAO;EACL;EACA,KAAK;EACL,KAAK;EACL,WAAW;EACX;EACA;EACA;EACD;;AASH,MAAa,sBAAsB;;;;AAcnC,SAAS,oBAA4B;AACnC,QAAO,GAAG,KAAK,KAAK,CAAC,GAAG,KAAK,QAAQ,CAAC,SAAS,GAAG,CAAC,MAAM,GAAG,EAAE;;;;;AAMhE,SAAS,eAAuB;AAC9B,KAAI,OAAO,iBAAiB,YAAa,QAAO,mBAAmB;CAEnE,IAAI,YAAY,eAAe,QAAQ,oBAAoB;AAC3D,KAAI,CAAC,WAAW;AACd,cAAY,mBAAmB;AAC/B,iBAAe,QAAQ,qBAAqB,UAAU;;AAExD,QAAO;;;;;;AAOT,SAAgB,iBACd,OACA,SACA,UACM;AACN,KAAI,OAAO,iBAAiB,YAAa;CAEzC,MAAMC,QAAsB;EAC1B;EACA,SAAS,WAAW;EACpB,WAAW,cAAc;EACzB,WAAW,KAAK,KAAK;EACrB,iBAAiB,UAAU;EAC3B,YAAY,UAAU;EACvB;AAED,cAAa,QAAQ,qBAAqB,KAAK,UAAU,MAAM,CAAC;;;;;AAMlE,SAAgB,mBAAwC;AACtD,KAAI,OAAO,iBAAiB,YAAa,QAAO;AAEhD,KAAI;EACF,MAAM,MAAM,aAAa,QAAQ,oBAAoB;AACrD,MAAI,CAAC,IAAK,QAAO;AACjB,SAAO,KAAK,MAAM,IAAI;SAChB;AACN,SAAO;;;;;;;;;AAUX,SAAgB,oBAMd;CACA,MAAM,YAAY,kBAAkB;CACpC,MAAM,mBAAmB,cAAc;AAEvC,KAAI,CAAC,UACH,QAAO,EAAE,SAAS,OAAO;CAK3B,MAAM,qBADkC;EAAC;EAAe;EAAW;EAAe,CACxC,SAAS,UAAU,MAAM;CACnE,MAAM,iBAAiB,UAAU,cAAc;CAC/C,MAAM,iBAAiB,KAAK,KAAK,GAAG,UAAU;AAK9C,KAAI,sBAAsB,kBAFN,iBAAiB,MAAS,KAEW;AAEvD,eAAa,WAAW,oBAAoB;AAE5C,SAAO;GACL,SAAS;GACT,OAAO,UAAU;GACjB,SAAS,UAAU,WAAW;GAC9B;GACA,gBACE,UAAU,WAAW,mBAAmB,KAAK,UAAU,QAAQ,GAC3D,uEACA;GACP;;AAGH,QAAO,EAAE,SAAS,OAAO;;;;;AAM3B,SAAgB,qBAA2B;AACzC,KAAI,OAAO,iBAAiB,YAAa;AACzC,cAAa,WAAW,oBAAoB;;;;;;AC3U9C,MAAa,mBAAmB,MAAM,OAAO;;AAG7C,MAAa,mBAAmB;AAChC,MAAM,sBAAsB;;;;AAmB5B,eAAe,iBAAuC;AACpD,QAAO,IAAI,SAAS,SAAS,WAAW;EACtC,MAAM,UAAU,UAAU,KAAK,kBAAkB,oBAAoB;AAErE,UAAQ,gBACN,uBAAO,IAAI,MAAM,+BAA+B,QAAQ,OAAO,UAAU,CAAC;AAE5E,UAAQ,kBAAkB,QAAQ,QAAQ,OAAO;AAEjD,UAAQ,mBAAmB,UAAU;GACnC,MAAM,KAAM,MAAM,OAA4B;AAG9C,OAAI,CAAC,GAAG,iBAAiB,SAAS,YAAY,CAC5C,IAAG,kBAAkB,aAAa,EAAE,SAAS,WAAW,CAAC;AAI3D,OAAI,CAAC,GAAG,iBAAiB,SAAS,SAAS,CACzC,IAAG,kBAAkB,SAAS;;GAGlC;;;;;AAMJ,eAAe,YAAY,IAAiB,SAAmD;AAC7F,QAAO,IAAI,SAAS,SAAS,WAAW;EAGtC,MAAM,UAFK,GAAG,YAAY,aAAa,WAAW,CACjC,YAAY,YAAY,CACnB,IAAI,QAAQ;AAElC,UAAQ,gBAAgB,uBAAO,IAAI,MAAM,2BAA2B,QAAQ,OAAO,UAAU,CAAC;AAC9F,UAAQ,kBAAkB,QAAQ,QAAQ,UAAU,KAAK;GACzD;;;;;AAMJ,eAAe,aAAa,IAAiB,UAA2C;AACtF,QAAO,IAAI,SAAS,SAAS,WAAW;EAGtC,MAAM,UAFK,GAAG,YAAY,aAAa,YAAY,CAClC,YAAY,YAAY,CACnB,IAAI,SAAS;AAEnC,UAAQ,gBAAgB,uBAAO,IAAI,MAAM,4BAA4B,QAAQ,OAAO,UAAU,CAAC;AAC/F,UAAQ,kBAAkB,SAAS;GACnC;;;;;AAMJ,eAAe,UACb,IACA,SACA,YACA,MACe;AACf,QAAO,IAAI,SAAS,SAAS,WAAW;EAEtC,MAAM,QADK,GAAG,YAAY,UAAU,YAAY,CAC/B,YAAY,SAAS;EACtC,MAAM,MAAM,GAAG,QAAQ,GAAG;EAC1B,MAAM,UAAU,MAAM,IAAI,MAAM,IAAI;AAEpC,UAAQ,gBACN,uBAAO,IAAI,MAAM,wBAAwB,WAAW,IAAI,QAAQ,OAAO,UAAU,CAAC;AACpF,UAAQ,kBAAkB,SAAS;GACnC;;;;;AAMJ,eAAe,SACb,IACA,SACA,YAC6B;AAC7B,QAAO,IAAI,SAAS,SAAS,WAAW;EAEtC,MAAM,QADK,GAAG,YAAY,UAAU,WAAW,CAC9B,YAAY,SAAS;EACtC,MAAM,MAAM,GAAG,QAAQ,GAAG;EAC1B,MAAM,UAAU,MAAM,IAAI,IAAI;AAE9B,UAAQ,gBACN,uBAAO,IAAI,MAAM,uBAAuB,WAAW,IAAI,QAAQ,OAAO,UAAU,CAAC;AACnF,UAAQ,kBAAkB,QAAQ,QAAQ,UAAU,KAAK;GACzD;;;;;AAMJ,eAAe,eAAe,IAAiB,SAAgC;CAE7E,MAAM,WAAW,MAAM,YAAY,IAAI,QAAQ;AAE/C,QAAO,IAAI,SAAS,SAAS,WAAW;EACtC,MAAM,KAAK,GAAG,YAAY,CAAC,aAAa,SAAS,EAAE,YAAY;AAG/D,KAAG,YAAY,YAAY,CAAC,OAAO,QAAQ;AAG3C,MAAI,UAAU;GACZ,MAAM,cAAc,KAAK,KAAK,SAAS,aAAa,SAAS,UAAU;GACvE,MAAM,aAAa,GAAG,YAAY,SAAS;AAC3C,QAAK,IAAI,IAAI,GAAG,IAAI,aAAa,IAC/B,YAAW,OAAO,GAAG,QAAQ,GAAG,IAAI;;AAIxC,KAAG,mBAAmB,SAAS;AAC/B,KAAG,gBAAgB,uBAAO,IAAI,MAAM,+BAA+B,GAAG,OAAO,UAAU,CAAC;GACxF;;;;;;AAOJ,eAAsB,qBACpB,KACA,SACA,UAQI,EAAE,EACgB;CACtB,MAAM,EAAE,YAAY,WAAW;AAG/B,kBAAiB,eAAe,QAAQ;CAExC,MAAM,KAAK,MAAM,gBAAgB;AAEjC,KAAI;EAEF,IAAI,WAAW,MAAM,YAAY,IAAI,QAAQ;EAG7C,MAAM,eAAe,MAAM,MAAM,KAAK;GAAE,QAAQ;GAAQ;GAAQ,CAAC;AACjE,MAAI,CAAC,aAAa,GAChB,OAAM,IAAI,MAAM,wBAAwB,aAAa,OAAO,GAAG,aAAa,aAAa;EAG3F,MAAM,gBAAgB,SAAS,aAAa,QAAQ,IAAI,iBAAiB,IAAI,KAAK,GAAG;EACrF,MAAM,OAAO,aAAa,QAAQ,IAAI,OAAO;EAC7C,MAAM,eAAe,aAAa,QAAQ,IAAI,gBAAgB;AAE9D,MAAI,CAAC,cACH,OAAM,IAAI,MAAM,wCAAwC;AAI1D,MAAI,YAAY,SAAS,SAAS,MAAM;AACtC,WAAQ,KAAK,SAAS,QAAQ,4DAA4D;AAC1F,SAAM,eAAe,IAAI,QAAQ;AACjC,cAAW;;AAMb,MAAI,EAFkB,iBAAiB,UAEnB;AAElB,WAAQ,KAAK,6CAA6C,QAAQ,2BAA2B;AAC7F,MAAG,OAAO;GAEV,MAAM,WAAW,MAAM,MAAM,KAAK,EAAE,QAAQ,CAAC;AAC7C,OAAI,CAAC,SAAS,GAAI,OAAM,IAAI,MAAM,oBAAoB,SAAS,SAAS;AAExE,oBAAiB,WAAW,QAAQ;GACpC,MAAM,SAAS,MAAM,SAAS,aAAa;AAC3C,oBAAiB,SAAS,QAAQ;AAClC,UAAO;;EAIT,MAAM,cAAc,KAAK,KAAK,gBAAgB,iBAAiB;AAE/D,MAAI,CAAC,UAAU;AACb,cAAW;IACT;IACA;IACA;IACA,YAAY;IACZ,WAAW;IACX,iBAAiB,EAAE;IACnB,WAAW,KAAK,KAAK;IACrB,WAAW,KAAK,KAAK;IACtB;AACD,SAAM,aAAa,IAAI,SAAS;;AAIlC,OAAK,IAAI,IAAI,GAAG,IAAI,aAAa,KAAK;AACpC,OAAI,QAAQ,QACV,OAAM,IAAI,MAAM,mBAAmB;AAIrC,OAAI,SAAS,gBAAgB,SAAS,EAAE,EAAE;IACxC,MAAMC,oBAAmB,SAAS,gBAAgB,SAAS,cAAe;AAC1E,iBAAa;KACX,OAAO;KACP;KACA,YAAY;KACZ,SAAS,KAAK,MAAOA,oBAAkB,gBAAiB,IAAI;KAC7D,CAAC;AACF;;GAGF,MAAM,QAAQ,IAAI;GAClB,MAAM,MAAM,KAAK,IAAI,QAAQ,mBAAmB,GAAG,gBAAgB,EAAE;GAGrE,MAAM,WAAW,MAAM,MAAM,KAAK;IAChC,SAAS,EAAE,OAAO,SAAS,MAAM,GAAG,OAAO;IAC3C;IACD,CAAC;AAEF,OAAI,SAAS,WAAW,IACtB,OAAM,IAAI,MAAM,yBAAyB,SAAS,OAAO,iBAAiB;GAG5E,MAAM,YAAY,MAAM,SAAS,aAAa;AAG9C,SAAM,UAAU,IAAI,SAAS,GAAG,UAAU;AAG1C,YAAS,gBAAgB,KAAK,EAAE;AAChC,YAAS,YAAY,KAAK,KAAK;AAC/B,SAAM,aAAa,IAAI,SAAS;GAGhC,MAAM,kBAAkB,SAAS,gBAAgB,SAAS;AAC1D,oBAAiB,eAAe,SAAS;IAAE;IAAiB,YAAY;IAAe,CAAC;AAExF,gBAAa;IACX,OAAO;IACP,iBAAiB,KAAK,IAAI,iBAAiB,cAAc;IACzD,YAAY;IACZ,SAAS,KAAK,MAAO,SAAS,gBAAgB,SAAS,cAAe,IAAI;IAC3E,CAAC;AAIF,YAAS,OAAO;;AAIlB,mBAAiB,WAAW,QAAQ;AACpC,eAAa;GACX,OAAO;GACP,iBAAiB;GACjB,YAAY;GACZ,SAAS;GACV,CAAC;EAIF,MAAM,cAAc,IAAI,YAAY,cAAc;EAClD,MAAM,YAAY,IAAI,WAAW,YAAY;AAE7C,OAAK,IAAI,IAAI,GAAG,IAAI,aAAa,KAAK;GACpC,MAAM,QAAQ,MAAM,SAAS,IAAI,SAAS,EAAE;AAC5C,OAAI,CAAC,MACH,OAAM,IAAI,MAAM,iBAAiB,EAAE,kBAAkB;GAGvD,MAAM,SAAS,IAAI;AACnB,aAAU,IAAI,IAAI,WAAW,MAAM,EAAE,OAAO;;AAO9C,QAAM,eAAe,IAAI,QAAQ;AACjC,KAAG,OAAO;AAEV,mBAAiB,SAAS,QAAQ;AAClC,SAAO;UACA,OAAO;AACd,mBAAiB,SAAS,QAAQ;AAClC,KAAG,OAAO;AACV,QAAM;;;;;;AAOV,eAAsB,sBAAsB,SAKzC;AACD,KAAI;EACF,MAAM,KAAK,MAAM,gBAAgB;EACjC,MAAM,WAAW,MAAM,YAAY,IAAI,QAAQ;AAC/C,KAAG,OAAO;AAEV,MAAI,CAAC,SACH,QAAO,EAAE,YAAY,OAAO;EAG9B,MAAM,cAAc,KAAK,KAAK,SAAS,aAAa,SAAS,UAAU;EACvE,MAAM,kBAAkB,SAAS,gBAAgB;AAEjD,MAAI,kBAAkB,YACpB,QAAO;GACL,YAAY;GACZ,iBAAiB,kBAAkB,SAAS;GAC5C,YAAY,SAAS;GACrB,SAAS,KAAK,MAAO,kBAAkB,cAAe,IAAI;GAC3D;AAGH,SAAO,EAAE,YAAY,OAAO;SACtB;AACN,SAAO,EAAE,YAAY,OAAO;;;;;;AAOhC,eAAsB,wBAAwB,SAAgC;AAC5E,KAAI;EACF,MAAM,KAAK,MAAM,gBAAgB;AACjC,QAAM,eAAe,IAAI,QAAQ;AACjC,KAAG,OAAO;SACJ;;;;;;AASV,eAAsB,kBAAkB,aAAqB,KAM1D;AACD,KAAI,OAAO,cAAc,eAAe,CAAC,UAAU,SAAS,SAC1D,QAAO;EACL,IAAI;EACJ,aAAa;EACb,QAAQ;EACR,SAAS;EACT,SAAS;EACV;AAGH,KAAI;EACF,MAAM,EAAE,OAAO,UAAU,MAAM,UAAU,QAAQ,UAAU;EAC3D,MAAM,UAAU,KAAK,OAAO,SAAS,KAAK,IAAU;EACpD,MAAM,SAAS,KAAK,OAAO,SAAS,KAAK,IAAU;EACnD,MAAM,cAAc,UAAU;AAE9B,MAAI,cAAc,WAChB,QAAO;GACL,IAAI;GACJ;GACA;GACA;GACA,SAAS,QAAQ,WAAW,cAAc,YAAY;GACvD;AAGH,SAAO;GAAE,IAAI;GAAM;GAAa;GAAQ;GAAS;UAC1CC,GAAQ;AACf,SAAO;GACL,IAAI;GACJ,aAAa;GACb,QAAQ;GACR,SAAS;GACT,SAAS,yBAAyB,EAAE;GACrC"}

package/dist/cli.mjs

@@ -1,11 +1,11 @@
 #!/usr/bin/env node
 import { t as __require } from "./chunk-B9cbKln6.mjs";
-import { n as BUILTIN_MODELS, r as DEFAULT_MODEL, t as Gerbil } from "./gerbil-BHrJJIa4.mjs";
+import { n as BUILTIN_MODELS, r as DEFAULT_MODEL, t as Gerbil } from "./gerbil-DNniplr4.mjs";
 import "./utils-DKO55ZmZ.mjs";
-import "./one-liner-DnQn7HJK.mjs";
+import "./one-liner-JhdIPxzF.mjs";
 import { n as isArchitectureSupported } from "./architectures-C1I5V3Dt.mjs";
-import { A as useSkill, E as getSkillInfo, O as listSkills, a as summarize, f as explain, h as commit, s as review, y as loadProjectSkills } from "./skills-DX8D59UH.mjs";
-import { r as startMCPServer } from "./mcp-1DaMsaBc.mjs";
+import { A as useSkill, E as getSkillInfo, O as listSkills, a as summarize, f as explain, h as commit, s as review, y as loadProjectSkills } from "./skills-CU694Dc8.mjs";
+import { r as startMCPServer } from "./mcp-D2vvH1Xc.mjs";
 import { a as getToolDefinitions, c as setToolContext, i as getTool, n as executeToolCall, o as loadProjectTools, r as formatToolsForPrompt, s as parseToolCall } from "./tools-DQ1mPUw5.mjs";
 import fs, { existsSync, readFileSync, unlinkSync } from "node:fs";
 import os, { tmpdir } from "node:os";
@@ -24,7 +24,7 @@ import TextInput from "ink-text-input";
 import http from "node:http";
 
 //#region package.json
-var version = "1.0.0";
+var version = "1.0.1";
 
 //#endregion
 //#region src/cli/repl/auto-update.ts
@@ -7701,7 +7701,7 @@ function App({ initialView = "menu" } = {}) {
 		return () => {
 			mounted = false;
 			if (gerbilRef.current) {
-				import("./repl-jV5gcJFA.mjs").then(({ setCleanupPromise: setCleanupPromise$1 }) => {
+				import("./repl-BDRkwPGX.mjs").then(({ setCleanupPromise: setCleanupPromise$1 }) => {
 					setCleanupPromise$1(gerbilRef.current?.dispose(true) ?? Promise.resolve());
 				});
 				gerbilRef.current = null;
@@ -8997,7 +8997,7 @@ program.command("serve").description("Start Gerbil server (use --mcp or --http f
 	}
 	if (opts.mcp) await startMCPServer({ model: opts.model });
 	else {
-		const { Gerbil: Gerbil$1 } = await import("./gerbil-DomNfIr1.mjs");
+		const { Gerbil: Gerbil$1 } = await import("./gerbil-CTZUa8EZ.mjs");
 		const g = new Gerbil$1();
 		const spinner = ora("Loading model...").start();
 		await g.loadModel(opts.model);