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

package/dist/gerbil-BfnsFWRE.mjs.map

@@ -1 +0,0 @@
-{"version":3,"file":"gerbil-BfnsFWRE.mjs","names":["pipeline","rawPipeline","isBrowser","tfDevice: \"webgpu\" | \"wasm\" | \"cpu\"","chromeErr: any","tokenQueue: string[]","resolveNext: ((value: string | null) => void) | null","results: EmbedResult[]","messages: Array<{ role: string; content: string }>"],"sources":["../src/core/gerbil.ts"],"sourcesContent":["/**\n * Gerbil - Local GPU-accelerated LLM inference\n */\n\nimport {\n  AutoModelForCausalLM,\n  AutoTokenizer,\n  env,\n  type FeatureExtractionPipeline,\n  type PreTrainedTokenizer,\n  pipeline as rawPipeline,\n  type TextGenerationPipeline,\n} from \"@huggingface/transformers\";\n\n// Wrapper to avoid TypeScript complexity issues with transformers.js types\nconst pipeline = rawPipeline as (task: string, model: string, options?: any) => Promise<any>;\n\n// Suppress noisy transformers.js warnings during model loading\nfunction suppressNoisyWarnings<T>(fn: () => Promise<T>): Promise<T> {\n  const originalWarn = console.warn;\n  console.warn = (...args: any[]) => {\n    const msg = args[0]?.toString?.() || \"\";\n    // Suppress \"Unable to determine content-length\" warnings from transformers.js\n    if (msg.includes(\"content-length\") || msg.includes(\"Unable to determine\")) {\n      return;\n    }\n    originalWarn.apply(console, args);\n  };\n\n  return fn().finally(() => {\n    console.warn = originalWarn;\n  });\n}\n\nimport {\n  BUILTIN_MODELS,\n  createExternalModelConfig,\n  getModelConfig,\n  resolveModel,\n} from \"./models.js\";\nimport type {\n  EmbedOptions,\n  EmbedResult,\n  GenerateOptions,\n  GenerateResult,\n  GerbilConfig,\n  JsonOptions,\n  LoadOptions,\n  ModelConfig,\n  SessionStats,\n  SystemInfo,\n} from \"./types.js\";\n\nimport { extractJson, zodToJsonSchema } from \"./utils.js\";\n\n// Configure transformers.js based on environment\nconst isBrowser = typeof window !== \"undefined\";\nenv.allowLocalModels = !isBrowser; // false in browser (fetch from HuggingFace)\nenv.useBrowserCache = isBrowser; // true in browser (cache in IndexedDB)\n\n// ============================================\n// Gerbil Class\n// ============================================\n\n// WebGPU initialization state for Node.js\nlet webgpuInitialized = false;\nlet webgpuAvailable = false;\n\n/**\n * Initialize WebGPU for Node.js environments\n * Called automatically before model loading\n */\nasync function initNodeWebGPU(): Promise<boolean> {\n  if (webgpuInitialized) {\n    return webgpuAvailable;\n  }\n  webgpuInitialized = true;\n\n  // Skip if in browser (already has WebGPU)\n  if (typeof window !== \"undefined\") {\n    webgpuAvailable = \"gpu\" in navigator;\n    return webgpuAvailable;\n  }\n\n  // Try to initialize WebGPU in Node.js via Dawn\n  // Use Function constructor to hide import from bundlers\n  try {\n    const dynamicImport = new Function(\"specifier\", \"return import(specifier)\");\n    const webgpuModule = await dynamicImport(\"webgpu\");\n    const { create, globals } = webgpuModule;\n\n    // Extend globalThis with WebGPU globals\n    Object.assign(globalThis, globals);\n\n    // Create navigator.gpu\n    if (!(globalThis as any).navigator) {\n      (globalThis as any).navigator = {};\n    }\n    (globalThis as any).navigator.gpu = create([]);\n\n    webgpuAvailable = true;\n  } catch {\n    // WebGPU not available, will fall back to CPU\n    webgpuAvailable = false;\n  }\n\n  return webgpuAvailable;\n}\n\n// ChromeGPUBackend is dynamically imported only in Node.js to avoid bundling puppeteer in browser\ntype ChromeGPUBackendType = import(\"./chrome-backend.js\").ChromeGPUBackend;\n\nexport class Gerbil {\n  private generator: TextGenerationPipeline | null = null;\n  private tokenizer: PreTrainedTokenizer | null = null;\n  private model: any = null; // AutoModelForCausalLM instance\n  private embedder: FeatureExtractionPipeline | null = null;\n  private currentModel: string | null = null;\n  private modelConfig: ModelConfig | null = null;\n  private readonly config: GerbilConfig;\n  private stats: SessionStats;\n  private useDirect = false; // Use direct model loading (for WebGPU)\n  private chromeBackend: ChromeGPUBackendType | null = null; // Chrome backend for Node.js WebGPU\n  private _deviceMode: \"webgpu\" | \"cpu\" | \"wasm\" = \"cpu\"; // Track which backend is active\n\n  constructor(config: GerbilConfig = {}) {\n    this.config = config;\n    this.stats = {\n      prompts: 0,\n      tokensIn: 0,\n      tokensOut: 0,\n      avgSpeed: 0,\n      totalTime: 0,\n      cacheHits: 0,\n      cacheMisses: 0,\n    };\n  }\n\n  // ============================================\n  // Static Methods\n  // ============================================\n\n  static listModels(): ModelConfig[] {\n    return Object.values(BUILTIN_MODELS);\n  }\n\n  static getModel(modelId: string): ModelConfig | undefined {\n    return BUILTIN_MODELS[modelId];\n  }\n\n  // ============================================\n  // Model Loading\n  // ============================================\n\n  /**\n   * Load a model\n   *\n   * @example\n   * ```ts\n   * // Built-in model\n   * await g.loadModel(\"qwen3-0.6b\");\n   *\n   * // HuggingFace model\n   * await g.loadModel(\"hf:microsoft/Phi-3-mini\");\n   *\n   * // Local model\n   * await g.loadModel(\"file:./models/my-model\");\n   * ```\n   */\n  async loadModel(modelId = \"qwen3-0.6b\", options: LoadOptions = {}): Promise<void> {\n    // Initialize WebGPU for Node.js if needed\n    await initNodeWebGPU();\n\n    const source = resolveModel(modelId);\n    const { onProgress, device = \"auto\", dtype: userDtype } = options;\n\n    // Get or create model config\n    let config = getModelConfig(modelId);\n    if (!config) {\n      config = createExternalModelConfig(modelId, source.path);\n    }\n\n    onProgress?.({ status: `Loading ${modelId}...` });\n\n    // Map device to transformers.js device\n    // Browser supports: webgpu, wasm (no cpu)\n    // Node supports: webgpu, cpu\n    const isBrowser = typeof window !== \"undefined\";\n    const fallbackDevice = isBrowser ? \"wasm\" : \"cpu\";\n    let tfDevice: \"webgpu\" | \"wasm\" | \"cpu\" = fallbackDevice;\n    if (device === \"webgpu\" || device === \"gpu\" || device === \"auto\") {\n      tfDevice = \"webgpu\";\n    }\n\n    // Use q4f16 for WebGPU (required for Qwen3), q4 for CPU/WASM\n    const dtype = userDtype ?? (tfDevice === \"webgpu\" ? \"q4f16\" : \"q4\");\n\n    // Track if we're still in loading phase (to suppress progress during inference)\n    let isLoading = true;\n    let lastFile = \"\";\n    let lastPct = -1;\n\n    const progressCallback = (progress: any) => {\n      if (!isLoading) {\n        return; // Suppress progress after initial load\n      }\n\n      if (progress.status === \"progress\" && progress.file) {\n        const pct = Math.round(progress.progress || 0);\n        // Only report if file changed or progress increased significantly\n        if (progress.file !== lastFile || pct >= lastPct + 5) {\n          lastFile = progress.file;\n          lastPct = pct;\n          onProgress?.({\n            status: `Downloading ${progress.file}`,\n            progress: pct,\n            file: progress.file,\n          });\n        }\n      }\n    };\n\n    try {\n      // Use direct model loading for browser WebGPU (like qwen-web does)\n      // This bypasses pipeline() which may have different ONNX session config\n      if (isBrowser && tfDevice === \"webgpu\") {\n        onProgress?.({ status: \"Loading tokenizer...\" });\n        this.tokenizer = (await suppressNoisyWarnings(() =>\n          AutoTokenizer.from_pretrained(source.path, {\n            progress_callback: progressCallback,\n          }),\n        )) as PreTrainedTokenizer;\n\n        onProgress?.({ status: \"Loading model...\" });\n        this.model = await suppressNoisyWarnings(() =>\n          AutoModelForCausalLM.from_pretrained(source.path, {\n            dtype,\n            device: tfDevice,\n            progress_callback: progressCallback,\n          }),\n        );\n\n        this.useDirect = true;\n        this._deviceMode = \"webgpu\";\n        isLoading = false;\n        this.currentModel = modelId;\n        this.modelConfig = config;\n        onProgress?.({ status: \"Ready (WebGPU)!\" });\n      } else if (!isBrowser && tfDevice === \"webgpu\") {\n        // Node.js + WebGPU: Use Chrome backend for real GPU acceleration\n        onProgress?.({ status: \"Starting Chrome WebGPU backend...\" });\n\n        // Dynamic import to avoid bundling puppeteer in browser builds\n        const { ChromeGPUBackend } = await import(\"./chrome-backend.js\");\n        this.chromeBackend = await ChromeGPUBackend.create({\n          modelId: source.path,\n          onProgress,\n        });\n\n        this.useDirect = false;\n        this._deviceMode = \"webgpu\";\n        isLoading = false;\n        this.currentModel = modelId;\n        this.modelConfig = config;\n        // Ready status is set by ChromeGPUBackend\n      } else {\n        // Use pipeline for CPU / WASM\n        const pipelineOptions = {\n          dtype,\n          device: tfDevice,\n          progress_callback: progressCallback,\n        };\n        this.generator = (await suppressNoisyWarnings(() =>\n          pipeline(\"text-generation\", source.path, pipelineOptions as any),\n        )) as TextGenerationPipeline;\n\n        this.useDirect = false;\n        this._deviceMode = tfDevice as \"cpu\" | \"wasm\";\n        isLoading = false;\n        this.currentModel = modelId;\n        this.modelConfig = config;\n        onProgress?.({ status: `Ready (${tfDevice.toUpperCase()})!` });\n      }\n    } catch (err) {\n      // Fallback to CPU/WASM if GPU fails (silently)\n      if (tfDevice !== fallbackDevice) {\n        onProgress?.({ status: `Using ${fallbackDevice.toUpperCase()}...` });\n\n        // Clean up Chrome backend if it was partially initialized\n        if (this.chromeBackend) {\n          await this.chromeBackend.dispose();\n          this.chromeBackend = null;\n        }\n\n        // Fallback always uses pipeline (WASM/CPU don't need direct loading)\n        this.generator = (await suppressNoisyWarnings(() =>\n          pipeline(\"text-generation\", source.path, {\n            dtype: \"q4\",\n            device: fallbackDevice,\n            progress_callback: progressCallback,\n          } as any),\n        )) as TextGenerationPipeline;\n\n        this.useDirect = false;\n        this._deviceMode = fallbackDevice as \"cpu\" | \"wasm\";\n        isLoading = false;\n        this.currentModel = modelId;\n        this.modelConfig = config;\n        onProgress?.({ status: `Ready (${fallbackDevice.toUpperCase()})!` });\n      } else {\n        throw err;\n      }\n    }\n  }\n\n  /**\n   * Check if a model is loaded\n   */\n  isLoaded(): boolean {\n    return (\n      this.generator !== null ||\n      (this.useDirect && this.model !== null) ||\n      this.chromeBackend !== null\n    );\n  }\n\n  /**\n   * Get current model info\n   */\n  getModelInfo(): ModelConfig | null {\n    return this.modelConfig;\n  }\n\n  /**\n   * Get current device mode (webgpu, cpu, or wasm)\n   */\n  getDeviceMode(): \"webgpu\" | \"cpu\" | \"wasm\" {\n    return this._deviceMode;\n  }\n\n  /**\n   * Get dtype used for current model\n   */\n  getDtype(): string {\n    // WebGPU uses q4f16, CPU/WASM use q4\n    return this._deviceMode === \"webgpu\" ? \"q4f16\" : \"q4\";\n  }\n\n  /**\n   * Get Chrome backend status (if using WebGPU via Chrome)\n   */\n  getChromeStatus(): {\n    pid: number | null;\n    port: number;\n    modelId: string;\n    startedAt: Date | null;\n  } | null {\n    if (!this.chromeBackend) {\n      return null;\n    }\n    return this.chromeBackend.getStatus();\n  }\n\n  /**\n   * Get Chrome memory usage (if using WebGPU via Chrome)\n   * Returns JS heap memory in bytes\n   */\n  async getChromeMemory(): Promise<{ jsHeapUsed: number; jsHeapTotal: number } | null> {\n    if (!this.chromeBackend) {\n      return null;\n    }\n    return this.chromeBackend.getMemoryUsage();\n  }\n\n  /**\n   * Get memory usage in GB (if using WebGPU via Chrome)\n   */\n  async getMemoryUsage(): Promise<{ usedGB: number; totalGB: number; usedPercent: number } | null> {\n    if (!this.chromeBackend) {\n      return null;\n    }\n    return this.chromeBackend.getMemoryStats();\n  }\n\n  /**\n   * Clear KV cache to free memory\n   * This will reset the conversation context but free up memory\n   */\n  async clearCache(): Promise<void> {\n    if (this.chromeBackend) {\n      await this.chromeBackend.reset();\n    }\n  }\n\n  /**\n   * Check memory usage and cleanup if needed\n   * @param thresholdGB Memory threshold in GB (default: 8)\n   * @returns true if cleanup was performed\n   */\n  async checkMemoryAndCleanup(thresholdGB = 8): Promise<boolean> {\n    if (!this.chromeBackend) {\n      return false;\n    }\n    return this.chromeBackend.checkMemoryAndCleanup(thresholdGB);\n  }\n\n  // ============================================\n  // Text Generation\n  // ============================================\n\n  /**\n   * Generate text\n   */\n  async generate(prompt: string, options: GenerateOptions = {}): Promise<GenerateResult> {\n    if (!this.isLoaded()) {\n      // Auto-load default model\n      await this.loadModel(this.config.model || \"qwen3-0.6b\");\n    }\n\n    const {\n      maxTokens = 256,\n      temperature = 0.7,\n      topP = 0.9,\n      topK = 50,\n      thinking = false,\n      system,\n    } = options;\n\n    const startTime = performance.now();\n\n    try {\n      let rawText = \"\";\n\n      if (this.chromeBackend) {\n        // Chrome backend approach (for Node.js WebGPU via Chrome)\n        try {\n          rawText = await this.chromeBackend.generate(prompt, {\n            maxTokens,\n            temperature,\n            topP,\n            topK,\n            thinking,\n            system,\n            // Wrap onToken to match Gerbil's simpler signature\n            onToken: options.onToken ? (t) => options.onToken!(t.text) : undefined,\n          });\n        } catch (chromeErr: any) {\n          // If Chrome died (OOM, crash), fall back to CPU silently\n          if (chromeErr?.message === \"CHROME_BACKEND_DEAD\" || !this.chromeBackend?.isAlive()) {\n            await this.chromeBackend?.dispose().catch(() => {});\n            this.chromeBackend = null;\n            this._deviceMode = \"cpu\";\n            // Load CPU fallback and retry\n            const modelPath = this.currentModel || \"qwen3-0.6b\";\n            this.generator = (await pipeline(\"text-generation\", modelPath, {\n              dtype: \"q4\",\n              device: \"cpu\",\n            } as any)) as TextGenerationPipeline;\n            // Retry with CPU\n            return this.generate(prompt, options);\n          }\n          throw chromeErr;\n        }\n      } else if (this.useDirect && this.model && this.tokenizer) {\n        // Direct model approach (for browser WebGPU)\n        const messages = this.buildMessages(prompt, { ...options, thinking });\n\n        const inputs = (this.tokenizer as any).apply_chat_template(messages, {\n          add_generation_prompt: true,\n          return_dict: true,\n          enable_thinking: thinking, // Qwen3 thinking mode\n        });\n\n        const output = await this.model.generate({\n          ...inputs,\n          max_new_tokens: maxTokens,\n          temperature: temperature > 0 ? temperature : undefined,\n          top_p: topP,\n          top_k: topK,\n          do_sample: temperature > 0,\n        });\n\n        // Get input length to extract only generated tokens\n        const inputLength = inputs.input_ids.dims?.[1] || inputs.input_ids.data?.length || 0;\n\n        // Slice output tensor to get only new tokens (skip prompt)\n        const outputTokens = output.slice(null, [inputLength, null]);\n        const decoded = this.tokenizer.batch_decode(outputTokens, {\n          skip_special_tokens: true,\n        });\n\n        rawText = decoded[0] || \"\";\n\n        // If we still have prompt artifacts, extract assistant response\n        if (rawText.toLowerCase().includes(\"assistant\")) {\n          const match = rawText.match(/assistant[:\\s]*([\\s\\S]*)/i);\n          if (match) {\n            rawText = match[1].trim();\n          }\n        }\n      } else if (this.generator) {\n        // Pipeline approach (for Node.js / CPU / WASM)\n        const formattedPrompt = this.formatPrompt(prompt, { ...options, thinking });\n\n        const output = await this.generator(formattedPrompt, {\n          max_new_tokens: maxTokens,\n          temperature,\n          top_p: topP,\n          top_k: topK,\n          do_sample: temperature > 0,\n          return_full_text: false,\n        });\n\n        // Extract text from pipeline output\n        if (Array.isArray(output) && output[0]) {\n          const result = output[0] as any;\n          if (Array.isArray(result.generated_text)) {\n            const last = result.generated_text.at(-1);\n            rawText = last?.content || \"\";\n          } else {\n            rawText = result.generated_text || \"\";\n          }\n        }\n      } else {\n        throw new Error(\"No model loaded\");\n      }\n\n      const endTime = performance.now();\n      const totalTime = endTime - startTime;\n\n      rawText = this.cleanOutput(rawText);\n\n      // Always parse thinking to strip <think> tags from output\n      // (model may generate them even without thinking mode enabled)\n      const { thinking: thinkingText, response } = this.parseThinking(rawText);\n\n      // Only include thinking in result if mode was enabled\n      const finalThinking = thinking ? thinkingText : undefined;\n\n      const tokensGenerated = Math.ceil(response.length / 4);\n\n      // Update stats\n      this.stats.prompts++;\n      this.stats.tokensOut += tokensGenerated;\n      this.stats.totalTime += totalTime;\n      this.stats.avgSpeed = (this.stats.tokensOut / this.stats.totalTime) * 1000;\n\n      return {\n        text: response,\n        thinking: finalThinking,\n        tokensGenerated,\n        tokensPerSecond: (tokensGenerated / totalTime) * 1000,\n        totalTime,\n        finishReason: \"stop\",\n        provider: \"local\",\n        cached: false,\n      };\n    } catch (_error) {\n      return {\n        text: \"\",\n        tokensGenerated: 0,\n        tokensPerSecond: 0,\n        totalTime: performance.now() - startTime,\n        finishReason: \"error\",\n        provider: \"local\",\n        cached: false,\n      };\n    }\n  }\n\n  /**\n   * Stream text generation (simulated token-by-token)\n   *\n   * Note: Yields the raw output including <think> tags if thinking mode is enabled.\n   * The final result has parsed thinking separated out.\n   */\n  async *stream(\n    prompt: string,\n    options: GenerateOptions = {},\n  ): AsyncGenerator<string, GenerateResult, unknown> {\n    if (!this.isLoaded()) {\n      await this.loadModel(this.config.model || \"qwen3-0.6b\");\n    }\n\n    const startTime = performance.now();\n\n    // For Chrome backend, use real streaming via onToken callback\n    if (this.chromeBackend) {\n      let fullText = \"\";\n      const tokenQueue: string[] = [];\n      let resolveNext: ((value: string | null) => void) | null = null;\n      let done = false;\n\n      // Start generation with streaming callback\n      const generatePromise = this.chromeBackend\n        .generate(prompt, {\n          ...options,\n          onToken: (token) => {\n            fullText += token.text;\n            if (resolveNext) {\n              resolveNext(token.text);\n              resolveNext = null;\n            } else {\n              tokenQueue.push(token.text);\n            }\n          },\n        })\n        .then(() => {\n          done = true;\n          if (resolveNext) {\n            resolveNext(null);\n          }\n        })\n        .catch((err) => {\n          done = true;\n          if (resolveNext) {\n            resolveNext(null);\n          }\n          throw err;\n        });\n\n      // Yield tokens as they arrive\n      while (!done || tokenQueue.length > 0) {\n        if (tokenQueue.length > 0) {\n          const token = tokenQueue.shift()!;\n          yield token;\n          options.onToken?.(token);\n        } else if (!done) {\n          const token = await new Promise<string | null>((resolve) => {\n            resolveNext = resolve;\n          });\n          if (token) {\n            yield token;\n            options.onToken?.(token);\n          }\n        }\n      }\n\n      await generatePromise;\n\n      const { thinking: thinkingText, response } = this.parseThinking(fullText);\n      const tokensGenerated = Math.ceil(response.length / 4);\n      const totalTime = performance.now() - startTime;\n\n      return {\n        text: response,\n        thinking: options.thinking ? thinkingText : undefined,\n        tokensGenerated,\n        totalTime,\n        tokensPerSecond: (tokensGenerated / totalTime) * 1000,\n        finishReason: \"stop\" as const,\n      };\n    }\n\n    // For pipeline/direct model, use fake streaming (generate then yield)\n    const result = await this.generateRaw(prompt, options);\n\n    // Yield word by word for more accurate token simulation\n    // (actual tokens average ~4 chars, words are a reasonable approximation)\n    const words = result.rawText.split(/(\\s+)/);\n    for (const word of words) {\n      if (word) {\n        yield word;\n        options.onToken?.(word);\n      }\n    }\n\n    return result.result;\n  }\n\n  /**\n   * Internal: Generate with raw text access for streaming\n   */\n  private async generateRaw(\n    prompt: string,\n    options: GenerateOptions = {},\n  ): Promise<{ rawText: string; result: GenerateResult }> {\n    const { maxTokens = 256, temperature = 0.7, topP = 0.9, topK = 50, thinking = false } = options;\n\n    const startTime = performance.now();\n    const formattedPrompt = this.formatPrompt(prompt, { ...options, thinking });\n\n    try {\n      const output = await this.generator!(formattedPrompt, {\n        max_new_tokens: maxTokens,\n        temperature,\n        top_p: topP,\n        top_k: topK,\n        do_sample: temperature > 0,\n        return_full_text: false,\n      });\n\n      const endTime = performance.now();\n      const totalTime = endTime - startTime;\n\n      // Extract text from output\n      let rawText = \"\";\n      if (Array.isArray(output) && output[0]) {\n        const result = output[0] as any;\n        if (Array.isArray(result.generated_text)) {\n          const last = result.generated_text.at(-1);\n          rawText = last?.content || \"\";\n        } else {\n          rawText = result.generated_text || \"\";\n        }\n      }\n\n      rawText = this.cleanOutput(rawText);\n      const { thinking: thinkingText, response } = this.parseThinking(rawText);\n      const finalThinking = thinking ? thinkingText : undefined;\n      const tokensGenerated = Math.ceil(response.length / 4);\n\n      // Update stats\n      this.stats.prompts++;\n      this.stats.tokensOut += tokensGenerated;\n      this.stats.totalTime += totalTime;\n      this.stats.avgSpeed = (this.stats.tokensOut / this.stats.totalTime) * 1000;\n\n      return {\n        rawText,\n        result: {\n          text: response,\n          thinking: finalThinking,\n          tokensGenerated,\n          tokensPerSecond: (tokensGenerated / totalTime) * 1000,\n          totalTime,\n          finishReason: \"stop\",\n          provider: \"local\",\n          cached: false,\n        },\n      };\n    } catch (_error) {\n      return {\n        rawText: \"\",\n        result: {\n          text: \"\",\n          tokensGenerated: 0,\n          tokensPerSecond: 0,\n          totalTime: performance.now() - startTime,\n          finishReason: \"error\",\n          provider: \"local\",\n          cached: false,\n        },\n      };\n    }\n  }\n\n  // ============================================\n  // Structured Output (JSON)\n  // ============================================\n\n  /**\n   * Generate structured JSON output\n   */\n  async json<T>(prompt: string, options: JsonOptions<T>): Promise<T> {\n    const { schema, retries = 3, temperature = 0.3 } = options;\n\n    const systemPrompt = `You are a JSON generator. You MUST respond with valid JSON only.\nNo explanations, no markdown, no code blocks. Just pure JSON.\nThe JSON must conform to this schema: ${JSON.stringify(zodToJsonSchema(schema))}`;\n\n    for (let attempt = 0; attempt < retries; attempt++) {\n      const result = await this.generate(prompt, {\n        system: options.system || systemPrompt,\n        temperature,\n        maxTokens: 1000,\n      });\n\n      try {\n        // Try to extract JSON from response\n        const jsonStr = extractJson(result.text);\n        const parsed = JSON.parse(jsonStr);\n        const validated = schema.parse(parsed);\n        return validated;\n      } catch (error) {\n        if (attempt === retries - 1) {\n          throw new Error(`Failed to generate valid JSON after ${retries} attempts: ${error}`);\n        }\n      }\n    }\n\n    throw new Error(\"Failed to generate valid JSON\");\n  }\n\n  // ============================================\n  // Embeddings\n  // ============================================\n\n  /**\n   * Generate embeddings\n   */\n  async embed(text: string, options: EmbedOptions = {}): Promise<EmbedResult> {\n    if (!this.embedder) {\n      // Load embedding model\n      const model = options.model || \"Xenova/all-MiniLM-L6-v2\";\n      this.embedder = (await pipeline(\"feature-extraction\", model)) as FeatureExtractionPipeline;\n    }\n\n    const startTime = performance.now();\n    const output = await this.embedder(text, {\n      pooling: \"mean\",\n      normalize: options.normalize !== false,\n    });\n\n    const vector = Array.from(output.data as Float32Array);\n\n    return {\n      vector,\n      text,\n      totalTime: performance.now() - startTime,\n    };\n  }\n\n  /**\n   * Generate embeddings for multiple texts\n   */\n  async embedBatch(texts: string[], options: EmbedOptions = {}): Promise<EmbedResult[]> {\n    const results: EmbedResult[] = [];\n    for (const text of texts) {\n      results.push(await this.embed(text, options));\n    }\n    return results;\n  }\n\n  // ============================================\n  // Stats & Info\n  // ============================================\n\n  /**\n   * Get session stats\n   */\n  getStats(): SessionStats {\n    return { ...this.stats };\n  }\n\n  /**\n   * Get system info\n   */\n  getInfo(): SystemInfo {\n    return {\n      version: \"1.0.0\",\n      model: this.modelConfig,\n      device: {\n        backend: \"transformers.js\",\n        gpu: null, // TODO: detect GPU\n        vram: null,\n        status: this.isLoaded() ? \"ready\" : \"loading\",\n      },\n      context: {\n        max: this.modelConfig?.contextLength || 0,\n        used: 0,\n        available: this.modelConfig?.contextLength || 0,\n      },\n      cache: {\n        location: \"~/.gerbil/models\",\n        size: \"0 MB\",\n        modelCount: 0,\n      },\n    };\n  }\n\n  /**\n   * Reset stats\n   */\n  resetStats(): void {\n    this.stats = {\n      prompts: 0,\n      tokensIn: 0,\n      tokensOut: 0,\n      avgSpeed: 0,\n      totalTime: 0,\n      cacheHits: 0,\n      cacheMisses: 0,\n    };\n  }\n\n  // ============================================\n  // Cleanup\n  // ============================================\n\n  /**\n   * Dispose of resources\n   */\n  async dispose(): Promise<void> {\n    // Clean up Chrome backend first (most important to release resources)\n    if (this.chromeBackend) {\n      try {\n        await this.chromeBackend.dispose();\n      } catch {\n        // Ignore errors during cleanup\n      }\n      this.chromeBackend = null;\n    }\n\n    if (this.generator) {\n      if (typeof (this.generator as any).dispose === \"function\") {\n        try {\n          await (this.generator as any).dispose();\n        } catch {\n          // Ignore errors during cleanup\n        }\n      }\n      this.generator = null;\n    }\n    if (this.embedder) {\n      if (typeof (this.embedder as any).dispose === \"function\") {\n        try {\n          await (this.embedder as any).dispose();\n        } catch {\n          // Ignore errors during cleanup\n        }\n      }\n      this.embedder = null;\n    }\n    this.currentModel = null;\n    this.modelConfig = null;\n  }\n\n  // ============================================\n  // Private Methods\n  // ============================================\n\n  private formatPrompt(prompt: string, options: GenerateOptions): string {\n    const system = options.system || \"You are a helpful assistant.\";\n    const isQwen = this.currentModel?.includes(\"qwen\");\n\n    if (options.thinking && this.modelConfig?.supportsThinking) {\n      const thinkSystem = `${system}\\n\\nThink step-by-step before answering. Wrap your reasoning in <think></think> tags, then provide your answer.`;\n      return `<|im_start|>system\\n${thinkSystem}<|im_end|>\\n<|im_start|>user\\n${prompt}<|im_end|>\\n<|im_start|>assistant\\n`;\n    }\n\n    if (isQwen) {\n      return `<|im_start|>system\\n${system}<|im_end|>\\n<|im_start|>user\\n${prompt} /no_think<|im_end|>\\n<|im_start|>assistant\\n`;\n    }\n\n    return `<|im_start|>system\\n${system}<|im_end|>\\n<|im_start|>user\\n${prompt}<|im_end|>\\n<|im_start|>assistant\\n`;\n  }\n\n  private buildMessages(\n    prompt: string,\n    options: GenerateOptions,\n  ): Array<{ role: string; content: string }> {\n    const system = options.system || \"You are a helpful assistant.\";\n    const messages: Array<{ role: string; content: string }> = [];\n\n    // For direct model (WebGPU), enable_thinking is passed to apply_chat_template\n    // so we don't need to add /no_think or modify the system prompt\n    messages.push({ role: \"system\", content: system });\n    messages.push({ role: \"user\", content: prompt });\n\n    return messages;\n  }\n\n  private parseThinking(text: string): {\n    thinking?: string;\n    response: string;\n  } {\n    // Handle complete <think>...</think> blocks\n    const match = text.match(/<think>([\\s\\S]*?)<\\/think>/);\n    if (match) {\n      const thinking = match[1].trim();\n      const response = text.replace(/<think>[\\s\\S]*?<\\/think>/, \"\").trim();\n      return { thinking, response };\n    }\n\n    // Handle unclosed <think> tags (model stopped mid-thought)\n    const unclosedMatch = text.match(/<think>([\\s\\S]*)$/);\n    if (unclosedMatch) {\n      const thinking = unclosedMatch[1].trim();\n      const response = text.replace(/<think>[\\s\\S]*$/, \"\").trim();\n      return { thinking: thinking || undefined, response };\n    }\n\n    // Handle any remaining think tags\n    const response = text.replace(/<\\/?think>/g, \"\").trim();\n    return { response };\n  }\n\n  private cleanOutput(text: string): string {\n    return (\n      text\n        .replace(/<\\|im_end\\|>/g, \"\")\n        .replace(/<\\|im_start\\|>/g, \"\")\n        .replace(/<\\|endoftext\\|>/g, \"\")\n        .replace(/<\\/s>/g, \"\")\n        // Clean up artifacts from direct model output\n        .replace(/^\\/no_think\\s*/i, \"\")\n        .replace(/^assistant\\s*/i, \"\")\n        .replace(/^\\s*\\/no_think\\s*/gim, \"\")\n        .replace(/^\\s*assistant\\s*/gim, \"\")\n        // Clean up role markers that might appear\n        .replace(/^(system|user|assistant):\\s*/gim, \"\")\n        .trim()\n    );\n  }\n}\n\nexport default Gerbil;\n"],"mappings":";;;;;;;;AAeA,MAAMA,aAAWC;AAGjB,SAAS,sBAAyB,IAAkC;CAClE,MAAM,eAAe,QAAQ;AAC7B,SAAQ,QAAQ,GAAG,SAAgB;EACjC,MAAM,MAAM,KAAK,IAAI,YAAY,IAAI;AAErC,MAAI,IAAI,SAAS,iBAAiB,IAAI,IAAI,SAAS,sBAAsB,CACvE;AAEF,eAAa,MAAM,SAAS,KAAK;;AAGnC,QAAO,IAAI,CAAC,cAAc;AACxB,UAAQ,OAAO;GACf;;AAyBJ,MAAM,YAAY,OAAO,WAAW;AACpC,IAAI,mBAAmB,CAAC;AACxB,IAAI,kBAAkB;AAOtB,IAAI,oBAAoB;AACxB,IAAI,kBAAkB;;;;;AAMtB,eAAe,iBAAmC;AAChD,KAAI,kBACF,QAAO;AAET,qBAAoB;AAGpB,KAAI,OAAO,WAAW,aAAa;AACjC,oBAAkB,SAAS;AAC3B,SAAO;;AAKT,KAAI;EAGF,MAAM,EAAE,QAAQ,YADK,MADC,IAAI,SAAS,aAAa,2BAA2B,CAClC,SAAS;AAIlD,SAAO,OAAO,YAAY,QAAQ;AAGlC,MAAI,CAAE,WAAmB,UACvB,CAAC,WAAmB,YAAY,EAAE;AAEpC,EAAC,WAAmB,UAAU,MAAM,OAAO,EAAE,CAAC;AAE9C,oBAAkB;SACZ;AAEN,oBAAkB;;AAGpB,QAAO;;AAMT,IAAa,SAAb,MAAoB;CAClB,AAAQ,YAA2C;CACnD,AAAQ,YAAwC;CAChD,AAAQ,QAAa;CACrB,AAAQ,WAA6C;CACrD,AAAQ,eAA8B;CACtC,AAAQ,cAAkC;CAC1C,AAAiB;CACjB,AAAQ;CACR,AAAQ,YAAY;CACpB,AAAQ,gBAA6C;CACrD,AAAQ,cAAyC;CAEjD,YAAY,SAAuB,EAAE,EAAE;AACrC,OAAK,SAAS;AACd,OAAK,QAAQ;GACX,SAAS;GACT,UAAU;GACV,WAAW;GACX,UAAU;GACV,WAAW;GACX,WAAW;GACX,aAAa;GACd;;CAOH,OAAO,aAA4B;AACjC,SAAO,OAAO,OAAO,eAAe;;CAGtC,OAAO,SAAS,SAA0C;AACxD,SAAO,eAAe;;;;;;;;;;;;;;;;;CAsBxB,MAAM,UAAU,UAAU,cAAc,UAAuB,EAAE,EAAiB;AAEhF,QAAM,gBAAgB;EAEtB,MAAM,SAAS,aAAa,QAAQ;EACpC,MAAM,EAAE,YAAY,SAAS,QAAQ,OAAO,cAAc;EAG1D,IAAI,SAAS,eAAe,QAAQ;AACpC,MAAI,CAAC,OACH,UAAS,0BAA0B,SAAS,OAAO,KAAK;AAG1D,eAAa,EAAE,QAAQ,WAAW,QAAQ,MAAM,CAAC;EAKjD,MAAMC,cAAY,OAAO,WAAW;EACpC,MAAM,iBAAiBA,cAAY,SAAS;EAC5C,IAAIC,WAAsC;AAC1C,MAAI,WAAW,YAAY,WAAW,SAAS,WAAW,OACxD,YAAW;EAIb,MAAM,QAAQ,cAAc,aAAa,WAAW,UAAU;EAG9D,IAAI,YAAY;EAChB,IAAI,WAAW;EACf,IAAI,UAAU;EAEd,MAAM,oBAAoB,aAAkB;AAC1C,OAAI,CAAC,UACH;AAGF,OAAI,SAAS,WAAW,cAAc,SAAS,MAAM;IACnD,MAAM,MAAM,KAAK,MAAM,SAAS,YAAY,EAAE;AAE9C,QAAI,SAAS,SAAS,YAAY,OAAO,UAAU,GAAG;AACpD,gBAAW,SAAS;AACpB,eAAU;AACV,kBAAa;MACX,QAAQ,eAAe,SAAS;MAChC,UAAU;MACV,MAAM,SAAS;MAChB,CAAC;;;;AAKR,MAAI;AAGF,OAAID,eAAa,aAAa,UAAU;AACtC,iBAAa,EAAE,QAAQ,wBAAwB,CAAC;AAChD,SAAK,YAAa,MAAM,4BACtB,cAAc,gBAAgB,OAAO,MAAM,EACzC,mBAAmB,kBACpB,CAAC,CACH;AAED,iBAAa,EAAE,QAAQ,oBAAoB,CAAC;AAC5C,SAAK,QAAQ,MAAM,4BACjB,qBAAqB,gBAAgB,OAAO,MAAM;KAChD;KACA,QAAQ;KACR,mBAAmB;KACpB,CAAC,CACH;AAED,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,gBAAY;AACZ,SAAK,eAAe;AACpB,SAAK,cAAc;AACnB,iBAAa,EAAE,QAAQ,mBAAmB,CAAC;cAClC,CAACA,eAAa,aAAa,UAAU;AAE9C,iBAAa,EAAE,QAAQ,qCAAqC,CAAC;IAG7D,MAAM,EAAE,qBAAqB,MAAM,OAAO;AAC1C,SAAK,gBAAgB,MAAM,iBAAiB,OAAO;KACjD,SAAS,OAAO;KAChB;KACD,CAAC;AAEF,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,gBAAY;AACZ,SAAK,eAAe;AACpB,SAAK,cAAc;UAEd;IAEL,MAAM,kBAAkB;KACtB;KACA,QAAQ;KACR,mBAAmB;KACpB;AACD,SAAK,YAAa,MAAM,4BACtBF,WAAS,mBAAmB,OAAO,MAAM,gBAAuB,CACjE;AAED,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,gBAAY;AACZ,SAAK,eAAe;AACpB,SAAK,cAAc;AACnB,iBAAa,EAAE,QAAQ,UAAU,SAAS,aAAa,CAAC,KAAK,CAAC;;WAEzD,KAAK;AAEZ,OAAI,aAAa,gBAAgB;AAC/B,iBAAa,EAAE,QAAQ,SAAS,eAAe,aAAa,CAAC,MAAM,CAAC;AAGpE,QAAI,KAAK,eAAe;AACtB,WAAM,KAAK,cAAc,SAAS;AAClC,UAAK,gBAAgB;;AAIvB,SAAK,YAAa,MAAM,4BACtBA,WAAS,mBAAmB,OAAO,MAAM;KACvC,OAAO;KACP,QAAQ;KACR,mBAAmB;KACpB,CAAQ,CACV;AAED,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,gBAAY;AACZ,SAAK,eAAe;AACpB,SAAK,cAAc;AACnB,iBAAa,EAAE,QAAQ,UAAU,eAAe,aAAa,CAAC,KAAK,CAAC;SAEpE,OAAM;;;;;;CAQZ,WAAoB;AAClB,SACE,KAAK,cAAc,QAClB,KAAK,aAAa,KAAK,UAAU,QAClC,KAAK,kBAAkB;;;;;CAO3B,eAAmC;AACjC,SAAO,KAAK;;;;;CAMd,gBAA2C;AACzC,SAAO,KAAK;;;;;CAMd,WAAmB;AAEjB,SAAO,KAAK,gBAAgB,WAAW,UAAU;;;;;CAMnD,kBAKS;AACP,MAAI,CAAC,KAAK,cACR,QAAO;AAET,SAAO,KAAK,cAAc,WAAW;;;;;;CAOvC,MAAM,kBAA+E;AACnF,MAAI,CAAC,KAAK,cACR,QAAO;AAET,SAAO,KAAK,cAAc,gBAAgB;;;;;CAM5C,MAAM,iBAA2F;AAC/F,MAAI,CAAC,KAAK,cACR,QAAO;AAET,SAAO,KAAK,cAAc,gBAAgB;;;;;;CAO5C,MAAM,aAA4B;AAChC,MAAI,KAAK,cACP,OAAM,KAAK,cAAc,OAAO;;;;;;;CASpC,MAAM,sBAAsB,cAAc,GAAqB;AAC7D,MAAI,CAAC,KAAK,cACR,QAAO;AAET,SAAO,KAAK,cAAc,sBAAsB,YAAY;;;;;CAU9D,MAAM,SAAS,QAAgB,UAA2B,EAAE,EAA2B;AACrF,MAAI,CAAC,KAAK,UAAU,CAElB,OAAM,KAAK,UAAU,KAAK,OAAO,SAAS,aAAa;EAGzD,MAAM,EACJ,YAAY,KACZ,cAAc,IACd,OAAO,IACP,OAAO,IACP,WAAW,OACX,WACE;EAEJ,MAAM,YAAY,YAAY,KAAK;AAEnC,MAAI;GACF,IAAI,UAAU;AAEd,OAAI,KAAK,cAEP,KAAI;AACF,cAAU,MAAM,KAAK,cAAc,SAAS,QAAQ;KAClD;KACA;KACA;KACA;KACA;KACA;KAEA,SAAS,QAAQ,WAAW,MAAM,QAAQ,QAAS,EAAE,KAAK,GAAG;KAC9D,CAAC;YACKI,WAAgB;AAEvB,QAAI,WAAW,YAAY,yBAAyB,CAAC,KAAK,eAAe,SAAS,EAAE;AAClF,WAAM,KAAK,eAAe,SAAS,CAAC,YAAY,GAAG;AACnD,UAAK,gBAAgB;AACrB,UAAK,cAAc;AAGnB,UAAK,YAAa,MAAMJ,WAAS,mBADf,KAAK,gBAAgB,cACwB;MAC7D,OAAO;MACP,QAAQ;MACT,CAAQ;AAET,YAAO,KAAK,SAAS,QAAQ,QAAQ;;AAEvC,UAAM;;YAEC,KAAK,aAAa,KAAK,SAAS,KAAK,WAAW;IAEzD,MAAM,WAAW,KAAK,cAAc,QAAQ;KAAE,GAAG;KAAS;KAAU,CAAC;IAErE,MAAM,SAAU,KAAK,UAAkB,oBAAoB,UAAU;KACnE,uBAAuB;KACvB,aAAa;KACb,iBAAiB;KAClB,CAAC;IAEF,MAAM,SAAS,MAAM,KAAK,MAAM,SAAS;KACvC,GAAG;KACH,gBAAgB;KAChB,aAAa,cAAc,IAAI,cAAc;KAC7C,OAAO;KACP,OAAO;KACP,WAAW,cAAc;KAC1B,CAAC;IAGF,MAAM,cAAc,OAAO,UAAU,OAAO,MAAM,OAAO,UAAU,MAAM,UAAU;IAGnF,MAAM,eAAe,OAAO,MAAM,MAAM,CAAC,aAAa,KAAK,CAAC;AAK5D,cAJgB,KAAK,UAAU,aAAa,cAAc,EACxD,qBAAqB,MACtB,CAAC,CAEgB,MAAM;AAGxB,QAAI,QAAQ,aAAa,CAAC,SAAS,YAAY,EAAE;KAC/C,MAAM,QAAQ,QAAQ,MAAM,4BAA4B;AACxD,SAAI,MACF,WAAU,MAAM,GAAG,MAAM;;cAGpB,KAAK,WAAW;IAEzB,MAAM,kBAAkB,KAAK,aAAa,QAAQ;KAAE,GAAG;KAAS;KAAU,CAAC;IAE3E,MAAM,SAAS,MAAM,KAAK,UAAU,iBAAiB;KACnD,gBAAgB;KAChB;KACA,OAAO;KACP,OAAO;KACP,WAAW,cAAc;KACzB,kBAAkB;KACnB,CAAC;AAGF,QAAI,MAAM,QAAQ,OAAO,IAAI,OAAO,IAAI;KACtC,MAAM,SAAS,OAAO;AACtB,SAAI,MAAM,QAAQ,OAAO,eAAe,CAEtC,WADa,OAAO,eAAe,GAAG,GAAG,EACzB,WAAW;SAE3B,WAAU,OAAO,kBAAkB;;SAIvC,OAAM,IAAI,MAAM,kBAAkB;GAIpC,MAAM,YADU,YAAY,KAAK,GACL;AAE5B,aAAU,KAAK,YAAY,QAAQ;GAInC,MAAM,EAAE,UAAU,cAAc,aAAa,KAAK,cAAc,QAAQ;GAGxE,MAAM,gBAAgB,WAAW,eAAe;GAEhD,MAAM,kBAAkB,KAAK,KAAK,SAAS,SAAS,EAAE;AAGtD,QAAK,MAAM;AACX,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,WAAY,KAAK,MAAM,YAAY,KAAK,MAAM,YAAa;AAEtE,UAAO;IACL,MAAM;IACN,UAAU;IACV;IACA,iBAAkB,kBAAkB,YAAa;IACjD;IACA,cAAc;IACd,UAAU;IACV,QAAQ;IACT;WACM,QAAQ;AACf,UAAO;IACL,MAAM;IACN,iBAAiB;IACjB,iBAAiB;IACjB,WAAW,YAAY,KAAK,GAAG;IAC/B,cAAc;IACd,UAAU;IACV,QAAQ;IACT;;;;;;;;;CAUL,OAAO,OACL,QACA,UAA2B,EAAE,EACoB;AACjD,MAAI,CAAC,KAAK,UAAU,CAClB,OAAM,KAAK,UAAU,KAAK,OAAO,SAAS,aAAa;EAGzD,MAAM,YAAY,YAAY,KAAK;AAGnC,MAAI,KAAK,eAAe;GACtB,IAAI,WAAW;GACf,MAAMK,aAAuB,EAAE;GAC/B,IAAIC,cAAuD;GAC3D,IAAI,OAAO;GAGX,MAAM,kBAAkB,KAAK,cAC1B,SAAS,QAAQ;IAChB,GAAG;IACH,UAAU,UAAU;AAClB,iBAAY,MAAM;AAClB,SAAI,aAAa;AACf,kBAAY,MAAM,KAAK;AACvB,oBAAc;WAEd,YAAW,KAAK,MAAM,KAAK;;IAGhC,CAAC,CACD,WAAW;AACV,WAAO;AACP,QAAI,YACF,aAAY,KAAK;KAEnB,CACD,OAAO,QAAQ;AACd,WAAO;AACP,QAAI,YACF,aAAY,KAAK;AAEnB,UAAM;KACN;AAGJ,UAAO,CAAC,QAAQ,WAAW,SAAS,EAClC,KAAI,WAAW,SAAS,GAAG;IACzB,MAAM,QAAQ,WAAW,OAAO;AAChC,UAAM;AACN,YAAQ,UAAU,MAAM;cACf,CAAC,MAAM;IAChB,MAAM,QAAQ,MAAM,IAAI,SAAwB,YAAY;AAC1D,mBAAc;MACd;AACF,QAAI,OAAO;AACT,WAAM;AACN,aAAQ,UAAU,MAAM;;;AAK9B,SAAM;GAEN,MAAM,EAAE,UAAU,cAAc,aAAa,KAAK,cAAc,SAAS;GACzE,MAAM,kBAAkB,KAAK,KAAK,SAAS,SAAS,EAAE;GACtD,MAAM,YAAY,YAAY,KAAK,GAAG;AAEtC,UAAO;IACL,MAAM;IACN,UAAU,QAAQ,WAAW,eAAe;IAC5C;IACA;IACA,iBAAkB,kBAAkB,YAAa;IACjD,cAAc;IACf;;EAIH,MAAM,SAAS,MAAM,KAAK,YAAY,QAAQ,QAAQ;EAItD,MAAM,QAAQ,OAAO,QAAQ,MAAM,QAAQ;AAC3C,OAAK,MAAM,QAAQ,MACjB,KAAI,MAAM;AACR,SAAM;AACN,WAAQ,UAAU,KAAK;;AAI3B,SAAO,OAAO;;;;;CAMhB,MAAc,YACZ,QACA,UAA2B,EAAE,EACyB;EACtD,MAAM,EAAE,YAAY,KAAK,cAAc,IAAK,OAAO,IAAK,OAAO,IAAI,WAAW,UAAU;EAExF,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,kBAAkB,KAAK,aAAa,QAAQ;GAAE,GAAG;GAAS;GAAU,CAAC;AAE3E,MAAI;GACF,MAAM,SAAS,MAAM,KAAK,UAAW,iBAAiB;IACpD,gBAAgB;IAChB;IACA,OAAO;IACP,OAAO;IACP,WAAW,cAAc;IACzB,kBAAkB;IACnB,CAAC;GAGF,MAAM,YADU,YAAY,KAAK,GACL;GAG5B,IAAI,UAAU;AACd,OAAI,MAAM,QAAQ,OAAO,IAAI,OAAO,IAAI;IACtC,MAAM,SAAS,OAAO;AACtB,QAAI,MAAM,QAAQ,OAAO,eAAe,CAEtC,WADa,OAAO,eAAe,GAAG,GAAG,EACzB,WAAW;QAE3B,WAAU,OAAO,kBAAkB;;AAIvC,aAAU,KAAK,YAAY,QAAQ;GACnC,MAAM,EAAE,UAAU,cAAc,aAAa,KAAK,cAAc,QAAQ;GACxE,MAAM,gBAAgB,WAAW,eAAe;GAChD,MAAM,kBAAkB,KAAK,KAAK,SAAS,SAAS,EAAE;AAGtD,QAAK,MAAM;AACX,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,aAAa;AACxB,QAAK,MAAM,WAAY,KAAK,MAAM,YAAY,KAAK,MAAM,YAAa;AAEtE,UAAO;IACL;IACA,QAAQ;KACN,MAAM;KACN,UAAU;KACV;KACA,iBAAkB,kBAAkB,YAAa;KACjD;KACA,cAAc;KACd,UAAU;KACV,QAAQ;KACT;IACF;WACM,QAAQ;AACf,UAAO;IACL,SAAS;IACT,QAAQ;KACN,MAAM;KACN,iBAAiB;KACjB,iBAAiB;KACjB,WAAW,YAAY,KAAK,GAAG;KAC/B,cAAc;KACd,UAAU;KACV,QAAQ;KACT;IACF;;;;;;CAWL,MAAM,KAAQ,QAAgB,SAAqC;EACjE,MAAM,EAAE,QAAQ,UAAU,GAAG,cAAc,OAAQ;EAEnD,MAAM,eAAe;;wCAEe,KAAK,UAAU,gBAAgB,OAAO,CAAC;AAE3E,OAAK,IAAI,UAAU,GAAG,UAAU,SAAS,WAAW;GAClD,MAAM,SAAS,MAAM,KAAK,SAAS,QAAQ;IACzC,QAAQ,QAAQ,UAAU;IAC1B;IACA,WAAW;IACZ,CAAC;AAEF,OAAI;IAEF,MAAM,UAAU,YAAY,OAAO,KAAK;IACxC,MAAM,SAAS,KAAK,MAAM,QAAQ;AAElC,WADkB,OAAO,MAAM,OAAO;YAE/B,OAAO;AACd,QAAI,YAAY,UAAU,EACxB,OAAM,IAAI,MAAM,uCAAuC,QAAQ,aAAa,QAAQ;;;AAK1F,QAAM,IAAI,MAAM,gCAAgC;;;;;CAUlD,MAAM,MAAM,MAAc,UAAwB,EAAE,EAAwB;AAC1E,MAAI,CAAC,KAAK,SAGR,MAAK,WAAY,MAAMN,WAAS,sBADlB,QAAQ,SAAS,0BAC6B;EAG9D,MAAM,YAAY,YAAY,KAAK;EACnC,MAAM,SAAS,MAAM,KAAK,SAAS,MAAM;GACvC,SAAS;GACT,WAAW,QAAQ,cAAc;GAClC,CAAC;AAIF,SAAO;GACL,QAHa,MAAM,KAAK,OAAO,KAAqB;GAIpD;GACA,WAAW,YAAY,KAAK,GAAG;GAChC;;;;;CAMH,MAAM,WAAW,OAAiB,UAAwB,EAAE,EAA0B;EACpF,MAAMO,UAAyB,EAAE;AACjC,OAAK,MAAM,QAAQ,MACjB,SAAQ,KAAK,MAAM,KAAK,MAAM,MAAM,QAAQ,CAAC;AAE/C,SAAO;;;;;CAUT,WAAyB;AACvB,SAAO,EAAE,GAAG,KAAK,OAAO;;;;;CAM1B,UAAsB;AACpB,SAAO;GACL,SAAS;GACT,OAAO,KAAK;GACZ,QAAQ;IACN,SAAS;IACT,KAAK;IACL,MAAM;IACN,QAAQ,KAAK,UAAU,GAAG,UAAU;IACrC;GACD,SAAS;IACP,KAAK,KAAK,aAAa,iBAAiB;IACxC,MAAM;IACN,WAAW,KAAK,aAAa,iBAAiB;IAC/C;GACD,OAAO;IACL,UAAU;IACV,MAAM;IACN,YAAY;IACb;GACF;;;;;CAMH,aAAmB;AACjB,OAAK,QAAQ;GACX,SAAS;GACT,UAAU;GACV,WAAW;GACX,UAAU;GACV,WAAW;GACX,WAAW;GACX,aAAa;GACd;;;;;CAUH,MAAM,UAAyB;AAE7B,MAAI,KAAK,eAAe;AACtB,OAAI;AACF,UAAM,KAAK,cAAc,SAAS;WAC5B;AAGR,QAAK,gBAAgB;;AAGvB,MAAI,KAAK,WAAW;AAClB,OAAI,OAAQ,KAAK,UAAkB,YAAY,WAC7C,KAAI;AACF,UAAO,KAAK,UAAkB,SAAS;WACjC;AAIV,QAAK,YAAY;;AAEnB,MAAI,KAAK,UAAU;AACjB,OAAI,OAAQ,KAAK,SAAiB,YAAY,WAC5C,KAAI;AACF,UAAO,KAAK,SAAiB,SAAS;WAChC;AAIV,QAAK,WAAW;;AAElB,OAAK,eAAe;AACpB,OAAK,cAAc;;CAOrB,AAAQ,aAAa,QAAgB,SAAkC;EACrE,MAAM,SAAS,QAAQ,UAAU;EACjC,MAAM,SAAS,KAAK,cAAc,SAAS,OAAO;AAElD,MAAI,QAAQ,YAAY,KAAK,aAAa,iBAExC,QAAO,uBADa,GAAG,OAAO,iHACY,gCAAgC,OAAO;AAGnF,MAAI,OACF,QAAO,uBAAuB,OAAO,gCAAgC,OAAO;AAG9E,SAAO,uBAAuB,OAAO,gCAAgC,OAAO;;CAG9E,AAAQ,cACN,QACA,SAC0C;EAC1C,MAAM,SAAS,QAAQ,UAAU;EACjC,MAAMC,WAAqD,EAAE;AAI7D,WAAS,KAAK;GAAE,MAAM;GAAU,SAAS;GAAQ,CAAC;AAClD,WAAS,KAAK;GAAE,MAAM;GAAQ,SAAS;GAAQ,CAAC;AAEhD,SAAO;;CAGT,AAAQ,cAAc,MAGpB;EAEA,MAAM,QAAQ,KAAK,MAAM,6BAA6B;AACtD,MAAI,MAGF,QAAO;GAAE,UAFQ,MAAM,GAAG,MAAM;GAEb,UADF,KAAK,QAAQ,4BAA4B,GAAG,CAAC,MAAM;GACvC;EAI/B,MAAM,gBAAgB,KAAK,MAAM,oBAAoB;AACrD,MAAI,eAAe;GACjB,MAAM,WAAW,cAAc,GAAG,MAAM;GACxC,MAAM,WAAW,KAAK,QAAQ,mBAAmB,GAAG,CAAC,MAAM;AAC3D,UAAO;IAAE,UAAU,YAAY;IAAW;IAAU;;AAKtD,SAAO,EAAE,UADQ,KAAK,QAAQ,eAAe,GAAG,CAAC,MAAM,EACpC;;CAGrB,AAAQ,YAAY,MAAsB;AACxC,SACE,KACG,QAAQ,iBAAiB,GAAG,CAC5B,QAAQ,mBAAmB,GAAG,CAC9B,QAAQ,oBAAoB,GAAG,CAC/B,QAAQ,UAAU,GAAG,CAErB,QAAQ,mBAAmB,GAAG,CAC9B,QAAQ,kBAAkB,GAAG,CAC7B,QAAQ,wBAAwB,GAAG,CACnC,QAAQ,uBAAuB,GAAG,CAElC,QAAQ,mCAAmC,GAAG,CAC9C,MAAM"}

package/dist/gerbil-BjW-z7Fq.mjs

@@ -1,5 +0,0 @@
-import "./models-DKULvhOr.mjs";
-import "./utils-7vXqtq2Q.mjs";
-import { t as Gerbil } from "./gerbil-BfnsFWRE.mjs";
-
-export { Gerbil };

package/dist/gerbil-DZ1k3ChC.d.mts

@@ -1,138 +0,0 @@
-import { _ as SystemInfo, a as GenerateOptions, d as LoadOptions, f as ModelConfig, g as SessionStats, n as EmbedOptions, o as GenerateResult, r as EmbedResult, s as GerbilConfig, u as JsonOptions } from "./types-BS1N92Jt.mjs";
-
-//#region src/core/gerbil.d.ts
-
-declare class Gerbil {
-  private generator;
-  private tokenizer;
-  private model;
-  private embedder;
-  private currentModel;
-  private modelConfig;
-  private readonly config;
-  private stats;
-  private useDirect;
-  private chromeBackend;
-  private _deviceMode;
-  constructor(config?: GerbilConfig);
-  static listModels(): ModelConfig[];
-  static getModel(modelId: string): ModelConfig | undefined;
-  /**
-   * Load a model
-   *
-   * @example
-   * ```ts
-   * // Built-in model
-   * await g.loadModel("qwen3-0.6b");
-   *
-   * // HuggingFace model
-   * await g.loadModel("hf:microsoft/Phi-3-mini");
-   *
-   * // Local model
-   * await g.loadModel("file:./models/my-model");
-   * ```
-   */
-  loadModel(modelId?: string, options?: LoadOptions): Promise<void>;
-  /**
-   * Check if a model is loaded
-   */
-  isLoaded(): boolean;
-  /**
-   * Get current model info
-   */
-  getModelInfo(): ModelConfig | null;
-  /**
-   * Get current device mode (webgpu, cpu, or wasm)
-   */
-  getDeviceMode(): "webgpu" | "cpu" | "wasm";
-  /**
-   * Get dtype used for current model
-   */
-  getDtype(): string;
-  /**
-   * Get Chrome backend status (if using WebGPU via Chrome)
-   */
-  getChromeStatus(): {
-    pid: number | null;
-    port: number;
-    modelId: string;
-    startedAt: Date | null;
-  } | null;
-  /**
-   * Get Chrome memory usage (if using WebGPU via Chrome)
-   * Returns JS heap memory in bytes
-   */
-  getChromeMemory(): Promise<{
-    jsHeapUsed: number;
-    jsHeapTotal: number;
-  } | null>;
-  /**
-   * Get memory usage in GB (if using WebGPU via Chrome)
-   */
-  getMemoryUsage(): Promise<{
-    usedGB: number;
-    totalGB: number;
-    usedPercent: number;
-  } | null>;
-  /**
-   * Clear KV cache to free memory
-   * This will reset the conversation context but free up memory
-   */
-  clearCache(): Promise<void>;
-  /**
-   * Check memory usage and cleanup if needed
-   * @param thresholdGB Memory threshold in GB (default: 8)
-   * @returns true if cleanup was performed
-   */
-  checkMemoryAndCleanup(thresholdGB?: number): Promise<boolean>;
-  /**
-   * Generate text
-   */
-  generate(prompt: string, options?: GenerateOptions): Promise<GenerateResult>;
-  /**
-   * Stream text generation (simulated token-by-token)
-   *
-   * Note: Yields the raw output including <think> tags if thinking mode is enabled.
-   * The final result has parsed thinking separated out.
-   */
-  stream(prompt: string, options?: GenerateOptions): AsyncGenerator<string, GenerateResult, unknown>;
-  /**
-   * Internal: Generate with raw text access for streaming
-   */
-  private generateRaw;
-  /**
-   * Generate structured JSON output
-   */
-  json<T>(prompt: string, options: JsonOptions<T>): Promise<T>;
-  /**
-   * Generate embeddings
-   */
-  embed(text: string, options?: EmbedOptions): Promise<EmbedResult>;
-  /**
-   * Generate embeddings for multiple texts
-   */
-  embedBatch(texts: string[], options?: EmbedOptions): Promise<EmbedResult[]>;
-  /**
-   * Get session stats
-   */
-  getStats(): SessionStats;
-  /**
-   * Get system info
-   */
-  getInfo(): SystemInfo;
-  /**
-   * Reset stats
-   */
-  resetStats(): void;
-  /**
-   * Dispose of resources
-   */
-  dispose(): Promise<void>;
-  private formatPrompt;
-  private buildMessages;
-  private parseThinking;
-  private cleanOutput;
-}
-//#endregion
-export { Gerbil as t };
-//# sourceMappingURL=gerbil-DZ1k3ChC.d.mts.map

package/dist/gerbil-DZ1k3ChC.d.mts.map

@@ -1 +0,0 @@
-{"version":3,"file":"gerbil-DZ1k3ChC.d.mts","names":[],"sources":["../src/core/gerbil.ts"],"sourcesContent":[],"mappings":";;;;AA6HsB,cAbT,MAAA,CAaS;EAiBC,QAAA,SAAA;EAIa,QAAA,SAAA;EAuBe,QAAA,KAAA;EAAmB,QAAA,QAAA;EAgKpD,QAAA,YAAA;EA0BH,QAAA,WAAA;EAYY,iBAAA,MAAA;EAUD,QAAA,KAAA;EAWJ,QAAA,SAAA;EAW0B,QAAA,aAAA;EAcN,QAAA,WAAA;EAA+B,WAAA,CAAA,MAAA,CAAA,EAhSnD,YAgSmD;EAAR,OAAA,UAAA,CAAA,CAAA,EA/Q1C,WA+Q0C,EAAA;EAqKpD,OAAA,QAAA,CAAA,OAAA,EAAA,MAAA,CAAA,EAhbuB,WAgbvB,GAAA,SAAA;EACe;;;;;;;;;;;;;;;EAgTF,SAAA,CAAA,OAAA,CAAA,EAAA,MAAA,EAAA,OAAA,CAAA,EA1sByB,WA0sBzB,CAAA,EA1sB4C,OA0sB5C,CAAA,IAAA,CAAA;;;;;;;;kBA1iBR;;;;;;;;;;;;;;;;eA0BH;;;;;;qBAYY;;;;;;;oBAUD;;;;;;;;;gBAWJ;;;;;;+CAW0B;;;;qCAcN,kBAAuB,QAAQ;;;;;;;mCAqK5D,kBACR,uBAAuB;;;;;;;;mCA+Ka,YAAY,KAAK,QAAQ;;;;gCAqC7B,eAAoB,QAAQ;;;;wCAyBpB,eAAoB,QAAQ;;;;cAe3D;;;;aAOD;;;;;;;;aA6CM"}

package/dist/mcp-R8kRLIKb.mjs.map

@@ -1 +0,0 @@
-{"version":3,"file":"mcp-R8kRLIKb.mjs","names":["handlers: Record<string, (args: any) => Promise<any>>","skills.summarize","skills.explain","skills.review","skills.commit","skills.translate","response: any"],"sources":["../src/integrations/mcp.ts"],"sourcesContent":["/**\n * Gerbil MCP Server\n *\n * Model Context Protocol server for Claude Desktop, Cursor, etc.\n *\n * @example\n * ```bash\n * gerbil serve --mcp\n * ```\n *\n * @example Claude Desktop config\n * ```json\n * {\n *   \"mcpServers\": {\n *     \"gerbil\": {\n *       \"command\": \"npx\",\n *       \"args\": [\"-y\", \"gerbil\", \"serve\", \"--mcp\"]\n *     }\n *   }\n * }\n * ```\n */\n\nimport { Gerbil } from \"../core/gerbil.js\";\nimport type { GerbilConfig } from \"../core/types.js\";\nimport * as skills from \"../skills/index.js\";\n\nexport interface MCPServerOptions extends GerbilConfig {\n  /** Port for HTTP transport (default: stdio) */\n  port?: number;\n\n  /** Tools to expose */\n  tools?: string[];\n}\n\n// Tool definitions\nconst TOOL_DEFINITIONS = {\n  generate: {\n    name: \"gerbil_generate\",\n    description: \"Generate text using a local LLM\",\n    inputSchema: {\n      type: \"object\",\n      properties: {\n        prompt: { type: \"string\", description: \"The prompt to generate from\" },\n        maxTokens: { type: \"number\", description: \"Maximum tokens to generate\" },\n        temperature: { type: \"number\", description: \"Sampling temperature (0-2)\" },\n        system: { type: \"string\", description: \"System prompt\" },\n        thinking: { type: \"boolean\", description: \"Enable thinking mode\" },\n      },\n      required: [\"prompt\"],\n    },\n  },\n  summarize: {\n    name: \"gerbil_summarize\",\n    description: \"Summarize content\",\n    inputSchema: {\n      type: \"object\",\n      properties: {\n        content: { type: \"string\", description: \"Content to summarize\" },\n        length: { type: \"string\", enum: [\"short\", \"medium\", \"long\"] },\n        format: { type: \"string\", enum: [\"paragraph\", \"bullets\"] },\n      },\n      required: [\"content\"],\n    },\n  },\n  explain: {\n    name: \"gerbil_explain\",\n    description: \"Explain code or concepts\",\n    inputSchema: {\n      type: \"object\",\n      properties: {\n        content: { type: \"string\", description: \"Code or concept to explain\" },\n        level: { type: \"string\", enum: [\"beginner\", \"intermediate\", \"expert\"] },\n        language: { type: \"string\", description: \"Programming language\" },\n      },\n      required: [\"content\"],\n    },\n  },\n  review: {\n    name: \"gerbil_review\",\n    description: \"Review code for issues\",\n    inputSchema: {\n      type: \"object\",\n      properties: {\n        code: { type: \"string\", description: \"Code to review\" },\n        focus: {\n          type: \"array\",\n          items: { type: \"string\", enum: [\"security\", \"performance\", \"style\", \"bugs\"] },\n        },\n      },\n      required: [\"code\"],\n    },\n  },\n  commit: {\n    name: \"gerbil_commit\",\n    description: \"Generate a commit message from a diff\",\n    inputSchema: {\n      type: \"object\",\n      properties: {\n        diff: { type: \"string\", description: \"Git diff\" },\n        type: { type: \"string\", enum: [\"conventional\", \"simple\", \"detailed\"] },\n      },\n      required: [\"diff\"],\n    },\n  },\n  translate: {\n    name: \"gerbil_translate\",\n    description: \"Translate text\",\n    inputSchema: {\n      type: \"object\",\n      properties: {\n        text: { type: \"string\", description: \"Text to translate\" },\n        to: { type: \"string\", description: \"Target language\" },\n        from: { type: \"string\", description: \"Source language (optional)\" },\n      },\n      required: [\"text\", \"to\"],\n    },\n  },\n  embed: {\n    name: \"gerbil_embed\",\n    description: \"Generate embeddings for text\",\n    inputSchema: {\n      type: \"object\",\n      properties: {\n        text: { type: \"string\", description: \"Text to embed\" },\n      },\n      required: [\"text\"],\n    },\n  },\n};\n\n/**\n * Create MCP server\n */\nexport async function createMCPServer(options: MCPServerOptions = {}) {\n  const g = new Gerbil(options);\n\n  // Load model\n  await g.loadModel(options.model || \"qwen3-0.6b\");\n\n  // Determine which tools to expose\n  const enabledTools = options.tools || Object.keys(TOOL_DEFINITIONS);\n\n  // Tool handlers\n  const handlers: Record<string, (args: any) => Promise<any>> = {\n    gerbil_generate: async (args) => {\n      const result = await g.generate(args.prompt, {\n        maxTokens: args.maxTokens,\n        temperature: args.temperature,\n        system: args.system,\n        thinking: args.thinking,\n      });\n      return {\n        content: [\n          {\n            type: \"text\",\n            text: result.thinking\n              ? `Thinking: ${result.thinking}\\n\\nAnswer: ${result.text}`\n              : result.text,\n          },\n        ],\n      };\n    },\n\n    gerbil_summarize: async (args) => {\n      const summary = await skills.summarize({\n        content: args.content,\n        length: args.length,\n        format: args.format,\n      });\n      return { content: [{ type: \"text\", text: summary }] };\n    },\n\n    gerbil_explain: async (args) => {\n      const explanation = await skills.explain({\n        content: args.content,\n        level: args.level,\n        language: args.language,\n      });\n      return { content: [{ type: \"text\", text: explanation }] };\n    },\n\n    gerbil_review: async (args) => {\n      const reviewResult = await skills.review({\n        code: args.code,\n        focus: args.focus,\n      });\n      return { content: [{ type: \"text\", text: reviewResult }] };\n    },\n\n    gerbil_commit: async (args) => {\n      const message = await skills.commit({\n        diff: args.diff,\n        type: args.type,\n      });\n      return { content: [{ type: \"text\", text: message }] };\n    },\n\n    gerbil_translate: async (args) => {\n      const translated = await skills.translate({\n        text: args.text,\n        to: args.to,\n        from: args.from,\n      });\n      return { content: [{ type: \"text\", text: translated }] };\n    },\n\n    gerbil_embed: async (args) => {\n      const result = await g.embed(args.text);\n      return {\n        content: [\n          {\n            type: \"text\",\n            text: JSON.stringify({\n              vector: result.vector.slice(0, 10),\n              dimensions: result.vector.length,\n            }),\n          },\n        ],\n      };\n    },\n  };\n\n  return {\n    // Server info\n    info: {\n      name: \"gerbil\",\n      version: \"1.0.0\",\n      description: \"Local LLM inference via Gerbil\",\n    },\n\n    // List available tools\n    listTools: () =>\n      enabledTools\n        .filter((t) => TOOL_DEFINITIONS[t as keyof typeof TOOL_DEFINITIONS])\n        .map((t) => TOOL_DEFINITIONS[t as keyof typeof TOOL_DEFINITIONS]),\n\n    // Call a tool\n    callTool: async (name: string, args: any) => {\n      const handler = handlers[name];\n      if (!handler) {\n        throw new Error(`Unknown tool: ${name}`);\n      }\n      return handler(args);\n    },\n\n    // Get Gerbil instance\n    gerbil: g,\n  };\n}\n\n/**\n * Start MCP server with stdio transport\n */\nexport async function startMCPServer(options: MCPServerOptions = {}) {\n  const server = await createMCPServer(options);\n\n  // Simple stdio protocol handler\n  const readline = await import(\"readline\");\n  const rl = readline.createInterface({\n    input: process.stdin,\n    output: process.stdout,\n    terminal: false,\n  });\n\n  console.error(\"🐹 Gerbil MCP Server\");\n  console.error(`   Model: ${options.model || \"qwen3-0.6b\"}`);\n  console.error(\n    `   Tools: ${server\n      .listTools()\n      .map((t) => t.name)\n      .join(\", \")}`,\n  );\n  console.error(\"   Ready for connections...\");\n\n  rl.on(\"line\", async (line) => {\n    try {\n      const request = JSON.parse(line);\n\n      let response: any;\n\n      switch (request.method) {\n        case \"initialize\":\n          response = {\n            protocolVersion: \"2024-11-05\",\n            serverInfo: server.info,\n            capabilities: {\n              tools: {},\n            },\n          };\n          break;\n\n        case \"tools/list\":\n          response = { tools: server.listTools() };\n          break;\n\n        case \"tools/call\":\n          response = await server.callTool(request.params.name, request.params.arguments);\n          break;\n\n        default:\n          response = { error: { code: -32_601, message: \"Method not found\" } };\n      }\n\n      console.log(JSON.stringify({ jsonrpc: \"2.0\", id: request.id, result: response }));\n    } catch (error) {\n      console.log(\n        JSON.stringify({\n          jsonrpc: \"2.0\",\n          id: null,\n          error: { code: -32_700, message: String(error) },\n        }),\n      );\n    }\n  });\n}\n\nexport default { createMCPServer, startMCPServer };\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;AAoCA,MAAM,mBAAmB;CACvB,UAAU;EACR,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,QAAQ;KAAE,MAAM;KAAU,aAAa;KAA+B;IACtE,WAAW;KAAE,MAAM;KAAU,aAAa;KAA8B;IACxE,aAAa;KAAE,MAAM;KAAU,aAAa;KAA8B;IAC1E,QAAQ;KAAE,MAAM;KAAU,aAAa;KAAiB;IACxD,UAAU;KAAE,MAAM;KAAW,aAAa;KAAwB;IACnE;GACD,UAAU,CAAC,SAAS;GACrB;EACF;CACD,WAAW;EACT,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,SAAS;KAAE,MAAM;KAAU,aAAa;KAAwB;IAChE,QAAQ;KAAE,MAAM;KAAU,MAAM;MAAC;MAAS;MAAU;MAAO;KAAE;IAC7D,QAAQ;KAAE,MAAM;KAAU,MAAM,CAAC,aAAa,UAAU;KAAE;IAC3D;GACD,UAAU,CAAC,UAAU;GACtB;EACF;CACD,SAAS;EACP,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,SAAS;KAAE,MAAM;KAAU,aAAa;KAA8B;IACtE,OAAO;KAAE,MAAM;KAAU,MAAM;MAAC;MAAY;MAAgB;MAAS;KAAE;IACvE,UAAU;KAAE,MAAM;KAAU,aAAa;KAAwB;IAClE;GACD,UAAU,CAAC,UAAU;GACtB;EACF;CACD,QAAQ;EACN,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;KAAkB;IACvD,OAAO;KACL,MAAM;KACN,OAAO;MAAE,MAAM;MAAU,MAAM;OAAC;OAAY;OAAe;OAAS;OAAO;MAAE;KAC9E;IACF;GACD,UAAU,CAAC,OAAO;GACnB;EACF;CACD,QAAQ;EACN,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;KAAY;IACjD,MAAM;KAAE,MAAM;KAAU,MAAM;MAAC;MAAgB;MAAU;MAAW;KAAE;IACvE;GACD,UAAU,CAAC,OAAO;GACnB;EACF;CACD,WAAW;EACT,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;KAAqB;IAC1D,IAAI;KAAE,MAAM;KAAU,aAAa;KAAmB;IACtD,MAAM;KAAE,MAAM;KAAU,aAAa;KAA8B;IACpE;GACD,UAAU,CAAC,QAAQ,KAAK;GACzB;EACF;CACD,OAAO;EACL,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY,EACV,MAAM;IAAE,MAAM;IAAU,aAAa;IAAiB,EACvD;GACD,UAAU,CAAC,OAAO;GACnB;EACF;CACF;;;;AAKD,eAAsB,gBAAgB,UAA4B,EAAE,EAAE;CACpE,MAAM,IAAI,IAAI,OAAO,QAAQ;AAG7B,OAAM,EAAE,UAAU,QAAQ,SAAS,aAAa;CAGhD,MAAM,eAAe,QAAQ,SAAS,OAAO,KAAK,iBAAiB;CAGnE,MAAMA,WAAwD;EAC5D,iBAAiB,OAAO,SAAS;GAC/B,MAAM,SAAS,MAAM,EAAE,SAAS,KAAK,QAAQ;IAC3C,WAAW,KAAK;IAChB,aAAa,KAAK;IAClB,QAAQ,KAAK;IACb,UAAU,KAAK;IAChB,CAAC;AACF,UAAO,EACL,SAAS,CACP;IACE,MAAM;IACN,MAAM,OAAO,WACT,aAAa,OAAO,SAAS,cAAc,OAAO,SAClD,OAAO;IACZ,CACF,EACF;;EAGH,kBAAkB,OAAO,SAAS;AAMhC,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MALnB,MAAMC,UAAiB;KACrC,SAAS,KAAK;KACd,QAAQ,KAAK;KACb,QAAQ,KAAK;KACd,CAAC;IACgD,CAAC,EAAE;;EAGvD,gBAAgB,OAAO,SAAS;AAM9B,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MALf,MAAMC,QAAe;KACvC,SAAS,KAAK;KACd,OAAO,KAAK;KACZ,UAAU,KAAK;KAChB,CAAC;IACoD,CAAC,EAAE;;EAG3D,eAAe,OAAO,SAAS;AAK7B,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MAJd,MAAMC,OAAc;KACvC,MAAM,KAAK;KACX,OAAO,KAAK;KACb,CAAC;IACqD,CAAC,EAAE;;EAG5D,eAAe,OAAO,SAAS;AAK7B,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MAJnB,MAAMC,OAAc;KAClC,MAAM,KAAK;KACX,MAAM,KAAK;KACZ,CAAC;IACgD,CAAC,EAAE;;EAGvD,kBAAkB,OAAO,SAAS;AAMhC,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MALhB,MAAMC,UAAiB;KACxC,MAAM,KAAK;KACX,IAAI,KAAK;KACT,MAAM,KAAK;KACZ,CAAC;IACmD,CAAC,EAAE;;EAG1D,cAAc,OAAO,SAAS;GAC5B,MAAM,SAAS,MAAM,EAAE,MAAM,KAAK,KAAK;AACvC,UAAO,EACL,SAAS,CACP;IACE,MAAM;IACN,MAAM,KAAK,UAAU;KACnB,QAAQ,OAAO,OAAO,MAAM,GAAG,GAAG;KAClC,YAAY,OAAO,OAAO;KAC3B,CAAC;IACH,CACF,EACF;;EAEJ;AAED,QAAO;EAEL,MAAM;GACJ,MAAM;GACN,SAAS;GACT,aAAa;GACd;EAGD,iBACE,aACG,QAAQ,MAAM,iBAAiB,GAAoC,CACnE,KAAK,MAAM,iBAAiB,GAAoC;EAGrE,UAAU,OAAO,MAAc,SAAc;GAC3C,MAAM,UAAU,SAAS;AACzB,OAAI,CAAC,QACH,OAAM,IAAI,MAAM,iBAAiB,OAAO;AAE1C,UAAO,QAAQ,KAAK;;EAItB,QAAQ;EACT;;;;;AAMH,eAAsB,eAAe,UAA4B,EAAE,EAAE;CACnE,MAAM,SAAS,MAAM,gBAAgB,QAAQ;CAI7C,MAAM,MADW,MAAM,OAAO,aACV,gBAAgB;EAClC,OAAO,QAAQ;EACf,QAAQ,QAAQ;EAChB,UAAU;EACX,CAAC;AAEF,SAAQ,MAAM,uBAAuB;AACrC,SAAQ,MAAM,aAAa,QAAQ,SAAS,eAAe;AAC3D,SAAQ,MACN,aAAa,OACV,WAAW,CACX,KAAK,MAAM,EAAE,KAAK,CAClB,KAAK,KAAK,GACd;AACD,SAAQ,MAAM,8BAA8B;AAE5C,IAAG,GAAG,QAAQ,OAAO,SAAS;AAC5B,MAAI;GACF,MAAM,UAAU,KAAK,MAAM,KAAK;GAEhC,IAAIC;AAEJ,WAAQ,QAAQ,QAAhB;IACE,KAAK;AACH,gBAAW;MACT,iBAAiB;MACjB,YAAY,OAAO;MACnB,cAAc,EACZ,OAAO,EAAE,EACV;MACF;AACD;IAEF,KAAK;AACH,gBAAW,EAAE,OAAO,OAAO,WAAW,EAAE;AACxC;IAEF,KAAK;AACH,gBAAW,MAAM,OAAO,SAAS,QAAQ,OAAO,MAAM,QAAQ,OAAO,UAAU;AAC/E;IAEF,QACE,YAAW,EAAE,OAAO;KAAE,MAAM;KAAS,SAAS;KAAoB,EAAE;;AAGxE,WAAQ,IAAI,KAAK,UAAU;IAAE,SAAS;IAAO,IAAI,QAAQ;IAAI,QAAQ;IAAU,CAAC,CAAC;WAC1E,OAAO;AACd,WAAQ,IACN,KAAK,UAAU;IACb,SAAS;IACT,IAAI;IACJ,OAAO;KAAE,MAAM;KAAS,SAAS,OAAO,MAAM;KAAE;IACjD,CAAC,CACH;;GAEH;;AAGJ,kBAAe;CAAE;CAAiB;CAAgB"}

package/dist/models-DKULvhOr.mjs.map

@@ -1 +0,0 @@
-{"version":3,"file":"models-DKULvhOr.mjs","names":["BUILTIN_MODELS: Record<string, ModelConfig>","family: ModelConfig[\"family\"]"],"sources":["../src/core/models.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// Built-in Models (curated & tested)\n// ============================================\n\nexport const BUILTIN_MODELS: Record<string, ModelConfig> = {\n  \"qwen3-0.6b\": {\n    id: \"qwen3-0.6b\",\n    repo: \"onnx-community/Qwen3-0.6B-ONNX\",\n    description: \"Qwen3 0.6B - Best balance of speed and quality, supports thinking\",\n    size: \"~400MB\",\n    contextLength: 32_768,\n    supportsThinking: true,\n    supportsJson: true,\n    family: \"qwen\",\n  },\n  \"qwen2.5-0.5b\": {\n    id: \"qwen2.5-0.5b\",\n    repo: \"onnx-community/Qwen2.5-0.5B-Instruct\",\n    description: \"Qwen2.5 0.5B - Fast and capable\",\n    size: \"~350MB\",\n    contextLength: 32_768,\n    supportsThinking: false,\n    supportsJson: true,\n    family: \"qwen\",\n  },\n  \"qwen2.5-coder-0.5b\": {\n    id: \"qwen2.5-coder-0.5b\",\n    repo: \"onnx-community/Qwen2.5-Coder-0.5B-Instruct\",\n    description: \"Qwen2.5 Coder 0.5B - Optimized for code\",\n    size: \"~400MB\",\n    contextLength: 32_768,\n    supportsThinking: false,\n    supportsJson: true,\n    family: \"qwen\",\n  },\n  \"smollm2-360m\": {\n    id: \"smollm2-360m\",\n    repo: \"HuggingFaceTB/SmolLM2-360M-Instruct\",\n    description: \"SmolLM2 360M - Fast, good for simple tasks\",\n    size: \"~250MB\",\n    contextLength: 8192,\n    supportsThinking: false,\n    supportsJson: false,\n    family: \"smollm\",\n  },\n  \"smollm2-135m\": {\n    id: \"smollm2-135m\",\n    repo: \"HuggingFaceTB/SmolLM2-135M-Instruct\",\n    description: \"SmolLM2 135M - Fastest, basic generation\",\n    size: \"~100MB\",\n    contextLength: 8192,\n    supportsThinking: false,\n    supportsJson: false,\n    family: \"smollm\",\n  },\n  \"phi-3-mini\": {\n    id: \"phi-3-mini\",\n    repo: \"microsoft/Phi-3-mini-4k-instruct-onnx\",\n    description: \"Phi-3 Mini - High quality, larger model\",\n    size: \"~2.1GB\",\n    contextLength: 4096,\n    supportsThinking: false,\n    supportsJson: true,\n    family: \"phi\",\n  },\n};\n\n// ============================================\n// Model Resolution\n// ============================================\n\n/**\n * Parse model identifier and resolve to source\n *\n * Supported formats:\n * - \"qwen3-0.6b\" (built-in)\n * - \"hf:org/model\" (HuggingFace shorthand)\n * - \"https://huggingface.co/org/model\" (full URL)\n * - \"file:./path/to/model\" (local path)\n */\nexport function resolveModel(modelId: string): ModelSource {\n  // Built-in model\n  if (BUILTIN_MODELS[modelId]) {\n    return {\n      type: \"builtin\",\n      path: BUILTIN_MODELS[modelId].repo,\n    };\n  }\n\n  // HuggingFace shorthand: hf:org/model\n  if (modelId.startsWith(\"hf:\")) {\n    const repo = modelId.slice(3);\n    return {\n      type: \"huggingface\",\n      path: repo,\n    };\n  }\n\n  // HuggingFace URL\n  if (modelId.startsWith(\"https://huggingface.co/\")) {\n    const repo = modelId.replace(\"https://huggingface.co/\", \"\");\n    return {\n      type: \"huggingface\",\n      path: repo,\n    };\n  }\n\n  // Local file\n  if (modelId.startsWith(\"file:\")) {\n    const path = modelId.slice(5);\n    return {\n      type: \"local\",\n      path,\n    };\n  }\n\n  // Assume it's a HuggingFace repo if it contains a slash\n  if (modelId.includes(\"/\")) {\n    return {\n      type: \"huggingface\",\n      path: modelId,\n    };\n  }\n\n  // Unknown - treat as HuggingFace\n  return {\n    type: \"huggingface\",\n    path: modelId,\n  };\n}\n\n/**\n * Get model config (built-in only)\n */\nexport function getModelConfig(modelId: string): ModelConfig | null {\n  return BUILTIN_MODELS[modelId] || null;\n}\n\n/**\n * Create model config for external model\n */\nexport function createExternalModelConfig(modelId: string, repo: string): ModelConfig {\n  // Try to infer family from repo name\n  let family: ModelConfig[\"family\"] = \"other\";\n  const repoLower = repo.toLowerCase();\n\n  if (repoLower.includes(\"qwen\")) {\n    family = \"qwen\";\n  } else if (repoLower.includes(\"smollm\")) {\n    family = \"smollm\";\n  } else if (repoLower.includes(\"phi\")) {\n    family = \"phi\";\n  } else if (repoLower.includes(\"mistral\")) {\n    family = \"mistral\";\n  } else if (repoLower.includes(\"llama\")) {\n    family = \"llama\";\n  }\n\n  return {\n    id: modelId,\n    repo,\n    description: `External model: ${repo}`,\n    size: \"Unknown\",\n    contextLength: 4096, // Conservative default\n    supportsThinking: family === \"qwen\",\n    supportsJson: family === \"qwen\" || family === \"phi\",\n    family,\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"],"mappings":";AAYA,MAAaA,iBAA8C;CACzD,cAAc;EACZ,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,gBAAgB;EACd,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,sBAAsB;EACpB,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,gBAAgB;EACd,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,gBAAgB;EACd,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACD,cAAc;EACZ,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,eAAe;EACf,kBAAkB;EAClB,cAAc;EACd,QAAQ;EACT;CACF;;;;;;;;;;AAeD,SAAgB,aAAa,SAA8B;AAEzD,KAAI,eAAe,SACjB,QAAO;EACL,MAAM;EACN,MAAM,eAAe,SAAS;EAC/B;AAIH,KAAI,QAAQ,WAAW,MAAM,CAE3B,QAAO;EACL,MAAM;EACN,MAHW,QAAQ,MAAM,EAAE;EAI5B;AAIH,KAAI,QAAQ,WAAW,0BAA0B,CAE/C,QAAO;EACL,MAAM;EACN,MAHW,QAAQ,QAAQ,2BAA2B,GAAG;EAI1D;AAIH,KAAI,QAAQ,WAAW,QAAQ,CAE7B,QAAO;EACL,MAAM;EACN,MAHW,QAAQ,MAAM,EAAE;EAI5B;AAIH,KAAI,QAAQ,SAAS,IAAI,CACvB,QAAO;EACL,MAAM;EACN,MAAM;EACP;AAIH,QAAO;EACL,MAAM;EACN,MAAM;EACP;;;;;AAMH,SAAgB,eAAe,SAAqC;AAClE,QAAO,eAAe,YAAY;;;;;AAMpC,SAAgB,0BAA0B,SAAiB,MAA2B;CAEpF,IAAIC,SAAgC;CACpC,MAAM,YAAY,KAAK,aAAa;AAEpC,KAAI,UAAU,SAAS,OAAO,CAC5B,UAAS;UACA,UAAU,SAAS,SAAS,CACrC,UAAS;UACA,UAAU,SAAS,MAAM,CAClC,UAAS;UACA,UAAU,SAAS,UAAU,CACtC,UAAS;UACA,UAAU,SAAS,QAAQ,CACpC,UAAS;AAGX,QAAO;EACL,IAAI;EACJ;EACA,aAAa,mBAAmB;EAChC,MAAM;EACN,eAAe;EACf,kBAAkB,WAAW;EAC7B,cAAc,WAAW,UAAU,WAAW;EAC9C;EACD;;;;;AAMH,SAAgB,oBAAmC;AACjD,QAAO,OAAO,OAAO,eAAe"}

package/dist/models-De2-_GmQ.d.mts.map

@@ -1 +0,0 @@
-{"version":3,"file":"models-De2-_GmQ.d.mts","names":[],"sources":["../src/core/models.ts"],"sourcesContent":[],"mappings":";;;;cAYa,gBAAgB,eAAe;;;;;;;;;;iBA4E5B,YAAA,mBAA+B;;;;iBA6F/B,iBAAA,CAAA,GAAqB"}