agentic-flow 3.0.0-alpha.2 → 3.1.0

package/dist/router/providers/onnx-phi4.js

@@ -5,6 +5,7 @@
  * when local ONNX model is not available
  */
 import { HfInference } from '@huggingface/inference';
+import { PHI4_MODEL_PATH } from '../../utils/model-downloader.js';
 export class ONNXPhi4Provider {
     name = 'onnx-phi4';
     type = 'custom';
@@ -13,7 +14,7 @@ export class ONNXPhi4Provider {
     supportsMCP = false;
     config;
     hf;
-    modelPath = './models/phi-4/cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/model.onnx';
+    modelPath = PHI4_MODEL_PATH;
     constructor(config = {}) {
         this.config = {
             modelId: config.modelId || 'microsoft/Phi-3-mini-4k-instruct',

package/dist/router/providers/onnx-phi4.js.map

@@ -1 +1 @@
-{"version":3,"file":"onnx-phi4.js","sourceRoot":"","sources":["../../../src/router/providers/onnx-phi4.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,wBAAwB,CAAC;AAmBrD,MAAM,OAAO,gBAAgB;IAC3B,IAAI,GAAG,WAAW,CAAC;IACnB,IAAI,GAAG,QAAiB,CAAC;IACzB,iBAAiB,GAAG,IAAI,CAAC;IACzB,aAAa,GAAG,KAAK,CAAC;IACtB,WAAW,GAAG,KAAK,CAAC;IAEZ,MAAM,CAA2B;IACjC,EAAE,CAAc;IAChB,SAAS,GAAG,4EAA4E,CAAC;IAEjG,YAAY,SAAyB,EAAE;QACrC,IAAI,CAAC,MAAM,GAAG;YACZ,OAAO,EAAE,MAAM,CAAC,OAAO,IAAI,kCAAkC;YAC7D,YAAY,EAAE,MAAM,CAAC,YAAY,IAAI,KAAK,EAAE,8CAA8C;YAC1F,iBAAiB,EAAE,MAAM,CAAC,iBAAiB,IAAI,OAAO,CAAC,GAAG,CAAC,mBAAmB,IAAI,EAAE;YACpF,SAAS,EAAE,MAAM,CAAC,SAAS,IAAI,GAAG;YAClC,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,GAAG;SACvC,CAAC;QAEF,IAAI,CAAC,EAAE,GAAG,IAAI,WAAW,CAAC,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC,CAAC;IAC3D,CAAC;IAED;;OAEG;IACK,cAAc,CAAC,QAAmB;QACxC,IAAI,MAAM,GAAG,EAAE,CAAC;QAEhB,KAAK,MAAM,GAAG,IAAI,QAAQ,EAAE,CAAC;YAC3B,MAAM,OAAO,GAAG,OAAO,GAAG,CAAC,OAAO,KAAK,QAAQ;gBAC7C,CAAC,CAAC,GAAG,CAAC,OAAO;gBACb,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAEnE,IAAI,GAAG,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;gBAC1B,MAAM,IAAI,eAAe,OAAO,WAAW,CAAC;YAC9C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBAC/B,MAAM,IAAI,aAAa,OAAO,WAAW,CAAC;YAC5C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;gBACpC,MAAM,IAAI,kBAAkB,OAAO,WAAW,CAAC;YACjD,CAAC;QACH,CAAC;QAED,MAAM,IAAI,iBAAiB,CAAC;QAC5B,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,UAAU,CAAC,MAAkB;QACzC,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,EAAE,CAAC,cAAc,CAAC;gBAC1C,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;gBAC1B,MAAM,EAAE,MAAM;gBACd,UAAU,EAAE;oBACV,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;oBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;oBAC1D,gBAAgB,EAAE,KAAK;oBACvB,SAAS,EAAE,IAAI;oBACf,KAAK,EAAE,GAAG;iBACX;aACF,CAAC,CAAC;YAEH,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;YACvC,MAAM,aAAa,GAAG,MAAM,CAAC,cAAc,CAAC;YAE5C,wBAAwB;YACxB,MAAM,iBAAiB,GAAG,aAAa;iBACpC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;iBACnB,IAAI,EAAE,CAAC;YAEV,wBAAwB;YACxB,MAAM,WAAW,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YACjD,MAAM,YAAY,GAAG,IAAI,CAAC,IAAI,CAAC,iBAAiB,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YAE7D,MAAM,OAAO,GAAmB,CAAC;oBAC/B,IAAI,EAAE,MAAM;oBACZ,IAAI,EAAE,iBAAiB;iBACxB,CAAC,CAAC;YAEH,OAAO;gBACL,EAAE,EAAE,aAAa,IAAI,CAAC,GAAG,EAAE,EAAE;gBAC7B,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;gBAC1B,OAAO;gBACP,UAAU,EAAE,UAAU;gBACtB,KAAK,EAAE;oBACL,WAAW;oBACX,YAAY;iBACb;gBACD,QAAQ,EAAE;oBACR,QAAQ,EAAE,WAAW;oBACrB,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;oBAC1B,OAAO;oBACP,IAAI,EAAE,YAAY,GAAG,QAAQ,EAAE,iBAAiB;oBAChD,IAAI,EAAE,KAAK;iBACZ;aACF,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,kBAAkB;gBACxB,OAAO,EAAE,qCAAqC,KAAK,EAAE;gBACrD,QAAQ,EAAE,WAAW;gBACrB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,WAAW,CAAC,MAAkB;QAC1C,MAAM,IAAI,KAAK,CAAC,2EAA2E,CAAC,CAAC;IAC/F,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,IAAI,CAAC,MAAkB;QAC3B,IAAI,IAAI,CAAC,MAAM,CAAC,YAAY,EAAE,CAAC;YAC7B,OAAO,IAAI,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC;QAClC,CAAC;aAAM,CAAC;YACN,OAAO,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;QACjC,CAAC;IACH,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,CAAC,MAAM,CAAC,MAAkB;QAC9B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,IAAI,CAAC,EAAE,CAAC,oBAAoB,CAAC;gBAC1C,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;gBAC1B,MAAM,EAAE,MAAM;gBACd,UAAU,EAAE;oBACV,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;oBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;oBAC1D,gBAAgB,EAAE,KAAK;iBACxB;aACF,CAAC,CAAC;YAEH,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;gBACjC,IAAI,KAAK,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;oBACrB,MAAM;wBACJ,IAAI,EAAE,qBAAqB;wBAC3B,KAAK,EAAE;4BACL,IAAI,EAAE,YAAY;4BAClB,IAAI,EAAE,KAAK,CAAC,KAAK,CAAC,IAAI;yBACvB;qBACF,CAAC;gBACJ,CAAC;YACH,CAAC;YAED,MAAM;gBACJ,IAAI,EAAE,cAAc;aACrB,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,qBAAqB;gBAC3B,OAAO,EAAE,qBAAqB,KAAK,EAAE;gBACrC,QAAQ,EAAE,WAAW;gBACrB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACH,oBAAoB,CAAC,QAAkB;QACrC,MAAM,SAAS,GAAG,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;QACrC,OAAO,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;IACpD,CAAC;IAED;;OAEG;IACH,YAAY;QACV,OAAO;YACL,OAAO,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;YAC5B,IAAI,EAAE,IAAI,CAAC,MAAM,CAAC,YAAY,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,KAAK;YACrD,sBAAsB,EAAE,KAAK,EAAE,+CAA+C;YAC9E,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,MAAM,EAAE,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS;SAC1D,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,OAAO,CAAC,YAAqB;QAC3B,IAAI,CAAC,MAAM,CAAC,YAAY,GAAG,YAAY,CAAC;IAC1C,CAAC;CACF","sourcesContent":["/**\n * ONNX Runtime Provider for Phi-4 Model\n *\n * Hybrid implementation with fallback to HuggingFace Inference API\n * when local ONNX model is not available\n */\n\nimport { HfInference } from '@huggingface/inference';\nimport type {\n  LLMProvider,\n  ChatParams,\n  ChatResponse,\n  StreamChunk,\n  ProviderError,\n  Message,\n  ContentBlock\n} from '../types.js';\n\nexport interface ONNXPhi4Config {\n  modelId?: string;\n  useLocalONNX?: boolean;\n  huggingfaceApiKey?: string;\n  maxTokens?: number;\n  temperature?: number;\n}\n\nexport class ONNXPhi4Provider implements LLMProvider {\n  name = 'onnx-phi4';\n  type = 'custom' as const;\n  supportsStreaming = true;\n  supportsTools = false;\n  supportsMCP = false;\n\n  private config: Required<ONNXPhi4Config>;\n  private hf: HfInference;\n  private modelPath = './models/phi-4/cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/model.onnx';\n\n  constructor(config: ONNXPhi4Config = {}) {\n    this.config = {\n      modelId: config.modelId || 'microsoft/Phi-3-mini-4k-instruct',\n      useLocalONNX: config.useLocalONNX ?? false, // Default to API until local model downloaded\n      huggingfaceApiKey: config.huggingfaceApiKey || process.env.HUGGINGFACE_API_KEY || '',\n      maxTokens: config.maxTokens || 512,\n      temperature: config.temperature || 0.7\n    };\n\n    this.hf = new HfInference(this.config.huggingfaceApiKey);\n  }\n\n  /**\n   * Format messages for Phi-4 chat template\n   */\n  private formatMessages(messages: Message[]): string {\n    let prompt = '';\n\n    for (const msg of messages) {\n      const content = typeof msg.content === 'string'\n        ? msg.content\n        : msg.content.map(c => c.type === 'text' ? c.text : '').join('');\n\n      if (msg.role === 'system') {\n        prompt += `<|system|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'user') {\n        prompt += `<|user|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'assistant') {\n        prompt += `<|assistant|>\\n${content}<|end|>\\n`;\n      }\n    }\n\n    prompt += '<|assistant|>\\n';\n    return prompt;\n  }\n\n  /**\n   * Chat completion via HuggingFace Inference API\n   */\n  private async chatViaAPI(params: ChatParams): Promise<ChatResponse> {\n    const startTime = Date.now();\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      const result = await this.hf.textGeneration({\n        model: this.config.modelId,\n        inputs: prompt,\n        parameters: {\n          max_new_tokens: params.maxTokens || this.config.maxTokens,\n          temperature: params.temperature || this.config.temperature,\n          return_full_text: false,\n          do_sample: true,\n          top_p: 0.9\n        }\n      });\n\n      const latency = Date.now() - startTime;\n      const generatedText = result.generated_text;\n\n      // Clean up the response\n      const assistantResponse = generatedText\n        .split('<|end|>')[0]\n        .trim();\n\n      // Estimate token counts\n      const inputTokens = Math.ceil(prompt.length / 4);\n      const outputTokens = Math.ceil(assistantResponse.length / 4);\n\n      const content: ContentBlock[] = [{\n        type: 'text',\n        text: assistantResponse\n      }];\n\n      return {\n        id: `onnx-phi4-${Date.now()}`,\n        model: this.config.modelId,\n        content,\n        stopReason: 'end_turn',\n        usage: {\n          inputTokens,\n          outputTokens\n        },\n        metadata: {\n          provider: 'onnx-phi4',\n          model: this.config.modelId,\n          latency,\n          cost: outputTokens * 0.000002, // Rough estimate\n          mode: 'api'\n        }\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXPhi4APIError',\n        message: `HuggingFace API inference failed: ${error}`,\n        provider: 'onnx-phi4',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Chat completion via local ONNX (not yet implemented)\n   */\n  private async chatViaONNX(params: ChatParams): Promise<ChatResponse> {\n    throw new Error('Local ONNX inference not yet implemented. Download model.onnx.data first.');\n  }\n\n  /**\n   * Chat completion (uses API or local ONNX based on config)\n   */\n  async chat(params: ChatParams): Promise<ChatResponse> {\n    if (this.config.useLocalONNX) {\n      return this.chatViaONNX(params);\n    } else {\n      return this.chatViaAPI(params);\n    }\n  }\n\n  /**\n   * Streaming generation via HuggingFace API\n   */\n  async *stream(params: ChatParams): AsyncGenerator<StreamChunk> {\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      const stream = this.hf.textGenerationStream({\n        model: this.config.modelId,\n        inputs: prompt,\n        parameters: {\n          max_new_tokens: params.maxTokens || this.config.maxTokens,\n          temperature: params.temperature || this.config.temperature,\n          return_full_text: false\n        }\n      });\n\n      for await (const chunk of stream) {\n        if (chunk.token.text) {\n          yield {\n            type: 'content_block_delta',\n            delta: {\n              type: 'text_delta',\n              text: chunk.token.text\n            }\n          };\n        }\n      }\n\n      yield {\n        type: 'message_stop'\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXPhi4StreamError',\n        message: `Streaming failed: ${error}`,\n        provider: 'onnx-phi4',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Validate capabilities\n   */\n  validateCapabilities(features: string[]): boolean {\n    const supported = ['chat', 'stream'];\n    return features.every(f => supported.includes(f));\n  }\n\n  /**\n   * Get model info\n   */\n  getModelInfo() {\n    return {\n      modelId: this.config.modelId,\n      mode: this.config.useLocalONNX ? 'local-onnx' : 'api',\n      supportsLocalInference: false, // Will be true when model.onnx.data downloaded\n      modelPath: this.modelPath,\n      apiKey: this.config.huggingfaceApiKey ? '***' : undefined\n    };\n  }\n\n  /**\n   * Switch between API and local ONNX\n   */\n  setMode(useLocalONNX: boolean) {\n    this.config.useLocalONNX = useLocalONNX;\n  }\n}\n"]}
+{"version":3,"file":"onnx-phi4.js","sourceRoot":"","sources":["../../../src/router/providers/onnx-phi4.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,wBAAwB,CAAC;AACrD,OAAO,EAAE,eAAe,EAAE,MAAM,iCAAiC,CAAC;AAmBlE,MAAM,OAAO,gBAAgB;IAC3B,IAAI,GAAG,WAAW,CAAC;IACnB,IAAI,GAAG,QAAiB,CAAC;IACzB,iBAAiB,GAAG,IAAI,CAAC;IACzB,aAAa,GAAG,KAAK,CAAC;IACtB,WAAW,GAAG,KAAK,CAAC;IAEZ,MAAM,CAA2B;IACjC,EAAE,CAAc;IAChB,SAAS,GAAG,eAAe,CAAC;IAEpC,YAAY,SAAyB,EAAE;QACrC,IAAI,CAAC,MAAM,GAAG;YACZ,OAAO,EAAE,MAAM,CAAC,OAAO,IAAI,kCAAkC;YAC7D,YAAY,EAAE,MAAM,CAAC,YAAY,IAAI,KAAK,EAAE,8CAA8C;YAC1F,iBAAiB,EAAE,MAAM,CAAC,iBAAiB,IAAI,OAAO,CAAC,GAAG,CAAC,mBAAmB,IAAI,EAAE;YACpF,SAAS,EAAE,MAAM,CAAC,SAAS,IAAI,GAAG;YAClC,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,GAAG;SACvC,CAAC;QAEF,IAAI,CAAC,EAAE,GAAG,IAAI,WAAW,CAAC,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC,CAAC;IAC3D,CAAC;IAED;;OAEG;IACK,cAAc,CAAC,QAAmB;QACxC,IAAI,MAAM,GAAG,EAAE,CAAC;QAEhB,KAAK,MAAM,GAAG,IAAI,QAAQ,EAAE,CAAC;YAC3B,MAAM,OAAO,GAAG,OAAO,GAAG,CAAC,OAAO,KAAK,QAAQ;gBAC7C,CAAC,CAAC,GAAG,CAAC,OAAO;gBACb,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAEnE,IAAI,GAAG,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;gBAC1B,MAAM,IAAI,eAAe,OAAO,WAAW,CAAC;YAC9C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBAC/B,MAAM,IAAI,aAAa,OAAO,WAAW,CAAC;YAC5C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;gBACpC,MAAM,IAAI,kBAAkB,OAAO,WAAW,CAAC;YACjD,CAAC;QACH,CAAC;QAED,MAAM,IAAI,iBAAiB,CAAC;QAC5B,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,UAAU,CAAC,MAAkB;QACzC,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,EAAE,CAAC,cAAc,CAAC;gBAC1C,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;gBAC1B,MAAM,EAAE,MAAM;gBACd,UAAU,EAAE;oBACV,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;oBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;oBAC1D,gBAAgB,EAAE,KAAK;oBACvB,SAAS,EAAE,IAAI;oBACf,KAAK,EAAE,GAAG;iBACX;aACF,CAAC,CAAC;YAEH,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;YACvC,MAAM,aAAa,GAAG,MAAM,CAAC,cAAc,CAAC;YAE5C,wBAAwB;YACxB,MAAM,iBAAiB,GAAG,aAAa;iBACpC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;iBACnB,IAAI,EAAE,CAAC;YAEV,wBAAwB;YACxB,MAAM,WAAW,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YACjD,MAAM,YAAY,GAAG,IAAI,CAAC,IAAI,CAAC,iBAAiB,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YAE7D,MAAM,OAAO,GAAmB,CAAC;oBAC/B,IAAI,EAAE,MAAM;oBACZ,IAAI,EAAE,iBAAiB;iBACxB,CAAC,CAAC;YAEH,OAAO;gBACL,EAAE,EAAE,aAAa,IAAI,CAAC,GAAG,EAAE,EAAE;gBAC7B,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;gBAC1B,OAAO;gBACP,UAAU,EAAE,UAAU;gBACtB,KAAK,EAAE;oBACL,WAAW;oBACX,YAAY;iBACb;gBACD,QAAQ,EAAE;oBACR,QAAQ,EAAE,WAAW;oBACrB,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;oBAC1B,OAAO;oBACP,IAAI,EAAE,YAAY,GAAG,QAAQ,EAAE,iBAAiB;oBAChD,IAAI,EAAE,KAAK;iBACZ;aACF,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,kBAAkB;gBACxB,OAAO,EAAE,qCAAqC,KAAK,EAAE;gBACrD,QAAQ,EAAE,WAAW;gBACrB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,WAAW,CAAC,MAAkB;QAC1C,MAAM,IAAI,KAAK,CAAC,2EAA2E,CAAC,CAAC;IAC/F,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,IAAI,CAAC,MAAkB;QAC3B,IAAI,IAAI,CAAC,MAAM,CAAC,YAAY,EAAE,CAAC;YAC7B,OAAO,IAAI,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC;QAClC,CAAC;aAAM,CAAC;YACN,OAAO,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;QACjC,CAAC;IACH,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,CAAC,MAAM,CAAC,MAAkB;QAC9B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,IAAI,CAAC,EAAE,CAAC,oBAAoB,CAAC;gBAC1C,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;gBAC1B,MAAM,EAAE,MAAM;gBACd,UAAU,EAAE;oBACV,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;oBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;oBAC1D,gBAAgB,EAAE,KAAK;iBACxB;aACF,CAAC,CAAC;YAEH,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;gBACjC,IAAI,KAAK,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;oBACrB,MAAM;wBACJ,IAAI,EAAE,qBAAqB;wBAC3B,KAAK,EAAE;4BACL,IAAI,EAAE,YAAY;4BAClB,IAAI,EAAE,KAAK,CAAC,KAAK,CAAC,IAAI;yBACvB;qBACF,CAAC;gBACJ,CAAC;YACH,CAAC;YAED,MAAM;gBACJ,IAAI,EAAE,cAAc;aACrB,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,qBAAqB;gBAC3B,OAAO,EAAE,qBAAqB,KAAK,EAAE;gBACrC,QAAQ,EAAE,WAAW;gBACrB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACH,oBAAoB,CAAC,QAAkB;QACrC,MAAM,SAAS,GAAG,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;QACrC,OAAO,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;IACpD,CAAC;IAED;;OAEG;IACH,YAAY;QACV,OAAO;YACL,OAAO,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;YAC5B,IAAI,EAAE,IAAI,CAAC,MAAM,CAAC,YAAY,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,KAAK;YACrD,sBAAsB,EAAE,KAAK,EAAE,+CAA+C;YAC9E,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,MAAM,EAAE,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS;SAC1D,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,OAAO,CAAC,YAAqB;QAC3B,IAAI,CAAC,MAAM,CAAC,YAAY,GAAG,YAAY,CAAC;IAC1C,CAAC;CACF","sourcesContent":["/**\n * ONNX Runtime Provider for Phi-4 Model\n *\n * Hybrid implementation with fallback to HuggingFace Inference API\n * when local ONNX model is not available\n */\n\nimport { HfInference } from '@huggingface/inference';\nimport { PHI4_MODEL_PATH } from '../../utils/model-downloader.js';\nimport type {\n  LLMProvider,\n  ChatParams,\n  ChatResponse,\n  StreamChunk,\n  ProviderError,\n  Message,\n  ContentBlock\n} from '../types.js';\n\nexport interface ONNXPhi4Config {\n  modelId?: string;\n  useLocalONNX?: boolean;\n  huggingfaceApiKey?: string;\n  maxTokens?: number;\n  temperature?: number;\n}\n\nexport class ONNXPhi4Provider implements LLMProvider {\n  name = 'onnx-phi4';\n  type = 'custom' as const;\n  supportsStreaming = true;\n  supportsTools = false;\n  supportsMCP = false;\n\n  private config: Required<ONNXPhi4Config>;\n  private hf: HfInference;\n  private modelPath = PHI4_MODEL_PATH;\n\n  constructor(config: ONNXPhi4Config = {}) {\n    this.config = {\n      modelId: config.modelId || 'microsoft/Phi-3-mini-4k-instruct',\n      useLocalONNX: config.useLocalONNX ?? false, // Default to API until local model downloaded\n      huggingfaceApiKey: config.huggingfaceApiKey || process.env.HUGGINGFACE_API_KEY || '',\n      maxTokens: config.maxTokens || 512,\n      temperature: config.temperature || 0.7\n    };\n\n    this.hf = new HfInference(this.config.huggingfaceApiKey);\n  }\n\n  /**\n   * Format messages for Phi-4 chat template\n   */\n  private formatMessages(messages: Message[]): string {\n    let prompt = '';\n\n    for (const msg of messages) {\n      const content = typeof msg.content === 'string'\n        ? msg.content\n        : msg.content.map(c => c.type === 'text' ? c.text : '').join('');\n\n      if (msg.role === 'system') {\n        prompt += `<|system|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'user') {\n        prompt += `<|user|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'assistant') {\n        prompt += `<|assistant|>\\n${content}<|end|>\\n`;\n      }\n    }\n\n    prompt += '<|assistant|>\\n';\n    return prompt;\n  }\n\n  /**\n   * Chat completion via HuggingFace Inference API\n   */\n  private async chatViaAPI(params: ChatParams): Promise<ChatResponse> {\n    const startTime = Date.now();\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      const result = await this.hf.textGeneration({\n        model: this.config.modelId,\n        inputs: prompt,\n        parameters: {\n          max_new_tokens: params.maxTokens || this.config.maxTokens,\n          temperature: params.temperature || this.config.temperature,\n          return_full_text: false,\n          do_sample: true,\n          top_p: 0.9\n        }\n      });\n\n      const latency = Date.now() - startTime;\n      const generatedText = result.generated_text;\n\n      // Clean up the response\n      const assistantResponse = generatedText\n        .split('<|end|>')[0]\n        .trim();\n\n      // Estimate token counts\n      const inputTokens = Math.ceil(prompt.length / 4);\n      const outputTokens = Math.ceil(assistantResponse.length / 4);\n\n      const content: ContentBlock[] = [{\n        type: 'text',\n        text: assistantResponse\n      }];\n\n      return {\n        id: `onnx-phi4-${Date.now()}`,\n        model: this.config.modelId,\n        content,\n        stopReason: 'end_turn',\n        usage: {\n          inputTokens,\n          outputTokens\n        },\n        metadata: {\n          provider: 'onnx-phi4',\n          model: this.config.modelId,\n          latency,\n          cost: outputTokens * 0.000002, // Rough estimate\n          mode: 'api'\n        }\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXPhi4APIError',\n        message: `HuggingFace API inference failed: ${error}`,\n        provider: 'onnx-phi4',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Chat completion via local ONNX (not yet implemented)\n   */\n  private async chatViaONNX(params: ChatParams): Promise<ChatResponse> {\n    throw new Error('Local ONNX inference not yet implemented. Download model.onnx.data first.');\n  }\n\n  /**\n   * Chat completion (uses API or local ONNX based on config)\n   */\n  async chat(params: ChatParams): Promise<ChatResponse> {\n    if (this.config.useLocalONNX) {\n      return this.chatViaONNX(params);\n    } else {\n      return this.chatViaAPI(params);\n    }\n  }\n\n  /**\n   * Streaming generation via HuggingFace API\n   */\n  async *stream(params: ChatParams): AsyncGenerator<StreamChunk> {\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      const stream = this.hf.textGenerationStream({\n        model: this.config.modelId,\n        inputs: prompt,\n        parameters: {\n          max_new_tokens: params.maxTokens || this.config.maxTokens,\n          temperature: params.temperature || this.config.temperature,\n          return_full_text: false\n        }\n      });\n\n      for await (const chunk of stream) {\n        if (chunk.token.text) {\n          yield {\n            type: 'content_block_delta',\n            delta: {\n              type: 'text_delta',\n              text: chunk.token.text\n            }\n          };\n        }\n      }\n\n      yield {\n        type: 'message_stop'\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXPhi4StreamError',\n        message: `Streaming failed: ${error}`,\n        provider: 'onnx-phi4',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Validate capabilities\n   */\n  validateCapabilities(features: string[]): boolean {\n    const supported = ['chat', 'stream'];\n    return features.every(f => supported.includes(f));\n  }\n\n  /**\n   * Get model info\n   */\n  getModelInfo() {\n    return {\n      modelId: this.config.modelId,\n      mode: this.config.useLocalONNX ? 'local-onnx' : 'api',\n      supportsLocalInference: false, // Will be true when model.onnx.data downloaded\n      modelPath: this.modelPath,\n      apiKey: this.config.huggingfaceApiKey ? '***' : undefined\n    };\n  }\n\n  /**\n   * Switch between API and local ONNX\n   */\n  setMode(useLocalONNX: boolean) {\n    this.config.useLocalONNX = useLocalONNX;\n  }\n}\n"]}

package/dist/router/providers/onnx.js

@@ -19,12 +19,12 @@ async function ensureOnnxDependencies() {
     }
     if (!transformers) {
         try {
-            const transformersModule = await import('@xenova/transformers');
+            const transformersModule = await import('@huggingface/transformers');
             transformers = transformersModule;
             transformers.env.allowLocalModels = true;
         }
         catch (e) {
-            throw new Error('@xenova/transformers not installed. Run: npm install @xenova/transformers');
+            throw new Error('@huggingface/transformers not installed. Run: npm install @huggingface/transformers');
         }
     }
 }

package/dist/router/providers/onnx.js.map

@@ -1 +1 @@
-{"version":3,"file":"onnx.js","sourceRoot":"","sources":["../../../src/router/providers/onnx.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAYH,iDAAiD;AACjD,IAAI,GAAQ,CAAC;AACb,IAAI,YAAiB,CAAC;AAEtB,KAAK,UAAU,sBAAsB;IACnC,IAAI,CAAC,GAAG,EAAE,CAAC;QACT,IAAI,CAAC;YACH,MAAM,SAAS,GAAG,MAAM,MAAM,CAAC,kBAAyB,CAAC,CAAC;YAC1D,GAAG,GAAG,SAAS,CAAC;QAClB,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,MAAM,IAAI,KAAK,CAAC,mEAAmE,CAAC,CAAC;QACvF,CAAC;IACH,CAAC;IACD,IAAI,CAAC,YAAY,EAAE,CAAC;QAClB,IAAI,CAAC;YACH,MAAM,kBAAkB,GAAG,MAAM,MAAM,CAAC,sBAA6B,CAAC,CAAC;YACvE,YAAY,GAAG,kBAAkB,CAAC;YAClC,YAAY,CAAC,GAAG,CAAC,gBAAgB,GAAG,IAAI,CAAC;QAC3C,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,MAAM,IAAI,KAAK,CAAC,2EAA2E,CAAC,CAAC;QAC/F,CAAC;IACH,CAAC;AACH,CAAC;AAWD,MAAM,OAAO,YAAY;IACvB,IAAI,GAAG,MAAM,CAAC;IACd,IAAI,GAAG,QAAiB,CAAC;IACzB,iBAAiB,GAAG,IAAI,CAAC;IACzB,aAAa,GAAG,KAAK,CAAC;IACtB,WAAW,GAAG,KAAK,CAAC;IAEZ,OAAO,GAAQ,IAAI,CAAC;IACpB,SAAS,GAAQ,IAAI,CAAC;IACtB,MAAM,CAAa;IACnB,kBAAkB,GAAa,EAAE,CAAC;IAE1C,YAAY,SAAqB,EAAE;QACjC,IAAI,CAAC,MAAM,GAAG;YACZ,OAAO,EAAE,MAAM,CAAC,OAAO,IAAI,+BAA+B;YAC1D,SAAS,EAAE,MAAM,CAAC,SAAS,IAAI,GAAG;YAClC,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,GAAG;YACtC,GAAG,MAAM;SACV,CAAC;IACJ,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,wBAAwB;QACpC,MAAM,SAAS,GAAa,EAAE,CAAC;QAE/B,2BAA2B;QAC3B,IAAI,CAAC;YACH,IAAI,OAAO,CAAC,QAAQ,KAAK,OAAO,EAAE,CAAC;gBACjC,SAAS,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;gBACvB,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;YACvC,CAAC;QACH,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,qBAAqB;QACvB,CAAC;QAED,gCAAgC;QAChC,IAAI,CAAC;YACH,IAAI,OAAO,CAAC,QAAQ,KAAK,OAAO,EAAE,CAAC;gBACjC,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;gBACtB,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YACtC,CAAC;QACH,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,yBAAyB;QAC3B,CAAC;QAED,yBAAyB;QACzB,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACtB,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QAEpC,OAAO,CAAC,GAAG,CAAC,gCAAgC,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QAElF,OAAO,SAAS,CAAC;IACnB,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,iBAAiB;QAC7B,IAAI,IAAI,CAAC,SAAS;YAAE,OAAO;QAE3B,IAAI,CAAC;YACH,MAAM,sBAAsB,EAAE,CAAC;YAE/B,OAAO,CAAC,GAAG,CAAC,0BAA0B,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;YAE7D,+CAA+C;YAC/C,IAAI,CAAC,SAAS,GAAG,MAAM,YAAY,CAAC,QAAQ,CAC1C,iBAAiB,EACjB,IAAI,CAAC,MAAM,CAAC,OAAO,EACnB;gBACE,SAAS,EAAE,IAAI,EAAE,kDAAkD;aACpE,CACF,CAAC;YAEF,OAAO,CAAC,GAAG,CAAC,kCAAkC,CAAC,CAAC;QAElD,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,eAAe;gBACrB,OAAO,EAAE,oCAAoC,KAAK,EAAE;gBACpD,QAAQ,EAAE,MAAM;gBAChB,SAAS,EAAE,KAAK;aACjB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACK,cAAc,CAAC,QAAmB;QACxC,iCAAiC;QACjC,IAAI,MAAM,GAAG,EAAE,CAAC;QAEhB,KAAK,MAAM,GAAG,IAAI,QAAQ,EAAE,CAAC;YAC3B,MAAM,OAAO,GAAG,OAAO,GAAG,CAAC,OAAO,KAAK,QAAQ;gBAC7C,CAAC,CAAC,GAAG,CAAC,OAAO;gBACb,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAEnE,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBACxB,MAAM,IAAI,aAAa,OAAO,WAAW,CAAC;YAC5C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;gBACpC,MAAM,IAAI,kBAAkB,OAAO,WAAW,CAAC;YACjD,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;gBACjC,MAAM,IAAI,eAAe,OAAO,WAAW,CAAC;YAC9C,CAAC;QACH,CAAC;QAED,MAAM,IAAI,iBAAiB,CAAC;QAC5B,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,IAAI,CAAC,MAAkB;QAC3B,MAAM,IAAI,CAAC,iBAAiB,EAAE,CAAC;QAE/B,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE;gBAC1C,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;gBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;gBAC1D,SAAS,EAAE,IAAI;gBACf,KAAK,EAAE,GAAG;aACX,CAAC,CAAC;YAEH,MAAM,aAAa,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,cAAc,CAAC;YAE/C,0CAA0C;YAC1C,MAAM,iBAAiB,GAAG,aAAa;iBACpC,KAAK,CAAC,eAAe,CAAC;iBACtB,GAAG,EAAE;gBACN,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;gBACrB,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC;YAEjB,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;YAEvC,8CAA8C;YAC9C,MAAM,WAAW,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YACjD,MAAM,YAAY,GAAG,IAAI,CAAC,IAAI,CAAC,iBAAiB,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YAE7D,MAAM,OAAO,GAAmB,CAAC;oBAC/B,IAAI,EAAE,MAAM;oBACZ,IAAI,EAAE,iBAAiB;iBACxB,CAAC,CAAC;YAEH,OAAO;gBACL,EAAE,EAAE,QAAQ,IAAI,CAAC,GAAG,EAAE,EAAE;gBACxB,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO,IAAI,YAAY;gBAC1C,OAAO;gBACP,UAAU,EAAE,UAAU;gBACtB,KAAK,EAAE;oBACL,WAAW;oBACX,YAAY;iBACb;gBACD,QAAQ,EAAE;oBACR,QAAQ,EAAE,MAAM;oBAChB,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;oBAC1B,OAAO;oBACP,IAAI,EAAE,CAAC,EAAE,0BAA0B;oBACnC,kBAAkB,EAAE,IAAI,CAAC,kBAAkB;iBAC5C;aACF,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,oBAAoB;gBAC1B,OAAO,EAAE,0BAA0B,KAAK,EAAE;gBAC1C,QAAQ,EAAE,MAAM;gBAChB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,CAAC,MAAM,CAAC,MAAkB;QAC9B,MAAM,IAAI,CAAC,iBAAiB,EAAE,CAAC;QAE/B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,2DAA2D;YAC3D,4DAA4D;YAC5D,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE;gBAC1C,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;gBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;gBAC1D,SAAS,EAAE,IAAI;gBACf,KAAK,EAAE,GAAG;aACX,CAAC,CAAC;YAEH,MAAM,aAAa,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,cAAc,CAAC;YAC/C,MAAM,iBAAiB,GAAG,aAAa;iBACpC,KAAK,CAAC,eAAe,CAAC;iBACtB,GAAG,EAAE;gBACN,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;gBACrB,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC;YAEjB,8CAA8C;YAC9C,MAAM,KAAK,GAAG,iBAAiB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;YAC3C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBACtC,MAAM,KAAK,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;gBAC3D,MAAM;oBACJ,IAAI,EAAE,qBAAqB;oBAC3B,KAAK,EAAE;wBACL,IAAI,EAAE,YAAY;wBAClB,IAAI,EAAE,KAAK;qBACZ;iBACF,CAAC;gBAEF,yCAAyC;gBACzC,MAAM,IAAI,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,CAAC;YACxD,CAAC;YAED,MAAM;gBACJ,IAAI,EAAE,cAAc;aACrB,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,iBAAiB;gBACvB,OAAO,EAAE,0BAA0B,KAAK,EAAE;gBAC1C,QAAQ,EAAE,MAAM;gBAChB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACH,oBAAoB,CAAC,QAAkB;QACrC,MAAM,SAAS,GAAG,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;QACrC,OAAO,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;IACpD,CAAC;IAED;;OAEG;IACH,YAAY;QACV,OAAO;YACL,OAAO,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;YAC5B,kBAAkB,EAAE,IAAI,CAAC,kBAAkB;YAC3C,WAAW,EAAE,IAAI,CAAC,kBAAkB,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,IAAI,CAAC,kBAAkB,CAAC,QAAQ,CAAC,KAAK,CAAC;YAChG,WAAW,EAAE,IAAI,CAAC,SAAS,KAAK,IAAI;SACrC,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,OAAO;QACX,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YACnB,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC;QACxB,CAAC;QACD,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACjB,MAAM,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC;YAC7B,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC;QACtB,CAAC;IACH,CAAC;CACF","sourcesContent":["/**\n * ONNX Runtime Provider for Local Model Inference\n *\n * Supports CPU and GPU execution providers for optimized local inference\n * Compatible with Phi-3, Llama, and other ONNX models\n */\n\nimport type {\n  LLMProvider,\n  ChatParams,\n  ChatResponse,\n  StreamChunk,\n  ProviderError,\n  Message,\n  ContentBlock\n} from '../types.js';\n\n// Dynamic imports for optional ONNX dependencies\nlet ort: any;\nlet transformers: any;\n\nasync function ensureOnnxDependencies() {\n  if (!ort) {\n    try {\n      const ortModule = await import('onnxruntime-node' as any);\n      ort = ortModule;\n    } catch (e) {\n      throw new Error('onnxruntime-node not installed. Run: npm install onnxruntime-node');\n    }\n  }\n  if (!transformers) {\n    try {\n      const transformersModule = await import('@xenova/transformers' as any);\n      transformers = transformersModule;\n      transformers.env.allowLocalModels = true;\n    } catch (e) {\n      throw new Error('@xenova/transformers not installed. Run: npm install @xenova/transformers');\n    }\n  }\n}\n\nexport interface ONNXConfig {\n  modelPath?: string;\n  modelId?: string; // HuggingFace model ID\n  executionProviders?: string[];\n  sessionOptions?: any;\n  maxTokens?: number;\n  temperature?: number;\n}\n\nexport class ONNXProvider implements LLMProvider {\n  name = 'onnx';\n  type = 'custom' as const;\n  supportsStreaming = true;\n  supportsTools = false;\n  supportsMCP = false;\n\n  private session: any = null;\n  private generator: any = null;\n  private config: ONNXConfig;\n  private executionProviders: string[] = [];\n\n  constructor(config: ONNXConfig = {}) {\n    this.config = {\n      modelId: config.modelId || 'Xenova/Phi-3-mini-4k-instruct',\n      maxTokens: config.maxTokens || 512,\n      temperature: config.temperature || 0.7,\n      ...config\n    };\n  }\n\n  /**\n   * Detect available execution providers\n   */\n  private async detectExecutionProviders(): Promise<string[]> {\n    const providers: string[] = [];\n\n    // Try CUDA for NVIDIA GPUs\n    try {\n      if (process.platform === 'linux') {\n        providers.push('cuda');\n        this.executionProviders.push('cuda');\n      }\n    } catch (e) {\n      // CUDA not available\n    }\n\n    // Try DirectML for Windows GPUs\n    try {\n      if (process.platform === 'win32') {\n        providers.push('dml');\n        this.executionProviders.push('dml');\n      }\n    } catch (e) {\n      // DirectML not available\n    }\n\n    // Always fallback to CPU\n    providers.push('cpu');\n    this.executionProviders.push('cpu');\n\n    console.log(`🔧 ONNX Execution Providers: ${this.executionProviders.join(', ')}`);\n\n    return providers;\n  }\n\n  /**\n   * Initialize ONNX session with model\n   */\n  private async initializeSession(): Promise<void> {\n    if (this.generator) return;\n\n    try {\n      await ensureOnnxDependencies();\n\n      console.log(`📦 Loading ONNX model: ${this.config.modelId}`);\n\n      // Use Transformers.js for easier model loading\n      this.generator = await transformers.pipeline(\n        'text-generation',\n        this.config.modelId,\n        {\n          quantized: true, // Use quantized models for better CPU performance\n        }\n      );\n\n      console.log(`✅ ONNX model loaded successfully`);\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXInitError',\n        message: `Failed to initialize ONNX model: ${error}`,\n        provider: 'onnx',\n        retryable: false\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Format messages for model input\n   */\n  private formatMessages(messages: Message[]): string {\n    // Simple chat template for Phi-3\n    let prompt = '';\n\n    for (const msg of messages) {\n      const content = typeof msg.content === 'string'\n        ? msg.content\n        : msg.content.map(c => c.type === 'text' ? c.text : '').join('');\n\n      if (msg.role === 'user') {\n        prompt += `<|user|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'assistant') {\n        prompt += `<|assistant|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'system') {\n        prompt += `<|system|>\\n${content}<|end|>\\n`;\n      }\n    }\n\n    prompt += '<|assistant|>\\n';\n    return prompt;\n  }\n\n  /**\n   * Chat completion\n   */\n  async chat(params: ChatParams): Promise<ChatResponse> {\n    await this.initializeSession();\n\n    const startTime = Date.now();\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      const result = await this.generator(prompt, {\n        max_new_tokens: params.maxTokens || this.config.maxTokens,\n        temperature: params.temperature || this.config.temperature,\n        do_sample: true,\n        top_p: 0.9,\n      });\n\n      const generatedText = result[0].generated_text;\n\n      // Extract only the new assistant response\n      const assistantResponse = generatedText\n        .split('<|assistant|>')\n        .pop()\n        ?.split('<|end|>')[0]\n        ?.trim() || '';\n\n      const latency = Date.now() - startTime;\n\n      // Estimate token counts (rough approximation)\n      const inputTokens = Math.ceil(prompt.length / 4);\n      const outputTokens = Math.ceil(assistantResponse.length / 4);\n\n      const content: ContentBlock[] = [{\n        type: 'text',\n        text: assistantResponse\n      }];\n\n      return {\n        id: `onnx-${Date.now()}`,\n        model: this.config.modelId || 'onnx-model',\n        content,\n        stopReason: 'end_turn',\n        usage: {\n          inputTokens,\n          outputTokens\n        },\n        metadata: {\n          provider: 'onnx',\n          model: this.config.modelId,\n          latency,\n          cost: 0, // Local inference is free\n          executionProviders: this.executionProviders\n        }\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXInferenceError',\n        message: `ONNX inference failed: ${error}`,\n        provider: 'onnx',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Streaming generation\n   */\n  async *stream(params: ChatParams): AsyncGenerator<StreamChunk> {\n    await this.initializeSession();\n\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      // Note: Transformers.js doesn't natively support streaming\n      // We'll simulate it by yielding tokens as they're generated\n      const result = await this.generator(prompt, {\n        max_new_tokens: params.maxTokens || this.config.maxTokens,\n        temperature: params.temperature || this.config.temperature,\n        do_sample: true,\n        top_p: 0.9,\n      });\n\n      const generatedText = result[0].generated_text;\n      const assistantResponse = generatedText\n        .split('<|assistant|>')\n        .pop()\n        ?.split('<|end|>')[0]\n        ?.trim() || '';\n\n      // Simulate streaming by chunking the response\n      const words = assistantResponse.split(' ');\n      for (let i = 0; i < words.length; i++) {\n        const chunk = words[i] + (i < words.length - 1 ? ' ' : '');\n        yield {\n          type: 'content_block_delta',\n          delta: {\n            type: 'text_delta',\n            text: chunk\n          }\n        };\n\n        // Small delay to simulate real streaming\n        await new Promise(resolve => setTimeout(resolve, 10));\n      }\n\n      yield {\n        type: 'message_stop'\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXStreamError',\n        message: `ONNX streaming failed: ${error}`,\n        provider: 'onnx',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Validate capabilities\n   */\n  validateCapabilities(features: string[]): boolean {\n    const supported = ['chat', 'stream'];\n    return features.every(f => supported.includes(f));\n  }\n\n  /**\n   * Get model info\n   */\n  getModelInfo() {\n    return {\n      modelId: this.config.modelId,\n      executionProviders: this.executionProviders,\n      supportsGPU: this.executionProviders.includes('cuda') || this.executionProviders.includes('dml'),\n      initialized: this.generator !== null\n    };\n  }\n\n  /**\n   * Cleanup resources\n   */\n  async dispose(): Promise<void> {\n    if (this.generator) {\n      this.generator = null;\n    }\n    if (this.session) {\n      await this.session.release();\n      this.session = null;\n    }\n  }\n}\n"]}
+{"version":3,"file":"onnx.js","sourceRoot":"","sources":["../../../src/router/providers/onnx.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAYH,iDAAiD;AACjD,IAAI,GAAQ,CAAC;AACb,IAAI,YAAiB,CAAC;AAEtB,KAAK,UAAU,sBAAsB;IACnC,IAAI,CAAC,GAAG,EAAE,CAAC;QACT,IAAI,CAAC;YACH,MAAM,SAAS,GAAG,MAAM,MAAM,CAAC,kBAAyB,CAAC,CAAC;YAC1D,GAAG,GAAG,SAAS,CAAC;QAClB,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,MAAM,IAAI,KAAK,CAAC,mEAAmE,CAAC,CAAC;QACvF,CAAC;IACH,CAAC;IACD,IAAI,CAAC,YAAY,EAAE,CAAC;QAClB,IAAI,CAAC;YACH,MAAM,kBAAkB,GAAG,MAAM,MAAM,CAAC,2BAAkC,CAAC,CAAC;YAC5E,YAAY,GAAG,kBAAkB,CAAC;YAClC,YAAY,CAAC,GAAG,CAAC,gBAAgB,GAAG,IAAI,CAAC;QAC3C,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,MAAM,IAAI,KAAK,CAAC,qFAAqF,CAAC,CAAC;QACzG,CAAC;IACH,CAAC;AACH,CAAC;AAWD,MAAM,OAAO,YAAY;IACvB,IAAI,GAAG,MAAM,CAAC;IACd,IAAI,GAAG,QAAiB,CAAC;IACzB,iBAAiB,GAAG,IAAI,CAAC;IACzB,aAAa,GAAG,KAAK,CAAC;IACtB,WAAW,GAAG,KAAK,CAAC;IAEZ,OAAO,GAAQ,IAAI,CAAC;IACpB,SAAS,GAAQ,IAAI,CAAC;IACtB,MAAM,CAAa;IACnB,kBAAkB,GAAa,EAAE,CAAC;IAE1C,YAAY,SAAqB,EAAE;QACjC,IAAI,CAAC,MAAM,GAAG;YACZ,OAAO,EAAE,MAAM,CAAC,OAAO,IAAI,+BAA+B;YAC1D,SAAS,EAAE,MAAM,CAAC,SAAS,IAAI,GAAG;YAClC,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,GAAG;YACtC,GAAG,MAAM;SACV,CAAC;IACJ,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,wBAAwB;QACpC,MAAM,SAAS,GAAa,EAAE,CAAC;QAE/B,2BAA2B;QAC3B,IAAI,CAAC;YACH,IAAI,OAAO,CAAC,QAAQ,KAAK,OAAO,EAAE,CAAC;gBACjC,SAAS,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;gBACvB,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;YACvC,CAAC;QACH,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,qBAAqB;QACvB,CAAC;QAED,gCAAgC;QAChC,IAAI,CAAC;YACH,IAAI,OAAO,CAAC,QAAQ,KAAK,OAAO,EAAE,CAAC;gBACjC,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;gBACtB,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YACtC,CAAC;QACH,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,yBAAyB;QAC3B,CAAC;QAED,yBAAyB;QACzB,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACtB,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QAEpC,OAAO,CAAC,GAAG,CAAC,gCAAgC,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QAElF,OAAO,SAAS,CAAC;IACnB,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,iBAAiB;QAC7B,IAAI,IAAI,CAAC,SAAS;YAAE,OAAO;QAE3B,IAAI,CAAC;YACH,MAAM,sBAAsB,EAAE,CAAC;YAE/B,OAAO,CAAC,GAAG,CAAC,0BAA0B,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;YAE7D,+CAA+C;YAC/C,IAAI,CAAC,SAAS,GAAG,MAAM,YAAY,CAAC,QAAQ,CAC1C,iBAAiB,EACjB,IAAI,CAAC,MAAM,CAAC,OAAO,EACnB;gBACE,SAAS,EAAE,IAAI,EAAE,kDAAkD;aACpE,CACF,CAAC;YAEF,OAAO,CAAC,GAAG,CAAC,kCAAkC,CAAC,CAAC;QAElD,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,eAAe;gBACrB,OAAO,EAAE,oCAAoC,KAAK,EAAE;gBACpD,QAAQ,EAAE,MAAM;gBAChB,SAAS,EAAE,KAAK;aACjB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACK,cAAc,CAAC,QAAmB;QACxC,iCAAiC;QACjC,IAAI,MAAM,GAAG,EAAE,CAAC;QAEhB,KAAK,MAAM,GAAG,IAAI,QAAQ,EAAE,CAAC;YAC3B,MAAM,OAAO,GAAG,OAAO,GAAG,CAAC,OAAO,KAAK,QAAQ;gBAC7C,CAAC,CAAC,GAAG,CAAC,OAAO;gBACb,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAEnE,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBACxB,MAAM,IAAI,aAAa,OAAO,WAAW,CAAC;YAC5C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;gBACpC,MAAM,IAAI,kBAAkB,OAAO,WAAW,CAAC;YACjD,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;gBACjC,MAAM,IAAI,eAAe,OAAO,WAAW,CAAC;YAC9C,CAAC;QACH,CAAC;QAED,MAAM,IAAI,iBAAiB,CAAC;QAC5B,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,IAAI,CAAC,MAAkB;QAC3B,MAAM,IAAI,CAAC,iBAAiB,EAAE,CAAC;QAE/B,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE;gBAC1C,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;gBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;gBAC1D,SAAS,EAAE,IAAI;gBACf,KAAK,EAAE,GAAG;aACX,CAAC,CAAC;YAEH,MAAM,aAAa,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,cAAc,CAAC;YAE/C,0CAA0C;YAC1C,MAAM,iBAAiB,GAAG,aAAa;iBACpC,KAAK,CAAC,eAAe,CAAC;iBACtB,GAAG,EAAE;gBACN,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;gBACrB,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC;YAEjB,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;YAEvC,8CAA8C;YAC9C,MAAM,WAAW,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YACjD,MAAM,YAAY,GAAG,IAAI,CAAC,IAAI,CAAC,iBAAiB,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YAE7D,MAAM,OAAO,GAAmB,CAAC;oBAC/B,IAAI,EAAE,MAAM;oBACZ,IAAI,EAAE,iBAAiB;iBACxB,CAAC,CAAC;YAEH,OAAO;gBACL,EAAE,EAAE,QAAQ,IAAI,CAAC,GAAG,EAAE,EAAE;gBACxB,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO,IAAI,YAAY;gBAC1C,OAAO;gBACP,UAAU,EAAE,UAAU;gBACtB,KAAK,EAAE;oBACL,WAAW;oBACX,YAAY;iBACb;gBACD,QAAQ,EAAE;oBACR,QAAQ,EAAE,MAAM;oBAChB,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;oBAC1B,OAAO;oBACP,IAAI,EAAE,CAAC,EAAE,0BAA0B;oBACnC,kBAAkB,EAAE,IAAI,CAAC,kBAAkB;iBAC5C;aACF,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,oBAAoB;gBAC1B,OAAO,EAAE,0BAA0B,KAAK,EAAE;gBAC1C,QAAQ,EAAE,MAAM;gBAChB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,CAAC,MAAM,CAAC,MAAkB;QAC9B,MAAM,IAAI,CAAC,iBAAiB,EAAE,CAAC;QAE/B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,2DAA2D;YAC3D,4DAA4D;YAC5D,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE;gBAC1C,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;gBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;gBAC1D,SAAS,EAAE,IAAI;gBACf,KAAK,EAAE,GAAG;aACX,CAAC,CAAC;YAEH,MAAM,aAAa,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,cAAc,CAAC;YAC/C,MAAM,iBAAiB,GAAG,aAAa;iBACpC,KAAK,CAAC,eAAe,CAAC;iBACtB,GAAG,EAAE;gBACN,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;gBACrB,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC;YAEjB,8CAA8C;YAC9C,MAAM,KAAK,GAAG,iBAAiB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;YAC3C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBACtC,MAAM,KAAK,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;gBAC3D,MAAM;oBACJ,IAAI,EAAE,qBAAqB;oBAC3B,KAAK,EAAE;wBACL,IAAI,EAAE,YAAY;wBAClB,IAAI,EAAE,KAAK;qBACZ;iBACF,CAAC;gBAEF,yCAAyC;gBACzC,MAAM,IAAI,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC,UAAU,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,CAAC;YACxD,CAAC;YAED,MAAM;gBACJ,IAAI,EAAE,cAAc;aACrB,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,iBAAiB;gBACvB,OAAO,EAAE,0BAA0B,KAAK,EAAE;gBAC1C,QAAQ,EAAE,MAAM;gBAChB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACH,oBAAoB,CAAC,QAAkB;QACrC,MAAM,SAAS,GAAG,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;QACrC,OAAO,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;IACpD,CAAC;IAED;;OAEG;IACH,YAAY;QACV,OAAO;YACL,OAAO,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;YAC5B,kBAAkB,EAAE,IAAI,CAAC,kBAAkB;YAC3C,WAAW,EAAE,IAAI,CAAC,kBAAkB,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,IAAI,CAAC,kBAAkB,CAAC,QAAQ,CAAC,KAAK,CAAC;YAChG,WAAW,EAAE,IAAI,CAAC,SAAS,KAAK,IAAI;SACrC,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,OAAO;QACX,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;YACnB,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC;QACxB,CAAC;QACD,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACjB,MAAM,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC;YAC7B,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC;QACtB,CAAC;IACH,CAAC;CACF","sourcesContent":["/**\n * ONNX Runtime Provider for Local Model Inference\n *\n * Supports CPU and GPU execution providers for optimized local inference\n * Compatible with Phi-3, Llama, and other ONNX models\n */\n\nimport type {\n  LLMProvider,\n  ChatParams,\n  ChatResponse,\n  StreamChunk,\n  ProviderError,\n  Message,\n  ContentBlock\n} from '../types.js';\n\n// Dynamic imports for optional ONNX dependencies\nlet ort: any;\nlet transformers: any;\n\nasync function ensureOnnxDependencies() {\n  if (!ort) {\n    try {\n      const ortModule = await import('onnxruntime-node' as any);\n      ort = ortModule;\n    } catch (e) {\n      throw new Error('onnxruntime-node not installed. Run: npm install onnxruntime-node');\n    }\n  }\n  if (!transformers) {\n    try {\n      const transformersModule = await import('@huggingface/transformers' as any);\n      transformers = transformersModule;\n      transformers.env.allowLocalModels = true;\n    } catch (e) {\n      throw new Error('@huggingface/transformers not installed. Run: npm install @huggingface/transformers');\n    }\n  }\n}\n\nexport interface ONNXConfig {\n  modelPath?: string;\n  modelId?: string; // HuggingFace model ID\n  executionProviders?: string[];\n  sessionOptions?: any;\n  maxTokens?: number;\n  temperature?: number;\n}\n\nexport class ONNXProvider implements LLMProvider {\n  name = 'onnx';\n  type = 'custom' as const;\n  supportsStreaming = true;\n  supportsTools = false;\n  supportsMCP = false;\n\n  private session: any = null;\n  private generator: any = null;\n  private config: ONNXConfig;\n  private executionProviders: string[] = [];\n\n  constructor(config: ONNXConfig = {}) {\n    this.config = {\n      modelId: config.modelId || 'Xenova/Phi-3-mini-4k-instruct',\n      maxTokens: config.maxTokens || 512,\n      temperature: config.temperature || 0.7,\n      ...config\n    };\n  }\n\n  /**\n   * Detect available execution providers\n   */\n  private async detectExecutionProviders(): Promise<string[]> {\n    const providers: string[] = [];\n\n    // Try CUDA for NVIDIA GPUs\n    try {\n      if (process.platform === 'linux') {\n        providers.push('cuda');\n        this.executionProviders.push('cuda');\n      }\n    } catch (e) {\n      // CUDA not available\n    }\n\n    // Try DirectML for Windows GPUs\n    try {\n      if (process.platform === 'win32') {\n        providers.push('dml');\n        this.executionProviders.push('dml');\n      }\n    } catch (e) {\n      // DirectML not available\n    }\n\n    // Always fallback to CPU\n    providers.push('cpu');\n    this.executionProviders.push('cpu');\n\n    console.log(`🔧 ONNX Execution Providers: ${this.executionProviders.join(', ')}`);\n\n    return providers;\n  }\n\n  /**\n   * Initialize ONNX session with model\n   */\n  private async initializeSession(): Promise<void> {\n    if (this.generator) return;\n\n    try {\n      await ensureOnnxDependencies();\n\n      console.log(`📦 Loading ONNX model: ${this.config.modelId}`);\n\n      // Use Transformers.js for easier model loading\n      this.generator = await transformers.pipeline(\n        'text-generation',\n        this.config.modelId,\n        {\n          quantized: true, // Use quantized models for better CPU performance\n        }\n      );\n\n      console.log(`✅ ONNX model loaded successfully`);\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXInitError',\n        message: `Failed to initialize ONNX model: ${error}`,\n        provider: 'onnx',\n        retryable: false\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Format messages for model input\n   */\n  private formatMessages(messages: Message[]): string {\n    // Simple chat template for Phi-3\n    let prompt = '';\n\n    for (const msg of messages) {\n      const content = typeof msg.content === 'string'\n        ? msg.content\n        : msg.content.map(c => c.type === 'text' ? c.text : '').join('');\n\n      if (msg.role === 'user') {\n        prompt += `<|user|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'assistant') {\n        prompt += `<|assistant|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'system') {\n        prompt += `<|system|>\\n${content}<|end|>\\n`;\n      }\n    }\n\n    prompt += '<|assistant|>\\n';\n    return prompt;\n  }\n\n  /**\n   * Chat completion\n   */\n  async chat(params: ChatParams): Promise<ChatResponse> {\n    await this.initializeSession();\n\n    const startTime = Date.now();\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      const result = await this.generator(prompt, {\n        max_new_tokens: params.maxTokens || this.config.maxTokens,\n        temperature: params.temperature || this.config.temperature,\n        do_sample: true,\n        top_p: 0.9,\n      });\n\n      const generatedText = result[0].generated_text;\n\n      // Extract only the new assistant response\n      const assistantResponse = generatedText\n        .split('<|assistant|>')\n        .pop()\n        ?.split('<|end|>')[0]\n        ?.trim() || '';\n\n      const latency = Date.now() - startTime;\n\n      // Estimate token counts (rough approximation)\n      const inputTokens = Math.ceil(prompt.length / 4);\n      const outputTokens = Math.ceil(assistantResponse.length / 4);\n\n      const content: ContentBlock[] = [{\n        type: 'text',\n        text: assistantResponse\n      }];\n\n      return {\n        id: `onnx-${Date.now()}`,\n        model: this.config.modelId || 'onnx-model',\n        content,\n        stopReason: 'end_turn',\n        usage: {\n          inputTokens,\n          outputTokens\n        },\n        metadata: {\n          provider: 'onnx',\n          model: this.config.modelId,\n          latency,\n          cost: 0, // Local inference is free\n          executionProviders: this.executionProviders\n        }\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXInferenceError',\n        message: `ONNX inference failed: ${error}`,\n        provider: 'onnx',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Streaming generation\n   */\n  async *stream(params: ChatParams): AsyncGenerator<StreamChunk> {\n    await this.initializeSession();\n\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      // Note: Transformers.js doesn't natively support streaming\n      // We'll simulate it by yielding tokens as they're generated\n      const result = await this.generator(prompt, {\n        max_new_tokens: params.maxTokens || this.config.maxTokens,\n        temperature: params.temperature || this.config.temperature,\n        do_sample: true,\n        top_p: 0.9,\n      });\n\n      const generatedText = result[0].generated_text;\n      const assistantResponse = generatedText\n        .split('<|assistant|>')\n        .pop()\n        ?.split('<|end|>')[0]\n        ?.trim() || '';\n\n      // Simulate streaming by chunking the response\n      const words = assistantResponse.split(' ');\n      for (let i = 0; i < words.length; i++) {\n        const chunk = words[i] + (i < words.length - 1 ? ' ' : '');\n        yield {\n          type: 'content_block_delta',\n          delta: {\n            type: 'text_delta',\n            text: chunk\n          }\n        };\n\n        // Small delay to simulate real streaming\n        await new Promise(resolve => setTimeout(resolve, 10));\n      }\n\n      yield {\n        type: 'message_stop'\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXStreamError',\n        message: `ONNX streaming failed: ${error}`,\n        provider: 'onnx',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Validate capabilities\n   */\n  validateCapabilities(features: string[]): boolean {\n    const supported = ['chat', 'stream'];\n    return features.every(f => supported.includes(f));\n  }\n\n  /**\n   * Get model info\n   */\n  getModelInfo() {\n    return {\n      modelId: this.config.modelId,\n      executionProviders: this.executionProviders,\n      supportsGPU: this.executionProviders.includes('cuda') || this.executionProviders.includes('dml'),\n      initialized: this.generator !== null\n    };\n  }\n\n  /**\n   * Cleanup resources\n   */\n  async dispose(): Promise<void> {\n    if (this.generator) {\n      this.generator = null;\n    }\n    if (this.session) {\n      await this.session.release();\n      this.session = null;\n    }\n  }\n}\n"]}

package/dist/router/router.d.ts

@@ -1,21 +1,65 @@
 import { LLMProvider, RouterConfig, ChatParams, ChatResponse, StreamChunk, ProviderType, RouterMetrics } from './types.js';
+import { CostOptimalRouter } from './cost-optimal-router.js';
 export declare class ModelRouter {
     private config;
     private providers;
     private metrics;
+    private costOptimalRouter?;
+    private embedQuery?;
+    private embedCache;
+    private static readonly EMBED_CACHE_MAX;
     constructor(configPath?: string);
-    private loadConfig;
     private createConfigFromEnv;
-    private substituteEnvVars;
     private initializeProviders;
     private initializeMetrics;
     chat(params: ChatParams, agentType?: string): Promise<ChatResponse>;
+    /**
+     * Stream a completion. Metrics are recorded once the stream completes. If
+     * the selected provider fails before yielding anything, the fallback chain
+     * is tried; once chunks have been yielded, an error is rethrown because the
+     * caller has already consumed part of the response.
+     */
     stream(params: ChatParams, agentType?: string): AsyncGenerator<StreamChunk>;
+    private streamFallback;
+    /**
+     * Token counts from a stream chunk. Providers report them as camelCase
+     * `usage` (OpenRouter, Ollama) or as Anthropic's snake_case event fields;
+     * counts are cumulative, so the largest seen wins.
+     */
+    private static collectStreamUsage;
+    /**
+     * Pick the provider for a request. Returns the params to send, which are a
+     * copy when routing steers the model; the caller's params are never mutated.
+     */
     private selectProvider;
     private getDefaultProvider;
     private selectByRules;
     private matchesRule;
     private selectByCost;
+    /**
+     * Enable learned cost-optimal routing (ADR-073). Attaches a
+     * {@link CostOptimalRouter} plus an embedder and switches the routing mode to
+     * `'cost-optimal'`. Opt-in: until called, routing behaves exactly as before.
+     */
+    enableCostOptimalRouting(opts: {
+        router: CostOptimalRouter;
+        embed: (text: string) => Promise<number[]> | number[];
+    }): void;
+    /**
+     * Embed `text`, caching the result. Bounded LRU: a cache hit refreshes
+     * recency; the oldest entry is evicted past {@link EMBED_CACHE_MAX}.
+     */
+    private embedCached;
+    /** Extract the most recent user-message text to embed for routing. */
+    private lastUserText;
+    /**
+     * Learned cost-optimal selection (ADR-073): embed the query, route to the
+     * cheapest model predicted to clear the quality bar, and return the matching
+     * provider with a copy of params whose model is the routed one. Falls back to the {@link selectByCost}
+     * heuristic when the learned router/embedder isn't configured or errors —
+     * routing never hard-fails on the cost-optimal path.
+     */
+    private selectByCostOptimal;
     private selectByPerformance;
     private handleProviderError;
     private updateMetrics;

package/dist/router/router.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"router.d.ts","sourceRoot":"","sources":["../../src/router/router.ts"],"names":[],"mappings":"AAIA,OAAO,EACL,WAAW,EACX,YAAY,EACZ,UAAU,EACV,YAAY,EACZ,WAAW,EACX,YAAY,EACZ,aAAa,EAEd,MAAM,YAAY,CAAC;AAMpB,qBAAa,WAAW;IACtB,OAAO,CAAC,MAAM,CAAe;IAC7B,OAAO,CAAC,SAAS,CAA6C;IAC9D,OAAO,CAAC,OAAO,CAAgB;gBAEnB,UAAU,CAAC,EAAE,MAAM;IAM/B,OAAO,CAAC,UAAU;IAwBlB,OAAO,CAAC,mBAAmB;IAyC3B,OAAO,CAAC,iBAAiB;IAwBzB,OAAO,CAAC,mBAAmB;IAwD3B,OAAO,CAAC,iBAAiB;IAUnB,IAAI,CAAC,MAAM,EAAE,UAAU,EAAE,SAAS,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;IAuBlE,MAAM,CAAC,MAAM,EAAE,UAAU,EAAE,SAAS,CAAC,EAAE,MAAM,GAAG,cAAc,CAAC,WAAW,CAAC;YAiBpE,cAAc;IA8B5B,OAAO,CAAC,kBAAkB;IAQ1B,OAAO,CAAC,aAAa;IAgBrB,OAAO,CAAC,WAAW;IAkBnB,OAAO,CAAC,YAAY;IAgBpB,OAAO,CAAC,mBAAmB;YAqBb,mBAAmB;IA4BjC,OAAO,CAAC,aAAa;IA6CrB,UAAU,IAAI,aAAa;IAI3B,SAAS,IAAI,YAAY;IAIzB,YAAY,IAAI,GAAG,CAAC,YAAY,EAAE,WAAW,CAAC;CAG/C"}
+{"version":3,"file":"router.d.ts","sourceRoot":"","sources":["../../src/router/router.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,WAAW,EACX,YAAY,EACZ,UAAU,EACV,YAAY,EACZ,WAAW,EACX,YAAY,EACZ,aAAa,EAGd,MAAM,YAAY,CAAC;AACpB,OAAO,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAS7D,qBAAa,WAAW;IACtB,OAAO,CAAC,MAAM,CAAe;IAC7B,OAAO,CAAC,SAAS,CAA6C;IAC9D,OAAO,CAAC,OAAO,CAAgB;IAE/B,OAAO,CAAC,iBAAiB,CAAC,CAAoB;IAC9C,OAAO,CAAC,UAAU,CAAC,CAAiD;IAGpE,OAAO,CAAC,UAAU,CAA+B;IACjD,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,eAAe,CAAO;gBAElC,UAAU,CAAC,EAAE,MAAM;IAS/B,OAAO,CAAC,mBAAmB;IAiD3B,OAAO,CAAC,mBAAmB;IAkE3B,OAAO,CAAC,iBAAiB;IAUnB,IAAI,CAAC,MAAM,EAAE,UAAU,EAAE,SAAS,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;IAuBzE;;;;;OAKG;IACI,MAAM,CAAC,MAAM,EAAE,UAAU,EAAE,SAAS,CAAC,EAAE,MAAM,GAAG,cAAc,CAAC,WAAW,CAAC;YAyBnE,cAAc;IAqC7B;;;;OAIG;IACH,OAAO,CAAC,MAAM,CAAC,kBAAkB;IAYjC;;;OAGG;YACW,cAAc;IAsC5B,OAAO,CAAC,kBAAkB;IAQ1B,OAAO,CAAC,aAAa;IAgBrB,OAAO,CAAC,WAAW;IAkBnB,OAAO,CAAC,YAAY;IAgBpB;;;;OAIG;IACH,wBAAwB,CAAC,IAAI,EAAE;QAC7B,MAAM,EAAE,iBAAiB,CAAC;QAC1B,KAAK,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,EAAE,CAAC,GAAG,MAAM,EAAE,CAAC;KACvD,GAAG,IAAI;IAMR;;;OAGG;YACW,WAAW;IAgBzB,sEAAsE;IACtE,OAAO,CAAC,YAAY;IAepB;;;;;;OAMG;YACW,mBAAmB;IA2BjC,OAAO,CAAC,mBAAmB;YAqBb,mBAAmB;IA4BjC,OAAO,CAAC,aAAa;IA8CrB,UAAU,IAAI,aAAa;IAI3B,SAAS,IAAI,YAAY;IAIzB,YAAY,IAAI,GAAG,CAAC,YAAY,EAAE,WAAW,CAAC;CAG/C"}

package/dist/router/router.js

@@ -1,95 +1,71 @@
-// Multi-model router core implementation
-import { readFileSync, existsSync } from 'fs';
-import { homedir } from 'os';
-import { join } from 'path';
+import { loadRouterConfig } from './config-loader.js';
 import { OpenRouterProvider } from './providers/openrouter.js';
 import { AnthropicProvider } from './providers/anthropic.js';
 import { ONNXLocalProvider } from './providers/onnx-local.js';
 import { GeminiProvider } from './providers/gemini.js';
+import { OllamaProvider } from './providers/ollama.js';
+import { PHI4_MODEL_PATH } from '../utils/model-downloader.js';
 export class ModelRouter {
     config;
     providers = new Map();
     metrics;
+    // ADR-073: learned cost-optimal routing (opt-in via enableCostOptimalRouting).
+    costOptimalRouter;
+    embedQuery;
+    // Bounded LRU embedding cache — embedding dominates the cost-optimal hot path
+    // (≈ms) vs the µs-scale k-NN, and recurring prompts are common.
+    embedCache = new Map();
+    static EMBED_CACHE_MAX = 512;
     constructor(configPath) {
-        this.config = this.loadConfig(configPath);
+        // Config comes only from an explicit path, AGENTIC_FLOW_ROUTER_CONFIG or
+        // ~/.agentic-flow (config-loader.ts): a repo's router.config*.json could
+        // point baseUrl at an attacker while substituting the real API key.
+        this.config = loadRouterConfig(configPath) ?? this.createConfigFromEnv();
         this.initializeProviders();
         this.metrics = this.initializeMetrics();
     }
-    loadConfig(configPath) {
-        const paths = [
-            configPath,
-            process.env.AGENTIC_FLOW_ROUTER_CONFIG,
-            join(homedir(), '.agentic-flow', 'router.config.json'),
-            join(process.cwd(), 'router.config.json'),
-            join(process.cwd(), 'config', 'router.config.json'),
-            join(process.cwd(), 'router.config.example.json')
-        ].filter(Boolean);
-        for (const path of paths) {
-            if (existsSync(path)) {
-                const content = readFileSync(path, 'utf-8');
-                const config = JSON.parse(content);
-                // Substitute environment variables
-                return this.substituteEnvVars(config);
-            }
-        }
-        // If no config file found, create config from environment variables
-        return this.createConfigFromEnv();
-    }
     createConfigFromEnv() {
         // Create minimal config from environment variables
         const config = {
             version: '1.0',
             defaultProvider: process.env.PROVIDER || 'anthropic',
             routing: { mode: 'manual' },
-            providers: {}
+            providers: {},
         };
         // Add Anthropic if API key exists
         if (process.env.ANTHROPIC_API_KEY) {
             config.providers.anthropic = {
                 apiKey: process.env.ANTHROPIC_API_KEY,
-                baseUrl: process.env.ANTHROPIC_BASE_URL
+                baseUrl: process.env.ANTHROPIC_BASE_URL,
             };
         }
         // Add OpenRouter if API key exists
         if (process.env.OPENROUTER_API_KEY) {
             config.providers.openrouter = {
                 apiKey: process.env.OPENROUTER_API_KEY,
-                baseUrl: process.env.OPENROUTER_BASE_URL
+                baseUrl: process.env.OPENROUTER_BASE_URL,
             };
         }
         // Add Gemini if API key exists
         if (process.env.GOOGLE_GEMINI_API_KEY) {
             config.providers.gemini = {
-                apiKey: process.env.GOOGLE_GEMINI_API_KEY
+                apiKey: process.env.GOOGLE_GEMINI_API_KEY,
+            };
+        }
+        // Add Ollama if API key OR explicit base URL is set (self-hosted needs no key)
+        if (process.env.OLLAMA_API_KEY || process.env.OLLAMA_BASE_URL) {
+            config.providers.ollama = {
+                apiKey: process.env.OLLAMA_API_KEY,
+                baseUrl: process.env.OLLAMA_BASE_URL,
             };
         }
         // ONNX is always available (no API key needed)
         config.providers.onnx = {
             modelPath: process.env.ONNX_MODEL_PATH,
-            executionProviders: ['cpu']
+            executionProviders: ['cpu'],
         };
         return config;
     }
-    substituteEnvVars(obj) {
-        if (typeof obj === 'string') {
-            // Replace ${VAR_NAME} with environment variable value
-            return obj.replace(/\$\{([^}]+)\}/g, (_, key) => {
-                const [varName, defaultValue] = key.split(':-');
-                return process.env[varName] || defaultValue || '';
-            });
-        }
-        if (Array.isArray(obj)) {
-            return obj.map(item => this.substituteEnvVars(item));
-        }
-        if (obj && typeof obj === 'object') {
-            const result = {};
-            for (const [key, value] of Object.entries(obj)) {
-                result[key] = this.substituteEnvVars(value);
-            }
-            return result;
-        }
-        return obj;
-    }
     initializeProviders() {
         const verbose = process.env.ROUTER_VERBOSE === 'true';
         // Initialize Anthropic
@@ -122,10 +98,11 @@ export class ModelRouter {
         if (this.config.providers.onnx) {
             try {
                 const provider = new ONNXLocalProvider({
-                    modelPath: this.config.providers.onnx.modelPath || './models/phi-4/cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/model.onnx',
+                    modelPath: this.config.providers.onnx.modelPath ||
+                        PHI4_MODEL_PATH,
                     executionProviders: this.config.providers.onnx.executionProviders || ['cpu'],
                     maxTokens: this.config.providers.onnx.maxTokens || 100,
-                    temperature: this.config.providers.onnx.temperature || 0.7
+                    temperature: this.config.providers.onnx.temperature || 0.7,
                 });
                 this.providers.set('onnx', provider);
                 if (verbose)
@@ -149,8 +126,19 @@ export class ModelRouter {
                     console.error('❌ Failed to initialize Gemini:', error);
             }
         }
-        // TODO: Initialize other providers (OpenAI, Ollama, LiteLLM)
-        // Will be implemented in Phase 1
+        // Initialize Ollama (OpenAI-compat, self-hosted or ollama.com Cloud)
+        if (this.config.providers.ollama) {
+            try {
+                const provider = new OllamaProvider(this.config.providers.ollama);
+                this.providers.set('ollama', provider);
+                if (verbose)
+                    console.log('✅ Ollama provider initialized');
+            }
+            catch (error) {
+                if (verbose)
+                    console.error('❌ Failed to initialize Ollama:', error);
+            }
+        }
     }
     initializeMetrics() {
         return {
@@ -158,64 +146,128 @@ export class ModelRouter {
             totalCost: 0,
             totalTokens: { input: 0, output: 0 },
             providerBreakdown: {},
-            agentBreakdown: {}
+            agentBreakdown: {},
         };
     }
     async chat(params, agentType) {
         const startTime = Date.now();
-        const provider = await this.selectProvider(params, agentType);
+        const { provider, params: routed } = await this.selectProvider(params, agentType);
         try {
-            const response = await provider.chat(params);
+            const response = await provider.chat(routed);
             // Update metrics
             this.updateMetrics(provider.name, response, Date.now() - startTime, agentType);
             // Add metadata
             response.metadata = {
                 ...response.metadata,
                 provider: provider.name,
-                latency: Date.now() - startTime
+                latency: Date.now() - startTime,
             };
             return response;
         }
         catch (error) {
-            return this.handleProviderError(error, params, agentType);
+            return this.handleProviderError(error, routed, agentType);
         }
     }
+    /**
+     * Stream a completion. Metrics are recorded once the stream completes. If
+     * the selected provider fails before yielding anything, the fallback chain
+     * is tried; once chunks have been yielded, an error is rethrown because the
+     * caller has already consumed part of the response.
+     */
     async *stream(params, agentType) {
-        const provider = await this.selectProvider(params, agentType);
+        const startTime = Date.now();
+        const { provider, params: routed } = await this.selectProvider(params, agentType);
         if (!provider.stream) {
             throw new Error(`Provider ${provider.name} does not support streaming`);
         }
+        const usage = { inputTokens: 0, outputTokens: 0 };
+        let yielded = false;
         try {
-            const iterator = provider.stream(params);
-            for await (const chunk of iterator) {
+            for await (const chunk of provider.stream(routed)) {
+                ModelRouter.collectStreamUsage(chunk, usage);
+                yielded = true;
                 yield chunk;
             }
         }
         catch (error) {
-            throw error;
+            if (yielded)
+                throw error;
+            yield* this.streamFallback(error, provider, routed, agentType, startTime);
+            return;
         }
+        this.updateMetrics(provider.name, { usage }, Date.now() - startTime, agentType);
     }
+    async *streamFallback(error, failed, params, agentType, startTime) {
+        console.error(`❌ Provider error from ${error.provider ?? failed.name}:`, error.message);
+        for (const providerType of this.config.fallbackChain || []) {
+            const provider = this.providers.get(providerType);
+            if (!provider || provider === failed || providerType === error.provider || !provider.stream) {
+                continue;
+            }
+            console.log(`🔄 Falling back to ${provider.name}`);
+            const usage = { inputTokens: 0, outputTokens: 0 };
+            let yielded = false;
+            try {
+                for await (const chunk of provider.stream(params)) {
+                    ModelRouter.collectStreamUsage(chunk, usage);
+                    yielded = true;
+                    yield chunk;
+                }
+            }
+            catch (fallbackError) {
+                if (yielded)
+                    throw fallbackError;
+                console.error(`❌ Fallback provider ${provider.name} also failed`);
+                continue;
+            }
+            this.updateMetrics(provider.name, { usage }, Date.now() - startTime, agentType);
+            return;
+        }
+        throw error; // No fallback succeeded
+    }
+    /**
+     * Token counts from a stream chunk. Providers report them as camelCase
+     * `usage` (OpenRouter, Ollama) or as Anthropic's snake_case event fields;
+     * counts are cumulative, so the largest seen wins.
+     */
+    static collectStreamUsage(chunk, usage) {
+        const reported = chunk?.usage ?? chunk?.message?.usage;
+        if (!reported)
+            return;
+        const input = reported.inputTokens ?? reported.input_tokens;
+        const output = reported.outputTokens ?? reported.output_tokens;
+        if (typeof input === 'number')
+            usage.inputTokens = Math.max(usage.inputTokens, input);
+        if (typeof output === 'number')
+            usage.outputTokens = Math.max(usage.outputTokens, output);
+    }
+    /**
+     * Pick the provider for a request. Returns the params to send, which are a
+     * copy when routing steers the model; the caller's params are never mutated.
+     */
     async selectProvider(params, agentType) {
         // If provider is explicitly specified in params, use it
         if (params.provider) {
             const forcedProvider = this.providers.get(params.provider);
             if (forcedProvider) {
-                return forcedProvider;
+                return { provider: forcedProvider, params };
             }
             console.warn(`⚠️  Requested provider '${params.provider}' not available, falling back to routing logic`);
         }
         const routingMode = this.config.routing?.mode || 'manual';
         switch (routingMode) {
             case 'manual':
-                return this.getDefaultProvider();
+                return { provider: this.getDefaultProvider(), params };
             case 'rule-based':
-                return this.selectByRules(params, agentType);
+                return { provider: this.selectByRules(params, agentType), params };
             case 'cost-optimized':
-                return this.selectByCost(params);
+                return { provider: this.selectByCost(params), params };
+            case 'cost-optimal':
+                return this.selectByCostOptimal(params);
             case 'performance-optimized':
-                return this.selectByPerformance(params);
+                return { provider: this.selectByPerformance(params), params };
             default:
-                return this.getDefaultProvider();
+                return { provider: this.getDefaultProvider(), params };
         }
     }
     getDefaultProvider() {
@@ -265,6 +317,80 @@ export class ModelRouter {
         }
         return this.getDefaultProvider();
     }
+    /**
+     * Enable learned cost-optimal routing (ADR-073). Attaches a
+     * {@link CostOptimalRouter} plus an embedder and switches the routing mode to
+     * `'cost-optimal'`. Opt-in: until called, routing behaves exactly as before.
+     */
+    enableCostOptimalRouting(opts) {
+        this.costOptimalRouter = opts.router;
+        this.embedQuery = opts.embed;
+        this.config.routing = { ...(this.config.routing ?? { mode: 'cost-optimal' }), mode: 'cost-optimal' };
+    }
+    /**
+     * Embed `text`, caching the result. Bounded LRU: a cache hit refreshes
+     * recency; the oldest entry is evicted past {@link EMBED_CACHE_MAX}.
+     */
+    async embedCached(text) {
+        const hit = this.embedCache.get(text);
+        if (hit) {
+            this.embedCache.delete(text); // move to most-recently-used
+            this.embedCache.set(text, hit);
+            return hit;
+        }
+        const embedding = await this.embedQuery(text);
+        this.embedCache.set(text, embedding);
+        if (this.embedCache.size > ModelRouter.EMBED_CACHE_MAX) {
+            const oldest = this.embedCache.keys().next().value;
+            if (oldest !== undefined)
+                this.embedCache.delete(oldest);
+        }
+        return embedding;
+    }
+    /** Extract the most recent user-message text to embed for routing. */
+    lastUserText(messages) {
+        for (let i = messages.length - 1; i >= 0; i--) {
+            const m = messages[i];
+            if (m.role !== 'user')
+                continue;
+            return typeof m.content === 'string'
+                ? m.content
+                : m.content.map((b) => b.text ?? '').join(' ');
+        }
+        // No user turn — fall back to all text content.
+        return messages
+            .map((m) => (typeof m.content === 'string' ? m.content : ''))
+            .join(' ')
+            .trim();
+    }
+    /**
+     * Learned cost-optimal selection (ADR-073): embed the query, route to the
+     * cheapest model predicted to clear the quality bar, and return the matching
+     * provider with a copy of params whose model is the routed one. Falls back to the {@link selectByCost}
+     * heuristic when the learned router/embedder isn't configured or errors —
+     * routing never hard-fails on the cost-optimal path.
+     */
+    async selectByCostOptimal(params) {
+        if (this.costOptimalRouter && this.embedQuery) {
+            try {
+                const text = this.lastUserText(params.messages);
+                const embedding = await this.embedCached(text);
+                const decision = this.costOptimalRouter.route(embedding);
+                const provider = this.providers.get(decision.provider);
+                if (provider) {
+                    console.log(`💡 Cost-optimal routing → ${decision.id} ` +
+                        `(q̂=${decision.predictedQuality.toFixed(2)}, $${decision.costPerMTok}/Mtok, metBar=${decision.metBar})`);
+                    // Steer the model, not just the provider, without mutating the caller's params.
+                    return { provider, params: { ...params, model: decision.model } };
+                }
+                console.warn(`⚠️  Cost-optimal pick '${decision.provider}' not initialized; falling back to heuristic`);
+            }
+            catch (err) {
+                console.warn('⚠️  Cost-optimal routing failed; falling back to heuristic', err);
+            }
+        }
+        return { provider: this.selectByCost(params), params };
+    }
     selectByPerformance(params) {
         // For now, use metrics to select fastest provider
         let fastestProvider = null;
@@ -318,13 +444,14 @@ export class ModelRouter {
                 requests: 0,
                 cost: 0,
                 avgLatency: 0,
-                errors: 0
+                errors: 0,
             };
         }
         const breakdown = this.metrics.providerBreakdown[providerName];
         breakdown.requests++;
         breakdown.cost += response.metadata?.cost || 0;
-        breakdown.avgLatency = (breakdown.avgLatency * (breakdown.requests - 1) + latency) / breakdown.requests;
+        breakdown.avgLatency =
+            (breakdown.avgLatency * (breakdown.requests - 1) + latency) / breakdown.requests;
         // Agent breakdown
         if (agentType) {
             if (!this.metrics.agentBreakdown) {