@almadar/llm 2.5.1 → 2.8.2

package/dist/{chunk-3OVQNNPN.js → chunk-5DRKGB5R.js}

@@ -2,7 +2,7 @@ import {
   RateLimiter,
   getGlobalRateLimiter,
   getGlobalTokenTracker
-} from "./chunk-MJS33AAS.js";
+} from "./chunk-ULT7T7O6.js";
 
 // src/structured-output.ts
 import OpenAI from "openai";
@@ -122,7 +122,7 @@ ${options.existingContext}
       totalTokens: response.usage?.total_tokens || 0
     };
     if (this.tokenTracker) {
-      this.tokenTracker.addUsage(usage.promptTokens, usage.completionTokens);
+      this.tokenTracker.addUsage(usage.promptTokens, usage.completionTokens, { provider: "structured-output" });
     }
     console.log(
       `[StructuredOutputClient] Generated in ${latencyMs}ms, ${usage.totalTokens} tokens`
@@ -171,4 +171,4 @@ export {
   resetStructuredOutputClient,
   isStructuredOutputAvailable
 };
-//# sourceMappingURL=chunk-3OVQNNPN.js.map
+//# sourceMappingURL=chunk-5DRKGB5R.js.map

package/dist/{chunk-3OVQNNPN.js.map → chunk-5DRKGB5R.js.map}

@@ -1 +1 @@
-{"version":3,"sources":["../src/structured-output.ts"],"sourcesContent":["/**\n * Structured Output Client for OpenAI\n *\n * Uses OpenAI's structured outputs feature (json_schema response_format)\n * to guarantee schema compliance at generation time.\n *\n * The system prompt builder is injectable so consumers can provide\n * domain-specific prompts (e.g., orbital schema references).\n *\n * @packageDocumentation\n */\n\nimport OpenAI from 'openai';\nimport type { ChatCompletionCreateParamsNonStreaming } from 'openai/resources/chat/completions';\nimport { z } from 'zod';\nimport {\n  RateLimiter,\n  getGlobalRateLimiter,\n  type RateLimiterOptions,\n} from './rate-limiter.js';\nimport { TokenTracker, getGlobalTokenTracker } from './token-tracker.js';\n\n// ============================================================================\n// Types\n// ============================================================================\n\n/**\n * JSON Schema type used for OpenAI structured outputs.\n */\nexport interface JsonSchema {\n  type?: string | string[];\n  properties?: Record<string, JsonSchema>;\n  required?: string[];\n  items?: JsonSchema;\n  enum?: unknown[];\n  const?: unknown;\n  anyOf?: JsonSchema[];\n  oneOf?: JsonSchema[];\n  allOf?: JsonSchema[];\n  $ref?: string;\n  $defs?: Record<string, JsonSchema>;\n  definitions?: Record<string, JsonSchema>;\n  additionalProperties?: boolean | JsonSchema;\n  description?: string;\n  default?: unknown;\n  minItems?: number;\n  maxItems?: number;\n  minLength?: number;\n}\n\nexport interface StructuredOutputOptions {\n  model?: string;\n  temperature?: number;\n  maxTokens?: number;\n  rateLimiter?: RateLimiterOptions;\n  useGlobalRateLimiter?: boolean;\n  trackTokens?: boolean;\n}\n\nexport interface StructuredGenerationOptions {\n  /** User's natural language request */\n  userRequest: string;\n  /** Model to use (overrides client default) */\n  model?: string;\n  /** Temperature (overrides client default) */\n  temperature?: number;\n  /** Maximum tokens (overrides client default) */\n  maxTokens?: number;\n  /** JSON Schema for structured output */\n  jsonSchema?: JsonSchema;\n  /** Schema name for the json_schema response format */\n  schemaName?: string;\n  /** System prompt override */\n  systemPrompt?: string;\n  /** System prompt builder function (called dynamically) */\n  buildSystemPrompt?: () => string;\n  /** Additional system prompt instructions */\n  additionalInstructions?: string;\n  /** Existing context for updates (e.g., existing schema JSON) */\n  existingContext?: string;\n  /** Skip post-generation validation (default: false) */\n  skipValidation?: boolean;\n}\n\nexport interface StructuredGenerationResult<T = unknown> {\n  /** Generated data (guaranteed to match JSON Schema structure) */\n  data: T;\n  /** Raw JSON string from API */\n  raw: string;\n  /** Token usage statistics */\n  usage: {\n    promptTokens: number;\n    completionTokens: number;\n    totalTokens: number;\n  };\n  /** Generation latency in milliseconds */\n  latencyMs: number;\n  /** Model used for generation */\n  model: string;\n  /** Zod validation result (if not skipped) */\n  zodValidation?: {\n    success: boolean;\n    errors?: z.ZodError['errors'];\n  };\n}\n\nexport const STRUCTURED_OUTPUT_MODELS = {\n  GPT5_MINI: 'gpt-5-mini',\n  GPT4O_MINI: 'gpt-4o-mini',\n  GPT4O: 'gpt-4o',\n  GPT4O_2024_08_06: 'gpt-4o-2024-08-06',\n} as const;\n\n// ============================================================================\n// Default System Prompt\n// ============================================================================\n\nconst DEFAULT_SYSTEM_PROMPT = `You are an expert application architect that generates structured schemas from natural language requirements.\n\nGenerate a complete, well-structured schema based on the user's requirements. Follow the JSON Schema structure exactly.`;\n\n// ============================================================================\n// Structured Output Client\n// ============================================================================\n\nexport class StructuredOutputClient {\n  private openai: OpenAI;\n  private rateLimiter: RateLimiter;\n  private tokenTracker: TokenTracker | null;\n  private defaultModel: string;\n  private defaultTemperature: number;\n  private defaultMaxTokens: number;\n\n  constructor(options: StructuredOutputOptions = {}) {\n    const apiKey = process.env.OPENAI_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'OPENAI_API_KEY environment variable is required for StructuredOutputClient',\n      );\n    }\n\n    this.openai = new OpenAI({ apiKey });\n    this.defaultModel = options.model || STRUCTURED_OUTPUT_MODELS.GPT5_MINI;\n    this.defaultTemperature = options.temperature ?? 0.3;\n    this.defaultMaxTokens = options.maxTokens ?? 16384;\n\n    this.rateLimiter =\n      options.useGlobalRateLimiter !== false\n        ? getGlobalRateLimiter(options.rateLimiter)\n        : new RateLimiter(options.rateLimiter);\n\n    this.tokenTracker =\n      options.trackTokens !== false\n        ? getGlobalTokenTracker(this.defaultModel)\n        : null;\n\n    console.log(\n      `[StructuredOutputClient] Initialized with model: ${this.defaultModel}`,\n    );\n  }\n\n  private usesMaxCompletionTokens(model: string): boolean {\n    const m = model.toLowerCase();\n    return (\n      m.startsWith('o1') ||\n      m.startsWith('gpt-5') ||\n      m.includes('o1-') ||\n      m.includes('o3')\n    );\n  }\n\n  /**\n   * Generate structured output with guaranteed JSON Schema compliance.\n   */\n  async generate<T = unknown>(\n    options: StructuredGenerationOptions,\n  ): Promise<StructuredGenerationResult<T>> {\n    const model = options.model || this.defaultModel;\n    const temperature = options.temperature ?? this.defaultTemperature;\n    const maxTokens = options.maxTokens ?? this.defaultMaxTokens;\n    const startTime = Date.now();\n\n    const jsonSchema: JsonSchema = options.jsonSchema || {\n      type: 'object',\n      properties: {},\n      required: [],\n      additionalProperties: false,\n    };\n\n    // Build system prompt\n    let systemPrompt: string;\n    if (options.systemPrompt) {\n      systemPrompt = options.systemPrompt;\n    } else if (options.buildSystemPrompt) {\n      systemPrompt = options.buildSystemPrompt();\n    } else {\n      systemPrompt = DEFAULT_SYSTEM_PROMPT;\n    }\n\n    if (options.additionalInstructions) {\n      systemPrompt += `\\n\\n## Additional Instructions\\n${options.additionalInstructions}`;\n    }\n\n    // Build user prompt\n    let userPrompt = options.userRequest;\n    if (options.existingContext) {\n      userPrompt += `\\n\\n## Existing Context\\nUpdate based on the above request:\\n\\`\\`\\`json\\n${options.existingContext}\\n\\`\\`\\``;\n    }\n\n    const schemaName = options.schemaName || 'structured_output';\n\n    console.log(\n      `[StructuredOutputClient] Generating with ${model}...`,\n    );\n    console.log(\n      `[StructuredOutputClient] Request: \"${options.userRequest.slice(0, 80)}...\"`,\n    );\n\n    const response = await this.rateLimiter.execute(async () => {\n      const isReasoningModel = this.usesMaxCompletionTokens(model);\n\n      const tokenParam = isReasoningModel\n        ? { max_completion_tokens: maxTokens }\n        : { max_tokens: maxTokens };\n\n      const tempParam = isReasoningModel ? {} : { temperature };\n\n      const params: ChatCompletionCreateParamsNonStreaming = {\n        model,\n        messages: [\n          { role: 'system', content: systemPrompt },\n          { role: 'user', content: userPrompt },\n        ],\n        response_format: {\n          type: 'json_schema',\n          json_schema: {\n            name: schemaName,\n            strict: true,\n            schema: jsonSchema as Record<string, unknown>,\n          },\n        },\n        ...tempParam,\n        ...tokenParam,\n      };\n\n      return this.openai.chat.completions.create(params);\n    });\n\n    const latencyMs = Date.now() - startTime;\n\n    const content = response.choices[0]?.message?.content;\n    if (!content) {\n      throw new Error('No content in OpenAI response');\n    }\n\n    let data: T;\n    try {\n      data = JSON.parse(content) as T;\n    } catch (error) {\n      throw new Error(`Failed to parse response JSON: ${error}`);\n    }\n\n    const usage = {\n      promptTokens: response.usage?.prompt_tokens || 0,\n      completionTokens: response.usage?.completion_tokens || 0,\n      totalTokens: response.usage?.total_tokens || 0,\n    };\n\n    if (this.tokenTracker) {\n      this.tokenTracker.addUsage(usage.promptTokens, usage.completionTokens);\n    }\n\n    console.log(\n      `[StructuredOutputClient] Generated in ${latencyMs}ms, ${usage.totalTokens} tokens`,\n    );\n\n    let zodValidation: StructuredGenerationResult['zodValidation'];\n    if (!options.skipValidation) {\n      zodValidation = { success: true };\n    }\n\n    return {\n      data,\n      raw: content,\n      usage,\n      latencyMs,\n      model,\n      zodValidation,\n    };\n  }\n\n  getModel(): string {\n    return this.defaultModel;\n  }\n\n  getRateLimiterStatus() {\n    return this.rateLimiter.getStatus();\n  }\n\n  getTokenUsage() {\n    return this.tokenTracker?.getSummary() ?? null;\n  }\n}\n\n// ============================================================================\n// Singleton Instance\n// ============================================================================\n\nlet sharedClient: StructuredOutputClient | null = null;\n\n/**\n * Get the singleton structured output client instance.\n *\n * Creates the instance on first call, returns cached instance thereafter.\n *\n * @param {StructuredOutputOptions} [options] - Client configuration options\n * @returns {StructuredOutputClient} The structured output client instance\n */\nexport function getStructuredOutputClient(\n  options?: StructuredOutputOptions,\n): StructuredOutputClient {\n  if (!sharedClient) {\n    sharedClient = new StructuredOutputClient(options);\n  }\n  return sharedClient;\n}\n\nexport function resetStructuredOutputClient(): void {\n  sharedClient = null;\n}\n\n// ============================================================================\n// Convenience Functions\n// ============================================================================\n\nexport function isStructuredOutputAvailable(): boolean {\n  return !!process.env.OPENAI_API_KEY;\n}\n"],"mappings":";;;;;;;AAYA,OAAO,YAAY;AA8FZ,IAAM,2BAA2B;AAAA,EACtC,WAAW;AAAA,EACX,YAAY;AAAA,EACZ,OAAO;AAAA,EACP,kBAAkB;AACpB;AAMA,IAAM,wBAAwB;AAAA;AAAA;AAQvB,IAAM,yBAAN,MAA6B;AAAA,EAQlC,YAAY,UAAmC,CAAC,GAAG;AACjD,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAEA,SAAK,SAAS,IAAI,OAAO,EAAE,OAAO,CAAC;AACnC,SAAK,eAAe,QAAQ,SAAS,yBAAyB;AAC9D,SAAK,qBAAqB,QAAQ,eAAe;AACjD,SAAK,mBAAmB,QAAQ,aAAa;AAE7C,SAAK,cACH,QAAQ,yBAAyB,QAC7B,qBAAqB,QAAQ,WAAW,IACxC,IAAI,YAAY,QAAQ,WAAW;AAEzC,SAAK,eACH,QAAQ,gBAAgB,QACpB,sBAAsB,KAAK,YAAY,IACvC;AAEN,YAAQ;AAAA,MACN,oDAAoD,KAAK,YAAY;AAAA,IACvE;AAAA,EACF;AAAA,EAEQ,wBAAwB,OAAwB;AACtD,UAAM,IAAI,MAAM,YAAY;AAC5B,WACE,EAAE,WAAW,IAAI,KACjB,EAAE,WAAW,OAAO,KACpB,EAAE,SAAS,KAAK,KAChB,EAAE,SAAS,IAAI;AAAA,EAEnB;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,SACJ,SACwC;AACxC,UAAM,QAAQ,QAAQ,SAAS,KAAK;AACpC,UAAM,cAAc,QAAQ,eAAe,KAAK;AAChD,UAAM,YAAY,QAAQ,aAAa,KAAK;AAC5C,UAAM,YAAY,KAAK,IAAI;AAE3B,UAAM,aAAyB,QAAQ,cAAc;AAAA,MACnD,MAAM;AAAA,MACN,YAAY,CAAC;AAAA,MACb,UAAU,CAAC;AAAA,MACX,sBAAsB;AAAA,IACxB;AAGA,QAAI;AACJ,QAAI,QAAQ,cAAc;AACxB,qBAAe,QAAQ;AAAA,IACzB,WAAW,QAAQ,mBAAmB;AACpC,qBAAe,QAAQ,kBAAkB;AAAA,IAC3C,OAAO;AACL,qBAAe;AAAA,IACjB;AAEA,QAAI,QAAQ,wBAAwB;AAClC,sBAAgB;AAAA;AAAA;AAAA,EAAmC,QAAQ,sBAAsB;AAAA,IACnF;AAGA,QAAI,aAAa,QAAQ;AACzB,QAAI,QAAQ,iBAAiB;AAC3B,oBAAc;AAAA;AAAA;AAAA;AAAA;AAAA,EAA4E,QAAQ,eAAe;AAAA;AAAA,IACnH;AAEA,UAAM,aAAa,QAAQ,cAAc;AAEzC,YAAQ;AAAA,MACN,4CAA4C,KAAK;AAAA,IACnD;AACA,YAAQ;AAAA,MACN,sCAAsC,QAAQ,YAAY,MAAM,GAAG,EAAE,CAAC;AAAA,IACxE;AAEA,UAAM,WAAW,MAAM,KAAK,YAAY,QAAQ,YAAY;AAC1D,YAAM,mBAAmB,KAAK,wBAAwB,KAAK;AAE3D,YAAM,aAAa,mBACf,EAAE,uBAAuB,UAAU,IACnC,EAAE,YAAY,UAAU;AAE5B,YAAM,YAAY,mBAAmB,CAAC,IAAI,EAAE,YAAY;AAExD,YAAM,SAAiD;AAAA,QACrD;AAAA,QACA,UAAU;AAAA,UACR,EAAE,MAAM,UAAU,SAAS,aAAa;AAAA,UACxC,EAAE,MAAM,QAAQ,SAAS,WAAW;AAAA,QACtC;AAAA,QACA,iBAAiB;AAAA,UACf,MAAM;AAAA,UACN,aAAa;AAAA,YACX,MAAM;AAAA,YACN,QAAQ;AAAA,YACR,QAAQ;AAAA,UACV;AAAA,QACF;AAAA,QACA,GAAG;AAAA,QACH,GAAG;AAAA,MACL;AAEA,aAAO,KAAK,OAAO,KAAK,YAAY,OAAO,MAAM;AAAA,IACnD,CAAC;AAED,UAAM,YAAY,KAAK,IAAI,IAAI;AAE/B,UAAM,UAAU,SAAS,QAAQ,CAAC,GAAG,SAAS;AAC9C,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI,MAAM,+BAA+B;AAAA,IACjD;AAEA,QAAI;AACJ,QAAI;AACF,aAAO,KAAK,MAAM,OAAO;AAAA,IAC3B,SAAS,OAAO;AACd,YAAM,IAAI,MAAM,kCAAkC,KAAK,EAAE;AAAA,IAC3D;AAEA,UAAM,QAAQ;AAAA,MACZ,cAAc,SAAS,OAAO,iBAAiB;AAAA,MAC/C,kBAAkB,SAAS,OAAO,qBAAqB;AAAA,MACvD,aAAa,SAAS,OAAO,gBAAgB;AAAA,IAC/C;AAEA,QAAI,KAAK,cAAc;AACrB,WAAK,aAAa,SAAS,MAAM,cAAc,MAAM,gBAAgB;AAAA,IACvE;AAEA,YAAQ;AAAA,MACN,yCAAyC,SAAS,OAAO,MAAM,WAAW;AAAA,IAC5E;AAEA,QAAI;AACJ,QAAI,CAAC,QAAQ,gBAAgB;AAC3B,sBAAgB,EAAE,SAAS,KAAK;AAAA,IAClC;AAEA,WAAO;AAAA,MACL;AAAA,MACA,KAAK;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA,EAEA,WAAmB;AACjB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,uBAAuB;AACrB,WAAO,KAAK,YAAY,UAAU;AAAA,EACpC;AAAA,EAEA,gBAAgB;AACd,WAAO,KAAK,cAAc,WAAW,KAAK;AAAA,EAC5C;AACF;AAMA,IAAI,eAA8C;AAU3C,SAAS,0BACd,SACwB;AACxB,MAAI,CAAC,cAAc;AACjB,mBAAe,IAAI,uBAAuB,OAAO;AAAA,EACnD;AACA,SAAO;AACT;AAEO,SAAS,8BAAoC;AAClD,iBAAe;AACjB;AAMO,SAAS,8BAAuC;AACrD,SAAO,CAAC,CAAC,QAAQ,IAAI;AACvB;","names":[]}
+{"version":3,"sources":["../src/structured-output.ts"],"sourcesContent":["/**\n * Structured Output Client for OpenAI\n *\n * Uses OpenAI's structured outputs feature (json_schema response_format)\n * to guarantee schema compliance at generation time.\n *\n * The system prompt builder is injectable so consumers can provide\n * domain-specific prompts (e.g., orbital schema references).\n *\n * @packageDocumentation\n */\n\nimport OpenAI from 'openai';\nimport type { ChatCompletionCreateParamsNonStreaming } from 'openai/resources/chat/completions';\nimport { z } from 'zod';\nimport {\n  RateLimiter,\n  getGlobalRateLimiter,\n  type RateLimiterOptions,\n} from './rate-limiter.js';\nimport { TokenTracker, getGlobalTokenTracker } from './token-tracker.js';\n\n// ============================================================================\n// Types\n// ============================================================================\n\n/**\n * JSON Schema type used for OpenAI structured outputs.\n */\nexport interface JsonSchema {\n  type?: string | string[];\n  properties?: Record<string, JsonSchema>;\n  required?: string[];\n  items?: JsonSchema;\n  enum?: unknown[];\n  const?: unknown;\n  anyOf?: JsonSchema[];\n  oneOf?: JsonSchema[];\n  allOf?: JsonSchema[];\n  $ref?: string;\n  $defs?: Record<string, JsonSchema>;\n  definitions?: Record<string, JsonSchema>;\n  additionalProperties?: boolean | JsonSchema;\n  description?: string;\n  default?: unknown;\n  minItems?: number;\n  maxItems?: number;\n  minLength?: number;\n}\n\nexport interface StructuredOutputOptions {\n  model?: string;\n  temperature?: number;\n  maxTokens?: number;\n  rateLimiter?: RateLimiterOptions;\n  useGlobalRateLimiter?: boolean;\n  trackTokens?: boolean;\n}\n\nexport interface StructuredGenerationOptions {\n  /** User's natural language request */\n  userRequest: string;\n  /** Model to use (overrides client default) */\n  model?: string;\n  /** Temperature (overrides client default) */\n  temperature?: number;\n  /** Maximum tokens (overrides client default) */\n  maxTokens?: number;\n  /** JSON Schema for structured output */\n  jsonSchema?: JsonSchema;\n  /** Schema name for the json_schema response format */\n  schemaName?: string;\n  /** System prompt override */\n  systemPrompt?: string;\n  /** System prompt builder function (called dynamically) */\n  buildSystemPrompt?: () => string;\n  /** Additional system prompt instructions */\n  additionalInstructions?: string;\n  /** Existing context for updates (e.g., existing schema JSON) */\n  existingContext?: string;\n  /** Skip post-generation validation (default: false) */\n  skipValidation?: boolean;\n}\n\nexport interface StructuredGenerationResult<T = unknown> {\n  /** Generated data (guaranteed to match JSON Schema structure) */\n  data: T;\n  /** Raw JSON string from API */\n  raw: string;\n  /** Token usage statistics */\n  usage: {\n    promptTokens: number;\n    completionTokens: number;\n    totalTokens: number;\n  };\n  /** Generation latency in milliseconds */\n  latencyMs: number;\n  /** Model used for generation */\n  model: string;\n  /** Zod validation result (if not skipped) */\n  zodValidation?: {\n    success: boolean;\n    errors?: z.ZodError['errors'];\n  };\n}\n\nexport const STRUCTURED_OUTPUT_MODELS = {\n  GPT5_MINI: 'gpt-5-mini',\n  GPT4O_MINI: 'gpt-4o-mini',\n  GPT4O: 'gpt-4o',\n  GPT4O_2024_08_06: 'gpt-4o-2024-08-06',\n} as const;\n\n// ============================================================================\n// Default System Prompt\n// ============================================================================\n\nconst DEFAULT_SYSTEM_PROMPT = `You are an expert application architect that generates structured schemas from natural language requirements.\n\nGenerate a complete, well-structured schema based on the user's requirements. Follow the JSON Schema structure exactly.`;\n\n// ============================================================================\n// Structured Output Client\n// ============================================================================\n\nexport class StructuredOutputClient {\n  private openai: OpenAI;\n  private rateLimiter: RateLimiter;\n  private tokenTracker: TokenTracker | null;\n  private defaultModel: string;\n  private defaultTemperature: number;\n  private defaultMaxTokens: number;\n\n  constructor(options: StructuredOutputOptions = {}) {\n    const apiKey = process.env.OPENAI_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'OPENAI_API_KEY environment variable is required for StructuredOutputClient',\n      );\n    }\n\n    this.openai = new OpenAI({ apiKey });\n    this.defaultModel = options.model || STRUCTURED_OUTPUT_MODELS.GPT5_MINI;\n    this.defaultTemperature = options.temperature ?? 0.3;\n    this.defaultMaxTokens = options.maxTokens ?? 16384;\n\n    this.rateLimiter =\n      options.useGlobalRateLimiter !== false\n        ? getGlobalRateLimiter(options.rateLimiter)\n        : new RateLimiter(options.rateLimiter);\n\n    this.tokenTracker =\n      options.trackTokens !== false\n        ? getGlobalTokenTracker(this.defaultModel)\n        : null;\n\n    console.log(\n      `[StructuredOutputClient] Initialized with model: ${this.defaultModel}`,\n    );\n  }\n\n  private usesMaxCompletionTokens(model: string): boolean {\n    const m = model.toLowerCase();\n    return (\n      m.startsWith('o1') ||\n      m.startsWith('gpt-5') ||\n      m.includes('o1-') ||\n      m.includes('o3')\n    );\n  }\n\n  /**\n   * Generate structured output with guaranteed JSON Schema compliance.\n   */\n  async generate<T = unknown>(\n    options: StructuredGenerationOptions,\n  ): Promise<StructuredGenerationResult<T>> {\n    const model = options.model || this.defaultModel;\n    const temperature = options.temperature ?? this.defaultTemperature;\n    const maxTokens = options.maxTokens ?? this.defaultMaxTokens;\n    const startTime = Date.now();\n\n    const jsonSchema: JsonSchema = options.jsonSchema || {\n      type: 'object',\n      properties: {},\n      required: [],\n      additionalProperties: false,\n    };\n\n    // Build system prompt\n    let systemPrompt: string;\n    if (options.systemPrompt) {\n      systemPrompt = options.systemPrompt;\n    } else if (options.buildSystemPrompt) {\n      systemPrompt = options.buildSystemPrompt();\n    } else {\n      systemPrompt = DEFAULT_SYSTEM_PROMPT;\n    }\n\n    if (options.additionalInstructions) {\n      systemPrompt += `\\n\\n## Additional Instructions\\n${options.additionalInstructions}`;\n    }\n\n    // Build user prompt\n    let userPrompt = options.userRequest;\n    if (options.existingContext) {\n      userPrompt += `\\n\\n## Existing Context\\nUpdate based on the above request:\\n\\`\\`\\`json\\n${options.existingContext}\\n\\`\\`\\``;\n    }\n\n    const schemaName = options.schemaName || 'structured_output';\n\n    console.log(\n      `[StructuredOutputClient] Generating with ${model}...`,\n    );\n    console.log(\n      `[StructuredOutputClient] Request: \"${options.userRequest.slice(0, 80)}...\"`,\n    );\n\n    const response = await this.rateLimiter.execute(async () => {\n      const isReasoningModel = this.usesMaxCompletionTokens(model);\n\n      const tokenParam = isReasoningModel\n        ? { max_completion_tokens: maxTokens }\n        : { max_tokens: maxTokens };\n\n      const tempParam = isReasoningModel ? {} : { temperature };\n\n      const params: ChatCompletionCreateParamsNonStreaming = {\n        model,\n        messages: [\n          { role: 'system', content: systemPrompt },\n          { role: 'user', content: userPrompt },\n        ],\n        response_format: {\n          type: 'json_schema',\n          json_schema: {\n            name: schemaName,\n            strict: true,\n            schema: jsonSchema as Record<string, unknown>,\n          },\n        },\n        ...tempParam,\n        ...tokenParam,\n      };\n\n      return this.openai.chat.completions.create(params);\n    });\n\n    const latencyMs = Date.now() - startTime;\n\n    const content = response.choices[0]?.message?.content;\n    if (!content) {\n      throw new Error('No content in OpenAI response');\n    }\n\n    let data: T;\n    try {\n      data = JSON.parse(content) as T;\n    } catch (error) {\n      throw new Error(`Failed to parse response JSON: ${error}`);\n    }\n\n    const usage = {\n      promptTokens: response.usage?.prompt_tokens || 0,\n      completionTokens: response.usage?.completion_tokens || 0,\n      totalTokens: response.usage?.total_tokens || 0,\n    };\n\n    if (this.tokenTracker) {\n      this.tokenTracker.addUsage(usage.promptTokens, usage.completionTokens, { provider: 'structured-output' });\n    }\n\n    console.log(\n      `[StructuredOutputClient] Generated in ${latencyMs}ms, ${usage.totalTokens} tokens`,\n    );\n\n    let zodValidation: StructuredGenerationResult['zodValidation'];\n    if (!options.skipValidation) {\n      zodValidation = { success: true };\n    }\n\n    return {\n      data,\n      raw: content,\n      usage,\n      latencyMs,\n      model,\n      zodValidation,\n    };\n  }\n\n  getModel(): string {\n    return this.defaultModel;\n  }\n\n  getRateLimiterStatus() {\n    return this.rateLimiter.getStatus();\n  }\n\n  getTokenUsage() {\n    return this.tokenTracker?.getSummary() ?? null;\n  }\n}\n\n// ============================================================================\n// Singleton Instance\n// ============================================================================\n\nlet sharedClient: StructuredOutputClient | null = null;\n\n/**\n * Get the singleton structured output client instance.\n *\n * Creates the instance on first call, returns cached instance thereafter.\n *\n * @param {StructuredOutputOptions} [options] - Client configuration options\n * @returns {StructuredOutputClient} The structured output client instance\n */\nexport function getStructuredOutputClient(\n  options?: StructuredOutputOptions,\n): StructuredOutputClient {\n  if (!sharedClient) {\n    sharedClient = new StructuredOutputClient(options);\n  }\n  return sharedClient;\n}\n\nexport function resetStructuredOutputClient(): void {\n  sharedClient = null;\n}\n\n// ============================================================================\n// Convenience Functions\n// ============================================================================\n\nexport function isStructuredOutputAvailable(): boolean {\n  return !!process.env.OPENAI_API_KEY;\n}\n"],"mappings":";;;;;;;AAYA,OAAO,YAAY;AA8FZ,IAAM,2BAA2B;AAAA,EACtC,WAAW;AAAA,EACX,YAAY;AAAA,EACZ,OAAO;AAAA,EACP,kBAAkB;AACpB;AAMA,IAAM,wBAAwB;AAAA;AAAA;AAQvB,IAAM,yBAAN,MAA6B;AAAA,EAQlC,YAAY,UAAmC,CAAC,GAAG;AACjD,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAEA,SAAK,SAAS,IAAI,OAAO,EAAE,OAAO,CAAC;AACnC,SAAK,eAAe,QAAQ,SAAS,yBAAyB;AAC9D,SAAK,qBAAqB,QAAQ,eAAe;AACjD,SAAK,mBAAmB,QAAQ,aAAa;AAE7C,SAAK,cACH,QAAQ,yBAAyB,QAC7B,qBAAqB,QAAQ,WAAW,IACxC,IAAI,YAAY,QAAQ,WAAW;AAEzC,SAAK,eACH,QAAQ,gBAAgB,QACpB,sBAAsB,KAAK,YAAY,IACvC;AAEN,YAAQ;AAAA,MACN,oDAAoD,KAAK,YAAY;AAAA,IACvE;AAAA,EACF;AAAA,EAEQ,wBAAwB,OAAwB;AACtD,UAAM,IAAI,MAAM,YAAY;AAC5B,WACE,EAAE,WAAW,IAAI,KACjB,EAAE,WAAW,OAAO,KACpB,EAAE,SAAS,KAAK,KAChB,EAAE,SAAS,IAAI;AAAA,EAEnB;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,SACJ,SACwC;AACxC,UAAM,QAAQ,QAAQ,SAAS,KAAK;AACpC,UAAM,cAAc,QAAQ,eAAe,KAAK;AAChD,UAAM,YAAY,QAAQ,aAAa,KAAK;AAC5C,UAAM,YAAY,KAAK,IAAI;AAE3B,UAAM,aAAyB,QAAQ,cAAc;AAAA,MACnD,MAAM;AAAA,MACN,YAAY,CAAC;AAAA,MACb,UAAU,CAAC;AAAA,MACX,sBAAsB;AAAA,IACxB;AAGA,QAAI;AACJ,QAAI,QAAQ,cAAc;AACxB,qBAAe,QAAQ;AAAA,IACzB,WAAW,QAAQ,mBAAmB;AACpC,qBAAe,QAAQ,kBAAkB;AAAA,IAC3C,OAAO;AACL,qBAAe;AAAA,IACjB;AAEA,QAAI,QAAQ,wBAAwB;AAClC,sBAAgB;AAAA;AAAA;AAAA,EAAmC,QAAQ,sBAAsB;AAAA,IACnF;AAGA,QAAI,aAAa,QAAQ;AACzB,QAAI,QAAQ,iBAAiB;AAC3B,oBAAc;AAAA;AAAA;AAAA;AAAA;AAAA,EAA4E,QAAQ,eAAe;AAAA;AAAA,IACnH;AAEA,UAAM,aAAa,QAAQ,cAAc;AAEzC,YAAQ;AAAA,MACN,4CAA4C,KAAK;AAAA,IACnD;AACA,YAAQ;AAAA,MACN,sCAAsC,QAAQ,YAAY,MAAM,GAAG,EAAE,CAAC;AAAA,IACxE;AAEA,UAAM,WAAW,MAAM,KAAK,YAAY,QAAQ,YAAY;AAC1D,YAAM,mBAAmB,KAAK,wBAAwB,KAAK;AAE3D,YAAM,aAAa,mBACf,EAAE,uBAAuB,UAAU,IACnC,EAAE,YAAY,UAAU;AAE5B,YAAM,YAAY,mBAAmB,CAAC,IAAI,EAAE,YAAY;AAExD,YAAM,SAAiD;AAAA,QACrD;AAAA,QACA,UAAU;AAAA,UACR,EAAE,MAAM,UAAU,SAAS,aAAa;AAAA,UACxC,EAAE,MAAM,QAAQ,SAAS,WAAW;AAAA,QACtC;AAAA,QACA,iBAAiB;AAAA,UACf,MAAM;AAAA,UACN,aAAa;AAAA,YACX,MAAM;AAAA,YACN,QAAQ;AAAA,YACR,QAAQ;AAAA,UACV;AAAA,QACF;AAAA,QACA,GAAG;AAAA,QACH,GAAG;AAAA,MACL;AAEA,aAAO,KAAK,OAAO,KAAK,YAAY,OAAO,MAAM;AAAA,IACnD,CAAC;AAED,UAAM,YAAY,KAAK,IAAI,IAAI;AAE/B,UAAM,UAAU,SAAS,QAAQ,CAAC,GAAG,SAAS;AAC9C,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI,MAAM,+BAA+B;AAAA,IACjD;AAEA,QAAI;AACJ,QAAI;AACF,aAAO,KAAK,MAAM,OAAO;AAAA,IAC3B,SAAS,OAAO;AACd,YAAM,IAAI,MAAM,kCAAkC,KAAK,EAAE;AAAA,IAC3D;AAEA,UAAM,QAAQ;AAAA,MACZ,cAAc,SAAS,OAAO,iBAAiB;AAAA,MAC/C,kBAAkB,SAAS,OAAO,qBAAqB;AAAA,MACvD,aAAa,SAAS,OAAO,gBAAgB;AAAA,IAC/C;AAEA,QAAI,KAAK,cAAc;AACrB,WAAK,aAAa,SAAS,MAAM,cAAc,MAAM,kBAAkB,EAAE,UAAU,oBAAoB,CAAC;AAAA,IAC1G;AAEA,YAAQ;AAAA,MACN,yCAAyC,SAAS,OAAO,MAAM,WAAW;AAAA,IAC5E;AAEA,QAAI;AACJ,QAAI,CAAC,QAAQ,gBAAgB;AAC3B,sBAAgB,EAAE,SAAS,KAAK;AAAA,IAClC;AAEA,WAAO;AAAA,MACL;AAAA,MACA,KAAK;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA,EAEA,WAAmB;AACjB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,uBAAuB;AACrB,WAAO,KAAK,YAAY,UAAU;AAAA,EACpC;AAAA,EAEA,gBAAgB;AACd,WAAO,KAAK,cAAc,WAAW,KAAK;AAAA,EAC5C;AACF;AAMA,IAAI,eAA8C;AAU3C,SAAS,0BACd,SACwB;AACxB,MAAI,CAAC,cAAc;AACjB,mBAAe,IAAI,uBAAuB,OAAO;AAAA,EACnD;AACA,SAAO;AACT;AAEO,SAAS,8BAAoC;AAClD,iBAAe;AACjB;AAMO,SAAS,8BAAuC;AACrD,SAAO,CAAC,CAAC,QAAQ,IAAI;AACvB;","names":[]}

package/dist/{chunk-F2DMHMRH.js → chunk-OBGKLC3H.js}

@@ -5,7 +5,7 @@ import {
   RateLimiter,
   getGlobalRateLimiter,
   getGlobalTokenTracker
-} from "./chunk-MJS33AAS.js";
+} from "./chunk-ULT7T7O6.js";
 
 // src/client.ts
 import { ChatOpenAI } from "@langchain/openai";
@@ -351,7 +351,8 @@ ${prompt}`;
             if (this.tokenTracker) {
               this.tokenTracker.addUsage(
                 usage.promptTokens,
-                usage.completionTokens
+                usage.completionTokens,
+                { provider: this.provider }
               );
             }
           }
@@ -417,7 +418,7 @@ Please output valid JSON that matches the expected schema.`;
     return response.raw;
   }
   async callRawWithMetadata(options) {
-    const { systemPrompt, userPrompt, maxTokens } = options;
+    const { systemPrompt, userPrompt, maxTokens, signal } = options;
     return this.rateLimiter.execute(async () => {
       const modelToUse = maxTokens ? this.getModelWithOptions({ maxTokens }) : this.model;
       const messages = [
@@ -425,7 +426,52 @@ Please output valid JSON that matches the expected schema.`;
         { role: "user", content: this.prepareUserPrompt(userPrompt) }
       ];
       const response = await modelToUse.invoke(
-        this.provider === "anthropic" ? addCacheControlToSystemMessages(messages) : messages
+        this.provider === "anthropic" ? addCacheControlToSystemMessages(messages) : messages,
+        signal ? { signal } : void 0
+      );
+      let usage = null;
+      if (response.usage_metadata) {
+        const usageMeta = response.usage_metadata;
+        usage = {
+          promptTokens: usageMeta.input_tokens || 0,
+          completionTokens: usageMeta.output_tokens || 0,
+          totalTokens: (usageMeta.input_tokens || 0) + (usageMeta.output_tokens || 0)
+        };
+        if (this.tokenTracker) {
+          this.tokenTracker.addUsage(
+            usage.promptTokens,
+            usage.completionTokens
+          );
+        }
+      }
+      const finishReason = this.extractFinishReason(response);
+      const content = typeof response.content === "string" ? response.content : JSON.stringify(response.content);
+      return { raw: content, finishReason, usage };
+    });
+  }
+  /**
+   * Call the LLM with a structured messages array.
+   *
+   * Unlike callRawWithMetadata (which takes systemPrompt + userPrompt strings),
+   * this accepts a full conversation history with proper role separation.
+   * This enables:
+   * - Anthropic prompt caching on message boundaries (not just system prompt)
+   * - Proper tool_use/tool_result role handling across providers
+   * - Reduced token waste from string concatenation
+   *
+   * All providers support the messages format:
+   * - Anthropic: native messages API with cache_control
+   * - DeepSeek: OpenAI-compatible messages via ChatOpenAI
+   * - OpenRouter: OpenAI-compatible messages via ChatOpenAI
+   */
+  async callWithMessages(options) {
+    const { messages, maxTokens, signal } = options;
+    return this.rateLimiter.execute(async () => {
+      const modelToUse = maxTokens ? this.getModelWithOptions({ maxTokens }) : this.model;
+      const langchainMessages = this.provider === "anthropic" ? addCacheControlToSystemMessages(messages) : messages;
+      const response = await modelToUse.invoke(
+        langchainMessages,
+        signal ? { signal } : void 0
       );
       let usage = null;
       if (response.usage_metadata) {
@@ -732,4 +778,4 @@ export {
   createOpenRouterClient,
   createZhipuClient
 };
-//# sourceMappingURL=chunk-F2DMHMRH.js.map
+//# sourceMappingURL=chunk-OBGKLC3H.js.map

package/dist/chunk-OBGKLC3H.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/client.ts"],"sourcesContent":["/**\n * Shared LLM Client\n *\n * Multi-provider LLM client with:\n * - OpenAI, DeepSeek, Anthropic, and Kimi support\n * - Anthropic prompt caching (CachingChatAnthropic)\n * - Rate limiting and retry logic\n * - Token tracking\n * - Structured output parsing with Zod\n *\n * @packageDocumentation\n */\n\nimport { ChatOpenAI } from '@langchain/openai';\nimport { ChatAnthropic } from '@langchain/anthropic';\nimport type { BaseMessageLike } from '@langchain/core/messages';\nimport Anthropic from '@anthropic-ai/sdk';\nimport { z } from 'zod';\nimport {\n  RateLimiter,\n  getGlobalRateLimiter,\n  type RateLimiterOptions,\n} from './rate-limiter.js';\nimport { TokenTracker, getGlobalTokenTracker } from './token-tracker.js';\nimport { parseJsonResponse } from './json-parser.js';\n\n// ============================================================================\n// Anthropic Cache Control Helper\n// ============================================================================\n\nfunction addCacheControlToSystemMessages(\n  messages: Array<{ role: string; content: string }>,\n): BaseMessageLike[] {\n  return messages.map((msg) => {\n    if (msg.role !== 'system') {\n      return msg as BaseMessageLike;\n    }\n\n    return {\n      role: msg.role,\n      content: [\n        {\n          type: 'text' as const,\n          text: msg.content,\n          cache_control: { type: 'ephemeral' },\n        },\n      ],\n    } as BaseMessageLike;\n  });\n}\n\ntype ChatModel = ChatOpenAI | ChatAnthropic;\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport type LLMProvider = 'openai' | 'deepseek' | 'anthropic' | 'kimi' | 'openrouter' | 'orbgen';\n\nexport interface ProviderConfig {\n  apiKey: string;\n  baseUrl?: string;\n  defaultModel: string;\n}\n\nexport interface LLMClientOptions {\n  provider?: LLMProvider;\n  model?: string;\n  temperature?: number;\n  streaming?: boolean;\n  rateLimiter?: RateLimiterOptions;\n  useGlobalRateLimiter?: boolean;\n  trackTokens?: boolean;\n}\n\nexport interface LLMCallOptions<T = unknown> {\n  systemPrompt: string;\n  userPrompt: string;\n  schema?: z.ZodSchema<T>;\n  maxRetries?: number;\n  retryWithContext?: boolean;\n  maxTokens?: number;\n  skipSchemaValidation?: boolean;\n  temperature?: number;\n}\n\nexport interface CacheableBlock {\n  type: 'text';\n  text: string;\n  cache_control?: { type: 'ephemeral' };\n}\n\nexport interface CacheAwareLLMCallOptions<T = unknown>\n  extends LLMCallOptions<T> {\n  systemBlocks?: CacheableBlock[];\n  userBlocks?: CacheableBlock[];\n  rawText?: boolean;\n}\n\nexport interface LLMUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n}\n\nexport type LLMFinishReason =\n  | 'stop'\n  | 'length'\n  | 'content_filter'\n  | 'tool_calls'\n  | null;\n\nexport interface LLMResponse<T> {\n  data: T;\n  raw: string;\n  finishReason: LLMFinishReason;\n  usage: LLMUsage | null;\n}\n\nexport interface LLMStreamOptions {\n  systemPrompt: string;\n  messages: Array<{ role: 'system' | 'user' | 'assistant'; content: string }>;\n  maxTokens?: number;\n  temperature?: number;\n}\n\nexport interface LLMStreamChunk {\n  content: string;\n  done: boolean;\n}\n\n// ============================================================================\n// Provider Configuration\n// ============================================================================\n\nconst PROVIDER_CONFIGS: Record<LLMProvider, () => ProviderConfig> = {\n  openai: () => {\n    const apiKey = process.env.OPENAI_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'OPENAI_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return { apiKey, baseUrl: undefined, defaultModel: 'gpt-4o' };\n  },\n  deepseek: () => {\n    const apiKey = process.env.DEEPSEEK_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'DEEPSEEK_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return {\n      apiKey,\n      baseUrl: 'https://api.deepseek.com/v1',\n      defaultModel: 'deepseek-chat',\n    };\n  },\n  anthropic: () => {\n    const apiKey = process.env.ANTHROPIC_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'ANTHROPIC_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return {\n      apiKey,\n      baseUrl: undefined,\n      defaultModel: 'claude-sonnet-4-5-20250929',\n    };\n  },\n  kimi: () => {\n    const apiKey = process.env.KIMI_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'KIMI_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return {\n      apiKey,\n      baseUrl: 'https://api.moonshot.ai/v1',\n      defaultModel: 'kimi-k2.5',\n    };\n  },\n  openrouter: () => {\n    const apiKey = process.env.OPEN_ROUTER_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'OPEN_ROUTER_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return {\n      apiKey,\n      baseUrl: 'https://openrouter.ai/api/v1',\n      defaultModel: 'qwen/qwen-2.5-72b-instruct', // Default to Qwen 2.5\n    };\n  },\n  orbgen: () => {\n    const baseUrl = process.env.ORBGEN_URL;\n    if (!baseUrl) {\n      throw new Error(\n        'ORBGEN_URL environment variable is not set. ' +\n          'Set it to the OrbGen Cloud Run URL (e.g., https://orbgen-v2-xxx.run.app)',\n      );\n    }\n    return {\n      apiKey: 'not-needed',\n      baseUrl: `${baseUrl}/v1`,\n      defaultModel: 'orbgen-v2',\n    };\n  },\n};\n\nexport const DEEPSEEK_MODELS = {\n  CHAT: 'deepseek-chat',\n  CODER: 'deepseek-coder',\n  REASONER: 'deepseek-reasoner',\n} as const;\n\nexport const OPENAI_MODELS = {\n  GPT4O: 'gpt-4o',\n  GPT4O_MINI: 'gpt-4o-mini',\n  GPT4_TURBO: 'gpt-4-turbo',\n  GPT35_TURBO: 'gpt-3.5-turbo',\n  GPT_5_1: 'gpt-5.1',\n} as const;\n\nexport const ANTHROPIC_MODELS = {\n  CLAUDE_SONNET_4_5: 'claude-sonnet-4-5-20250929',\n  CLAUDE_SONNET_4: 'claude-sonnet-4-20250514',\n  CLAUDE_OPUS_4_5: 'claude-opus-4-5-20250929',\n  CLAUDE_3_5_HAIKU: 'claude-3-5-haiku-20241022',\n} as const;\n\nexport const KIMI_MODELS = {\n  K2_5: 'kimi-k2.5',\n} as const;\n\nexport const OPENROUTER_MODELS = {\n  // Qwen models - JSON/structured data specialists\n  QWEN_2_5_72B: 'qwen/qwen-2.5-72b-instruct',\n  QWEN_2_5_CODER_32B: 'qwen/qwen-2.5-coder-32b-instruct',\n  QWEN_3_235B: 'qwen/qwen3-235b-a22b',\n\n  // Gemma models - best small models for structured JSON output\n  // Gemma 3 4B: 6/6 on complex decomposition, 100% behavior matching, free, fastest\n  GEMMA_3_4B: 'google/gemma-3-4b-it',\n  GEMMA_3_12B: 'google/gemma-3-12b-it',\n  GEMMA_3_27B: 'google/gemma-3-27b-it',\n\n  // Mistral models - strong structured output, function calling\n  MINISTRAL_8B: 'mistralai/ministral-8b-2512',\n  // Mistral Small 3.1: 6/6 on complex decomposition, picked std-kanban for tasks\n  MISTRAL_SMALL_3_1: 'mistralai/mistral-small-3.1-24b-instruct',\n  // Mistral Medium 3.1: next tier up from Small, stronger reasoning, tool calling\n  MISTRAL_MEDIUM_3_1: 'mistralai/mistral-medium-3.1',\n\n  // Llama models - agentic workhorses\n  LLAMA_3_3_70B: 'meta-llama/llama-3.3-70b-instruct',\n  LLAMA_3_1_405B: 'meta-llama/llama-3.1-405b-instruct',\n  LLAMA_4_MAVERICK: 'meta-llama/llama-4-maverick',\n  LLAMA_4_SCOUT: 'meta-llama/llama-4-scout',\n\n  // Kimi models - strong reasoning\n  KIMI_K2: 'moonshotai/kimi-k2',\n\n  // Zhipu GLM models - via OpenRouter\n  GLM_4_7: 'z-ai/glm-4.7',\n} as const;\n\nconst DEFAULT_TEMPERATURE = 0.3;\n\n// ============================================================================\n// LLM Client\n// ============================================================================\n\nexport class LLMClient {\n  private model: ChatModel;\n  private rateLimiter: RateLimiter;\n  private tokenTracker: TokenTracker | null;\n  private modelName: string;\n  private provider: LLMProvider;\n  private providerConfig: ProviderConfig;\n  private temperature: number;\n  private streaming: boolean;\n\n  constructor(options: LLMClientOptions = {}) {\n    this.provider = options.provider || 'openai';\n    // Kimi: 0.6 when thinking disabled (our default), 1.0 when thinking enabled\n    this.temperature = options.temperature ?? \n      (this.provider === 'kimi' ? 0.6 : DEFAULT_TEMPERATURE);\n    this.streaming = options.streaming ?? false;\n\n    this.providerConfig = PROVIDER_CONFIGS[this.provider]();\n    this.modelName = options.model || this.providerConfig.defaultModel;\n\n    const keyPreview = this.providerConfig.apiKey.slice(-4);\n    console.log(\n      `[LLMClient] Provider: ${this.provider}, Model: ${this.modelName}, Key: ****${keyPreview}`,\n    );\n    if (this.providerConfig.baseUrl) {\n      console.log(\n        `[LLMClient] Using custom base URL: ${this.providerConfig.baseUrl}`,\n      );\n    }\n\n    this.model = this.createModel();\n\n    this.rateLimiter =\n      options.useGlobalRateLimiter !== false\n        ? getGlobalRateLimiter(options.rateLimiter)\n        : new RateLimiter(options.rateLimiter);\n\n    this.tokenTracker =\n      options.trackTokens !== false\n        ? getGlobalTokenTracker(this.modelName)\n        : null;\n  }\n\n  private usesMaxCompletionTokens(): boolean {\n    const model = this.modelName.toLowerCase();\n    return (\n      model.startsWith('o1') ||\n      model.startsWith('gpt-5') ||\n      model.includes('o1-') ||\n      model.includes('o3')\n    );\n  }\n\n  private createModel(options?: {\n    maxTokens?: number;\n    temperature?: number;\n  }): ChatModel {\n    const maxTokens = options?.maxTokens;\n    const temperature = options?.temperature ?? this.temperature;\n\n    if (this.provider === 'anthropic') {\n      return new ChatAnthropic({\n        apiKey: this.providerConfig.apiKey,\n        model: this.modelName,\n        temperature,\n        streaming: this.streaming,\n        maxTokens: maxTokens || 8192,\n        callbacks: [\n          {\n            handleLLMEnd: (output) => {\n              const generation = output.generations?.[0]?.[0];\n              const usage = (\n                generation as unknown as {\n                  message?: {\n                    usage_metadata?: {\n                      cache_creation_input_tokens?: number;\n                      cache_read_input_tokens?: number;\n                      input_tokens?: number;\n                      output_tokens?: number;\n                    };\n                  };\n                }\n              )?.message?.usage_metadata;\n\n              if (usage) {\n                const cacheCreated = usage.cache_creation_input_tokens ?? 0;\n                const cacheRead = usage.cache_read_input_tokens ?? 0;\n                const inputTokens = usage.input_tokens ?? 0;\n                const outputTokens = usage.output_tokens ?? 0;\n\n                if (cacheCreated > 0) {\n                  console.log(\n                    `[LLMClient:Anthropic] Cache WRITE: ${cacheCreated} tokens cached`,\n                  );\n                }\n                if (cacheRead > 0) {\n                  const savingsPercent = Math.round(\n                    (cacheRead / (cacheRead + inputTokens)) * 100,\n                  );\n                  console.log(\n                    `[LLMClient:Anthropic] Cache HIT: ${cacheRead} tokens (~${savingsPercent}% of prompt)`,\n                  );\n                }\n                if (cacheCreated === 0 && cacheRead === 0 && inputTokens > 0) {\n                  if (inputTokens < 500) {\n                    console.log(\n                      `[LLMClient:Anthropic] ${inputTokens} input, ${outputTokens} output tokens (likely cached)`,\n                    );\n                  } else {\n                    console.log(\n                      `[LLMClient:Anthropic] ${inputTokens} input, ${outputTokens} output tokens`,\n                    );\n                  }\n                }\n              }\n            },\n          },\n        ],\n      });\n    }\n\n    const useCompletionTokens = this.usesMaxCompletionTokens();\n\n    const tokenConfig = maxTokens\n      ? useCompletionTokens\n        ? { modelKwargs: { max_completion_tokens: maxTokens } }\n        : { maxTokens }\n      : {};\n\n    const timeout = this.provider === 'deepseek' ? 600000 : undefined;\n\n    // Kimi-k2.5: disable thinking to avoid reasoning_content issues with tool calls\n    // When thinking is disabled, temperature must be 0.6 (not 1.0)\n    const isKimi = this.provider === 'kimi';\n    const effectiveTemp = isKimi ? 0.6 : temperature;\n\n    // Build modelKwargs incrementally to avoid spread conflicts\n    const modelKwargs: Record<string, unknown> = {};\n    if (useCompletionTokens && maxTokens) {\n      modelKwargs.max_completion_tokens = maxTokens;\n    }\n    if (isKimi) {\n      modelKwargs.thinking = { type: 'disabled' };\n    }\n    // OpenRouter (Qwen): explicit tool_choice so the model doesn't ignore tool definitions\n    if (this.provider === 'openrouter') {\n      modelKwargs.tool_choice = 'auto';\n    }\n\n    return new ChatOpenAI({\n      apiKey: this.providerConfig.apiKey,\n      model: this.modelName,\n      temperature: useCompletionTokens ? undefined : effectiveTemp,\n      streaming: this.streaming,\n      timeout,\n      ...(Object.keys(modelKwargs).length > 0 ? { modelKwargs } : {}),\n      ...(useCompletionTokens ? {} : maxTokens ? { maxTokens } : {}),\n      configuration: {\n        apiKey: this.providerConfig.apiKey,\n        ...(this.providerConfig.baseUrl\n          ? { baseURL: this.providerConfig.baseUrl }\n          : {}),\n      },\n    });\n  }\n\n  private getModelWithOptions(options: {\n    maxTokens?: number;\n    temperature?: number;\n  }): ChatModel {\n    return this.createModel(options);\n  }\n\n  /**\n   * Check if this model is a Qwen3.5 thinking model.\n   * These models burn all output tokens on internal reasoning\n   * unless thinking is explicitly disabled via /no_think prefix.\n   */\n  private isQwenThinkingModel(): boolean {\n    return this.modelName.includes('qwen3.5');\n  }\n\n  /**\n   * Prepare user prompt with provider-specific adjustments.\n   * Qwen3.5 models require /no_think to disable reasoning mode.\n   */\n  private prepareUserPrompt(prompt: string): string {\n    if (this.isQwenThinkingModel()) {\n      return `/no_think\\n${prompt}`;\n    }\n    return prompt;\n  }\n\n  getProvider(): LLMProvider {\n    return this.provider;\n  }\n\n  getModelName(): string {\n    return this.modelName;\n  }\n\n  getModel(): ChatModel {\n    return this.model;\n  }\n\n  getRateLimiterStatus() {\n    return this.rateLimiter.getStatus();\n  }\n\n  getTokenUsage() {\n    return this.tokenTracker?.getSummary() ?? null;\n  }\n\n  async call<T>(options: LLMCallOptions<T>): Promise<T> {\n    const response = await this.callWithMetadata(options);\n    return response.data;\n  }\n\n  async callWithMetadata<T>(options: LLMCallOptions<T>): Promise<LLMResponse<T>> {\n    const {\n      systemPrompt,\n      userPrompt,\n      schema,\n      maxRetries = 2,\n      retryWithContext = true,\n      maxTokens,\n      skipSchemaValidation = false,\n      temperature,\n    } = options;\n\n    let currentPrompt = userPrompt;\n    let lastError: Error | null = null;\n\n    console.log(\n      `[LLMClient:call] Starting call to ${this.provider}/${this.modelName}`,\n    );\n    console.log(`[LLMClient:call] Prompt length: ${userPrompt.length} chars`);\n    if (maxTokens) {\n      console.log(`[LLMClient:call] Max tokens: ${maxTokens}`);\n    }\n\n    for (let attempt = 0; attempt <= maxRetries; attempt++) {\n      try {\n        console.log(\n          `[LLMClient:call] Attempt ${attempt + 1}/${maxRetries + 1}...`,\n        );\n        const attemptStartTime = Date.now();\n\n        const result = await this.rateLimiter.execute(async () => {\n          console.log(`[LLMClient:call] Invoking model...`);\n          const invokeStartTime = Date.now();\n\n          const modelToUse =\n            maxTokens || temperature !== undefined\n              ? this.getModelWithOptions({ maxTokens, temperature })\n              : this.model;\n\n          const messages = [\n            { role: 'system', content: systemPrompt },\n            { role: 'user', content: this.prepareUserPrompt(currentPrompt) },\n          ];\n          const response = await modelToUse.invoke(\n            this.provider === 'anthropic'\n              ? addCacheControlToSystemMessages(messages)\n              : messages,\n          );\n\n          console.log(\n            `[LLMClient:call] Model responded in ${Date.now() - invokeStartTime}ms`,\n          );\n\n          let usage: LLMUsage | null = null;\n          if (response.usage_metadata) {\n            const usageMeta = response.usage_metadata as {\n              input_tokens?: number;\n              output_tokens?: number;\n            };\n            usage = {\n              promptTokens: usageMeta.input_tokens || 0,\n              completionTokens: usageMeta.output_tokens || 0,\n              totalTokens:\n                (usageMeta.input_tokens || 0) +\n                (usageMeta.output_tokens || 0),\n            };\n            console.log(\n              `[LLMClient:call] Tokens used: ${usage.promptTokens} in, ${usage.completionTokens} out`,\n            );\n\n            if (this.tokenTracker) {\n              this.tokenTracker.addUsage(\n                usage.promptTokens,\n                usage.completionTokens,\n                { provider: this.provider },\n              );\n            }\n          }\n\n          const finishReason = this.extractFinishReason(response);\n          if (finishReason === 'length') {\n            console.warn(\n              `[LLMClient:call] Response truncated (finish_reason=length)`,\n            );\n          }\n\n          const content =\n            typeof response.content === 'string'\n              ? response.content\n              : JSON.stringify(response.content);\n\n          console.log(\n            `[LLMClient:call] Response length: ${content.length} chars, finish_reason: ${finishReason}`,\n          );\n\n          return { content, finishReason, usage };\n        });\n\n        console.log(\n          `[LLMClient:call] Attempt ${attempt + 1} completed in ${Date.now() - attemptStartTime}ms, parsing response...`,\n        );\n\n        const parsed = skipSchemaValidation\n          ? (parseJsonResponse(result.content, undefined) as T)\n          : parseJsonResponse(result.content, schema);\n        console.log(\n          `[LLMClient:call] Response parsed successfully${skipSchemaValidation ? ' (schema validation skipped)' : ''}`,\n        );\n\n        return {\n          data: parsed,\n          raw: result.content,\n          finishReason: result.finishReason,\n          usage: result.usage,\n        };\n      } catch (error) {\n        lastError = error instanceof Error ? error : new Error(String(error));\n        console.error(\n          `[LLMClient:call] Attempt ${attempt + 1} failed:`,\n          lastError.message,\n        );\n\n        if (this.isRateLimitError(lastError)) {\n          console.error(`[LLMClient:call] Rate limit error, not retrying`);\n          throw lastError;\n        }\n\n        if (attempt < maxRetries && retryWithContext) {\n          console.log(`[LLMClient:call] Will retry with error context`);\n          currentPrompt =\n            `${userPrompt}\\n\\n` +\n            `[Previous attempt failed with: ${lastError.message}]\\n` +\n            `Please output valid JSON that matches the expected schema.`;\n        }\n      }\n    }\n\n    console.error(`[LLMClient:call] All attempts exhausted, throwing error`);\n    throw lastError;\n  }\n\n  private extractFinishReason(\n    response: Awaited<ReturnType<ChatOpenAI['invoke']>>,\n  ): LLMFinishReason {\n    const metadata = response.response_metadata as\n      | Record<string, unknown>\n      | undefined;\n    if (metadata?.finish_reason) {\n      const reason = metadata.finish_reason as string;\n      if (\n        reason === 'stop' ||\n        reason === 'length' ||\n        reason === 'content_filter' ||\n        reason === 'tool_calls'\n      ) {\n        return reason;\n      }\n    }\n    return null;\n  }\n\n  async callRaw(options: {\n    systemPrompt: string;\n    userPrompt: string;\n    maxTokens?: number;\n    signal?: AbortSignal;\n  }): Promise<string> {\n    const response = await this.callRawWithMetadata(options);\n    return response.raw;\n  }\n\n  async callRawWithMetadata(options: {\n    systemPrompt: string;\n    userPrompt: string;\n    maxTokens?: number;\n    signal?: AbortSignal;\n  }): Promise<Omit<LLMResponse<string>, 'data'> & { raw: string }> {\n    const { systemPrompt, userPrompt, maxTokens, signal } = options;\n\n    return this.rateLimiter.execute(async () => {\n      const modelToUse = maxTokens\n        ? this.getModelWithOptions({ maxTokens })\n        : this.model;\n\n      const messages = [\n        { role: 'system', content: systemPrompt },\n        { role: 'user', content: this.prepareUserPrompt(userPrompt) },\n      ];\n      const response = await modelToUse.invoke(\n        this.provider === 'anthropic'\n          ? addCacheControlToSystemMessages(messages)\n          : messages,\n        signal ? { signal } : undefined,\n      );\n\n      let usage: LLMUsage | null = null;\n      if (response.usage_metadata) {\n        const usageMeta = response.usage_metadata as {\n          input_tokens?: number;\n          output_tokens?: number;\n        };\n        usage = {\n          promptTokens: usageMeta.input_tokens || 0,\n          completionTokens: usageMeta.output_tokens || 0,\n          totalTokens:\n            (usageMeta.input_tokens || 0) + (usageMeta.output_tokens || 0),\n        };\n\n        if (this.tokenTracker) {\n          this.tokenTracker.addUsage(\n            usage.promptTokens,\n            usage.completionTokens,\n          );\n        }\n      }\n\n      const finishReason = this.extractFinishReason(response);\n      const content =\n        typeof response.content === 'string'\n          ? response.content\n          : JSON.stringify(response.content);\n\n      return { raw: content, finishReason, usage };\n    });\n  }\n\n  /**\n   * Call the LLM with a structured messages array.\n   *\n   * Unlike callRawWithMetadata (which takes systemPrompt + userPrompt strings),\n   * this accepts a full conversation history with proper role separation.\n   * This enables:\n   * - Anthropic prompt caching on message boundaries (not just system prompt)\n   * - Proper tool_use/tool_result role handling across providers\n   * - Reduced token waste from string concatenation\n   *\n   * All providers support the messages format:\n   * - Anthropic: native messages API with cache_control\n   * - DeepSeek: OpenAI-compatible messages via ChatOpenAI\n   * - OpenRouter: OpenAI-compatible messages via ChatOpenAI\n   */\n  async callWithMessages(options: {\n    messages: Array<{ role: string; content: string }>;\n    maxTokens?: number;\n    signal?: AbortSignal;\n  }): Promise<Omit<LLMResponse<string>, 'data'> & { raw: string }> {\n    const { messages, maxTokens, signal } = options;\n\n    return this.rateLimiter.execute(async () => {\n      const modelToUse = maxTokens\n        ? this.getModelWithOptions({ maxTokens })\n        : this.model;\n\n      const langchainMessages = this.provider === 'anthropic'\n        ? addCacheControlToSystemMessages(messages)\n        : (messages as BaseMessageLike[]);\n\n      const response = await modelToUse.invoke(\n        langchainMessages,\n        signal ? { signal } : undefined,\n      );\n\n      let usage: LLMUsage | null = null;\n      if (response.usage_metadata) {\n        const usageMeta = response.usage_metadata as {\n          input_tokens?: number;\n          output_tokens?: number;\n        };\n        usage = {\n          promptTokens: usageMeta.input_tokens || 0,\n          completionTokens: usageMeta.output_tokens || 0,\n          totalTokens:\n            (usageMeta.input_tokens || 0) + (usageMeta.output_tokens || 0),\n        };\n\n        if (this.tokenTracker) {\n          this.tokenTracker.addUsage(\n            usage.promptTokens,\n            usage.completionTokens,\n          );\n        }\n      }\n\n      const finishReason = this.extractFinishReason(response);\n      const content =\n        typeof response.content === 'string'\n          ? response.content\n          : JSON.stringify(response.content);\n\n      return { raw: content, finishReason, usage };\n    });\n  }\n\n  /**\n   * Stream a raw text response as an async iterator of content chunks.\n   * Uses the underlying LangChain model's .stream() method.\n   *\n   * @param options - System prompt plus full message history\n   * @yields LLMStreamChunk with content deltas and a done flag\n   */\n  async *streamRaw(options: LLMStreamOptions): AsyncGenerator<LLMStreamChunk> {\n    const { messages, maxTokens, temperature } = options;\n\n    const modelToUse = (maxTokens || temperature !== undefined)\n      ? this.getModelWithOptions({ maxTokens, temperature })\n      : this.model;\n\n    const langchainMessages = this.provider === 'anthropic'\n      ? addCacheControlToSystemMessages(messages)\n      : messages;\n\n    const stream = await modelToUse.stream(langchainMessages);\n\n    for await (const chunk of stream) {\n      const content = typeof chunk.content === 'string'\n        ? chunk.content\n        : Array.isArray(chunk.content)\n          ? chunk.content\n              .filter((c): c is { type: 'text'; text: string } => typeof c === 'object' && c !== null && 'text' in c)\n              .map((c) => c.text)\n              .join('')\n          : '';\n\n      if (content) {\n        yield { content, done: false };\n      }\n    }\n\n    yield { content: '', done: true };\n  }\n\n  private isRateLimitError(error: Error): boolean {\n    const message = error.message.toLowerCase();\n    return (\n      message.includes('rate limit') ||\n      message.includes('429') ||\n      message.includes('quota exceeded')\n    );\n  }\n\n  // ==========================================================================\n  // Anthropic Cache Control Support\n  // ==========================================================================\n\n  async callWithCache<T>(\n    options: CacheAwareLLMCallOptions<T>,\n  ): Promise<LLMResponse<T>> {\n    const {\n      systemPrompt,\n      userPrompt,\n      systemBlocks,\n      userBlocks,\n      schema,\n      maxRetries = 2,\n      maxTokens,\n      skipSchemaValidation = false,\n      temperature,\n      rawText = false,\n    } = options;\n\n    if (this.provider !== 'anthropic') {\n      console.log(\n        `[LLMClient:callWithCache] Provider ${this.provider} doesn't support caching, using regular call`,\n      );\n      return this.callWithMetadata(options);\n    }\n\n    const cacheableCount =\n      (systemBlocks || []).filter((b) => b.cache_control).length +\n      (userBlocks || []).filter((b) => b.cache_control).length;\n    console.log(\n      `[LLMClient:callWithCache] ${cacheableCount} cacheable block(s)`,\n    );\n\n    let lastError: Error | null = null;\n\n    for (let attempt = 0; attempt <= maxRetries; attempt++) {\n      try {\n        console.log(\n          `[LLMClient:callWithCache] Attempt ${attempt + 1}/${maxRetries + 1}...`,\n        );\n\n        const result = await this.rateLimiter.execute(async () => {\n          const anthropic = new Anthropic();\n\n          const systemContent =\n            systemBlocks && systemBlocks.length > 0\n              ? systemBlocks.map((b) => ({\n                  type: 'text' as const,\n                  text: b.text,\n                  ...(b.cache_control\n                    ? { cache_control: b.cache_control }\n                    : {}),\n                }))\n              : systemPrompt\n                ? [{ type: 'text' as const, text: systemPrompt }]\n                : [];\n\n          const userContent =\n            userBlocks && userBlocks.length > 0\n              ? userBlocks.map((b) => ({\n                  type: 'text' as const,\n                  text: b.text,\n                  ...(b.cache_control\n                    ? { cache_control: b.cache_control }\n                    : {}),\n                }))\n              : userPrompt\n                ? [{ type: 'text' as const, text: userPrompt }]\n                : [];\n\n          const response = await anthropic.messages.create({\n            model: this.modelName,\n            max_tokens: maxTokens || 8192,\n            temperature: temperature ?? 0,\n            system: systemContent,\n            messages: [{ role: 'user', content: userContent }],\n          });\n\n          const textContent = response.content.find((c) => c.type === 'text');\n          const content =\n            textContent && 'text' in textContent ? textContent.text : '';\n\n          const apiUsage = response.usage as {\n            input_tokens: number;\n            output_tokens: number;\n            cache_creation_input_tokens?: number;\n            cache_read_input_tokens?: number;\n          };\n\n          const cacheRead = apiUsage.cache_read_input_tokens || 0;\n          const cacheCreation = apiUsage.cache_creation_input_tokens || 0;\n\n          if (cacheCreation > 0) {\n            console.log(\n              `[LLMClient:callWithCache] Cache WRITE: ${cacheCreation} tokens`,\n            );\n          }\n          if (cacheRead > 0) {\n            const savingsPercent = Math.round(\n              (cacheRead / (cacheRead + apiUsage.input_tokens)) * 100,\n            );\n            console.log(\n              `[LLMClient:callWithCache] Cache HIT: ${cacheRead} tokens (~${savingsPercent}% of prompt)`,\n            );\n          }\n          if (cacheCreation === 0 && cacheRead === 0) {\n            console.log(\n              `[LLMClient:callWithCache] No caching: ${apiUsage.input_tokens} input tokens`,\n            );\n          }\n\n          const usage: LLMUsage = {\n            promptTokens: apiUsage.input_tokens,\n            completionTokens: apiUsage.output_tokens,\n            totalTokens: apiUsage.input_tokens + apiUsage.output_tokens,\n          };\n\n          if (this.tokenTracker) {\n            this.tokenTracker.addUsage(\n              usage.promptTokens,\n              usage.completionTokens,\n            );\n          }\n\n          const finishReason =\n            response.stop_reason === 'end_turn'\n              ? 'stop'\n              : response.stop_reason;\n\n          return {\n            content,\n            finishReason: finishReason as LLMFinishReason,\n            usage,\n          };\n        });\n\n        let parsed: T;\n        if (rawText) {\n          parsed = result.content as unknown as T;\n        } else if (skipSchemaValidation) {\n          parsed = parseJsonResponse(result.content, undefined) as T;\n        } else {\n          parsed = parseJsonResponse(result.content, schema);\n        }\n\n        return {\n          data: parsed,\n          raw: result.content,\n          finishReason: result.finishReason,\n          usage: result.usage,\n        };\n      } catch (error) {\n        lastError = error instanceof Error ? error : new Error(String(error));\n        console.error(\n          `[LLMClient:callWithCache] Attempt ${attempt + 1} failed:`,\n          lastError.message,\n        );\n\n        if (this.isRateLimitError(lastError)) {\n          throw lastError;\n        }\n      }\n    }\n\n    throw lastError;\n  }\n\n  static cacheableBlock(text: string, cache = true): CacheableBlock {\n    return cache\n      ? { type: 'text', text, cache_control: { type: 'ephemeral' } }\n      : { type: 'text', text };\n  }\n}\n\n// ============================================================================\n// Singleton Instances\n// ============================================================================\n\nconst sharedClients: Partial<Record<LLMProvider, LLMClient>> = {};\n\nexport function getSharedLLMClient(options?: LLMClientOptions): LLMClient {\n  const provider = options?.provider || 'openai';\n  if (!sharedClients[provider]) {\n    sharedClients[provider] = new LLMClient(options);\n  }\n  return sharedClients[provider]!;\n}\n\nexport function resetSharedLLMClient(provider?: LLMProvider): void {\n  if (provider) {\n    delete sharedClients[provider];\n  } else {\n    for (const key of Object.keys(sharedClients) as LLMProvider[]) {\n      delete sharedClients[key];\n    }\n  }\n}\n\n// ============================================================================\n// Provider Detection\n// ============================================================================\n\nexport function getAvailableProvider(): LLMProvider {\n  if (process.env.ANTHROPIC_API_KEY) return 'anthropic';\n  if (process.env.DEEPSEEK_API_KEY) return 'deepseek';\n  if (process.env.KIMI_API_KEY) return 'kimi';\n  if (process.env.OPENAI_API_KEY) return 'openai';\n  throw new Error(\n    'No LLM API key found. Please set ANTHROPIC_API_KEY, OPENAI_API_KEY, DEEPSEEK_API_KEY, or KIMI_API_KEY.',\n  );\n}\n\nexport function isProviderAvailable(provider: LLMProvider): boolean {\n  switch (provider) {\n    case 'openai':\n      return !!process.env.OPENAI_API_KEY;\n    case 'deepseek':\n      return !!process.env.DEEPSEEK_API_KEY;\n    case 'anthropic':\n      return !!process.env.ANTHROPIC_API_KEY;\n    case 'kimi':\n      return !!process.env.KIMI_API_KEY;\n    case 'openrouter':\n      return !!process.env.OPEN_ROUTER_API_KEY;\n    case 'orbgen':\n      return !!process.env.ORBGEN_URL;\n    default:\n      return false;\n  }\n}\n\n// ============================================================================\n// Convenience Functions\n// ============================================================================\n\n/**\n * Create an LLM client optimized for requirements analysis.\n *\n * Uses lower temperature (0.3) for more deterministic output.\n * Defaults to GPT-5.1 for OpenAI or DeepSeek Chat.\n *\n * @param {Partial<LLMClientOptions>} [options] - Optional client configuration\n * @returns {LLMClient} Configured LLM client\n */\nexport function createRequirementsClient(\n  options?: Partial<LLMClientOptions>,\n): LLMClient {\n  const provider = options?.provider || getAvailableProvider();\n  const defaultModel =\n    provider === 'deepseek' ? DEEPSEEK_MODELS.CHAT : OPENAI_MODELS.GPT_5_1;\n  return new LLMClient({\n    provider,\n    model: defaultModel,\n    temperature: 0.3,\n    ...options,\n  });\n}\n\n/**\n * Create an LLM client optimized for creative tasks.\n *\n * Uses higher temperature (0.7) for more varied output.\n * Defaults to GPT-4o or DeepSeek Reasoner.\n *\n * @param {Partial<LLMClientOptions>} [options] - Optional client configuration\n * @returns {LLMClient} Configured LLM client\n */\nexport function createCreativeClient(\n  options?: Partial<LLMClientOptions>,\n): LLMClient {\n  const provider = options?.provider || getAvailableProvider();\n  const defaultModel =\n    provider === 'deepseek' ? DEEPSEEK_MODELS.REASONER : OPENAI_MODELS.GPT4O;\n  return new LLMClient({\n    provider,\n    model: defaultModel,\n    temperature: 0.7,\n    ...options,\n  });\n}\n\n/**\n * Create an LLM client optimized for code fixing.\n *\n * Uses low temperature (0.2) for precise, deterministic fixes.\n * Defaults to GPT-4o Mini or DeepSeek Chat for cost efficiency.\n *\n * @param {Partial<LLMClientOptions>} [options] - Optional client configuration\n * @returns {LLMClient} Configured LLM client\n */\nexport function createFixClient(\n  options?: Partial<LLMClientOptions>,\n): LLMClient {\n  const provider = options?.provider || getAvailableProvider();\n  const defaultModel =\n    provider === 'deepseek'\n      ? DEEPSEEK_MODELS.CHAT\n      : OPENAI_MODELS.GPT4O_MINI;\n  return new LLMClient({\n    provider,\n    model: defaultModel,\n    temperature: 0.2,\n    ...options,\n  });\n}\n\n/**\n * Create a DeepSeek LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured DeepSeek client\n */\nexport function createDeepSeekClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'deepseek',\n    model: DEEPSEEK_MODELS.CHAT,\n    ...options,\n  });\n}\n\n/**\n * Create an OpenAI LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured OpenAI client\n */\nexport function createOpenAIClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'openai',\n    model: OPENAI_MODELS.GPT4O,\n    ...options,\n  });\n}\n\n/**\n * Create an Anthropic LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured Anthropic client\n */\nexport function createAnthropicClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'anthropic',\n    model: ANTHROPIC_MODELS.CLAUDE_SONNET_4_5,\n    ...options,\n  });\n}\n\n/**\n * Create a Kimi LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured Kimi client\n */\nexport function createKimiClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'kimi',\n    model: KIMI_MODELS.K2_5,\n    ...options,\n  });\n}\n\n/**\n * Create an OpenRouter LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured OpenRouter client\n */\nexport function createOpenRouterClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'openrouter',\n    model: OPENROUTER_MODELS.QWEN_2_5_72B,\n    ...options,\n  });\n}\n\n/**\n * Create a Zhipu (GLM) LLM client via OpenRouter.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured Zhipu client\n */\nexport function createZhipuClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'openrouter',\n    model: OPENROUTER_MODELS.GLM_4_7,\n    ...options,\n  });\n}\n"],"mappings":";;;;;;;;;;AAaA,SAAS,kBAAkB;AAC3B,SAAS,qBAAqB;AAE9B,OAAO,eAAe;AActB,SAAS,gCACP,UACmB;AACnB,SAAO,SAAS,IAAI,CAAC,QAAQ;AAC3B,QAAI,IAAI,SAAS,UAAU;AACzB,aAAO;AAAA,IACT;AAEA,WAAO;AAAA,MACL,MAAM,IAAI;AAAA,MACV,SAAS;AAAA,QACP;AAAA,UACE,MAAM;AAAA,UACN,MAAM,IAAI;AAAA,UACV,eAAe,EAAE,MAAM,YAAY;AAAA,QACrC;AAAA,MACF;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAsFA,IAAM,mBAA8D;AAAA,EAClE,QAAQ,MAAM;AACZ,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO,EAAE,QAAQ,SAAS,QAAW,cAAc,SAAS;AAAA,EAC9D;AAAA,EACA,UAAU,MAAM;AACd,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL;AAAA,MACA,SAAS;AAAA,MACT,cAAc;AAAA,IAChB;AAAA,EACF;AAAA,EACA,WAAW,MAAM;AACf,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL;AAAA,MACA,SAAS;AAAA,MACT,cAAc;AAAA,IAChB;AAAA,EACF;AAAA,EACA,MAAM,MAAM;AACV,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL;AAAA,MACA,SAAS;AAAA,MACT,cAAc;AAAA,IAChB;AAAA,EACF;AAAA,EACA,YAAY,MAAM;AAChB,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL;AAAA,MACA,SAAS;AAAA,MACT,cAAc;AAAA;AAAA,IAChB;AAAA,EACF;AAAA,EACA,QAAQ,MAAM;AACZ,UAAM,UAAU,QAAQ,IAAI;AAC5B,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,SAAS,GAAG,OAAO;AAAA,MACnB,cAAc;AAAA,IAChB;AAAA,EACF;AACF;AAEO,IAAM,kBAAkB;AAAA,EAC7B,MAAM;AAAA,EACN,OAAO;AAAA,EACP,UAAU;AACZ;AAEO,IAAM,gBAAgB;AAAA,EAC3B,OAAO;AAAA,EACP,YAAY;AAAA,EACZ,YAAY;AAAA,EACZ,aAAa;AAAA,EACb,SAAS;AACX;AAEO,IAAM,mBAAmB;AAAA,EAC9B,mBAAmB;AAAA,EACnB,iBAAiB;AAAA,EACjB,iBAAiB;AAAA,EACjB,kBAAkB;AACpB;AAEO,IAAM,cAAc;AAAA,EACzB,MAAM;AACR;AAEO,IAAM,oBAAoB;AAAA;AAAA,EAE/B,cAAc;AAAA,EACd,oBAAoB;AAAA,EACpB,aAAa;AAAA;AAAA;AAAA,EAIb,YAAY;AAAA,EACZ,aAAa;AAAA,EACb,aAAa;AAAA;AAAA,EAGb,cAAc;AAAA;AAAA,EAEd,mBAAmB;AAAA;AAAA,EAEnB,oBAAoB;AAAA;AAAA,EAGpB,eAAe;AAAA,EACf,gBAAgB;AAAA,EAChB,kBAAkB;AAAA,EAClB,eAAe;AAAA;AAAA,EAGf,SAAS;AAAA;AAAA,EAGT,SAAS;AACX;AAEA,IAAM,sBAAsB;AAMrB,IAAM,YAAN,MAAgB;AAAA,EAUrB,YAAY,UAA4B,CAAC,GAAG;AAC1C,SAAK,WAAW,QAAQ,YAAY;AAEpC,SAAK,cAAc,QAAQ,gBACxB,KAAK,aAAa,SAAS,MAAM;AACpC,SAAK,YAAY,QAAQ,aAAa;AAEtC,SAAK,iBAAiB,iBAAiB,KAAK,QAAQ,EAAE;AACtD,SAAK,YAAY,QAAQ,SAAS,KAAK,eAAe;AAEtD,UAAM,aAAa,KAAK,eAAe,OAAO,MAAM,EAAE;AACtD,YAAQ;AAAA,MACN,yBAAyB,KAAK,QAAQ,YAAY,KAAK,SAAS,cAAc,UAAU;AAAA,IAC1F;AACA,QAAI,KAAK,eAAe,SAAS;AAC/B,cAAQ;AAAA,QACN,sCAAsC,KAAK,eAAe,OAAO;AAAA,MACnE;AAAA,IACF;AAEA,SAAK,QAAQ,KAAK,YAAY;AAE9B,SAAK,cACH,QAAQ,yBAAyB,QAC7B,qBAAqB,QAAQ,WAAW,IACxC,IAAI,YAAY,QAAQ,WAAW;AAEzC,SAAK,eACH,QAAQ,gBAAgB,QACpB,sBAAsB,KAAK,SAAS,IACpC;AAAA,EACR;AAAA,EAEQ,0BAAmC;AACzC,UAAM,QAAQ,KAAK,UAAU,YAAY;AACzC,WACE,MAAM,WAAW,IAAI,KACrB,MAAM,WAAW,OAAO,KACxB,MAAM,SAAS,KAAK,KACpB,MAAM,SAAS,IAAI;AAAA,EAEvB;AAAA,EAEQ,YAAY,SAGN;AACZ,UAAM,YAAY,SAAS;AAC3B,UAAM,cAAc,SAAS,eAAe,KAAK;AAEjD,QAAI,KAAK,aAAa,aAAa;AACjC,aAAO,IAAI,cAAc;AAAA,QACvB,QAAQ,KAAK,eAAe;AAAA,QAC5B,OAAO,KAAK;AAAA,QACZ;AAAA,QACA,WAAW,KAAK;AAAA,QAChB,WAAW,aAAa;AAAA,QACxB,WAAW;AAAA,UACT;AAAA,YACE,cAAc,CAAC,WAAW;AACxB,oBAAM,aAAa,OAAO,cAAc,CAAC,IAAI,CAAC;AAC9C,oBAAM,QACJ,YAUC,SAAS;AAEZ,kBAAI,OAAO;AACT,sBAAM,eAAe,MAAM,+BAA+B;AAC1D,sBAAM,YAAY,MAAM,2BAA2B;AACnD,sBAAM,cAAc,MAAM,gBAAgB;AAC1C,sBAAM,eAAe,MAAM,iBAAiB;AAE5C,oBAAI,eAAe,GAAG;AACpB,0BAAQ;AAAA,oBACN,sCAAsC,YAAY;AAAA,kBACpD;AAAA,gBACF;AACA,oBAAI,YAAY,GAAG;AACjB,wBAAM,iBAAiB,KAAK;AAAA,oBACzB,aAAa,YAAY,eAAgB;AAAA,kBAC5C;AACA,0BAAQ;AAAA,oBACN,oCAAoC,SAAS,aAAa,cAAc;AAAA,kBAC1E;AAAA,gBACF;AACA,oBAAI,iBAAiB,KAAK,cAAc,KAAK,cAAc,GAAG;AAC5D,sBAAI,cAAc,KAAK;AACrB,4BAAQ;AAAA,sBACN,yBAAyB,WAAW,WAAW,YAAY;AAAA,oBAC7D;AAAA,kBACF,OAAO;AACL,4BAAQ;AAAA,sBACN,yBAAyB,WAAW,WAAW,YAAY;AAAA,oBAC7D;AAAA,kBACF;AAAA,gBACF;AAAA,cACF;AAAA,YACF;AAAA,UACF;AAAA,QACF;AAAA,MACF,CAAC;AAAA,IACH;AAEA,UAAM,sBAAsB,KAAK,wBAAwB;AAEzD,UAAM,cAAc,YAChB,sBACE,EAAE,aAAa,EAAE,uBAAuB,UAAU,EAAE,IACpD,EAAE,UAAU,IACd,CAAC;AAEL,UAAM,UAAU,KAAK,aAAa,aAAa,MAAS;AAIxD,UAAM,SAAS,KAAK,aAAa;AACjC,UAAM,gBAAgB,SAAS,MAAM;AAGrC,UAAM,cAAuC,CAAC;AAC9C,QAAI,uBAAuB,WAAW;AACpC,kBAAY,wBAAwB;AAAA,IACtC;AACA,QAAI,QAAQ;AACV,kBAAY,WAAW,EAAE,MAAM,WAAW;AAAA,IAC5C;AAEA,QAAI,KAAK,aAAa,cAAc;AAClC,kBAAY,cAAc;AAAA,IAC5B;AAEA,WAAO,IAAI,WAAW;AAAA,MACpB,QAAQ,KAAK,eAAe;AAAA,MAC5B,OAAO,KAAK;AAAA,MACZ,aAAa,sBAAsB,SAAY;AAAA,MAC/C,WAAW,KAAK;AAAA,MAChB;AAAA,MACA,GAAI,OAAO,KAAK,WAAW,EAAE,SAAS,IAAI,EAAE,YAAY,IAAI,CAAC;AAAA,MAC7D,GAAI,sBAAsB,CAAC,IAAI,YAAY,EAAE,UAAU,IAAI,CAAC;AAAA,MAC5D,eAAe;AAAA,QACb,QAAQ,KAAK,eAAe;AAAA,QAC5B,GAAI,KAAK,eAAe,UACpB,EAAE,SAAS,KAAK,eAAe,QAAQ,IACvC,CAAC;AAAA,MACP;AAAA,IACF,CAAC;AAAA,EACH;AAAA,EAEQ,oBAAoB,SAGd;AACZ,WAAO,KAAK,YAAY,OAAO;AAAA,EACjC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOQ,sBAA+B;AACrC,WAAO,KAAK,UAAU,SAAS,SAAS;AAAA,EAC1C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMQ,kBAAkB,QAAwB;AAChD,QAAI,KAAK,oBAAoB,GAAG;AAC9B,aAAO;AAAA,EAAc,MAAM;AAAA,IAC7B;AACA,WAAO;AAAA,EACT;AAAA,EAEA,cAA2B;AACzB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,eAAuB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,WAAsB;AACpB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,uBAAuB;AACrB,WAAO,KAAK,YAAY,UAAU;AAAA,EACpC;AAAA,EAEA,gBAAgB;AACd,WAAO,KAAK,cAAc,WAAW,KAAK;AAAA,EAC5C;AAAA,EAEA,MAAM,KAAQ,SAAwC;AACpD,UAAM,WAAW,MAAM,KAAK,iBAAiB,OAAO;AACpD,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,iBAAoB,SAAqD;AAC7E,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA,aAAa;AAAA,MACb,mBAAmB;AAAA,MACnB;AAAA,MACA,uBAAuB;AAAA,MACvB;AAAA,IACF,IAAI;AAEJ,QAAI,gBAAgB;AACpB,QAAI,YAA0B;AAE9B,YAAQ;AAAA,MACN,qCAAqC,KAAK,QAAQ,IAAI,KAAK,SAAS;AAAA,IACtE;AACA,YAAQ,IAAI,mCAAmC,WAAW,MAAM,QAAQ;AACxE,QAAI,WAAW;AACb,cAAQ,IAAI,gCAAgC,SAAS,EAAE;AAAA,IACzD;AAEA,aAAS,UAAU,GAAG,WAAW,YAAY,WAAW;AACtD,UAAI;AACF,gBAAQ;AAAA,UACN,4BAA4B,UAAU,CAAC,IAAI,aAAa,CAAC;AAAA,QAC3D;AACA,cAAM,mBAAmB,KAAK,IAAI;AAElC,cAAM,SAAS,MAAM,KAAK,YAAY,QAAQ,YAAY;AACxD,kBAAQ,IAAI,oCAAoC;AAChD,gBAAM,kBAAkB,KAAK,IAAI;AAEjC,gBAAM,aACJ,aAAa,gBAAgB,SACzB,KAAK,oBAAoB,EAAE,WAAW,YAAY,CAAC,IACnD,KAAK;AAEX,gBAAM,WAAW;AAAA,YACf,EAAE,MAAM,UAAU,SAAS,aAAa;AAAA,YACxC,EAAE,MAAM,QAAQ,SAAS,KAAK,kBAAkB,aAAa,EAAE;AAAA,UACjE;AACA,gBAAM,WAAW,MAAM,WAAW;AAAA,YAChC,KAAK,aAAa,cACd,gCAAgC,QAAQ,IACxC;AAAA,UACN;AAEA,kBAAQ;AAAA,YACN,uCAAuC,KAAK,IAAI,IAAI,eAAe;AAAA,UACrE;AAEA,cAAI,QAAyB;AAC7B,cAAI,SAAS,gBAAgB;AAC3B,kBAAM,YAAY,SAAS;AAI3B,oBAAQ;AAAA,cACN,cAAc,UAAU,gBAAgB;AAAA,cACxC,kBAAkB,UAAU,iBAAiB;AAAA,cAC7C,cACG,UAAU,gBAAgB,MAC1B,UAAU,iBAAiB;AAAA,YAChC;AACA,oBAAQ;AAAA,cACN,iCAAiC,MAAM,YAAY,QAAQ,MAAM,gBAAgB;AAAA,YACnF;AAEA,gBAAI,KAAK,cAAc;AACrB,mBAAK,aAAa;AAAA,gBAChB,MAAM;AAAA,gBACN,MAAM;AAAA,gBACN,EAAE,UAAU,KAAK,SAAS;AAAA,cAC5B;AAAA,YACF;AAAA,UACF;AAEA,gBAAM,eAAe,KAAK,oBAAoB,QAAQ;AACtD,cAAI,iBAAiB,UAAU;AAC7B,oBAAQ;AAAA,cACN;AAAA,YACF;AAAA,UACF;AAEA,gBAAM,UACJ,OAAO,SAAS,YAAY,WACxB,SAAS,UACT,KAAK,UAAU,SAAS,OAAO;AAErC,kBAAQ;AAAA,YACN,qCAAqC,QAAQ,MAAM,0BAA0B,YAAY;AAAA,UAC3F;AAEA,iBAAO,EAAE,SAAS,cAAc,MAAM;AAAA,QACxC,CAAC;AAED,gBAAQ;AAAA,UACN,4BAA4B,UAAU,CAAC,iBAAiB,KAAK,IAAI,IAAI,gBAAgB;AAAA,QACvF;AAEA,cAAM,SAAS,uBACV,kBAAkB,OAAO,SAAS,MAAS,IAC5C,kBAAkB,OAAO,SAAS,MAAM;AAC5C,gBAAQ;AAAA,UACN,gDAAgD,uBAAuB,iCAAiC,EAAE;AAAA,QAC5G;AAEA,eAAO;AAAA,UACL,MAAM;AAAA,UACN,KAAK,OAAO;AAAA,UACZ,cAAc,OAAO;AAAA,UACrB,OAAO,OAAO;AAAA,QAChB;AAAA,MACF,SAAS,OAAO;AACd,oBAAY,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;AACpE,gBAAQ;AAAA,UACN,4BAA4B,UAAU,CAAC;AAAA,UACvC,UAAU;AAAA,QACZ;AAEA,YAAI,KAAK,iBAAiB,SAAS,GAAG;AACpC,kBAAQ,MAAM,iDAAiD;AAC/D,gBAAM;AAAA,QACR;AAEA,YAAI,UAAU,cAAc,kBAAkB;AAC5C,kBAAQ,IAAI,gDAAgD;AAC5D,0BACE,GAAG,UAAU;AAAA;AAAA,iCACqB,UAAU,OAAO;AAAA;AAAA,QAEvD;AAAA,MACF;AAAA,IACF;AAEA,YAAQ,MAAM,yDAAyD;AACvE,UAAM;AAAA,EACR;AAAA,EAEQ,oBACN,UACiB;AACjB,UAAM,WAAW,SAAS;AAG1B,QAAI,UAAU,eAAe;AAC3B,YAAM,SAAS,SAAS;AACxB,UACE,WAAW,UACX,WAAW,YACX,WAAW,oBACX,WAAW,cACX;AACA,eAAO;AAAA,MACT;AAAA,IACF;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,QAAQ,SAKM;AAClB,UAAM,WAAW,MAAM,KAAK,oBAAoB,OAAO;AACvD,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,oBAAoB,SAKuC;AAC/D,UAAM,EAAE,cAAc,YAAY,WAAW,OAAO,IAAI;AAExD,WAAO,KAAK,YAAY,QAAQ,YAAY;AAC1C,YAAM,aAAa,YACf,KAAK,oBAAoB,EAAE,UAAU,CAAC,IACtC,KAAK;AAET,YAAM,WAAW;AAAA,QACf,EAAE,MAAM,UAAU,SAAS,aAAa;AAAA,QACxC,EAAE,MAAM,QAAQ,SAAS,KAAK,kBAAkB,UAAU,EAAE;AAAA,MAC9D;AACA,YAAM,WAAW,MAAM,WAAW;AAAA,QAChC,KAAK,aAAa,cACd,gCAAgC,QAAQ,IACxC;AAAA,QACJ,SAAS,EAAE,OAAO,IAAI;AAAA,MACxB;AAEA,UAAI,QAAyB;AAC7B,UAAI,SAAS,gBAAgB;AAC3B,cAAM,YAAY,SAAS;AAI3B,gBAAQ;AAAA,UACN,cAAc,UAAU,gBAAgB;AAAA,UACxC,kBAAkB,UAAU,iBAAiB;AAAA,UAC7C,cACG,UAAU,gBAAgB,MAAM,UAAU,iBAAiB;AAAA,QAChE;AAEA,YAAI,KAAK,cAAc;AACrB,eAAK,aAAa;AAAA,YAChB,MAAM;AAAA,YACN,MAAM;AAAA,UACR;AAAA,QACF;AAAA,MACF;AAEA,YAAM,eAAe,KAAK,oBAAoB,QAAQ;AACtD,YAAM,UACJ,OAAO,SAAS,YAAY,WACxB,SAAS,UACT,KAAK,UAAU,SAAS,OAAO;AAErC,aAAO,EAAE,KAAK,SAAS,cAAc,MAAM;AAAA,IAC7C,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAiBA,MAAM,iBAAiB,SAI0C;AAC/D,UAAM,EAAE,UAAU,WAAW,OAAO,IAAI;AAExC,WAAO,KAAK,YAAY,QAAQ,YAAY;AAC1C,YAAM,aAAa,YACf,KAAK,oBAAoB,EAAE,UAAU,CAAC,IACtC,KAAK;AAET,YAAM,oBAAoB,KAAK,aAAa,cACxC,gCAAgC,QAAQ,IACvC;AAEL,YAAM,WAAW,MAAM,WAAW;AAAA,QAChC;AAAA,QACA,SAAS,EAAE,OAAO,IAAI;AAAA,MACxB;AAEA,UAAI,QAAyB;AAC7B,UAAI,SAAS,gBAAgB;AAC3B,cAAM,YAAY,SAAS;AAI3B,gBAAQ;AAAA,UACN,cAAc,UAAU,gBAAgB;AAAA,UACxC,kBAAkB,UAAU,iBAAiB;AAAA,UAC7C,cACG,UAAU,gBAAgB,MAAM,UAAU,iBAAiB;AAAA,QAChE;AAEA,YAAI,KAAK,cAAc;AACrB,eAAK,aAAa;AAAA,YAChB,MAAM;AAAA,YACN,MAAM;AAAA,UACR;AAAA,QACF;AAAA,MACF;AAEA,YAAM,eAAe,KAAK,oBAAoB,QAAQ;AACtD,YAAM,UACJ,OAAO,SAAS,YAAY,WACxB,SAAS,UACT,KAAK,UAAU,SAAS,OAAO;AAErC,aAAO,EAAE,KAAK,SAAS,cAAc,MAAM;AAAA,IAC7C,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,OAAO,UAAU,SAA2D;AAC1E,UAAM,EAAE,UAAU,WAAW,YAAY,IAAI;AAE7C,UAAM,aAAc,aAAa,gBAAgB,SAC7C,KAAK,oBAAoB,EAAE,WAAW,YAAY,CAAC,IACnD,KAAK;AAET,UAAM,oBAAoB,KAAK,aAAa,cACxC,gCAAgC,QAAQ,IACxC;AAEJ,UAAM,SAAS,MAAM,WAAW,OAAO,iBAAiB;AAExD,qBAAiB,SAAS,QAAQ;AAChC,YAAM,UAAU,OAAO,MAAM,YAAY,WACrC,MAAM,UACN,MAAM,QAAQ,MAAM,OAAO,IACzB,MAAM,QACH,OAAO,CAAC,MAA2C,OAAO,MAAM,YAAY,MAAM,QAAQ,UAAU,CAAC,EACrG,IAAI,CAAC,MAAM,EAAE,IAAI,EACjB,KAAK,EAAE,IACV;AAEN,UAAI,SAAS;AACX,cAAM,EAAE,SAAS,MAAM,MAAM;AAAA,MAC/B;AAAA,IACF;AAEA,UAAM,EAAE,SAAS,IAAI,MAAM,KAAK;AAAA,EAClC;AAAA,EAEQ,iBAAiB,OAAuB;AAC9C,UAAM,UAAU,MAAM,QAAQ,YAAY;AAC1C,WACE,QAAQ,SAAS,YAAY,KAC7B,QAAQ,SAAS,KAAK,KACtB,QAAQ,SAAS,gBAAgB;AAAA,EAErC;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,cACJ,SACyB;AACzB,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,aAAa;AAAA,MACb;AAAA,MACA,uBAAuB;AAAA,MACvB;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,QAAI,KAAK,aAAa,aAAa;AACjC,cAAQ;AAAA,QACN,sCAAsC,KAAK,QAAQ;AAAA,MACrD;AACA,aAAO,KAAK,iBAAiB,OAAO;AAAA,IACtC;AAEA,UAAM,kBACH,gBAAgB,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,EAAE,UACnD,cAAc,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,EAAE;AACpD,YAAQ;AAAA,MACN,6BAA6B,cAAc;AAAA,IAC7C;AAEA,QAAI,YAA0B;AAE9B,aAAS,UAAU,GAAG,WAAW,YAAY,WAAW;AACtD,UAAI;AACF,gBAAQ;AAAA,UACN,qCAAqC,UAAU,CAAC,IAAI,aAAa,CAAC;AAAA,QACpE;AAEA,cAAM,SAAS,MAAM,KAAK,YAAY,QAAQ,YAAY;AACxD,gBAAM,YAAY,IAAI,UAAU;AAEhC,gBAAM,gBACJ,gBAAgB,aAAa,SAAS,IAClC,aAAa,IAAI,CAAC,OAAO;AAAA,YACvB,MAAM;AAAA,YACN,MAAM,EAAE;AAAA,YACR,GAAI,EAAE,gBACF,EAAE,eAAe,EAAE,cAAc,IACjC,CAAC;AAAA,UACP,EAAE,IACF,eACE,CAAC,EAAE,MAAM,QAAiB,MAAM,aAAa,CAAC,IAC9C,CAAC;AAET,gBAAM,cACJ,cAAc,WAAW,SAAS,IAC9B,WAAW,IAAI,CAAC,OAAO;AAAA,YACrB,MAAM;AAAA,YACN,MAAM,EAAE;AAAA,YACR,GAAI,EAAE,gBACF,EAAE,eAAe,EAAE,cAAc,IACjC,CAAC;AAAA,UACP,EAAE,IACF,aACE,CAAC,EAAE,MAAM,QAAiB,MAAM,WAAW,CAAC,IAC5C,CAAC;AAET,gBAAM,WAAW,MAAM,UAAU,SAAS,OAAO;AAAA,YAC/C,OAAO,KAAK;AAAA,YACZ,YAAY,aAAa;AAAA,YACzB,aAAa,eAAe;AAAA,YAC5B,QAAQ;AAAA,YACR,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,YAAY,CAAC;AAAA,UACnD,CAAC;AAED,gBAAM,cAAc,SAAS,QAAQ,KAAK,CAAC,MAAM,EAAE,SAAS,MAAM;AAClE,gBAAM,UACJ,eAAe,UAAU,cAAc,YAAY,OAAO;AAE5D,gBAAM,WAAW,SAAS;AAO1B,gBAAM,YAAY,SAAS,2BAA2B;AACtD,gBAAM,gBAAgB,SAAS,+BAA+B;AAE9D,cAAI,gBAAgB,GAAG;AACrB,oBAAQ;AAAA,cACN,0CAA0C,aAAa;AAAA,YACzD;AAAA,UACF;AACA,cAAI,YAAY,GAAG;AACjB,kBAAM,iBAAiB,KAAK;AAAA,cACzB,aAAa,YAAY,SAAS,gBAAiB;AAAA,YACtD;AACA,oBAAQ;AAAA,cACN,wCAAwC,SAAS,aAAa,cAAc;AAAA,YAC9E;AAAA,UACF;AACA,cAAI,kBAAkB,KAAK,cAAc,GAAG;AAC1C,oBAAQ;AAAA,cACN,yCAAyC,SAAS,YAAY;AAAA,YAChE;AAAA,UACF;AAEA,gBAAM,QAAkB;AAAA,YACtB,cAAc,SAAS;AAAA,YACvB,kBAAkB,SAAS;AAAA,YAC3B,aAAa,SAAS,eAAe,SAAS;AAAA,UAChD;AAEA,cAAI,KAAK,cAAc;AACrB,iBAAK,aAAa;AAAA,cAChB,MAAM;AAAA,cACN,MAAM;AAAA,YACR;AAAA,UACF;AAEA,gBAAM,eACJ,SAAS,gBAAgB,aACrB,SACA,SAAS;AAEf,iBAAO;AAAA,YACL;AAAA,YACA;AAAA,YACA;AAAA,UACF;AAAA,QACF,CAAC;AAED,YAAI;AACJ,YAAI,SAAS;AACX,mBAAS,OAAO;AAAA,QAClB,WAAW,sBAAsB;AAC/B,mBAAS,kBAAkB,OAAO,SAAS,MAAS;AAAA,QACtD,OAAO;AACL,mBAAS,kBAAkB,OAAO,SAAS,MAAM;AAAA,QACnD;AAEA,eAAO;AAAA,UACL,MAAM;AAAA,UACN,KAAK,OAAO;AAAA,UACZ,cAAc,OAAO;AAAA,UACrB,OAAO,OAAO;AAAA,QAChB;AAAA,MACF,SAAS,OAAO;AACd,oBAAY,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;AACpE,gBAAQ;AAAA,UACN,qCAAqC,UAAU,CAAC;AAAA,UAChD,UAAU;AAAA,QACZ;AAEA,YAAI,KAAK,iBAAiB,SAAS,GAAG;AACpC,gBAAM;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAEA,UAAM;AAAA,EACR;AAAA,EAEA,OAAO,eAAe,MAAc,QAAQ,MAAsB;AAChE,WAAO,QACH,EAAE,MAAM,QAAQ,MAAM,eAAe,EAAE,MAAM,YAAY,EAAE,IAC3D,EAAE,MAAM,QAAQ,KAAK;AAAA,EAC3B;AACF;AAMA,IAAM,gBAAyD,CAAC;AAEzD,SAAS,mBAAmB,SAAuC;AACxE,QAAM,WAAW,SAAS,YAAY;AACtC,MAAI,CAAC,cAAc,QAAQ,GAAG;AAC5B,kBAAc,QAAQ,IAAI,IAAI,UAAU,OAAO;AAAA,EACjD;AACA,SAAO,cAAc,QAAQ;AAC/B;AAEO,SAAS,qBAAqB,UAA8B;AACjE,MAAI,UAAU;AACZ,WAAO,cAAc,QAAQ;AAAA,EAC/B,OAAO;AACL,eAAW,OAAO,OAAO,KAAK,aAAa,GAAoB;AAC7D,aAAO,cAAc,GAAG;AAAA,IAC1B;AAAA,EACF;AACF;AAMO,SAAS,uBAAoC;AAClD,MAAI,QAAQ,IAAI,kBAAmB,QAAO;AAC1C,MAAI,QAAQ,IAAI,iBAAkB,QAAO;AACzC,MAAI,QAAQ,IAAI,aAAc,QAAO;AACrC,MAAI,QAAQ,IAAI,eAAgB,QAAO;AACvC,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEO,SAAS,oBAAoB,UAAgC;AAClE,UAAQ,UAAU;AAAA,IAChB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB;AACE,aAAO;AAAA,EACX;AACF;AAeO,SAAS,yBACd,SACW;AACX,QAAM,WAAW,SAAS,YAAY,qBAAqB;AAC3D,QAAM,eACJ,aAAa,aAAa,gBAAgB,OAAO,cAAc;AACjE,SAAO,IAAI,UAAU;AAAA,IACnB;AAAA,IACA,OAAO;AAAA,IACP,aAAa;AAAA,IACb,GAAG;AAAA,EACL,CAAC;AACH;AAWO,SAAS,qBACd,SACW;AACX,QAAM,WAAW,SAAS,YAAY,qBAAqB;AAC3D,QAAM,eACJ,aAAa,aAAa,gBAAgB,WAAW,cAAc;AACrE,SAAO,IAAI,UAAU;AAAA,IACnB;AAAA,IACA,OAAO;AAAA,IACP,aAAa;AAAA,IACb,GAAG;AAAA,EACL,CAAC;AACH;AAWO,SAAS,gBACd,SACW;AACX,QAAM,WAAW,SAAS,YAAY,qBAAqB;AAC3D,QAAM,eACJ,aAAa,aACT,gBAAgB,OAChB,cAAc;AACpB,SAAO,IAAI,UAAU;AAAA,IACnB;AAAA,IACA,OAAO;AAAA,IACP,aAAa;AAAA,IACb,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,qBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,gBAAgB;AAAA,IACvB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,mBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,cAAc;AAAA,IACrB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,sBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,iBAAiB;AAAA,IACxB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,iBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,YAAY;AAAA,IACnB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,uBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,kBAAkB;AAAA,IACzB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,kBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,kBAAkB;AAAA,IACzB,GAAG;AAAA,EACL,CAAC;AACH;","names":[]}

package/dist/{chunk-MJS33AAS.js → chunk-ULT7T7O6.js}

@@ -146,18 +146,88 @@ function resetGlobalRateLimiter() {
 }
 
 // src/token-tracker.ts
-var MODEL_COSTS = {
-  "gpt-4o": { promptCostPer1K: 5e-3, completionCostPer1K: 0.015 },
-  "gpt-4o-mini": { promptCostPer1K: 15e-5, completionCostPer1K: 6e-4 },
-  "gpt-4-turbo": { promptCostPer1K: 0.01, completionCostPer1K: 0.03 },
-  "gpt-4": { promptCostPer1K: 0.03, completionCostPer1K: 0.06 },
-  "gpt-3.5-turbo": {
-    promptCostPer1K: 5e-4,
-    completionCostPer1K: 15e-4
-  }
+import { appendFileSync, mkdirSync, readFileSync, writeFileSync } from "fs";
+import { dirname, join } from "path";
+var ALMADAR_ROOT = process.env["ALMADAR_ROOT"] ?? process.cwd();
+var PRICING_CACHE_PATH = join(ALMADAR_ROOT, ".llm-pricing-cache.json");
+var CALL_LOG_PATH = join(ALMADAR_ROOT, ".llm-call-log.jsonl");
+var CACHE_TTL_MS = 24 * 60 * 60 * 1e3;
+var MODEL_ID_MAP = {
+  // Anthropic
+  "claude-opus-4-5-20250929": "anthropic/claude-opus-4.5",
+  "claude-sonnet-4-5-20250929": "anthropic/claude-sonnet-4.5",
+  "claude-sonnet-4-20250514": "anthropic/claude-sonnet-4",
+  "claude-3-5-haiku-20241022": "anthropic/claude-3.5-haiku",
+  // DeepSeek — map to current versions on OpenRouter
+  "deepseek-chat": "deepseek/deepseek-v3.2",
+  "deepseek-coder": "deepseek/deepseek-v3.2",
+  "deepseek-reasoner": "deepseek/deepseek-r1-0528",
+  // Kimi
+  "kimi-k2.5": "moonshotai/kimi-k2.5"
 };
+var FALLBACK_COSTS = {};
+var pricingCache = null;
+function loadCachedPricing() {
+  try {
+    const raw = readFileSync(PRICING_CACHE_PATH, "utf-8");
+    const parsed = JSON.parse(raw);
+    if (Date.now() - parsed.fetchedAt < CACHE_TTL_MS) {
+      return parsed;
+    }
+  } catch {
+  }
+  return null;
+}
+async function fetchPricingFromOpenRouter() {
+  const res = await fetch("https://openrouter.ai/api/v1/models");
+  if (!res.ok) throw new Error(`OpenRouter models API: HTTP ${res.status}`);
+  const json = await res.json();
+  const models = {};
+  for (const m of json.data ?? []) {
+    const promptPerToken = parseFloat(m.pricing?.prompt ?? "0");
+    const completionPerToken = parseFloat(m.pricing?.completion ?? "0");
+    if (promptPerToken > 0 || completionPerToken > 0) {
+      models[m.id] = {
+        promptCostPer1K: promptPerToken * 1e3,
+        completionCostPer1K: completionPerToken * 1e3
+      };
+    }
+  }
+  return models;
+}
+function getPricing() {
+  if (pricingCache) return pricingCache.models;
+  const diskCache = loadCachedPricing();
+  if (diskCache) {
+    pricingCache = diskCache;
+    return diskCache.models;
+  }
+  refreshPricingCache();
+  return FALLBACK_COSTS;
+}
+function refreshPricingCache() {
+  fetchPricingFromOpenRouter().then((models) => {
+    pricingCache = { fetchedAt: Date.now(), models };
+    try {
+      mkdirSync(dirname(PRICING_CACHE_PATH), { recursive: true });
+      writeFileSync(PRICING_CACHE_PATH, JSON.stringify(pricingCache));
+    } catch {
+    }
+  }).catch(() => {
+  });
+}
+function getCostForModel(model) {
+  const pricing = getPricing();
+  const orId = MODEL_ID_MAP[model];
+  if (orId && pricing[orId]) return pricing[orId];
+  if (pricing[model]) return pricing[model];
+  for (const [key, cost] of Object.entries(pricing)) {
+    if (key.includes(model) || model.includes(key.split("/")[1] ?? "")) return cost;
+  }
+  return { promptCostPer1K: 0, completionCostPer1K: 0 };
+}
 var TokenTracker = class {
-  constructor(model = "gpt-4o") {
+  constructor(model = "claude-sonnet-4-5-20250929") {
     this.usage = {
       promptTokens: 0,
       completionTokens: 0,
@@ -166,17 +236,35 @@ var TokenTracker = class {
     };
     this.model = model;
   }
-  addUsage(promptTokens, completionTokens) {
+  addUsage(promptTokens, completionTokens, options) {
     this.usage.promptTokens += promptTokens;
     this.usage.completionTokens += completionTokens;
     this.usage.totalTokens += promptTokens + completionTokens;
     this.usage.callCount++;
+    const costs = getCostForModel(this.model);
+    const estimatedCost = promptTokens / 1e3 * costs.promptCostPer1K + completionTokens / 1e3 * costs.completionCostPer1K;
+    const entry = {
+      timestamp: (/* @__PURE__ */ new Date()).toISOString(),
+      provider: options?.provider ?? "unknown",
+      model: this.model,
+      promptTokens,
+      completionTokens,
+      totalTokens: promptTokens + completionTokens,
+      estimatedCost,
+      durationMs: options?.durationMs,
+      source: "local-log"
+    };
+    try {
+      mkdirSync(dirname(CALL_LOG_PATH), { recursive: true });
+      appendFileSync(CALL_LOG_PATH, JSON.stringify(entry) + "\n");
+    } catch {
+    }
   }
   getSummary() {
     return { ...this.usage };
   }
   getEstimatedCost() {
-    const costs = MODEL_COSTS[this.model] || MODEL_COSTS["gpt-4o"];
+    const costs = getCostForModel(this.model);
     const promptCost = this.usage.promptTokens / 1e3 * costs.promptCostPer1K;
     const completionCost = this.usage.completionTokens / 1e3 * costs.completionCostPer1K;
     return promptCost + completionCost;
@@ -231,4 +319,4 @@ export {
   getGlobalTokenTracker,
   resetGlobalTokenTracker
 };
-//# sourceMappingURL=chunk-MJS33AAS.js.map
+//# sourceMappingURL=chunk-ULT7T7O6.js.map

package/dist/chunk-ULT7T7O6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/rate-limiter.ts","../src/token-tracker.ts"],"sourcesContent":["/**\n * Rate Limiter for LLM API Calls\n *\n * Implements token bucket algorithm with:\n * - Configurable requests per minute/second\n * - Automatic backoff on 429 errors\n * - Queue for pending requests\n *\n * @packageDocumentation\n */\n\nexport interface RateLimiterOptions {\n  /** Maximum requests per minute (default: 60) */\n  requestsPerMinute?: number;\n  /** Maximum requests per second (default: 3) */\n  requestsPerSecond?: number;\n  /** Maximum concurrent requests (default: 5) */\n  maxConcurrent?: number;\n  /** Base delay for exponential backoff in ms (default: 1000) */\n  baseBackoffMs?: number;\n  /** Maximum backoff delay in ms (default: 60000) */\n  maxBackoffMs?: number;\n}\n\ninterface QueuedRequest<T> {\n  execute: () => Promise<T>;\n  resolve: (value: T) => void;\n  reject: (error: Error) => void;\n  retryCount: number;\n}\n\n/**\n * Rate limiter for LLM API calls using token bucket algorithm.\n *\n * @example\n * ```typescript\n * const limiter = new RateLimiter({ requestsPerMinute: 30 });\n * const result = await limiter.execute(() => llm.invoke(messages));\n * ```\n */\nexport class RateLimiter {\n  private requestsPerMinute: number;\n  private requestsPerSecond: number;\n  private maxConcurrent: number;\n  private baseBackoffMs: number;\n  private maxBackoffMs: number;\n\n  private minuteTokens: number;\n  private secondTokens: number;\n  private activeRequests: number = 0;\n  private queue: QueuedRequest<unknown>[] = [];\n  private lastMinuteReset: number = Date.now();\n  private lastSecondReset: number = Date.now();\n  private processing: boolean = false;\n  private currentBackoffMs: number = 0;\n\n  constructor(options: RateLimiterOptions = {}) {\n    this.requestsPerMinute = options.requestsPerMinute ?? 60;\n    this.requestsPerSecond = options.requestsPerSecond ?? 3;\n    this.maxConcurrent = options.maxConcurrent ?? 5;\n    this.baseBackoffMs = options.baseBackoffMs ?? 1000;\n    this.maxBackoffMs = options.maxBackoffMs ?? 60000;\n\n    this.minuteTokens = this.requestsPerMinute;\n    this.secondTokens = this.requestsPerSecond;\n  }\n\n  async execute<T>(fn: () => Promise<T>, _maxRetries: number = 3): Promise<T> {\n    return new Promise<T>((resolve, reject) => {\n      this.queue.push({\n        execute: fn as () => Promise<unknown>,\n        resolve: resolve as (value: unknown) => void,\n        reject,\n        retryCount: 0,\n      });\n      this.processQueue();\n    });\n  }\n\n  getStatus(): {\n    queueLength: number;\n    activeRequests: number;\n    minuteTokens: number;\n    secondTokens: number;\n    backoffMs: number;\n  } {\n    return {\n      queueLength: this.queue.length,\n      activeRequests: this.activeRequests,\n      minuteTokens: this.minuteTokens,\n      secondTokens: this.secondTokens,\n      backoffMs: this.currentBackoffMs,\n    };\n  }\n\n  reset(): void {\n    this.minuteTokens = this.requestsPerMinute;\n    this.secondTokens = this.requestsPerSecond;\n    this.activeRequests = 0;\n    this.queue = [];\n    this.currentBackoffMs = 0;\n    this.lastMinuteReset = Date.now();\n    this.lastSecondReset = Date.now();\n  }\n\n  private async processQueue(): Promise<void> {\n    if (this.processing) return;\n    this.processing = true;\n\n    while (this.queue.length > 0) {\n      this.refillTokens();\n\n      if (!this.canMakeRequest()) {\n        const waitTime = this.getWaitTime();\n        await this.sleep(waitTime);\n        continue;\n      }\n\n      if (this.currentBackoffMs > 0) {\n        await this.sleep(this.currentBackoffMs);\n        this.currentBackoffMs = 0;\n      }\n\n      const request = this.queue.shift();\n      if (!request) continue;\n\n      this.consumeTokens();\n      this.activeRequests++;\n\n      try {\n        const result = await request.execute();\n        request.resolve(result);\n        this.currentBackoffMs = 0;\n      } catch (error) {\n        const err = error instanceof Error ? error : new Error(String(error));\n\n        if (this.isRateLimitError(err)) {\n          this.currentBackoffMs = Math.min(\n            this.baseBackoffMs * Math.pow(2, request.retryCount),\n            this.maxBackoffMs,\n          );\n\n          console.warn(\n            `[RateLimiter] Rate limited. Backing off for ${this.currentBackoffMs}ms ` +\n              `(retry ${request.retryCount + 1})`,\n          );\n\n          if (request.retryCount < 3) {\n            this.queue.unshift({\n              ...request,\n              retryCount: request.retryCount + 1,\n            });\n          } else {\n            request.reject(\n              new Error(\n                `Rate limit exceeded after ${request.retryCount + 1} retries: ${err.message}`,\n              ),\n            );\n          }\n        } else {\n          request.reject(err);\n        }\n      } finally {\n        this.activeRequests--;\n      }\n    }\n\n    this.processing = false;\n  }\n\n  private refillTokens(): void {\n    const now = Date.now();\n    if (now - this.lastMinuteReset >= 60000) {\n      this.minuteTokens = this.requestsPerMinute;\n      this.lastMinuteReset = now;\n    }\n    if (now - this.lastSecondReset >= 1000) {\n      this.secondTokens = this.requestsPerSecond;\n      this.lastSecondReset = now;\n    }\n  }\n\n  private canMakeRequest(): boolean {\n    return (\n      this.minuteTokens > 0 &&\n      this.secondTokens > 0 &&\n      this.activeRequests < this.maxConcurrent\n    );\n  }\n\n  private consumeTokens(): void {\n    this.minuteTokens--;\n    this.secondTokens--;\n  }\n\n  private getWaitTime(): number {\n    const now = Date.now();\n    if (this.secondTokens <= 0) {\n      return Math.max(0, 1000 - (now - this.lastSecondReset));\n    }\n    if (this.minuteTokens <= 0) {\n      return Math.max(0, 60000 - (now - this.lastMinuteReset));\n    }\n    return 100;\n  }\n\n  private isRateLimitError(error: Error): boolean {\n    const message = error.message.toLowerCase();\n    return (\n      message.includes('429') ||\n      message.includes('rate limit') ||\n      message.includes('too many requests') ||\n      message.includes('quota exceeded')\n    );\n  }\n\n  private sleep(ms: number): Promise<void> {\n    return new Promise((resolve) => setTimeout(resolve, ms));\n  }\n}\n\n// Singleton instance\nlet globalRateLimiter: RateLimiter | null = null;\n\nexport function getGlobalRateLimiter(\n  options?: RateLimiterOptions,\n): RateLimiter {\n  if (!globalRateLimiter) {\n    globalRateLimiter = new RateLimiter(options);\n  }\n  return globalRateLimiter;\n}\n\nexport function resetGlobalRateLimiter(): void {\n  globalRateLimiter?.reset();\n  globalRateLimiter = null;\n}\n","/**\n * Token Tracker for LLM Usage\n *\n * Tracks token usage across multiple LLM calls for:\n * - Cost estimation (pricing fetched from OpenRouter models API)\n * - Usage monitoring\n * - Quota management\n * - Per-call JSONL logging\n *\n * @packageDocumentation\n */\n\nimport { appendFileSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';\nimport { dirname, join } from 'node:path';\n\nexport interface TokenUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  callCount: number;\n}\n\nexport interface TokenCost {\n  promptCostPer1K: number;\n  completionCostPer1K: number;\n}\n\nexport interface CallLogEntry {\n  timestamp: string;\n  provider: string;\n  model: string;\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  estimatedCost: number;\n  durationMs?: number;\n  source: 'local-log';\n}\n\n// ---------------------------------------------------------------------------\n// Pricing: fetched from OpenRouter /api/v1/models, cached to disk for 24h\n// ---------------------------------------------------------------------------\n\nconst ALMADAR_ROOT = process.env['ALMADAR_ROOT'] ?? process.cwd();\nconst PRICING_CACHE_PATH = join(ALMADAR_ROOT, '.llm-pricing-cache.json');\nconst CALL_LOG_PATH = join(ALMADAR_ROOT, '.llm-call-log.jsonl');\nconst CACHE_TTL_MS = 24 * 60 * 60 * 1000; // 24 hours\n\n/** Map from our local model name to OpenRouter model ID */\nconst MODEL_ID_MAP: Record<string, string> = {\n  // Anthropic\n  'claude-opus-4-5-20250929': 'anthropic/claude-opus-4.5',\n  'claude-sonnet-4-5-20250929': 'anthropic/claude-sonnet-4.5',\n  'claude-sonnet-4-20250514': 'anthropic/claude-sonnet-4',\n  'claude-3-5-haiku-20241022': 'anthropic/claude-3.5-haiku',\n  // DeepSeek — map to current versions on OpenRouter\n  'deepseek-chat': 'deepseek/deepseek-v3.2',\n  'deepseek-coder': 'deepseek/deepseek-v3.2',\n  'deepseek-reasoner': 'deepseek/deepseek-r1-0528',\n  // Kimi\n  'kimi-k2.5': 'moonshotai/kimi-k2.5',\n};\n\n// Fallback: zero cost — forces OpenRouter fetch for real pricing\nconst FALLBACK_COSTS: Record<string, TokenCost> = {};\n\ninterface PricingCache {\n  fetchedAt: number;\n  models: Record<string, TokenCost>;\n}\n\nlet pricingCache: PricingCache | null = null;\n\nfunction loadCachedPricing(): PricingCache | null {\n  try {\n    const raw = readFileSync(PRICING_CACHE_PATH, 'utf-8');\n    const parsed = JSON.parse(raw) as PricingCache;\n    if (Date.now() - parsed.fetchedAt < CACHE_TTL_MS) {\n      return parsed;\n    }\n  } catch {\n    // No cache or expired\n  }\n  return null;\n}\n\nasync function fetchPricingFromOpenRouter(): Promise<Record<string, TokenCost>> {\n  const res = await fetch('https://openrouter.ai/api/v1/models');\n  if (!res.ok) throw new Error(`OpenRouter models API: HTTP ${res.status}`);\n  const json = await res.json() as { data?: Array<{ id: string; pricing?: { prompt?: string; completion?: string } }> };\n  const models: Record<string, TokenCost> = {};\n  for (const m of json.data ?? []) {\n    const promptPerToken = parseFloat(m.pricing?.prompt ?? '0');\n    const completionPerToken = parseFloat(m.pricing?.completion ?? '0');\n    if (promptPerToken > 0 || completionPerToken > 0) {\n      models[m.id] = {\n        promptCostPer1K: promptPerToken * 1000,\n        completionCostPer1K: completionPerToken * 1000,\n      };\n    }\n  }\n  return models;\n}\n\n/**\n * Get pricing for all models. Uses 24h disk cache, fetches from OpenRouter on miss.\n * Non-blocking: returns cached/fallback immediately, refreshes in background if stale.\n */\nfunction getPricing(): Record<string, TokenCost> {\n  if (pricingCache) return pricingCache.models;\n\n  const diskCache = loadCachedPricing();\n  if (diskCache) {\n    pricingCache = diskCache;\n    return diskCache.models;\n  }\n\n  // Trigger background fetch, return fallback for now\n  refreshPricingCache();\n  return FALLBACK_COSTS;\n}\n\nfunction refreshPricingCache(): void {\n  fetchPricingFromOpenRouter()\n    .then((models) => {\n      pricingCache = { fetchedAt: Date.now(), models };\n      try {\n        mkdirSync(dirname(PRICING_CACHE_PATH), { recursive: true });\n        writeFileSync(PRICING_CACHE_PATH, JSON.stringify(pricingCache));\n      } catch {\n        // Non-critical\n      }\n    })\n    .catch(() => {\n      // Silently fail, use fallback\n    });\n}\n\nfunction getCostForModel(model: string): TokenCost {\n  const pricing = getPricing();\n  // Try direct match on OpenRouter ID\n  const orId = MODEL_ID_MAP[model];\n  if (orId && pricing[orId]) return pricing[orId];\n  // Try direct key match (e.g., user passed \"openai/gpt-4o\")\n  if (pricing[model]) return pricing[model];\n  // Fuzzy: find first key containing the model name\n  for (const [key, cost] of Object.entries(pricing)) {\n    if (key.includes(model) || model.includes(key.split('/')[1] ?? '')) return cost;\n  }\n  // No pricing available — return zero (OpenRouter fetch pending or model not listed)\n  return { promptCostPer1K: 0, completionCostPer1K: 0 };\n}\n\n// ---------------------------------------------------------------------------\n// TokenTracker\n// ---------------------------------------------------------------------------\n\nexport class TokenTracker {\n  private model: string;\n  private usage: TokenUsage = {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n    callCount: 0,\n  };\n\n  constructor(model: string = 'claude-sonnet-4-5-20250929') {\n    this.model = model;\n  }\n\n  addUsage(promptTokens: number, completionTokens: number, options?: { provider?: string; durationMs?: number }): void {\n    this.usage.promptTokens += promptTokens;\n    this.usage.completionTokens += completionTokens;\n    this.usage.totalTokens += promptTokens + completionTokens;\n    this.usage.callCount++;\n\n    const costs = getCostForModel(this.model);\n    const estimatedCost =\n      (promptTokens / 1000) * costs.promptCostPer1K +\n      (completionTokens / 1000) * costs.completionCostPer1K;\n\n    const entry: CallLogEntry = {\n      timestamp: new Date().toISOString(),\n      provider: options?.provider ?? 'unknown',\n      model: this.model,\n      promptTokens,\n      completionTokens,\n      totalTokens: promptTokens + completionTokens,\n      estimatedCost,\n      durationMs: options?.durationMs,\n      source: 'local-log',\n    };\n\n    try {\n      mkdirSync(dirname(CALL_LOG_PATH), { recursive: true });\n      appendFileSync(CALL_LOG_PATH, JSON.stringify(entry) + '\\n');\n    } catch {\n      // Non-critical: don't break LLM calls if logging fails\n    }\n  }\n\n  getSummary(): TokenUsage {\n    return { ...this.usage };\n  }\n\n  getEstimatedCost(): number {\n    const costs = getCostForModel(this.model);\n    const promptCost =\n      (this.usage.promptTokens / 1000) * costs.promptCostPer1K;\n    const completionCost =\n      (this.usage.completionTokens / 1000) * costs.completionCostPer1K;\n    return promptCost + completionCost;\n  }\n\n  getFormattedCost(): string {\n    const cost = this.getEstimatedCost();\n    return `$${cost.toFixed(4)}`;\n  }\n\n  getReport(): string {\n    const summary = this.getSummary();\n    const cost = this.getEstimatedCost();\n    return [\n      `Token Usage Report (${this.model})`,\n      `─────────────────────────────`,\n      `Calls:              ${summary.callCount}`,\n      `Prompt Tokens:      ${summary.promptTokens.toLocaleString()}`,\n      `Completion Tokens:  ${summary.completionTokens.toLocaleString()}`,\n      `Total Tokens:       ${summary.totalTokens.toLocaleString()}`,\n      `Estimated Cost:     $${cost.toFixed(4)}`,\n    ].join('\\n');\n  }\n\n  reset(): void {\n    this.usage = {\n      promptTokens: 0,\n      completionTokens: 0,\n      totalTokens: 0,\n      callCount: 0,\n    };\n  }\n\n  setModel(model: string): void {\n    this.model = model;\n  }\n}\n\n// Global tracker instance\nlet globalTracker: TokenTracker | null = null;\n\nexport function getGlobalTokenTracker(model?: string): TokenTracker {\n  if (!globalTracker) {\n    globalTracker = new TokenTracker(model);\n  } else if (model) {\n    globalTracker.setModel(model);\n  }\n  return globalTracker;\n}\n\nexport function resetGlobalTokenTracker(): void {\n  globalTracker?.reset();\n}\n\nexport function getCallLogPath(): string {\n  return CALL_LOG_PATH;\n}\n\n/** Force-refresh the pricing cache from OpenRouter. */\nexport async function refreshPricing(): Promise<void> {\n  const models = await fetchPricingFromOpenRouter();\n  pricingCache = { fetchedAt: Date.now(), models };\n  mkdirSync(dirname(PRICING_CACHE_PATH), { recursive: true });\n  writeFileSync(PRICING_CACHE_PATH, JSON.stringify(pricingCache));\n}\n"],"mappings":";AAwCO,IAAM,cAAN,MAAkB;AAAA,EAgBvB,YAAY,UAA8B,CAAC,GAAG;AAP9C,SAAQ,iBAAyB;AACjC,SAAQ,QAAkC,CAAC;AAC3C,SAAQ,kBAA0B,KAAK,IAAI;AAC3C,SAAQ,kBAA0B,KAAK,IAAI;AAC3C,SAAQ,aAAsB;AAC9B,SAAQ,mBAA2B;AAGjC,SAAK,oBAAoB,QAAQ,qBAAqB;AACtD,SAAK,oBAAoB,QAAQ,qBAAqB;AACtD,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,eAAe,QAAQ,gBAAgB;AAE5C,SAAK,eAAe,KAAK;AACzB,SAAK,eAAe,KAAK;AAAA,EAC3B;AAAA,EAEA,MAAM,QAAW,IAAsB,cAAsB,GAAe;AAC1E,WAAO,IAAI,QAAW,CAAC,SAAS,WAAW;AACzC,WAAK,MAAM,KAAK;AAAA,QACd,SAAS;AAAA,QACT;AAAA,QACA;AAAA,QACA,YAAY;AAAA,MACd,CAAC;AACD,WAAK,aAAa;AAAA,IACpB,CAAC;AAAA,EACH;AAAA,EAEA,YAME;AACA,WAAO;AAAA,MACL,aAAa,KAAK,MAAM;AAAA,MACxB,gBAAgB,KAAK;AAAA,MACrB,cAAc,KAAK;AAAA,MACnB,cAAc,KAAK;AAAA,MACnB,WAAW,KAAK;AAAA,IAClB;AAAA,EACF;AAAA,EAEA,QAAc;AACZ,SAAK,eAAe,KAAK;AACzB,SAAK,eAAe,KAAK;AACzB,SAAK,iBAAiB;AACtB,SAAK,QAAQ,CAAC;AACd,SAAK,mBAAmB;AACxB,SAAK,kBAAkB,KAAK,IAAI;AAChC,SAAK,kBAAkB,KAAK,IAAI;AAAA,EAClC;AAAA,EAEA,MAAc,eAA8B;AAC1C,QAAI,KAAK,WAAY;AACrB,SAAK,aAAa;AAElB,WAAO,KAAK,MAAM,SAAS,GAAG;AAC5B,WAAK,aAAa;AAElB,UAAI,CAAC,KAAK,eAAe,GAAG;AAC1B,cAAM,WAAW,KAAK,YAAY;AAClC,cAAM,KAAK,MAAM,QAAQ;AACzB;AAAA,MACF;AAEA,UAAI,KAAK,mBAAmB,GAAG;AAC7B,cAAM,KAAK,MAAM,KAAK,gBAAgB;AACtC,aAAK,mBAAmB;AAAA,MAC1B;AAEA,YAAM,UAAU,KAAK,MAAM,MAAM;AACjC,UAAI,CAAC,QAAS;AAEd,WAAK,cAAc;AACnB,WAAK;AAEL,UAAI;AACF,cAAM,SAAS,MAAM,QAAQ,QAAQ;AACrC,gBAAQ,QAAQ,MAAM;AACtB,aAAK,mBAAmB;AAAA,MAC1B,SAAS,OAAO;AACd,cAAM,MAAM,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;AAEpE,YAAI,KAAK,iBAAiB,GAAG,GAAG;AAC9B,eAAK,mBAAmB,KAAK;AAAA,YAC3B,KAAK,gBAAgB,KAAK,IAAI,GAAG,QAAQ,UAAU;AAAA,YACnD,KAAK;AAAA,UACP;AAEA,kBAAQ;AAAA,YACN,+CAA+C,KAAK,gBAAgB,aACxD,QAAQ,aAAa,CAAC;AAAA,UACpC;AAEA,cAAI,QAAQ,aAAa,GAAG;AAC1B,iBAAK,MAAM,QAAQ;AAAA,cACjB,GAAG;AAAA,cACH,YAAY,QAAQ,aAAa;AAAA,YACnC,CAAC;AAAA,UACH,OAAO;AACL,oBAAQ;AAAA,cACN,IAAI;AAAA,gBACF,6BAA6B,QAAQ,aAAa,CAAC,aAAa,IAAI,OAAO;AAAA,cAC7E;AAAA,YACF;AAAA,UACF;AAAA,QACF,OAAO;AACL,kBAAQ,OAAO,GAAG;AAAA,QACpB;AAAA,MACF,UAAE;AACA,aAAK;AAAA,MACP;AAAA,IACF;AAEA,SAAK,aAAa;AAAA,EACpB;AAAA,EAEQ,eAAqB;AAC3B,UAAM,MAAM,KAAK,IAAI;AACrB,QAAI,MAAM,KAAK,mBAAmB,KAAO;AACvC,WAAK,eAAe,KAAK;AACzB,WAAK,kBAAkB;AAAA,IACzB;AACA,QAAI,MAAM,KAAK,mBAAmB,KAAM;AACtC,WAAK,eAAe,KAAK;AACzB,WAAK,kBAAkB;AAAA,IACzB;AAAA,EACF;AAAA,EAEQ,iBAA0B;AAChC,WACE,KAAK,eAAe,KACpB,KAAK,eAAe,KACpB,KAAK,iBAAiB,KAAK;AAAA,EAE/B;AAAA,EAEQ,gBAAsB;AAC5B,SAAK;AACL,SAAK;AAAA,EACP;AAAA,EAEQ,cAAsB;AAC5B,UAAM,MAAM,KAAK,IAAI;AACrB,QAAI,KAAK,gBAAgB,GAAG;AAC1B,aAAO,KAAK,IAAI,GAAG,OAAQ,MAAM,KAAK,gBAAgB;AAAA,IACxD;AACA,QAAI,KAAK,gBAAgB,GAAG;AAC1B,aAAO,KAAK,IAAI,GAAG,OAAS,MAAM,KAAK,gBAAgB;AAAA,IACzD;AACA,WAAO;AAAA,EACT;AAAA,EAEQ,iBAAiB,OAAuB;AAC9C,UAAM,UAAU,MAAM,QAAQ,YAAY;AAC1C,WACE,QAAQ,SAAS,KAAK,KACtB,QAAQ,SAAS,YAAY,KAC7B,QAAQ,SAAS,mBAAmB,KACpC,QAAQ,SAAS,gBAAgB;AAAA,EAErC;AAAA,EAEQ,MAAM,IAA2B;AACvC,WAAO,IAAI,QAAQ,CAAC,YAAY,WAAW,SAAS,EAAE,CAAC;AAAA,EACzD;AACF;AAGA,IAAI,oBAAwC;AAErC,SAAS,qBACd,SACa;AACb,MAAI,CAAC,mBAAmB;AACtB,wBAAoB,IAAI,YAAY,OAAO;AAAA,EAC7C;AACA,SAAO;AACT;AAEO,SAAS,yBAA+B;AAC7C,qBAAmB,MAAM;AACzB,sBAAoB;AACtB;;;AChOA,SAAS,gBAAgB,WAAW,cAAc,qBAAqB;AACvE,SAAS,SAAS,YAAY;AA8B9B,IAAM,eAAe,QAAQ,IAAI,cAAc,KAAK,QAAQ,IAAI;AAChE,IAAM,qBAAqB,KAAK,cAAc,yBAAyB;AACvE,IAAM,gBAAgB,KAAK,cAAc,qBAAqB;AAC9D,IAAM,eAAe,KAAK,KAAK,KAAK;AAGpC,IAAM,eAAuC;AAAA;AAAA,EAE3C,4BAA4B;AAAA,EAC5B,8BAA8B;AAAA,EAC9B,4BAA4B;AAAA,EAC5B,6BAA6B;AAAA;AAAA,EAE7B,iBAAiB;AAAA,EACjB,kBAAkB;AAAA,EAClB,qBAAqB;AAAA;AAAA,EAErB,aAAa;AACf;AAGA,IAAM,iBAA4C,CAAC;AAOnD,IAAI,eAAoC;AAExC,SAAS,oBAAyC;AAChD,MAAI;AACF,UAAM,MAAM,aAAa,oBAAoB,OAAO;AACpD,UAAM,SAAS,KAAK,MAAM,GAAG;AAC7B,QAAI,KAAK,IAAI,IAAI,OAAO,YAAY,cAAc;AAChD,aAAO;AAAA,IACT;AAAA,EACF,QAAQ;AAAA,EAER;AACA,SAAO;AACT;AAEA,eAAe,6BAAiE;AAC9E,QAAM,MAAM,MAAM,MAAM,qCAAqC;AAC7D,MAAI,CAAC,IAAI,GAAI,OAAM,IAAI,MAAM,+BAA+B,IAAI,MAAM,EAAE;AACxE,QAAM,OAAO,MAAM,IAAI,KAAK;AAC5B,QAAM,SAAoC,CAAC;AAC3C,aAAW,KAAK,KAAK,QAAQ,CAAC,GAAG;AAC/B,UAAM,iBAAiB,WAAW,EAAE,SAAS,UAAU,GAAG;AAC1D,UAAM,qBAAqB,WAAW,EAAE,SAAS,cAAc,GAAG;AAClE,QAAI,iBAAiB,KAAK,qBAAqB,GAAG;AAChD,aAAO,EAAE,EAAE,IAAI;AAAA,QACb,iBAAiB,iBAAiB;AAAA,QAClC,qBAAqB,qBAAqB;AAAA,MAC5C;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAMA,SAAS,aAAwC;AAC/C,MAAI,aAAc,QAAO,aAAa;AAEtC,QAAM,YAAY,kBAAkB;AACpC,MAAI,WAAW;AACb,mBAAe;AACf,WAAO,UAAU;AAAA,EACnB;AAGA,sBAAoB;AACpB,SAAO;AACT;AAEA,SAAS,sBAA4B;AACnC,6BAA2B,EACxB,KAAK,CAAC,WAAW;AAChB,mBAAe,EAAE,WAAW,KAAK,IAAI,GAAG,OAAO;AAC/C,QAAI;AACF,gBAAU,QAAQ,kBAAkB,GAAG,EAAE,WAAW,KAAK,CAAC;AAC1D,oBAAc,oBAAoB,KAAK,UAAU,YAAY,CAAC;AAAA,IAChE,QAAQ;AAAA,IAER;AAAA,EACF,CAAC,EACA,MAAM,MAAM;AAAA,EAEb,CAAC;AACL;AAEA,SAAS,gBAAgB,OAA0B;AACjD,QAAM,UAAU,WAAW;AAE3B,QAAM,OAAO,aAAa,KAAK;AAC/B,MAAI,QAAQ,QAAQ,IAAI,EAAG,QAAO,QAAQ,IAAI;AAE9C,MAAI,QAAQ,KAAK,EAAG,QAAO,QAAQ,KAAK;AAExC,aAAW,CAAC,KAAK,IAAI,KAAK,OAAO,QAAQ,OAAO,GAAG;AACjD,QAAI,IAAI,SAAS,KAAK,KAAK,MAAM,SAAS,IAAI,MAAM,GAAG,EAAE,CAAC,KAAK,EAAE,EAAG,QAAO;AAAA,EAC7E;AAEA,SAAO,EAAE,iBAAiB,GAAG,qBAAqB,EAAE;AACtD;AAMO,IAAM,eAAN,MAAmB;AAAA,EASxB,YAAY,QAAgB,8BAA8B;AAP1D,SAAQ,QAAoB;AAAA,MAC1B,cAAc;AAAA,MACd,kBAAkB;AAAA,MAClB,aAAa;AAAA,MACb,WAAW;AAAA,IACb;AAGE,SAAK,QAAQ;AAAA,EACf;AAAA,EAEA,SAAS,cAAsB,kBAA0B,SAA4D;AACnH,SAAK,MAAM,gBAAgB;AAC3B,SAAK,MAAM,oBAAoB;AAC/B,SAAK,MAAM,eAAe,eAAe;AACzC,SAAK,MAAM;AAEX,UAAM,QAAQ,gBAAgB,KAAK,KAAK;AACxC,UAAM,gBACH,eAAe,MAAQ,MAAM,kBAC7B,mBAAmB,MAAQ,MAAM;AAEpC,UAAM,QAAsB;AAAA,MAC1B,YAAW,oBAAI,KAAK,GAAE,YAAY;AAAA,MAClC,UAAU,SAAS,YAAY;AAAA,MAC/B,OAAO,KAAK;AAAA,MACZ;AAAA,MACA;AAAA,MACA,aAAa,eAAe;AAAA,MAC5B;AAAA,MACA,YAAY,SAAS;AAAA,MACrB,QAAQ;AAAA,IACV;AAEA,QAAI;AACF,gBAAU,QAAQ,aAAa,GAAG,EAAE,WAAW,KAAK,CAAC;AACrD,qBAAe,eAAe,KAAK,UAAU,KAAK,IAAI,IAAI;AAAA,IAC5D,QAAQ;AAAA,IAER;AAAA,EACF;AAAA,EAEA,aAAyB;AACvB,WAAO,EAAE,GAAG,KAAK,MAAM;AAAA,EACzB;AAAA,EAEA,mBAA2B;AACzB,UAAM,QAAQ,gBAAgB,KAAK,KAAK;AACxC,UAAM,aACH,KAAK,MAAM,eAAe,MAAQ,MAAM;AAC3C,UAAM,iBACH,KAAK,MAAM,mBAAmB,MAAQ,MAAM;AAC/C,WAAO,aAAa;AAAA,EACtB;AAAA,EAEA,mBAA2B;AACzB,UAAM,OAAO,KAAK,iBAAiB;AACnC,WAAO,IAAI,KAAK,QAAQ,CAAC,CAAC;AAAA,EAC5B;AAAA,EAEA,YAAoB;AAClB,UAAM,UAAU,KAAK,WAAW;AAChC,UAAM,OAAO,KAAK,iBAAiB;AACnC,WAAO;AAAA,MACL,uBAAuB,KAAK,KAAK;AAAA,MACjC;AAAA,MACA,uBAAuB,QAAQ,SAAS;AAAA,MACxC,uBAAuB,QAAQ,aAAa,eAAe,CAAC;AAAA,MAC5D,uBAAuB,QAAQ,iBAAiB,eAAe,CAAC;AAAA,MAChE,uBAAuB,QAAQ,YAAY,eAAe,CAAC;AAAA,MAC3D,wBAAwB,KAAK,QAAQ,CAAC,CAAC;AAAA,IACzC,EAAE,KAAK,IAAI;AAAA,EACb;AAAA,EAEA,QAAc;AACZ,SAAK,QAAQ;AAAA,MACX,cAAc;AAAA,MACd,kBAAkB;AAAA,MAClB,aAAa;AAAA,MACb,WAAW;AAAA,IACb;AAAA,EACF;AAAA,EAEA,SAAS,OAAqB;AAC5B,SAAK,QAAQ;AAAA,EACf;AACF;AAGA,IAAI,gBAAqC;AAElC,SAAS,sBAAsB,OAA8B;AAClE,MAAI,CAAC,eAAe;AAClB,oBAAgB,IAAI,aAAa,KAAK;AAAA,EACxC,WAAW,OAAO;AAChB,kBAAc,SAAS,KAAK;AAAA,EAC9B;AACA,SAAO;AACT;AAEO,SAAS,0BAAgC;AAC9C,iBAAe,MAAM;AACvB;","names":[]}

package/dist/client.d.ts

@@ -1,4 +1,4 @@
-import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-DDH7JH5p.js';
+import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-B9tDNSMl.js';
 import { ChatOpenAI } from '@langchain/openai';
 import { ChatAnthropic } from '@langchain/anthropic';
 import { z } from 'zod';
@@ -146,11 +146,38 @@ declare class LLMClient {
         systemPrompt: string;
         userPrompt: string;
         maxTokens?: number;
+        signal?: AbortSignal;
     }): Promise<string>;
     callRawWithMetadata(options: {
         systemPrompt: string;
         userPrompt: string;
         maxTokens?: number;
+        signal?: AbortSignal;
+    }): Promise<Omit<LLMResponse<string>, 'data'> & {
+        raw: string;
+    }>;
+    /**
+     * Call the LLM with a structured messages array.
+     *
+     * Unlike callRawWithMetadata (which takes systemPrompt + userPrompt strings),
+     * this accepts a full conversation history with proper role separation.
+     * This enables:
+     * - Anthropic prompt caching on message boundaries (not just system prompt)
+     * - Proper tool_use/tool_result role handling across providers
+     * - Reduced token waste from string concatenation
+     *
+     * All providers support the messages format:
+     * - Anthropic: native messages API with cache_control
+     * - DeepSeek: OpenAI-compatible messages via ChatOpenAI
+     * - OpenRouter: OpenAI-compatible messages via ChatOpenAI
+     */
+    callWithMessages(options: {
+        messages: Array<{
+            role: string;
+            content: string;
+        }>;
+        maxTokens?: number;
+        signal?: AbortSignal;
     }): Promise<Omit<LLMResponse<string>, 'data'> & {
         raw: string;
     }>;

package/dist/client.js

@@ -18,9 +18,9 @@ import {
   getSharedLLMClient,
   isProviderAvailable,
   resetSharedLLMClient
-} from "./chunk-F2DMHMRH.js";
+} from "./chunk-OBGKLC3H.js";
 import "./chunk-LZGCEPHN.js";
-import "./chunk-MJS33AAS.js";
+import "./chunk-ULT7T7O6.js";
 export {
   ANTHROPIC_MODELS,
   DEEPSEEK_MODELS,

package/dist/index.d.ts

@@ -1,6 +1,6 @@
 import { LLMFinishReason, LLMClient } from './client.js';
 export { ANTHROPIC_MODELS, CacheAwareLLMCallOptions, CacheableBlock, DEEPSEEK_MODELS, KIMI_MODELS, LLMCallOptions, LLMClientOptions, LLMProvider, LLMResponse, LLMStreamChunk, LLMStreamOptions, LLMUsage, OPENAI_MODELS, OPENROUTER_MODELS, ProviderConfig, createAnthropicClient, createCreativeClient, createDeepSeekClient, createFixClient, createKimiClient, createOpenAIClient, createOpenRouterClient, createRequirementsClient, createZhipuClient, getAvailableProvider, getSharedLLMClient, isProviderAvailable, resetSharedLLMClient } from './client.js';
-export { a as RateLimiter, R as RateLimiterOptions, b as TokenTracker, T as TokenUsage, g as getGlobalRateLimiter, c as getGlobalTokenTracker, r as resetGlobalRateLimiter, d as resetGlobalTokenTracker } from './rate-limiter-DDH7JH5p.js';
+export { a as RateLimiter, R as RateLimiterOptions, b as TokenTracker, T as TokenUsage, g as getGlobalRateLimiter, c as getGlobalTokenTracker, r as resetGlobalRateLimiter, d as resetGlobalTokenTracker } from './rate-limiter-B9tDNSMl.js';
 export { autoCloseJson, extractJsonFromText, isValidJson, parseJsonResponse, safeParseJson } from './json-parser.js';
 import { z } from 'zod';
 export { JsonSchema, STRUCTURED_OUTPUT_MODELS, StructuredGenerationOptions, StructuredGenerationResult, StructuredOutputClient, StructuredOutputOptions, getStructuredOutputClient, isStructuredOutputAvailable, resetStructuredOutputClient } from './structured-output.js';

package/dist/index.js

@@ -18,7 +18,7 @@ import {
   getSharedLLMClient,
   isProviderAvailable,
   resetSharedLLMClient
-} from "./chunk-F2DMHMRH.js";
+} from "./chunk-OBGKLC3H.js";
 import {
   autoCloseJson,
   extractJsonFromText,
@@ -32,7 +32,7 @@ import {
   getStructuredOutputClient,
   isStructuredOutputAvailable,
   resetStructuredOutputClient
-} from "./chunk-3OVQNNPN.js";
+} from "./chunk-5DRKGB5R.js";
 import {
   RateLimiter,
   TokenTracker,
@@ -40,7 +40,7 @@ import {
   getGlobalTokenTracker,
   resetGlobalRateLimiter,
   resetGlobalTokenTracker
-} from "./chunk-MJS33AAS.js";
+} from "./chunk-ULT7T7O6.js";
 import {
   MasarError,
   MasarProvider,

package/dist/{rate-limiter-DDH7JH5p.d.ts → rate-limiter-B9tDNSMl.d.ts}

@@ -2,9 +2,10 @@
  * Token Tracker for LLM Usage
  *
  * Tracks token usage across multiple LLM calls for:
- * - Cost estimation
+ * - Cost estimation (pricing fetched from OpenRouter models API)
  * - Usage monitoring
  * - Quota management
+ * - Per-call JSONL logging
  *
  * @packageDocumentation
  */
@@ -18,7 +19,10 @@ declare class TokenTracker {
     private model;
     private usage;
     constructor(model?: string);
-    addUsage(promptTokens: number, completionTokens: number): void;
+    addUsage(promptTokens: number, completionTokens: number, options?: {
+        provider?: string;
+        durationMs?: number;
+    }): void;
     getSummary(): TokenUsage;
     getEstimatedCost(): number;
     getFormattedCost(): string;

package/dist/structured-output.d.ts

@@ -1,4 +1,4 @@
-import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-DDH7JH5p.js';
+import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-B9tDNSMl.js';
 import { z } from 'zod';
 
 /**

package/dist/structured-output.js

@@ -4,8 +4,8 @@ import {
   getStructuredOutputClient,
   isStructuredOutputAvailable,
   resetStructuredOutputClient
-} from "./chunk-3OVQNNPN.js";
-import "./chunk-MJS33AAS.js";
+} from "./chunk-5DRKGB5R.js";
+import "./chunk-ULT7T7O6.js";
 export {
   STRUCTURED_OUTPUT_MODELS,
   StructuredOutputClient,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@almadar/llm",
-  "version": "2.5.1",
+  "version": "2.8.2",
   "description": "Multi-provider LLM client with rate limiting, token tracking, structured outputs, and continuation handling",
   "type": "module",
   "main": "./dist/index.js",

package/src/client.ts

@@ -571,6 +571,7 @@ export class LLMClient {
               this.tokenTracker.addUsage(
                 usage.promptTokens,
                 usage.completionTokens,
+                { provider: this.provider },
               );
             }
           }
@@ -661,6 +662,7 @@ export class LLMClient {
     systemPrompt: string;
     userPrompt: string;
     maxTokens?: number;
+    signal?: AbortSignal;
   }): Promise<string> {
     const response = await this.callRawWithMetadata(options);
     return response.raw;
@@ -670,8 +672,9 @@ export class LLMClient {
     systemPrompt: string;
     userPrompt: string;
     maxTokens?: number;
+    signal?: AbortSignal;
   }): Promise<Omit<LLMResponse<string>, 'data'> & { raw: string }> {
-    const { systemPrompt, userPrompt, maxTokens } = options;
+    const { systemPrompt, userPrompt, maxTokens, signal } = options;
 
     return this.rateLimiter.execute(async () => {
       const modelToUse = maxTokens
@@ -686,6 +689,74 @@ export class LLMClient {
         this.provider === 'anthropic'
           ? addCacheControlToSystemMessages(messages)
           : messages,
+        signal ? { signal } : undefined,
+      );
+
+      let usage: LLMUsage | null = null;
+      if (response.usage_metadata) {
+        const usageMeta = response.usage_metadata as {
+          input_tokens?: number;
+          output_tokens?: number;
+        };
+        usage = {
+          promptTokens: usageMeta.input_tokens || 0,
+          completionTokens: usageMeta.output_tokens || 0,
+          totalTokens:
+            (usageMeta.input_tokens || 0) + (usageMeta.output_tokens || 0),
+        };
+
+        if (this.tokenTracker) {
+          this.tokenTracker.addUsage(
+            usage.promptTokens,
+            usage.completionTokens,
+          );
+        }
+      }
+
+      const finishReason = this.extractFinishReason(response);
+      const content =
+        typeof response.content === 'string'
+          ? response.content
+          : JSON.stringify(response.content);
+
+      return { raw: content, finishReason, usage };
+    });
+  }
+
+  /**
+   * Call the LLM with a structured messages array.
+   *
+   * Unlike callRawWithMetadata (which takes systemPrompt + userPrompt strings),
+   * this accepts a full conversation history with proper role separation.
+   * This enables:
+   * - Anthropic prompt caching on message boundaries (not just system prompt)
+   * - Proper tool_use/tool_result role handling across providers
+   * - Reduced token waste from string concatenation
+   *
+   * All providers support the messages format:
+   * - Anthropic: native messages API with cache_control
+   * - DeepSeek: OpenAI-compatible messages via ChatOpenAI
+   * - OpenRouter: OpenAI-compatible messages via ChatOpenAI
+   */
+  async callWithMessages(options: {
+    messages: Array<{ role: string; content: string }>;
+    maxTokens?: number;
+    signal?: AbortSignal;
+  }): Promise<Omit<LLMResponse<string>, 'data'> & { raw: string }> {
+    const { messages, maxTokens, signal } = options;
+
+    return this.rateLimiter.execute(async () => {
+      const modelToUse = maxTokens
+        ? this.getModelWithOptions({ maxTokens })
+        : this.model;
+
+      const langchainMessages = this.provider === 'anthropic'
+        ? addCacheControlToSystemMessages(messages)
+        : (messages as BaseMessageLike[]);
+
+      const response = await modelToUse.invoke(
+        langchainMessages,
+        signal ? { signal } : undefined,
       );
 
       let usage: LLMUsage | null = null;

package/src/structured-output.ts

@@ -267,7 +267,7 @@ export class StructuredOutputClient {
     };
 
     if (this.tokenTracker) {
-      this.tokenTracker.addUsage(usage.promptTokens, usage.completionTokens);
+      this.tokenTracker.addUsage(usage.promptTokens, usage.completionTokens, { provider: 'structured-output' });
     }
 
     console.log(

package/src/token-tracker.ts

@@ -2,13 +2,17 @@
  * Token Tracker for LLM Usage
  *
  * Tracks token usage across multiple LLM calls for:
- * - Cost estimation
+ * - Cost estimation (pricing fetched from OpenRouter models API)
  * - Usage monitoring
  * - Quota management
+ * - Per-call JSONL logging
  *
  * @packageDocumentation
  */
 
+import { appendFileSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';
+import { dirname, join } from 'node:path';
+
 export interface TokenUsage {
   promptTokens: number;
   completionTokens: number;
@@ -21,18 +25,136 @@ export interface TokenCost {
   completionCostPer1K: number;
 }
 
-// Pricing as of 2024 (update as needed)
-const MODEL_COSTS: Record<string, TokenCost> = {
-  'gpt-4o': { promptCostPer1K: 0.005, completionCostPer1K: 0.015 },
-  'gpt-4o-mini': { promptCostPer1K: 0.00015, completionCostPer1K: 0.0006 },
-  'gpt-4-turbo': { promptCostPer1K: 0.01, completionCostPer1K: 0.03 },
-  'gpt-4': { promptCostPer1K: 0.03, completionCostPer1K: 0.06 },
-  'gpt-3.5-turbo': {
-    promptCostPer1K: 0.0005,
-    completionCostPer1K: 0.0015,
-  },
+export interface CallLogEntry {
+  timestamp: string;
+  provider: string;
+  model: string;
+  promptTokens: number;
+  completionTokens: number;
+  totalTokens: number;
+  estimatedCost: number;
+  durationMs?: number;
+  source: 'local-log';
+}
+
+// ---------------------------------------------------------------------------
+// Pricing: fetched from OpenRouter /api/v1/models, cached to disk for 24h
+// ---------------------------------------------------------------------------
+
+const ALMADAR_ROOT = process.env['ALMADAR_ROOT'] ?? process.cwd();
+const PRICING_CACHE_PATH = join(ALMADAR_ROOT, '.llm-pricing-cache.json');
+const CALL_LOG_PATH = join(ALMADAR_ROOT, '.llm-call-log.jsonl');
+const CACHE_TTL_MS = 24 * 60 * 60 * 1000; // 24 hours
+
+/** Map from our local model name to OpenRouter model ID */
+const MODEL_ID_MAP: Record<string, string> = {
+  // Anthropic
+  'claude-opus-4-5-20250929': 'anthropic/claude-opus-4.5',
+  'claude-sonnet-4-5-20250929': 'anthropic/claude-sonnet-4.5',
+  'claude-sonnet-4-20250514': 'anthropic/claude-sonnet-4',
+  'claude-3-5-haiku-20241022': 'anthropic/claude-3.5-haiku',
+  // DeepSeek — map to current versions on OpenRouter
+  'deepseek-chat': 'deepseek/deepseek-v3.2',
+  'deepseek-coder': 'deepseek/deepseek-v3.2',
+  'deepseek-reasoner': 'deepseek/deepseek-r1-0528',
+  // Kimi
+  'kimi-k2.5': 'moonshotai/kimi-k2.5',
 };
 
+// Fallback: zero cost — forces OpenRouter fetch for real pricing
+const FALLBACK_COSTS: Record<string, TokenCost> = {};
+
+interface PricingCache {
+  fetchedAt: number;
+  models: Record<string, TokenCost>;
+}
+
+let pricingCache: PricingCache | null = null;
+
+function loadCachedPricing(): PricingCache | null {
+  try {
+    const raw = readFileSync(PRICING_CACHE_PATH, 'utf-8');
+    const parsed = JSON.parse(raw) as PricingCache;
+    if (Date.now() - parsed.fetchedAt < CACHE_TTL_MS) {
+      return parsed;
+    }
+  } catch {
+    // No cache or expired
+  }
+  return null;
+}
+
+async function fetchPricingFromOpenRouter(): Promise<Record<string, TokenCost>> {
+  const res = await fetch('https://openrouter.ai/api/v1/models');
+  if (!res.ok) throw new Error(`OpenRouter models API: HTTP ${res.status}`);
+  const json = await res.json() as { data?: Array<{ id: string; pricing?: { prompt?: string; completion?: string } }> };
+  const models: Record<string, TokenCost> = {};
+  for (const m of json.data ?? []) {
+    const promptPerToken = parseFloat(m.pricing?.prompt ?? '0');
+    const completionPerToken = parseFloat(m.pricing?.completion ?? '0');
+    if (promptPerToken > 0 || completionPerToken > 0) {
+      models[m.id] = {
+        promptCostPer1K: promptPerToken * 1000,
+        completionCostPer1K: completionPerToken * 1000,
+      };
+    }
+  }
+  return models;
+}
+
+/**
+ * Get pricing for all models. Uses 24h disk cache, fetches from OpenRouter on miss.
+ * Non-blocking: returns cached/fallback immediately, refreshes in background if stale.
+ */
+function getPricing(): Record<string, TokenCost> {
+  if (pricingCache) return pricingCache.models;
+
+  const diskCache = loadCachedPricing();
+  if (diskCache) {
+    pricingCache = diskCache;
+    return diskCache.models;
+  }
+
+  // Trigger background fetch, return fallback for now
+  refreshPricingCache();
+  return FALLBACK_COSTS;
+}
+
+function refreshPricingCache(): void {
+  fetchPricingFromOpenRouter()
+    .then((models) => {
+      pricingCache = { fetchedAt: Date.now(), models };
+      try {
+        mkdirSync(dirname(PRICING_CACHE_PATH), { recursive: true });
+        writeFileSync(PRICING_CACHE_PATH, JSON.stringify(pricingCache));
+      } catch {
+        // Non-critical
+      }
+    })
+    .catch(() => {
+      // Silently fail, use fallback
+    });
+}
+
+function getCostForModel(model: string): TokenCost {
+  const pricing = getPricing();
+  // Try direct match on OpenRouter ID
+  const orId = MODEL_ID_MAP[model];
+  if (orId && pricing[orId]) return pricing[orId];
+  // Try direct key match (e.g., user passed "openai/gpt-4o")
+  if (pricing[model]) return pricing[model];
+  // Fuzzy: find first key containing the model name
+  for (const [key, cost] of Object.entries(pricing)) {
+    if (key.includes(model) || model.includes(key.split('/')[1] ?? '')) return cost;
+  }
+  // No pricing available — return zero (OpenRouter fetch pending or model not listed)
+  return { promptCostPer1K: 0, completionCostPer1K: 0 };
+}
+
+// ---------------------------------------------------------------------------
+// TokenTracker
+// ---------------------------------------------------------------------------
+
 export class TokenTracker {
   private model: string;
   private usage: TokenUsage = {
@@ -42,15 +164,39 @@ export class TokenTracker {
     callCount: 0,
   };
 
-  constructor(model: string = 'gpt-4o') {
+  constructor(model: string = 'claude-sonnet-4-5-20250929') {
     this.model = model;
   }
 
-  addUsage(promptTokens: number, completionTokens: number): void {
+  addUsage(promptTokens: number, completionTokens: number, options?: { provider?: string; durationMs?: number }): void {
     this.usage.promptTokens += promptTokens;
     this.usage.completionTokens += completionTokens;
     this.usage.totalTokens += promptTokens + completionTokens;
     this.usage.callCount++;
+
+    const costs = getCostForModel(this.model);
+    const estimatedCost =
+      (promptTokens / 1000) * costs.promptCostPer1K +
+      (completionTokens / 1000) * costs.completionCostPer1K;
+
+    const entry: CallLogEntry = {
+      timestamp: new Date().toISOString(),
+      provider: options?.provider ?? 'unknown',
+      model: this.model,
+      promptTokens,
+      completionTokens,
+      totalTokens: promptTokens + completionTokens,
+      estimatedCost,
+      durationMs: options?.durationMs,
+      source: 'local-log',
+    };
+
+    try {
+      mkdirSync(dirname(CALL_LOG_PATH), { recursive: true });
+      appendFileSync(CALL_LOG_PATH, JSON.stringify(entry) + '\n');
+    } catch {
+      // Non-critical: don't break LLM calls if logging fails
+    }
   }
 
   getSummary(): TokenUsage {
@@ -58,7 +204,7 @@ export class TokenTracker {
   }
 
   getEstimatedCost(): number {
-    const costs = MODEL_COSTS[this.model] || MODEL_COSTS['gpt-4o'];
+    const costs = getCostForModel(this.model);
     const promptCost =
       (this.usage.promptTokens / 1000) * costs.promptCostPer1K;
     const completionCost =
@@ -114,3 +260,15 @@ export function getGlobalTokenTracker(model?: string): TokenTracker {
 export function resetGlobalTokenTracker(): void {
   globalTracker?.reset();
 }
+
+export function getCallLogPath(): string {
+  return CALL_LOG_PATH;
+}
+
+/** Force-refresh the pricing cache from OpenRouter. */
+export async function refreshPricing(): Promise<void> {
+  const models = await fetchPricingFromOpenRouter();
+  pricingCache = { fetchedAt: Date.now(), models };
+  mkdirSync(dirname(PRICING_CACHE_PATH), { recursive: true });
+  writeFileSync(PRICING_CACHE_PATH, JSON.stringify(pricingCache));
+}

package/dist/chunk-F2DMHMRH.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/client.ts"],"sourcesContent":["/**\n * Shared LLM Client\n *\n * Multi-provider LLM client with:\n * - OpenAI, DeepSeek, Anthropic, and Kimi support\n * - Anthropic prompt caching (CachingChatAnthropic)\n * - Rate limiting and retry logic\n * - Token tracking\n * - Structured output parsing with Zod\n *\n * @packageDocumentation\n */\n\nimport { ChatOpenAI } from '@langchain/openai';\nimport { ChatAnthropic } from '@langchain/anthropic';\nimport type { BaseMessageLike } from '@langchain/core/messages';\nimport Anthropic from '@anthropic-ai/sdk';\nimport { z } from 'zod';\nimport {\n  RateLimiter,\n  getGlobalRateLimiter,\n  type RateLimiterOptions,\n} from './rate-limiter.js';\nimport { TokenTracker, getGlobalTokenTracker } from './token-tracker.js';\nimport { parseJsonResponse } from './json-parser.js';\n\n// ============================================================================\n// Anthropic Cache Control Helper\n// ============================================================================\n\nfunction addCacheControlToSystemMessages(\n  messages: Array<{ role: string; content: string }>,\n): BaseMessageLike[] {\n  return messages.map((msg) => {\n    if (msg.role !== 'system') {\n      return msg as BaseMessageLike;\n    }\n\n    return {\n      role: msg.role,\n      content: [\n        {\n          type: 'text' as const,\n          text: msg.content,\n          cache_control: { type: 'ephemeral' },\n        },\n      ],\n    } as BaseMessageLike;\n  });\n}\n\ntype ChatModel = ChatOpenAI | ChatAnthropic;\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport type LLMProvider = 'openai' | 'deepseek' | 'anthropic' | 'kimi' | 'openrouter' | 'orbgen';\n\nexport interface ProviderConfig {\n  apiKey: string;\n  baseUrl?: string;\n  defaultModel: string;\n}\n\nexport interface LLMClientOptions {\n  provider?: LLMProvider;\n  model?: string;\n  temperature?: number;\n  streaming?: boolean;\n  rateLimiter?: RateLimiterOptions;\n  useGlobalRateLimiter?: boolean;\n  trackTokens?: boolean;\n}\n\nexport interface LLMCallOptions<T = unknown> {\n  systemPrompt: string;\n  userPrompt: string;\n  schema?: z.ZodSchema<T>;\n  maxRetries?: number;\n  retryWithContext?: boolean;\n  maxTokens?: number;\n  skipSchemaValidation?: boolean;\n  temperature?: number;\n}\n\nexport interface CacheableBlock {\n  type: 'text';\n  text: string;\n  cache_control?: { type: 'ephemeral' };\n}\n\nexport interface CacheAwareLLMCallOptions<T = unknown>\n  extends LLMCallOptions<T> {\n  systemBlocks?: CacheableBlock[];\n  userBlocks?: CacheableBlock[];\n  rawText?: boolean;\n}\n\nexport interface LLMUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n}\n\nexport type LLMFinishReason =\n  | 'stop'\n  | 'length'\n  | 'content_filter'\n  | 'tool_calls'\n  | null;\n\nexport interface LLMResponse<T> {\n  data: T;\n  raw: string;\n  finishReason: LLMFinishReason;\n  usage: LLMUsage | null;\n}\n\nexport interface LLMStreamOptions {\n  systemPrompt: string;\n  messages: Array<{ role: 'system' | 'user' | 'assistant'; content: string }>;\n  maxTokens?: number;\n  temperature?: number;\n}\n\nexport interface LLMStreamChunk {\n  content: string;\n  done: boolean;\n}\n\n// ============================================================================\n// Provider Configuration\n// ============================================================================\n\nconst PROVIDER_CONFIGS: Record<LLMProvider, () => ProviderConfig> = {\n  openai: () => {\n    const apiKey = process.env.OPENAI_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'OPENAI_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return { apiKey, baseUrl: undefined, defaultModel: 'gpt-4o' };\n  },\n  deepseek: () => {\n    const apiKey = process.env.DEEPSEEK_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'DEEPSEEK_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return {\n      apiKey,\n      baseUrl: 'https://api.deepseek.com/v1',\n      defaultModel: 'deepseek-chat',\n    };\n  },\n  anthropic: () => {\n    const apiKey = process.env.ANTHROPIC_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'ANTHROPIC_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return {\n      apiKey,\n      baseUrl: undefined,\n      defaultModel: 'claude-sonnet-4-5-20250929',\n    };\n  },\n  kimi: () => {\n    const apiKey = process.env.KIMI_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'KIMI_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return {\n      apiKey,\n      baseUrl: 'https://api.moonshot.ai/v1',\n      defaultModel: 'kimi-k2.5',\n    };\n  },\n  openrouter: () => {\n    const apiKey = process.env.OPEN_ROUTER_API_KEY;\n    if (!apiKey) {\n      throw new Error(\n        'OPEN_ROUTER_API_KEY environment variable is not set. ' +\n          'Please set it in your .env file or environment.',\n      );\n    }\n    return {\n      apiKey,\n      baseUrl: 'https://openrouter.ai/api/v1',\n      defaultModel: 'qwen/qwen-2.5-72b-instruct', // Default to Qwen 2.5\n    };\n  },\n  orbgen: () => {\n    const baseUrl = process.env.ORBGEN_URL;\n    if (!baseUrl) {\n      throw new Error(\n        'ORBGEN_URL environment variable is not set. ' +\n          'Set it to the OrbGen Cloud Run URL (e.g., https://orbgen-v2-xxx.run.app)',\n      );\n    }\n    return {\n      apiKey: 'not-needed',\n      baseUrl: `${baseUrl}/v1`,\n      defaultModel: 'orbgen-v2',\n    };\n  },\n};\n\nexport const DEEPSEEK_MODELS = {\n  CHAT: 'deepseek-chat',\n  CODER: 'deepseek-coder',\n  REASONER: 'deepseek-reasoner',\n} as const;\n\nexport const OPENAI_MODELS = {\n  GPT4O: 'gpt-4o',\n  GPT4O_MINI: 'gpt-4o-mini',\n  GPT4_TURBO: 'gpt-4-turbo',\n  GPT35_TURBO: 'gpt-3.5-turbo',\n  GPT_5_1: 'gpt-5.1',\n} as const;\n\nexport const ANTHROPIC_MODELS = {\n  CLAUDE_SONNET_4_5: 'claude-sonnet-4-5-20250929',\n  CLAUDE_SONNET_4: 'claude-sonnet-4-20250514',\n  CLAUDE_OPUS_4_5: 'claude-opus-4-5-20250929',\n  CLAUDE_3_5_HAIKU: 'claude-3-5-haiku-20241022',\n} as const;\n\nexport const KIMI_MODELS = {\n  K2_5: 'kimi-k2.5',\n} as const;\n\nexport const OPENROUTER_MODELS = {\n  // Qwen models - JSON/structured data specialists\n  QWEN_2_5_72B: 'qwen/qwen-2.5-72b-instruct',\n  QWEN_2_5_CODER_32B: 'qwen/qwen-2.5-coder-32b-instruct',\n  QWEN_3_235B: 'qwen/qwen3-235b-a22b',\n\n  // Gemma models - best small models for structured JSON output\n  // Gemma 3 4B: 6/6 on complex decomposition, 100% behavior matching, free, fastest\n  GEMMA_3_4B: 'google/gemma-3-4b-it',\n  GEMMA_3_12B: 'google/gemma-3-12b-it',\n  GEMMA_3_27B: 'google/gemma-3-27b-it',\n\n  // Mistral models - strong structured output, function calling\n  MINISTRAL_8B: 'mistralai/ministral-8b-2512',\n  // Mistral Small 3.1: 6/6 on complex decomposition, picked std-kanban for tasks\n  MISTRAL_SMALL_3_1: 'mistralai/mistral-small-3.1-24b-instruct',\n  // Mistral Medium 3.1: next tier up from Small, stronger reasoning, tool calling\n  MISTRAL_MEDIUM_3_1: 'mistralai/mistral-medium-3.1',\n\n  // Llama models - agentic workhorses\n  LLAMA_3_3_70B: 'meta-llama/llama-3.3-70b-instruct',\n  LLAMA_3_1_405B: 'meta-llama/llama-3.1-405b-instruct',\n  LLAMA_4_MAVERICK: 'meta-llama/llama-4-maverick',\n  LLAMA_4_SCOUT: 'meta-llama/llama-4-scout',\n\n  // Kimi models - strong reasoning\n  KIMI_K2: 'moonshotai/kimi-k2',\n\n  // Zhipu GLM models - via OpenRouter\n  GLM_4_7: 'z-ai/glm-4.7',\n} as const;\n\nconst DEFAULT_TEMPERATURE = 0.3;\n\n// ============================================================================\n// LLM Client\n// ============================================================================\n\nexport class LLMClient {\n  private model: ChatModel;\n  private rateLimiter: RateLimiter;\n  private tokenTracker: TokenTracker | null;\n  private modelName: string;\n  private provider: LLMProvider;\n  private providerConfig: ProviderConfig;\n  private temperature: number;\n  private streaming: boolean;\n\n  constructor(options: LLMClientOptions = {}) {\n    this.provider = options.provider || 'openai';\n    // Kimi: 0.6 when thinking disabled (our default), 1.0 when thinking enabled\n    this.temperature = options.temperature ?? \n      (this.provider === 'kimi' ? 0.6 : DEFAULT_TEMPERATURE);\n    this.streaming = options.streaming ?? false;\n\n    this.providerConfig = PROVIDER_CONFIGS[this.provider]();\n    this.modelName = options.model || this.providerConfig.defaultModel;\n\n    const keyPreview = this.providerConfig.apiKey.slice(-4);\n    console.log(\n      `[LLMClient] Provider: ${this.provider}, Model: ${this.modelName}, Key: ****${keyPreview}`,\n    );\n    if (this.providerConfig.baseUrl) {\n      console.log(\n        `[LLMClient] Using custom base URL: ${this.providerConfig.baseUrl}`,\n      );\n    }\n\n    this.model = this.createModel();\n\n    this.rateLimiter =\n      options.useGlobalRateLimiter !== false\n        ? getGlobalRateLimiter(options.rateLimiter)\n        : new RateLimiter(options.rateLimiter);\n\n    this.tokenTracker =\n      options.trackTokens !== false\n        ? getGlobalTokenTracker(this.modelName)\n        : null;\n  }\n\n  private usesMaxCompletionTokens(): boolean {\n    const model = this.modelName.toLowerCase();\n    return (\n      model.startsWith('o1') ||\n      model.startsWith('gpt-5') ||\n      model.includes('o1-') ||\n      model.includes('o3')\n    );\n  }\n\n  private createModel(options?: {\n    maxTokens?: number;\n    temperature?: number;\n  }): ChatModel {\n    const maxTokens = options?.maxTokens;\n    const temperature = options?.temperature ?? this.temperature;\n\n    if (this.provider === 'anthropic') {\n      return new ChatAnthropic({\n        apiKey: this.providerConfig.apiKey,\n        model: this.modelName,\n        temperature,\n        streaming: this.streaming,\n        maxTokens: maxTokens || 8192,\n        callbacks: [\n          {\n            handleLLMEnd: (output) => {\n              const generation = output.generations?.[0]?.[0];\n              const usage = (\n                generation as unknown as {\n                  message?: {\n                    usage_metadata?: {\n                      cache_creation_input_tokens?: number;\n                      cache_read_input_tokens?: number;\n                      input_tokens?: number;\n                      output_tokens?: number;\n                    };\n                  };\n                }\n              )?.message?.usage_metadata;\n\n              if (usage) {\n                const cacheCreated = usage.cache_creation_input_tokens ?? 0;\n                const cacheRead = usage.cache_read_input_tokens ?? 0;\n                const inputTokens = usage.input_tokens ?? 0;\n                const outputTokens = usage.output_tokens ?? 0;\n\n                if (cacheCreated > 0) {\n                  console.log(\n                    `[LLMClient:Anthropic] Cache WRITE: ${cacheCreated} tokens cached`,\n                  );\n                }\n                if (cacheRead > 0) {\n                  const savingsPercent = Math.round(\n                    (cacheRead / (cacheRead + inputTokens)) * 100,\n                  );\n                  console.log(\n                    `[LLMClient:Anthropic] Cache HIT: ${cacheRead} tokens (~${savingsPercent}% of prompt)`,\n                  );\n                }\n                if (cacheCreated === 0 && cacheRead === 0 && inputTokens > 0) {\n                  if (inputTokens < 500) {\n                    console.log(\n                      `[LLMClient:Anthropic] ${inputTokens} input, ${outputTokens} output tokens (likely cached)`,\n                    );\n                  } else {\n                    console.log(\n                      `[LLMClient:Anthropic] ${inputTokens} input, ${outputTokens} output tokens`,\n                    );\n                  }\n                }\n              }\n            },\n          },\n        ],\n      });\n    }\n\n    const useCompletionTokens = this.usesMaxCompletionTokens();\n\n    const tokenConfig = maxTokens\n      ? useCompletionTokens\n        ? { modelKwargs: { max_completion_tokens: maxTokens } }\n        : { maxTokens }\n      : {};\n\n    const timeout = this.provider === 'deepseek' ? 600000 : undefined;\n\n    // Kimi-k2.5: disable thinking to avoid reasoning_content issues with tool calls\n    // When thinking is disabled, temperature must be 0.6 (not 1.0)\n    const isKimi = this.provider === 'kimi';\n    const effectiveTemp = isKimi ? 0.6 : temperature;\n\n    // Build modelKwargs incrementally to avoid spread conflicts\n    const modelKwargs: Record<string, unknown> = {};\n    if (useCompletionTokens && maxTokens) {\n      modelKwargs.max_completion_tokens = maxTokens;\n    }\n    if (isKimi) {\n      modelKwargs.thinking = { type: 'disabled' };\n    }\n    // OpenRouter (Qwen): explicit tool_choice so the model doesn't ignore tool definitions\n    if (this.provider === 'openrouter') {\n      modelKwargs.tool_choice = 'auto';\n    }\n\n    return new ChatOpenAI({\n      apiKey: this.providerConfig.apiKey,\n      model: this.modelName,\n      temperature: useCompletionTokens ? undefined : effectiveTemp,\n      streaming: this.streaming,\n      timeout,\n      ...(Object.keys(modelKwargs).length > 0 ? { modelKwargs } : {}),\n      ...(useCompletionTokens ? {} : maxTokens ? { maxTokens } : {}),\n      configuration: {\n        apiKey: this.providerConfig.apiKey,\n        ...(this.providerConfig.baseUrl\n          ? { baseURL: this.providerConfig.baseUrl }\n          : {}),\n      },\n    });\n  }\n\n  private getModelWithOptions(options: {\n    maxTokens?: number;\n    temperature?: number;\n  }): ChatModel {\n    return this.createModel(options);\n  }\n\n  /**\n   * Check if this model is a Qwen3.5 thinking model.\n   * These models burn all output tokens on internal reasoning\n   * unless thinking is explicitly disabled via /no_think prefix.\n   */\n  private isQwenThinkingModel(): boolean {\n    return this.modelName.includes('qwen3.5');\n  }\n\n  /**\n   * Prepare user prompt with provider-specific adjustments.\n   * Qwen3.5 models require /no_think to disable reasoning mode.\n   */\n  private prepareUserPrompt(prompt: string): string {\n    if (this.isQwenThinkingModel()) {\n      return `/no_think\\n${prompt}`;\n    }\n    return prompt;\n  }\n\n  getProvider(): LLMProvider {\n    return this.provider;\n  }\n\n  getModelName(): string {\n    return this.modelName;\n  }\n\n  getModel(): ChatModel {\n    return this.model;\n  }\n\n  getRateLimiterStatus() {\n    return this.rateLimiter.getStatus();\n  }\n\n  getTokenUsage() {\n    return this.tokenTracker?.getSummary() ?? null;\n  }\n\n  async call<T>(options: LLMCallOptions<T>): Promise<T> {\n    const response = await this.callWithMetadata(options);\n    return response.data;\n  }\n\n  async callWithMetadata<T>(options: LLMCallOptions<T>): Promise<LLMResponse<T>> {\n    const {\n      systemPrompt,\n      userPrompt,\n      schema,\n      maxRetries = 2,\n      retryWithContext = true,\n      maxTokens,\n      skipSchemaValidation = false,\n      temperature,\n    } = options;\n\n    let currentPrompt = userPrompt;\n    let lastError: Error | null = null;\n\n    console.log(\n      `[LLMClient:call] Starting call to ${this.provider}/${this.modelName}`,\n    );\n    console.log(`[LLMClient:call] Prompt length: ${userPrompt.length} chars`);\n    if (maxTokens) {\n      console.log(`[LLMClient:call] Max tokens: ${maxTokens}`);\n    }\n\n    for (let attempt = 0; attempt <= maxRetries; attempt++) {\n      try {\n        console.log(\n          `[LLMClient:call] Attempt ${attempt + 1}/${maxRetries + 1}...`,\n        );\n        const attemptStartTime = Date.now();\n\n        const result = await this.rateLimiter.execute(async () => {\n          console.log(`[LLMClient:call] Invoking model...`);\n          const invokeStartTime = Date.now();\n\n          const modelToUse =\n            maxTokens || temperature !== undefined\n              ? this.getModelWithOptions({ maxTokens, temperature })\n              : this.model;\n\n          const messages = [\n            { role: 'system', content: systemPrompt },\n            { role: 'user', content: this.prepareUserPrompt(currentPrompt) },\n          ];\n          const response = await modelToUse.invoke(\n            this.provider === 'anthropic'\n              ? addCacheControlToSystemMessages(messages)\n              : messages,\n          );\n\n          console.log(\n            `[LLMClient:call] Model responded in ${Date.now() - invokeStartTime}ms`,\n          );\n\n          let usage: LLMUsage | null = null;\n          if (response.usage_metadata) {\n            const usageMeta = response.usage_metadata as {\n              input_tokens?: number;\n              output_tokens?: number;\n            };\n            usage = {\n              promptTokens: usageMeta.input_tokens || 0,\n              completionTokens: usageMeta.output_tokens || 0,\n              totalTokens:\n                (usageMeta.input_tokens || 0) +\n                (usageMeta.output_tokens || 0),\n            };\n            console.log(\n              `[LLMClient:call] Tokens used: ${usage.promptTokens} in, ${usage.completionTokens} out`,\n            );\n\n            if (this.tokenTracker) {\n              this.tokenTracker.addUsage(\n                usage.promptTokens,\n                usage.completionTokens,\n              );\n            }\n          }\n\n          const finishReason = this.extractFinishReason(response);\n          if (finishReason === 'length') {\n            console.warn(\n              `[LLMClient:call] Response truncated (finish_reason=length)`,\n            );\n          }\n\n          const content =\n            typeof response.content === 'string'\n              ? response.content\n              : JSON.stringify(response.content);\n\n          console.log(\n            `[LLMClient:call] Response length: ${content.length} chars, finish_reason: ${finishReason}`,\n          );\n\n          return { content, finishReason, usage };\n        });\n\n        console.log(\n          `[LLMClient:call] Attempt ${attempt + 1} completed in ${Date.now() - attemptStartTime}ms, parsing response...`,\n        );\n\n        const parsed = skipSchemaValidation\n          ? (parseJsonResponse(result.content, undefined) as T)\n          : parseJsonResponse(result.content, schema);\n        console.log(\n          `[LLMClient:call] Response parsed successfully${skipSchemaValidation ? ' (schema validation skipped)' : ''}`,\n        );\n\n        return {\n          data: parsed,\n          raw: result.content,\n          finishReason: result.finishReason,\n          usage: result.usage,\n        };\n      } catch (error) {\n        lastError = error instanceof Error ? error : new Error(String(error));\n        console.error(\n          `[LLMClient:call] Attempt ${attempt + 1} failed:`,\n          lastError.message,\n        );\n\n        if (this.isRateLimitError(lastError)) {\n          console.error(`[LLMClient:call] Rate limit error, not retrying`);\n          throw lastError;\n        }\n\n        if (attempt < maxRetries && retryWithContext) {\n          console.log(`[LLMClient:call] Will retry with error context`);\n          currentPrompt =\n            `${userPrompt}\\n\\n` +\n            `[Previous attempt failed with: ${lastError.message}]\\n` +\n            `Please output valid JSON that matches the expected schema.`;\n        }\n      }\n    }\n\n    console.error(`[LLMClient:call] All attempts exhausted, throwing error`);\n    throw lastError;\n  }\n\n  private extractFinishReason(\n    response: Awaited<ReturnType<ChatOpenAI['invoke']>>,\n  ): LLMFinishReason {\n    const metadata = response.response_metadata as\n      | Record<string, unknown>\n      | undefined;\n    if (metadata?.finish_reason) {\n      const reason = metadata.finish_reason as string;\n      if (\n        reason === 'stop' ||\n        reason === 'length' ||\n        reason === 'content_filter' ||\n        reason === 'tool_calls'\n      ) {\n        return reason;\n      }\n    }\n    return null;\n  }\n\n  async callRaw(options: {\n    systemPrompt: string;\n    userPrompt: string;\n    maxTokens?: number;\n  }): Promise<string> {\n    const response = await this.callRawWithMetadata(options);\n    return response.raw;\n  }\n\n  async callRawWithMetadata(options: {\n    systemPrompt: string;\n    userPrompt: string;\n    maxTokens?: number;\n  }): Promise<Omit<LLMResponse<string>, 'data'> & { raw: string }> {\n    const { systemPrompt, userPrompt, maxTokens } = options;\n\n    return this.rateLimiter.execute(async () => {\n      const modelToUse = maxTokens\n        ? this.getModelWithOptions({ maxTokens })\n        : this.model;\n\n      const messages = [\n        { role: 'system', content: systemPrompt },\n        { role: 'user', content: this.prepareUserPrompt(userPrompt) },\n      ];\n      const response = await modelToUse.invoke(\n        this.provider === 'anthropic'\n          ? addCacheControlToSystemMessages(messages)\n          : messages,\n      );\n\n      let usage: LLMUsage | null = null;\n      if (response.usage_metadata) {\n        const usageMeta = response.usage_metadata as {\n          input_tokens?: number;\n          output_tokens?: number;\n        };\n        usage = {\n          promptTokens: usageMeta.input_tokens || 0,\n          completionTokens: usageMeta.output_tokens || 0,\n          totalTokens:\n            (usageMeta.input_tokens || 0) + (usageMeta.output_tokens || 0),\n        };\n\n        if (this.tokenTracker) {\n          this.tokenTracker.addUsage(\n            usage.promptTokens,\n            usage.completionTokens,\n          );\n        }\n      }\n\n      const finishReason = this.extractFinishReason(response);\n      const content =\n        typeof response.content === 'string'\n          ? response.content\n          : JSON.stringify(response.content);\n\n      return { raw: content, finishReason, usage };\n    });\n  }\n\n  /**\n   * Stream a raw text response as an async iterator of content chunks.\n   * Uses the underlying LangChain model's .stream() method.\n   *\n   * @param options - System prompt plus full message history\n   * @yields LLMStreamChunk with content deltas and a done flag\n   */\n  async *streamRaw(options: LLMStreamOptions): AsyncGenerator<LLMStreamChunk> {\n    const { messages, maxTokens, temperature } = options;\n\n    const modelToUse = (maxTokens || temperature !== undefined)\n      ? this.getModelWithOptions({ maxTokens, temperature })\n      : this.model;\n\n    const langchainMessages = this.provider === 'anthropic'\n      ? addCacheControlToSystemMessages(messages)\n      : messages;\n\n    const stream = await modelToUse.stream(langchainMessages);\n\n    for await (const chunk of stream) {\n      const content = typeof chunk.content === 'string'\n        ? chunk.content\n        : Array.isArray(chunk.content)\n          ? chunk.content\n              .filter((c): c is { type: 'text'; text: string } => typeof c === 'object' && c !== null && 'text' in c)\n              .map((c) => c.text)\n              .join('')\n          : '';\n\n      if (content) {\n        yield { content, done: false };\n      }\n    }\n\n    yield { content: '', done: true };\n  }\n\n  private isRateLimitError(error: Error): boolean {\n    const message = error.message.toLowerCase();\n    return (\n      message.includes('rate limit') ||\n      message.includes('429') ||\n      message.includes('quota exceeded')\n    );\n  }\n\n  // ==========================================================================\n  // Anthropic Cache Control Support\n  // ==========================================================================\n\n  async callWithCache<T>(\n    options: CacheAwareLLMCallOptions<T>,\n  ): Promise<LLMResponse<T>> {\n    const {\n      systemPrompt,\n      userPrompt,\n      systemBlocks,\n      userBlocks,\n      schema,\n      maxRetries = 2,\n      maxTokens,\n      skipSchemaValidation = false,\n      temperature,\n      rawText = false,\n    } = options;\n\n    if (this.provider !== 'anthropic') {\n      console.log(\n        `[LLMClient:callWithCache] Provider ${this.provider} doesn't support caching, using regular call`,\n      );\n      return this.callWithMetadata(options);\n    }\n\n    const cacheableCount =\n      (systemBlocks || []).filter((b) => b.cache_control).length +\n      (userBlocks || []).filter((b) => b.cache_control).length;\n    console.log(\n      `[LLMClient:callWithCache] ${cacheableCount} cacheable block(s)`,\n    );\n\n    let lastError: Error | null = null;\n\n    for (let attempt = 0; attempt <= maxRetries; attempt++) {\n      try {\n        console.log(\n          `[LLMClient:callWithCache] Attempt ${attempt + 1}/${maxRetries + 1}...`,\n        );\n\n        const result = await this.rateLimiter.execute(async () => {\n          const anthropic = new Anthropic();\n\n          const systemContent =\n            systemBlocks && systemBlocks.length > 0\n              ? systemBlocks.map((b) => ({\n                  type: 'text' as const,\n                  text: b.text,\n                  ...(b.cache_control\n                    ? { cache_control: b.cache_control }\n                    : {}),\n                }))\n              : systemPrompt\n                ? [{ type: 'text' as const, text: systemPrompt }]\n                : [];\n\n          const userContent =\n            userBlocks && userBlocks.length > 0\n              ? userBlocks.map((b) => ({\n                  type: 'text' as const,\n                  text: b.text,\n                  ...(b.cache_control\n                    ? { cache_control: b.cache_control }\n                    : {}),\n                }))\n              : userPrompt\n                ? [{ type: 'text' as const, text: userPrompt }]\n                : [];\n\n          const response = await anthropic.messages.create({\n            model: this.modelName,\n            max_tokens: maxTokens || 8192,\n            temperature: temperature ?? 0,\n            system: systemContent,\n            messages: [{ role: 'user', content: userContent }],\n          });\n\n          const textContent = response.content.find((c) => c.type === 'text');\n          const content =\n            textContent && 'text' in textContent ? textContent.text : '';\n\n          const apiUsage = response.usage as {\n            input_tokens: number;\n            output_tokens: number;\n            cache_creation_input_tokens?: number;\n            cache_read_input_tokens?: number;\n          };\n\n          const cacheRead = apiUsage.cache_read_input_tokens || 0;\n          const cacheCreation = apiUsage.cache_creation_input_tokens || 0;\n\n          if (cacheCreation > 0) {\n            console.log(\n              `[LLMClient:callWithCache] Cache WRITE: ${cacheCreation} tokens`,\n            );\n          }\n          if (cacheRead > 0) {\n            const savingsPercent = Math.round(\n              (cacheRead / (cacheRead + apiUsage.input_tokens)) * 100,\n            );\n            console.log(\n              `[LLMClient:callWithCache] Cache HIT: ${cacheRead} tokens (~${savingsPercent}% of prompt)`,\n            );\n          }\n          if (cacheCreation === 0 && cacheRead === 0) {\n            console.log(\n              `[LLMClient:callWithCache] No caching: ${apiUsage.input_tokens} input tokens`,\n            );\n          }\n\n          const usage: LLMUsage = {\n            promptTokens: apiUsage.input_tokens,\n            completionTokens: apiUsage.output_tokens,\n            totalTokens: apiUsage.input_tokens + apiUsage.output_tokens,\n          };\n\n          if (this.tokenTracker) {\n            this.tokenTracker.addUsage(\n              usage.promptTokens,\n              usage.completionTokens,\n            );\n          }\n\n          const finishReason =\n            response.stop_reason === 'end_turn'\n              ? 'stop'\n              : response.stop_reason;\n\n          return {\n            content,\n            finishReason: finishReason as LLMFinishReason,\n            usage,\n          };\n        });\n\n        let parsed: T;\n        if (rawText) {\n          parsed = result.content as unknown as T;\n        } else if (skipSchemaValidation) {\n          parsed = parseJsonResponse(result.content, undefined) as T;\n        } else {\n          parsed = parseJsonResponse(result.content, schema);\n        }\n\n        return {\n          data: parsed,\n          raw: result.content,\n          finishReason: result.finishReason,\n          usage: result.usage,\n        };\n      } catch (error) {\n        lastError = error instanceof Error ? error : new Error(String(error));\n        console.error(\n          `[LLMClient:callWithCache] Attempt ${attempt + 1} failed:`,\n          lastError.message,\n        );\n\n        if (this.isRateLimitError(lastError)) {\n          throw lastError;\n        }\n      }\n    }\n\n    throw lastError;\n  }\n\n  static cacheableBlock(text: string, cache = true): CacheableBlock {\n    return cache\n      ? { type: 'text', text, cache_control: { type: 'ephemeral' } }\n      : { type: 'text', text };\n  }\n}\n\n// ============================================================================\n// Singleton Instances\n// ============================================================================\n\nconst sharedClients: Partial<Record<LLMProvider, LLMClient>> = {};\n\nexport function getSharedLLMClient(options?: LLMClientOptions): LLMClient {\n  const provider = options?.provider || 'openai';\n  if (!sharedClients[provider]) {\n    sharedClients[provider] = new LLMClient(options);\n  }\n  return sharedClients[provider]!;\n}\n\nexport function resetSharedLLMClient(provider?: LLMProvider): void {\n  if (provider) {\n    delete sharedClients[provider];\n  } else {\n    for (const key of Object.keys(sharedClients) as LLMProvider[]) {\n      delete sharedClients[key];\n    }\n  }\n}\n\n// ============================================================================\n// Provider Detection\n// ============================================================================\n\nexport function getAvailableProvider(): LLMProvider {\n  if (process.env.ANTHROPIC_API_KEY) return 'anthropic';\n  if (process.env.DEEPSEEK_API_KEY) return 'deepseek';\n  if (process.env.KIMI_API_KEY) return 'kimi';\n  if (process.env.OPENAI_API_KEY) return 'openai';\n  throw new Error(\n    'No LLM API key found. Please set ANTHROPIC_API_KEY, OPENAI_API_KEY, DEEPSEEK_API_KEY, or KIMI_API_KEY.',\n  );\n}\n\nexport function isProviderAvailable(provider: LLMProvider): boolean {\n  switch (provider) {\n    case 'openai':\n      return !!process.env.OPENAI_API_KEY;\n    case 'deepseek':\n      return !!process.env.DEEPSEEK_API_KEY;\n    case 'anthropic':\n      return !!process.env.ANTHROPIC_API_KEY;\n    case 'kimi':\n      return !!process.env.KIMI_API_KEY;\n    case 'openrouter':\n      return !!process.env.OPEN_ROUTER_API_KEY;\n    case 'orbgen':\n      return !!process.env.ORBGEN_URL;\n    default:\n      return false;\n  }\n}\n\n// ============================================================================\n// Convenience Functions\n// ============================================================================\n\n/**\n * Create an LLM client optimized for requirements analysis.\n *\n * Uses lower temperature (0.3) for more deterministic output.\n * Defaults to GPT-5.1 for OpenAI or DeepSeek Chat.\n *\n * @param {Partial<LLMClientOptions>} [options] - Optional client configuration\n * @returns {LLMClient} Configured LLM client\n */\nexport function createRequirementsClient(\n  options?: Partial<LLMClientOptions>,\n): LLMClient {\n  const provider = options?.provider || getAvailableProvider();\n  const defaultModel =\n    provider === 'deepseek' ? DEEPSEEK_MODELS.CHAT : OPENAI_MODELS.GPT_5_1;\n  return new LLMClient({\n    provider,\n    model: defaultModel,\n    temperature: 0.3,\n    ...options,\n  });\n}\n\n/**\n * Create an LLM client optimized for creative tasks.\n *\n * Uses higher temperature (0.7) for more varied output.\n * Defaults to GPT-4o or DeepSeek Reasoner.\n *\n * @param {Partial<LLMClientOptions>} [options] - Optional client configuration\n * @returns {LLMClient} Configured LLM client\n */\nexport function createCreativeClient(\n  options?: Partial<LLMClientOptions>,\n): LLMClient {\n  const provider = options?.provider || getAvailableProvider();\n  const defaultModel =\n    provider === 'deepseek' ? DEEPSEEK_MODELS.REASONER : OPENAI_MODELS.GPT4O;\n  return new LLMClient({\n    provider,\n    model: defaultModel,\n    temperature: 0.7,\n    ...options,\n  });\n}\n\n/**\n * Create an LLM client optimized for code fixing.\n *\n * Uses low temperature (0.2) for precise, deterministic fixes.\n * Defaults to GPT-4o Mini or DeepSeek Chat for cost efficiency.\n *\n * @param {Partial<LLMClientOptions>} [options] - Optional client configuration\n * @returns {LLMClient} Configured LLM client\n */\nexport function createFixClient(\n  options?: Partial<LLMClientOptions>,\n): LLMClient {\n  const provider = options?.provider || getAvailableProvider();\n  const defaultModel =\n    provider === 'deepseek'\n      ? DEEPSEEK_MODELS.CHAT\n      : OPENAI_MODELS.GPT4O_MINI;\n  return new LLMClient({\n    provider,\n    model: defaultModel,\n    temperature: 0.2,\n    ...options,\n  });\n}\n\n/**\n * Create a DeepSeek LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured DeepSeek client\n */\nexport function createDeepSeekClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'deepseek',\n    model: DEEPSEEK_MODELS.CHAT,\n    ...options,\n  });\n}\n\n/**\n * Create an OpenAI LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured OpenAI client\n */\nexport function createOpenAIClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'openai',\n    model: OPENAI_MODELS.GPT4O,\n    ...options,\n  });\n}\n\n/**\n * Create an Anthropic LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured Anthropic client\n */\nexport function createAnthropicClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'anthropic',\n    model: ANTHROPIC_MODELS.CLAUDE_SONNET_4_5,\n    ...options,\n  });\n}\n\n/**\n * Create a Kimi LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured Kimi client\n */\nexport function createKimiClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'kimi',\n    model: KIMI_MODELS.K2_5,\n    ...options,\n  });\n}\n\n/**\n * Create an OpenRouter LLM client.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured OpenRouter client\n */\nexport function createOpenRouterClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'openrouter',\n    model: OPENROUTER_MODELS.QWEN_2_5_72B,\n    ...options,\n  });\n}\n\n/**\n * Create a Zhipu (GLM) LLM client via OpenRouter.\n *\n * @param {Partial<Omit<LLMClientOptions, 'provider'>>} [options] - Optional client configuration\n * @returns {LLMClient} Configured Zhipu client\n */\nexport function createZhipuClient(\n  options?: Partial<Omit<LLMClientOptions, 'provider'>>,\n): LLMClient {\n  return new LLMClient({\n    provider: 'openrouter',\n    model: OPENROUTER_MODELS.GLM_4_7,\n    ...options,\n  });\n}\n"],"mappings":";;;;;;;;;;AAaA,SAAS,kBAAkB;AAC3B,SAAS,qBAAqB;AAE9B,OAAO,eAAe;AActB,SAAS,gCACP,UACmB;AACnB,SAAO,SAAS,IAAI,CAAC,QAAQ;AAC3B,QAAI,IAAI,SAAS,UAAU;AACzB,aAAO;AAAA,IACT;AAEA,WAAO;AAAA,MACL,MAAM,IAAI;AAAA,MACV,SAAS;AAAA,QACP;AAAA,UACE,MAAM;AAAA,UACN,MAAM,IAAI;AAAA,UACV,eAAe,EAAE,MAAM,YAAY;AAAA,QACrC;AAAA,MACF;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAsFA,IAAM,mBAA8D;AAAA,EAClE,QAAQ,MAAM;AACZ,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO,EAAE,QAAQ,SAAS,QAAW,cAAc,SAAS;AAAA,EAC9D;AAAA,EACA,UAAU,MAAM;AACd,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL;AAAA,MACA,SAAS;AAAA,MACT,cAAc;AAAA,IAChB;AAAA,EACF;AAAA,EACA,WAAW,MAAM;AACf,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL;AAAA,MACA,SAAS;AAAA,MACT,cAAc;AAAA,IAChB;AAAA,EACF;AAAA,EACA,MAAM,MAAM;AACV,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL;AAAA,MACA,SAAS;AAAA,MACT,cAAc;AAAA,IAChB;AAAA,EACF;AAAA,EACA,YAAY,MAAM;AAChB,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL;AAAA,MACA,SAAS;AAAA,MACT,cAAc;AAAA;AAAA,IAChB;AAAA,EACF;AAAA,EACA,QAAQ,MAAM;AACZ,UAAM,UAAU,QAAQ,IAAI;AAC5B,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI;AAAA,QACR;AAAA,MAEF;AAAA,IACF;AACA,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,SAAS,GAAG,OAAO;AAAA,MACnB,cAAc;AAAA,IAChB;AAAA,EACF;AACF;AAEO,IAAM,kBAAkB;AAAA,EAC7B,MAAM;AAAA,EACN,OAAO;AAAA,EACP,UAAU;AACZ;AAEO,IAAM,gBAAgB;AAAA,EAC3B,OAAO;AAAA,EACP,YAAY;AAAA,EACZ,YAAY;AAAA,EACZ,aAAa;AAAA,EACb,SAAS;AACX;AAEO,IAAM,mBAAmB;AAAA,EAC9B,mBAAmB;AAAA,EACnB,iBAAiB;AAAA,EACjB,iBAAiB;AAAA,EACjB,kBAAkB;AACpB;AAEO,IAAM,cAAc;AAAA,EACzB,MAAM;AACR;AAEO,IAAM,oBAAoB;AAAA;AAAA,EAE/B,cAAc;AAAA,EACd,oBAAoB;AAAA,EACpB,aAAa;AAAA;AAAA;AAAA,EAIb,YAAY;AAAA,EACZ,aAAa;AAAA,EACb,aAAa;AAAA;AAAA,EAGb,cAAc;AAAA;AAAA,EAEd,mBAAmB;AAAA;AAAA,EAEnB,oBAAoB;AAAA;AAAA,EAGpB,eAAe;AAAA,EACf,gBAAgB;AAAA,EAChB,kBAAkB;AAAA,EAClB,eAAe;AAAA;AAAA,EAGf,SAAS;AAAA;AAAA,EAGT,SAAS;AACX;AAEA,IAAM,sBAAsB;AAMrB,IAAM,YAAN,MAAgB;AAAA,EAUrB,YAAY,UAA4B,CAAC,GAAG;AAC1C,SAAK,WAAW,QAAQ,YAAY;AAEpC,SAAK,cAAc,QAAQ,gBACxB,KAAK,aAAa,SAAS,MAAM;AACpC,SAAK,YAAY,QAAQ,aAAa;AAEtC,SAAK,iBAAiB,iBAAiB,KAAK,QAAQ,EAAE;AACtD,SAAK,YAAY,QAAQ,SAAS,KAAK,eAAe;AAEtD,UAAM,aAAa,KAAK,eAAe,OAAO,MAAM,EAAE;AACtD,YAAQ;AAAA,MACN,yBAAyB,KAAK,QAAQ,YAAY,KAAK,SAAS,cAAc,UAAU;AAAA,IAC1F;AACA,QAAI,KAAK,eAAe,SAAS;AAC/B,cAAQ;AAAA,QACN,sCAAsC,KAAK,eAAe,OAAO;AAAA,MACnE;AAAA,IACF;AAEA,SAAK,QAAQ,KAAK,YAAY;AAE9B,SAAK,cACH,QAAQ,yBAAyB,QAC7B,qBAAqB,QAAQ,WAAW,IACxC,IAAI,YAAY,QAAQ,WAAW;AAEzC,SAAK,eACH,QAAQ,gBAAgB,QACpB,sBAAsB,KAAK,SAAS,IACpC;AAAA,EACR;AAAA,EAEQ,0BAAmC;AACzC,UAAM,QAAQ,KAAK,UAAU,YAAY;AACzC,WACE,MAAM,WAAW,IAAI,KACrB,MAAM,WAAW,OAAO,KACxB,MAAM,SAAS,KAAK,KACpB,MAAM,SAAS,IAAI;AAAA,EAEvB;AAAA,EAEQ,YAAY,SAGN;AACZ,UAAM,YAAY,SAAS;AAC3B,UAAM,cAAc,SAAS,eAAe,KAAK;AAEjD,QAAI,KAAK,aAAa,aAAa;AACjC,aAAO,IAAI,cAAc;AAAA,QACvB,QAAQ,KAAK,eAAe;AAAA,QAC5B,OAAO,KAAK;AAAA,QACZ;AAAA,QACA,WAAW,KAAK;AAAA,QAChB,WAAW,aAAa;AAAA,QACxB,WAAW;AAAA,UACT;AAAA,YACE,cAAc,CAAC,WAAW;AACxB,oBAAM,aAAa,OAAO,cAAc,CAAC,IAAI,CAAC;AAC9C,oBAAM,QACJ,YAUC,SAAS;AAEZ,kBAAI,OAAO;AACT,sBAAM,eAAe,MAAM,+BAA+B;AAC1D,sBAAM,YAAY,MAAM,2BAA2B;AACnD,sBAAM,cAAc,MAAM,gBAAgB;AAC1C,sBAAM,eAAe,MAAM,iBAAiB;AAE5C,oBAAI,eAAe,GAAG;AACpB,0BAAQ;AAAA,oBACN,sCAAsC,YAAY;AAAA,kBACpD;AAAA,gBACF;AACA,oBAAI,YAAY,GAAG;AACjB,wBAAM,iBAAiB,KAAK;AAAA,oBACzB,aAAa,YAAY,eAAgB;AAAA,kBAC5C;AACA,0BAAQ;AAAA,oBACN,oCAAoC,SAAS,aAAa,cAAc;AAAA,kBAC1E;AAAA,gBACF;AACA,oBAAI,iBAAiB,KAAK,cAAc,KAAK,cAAc,GAAG;AAC5D,sBAAI,cAAc,KAAK;AACrB,4BAAQ;AAAA,sBACN,yBAAyB,WAAW,WAAW,YAAY;AAAA,oBAC7D;AAAA,kBACF,OAAO;AACL,4BAAQ;AAAA,sBACN,yBAAyB,WAAW,WAAW,YAAY;AAAA,oBAC7D;AAAA,kBACF;AAAA,gBACF;AAAA,cACF;AAAA,YACF;AAAA,UACF;AAAA,QACF;AAAA,MACF,CAAC;AAAA,IACH;AAEA,UAAM,sBAAsB,KAAK,wBAAwB;AAEzD,UAAM,cAAc,YAChB,sBACE,EAAE,aAAa,EAAE,uBAAuB,UAAU,EAAE,IACpD,EAAE,UAAU,IACd,CAAC;AAEL,UAAM,UAAU,KAAK,aAAa,aAAa,MAAS;AAIxD,UAAM,SAAS,KAAK,aAAa;AACjC,UAAM,gBAAgB,SAAS,MAAM;AAGrC,UAAM,cAAuC,CAAC;AAC9C,QAAI,uBAAuB,WAAW;AACpC,kBAAY,wBAAwB;AAAA,IACtC;AACA,QAAI,QAAQ;AACV,kBAAY,WAAW,EAAE,MAAM,WAAW;AAAA,IAC5C;AAEA,QAAI,KAAK,aAAa,cAAc;AAClC,kBAAY,cAAc;AAAA,IAC5B;AAEA,WAAO,IAAI,WAAW;AAAA,MACpB,QAAQ,KAAK,eAAe;AAAA,MAC5B,OAAO,KAAK;AAAA,MACZ,aAAa,sBAAsB,SAAY;AAAA,MAC/C,WAAW,KAAK;AAAA,MAChB;AAAA,MACA,GAAI,OAAO,KAAK,WAAW,EAAE,SAAS,IAAI,EAAE,YAAY,IAAI,CAAC;AAAA,MAC7D,GAAI,sBAAsB,CAAC,IAAI,YAAY,EAAE,UAAU,IAAI,CAAC;AAAA,MAC5D,eAAe;AAAA,QACb,QAAQ,KAAK,eAAe;AAAA,QAC5B,GAAI,KAAK,eAAe,UACpB,EAAE,SAAS,KAAK,eAAe,QAAQ,IACvC,CAAC;AAAA,MACP;AAAA,IACF,CAAC;AAAA,EACH;AAAA,EAEQ,oBAAoB,SAGd;AACZ,WAAO,KAAK,YAAY,OAAO;AAAA,EACjC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOQ,sBAA+B;AACrC,WAAO,KAAK,UAAU,SAAS,SAAS;AAAA,EAC1C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMQ,kBAAkB,QAAwB;AAChD,QAAI,KAAK,oBAAoB,GAAG;AAC9B,aAAO;AAAA,EAAc,MAAM;AAAA,IAC7B;AACA,WAAO;AAAA,EACT;AAAA,EAEA,cAA2B;AACzB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,eAAuB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,WAAsB;AACpB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,uBAAuB;AACrB,WAAO,KAAK,YAAY,UAAU;AAAA,EACpC;AAAA,EAEA,gBAAgB;AACd,WAAO,KAAK,cAAc,WAAW,KAAK;AAAA,EAC5C;AAAA,EAEA,MAAM,KAAQ,SAAwC;AACpD,UAAM,WAAW,MAAM,KAAK,iBAAiB,OAAO;AACpD,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,iBAAoB,SAAqD;AAC7E,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA,aAAa;AAAA,MACb,mBAAmB;AAAA,MACnB;AAAA,MACA,uBAAuB;AAAA,MACvB;AAAA,IACF,IAAI;AAEJ,QAAI,gBAAgB;AACpB,QAAI,YAA0B;AAE9B,YAAQ;AAAA,MACN,qCAAqC,KAAK,QAAQ,IAAI,KAAK,SAAS;AAAA,IACtE;AACA,YAAQ,IAAI,mCAAmC,WAAW,MAAM,QAAQ;AACxE,QAAI,WAAW;AACb,cAAQ,IAAI,gCAAgC,SAAS,EAAE;AAAA,IACzD;AAEA,aAAS,UAAU,GAAG,WAAW,YAAY,WAAW;AACtD,UAAI;AACF,gBAAQ;AAAA,UACN,4BAA4B,UAAU,CAAC,IAAI,aAAa,CAAC;AAAA,QAC3D;AACA,cAAM,mBAAmB,KAAK,IAAI;AAElC,cAAM,SAAS,MAAM,KAAK,YAAY,QAAQ,YAAY;AACxD,kBAAQ,IAAI,oCAAoC;AAChD,gBAAM,kBAAkB,KAAK,IAAI;AAEjC,gBAAM,aACJ,aAAa,gBAAgB,SACzB,KAAK,oBAAoB,EAAE,WAAW,YAAY,CAAC,IACnD,KAAK;AAEX,gBAAM,WAAW;AAAA,YACf,EAAE,MAAM,UAAU,SAAS,aAAa;AAAA,YACxC,EAAE,MAAM,QAAQ,SAAS,KAAK,kBAAkB,aAAa,EAAE;AAAA,UACjE;AACA,gBAAM,WAAW,MAAM,WAAW;AAAA,YAChC,KAAK,aAAa,cACd,gCAAgC,QAAQ,IACxC;AAAA,UACN;AAEA,kBAAQ;AAAA,YACN,uCAAuC,KAAK,IAAI,IAAI,eAAe;AAAA,UACrE;AAEA,cAAI,QAAyB;AAC7B,cAAI,SAAS,gBAAgB;AAC3B,kBAAM,YAAY,SAAS;AAI3B,oBAAQ;AAAA,cACN,cAAc,UAAU,gBAAgB;AAAA,cACxC,kBAAkB,UAAU,iBAAiB;AAAA,cAC7C,cACG,UAAU,gBAAgB,MAC1B,UAAU,iBAAiB;AAAA,YAChC;AACA,oBAAQ;AAAA,cACN,iCAAiC,MAAM,YAAY,QAAQ,MAAM,gBAAgB;AAAA,YACnF;AAEA,gBAAI,KAAK,cAAc;AACrB,mBAAK,aAAa;AAAA,gBAChB,MAAM;AAAA,gBACN,MAAM;AAAA,cACR;AAAA,YACF;AAAA,UACF;AAEA,gBAAM,eAAe,KAAK,oBAAoB,QAAQ;AACtD,cAAI,iBAAiB,UAAU;AAC7B,oBAAQ;AAAA,cACN;AAAA,YACF;AAAA,UACF;AAEA,gBAAM,UACJ,OAAO,SAAS,YAAY,WACxB,SAAS,UACT,KAAK,UAAU,SAAS,OAAO;AAErC,kBAAQ;AAAA,YACN,qCAAqC,QAAQ,MAAM,0BAA0B,YAAY;AAAA,UAC3F;AAEA,iBAAO,EAAE,SAAS,cAAc,MAAM;AAAA,QACxC,CAAC;AAED,gBAAQ;AAAA,UACN,4BAA4B,UAAU,CAAC,iBAAiB,KAAK,IAAI,IAAI,gBAAgB;AAAA,QACvF;AAEA,cAAM,SAAS,uBACV,kBAAkB,OAAO,SAAS,MAAS,IAC5C,kBAAkB,OAAO,SAAS,MAAM;AAC5C,gBAAQ;AAAA,UACN,gDAAgD,uBAAuB,iCAAiC,EAAE;AAAA,QAC5G;AAEA,eAAO;AAAA,UACL,MAAM;AAAA,UACN,KAAK,OAAO;AAAA,UACZ,cAAc,OAAO;AAAA,UACrB,OAAO,OAAO;AAAA,QAChB;AAAA,MACF,SAAS,OAAO;AACd,oBAAY,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;AACpE,gBAAQ;AAAA,UACN,4BAA4B,UAAU,CAAC;AAAA,UACvC,UAAU;AAAA,QACZ;AAEA,YAAI,KAAK,iBAAiB,SAAS,GAAG;AACpC,kBAAQ,MAAM,iDAAiD;AAC/D,gBAAM;AAAA,QACR;AAEA,YAAI,UAAU,cAAc,kBAAkB;AAC5C,kBAAQ,IAAI,gDAAgD;AAC5D,0BACE,GAAG,UAAU;AAAA;AAAA,iCACqB,UAAU,OAAO;AAAA;AAAA,QAEvD;AAAA,MACF;AAAA,IACF;AAEA,YAAQ,MAAM,yDAAyD;AACvE,UAAM;AAAA,EACR;AAAA,EAEQ,oBACN,UACiB;AACjB,UAAM,WAAW,SAAS;AAG1B,QAAI,UAAU,eAAe;AAC3B,YAAM,SAAS,SAAS;AACxB,UACE,WAAW,UACX,WAAW,YACX,WAAW,oBACX,WAAW,cACX;AACA,eAAO;AAAA,MACT;AAAA,IACF;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,QAAQ,SAIM;AAClB,UAAM,WAAW,MAAM,KAAK,oBAAoB,OAAO;AACvD,WAAO,SAAS;AAAA,EAClB;AAAA,EAEA,MAAM,oBAAoB,SAIuC;AAC/D,UAAM,EAAE,cAAc,YAAY,UAAU,IAAI;AAEhD,WAAO,KAAK,YAAY,QAAQ,YAAY;AAC1C,YAAM,aAAa,YACf,KAAK,oBAAoB,EAAE,UAAU,CAAC,IACtC,KAAK;AAET,YAAM,WAAW;AAAA,QACf,EAAE,MAAM,UAAU,SAAS,aAAa;AAAA,QACxC,EAAE,MAAM,QAAQ,SAAS,KAAK,kBAAkB,UAAU,EAAE;AAAA,MAC9D;AACA,YAAM,WAAW,MAAM,WAAW;AAAA,QAChC,KAAK,aAAa,cACd,gCAAgC,QAAQ,IACxC;AAAA,MACN;AAEA,UAAI,QAAyB;AAC7B,UAAI,SAAS,gBAAgB;AAC3B,cAAM,YAAY,SAAS;AAI3B,gBAAQ;AAAA,UACN,cAAc,UAAU,gBAAgB;AAAA,UACxC,kBAAkB,UAAU,iBAAiB;AAAA,UAC7C,cACG,UAAU,gBAAgB,MAAM,UAAU,iBAAiB;AAAA,QAChE;AAEA,YAAI,KAAK,cAAc;AACrB,eAAK,aAAa;AAAA,YAChB,MAAM;AAAA,YACN,MAAM;AAAA,UACR;AAAA,QACF;AAAA,MACF;AAEA,YAAM,eAAe,KAAK,oBAAoB,QAAQ;AACtD,YAAM,UACJ,OAAO,SAAS,YAAY,WACxB,SAAS,UACT,KAAK,UAAU,SAAS,OAAO;AAErC,aAAO,EAAE,KAAK,SAAS,cAAc,MAAM;AAAA,IAC7C,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,OAAO,UAAU,SAA2D;AAC1E,UAAM,EAAE,UAAU,WAAW,YAAY,IAAI;AAE7C,UAAM,aAAc,aAAa,gBAAgB,SAC7C,KAAK,oBAAoB,EAAE,WAAW,YAAY,CAAC,IACnD,KAAK;AAET,UAAM,oBAAoB,KAAK,aAAa,cACxC,gCAAgC,QAAQ,IACxC;AAEJ,UAAM,SAAS,MAAM,WAAW,OAAO,iBAAiB;AAExD,qBAAiB,SAAS,QAAQ;AAChC,YAAM,UAAU,OAAO,MAAM,YAAY,WACrC,MAAM,UACN,MAAM,QAAQ,MAAM,OAAO,IACzB,MAAM,QACH,OAAO,CAAC,MAA2C,OAAO,MAAM,YAAY,MAAM,QAAQ,UAAU,CAAC,EACrG,IAAI,CAAC,MAAM,EAAE,IAAI,EACjB,KAAK,EAAE,IACV;AAEN,UAAI,SAAS;AACX,cAAM,EAAE,SAAS,MAAM,MAAM;AAAA,MAC/B;AAAA,IACF;AAEA,UAAM,EAAE,SAAS,IAAI,MAAM,KAAK;AAAA,EAClC;AAAA,EAEQ,iBAAiB,OAAuB;AAC9C,UAAM,UAAU,MAAM,QAAQ,YAAY;AAC1C,WACE,QAAQ,SAAS,YAAY,KAC7B,QAAQ,SAAS,KAAK,KACtB,QAAQ,SAAS,gBAAgB;AAAA,EAErC;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,cACJ,SACyB;AACzB,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,aAAa;AAAA,MACb;AAAA,MACA,uBAAuB;AAAA,MACvB;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,QAAI,KAAK,aAAa,aAAa;AACjC,cAAQ;AAAA,QACN,sCAAsC,KAAK,QAAQ;AAAA,MACrD;AACA,aAAO,KAAK,iBAAiB,OAAO;AAAA,IACtC;AAEA,UAAM,kBACH,gBAAgB,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,EAAE,UACnD,cAAc,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,EAAE;AACpD,YAAQ;AAAA,MACN,6BAA6B,cAAc;AAAA,IAC7C;AAEA,QAAI,YAA0B;AAE9B,aAAS,UAAU,GAAG,WAAW,YAAY,WAAW;AACtD,UAAI;AACF,gBAAQ;AAAA,UACN,qCAAqC,UAAU,CAAC,IAAI,aAAa,CAAC;AAAA,QACpE;AAEA,cAAM,SAAS,MAAM,KAAK,YAAY,QAAQ,YAAY;AACxD,gBAAM,YAAY,IAAI,UAAU;AAEhC,gBAAM,gBACJ,gBAAgB,aAAa,SAAS,IAClC,aAAa,IAAI,CAAC,OAAO;AAAA,YACvB,MAAM;AAAA,YACN,MAAM,EAAE;AAAA,YACR,GAAI,EAAE,gBACF,EAAE,eAAe,EAAE,cAAc,IACjC,CAAC;AAAA,UACP,EAAE,IACF,eACE,CAAC,EAAE,MAAM,QAAiB,MAAM,aAAa,CAAC,IAC9C,CAAC;AAET,gBAAM,cACJ,cAAc,WAAW,SAAS,IAC9B,WAAW,IAAI,CAAC,OAAO;AAAA,YACrB,MAAM;AAAA,YACN,MAAM,EAAE;AAAA,YACR,GAAI,EAAE,gBACF,EAAE,eAAe,EAAE,cAAc,IACjC,CAAC;AAAA,UACP,EAAE,IACF,aACE,CAAC,EAAE,MAAM,QAAiB,MAAM,WAAW,CAAC,IAC5C,CAAC;AAET,gBAAM,WAAW,MAAM,UAAU,SAAS,OAAO;AAAA,YAC/C,OAAO,KAAK;AAAA,YACZ,YAAY,aAAa;AAAA,YACzB,aAAa,eAAe;AAAA,YAC5B,QAAQ;AAAA,YACR,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,YAAY,CAAC;AAAA,UACnD,CAAC;AAED,gBAAM,cAAc,SAAS,QAAQ,KAAK,CAAC,MAAM,EAAE,SAAS,MAAM;AAClE,gBAAM,UACJ,eAAe,UAAU,cAAc,YAAY,OAAO;AAE5D,gBAAM,WAAW,SAAS;AAO1B,gBAAM,YAAY,SAAS,2BAA2B;AACtD,gBAAM,gBAAgB,SAAS,+BAA+B;AAE9D,cAAI,gBAAgB,GAAG;AACrB,oBAAQ;AAAA,cACN,0CAA0C,aAAa;AAAA,YACzD;AAAA,UACF;AACA,cAAI,YAAY,GAAG;AACjB,kBAAM,iBAAiB,KAAK;AAAA,cACzB,aAAa,YAAY,SAAS,gBAAiB;AAAA,YACtD;AACA,oBAAQ;AAAA,cACN,wCAAwC,SAAS,aAAa,cAAc;AAAA,YAC9E;AAAA,UACF;AACA,cAAI,kBAAkB,KAAK,cAAc,GAAG;AAC1C,oBAAQ;AAAA,cACN,yCAAyC,SAAS,YAAY;AAAA,YAChE;AAAA,UACF;AAEA,gBAAM,QAAkB;AAAA,YACtB,cAAc,SAAS;AAAA,YACvB,kBAAkB,SAAS;AAAA,YAC3B,aAAa,SAAS,eAAe,SAAS;AAAA,UAChD;AAEA,cAAI,KAAK,cAAc;AACrB,iBAAK,aAAa;AAAA,cAChB,MAAM;AAAA,cACN,MAAM;AAAA,YACR;AAAA,UACF;AAEA,gBAAM,eACJ,SAAS,gBAAgB,aACrB,SACA,SAAS;AAEf,iBAAO;AAAA,YACL;AAAA,YACA;AAAA,YACA;AAAA,UACF;AAAA,QACF,CAAC;AAED,YAAI;AACJ,YAAI,SAAS;AACX,mBAAS,OAAO;AAAA,QAClB,WAAW,sBAAsB;AAC/B,mBAAS,kBAAkB,OAAO,SAAS,MAAS;AAAA,QACtD,OAAO;AACL,mBAAS,kBAAkB,OAAO,SAAS,MAAM;AAAA,QACnD;AAEA,eAAO;AAAA,UACL,MAAM;AAAA,UACN,KAAK,OAAO;AAAA,UACZ,cAAc,OAAO;AAAA,UACrB,OAAO,OAAO;AAAA,QAChB;AAAA,MACF,SAAS,OAAO;AACd,oBAAY,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;AACpE,gBAAQ;AAAA,UACN,qCAAqC,UAAU,CAAC;AAAA,UAChD,UAAU;AAAA,QACZ;AAEA,YAAI,KAAK,iBAAiB,SAAS,GAAG;AACpC,gBAAM;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAEA,UAAM;AAAA,EACR;AAAA,EAEA,OAAO,eAAe,MAAc,QAAQ,MAAsB;AAChE,WAAO,QACH,EAAE,MAAM,QAAQ,MAAM,eAAe,EAAE,MAAM,YAAY,EAAE,IAC3D,EAAE,MAAM,QAAQ,KAAK;AAAA,EAC3B;AACF;AAMA,IAAM,gBAAyD,CAAC;AAEzD,SAAS,mBAAmB,SAAuC;AACxE,QAAM,WAAW,SAAS,YAAY;AACtC,MAAI,CAAC,cAAc,QAAQ,GAAG;AAC5B,kBAAc,QAAQ,IAAI,IAAI,UAAU,OAAO;AAAA,EACjD;AACA,SAAO,cAAc,QAAQ;AAC/B;AAEO,SAAS,qBAAqB,UAA8B;AACjE,MAAI,UAAU;AACZ,WAAO,cAAc,QAAQ;AAAA,EAC/B,OAAO;AACL,eAAW,OAAO,OAAO,KAAK,aAAa,GAAoB;AAC7D,aAAO,cAAc,GAAG;AAAA,IAC1B;AAAA,EACF;AACF;AAMO,SAAS,uBAAoC;AAClD,MAAI,QAAQ,IAAI,kBAAmB,QAAO;AAC1C,MAAI,QAAQ,IAAI,iBAAkB,QAAO;AACzC,MAAI,QAAQ,IAAI,aAAc,QAAO;AACrC,MAAI,QAAQ,IAAI,eAAgB,QAAO;AACvC,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEO,SAAS,oBAAoB,UAAgC;AAClE,UAAQ,UAAU;AAAA,IAChB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB,KAAK;AACH,aAAO,CAAC,CAAC,QAAQ,IAAI;AAAA,IACvB;AACE,aAAO;AAAA,EACX;AACF;AAeO,SAAS,yBACd,SACW;AACX,QAAM,WAAW,SAAS,YAAY,qBAAqB;AAC3D,QAAM,eACJ,aAAa,aAAa,gBAAgB,OAAO,cAAc;AACjE,SAAO,IAAI,UAAU;AAAA,IACnB;AAAA,IACA,OAAO;AAAA,IACP,aAAa;AAAA,IACb,GAAG;AAAA,EACL,CAAC;AACH;AAWO,SAAS,qBACd,SACW;AACX,QAAM,WAAW,SAAS,YAAY,qBAAqB;AAC3D,QAAM,eACJ,aAAa,aAAa,gBAAgB,WAAW,cAAc;AACrE,SAAO,IAAI,UAAU;AAAA,IACnB;AAAA,IACA,OAAO;AAAA,IACP,aAAa;AAAA,IACb,GAAG;AAAA,EACL,CAAC;AACH;AAWO,SAAS,gBACd,SACW;AACX,QAAM,WAAW,SAAS,YAAY,qBAAqB;AAC3D,QAAM,eACJ,aAAa,aACT,gBAAgB,OAChB,cAAc;AACpB,SAAO,IAAI,UAAU;AAAA,IACnB;AAAA,IACA,OAAO;AAAA,IACP,aAAa;AAAA,IACb,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,qBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,gBAAgB;AAAA,IACvB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,mBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,cAAc;AAAA,IACrB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,sBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,iBAAiB;AAAA,IACxB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,iBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,YAAY;AAAA,IACnB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,uBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,kBAAkB;AAAA,IACzB,GAAG;AAAA,EACL,CAAC;AACH;AAQO,SAAS,kBACd,SACW;AACX,SAAO,IAAI,UAAU;AAAA,IACnB,UAAU;AAAA,IACV,OAAO,kBAAkB;AAAA,IACzB,GAAG;AAAA,EACL,CAAC;AACH;","names":[]}

package/dist/chunk-MJS33AAS.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/rate-limiter.ts","../src/token-tracker.ts"],"sourcesContent":["/**\n * Rate Limiter for LLM API Calls\n *\n * Implements token bucket algorithm with:\n * - Configurable requests per minute/second\n * - Automatic backoff on 429 errors\n * - Queue for pending requests\n *\n * @packageDocumentation\n */\n\nexport interface RateLimiterOptions {\n  /** Maximum requests per minute (default: 60) */\n  requestsPerMinute?: number;\n  /** Maximum requests per second (default: 3) */\n  requestsPerSecond?: number;\n  /** Maximum concurrent requests (default: 5) */\n  maxConcurrent?: number;\n  /** Base delay for exponential backoff in ms (default: 1000) */\n  baseBackoffMs?: number;\n  /** Maximum backoff delay in ms (default: 60000) */\n  maxBackoffMs?: number;\n}\n\ninterface QueuedRequest<T> {\n  execute: () => Promise<T>;\n  resolve: (value: T) => void;\n  reject: (error: Error) => void;\n  retryCount: number;\n}\n\n/**\n * Rate limiter for LLM API calls using token bucket algorithm.\n *\n * @example\n * ```typescript\n * const limiter = new RateLimiter({ requestsPerMinute: 30 });\n * const result = await limiter.execute(() => llm.invoke(messages));\n * ```\n */\nexport class RateLimiter {\n  private requestsPerMinute: number;\n  private requestsPerSecond: number;\n  private maxConcurrent: number;\n  private baseBackoffMs: number;\n  private maxBackoffMs: number;\n\n  private minuteTokens: number;\n  private secondTokens: number;\n  private activeRequests: number = 0;\n  private queue: QueuedRequest<unknown>[] = [];\n  private lastMinuteReset: number = Date.now();\n  private lastSecondReset: number = Date.now();\n  private processing: boolean = false;\n  private currentBackoffMs: number = 0;\n\n  constructor(options: RateLimiterOptions = {}) {\n    this.requestsPerMinute = options.requestsPerMinute ?? 60;\n    this.requestsPerSecond = options.requestsPerSecond ?? 3;\n    this.maxConcurrent = options.maxConcurrent ?? 5;\n    this.baseBackoffMs = options.baseBackoffMs ?? 1000;\n    this.maxBackoffMs = options.maxBackoffMs ?? 60000;\n\n    this.minuteTokens = this.requestsPerMinute;\n    this.secondTokens = this.requestsPerSecond;\n  }\n\n  async execute<T>(fn: () => Promise<T>, _maxRetries: number = 3): Promise<T> {\n    return new Promise<T>((resolve, reject) => {\n      this.queue.push({\n        execute: fn as () => Promise<unknown>,\n        resolve: resolve as (value: unknown) => void,\n        reject,\n        retryCount: 0,\n      });\n      this.processQueue();\n    });\n  }\n\n  getStatus(): {\n    queueLength: number;\n    activeRequests: number;\n    minuteTokens: number;\n    secondTokens: number;\n    backoffMs: number;\n  } {\n    return {\n      queueLength: this.queue.length,\n      activeRequests: this.activeRequests,\n      minuteTokens: this.minuteTokens,\n      secondTokens: this.secondTokens,\n      backoffMs: this.currentBackoffMs,\n    };\n  }\n\n  reset(): void {\n    this.minuteTokens = this.requestsPerMinute;\n    this.secondTokens = this.requestsPerSecond;\n    this.activeRequests = 0;\n    this.queue = [];\n    this.currentBackoffMs = 0;\n    this.lastMinuteReset = Date.now();\n    this.lastSecondReset = Date.now();\n  }\n\n  private async processQueue(): Promise<void> {\n    if (this.processing) return;\n    this.processing = true;\n\n    while (this.queue.length > 0) {\n      this.refillTokens();\n\n      if (!this.canMakeRequest()) {\n        const waitTime = this.getWaitTime();\n        await this.sleep(waitTime);\n        continue;\n      }\n\n      if (this.currentBackoffMs > 0) {\n        await this.sleep(this.currentBackoffMs);\n        this.currentBackoffMs = 0;\n      }\n\n      const request = this.queue.shift();\n      if (!request) continue;\n\n      this.consumeTokens();\n      this.activeRequests++;\n\n      try {\n        const result = await request.execute();\n        request.resolve(result);\n        this.currentBackoffMs = 0;\n      } catch (error) {\n        const err = error instanceof Error ? error : new Error(String(error));\n\n        if (this.isRateLimitError(err)) {\n          this.currentBackoffMs = Math.min(\n            this.baseBackoffMs * Math.pow(2, request.retryCount),\n            this.maxBackoffMs,\n          );\n\n          console.warn(\n            `[RateLimiter] Rate limited. Backing off for ${this.currentBackoffMs}ms ` +\n              `(retry ${request.retryCount + 1})`,\n          );\n\n          if (request.retryCount < 3) {\n            this.queue.unshift({\n              ...request,\n              retryCount: request.retryCount + 1,\n            });\n          } else {\n            request.reject(\n              new Error(\n                `Rate limit exceeded after ${request.retryCount + 1} retries: ${err.message}`,\n              ),\n            );\n          }\n        } else {\n          request.reject(err);\n        }\n      } finally {\n        this.activeRequests--;\n      }\n    }\n\n    this.processing = false;\n  }\n\n  private refillTokens(): void {\n    const now = Date.now();\n    if (now - this.lastMinuteReset >= 60000) {\n      this.minuteTokens = this.requestsPerMinute;\n      this.lastMinuteReset = now;\n    }\n    if (now - this.lastSecondReset >= 1000) {\n      this.secondTokens = this.requestsPerSecond;\n      this.lastSecondReset = now;\n    }\n  }\n\n  private canMakeRequest(): boolean {\n    return (\n      this.minuteTokens > 0 &&\n      this.secondTokens > 0 &&\n      this.activeRequests < this.maxConcurrent\n    );\n  }\n\n  private consumeTokens(): void {\n    this.minuteTokens--;\n    this.secondTokens--;\n  }\n\n  private getWaitTime(): number {\n    const now = Date.now();\n    if (this.secondTokens <= 0) {\n      return Math.max(0, 1000 - (now - this.lastSecondReset));\n    }\n    if (this.minuteTokens <= 0) {\n      return Math.max(0, 60000 - (now - this.lastMinuteReset));\n    }\n    return 100;\n  }\n\n  private isRateLimitError(error: Error): boolean {\n    const message = error.message.toLowerCase();\n    return (\n      message.includes('429') ||\n      message.includes('rate limit') ||\n      message.includes('too many requests') ||\n      message.includes('quota exceeded')\n    );\n  }\n\n  private sleep(ms: number): Promise<void> {\n    return new Promise((resolve) => setTimeout(resolve, ms));\n  }\n}\n\n// Singleton instance\nlet globalRateLimiter: RateLimiter | null = null;\n\nexport function getGlobalRateLimiter(\n  options?: RateLimiterOptions,\n): RateLimiter {\n  if (!globalRateLimiter) {\n    globalRateLimiter = new RateLimiter(options);\n  }\n  return globalRateLimiter;\n}\n\nexport function resetGlobalRateLimiter(): void {\n  globalRateLimiter?.reset();\n  globalRateLimiter = null;\n}\n","/**\n * Token Tracker for LLM Usage\n *\n * Tracks token usage across multiple LLM calls for:\n * - Cost estimation\n * - Usage monitoring\n * - Quota management\n *\n * @packageDocumentation\n */\n\nexport interface TokenUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  callCount: number;\n}\n\nexport interface TokenCost {\n  promptCostPer1K: number;\n  completionCostPer1K: number;\n}\n\n// Pricing as of 2024 (update as needed)\nconst MODEL_COSTS: Record<string, TokenCost> = {\n  'gpt-4o': { promptCostPer1K: 0.005, completionCostPer1K: 0.015 },\n  'gpt-4o-mini': { promptCostPer1K: 0.00015, completionCostPer1K: 0.0006 },\n  'gpt-4-turbo': { promptCostPer1K: 0.01, completionCostPer1K: 0.03 },\n  'gpt-4': { promptCostPer1K: 0.03, completionCostPer1K: 0.06 },\n  'gpt-3.5-turbo': {\n    promptCostPer1K: 0.0005,\n    completionCostPer1K: 0.0015,\n  },\n};\n\nexport class TokenTracker {\n  private model: string;\n  private usage: TokenUsage = {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n    callCount: 0,\n  };\n\n  constructor(model: string = 'gpt-4o') {\n    this.model = model;\n  }\n\n  addUsage(promptTokens: number, completionTokens: number): void {\n    this.usage.promptTokens += promptTokens;\n    this.usage.completionTokens += completionTokens;\n    this.usage.totalTokens += promptTokens + completionTokens;\n    this.usage.callCount++;\n  }\n\n  getSummary(): TokenUsage {\n    return { ...this.usage };\n  }\n\n  getEstimatedCost(): number {\n    const costs = MODEL_COSTS[this.model] || MODEL_COSTS['gpt-4o'];\n    const promptCost =\n      (this.usage.promptTokens / 1000) * costs.promptCostPer1K;\n    const completionCost =\n      (this.usage.completionTokens / 1000) * costs.completionCostPer1K;\n    return promptCost + completionCost;\n  }\n\n  getFormattedCost(): string {\n    const cost = this.getEstimatedCost();\n    return `$${cost.toFixed(4)}`;\n  }\n\n  getReport(): string {\n    const summary = this.getSummary();\n    const cost = this.getEstimatedCost();\n    return [\n      `Token Usage Report (${this.model})`,\n      `─────────────────────────────`,\n      `Calls:              ${summary.callCount}`,\n      `Prompt Tokens:      ${summary.promptTokens.toLocaleString()}`,\n      `Completion Tokens:  ${summary.completionTokens.toLocaleString()}`,\n      `Total Tokens:       ${summary.totalTokens.toLocaleString()}`,\n      `Estimated Cost:     $${cost.toFixed(4)}`,\n    ].join('\\n');\n  }\n\n  reset(): void {\n    this.usage = {\n      promptTokens: 0,\n      completionTokens: 0,\n      totalTokens: 0,\n      callCount: 0,\n    };\n  }\n\n  setModel(model: string): void {\n    this.model = model;\n  }\n}\n\n// Global tracker instance\nlet globalTracker: TokenTracker | null = null;\n\nexport function getGlobalTokenTracker(model?: string): TokenTracker {\n  if (!globalTracker) {\n    globalTracker = new TokenTracker(model);\n  } else if (model) {\n    globalTracker.setModel(model);\n  }\n  return globalTracker;\n}\n\nexport function resetGlobalTokenTracker(): void {\n  globalTracker?.reset();\n}\n"],"mappings":";AAwCO,IAAM,cAAN,MAAkB;AAAA,EAgBvB,YAAY,UAA8B,CAAC,GAAG;AAP9C,SAAQ,iBAAyB;AACjC,SAAQ,QAAkC,CAAC;AAC3C,SAAQ,kBAA0B,KAAK,IAAI;AAC3C,SAAQ,kBAA0B,KAAK,IAAI;AAC3C,SAAQ,aAAsB;AAC9B,SAAQ,mBAA2B;AAGjC,SAAK,oBAAoB,QAAQ,qBAAqB;AACtD,SAAK,oBAAoB,QAAQ,qBAAqB;AACtD,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,eAAe,QAAQ,gBAAgB;AAE5C,SAAK,eAAe,KAAK;AACzB,SAAK,eAAe,KAAK;AAAA,EAC3B;AAAA,EAEA,MAAM,QAAW,IAAsB,cAAsB,GAAe;AAC1E,WAAO,IAAI,QAAW,CAAC,SAAS,WAAW;AACzC,WAAK,MAAM,KAAK;AAAA,QACd,SAAS;AAAA,QACT;AAAA,QACA;AAAA,QACA,YAAY;AAAA,MACd,CAAC;AACD,WAAK,aAAa;AAAA,IACpB,CAAC;AAAA,EACH;AAAA,EAEA,YAME;AACA,WAAO;AAAA,MACL,aAAa,KAAK,MAAM;AAAA,MACxB,gBAAgB,KAAK;AAAA,MACrB,cAAc,KAAK;AAAA,MACnB,cAAc,KAAK;AAAA,MACnB,WAAW,KAAK;AAAA,IAClB;AAAA,EACF;AAAA,EAEA,QAAc;AACZ,SAAK,eAAe,KAAK;AACzB,SAAK,eAAe,KAAK;AACzB,SAAK,iBAAiB;AACtB,SAAK,QAAQ,CAAC;AACd,SAAK,mBAAmB;AACxB,SAAK,kBAAkB,KAAK,IAAI;AAChC,SAAK,kBAAkB,KAAK,IAAI;AAAA,EAClC;AAAA,EAEA,MAAc,eAA8B;AAC1C,QAAI,KAAK,WAAY;AACrB,SAAK,aAAa;AAElB,WAAO,KAAK,MAAM,SAAS,GAAG;AAC5B,WAAK,aAAa;AAElB,UAAI,CAAC,KAAK,eAAe,GAAG;AAC1B,cAAM,WAAW,KAAK,YAAY;AAClC,cAAM,KAAK,MAAM,QAAQ;AACzB;AAAA,MACF;AAEA,UAAI,KAAK,mBAAmB,GAAG;AAC7B,cAAM,KAAK,MAAM,KAAK,gBAAgB;AACtC,aAAK,mBAAmB;AAAA,MAC1B;AAEA,YAAM,UAAU,KAAK,MAAM,MAAM;AACjC,UAAI,CAAC,QAAS;AAEd,WAAK,cAAc;AACnB,WAAK;AAEL,UAAI;AACF,cAAM,SAAS,MAAM,QAAQ,QAAQ;AACrC,gBAAQ,QAAQ,MAAM;AACtB,aAAK,mBAAmB;AAAA,MAC1B,SAAS,OAAO;AACd,cAAM,MAAM,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;AAEpE,YAAI,KAAK,iBAAiB,GAAG,GAAG;AAC9B,eAAK,mBAAmB,KAAK;AAAA,YAC3B,KAAK,gBAAgB,KAAK,IAAI,GAAG,QAAQ,UAAU;AAAA,YACnD,KAAK;AAAA,UACP;AAEA,kBAAQ;AAAA,YACN,+CAA+C,KAAK,gBAAgB,aACxD,QAAQ,aAAa,CAAC;AAAA,UACpC;AAEA,cAAI,QAAQ,aAAa,GAAG;AAC1B,iBAAK,MAAM,QAAQ;AAAA,cACjB,GAAG;AAAA,cACH,YAAY,QAAQ,aAAa;AAAA,YACnC,CAAC;AAAA,UACH,OAAO;AACL,oBAAQ;AAAA,cACN,IAAI;AAAA,gBACF,6BAA6B,QAAQ,aAAa,CAAC,aAAa,IAAI,OAAO;AAAA,cAC7E;AAAA,YACF;AAAA,UACF;AAAA,QACF,OAAO;AACL,kBAAQ,OAAO,GAAG;AAAA,QACpB;AAAA,MACF,UAAE;AACA,aAAK;AAAA,MACP;AAAA,IACF;AAEA,SAAK,aAAa;AAAA,EACpB;AAAA,EAEQ,eAAqB;AAC3B,UAAM,MAAM,KAAK,IAAI;AACrB,QAAI,MAAM,KAAK,mBAAmB,KAAO;AACvC,WAAK,eAAe,KAAK;AACzB,WAAK,kBAAkB;AAAA,IACzB;AACA,QAAI,MAAM,KAAK,mBAAmB,KAAM;AACtC,WAAK,eAAe,KAAK;AACzB,WAAK,kBAAkB;AAAA,IACzB;AAAA,EACF;AAAA,EAEQ,iBAA0B;AAChC,WACE,KAAK,eAAe,KACpB,KAAK,eAAe,KACpB,KAAK,iBAAiB,KAAK;AAAA,EAE/B;AAAA,EAEQ,gBAAsB;AAC5B,SAAK;AACL,SAAK;AAAA,EACP;AAAA,EAEQ,cAAsB;AAC5B,UAAM,MAAM,KAAK,IAAI;AACrB,QAAI,KAAK,gBAAgB,GAAG;AAC1B,aAAO,KAAK,IAAI,GAAG,OAAQ,MAAM,KAAK,gBAAgB;AAAA,IACxD;AACA,QAAI,KAAK,gBAAgB,GAAG;AAC1B,aAAO,KAAK,IAAI,GAAG,OAAS,MAAM,KAAK,gBAAgB;AAAA,IACzD;AACA,WAAO;AAAA,EACT;AAAA,EAEQ,iBAAiB,OAAuB;AAC9C,UAAM,UAAU,MAAM,QAAQ,YAAY;AAC1C,WACE,QAAQ,SAAS,KAAK,KACtB,QAAQ,SAAS,YAAY,KAC7B,QAAQ,SAAS,mBAAmB,KACpC,QAAQ,SAAS,gBAAgB;AAAA,EAErC;AAAA,EAEQ,MAAM,IAA2B;AACvC,WAAO,IAAI,QAAQ,CAAC,YAAY,WAAW,SAAS,EAAE,CAAC;AAAA,EACzD;AACF;AAGA,IAAI,oBAAwC;AAErC,SAAS,qBACd,SACa;AACb,MAAI,CAAC,mBAAmB;AACtB,wBAAoB,IAAI,YAAY,OAAO;AAAA,EAC7C;AACA,SAAO;AACT;AAEO,SAAS,yBAA+B;AAC7C,qBAAmB,MAAM;AACzB,sBAAoB;AACtB;;;ACpNA,IAAM,cAAyC;AAAA,EAC7C,UAAU,EAAE,iBAAiB,MAAO,qBAAqB,MAAM;AAAA,EAC/D,eAAe,EAAE,iBAAiB,OAAS,qBAAqB,KAAO;AAAA,EACvE,eAAe,EAAE,iBAAiB,MAAM,qBAAqB,KAAK;AAAA,EAClE,SAAS,EAAE,iBAAiB,MAAM,qBAAqB,KAAK;AAAA,EAC5D,iBAAiB;AAAA,IACf,iBAAiB;AAAA,IACjB,qBAAqB;AAAA,EACvB;AACF;AAEO,IAAM,eAAN,MAAmB;AAAA,EASxB,YAAY,QAAgB,UAAU;AAPtC,SAAQ,QAAoB;AAAA,MAC1B,cAAc;AAAA,MACd,kBAAkB;AAAA,MAClB,aAAa;AAAA,MACb,WAAW;AAAA,IACb;AAGE,SAAK,QAAQ;AAAA,EACf;AAAA,EAEA,SAAS,cAAsB,kBAAgC;AAC7D,SAAK,MAAM,gBAAgB;AAC3B,SAAK,MAAM,oBAAoB;AAC/B,SAAK,MAAM,eAAe,eAAe;AACzC,SAAK,MAAM;AAAA,EACb;AAAA,EAEA,aAAyB;AACvB,WAAO,EAAE,GAAG,KAAK,MAAM;AAAA,EACzB;AAAA,EAEA,mBAA2B;AACzB,UAAM,QAAQ,YAAY,KAAK,KAAK,KAAK,YAAY,QAAQ;AAC7D,UAAM,aACH,KAAK,MAAM,eAAe,MAAQ,MAAM;AAC3C,UAAM,iBACH,KAAK,MAAM,mBAAmB,MAAQ,MAAM;AAC/C,WAAO,aAAa;AAAA,EACtB;AAAA,EAEA,mBAA2B;AACzB,UAAM,OAAO,KAAK,iBAAiB;AACnC,WAAO,IAAI,KAAK,QAAQ,CAAC,CAAC;AAAA,EAC5B;AAAA,EAEA,YAAoB;AAClB,UAAM,UAAU,KAAK,WAAW;AAChC,UAAM,OAAO,KAAK,iBAAiB;AACnC,WAAO;AAAA,MACL,uBAAuB,KAAK,KAAK;AAAA,MACjC;AAAA,MACA,uBAAuB,QAAQ,SAAS;AAAA,MACxC,uBAAuB,QAAQ,aAAa,eAAe,CAAC;AAAA,MAC5D,uBAAuB,QAAQ,iBAAiB,eAAe,CAAC;AAAA,MAChE,uBAAuB,QAAQ,YAAY,eAAe,CAAC;AAAA,MAC3D,wBAAwB,KAAK,QAAQ,CAAC,CAAC;AAAA,IACzC,EAAE,KAAK,IAAI;AAAA,EACb;AAAA,EAEA,QAAc;AACZ,SAAK,QAAQ;AAAA,MACX,cAAc;AAAA,MACd,kBAAkB;AAAA,MAClB,aAAa;AAAA,MACb,WAAW;AAAA,IACb;AAAA,EACF;AAAA,EAEA,SAAS,OAAqB;AAC5B,SAAK,QAAQ;AAAA,EACf;AACF;AAGA,IAAI,gBAAqC;AAElC,SAAS,sBAAsB,OAA8B;AAClE,MAAI,CAAC,eAAe;AAClB,oBAAgB,IAAI,aAAa,KAAK;AAAA,EACxC,WAAW,OAAO;AAChB,kBAAc,SAAS,KAAK;AAAA,EAC9B;AACA,SAAO;AACT;AAEO,SAAS,0BAAgC;AAC9C,iBAAe,MAAM;AACvB;","names":[]}