@posthog/ai/dist/vercel/index.mjs.map

PostHog Node.js AI integrations

{"version":3,"file":"index.mjs","sources":["../../src/typeGuards.ts","../../src/sanitization/base64_recognizer.ts","../../src/sanitization/media_type_context.ts","../../src/sanitization/binary_content_redactor.ts","../../src/sanitization.ts","../../src/utils.ts","../../src/serializeError.ts","../../src/captureAiGeneration.ts","../../src/vercel/middleware.ts"],"sourcesContent":["// Type guards for safer type checking\n\nexport const isString = (value: unknown): value is string => {\n  return typeof value === 'string'\n}\n\nexport const isObject = (value: unknown): value is Record<string, unknown> => {\n  return value !== null && typeof value === 'object' && !Array.isArray(value)\n}\n","const DATA_URL_PREFIX_RE = /^data:([^;,\\s]+)(?:;[^;,\\s]+)*;base64,/i\nconst BASE64_ALPHABET_RE = /^[A-Za-z0-9+/_=-]+$/\n\nexport type Base64Recognition = { kind: 'data-url'; mediaType: string } | { kind: 'raw' } | { kind: 'none' }\n\nexport class Base64Recognizer {\n  recognize(value: string, minLength: number): Base64Recognition {\n    const dataUrl = DATA_URL_PREFIX_RE.exec(value)\n    if (dataUrl) return { kind: 'data-url', mediaType: dataUrl[1] }\n\n    if (value.length < minLength) return { kind: 'none' }\n\n    const confidencePrefix = value.slice(0, minLength)\n    if (BASE64_ALPHABET_RE.test(confidencePrefix)) {\n      return { kind: 'raw' }\n    } else {\n      return { kind: 'none' }\n    }\n  }\n}\n","const MIME_HINT_KEYS = ['mediaType', 'media_type', 'mimeType', 'mime_type'] as const\n\nconst STRONG_CONTEXT_KEYS = new Set([\n  'data',\n  'file_data',\n  'fileData',\n  'image_url',\n  'imageUrl',\n  'video_url',\n  'videoUrl',\n  'audio',\n  'audio_data',\n  'audioData',\n  'inline_data',\n  'inlineData',\n  'source',\n  'result',\n])\n\nconst STRONG_CONTEXT_TYPES = new Set([\n  'image',\n  'image_url',\n  'input_image',\n  'audio',\n  'input_audio',\n  'video',\n  'video_url',\n  'file',\n  'input_file',\n  'document',\n  'media',\n  'file-data',\n])\n\nconst FILE_FAMILY_TYPES = new Set(['file', 'input_file', 'document', 'media', 'file-data'])\n\nconst KNOWN_AUDIO_FORMATS = new Set(['wav', 'mp3', 'ogg', 'flac', 'm4a', 'aac', 'webm'])\n\nexport class MediaTypeContext {\n  static readonly EMPTY = new MediaTypeContext(undefined, undefined)\n\n  constructor(\n    private readonly parent: Record<string, unknown> | undefined,\n    private readonly key: string | undefined\n  ) {}\n\n  inferMediaType(): string | undefined {\n    return (\n      this.inferFromSiblingMime() ?? this.inferFromSiblingFormat() ?? this.inferFromParentType() ?? this.inferFromKey()\n    )\n  }\n\n  inferFromSiblingMime(): string | undefined {\n    if (!this.parent) return undefined\n    for (const hint of MIME_HINT_KEYS) {\n      const v = this.parent[hint]\n      if (typeof v === 'string') return v\n    }\n    return undefined\n  }\n\n  inferFromSiblingFormat(): string | undefined {\n    if (!this.parent) return undefined\n    const fmt = this.parent.format\n    if (typeof fmt === 'string' && KNOWN_AUDIO_FORMATS.has(fmt.toLowerCase())) {\n      return `audio/${fmt.toLowerCase()}`\n    }\n    return undefined\n  }\n\n  inferFromParentType(): string | undefined {\n    if (!this.parent) return undefined\n    const t = this.parent.type\n    if (typeof t !== 'string') return undefined\n    if (t === 'image' || t === 'image_url' || t === 'input_image') return 'image'\n    if (t === 'audio' || t === 'input_audio') return 'audio'\n    if (t === 'video' || t === 'video_url') return 'video'\n    if (FILE_FAMILY_TYPES.has(t)) return 'application/octet-stream'\n    return undefined\n  }\n\n  inferFromKey(): string | undefined {\n    if (!this.key) return undefined\n    const key = this.key.toLowerCase()\n    if (key.includes('audio')) return 'audio'\n    if (key.includes('video')) return 'video'\n    if (key.includes('image')) return 'image'\n    if (key.includes('file') || key.includes('document')) return 'application/octet-stream'\n    return undefined\n  }\n\n  signalsBinary(): boolean {\n    if (this.parent) {\n      for (const hint of MIME_HINT_KEYS) {\n        if (typeof this.parent[hint] === 'string') return true\n      }\n      const fmt = this.parent.format\n      if (typeof fmt === 'string' && KNOWN_AUDIO_FORMATS.has(fmt.toLowerCase())) return true\n      const t = this.parent.type\n      if (typeof t === 'string' && STRONG_CONTEXT_TYPES.has(t)) return true\n    }\n    if (this.key && STRONG_CONTEXT_KEYS.has(this.key)) return true\n    return false\n  }\n}\n","import { Base64Recognizer } from './base64_recognizer'\nimport { MediaTypeContext } from './media_type_context'\n\nconst STRONG_CONTEXT_MIN_LENGTH = 64\nconst WEAK_CONTEXT_MIN_LENGTH = 1024\n\nexport class BinaryContentRedactor {\n  private visited: WeakSet<object> = new WeakSet()\n\n  constructor(private readonly recognizer: Base64Recognizer = new Base64Recognizer()) {}\n\n  redact<T>(value: T): T\n  redact(value: unknown): unknown {\n    if (this.isMultimodalEnabled()) return value\n    this.visited = new WeakSet()\n    return this.walk(value, MediaTypeContext.EMPTY)\n  }\n\n  private walk(value: unknown, ctx: MediaTypeContext): unknown {\n    if (value === null || value === undefined) return value\n    if (typeof value === 'string') return this.redactString(value, ctx)\n    if (typeof value !== 'object') return value\n\n    // Buffer extends Uint8Array, so this branch catches both.\n    if (typeof Uint8Array !== 'undefined' && value instanceof Uint8Array) {\n      return this.placeholderFor(ctx.inferMediaType())\n    }\n\n    if (this.visited.has(value)) return null\n    this.visited.add(value)\n\n    if (Array.isArray(value)) {\n      return value.map((item) => this.walk(item, ctx))\n    }\n\n    const obj = value as Record<string, unknown>\n    const out: Record<string, unknown> = {}\n    for (const k of Object.keys(obj)) {\n      out[k] = this.walk(obj[k], new MediaTypeContext(obj, k))\n    }\n    return out\n  }\n\n  private redactString(value: string, ctx: MediaTypeContext): string {\n    const minLength = ctx.signalsBinary() ? STRONG_CONTEXT_MIN_LENGTH : WEAK_CONTEXT_MIN_LENGTH\n    const recognition = this.recognizer.recognize(value, minLength)\n    switch (recognition.kind) {\n      case 'data-url':\n        return this.placeholderFor(recognition.mediaType)\n      case 'raw':\n        return this.placeholderFor(ctx.inferMediaType())\n      case 'none':\n        return value\n    }\n  }\n\n  private placeholderFor(mediaType?: string): string {\n    if (!mediaType) return '[base64 redacted]'\n    if (mediaType === 'application/octet-stream') return '[base64 file redacted]'\n    return `[base64 ${mediaType} redacted]`\n  }\n\n  private isMultimodalEnabled(): boolean {\n    const val = process.env._INTERNAL_LLMA_MULTIMODAL || ''\n    return val.toLowerCase() === 'true' || val === '1' || val.toLowerCase() === 'yes'\n  }\n}\n","import { BinaryContentRedactor } from './sanitization/binary_content_redactor'\n\nconst redactor = new BinaryContentRedactor()\n\nexport function redactBase64DataUrl(str: string): string\nexport function redactBase64DataUrl(str: unknown): unknown\nexport function redactBase64DataUrl(str: unknown): unknown {\n  return redactor.redact(str)\n}\n\nexport const sanitizeOpenAI = (data: unknown): unknown => redactor.redact(data)\nexport const sanitizeOpenAIResponse = (data: unknown): unknown => redactor.redact(data)\nexport const sanitizeAnthropic = (data: unknown): unknown => redactor.redact(data)\nexport const sanitizeGemini = (data: unknown): unknown => redactor.redact(data)\nexport const sanitizeLangChain = (data: unknown): unknown => redactor.redact(data)\n","import { PostHog } from 'posthog-node'\nimport OpenAIOrignal from 'openai'\nimport AnthropicOriginal from '@anthropic-ai/sdk'\nimport type { ChatCompletionTool } from 'openai/resources/chat/completions'\nimport type { ResponseCreateParamsWithTools } from 'openai/lib/ResponsesParser'\nimport type { Tool as GeminiTool } from '@google/genai'\nimport type {\n  FormattedMessage,\n  FormattedContent,\n  FormattedAudioContent,\n  FormattedImageContent,\n  FormattedDocumentContent,\n} from './types'\nimport { v4 as uuidv4 } from 'uuid'\nimport { isString } from './typeGuards'\nimport { redactBase64DataUrl } from './sanitization'\n\ntype ChatCompletionCreateParamsBase = OpenAIOrignal.Chat.Completions.ChatCompletionCreateParams\ntype MessageCreateParams = AnthropicOriginal.Messages.MessageCreateParams\ntype ResponseCreateParams = OpenAIOrignal.Responses.ResponseCreateParams\ntype EmbeddingCreateParams = OpenAIOrignal.EmbeddingCreateParams\ntype TranscriptionCreateParams = OpenAIOrignal.Audio.Transcriptions.TranscriptionCreateParams\ntype AnthropicTool = AnthropicOriginal.Tool\n\nconst TOKEN_PROPERTY_KEYS = new Set([\n  '$ai_input_tokens',\n  '$ai_output_tokens',\n  '$ai_cache_read_input_tokens',\n  '$ai_cache_creation_input_tokens',\n  '$ai_total_tokens',\n  '$ai_reasoning_tokens',\n])\n\nexport function getTokensSource(posthogProperties?: Record<string, unknown>): string {\n  if (posthogProperties && Object.keys(posthogProperties).some((key) => TOKEN_PROPERTY_KEYS.has(key))) {\n    return 'passthrough'\n  }\n  return 'sdk'\n}\n\n// limit large outputs by truncating to 200kb (approx 200k bytes)\nexport const MAX_OUTPUT_SIZE = 200000\nconst STRING_FORMAT = 'utf8'\n\n// Reused across calls to avoid per-invocation allocation; truncate() runs\n// hundreds of times for prompts with many parts.\nconst sharedTextEncoder = new TextEncoder()\nconst sharedTextDecoder = new TextDecoder(STRING_FORMAT, { fatal: false })\n\nexport const utf8ByteLength = (str: string): number => sharedTextEncoder.encode(str).byteLength\n\n/**\n * Safely converts content to a string, preserving structure for objects/arrays.\n * - If content is already a string, returns it as-is\n * - If content is an object or array, stringifies it with JSON.stringify to preserve structure\n * - Otherwise, converts to string with String()\n *\n * This prevents the \"[object Object]\" bug when objects are naively converted to strings.\n *\n * @param content - The content to convert to a string\n * @returns A string representation that preserves structure for complex types\n */\nexport function toContentString(content: unknown): string {\n  if (typeof content === 'string') {\n    return content\n  }\n  if (content !== undefined && content !== null && typeof content === 'object') {\n    try {\n      return JSON.stringify(content)\n    } catch {\n      // Fallback for circular refs, BigInt, or objects with throwing toJSON\n      return String(content)\n    }\n  }\n  return String(content)\n}\n\nexport interface MonitoringEventPropertiesWithDefaults {\n  distinctId?: string\n  traceId: string\n  properties?: Record<string, any>\n  privacyMode: boolean\n  groups?: Record<string, any>\n  modelOverride?: string\n  providerOverride?: string\n  costOverride?: CostOverride\n  captureImmediate?: boolean\n}\n\nexport type MonitoringEventProperties = Partial<MonitoringEventPropertiesWithDefaults>\n\nexport type MonitoringParams = {\n  [K in keyof MonitoringEventProperties as `posthog${Capitalize<string & K>}`]: MonitoringEventProperties[K]\n}\n\nexport interface CostOverride {\n  inputCost: number\n  outputCost: number\n}\n\nexport const getModelParams = (\n  params:\n    | ((\n        | ChatCompletionCreateParamsBase\n        | MessageCreateParams\n        | ResponseCreateParams\n        | ResponseCreateParamsWithTools\n        | EmbeddingCreateParams\n        | TranscriptionCreateParams\n      ) &\n        MonitoringParams)\n    | null\n): Record<string, any> => {\n  if (!params) {\n    return {}\n  }\n  const modelParams: Record<string, any> = {}\n  const paramKeys = [\n    'temperature',\n    'max_tokens',\n    'max_completion_tokens',\n    'top_p',\n    'frequency_penalty',\n    'presence_penalty',\n    'n',\n    'stop',\n    'stream',\n    'streaming',\n    'language',\n    'response_format',\n    'timestamp_granularities',\n  ] as const\n\n  for (const key of paramKeys) {\n    if (key in params && (params as any)[key] !== undefined) {\n      modelParams[key] = (params as any)[key]\n    }\n  }\n  return modelParams\n}\n\n/**\n * Helper to format responses (non-streaming) for consumption\n */\nexport const formatResponse = (response: any, provider: string): FormattedMessage[] => {\n  if (!response) {\n    return []\n  }\n  if (provider === 'anthropic') {\n    return formatResponseAnthropic(response)\n  } else if (provider === 'openai') {\n    return formatResponseOpenAI(response)\n  } else if (provider === 'gemini') {\n    return formatResponseGemini(response)\n  }\n  return []\n}\n\nexport const formatResponseAnthropic = (response: any): FormattedMessage[] => {\n  const output: FormattedMessage[] = []\n  const content: FormattedContent = []\n\n  for (const choice of response.content ?? []) {\n    if (choice?.type === 'text' && choice?.text) {\n      content.push({ type: 'text', text: choice.text })\n    } else if (choice?.type === 'tool_use' && choice?.name && choice?.id) {\n      content.push({\n        type: 'function',\n        id: choice.id,\n        function: {\n          name: choice.name,\n          arguments: choice.input || {},\n        },\n      })\n    }\n  }\n\n  if (content.length > 0) {\n    output.push({\n      role: 'assistant',\n      content,\n    })\n  }\n\n  return output\n}\n\nexport const formatResponseOpenAI = (response: any): FormattedMessage[] => {\n  const output: FormattedMessage[] = []\n\n  if (response.choices) {\n    for (const choice of response.choices) {\n      const content: FormattedContent = []\n      let role = 'assistant'\n\n      if (choice.message) {\n        if (choice.message.role) {\n          role = choice.message.role\n        }\n\n        if (choice.message.content) {\n          content.push({ type: 'text', text: choice.message.content })\n        }\n\n        if (choice.message.tool_calls) {\n          for (const toolCall of choice.message.tool_calls) {\n            content.push({\n              type: 'function',\n              id: toolCall.id,\n              function: {\n                name: toolCall.function.name,\n                arguments: toolCall.function.arguments,\n              },\n            })\n          }\n        }\n\n        // Handle audio output (gpt-4o-audio-preview)\n        if (choice.message.audio) {\n          content.push({\n            type: 'audio',\n            ...choice.message.audio,\n          })\n        }\n      }\n\n      if (content.length > 0) {\n        output.push({\n          role,\n          content,\n        })\n      }\n    }\n  }\n\n  // Handle Responses API format\n  if (response.output) {\n    const content: FormattedContent = []\n    let role = 'assistant'\n\n    for (const item of response.output) {\n      if (item.type === 'message') {\n        role = item.role\n\n        if (item.content && Array.isArray(item.content)) {\n          for (const contentItem of item.content) {\n            if (contentItem.type === 'output_text' && contentItem.text) {\n              content.push({ type: 'text', text: contentItem.text })\n            } else if (contentItem.text) {\n              content.push({ type: 'text', text: contentItem.text })\n            } else if (contentItem.type === 'input_image' && contentItem.image_url) {\n              content.push({\n                type: 'image',\n                image: contentItem.image_url,\n              })\n            }\n          }\n        } else if (item.content) {\n          content.push({ type: 'text', text: String(item.content) })\n        }\n      } else if (item.type === 'function_call') {\n        content.push({\n          type: 'function',\n          id: item.call_id || item.id || '',\n          function: {\n            name: item.name,\n            arguments: item.arguments || {},\n          },\n        })\n      }\n    }\n\n    if (content.length > 0) {\n      output.push({\n        role,\n        content,\n      })\n    }\n  }\n\n  return output\n}\n\nexport const buildInlineDataBlock = (\n  mimeType: string,\n  data: string\n): FormattedAudioContent | FormattedImageContent | FormattedDocumentContent => {\n  if (mimeType.startsWith('audio/')) {\n    return { type: 'audio', mime_type: mimeType, data }\n  }\n  if (mimeType.startsWith('image/')) {\n    return { type: 'image', inline_data: { mime_type: mimeType, data } }\n  }\n  return { type: 'document', inline_data: { mime_type: mimeType, data } }\n}\n\nexport const formatResponseGemini = (response: any): FormattedMessage[] => {\n  const output: FormattedMessage[] = []\n\n  if (response.candidates && Array.isArray(response.candidates)) {\n    for (const candidate of response.candidates) {\n      if (candidate.content && candidate.content.parts) {\n        const content: FormattedContent = []\n\n        for (const part of candidate.content.parts) {\n          if (part.text) {\n            content.push({ type: 'text', text: part.text })\n          } else if (part.functionCall) {\n            content.push({\n              type: 'function',\n              function: {\n                name: part.functionCall.name,\n                arguments: part.functionCall.args,\n              },\n            })\n          } else if (part.inlineData) {\n            // Handle inline data (images, audio, documents)\n            const mimeType = part.inlineData.mimeType || part.inlineData.mime_type || 'application/octet-stream'\n            let data = part.inlineData.data\n\n            // Handle binary data (Uint8Array/Buffer -> base64)\n            if (data instanceof Uint8Array) {\n              if (typeof Buffer !== 'undefined') {\n                data = Buffer.from(data).toString('base64')\n              } else {\n                let binary = ''\n                for (let i = 0; i < data.length; i++) {\n                  binary += String.fromCharCode(data[i])\n                }\n                data = btoa(binary)\n              }\n            }\n\n            // Sanitize base64 data for images and other large inline data\n            data = redactBase64DataUrl(data)\n\n            content.push(buildInlineDataBlock(mimeType, data))\n          }\n        }\n\n        if (content.length > 0) {\n          output.push({\n            role: 'assistant',\n            content,\n          })\n        }\n      } else if (candidate.text) {\n        output.push({\n          role: 'assistant',\n          content: [{ type: 'text', text: candidate.text }],\n        })\n      }\n    }\n  } else if (response.text) {\n    output.push({\n      role: 'assistant',\n      content: [{ type: 'text', text: response.text }],\n    })\n  }\n\n  return output\n}\n\nexport const mergeSystemPrompt = (params: MessageCreateParams & MonitoringParams, provider: string): any => {\n  if (provider == 'anthropic') {\n    const messages = params.messages || []\n    if (!(params as any).system) {\n      return messages\n    }\n    const systemMessage = (params as any).system\n    return [{ role: 'system', content: systemMessage }, ...messages]\n  }\n  return params.messages\n}\n\nexport const withPrivacyMode = (client: PostHog, privacyMode: boolean, input: any): any => {\n  return (client as any).privacy_mode || privacyMode ? null : input\n}\n\nfunction toSafeString(input: unknown): string {\n  if (input === undefined || input === null) {\n    return ''\n  }\n  if (typeof input === 'string') {\n    return input\n  }\n  try {\n    return JSON.stringify(input)\n  } catch {\n    console.warn('Failed to stringify input', input)\n    return ''\n  }\n}\n\nexport const truncate = (input: unknown): string => {\n  const str = toSafeString(input)\n  if (str === '') {\n    return ''\n  }\n\n  // Check if we need to truncate and ensure STRING_FORMAT is respected\n  const buffer = sharedTextEncoder.encode(str)\n  if (buffer.length <= MAX_OUTPUT_SIZE) {\n    // Ensure STRING_FORMAT is respected\n    return sharedTextDecoder.decode(buffer)\n  }\n\n  // Truncate the buffer and ensure a valid string is returned.\n  // fatal: false means we get U+FFFD at the end if truncation broke the encoding.\n  const truncatedBuffer = buffer.slice(0, MAX_OUTPUT_SIZE)\n  let truncatedStr = sharedTextDecoder.decode(truncatedBuffer)\n  if (truncatedStr.endsWith('\\uFFFD')) {\n    truncatedStr = truncatedStr.slice(0, -1)\n  }\n  return `${truncatedStr}... [truncated]`\n}\n\n/**\n * Calculate web search count from raw API response.\n *\n * Uses a two-tier detection strategy:\n * Priority 1 (Exact Count): Count actual web search calls when available\n * Priority 2 (Binary Detection): Return 1 if web search indicators are present, 0 otherwise\n *\n * @param result - Raw API response from any provider (OpenAI, Perplexity, OpenRouter, Gemini, etc.)\n * @returns Number of web searches performed (exact count or binary 1/0)\n */\nexport function calculateWebSearchCount(result: unknown): number {\n  if (!result || typeof result !== 'object') {\n    return 0\n  }\n\n  // Priority 1: Exact Count\n  // Check for OpenAI Responses API web_search_call items\n  if ('output' in result && Array.isArray(result.output)) {\n    let count = 0\n\n    for (const item of result.output) {\n      if (typeof item === 'object' && item !== null && 'type' in item && item.type === 'web_search_call') {\n        count++\n      }\n    }\n\n    if (count > 0) {\n      return count\n    }\n  }\n\n  // Priority 2: Binary Detection (1 or 0)\n\n  // Check for citations at root level (Perplexity)\n  if ('citations' in result && Array.isArray(result.citations) && result.citations.length > 0) {\n    return 1\n  }\n\n  // Check for search_results at root level (Perplexity via OpenRouter)\n  if ('search_results' in result && Array.isArray(result.search_results) && result.search_results.length > 0) {\n    return 1\n  }\n\n  // Check for usage.search_context_size (Perplexity via OpenRouter)\n  if ('usage' in result && typeof result.usage === 'object' && result.usage !== null) {\n    if ('search_context_size' in result.usage && result.usage.search_context_size) {\n      return 1\n    }\n  }\n\n  // Check for annotations with url_citation in choices[].message or choices[].delta (OpenAI/Perplexity)\n  if ('choices' in result && Array.isArray(result.choices)) {\n    for (const choice of result.choices) {\n      if (typeof choice === 'object' && choice !== null) {\n        // Check both message (non-streaming) and delta (streaming) for annotations\n        const content = ('message' in choice ? choice.message : null) || ('delta' in choice ? choice.delta : null)\n\n        if (typeof content === 'object' && content !== null && 'annotations' in content) {\n          const annotations = content.annotations\n\n          if (Array.isArray(annotations)) {\n            const hasUrlCitation = annotations.some((ann: unknown) => {\n              return typeof ann === 'object' && ann !== null && 'type' in ann && ann.type === 'url_citation'\n            })\n\n            if (hasUrlCitation) {\n              return 1\n            }\n          }\n        }\n      }\n    }\n  }\n\n  // Check for annotations in output[].content[] (OpenAI Responses API)\n  if ('output' in result && Array.isArray(result.output)) {\n    for (const item of result.output) {\n      if (typeof item === 'object' && item !== null && 'content' in item) {\n        const content = item.content\n\n        if (Array.isArray(content)) {\n          for (const contentItem of content) {\n            if (typeof contentItem === 'object' && contentItem !== null && 'annotations' in contentItem) {\n              const annotations = contentItem.annotations\n\n              if (Array.isArray(annotations)) {\n                const hasUrlCitation = annotations.some((ann: unknown) => {\n                  return typeof ann === 'object' && ann !== null && 'type' in ann && ann.type === 'url_citation'\n                })\n\n                if (hasUrlCitation) {\n                  return 1\n                }\n              }\n            }\n          }\n        }\n      }\n    }\n  }\n\n  // Check for grounding_metadata (Gemini)\n  if ('candidates' in result && Array.isArray(result.candidates)) {\n    for (const candidate of result.candidates) {\n      if (\n        typeof candidate === 'object' &&\n        candidate !== null &&\n        'grounding_metadata' in candidate &&\n        candidate.grounding_metadata\n      ) {\n        return 1\n      }\n    }\n  }\n\n  return 0\n}\n\n/**\n * Extract available tool calls from the request parameters.\n * These are the tools provided to the LLM, not the tool calls in the response.\n */\nexport const extractAvailableToolCalls = (\n  provider: string,\n  params: any\n): ChatCompletionTool[] | AnthropicTool[] | GeminiTool[] | null => {\n  if (provider === 'anthropic') {\n    if (params.tools) {\n      return params.tools\n    }\n\n    return null\n  } else if (provider === 'gemini') {\n    if (params.config && params.config.tools) {\n      return params.config.tools\n    }\n\n    return null\n  } else if (provider === 'openai') {\n    if (params.tools) {\n      return params.tools\n    }\n\n    return null\n  } else if (provider === 'vercel') {\n    if (params.tools) {\n      return params.tools\n    }\n\n    return null\n  }\n\n  return null\n}\n\nexport enum AIEvent {\n  Generation = '$ai_generation',\n  Embedding = '$ai_embedding',\n}\n\nexport function sanitizeValues(obj: any): any {\n  if (obj === undefined || obj === null) {\n    return obj\n  }\n  const jsonSafe = JSON.parse(JSON.stringify(obj))\n  if (typeof jsonSafe === 'string') {\n    // Sanitize lone surrogates by round-tripping through UTF-8\n    return new TextDecoder().decode(new TextEncoder().encode(jsonSafe))\n  } else if (Array.isArray(jsonSafe)) {\n    return jsonSafe.map(sanitizeValues)\n  } else if (jsonSafe && typeof jsonSafe === 'object') {\n    return Object.fromEntries(Object.entries(jsonSafe).map(([k, v]) => [k, sanitizeValues(v)]))\n  }\n  return jsonSafe\n}\n\nconst POSTHOG_PARAMS_MAP: Record<keyof MonitoringParams, string> = {\n  posthogDistinctId: 'distinctId',\n  posthogTraceId: 'traceId',\n  posthogProperties: 'properties',\n  posthogPrivacyMode: 'privacyMode',\n  posthogGroups: 'groups',\n  posthogModelOverride: 'modelOverride',\n  posthogProviderOverride: 'providerOverride',\n  posthogCostOverride: 'costOverride',\n  posthogCaptureImmediate: 'captureImmediate',\n}\n\nexport function extractPosthogParams<T>(body: T & MonitoringParams): {\n  providerParams: T\n  posthogParams: MonitoringEventPropertiesWithDefaults\n} {\n  const providerParams: Record<string, unknown> = {}\n  const posthogParams: Record<string, unknown> = {}\n\n  for (const [key, value] of Object.entries(body)) {\n    if (POSTHOG_PARAMS_MAP[key as keyof MonitoringParams]) {\n      posthogParams[POSTHOG_PARAMS_MAP[key as keyof MonitoringParams]] = value\n    } else if (key.startsWith('posthog')) {\n      console.warn(`Unknown Posthog parameter ${key}`)\n    } else {\n      providerParams[key] = value\n    }\n  }\n\n  return {\n    providerParams: providerParams as T,\n    posthogParams: addDefaults(posthogParams),\n  }\n}\n\nfunction addDefaults(params: MonitoringEventProperties): MonitoringEventPropertiesWithDefaults {\n  return {\n    ...params,\n    privacyMode: params.privacyMode ?? false,\n    traceId: params.traceId ?? uuidv4(),\n  }\n}\n\nexport function formatOpenAIResponsesInput(input: unknown, instructions?: string | null): FormattedMessage[] {\n  const messages: FormattedMessage[] = []\n\n  if (instructions) {\n    messages.push({\n      role: 'system',\n      content: instructions,\n    })\n  }\n\n  if (Array.isArray(input)) {\n    for (const item of input) {\n      if (typeof item === 'string') {\n        messages.push({ role: 'user', content: item })\n      } else if (item && typeof item === 'object') {\n        const obj = item as Record<string, unknown>\n        const role = isString(obj.role) ? obj.role : 'user'\n\n        // Handle content properly - preserve structure for objects/arrays\n        const content = obj.content ?? obj.text ?? item\n        messages.push({ role, content: toContentString(content) })\n      } else {\n        messages.push({ role: 'user', content: toContentString(item) })\n      }\n    }\n  } else if (typeof input === 'string') {\n    messages.push({ role: 'user', content: input })\n  } else if (input) {\n    messages.push({ role: 'user', content: toContentString(input) })\n  }\n\n  return messages\n}\n","import { sanitizeValues } from './utils'\n\nconst DEFAULT_MAX_DEPTH = 3\nconst MAX_STACK_LINES = 20\n\nexport function serializeError(value: unknown, depth = DEFAULT_MAX_DEPTH): unknown {\n  if (depth < 0 || value === null || typeof value !== 'object') {\n    return value\n  }\n  if (value instanceof Error) {\n    const out: Record<string, unknown> = {\n      name: value.name,\n      message: value.message,\n      stack: truncateStack(value.stack),\n    }\n    for (const key of Object.keys(value)) {\n      out[key] = serializeError((value as unknown as Record<string, unknown>)[key], depth - 1)\n    }\n    if (value.cause !== undefined) {\n      out.cause = serializeError(value.cause, depth - 1)\n    }\n    return out\n  }\n  if (Array.isArray(value)) {\n    return value.map((item) => serializeError(item, depth - 1))\n  }\n  return value\n}\n\nexport function stringifyError(error: unknown): string {\n  try {\n    return JSON.stringify(sanitizeValues(serializeError(error)))\n  } catch {\n    if (error instanceof Error) {\n      return JSON.stringify({ name: error.name, message: error.message })\n    }\n    return JSON.stringify({ message: String(error) })\n  }\n}\n\nfunction truncateStack(stack: string | undefined): string | undefined {\n  if (!stack) {\n    return stack\n  }\n  const lines = stack.split('\\n')\n  if (lines.length <= MAX_STACK_LINES) {\n    return stack\n  }\n  return [...lines.slice(0, MAX_STACK_LINES), '... (truncated)'].join('\\n')\n}\n","import { EventMessage, PostHog } from 'posthog-node'\nimport type { ChatCompletionTool } from 'openai/resources/chat/completions'\nimport type { Tool as GeminiTool } from '@google/genai'\nimport AnthropicOriginal from '@anthropic-ai/sdk'\nimport { v4 as uuidv4 } from 'uuid'\nimport { uuidv7, ErrorTracking as CoreErrorTracking } from '@posthog/core'\nimport { version } from '../package.json'\nimport type { TokenUsage } from './types'\nimport { stringifyError } from './serializeError'\nimport { AIEvent, CostOverride, getTokensSource, sanitizeValues, withPrivacyMode } from './utils'\n\ntype AnthropicTool = AnthropicOriginal.Tool\n\n/**\n * Options for `captureAiGeneration`. Mirrors the `$ai_generation` event shape\n * directly so that any caller — first-party SDK wrappers and external code\n * alike — produces an identical event.\n */\nexport interface CaptureAiGenerationOptions {\n  distinctId?: string\n  /** Auto-generated when omitted. */\n  traceId?: string\n  /** Defaults to `$ai_generation`. */\n  eventType?: AIEvent\n\n  /** Required for the event to be useful, but accepted as optional so SDK wrappers can pass through whatever they detect. */\n  model?: string\n  provider: string\n  input: unknown\n  output: unknown\n\n  /** Maps to `$ai_model_parameters` (temperature, max_tokens, top_p, …). */\n  modelParameters?: Record<string, unknown>\n\n  baseURL?: string\n  httpStatus?: number\n  /** Wall-clock latency in seconds. */\n  latency?: number\n  /** Time from request start to the first streamed token, in seconds. */\n  timeToFirstToken?: number\n\n  usage?: TokenUsage\n\n  /** Extra event properties merged into the captured event. */\n  properties?: Record<string, unknown>\n  /** Mapping of group type to group id, matching `EventMessage.groups`. */\n  groups?: Record<string, string | number>\n  privacyMode?: boolean\n\n  /**\n   * For SDK wrappers: overrides the auto-detected model. External callers\n   * should pass `model` directly instead.\n   */\n  modelOverride?: string\n  /**\n   * For SDK wrappers: overrides the auto-detected provider. External callers\n   * should pass `provider` directly instead.\n   */\n  providerOverride?: string\n  costOverride?: CostOverride\n\n  tools?: ChatCompletionTool[] | AnthropicTool[] | GeminiTool[] | null\n  stopReason?: string\n  /** When set, the event is captured as an error. */\n  error?: unknown\n\n  /** Awaits delivery instead of batching. Useful in serverless environments. */\n  captureImmediate?: boolean\n}\n\n/**\n * Capture an `$ai_generation` (or `$ai_embedding`) event to PostHog.\n *\n * This is the canonical primitive that every `@posthog/ai` wrapper\n * (`withTracing`, `OpenAI`, `Anthropic`, `GoogleGenAI`, …) funnels through, so\n * external code can use it directly to instrument LLM calls made through\n * arbitrary clients (Cloudflare Workers AI, custom HTTP, etc.) and get the\n * same events the SDK wrappers produce.\n *\n * When `error` is set, the event is captured as an error. If the error is an\n * object, it is mutated in place to set `__posthog_previously_captured_error`\n * so callers can re-throw the original error reference safely.\n */\nexport const captureAiGeneration = async (client: PostHog, options: CaptureAiGenerationOptions): Promise<void> => {\n  if (!client.capture) {\n    return\n  }\n\n  const traceId = options.traceId ?? uuidv4()\n  const eventType = options.eventType ?? AIEvent.Generation\n  const privacyMode = options.privacyMode ?? false\n  const usage = options.usage ?? {}\n\n  const safeInput = sanitizeValues(options.input)\n  const safeOutput = sanitizeValues(options.output)\n\n  let httpStatus = options.httpStatus\n  let errorData: Record<string, unknown> = {}\n  if (options.error) {\n    if (httpStatus === undefined) {\n      if (typeof options.error === 'object' && 'status' in options.error && typeof options.error.status === 'number') {\n        httpStatus = options.error.status\n      } else {\n        httpStatus = 500\n      }\n    }\n\n    let exceptionId: string | undefined\n    if (client.options?.enableExceptionAutocapture) {\n      exceptionId = uuidv7()\n      client.captureException(options.error, undefined, { $ai_trace_id: traceId }, exceptionId)\n      if (typeof options.error === 'object') {\n        ;(options.error as CoreErrorTracking.PreviouslyCapturedError).__posthog_previously_captured_error = true\n      }\n    }\n\n    errorData = {\n      $ai_is_error: true,\n      $ai_error: stringifyError(options.error),\n      $exception_event_id: exceptionId,\n    }\n  }\n  httpStatus = httpStatus ?? 200\n\n  let costOverrideData: Record<string, number> = {}\n  if (options.costOverride) {\n    const inputCostUSD = (options.costOverride.inputCost ?? 0) * (usage.inputTokens ?? 0)\n    const outputCostUSD = (options.costOverride.outputCost ?? 0) * (usage.outputTokens ?? 0)\n    costOverrideData = {\n      $ai_input_cost_usd: inputCostUSD,\n      $ai_output_cost_usd: outputCostUSD,\n      $ai_total_cost_usd: inputCostUSD + outputCostUSD,\n    }\n  }\n\n  const additionalTokenValues = {\n    ...(usage.reasoningTokens ? { $ai_reasoning_tokens: usage.reasoningTokens } : {}),\n    ...(usage.cacheReadInputTokens ? { $ai_cache_read_input_tokens: usage.cacheReadInputTokens } : {}),\n    ...(usage.cacheCreationInputTokens ? { $ai_cache_creation_input_tokens: usage.cacheCreationInputTokens } : {}),\n    ...(usage.webSearchCount ? { $ai_web_search_count: usage.webSearchCount } : {}),\n    ...(usage.rawUsage ? { $ai_usage: usage.rawUsage } : {}),\n  }\n\n  const properties: Record<string, unknown> = {\n    $ai_lib: 'posthog-ai',\n    $ai_lib_version: version,\n    $ai_provider: options.providerOverride ?? options.provider,\n    $ai_model: options.modelOverride ?? options.model,\n    $ai_model_parameters: options.modelParameters ?? {},\n    $ai_input: withPrivacyMode(client, privacyMode, safeInput),\n    $ai_output_choices: withPrivacyMode(client, privacyMode, safeOutput),\n    $ai_http_status: httpStatus,\n    $ai_input_tokens: usage.inputTokens ?? 0,\n    ...(usage.outputTokens !== undefined ? { $ai_output_tokens: usage.outputTokens } : {}),\n    ...additionalTokenValues,\n    $ai_latency: options.latency ?? 0,\n    ...(options.timeToFirstToken !== undefined ? { $ai_time_to_first_token: options.timeToFirstToken } : {}),\n    $ai_trace_id: traceId,\n    $ai_base_url: options.baseURL ?? '',\n    ...options.properties,\n    $ai_tokens_source: getTokensSource(options.properties),\n    ...(options.distinctId ? {} : { $process_person_profile: false }),\n    ...(options.stopReason ? { $ai_stop_reason: options.stopReason } : {}),\n    ...(options.tools ? { $ai_tools: options.tools } : {}),\n    ...errorData,\n    ...costOverrideData,\n  }\n\n  const event: EventMessage = {\n    distinctId: options.distinctId ?? traceId,\n    event: eventType,\n    properties,\n    groups: options.groups,\n  }\n\n  if (options.captureImmediate) {\n    await client.captureImmediate(event)\n  } else {\n    client.capture(event)\n  }\n}\n","import type {\n  LanguageModelV2,\n  LanguageModelV2CallOptions,\n  LanguageModelV2Content,\n  LanguageModelV2Prompt,\n  LanguageModelV2StreamPart,\n  LanguageModelV3,\n  LanguageModelV3CallOptions,\n  LanguageModelV3Content,\n  LanguageModelV3Prompt,\n  LanguageModelV3StreamPart,\n} from '@ai-sdk/provider'\nimport { v4 as uuidv4 } from 'uuid'\nimport { PostHog } from 'posthog-node'\nimport {\n  CostOverride,\n  truncate,\n  MAX_OUTPUT_SIZE,\n  utf8ByteLength,\n  extractAvailableToolCalls,\n  toContentString,\n  calculateWebSearchCount,\n  getModelParams,\n} from '../utils'\nimport { captureAiGeneration } from '../captureAiGeneration'\nimport { redactBase64DataUrl } from '../sanitization'\nimport { isString } from '../typeGuards'\n\n// Union types for dual version support\ntype LanguageModel = LanguageModelV2 | LanguageModelV3\ntype LanguageModelCallOptions = LanguageModelV2CallOptions | LanguageModelV3CallOptions\ntype LanguageModelPrompt = LanguageModelV2Prompt | LanguageModelV3Prompt\ntype LanguageModelContent = LanguageModelV2Content | LanguageModelV3Content\ntype LanguageModelStreamPart = LanguageModelV2StreamPart | LanguageModelV3StreamPart\n\n// Type guards\nfunction isV3Model(model: LanguageModel): model is LanguageModelV3 {\n  return model.specificationVersion === 'v3'\n}\n\nfunction isV2Model(model: LanguageModel): model is LanguageModelV2 {\n  return model.specificationVersion === 'v2'\n}\n\ninterface ClientOptions {\n  posthogDistinctId?: string\n  posthogTraceId?: string\n  posthogProperties?: Record<string, any>\n  posthogPrivacyMode?: boolean\n  posthogGroups?: Record<string, any>\n  posthogModelOverride?: string\n  posthogProviderOverride?: string\n  posthogCostOverride?: CostOverride\n  posthogCaptureImmediate?: boolean\n}\n\ninterface PostHogInput {\n  role: string\n  type?: string\n  content?:\n    | string\n    | {\n        [key: string]: any\n      }\n}\n\n// Content types for the output array\ntype OutputContentItem =\n  | { type: 'text'; text: string }\n  | { type: 'reasoning'; text: string }\n  | { type: 'tool-call'; id: string; function: { name: string; arguments: string } }\n  | { type: 'file'; name: string; mediaType: string; data: string }\n  | { type: 'source'; sourceType: string; id: string; url: string; title: string }\n\nconst mapVercelParams = (params: any): Record<string, any> => {\n  return {\n    temperature: params.temperature,\n    max_output_tokens: params.maxOutputTokens,\n    top_p: params.topP,\n    frequency_penalty: params.frequencyPenalty,\n    presence_penalty: params.presencePenalty,\n    stop: params.stopSequences,\n    stream: params.stream,\n  }\n}\n\nconst mapVercelPrompt = (messages: LanguageModelPrompt): PostHogInput[] => {\n  // Map and truncate individual content\n  const inputs: PostHogInput[] = messages.map((message) => {\n    let content: any\n\n    // Handle system role which has string content\n    if (message.role === 'system') {\n      content = [\n        {\n          type: 'text',\n          text: truncate(toContentString(message.content)),\n        },\n      ]\n    } else {\n      // Handle other roles which have array content\n      if (Array.isArray(message.content)) {\n        content = message.content.map((c: any) => {\n          if (c.type === 'text') {\n            return {\n              type: 'text',\n              text: truncate(c.text),\n            }\n          } else if (c.type === 'file') {\n            // For file type, check if it's a data URL and redact if needed\n            let fileData: string\n\n            const contentData: unknown = c.data\n\n            if (contentData instanceof URL) {\n              fileData = contentData.toString()\n            } else if (isString(contentData)) {\n              // Redact base64 data URLs and raw base64 to prevent oversized events\n              fileData = redactBase64DataUrl(contentData)\n            } else {\n              fileData = 'raw files not supported'\n            }\n\n            return {\n              type: 'file',\n              file: fileData,\n              mediaType: c.mediaType,\n            }\n          } else if (c.type === 'reasoning') {\n            return {\n              type: 'reasoning',\n              text: truncate(c.reasoning),\n            }\n          } else if (c.type === 'tool-call') {\n            return {\n              type: 'tool-call',\n              toolCallId: c.toolCallId,\n              toolName: c.toolName,\n              input: c.input,\n            }\n          } else if (c.type === 'tool-result') {\n            return {\n              type: 'tool-result',\n              toolCallId: c.toolCallId,\n              toolName: c.toolName,\n              output: c.output,\n              isError: c.isError,\n            }\n          }\n          return {\n            type: 'text',\n            text: '',\n          }\n        })\n      } else {\n        // Fallback for non-array content\n        content = [\n          {\n            type: 'text',\n            text: truncate(toContentString(message.content)),\n          },\n        ]\n      }\n    }\n\n    return {\n      role: message.role,\n      content,\n    }\n  })\n\n  try {\n    // Trim the inputs array until its serialized JSON size fits within MAX_OUTPUT_SIZE.\n    // Pre-compute each message's byte size once so we can shift by accumulated budget\n    // in a single linear pass, instead of re-stringifying the whole array per iteration.\n    const messageSizes = inputs.map((m) => utf8ByteLength(JSON.stringify(m)))\n    // Account for the surrounding `[` `]` plus a comma between each pair of elements.\n    let totalBytes = 2 + Math.max(0, messageSizes.length - 1)\n    for (const size of messageSizes) {\n      totalBytes += size\n    }\n    let removedCount = 0\n    while (totalBytes > MAX_OUTPUT_SIZE && removedCount < messageSizes.length) {\n      totalBytes -= messageSizes[removedCount]\n      // Each removed message past the first also drops the comma that joined it.\n      if (removedCount < messageSizes.length - 1) {\n        totalBytes -= 1\n      }\n      removedCount++\n    }\n    if (removedCount > 0) {\n      inputs.splice(0, removedCount)\n      // Add one placeholder to indicate how many were removed\n      inputs.unshift({\n        role: 'posthog',\n        content: `[${removedCount} message${removedCount === 1 ? '' : 's'} removed due to size limit]`,\n      })\n    }\n  } catch (error) {\n    console.error('Error stringifying inputs', error)\n    return [{ role: 'posthog', content: 'An error occurred while processing your request. Please try again.' }]\n  }\n  return inputs\n}\n\nconst mapVercelOutput = (result: LanguageModelContent[]): PostHogInput[] => {\n  const content: OutputContentItem[] = result.map((item) => {\n    if (item.type === 'text') {\n      return { type: 'text', text: truncate(item.text) }\n    }\n    if (item.type === 'tool-call') {\n      const toolCall = item as { input?: unknown; args?: unknown; arguments?: unknown }\n      const rawArgs = toolCall.input ?? toolCall.args ?? toolCall.arguments ?? {}\n      return {\n        type: 'tool-call',\n        id: item.toolCallId,\n        function: {\n          name: item.toolName,\n          arguments: typeof rawArgs === 'string' ? rawArgs : JSON.stringify(rawArgs),\n        },\n      }\n    }\n    if (item.type === 'reasoning') {\n      return { type: 'reasoning', text: truncate(item.text) }\n    }\n    if (item.type === 'file') {\n      // Handle files similar to input mapping - avoid large base64 data\n      let fileData: string\n      if (item.data instanceof URL) {\n        fileData = item.data.toString()\n      } else if (typeof item.data === 'string') {\n        fileData = redactBase64DataUrl(item.data)\n\n        // If not redacted and still large, replace with size indicator\n        if (fileData === item.data && item.data.length > 1000) {\n          fileData = `[${item.mediaType} file - ${item.data.length} bytes]`\n        }\n      } else {\n        fileData = `[binary ${item.mediaType} file]`\n      }\n\n      return {\n        type: 'file',\n        name: 'generated_file',\n        mediaType: item.mediaType,\n        data: fileData,\n      }\n    }\n    if (item.type === 'source') {\n      return {\n        type: 'source',\n        sourceType: item.sourceType,\n        id: item.id,\n        url: (item as any).url || '',\n        title: item.title || '',\n      }\n    }\n    // Fallback for unknown types - try to extract text if possible\n    return { type: 'text', text: truncate(JSON.stringify(item)) }\n  })\n\n  if (content.length > 0) {\n    return [\n      {\n        role: 'assistant',\n        content: content.length === 1 && content[0].type === 'text' ? content[0].text : content,\n      },\n    ]\n  }\n  // otherwise stringify and truncate\n  try {\n    const jsonOutput = JSON.stringify(result)\n    return [{ content: truncate(jsonOutput), role: 'assistant' }]\n  } catch {\n    console.error('Error stringifying output')\n    return []\n  }\n}\n\nconst extractProvider = (model: LanguageModel): string => {\n  const provider = model.provider.toLowerCase()\n  const providerName = provider.split('.')[0]\n  return providerName\n}\n\n// Extract web search count from provider metadata (works for both V2 and V3)\nconst extractWebSearchCount = (providerMetadata: unknown, usage: any): number => {\n  // Try Anthropic-specific extraction\n  if (\n    providerMetadata &&\n    typeof providerMetadata === 'object' &&\n    'anthropic' in providerMetadata &&\n    providerMetadata.anthropic &&\n    typeof providerMetadata.anthropic === 'object' &&\n    'server_tool_use' in providerMetadata.anthropic\n  ) {\n    const serverToolUse = providerMetadata.anthropic.server_tool_use\n    if (\n      serverToolUse &&\n      typeof serverToolUse === 'object' &&\n      'web_search_requests' in serverToolUse &&\n      typeof serverToolUse.web_search_requests === 'number'\n    ) {\n      return serverToolUse.web_search_requests\n    }\n  }\n\n  // Fall back to generic calculation\n  return calculateWebSearchCount({\n    usage,\n    providerMetadata,\n  })\n}\n\n// Helper to extract numeric token value from V2 (number) or V3 (object with .total) usage formats\nconst extractTokenCount = (value: unknown): number | undefined => {\n  if (typeof value === 'number') {\n    return value\n  }\n  if (\n    value &&\n    typeof value === 'object' &&\n    'total' in value &&\n    typeof (value as { total: unknown }).total === 'number'\n  ) {\n    return (value as { total: number }).total\n  }\n  return undefined\n}\n\n// Helper to extract reasoning tokens from V2 (usage.reasoningTokens) or V3 (usage.outputTokens.reasoning)\nconst extractReasoningTokens = (usage: Record<string, unknown>): unknown => {\n  // V2 style: top-level reasoningTokens\n  if ('reasoningTokens' in usage) {\n    return usage.reasoningTokens\n  }\n  // V3 style: nested in outputTokens.reasoning\n  if (\n    'outputTokens' in usage &&\n    usage.outputTokens &&\n    typeof usage.outputTokens === 'object' &&\n    'reasoning' in usage.outputTokens\n  ) {\n    return (usage.outputTokens as { reasoning: unknown }).reasoning\n  }\n  return undefined\n}\n\n// Helper to extract cached input tokens from V2 (usage.cachedInputTokens) or V3 (usage.inputTokens.cacheRead)\nconst extractCacheReadTokens = (usage: Record<string, unknown>): unknown => {\n  // V2 style: top-level cachedInputTokens\n  if ('cachedInputTokens' in usage) {\n    return usage.cachedInputTokens\n  }\n  // V3 style: nested in inputTokens.cacheRead\n  if (\n    'inputTokens' in usage &&\n    usage.inputTokens &&\n    typeof usage.inputTokens === 'object' &&\n    'cacheRead' in usage.inputTokens\n  ) {\n    return (usage.inputTokens as { cacheRead: unknown }).cacheRead\n  }\n  return undefined\n}\n\n// Helper to extract cache write tokens from V3 (usage.inputTokens.cacheWrite). Providers like\n// Amazon Bedrock populate this standardized field instead of providerMetadata.anthropic.\nconst extractCacheWriteTokens = (usage: Record<string, unknown>): unknown => {\n  if (\n    'inputTokens' in usage &&\n    usage.inputTokens &&\n    typeof usage.inputTokens === 'object' &&\n    'cacheWrite' in usage.inputTokens\n  ) {\n    return (usage.inputTokens as { cacheWrite: unknown }).cacheWrite\n  }\n  return undefined\n}\n\n// Extract additional token values from provider metadata, with a V3 standardized fallback\n// (e.g. Amazon Bedrock exposes cache write tokens via usage.inputTokens.cacheWrite rather\n// than providerMetadata.anthropic.cacheCreationInputTokens). A cacheWrite of 0 is treated\n// as absent so we preserve the pre-fallback event shape on providers that simply omit the\n// field — consumers downstream saw `$ai_cache_creation_input_tokens` missing, not 0.\nconst extractAdditionalTokenValues = (providerMetadata: unknown, usage: unknown): Record<string, any> => {\n  if (\n    providerMetadata &&\n    typeof providerMetadata === 'object' &&\n    'anthropic'