@elizaos/plugin-openai 2.0.3-beta.4 → 2.0.3-beta.5

package/dist/browser/index.browser.js.map

@@ -11,7 +11,7 @@
     "import type { IAgentRuntime, ModelTypeName } from \"@elizaos/core\";\nimport { EventType } from \"@elizaos/core\";\nimport type { TokenUsage } from \"../types\";\n\nconst MAX_PROMPT_LENGTH = 200;\n\ninterface ModelUsageEventPayload {\n  runtime: IAgentRuntime;\n  source: \"openai\";\n  provider: \"openai\";\n  type: ModelTypeName;\n  prompt: string;\n  tokens: {\n    prompt: number;\n    completion: number;\n    total: number;\n    cached?: number;\n  };\n}\n\ninterface AISDKUsage {\n  inputTokens?: number;\n  outputTokens?: number;\n  totalTokens?: number;\n  cachedInputTokens?: number;\n}\n\ninterface OpenAIAPIUsage {\n  promptTokens?: number;\n  completionTokens?: number;\n  totalTokens?: number;\n  cachedPromptTokens?: number;\n  promptTokensDetails?: {\n    cachedTokens?: number;\n  };\n}\n\ntype ModelUsage = TokenUsage | AISDKUsage | OpenAIAPIUsage;\n\nfunction truncatePrompt(prompt: string): string {\n  if (prompt.length <= MAX_PROMPT_LENGTH) {\n    return prompt;\n  }\n  return `${prompt.slice(0, MAX_PROMPT_LENGTH)}…`;\n}\n\nfunction normalizeUsage(usage: ModelUsage): TokenUsage {\n  if (\"promptTokens\" in usage) {\n    const promptTokensDetails =\n      \"promptTokensDetails\" in usage ? usage.promptTokensDetails : undefined;\n    const cachedPromptTokens = usage.cachedPromptTokens ?? promptTokensDetails?.cachedTokens;\n    return {\n      promptTokens: usage.promptTokens ?? 0,\n      completionTokens: usage.completionTokens ?? 0,\n      totalTokens: usage.totalTokens ?? (usage.promptTokens ?? 0) + (usage.completionTokens ?? 0),\n      cachedPromptTokens,\n    };\n  }\n  if (\"inputTokens\" in usage || \"outputTokens\" in usage) {\n    const input = (usage as AISDKUsage).inputTokens ?? 0;\n    const output = (usage as AISDKUsage).outputTokens ?? 0;\n    const total = (usage as AISDKUsage).totalTokens ?? input + output;\n    return {\n      promptTokens: input,\n      completionTokens: output,\n      totalTokens: total,\n      cachedPromptTokens: (usage as AISDKUsage).cachedInputTokens,\n    };\n  }\n  return {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n  };\n}\n\nexport function emitModelUsageEvent(\n  runtime: IAgentRuntime,\n  type: ModelTypeName,\n  prompt: string,\n  usage: ModelUsage\n): void {\n  const normalized = normalizeUsage(usage);\n\n  const payload: ModelUsageEventPayload = {\n    runtime,\n    source: \"openai\",\n    provider: \"openai\",\n    type,\n    prompt: truncatePrompt(prompt),\n    tokens: {\n      prompt: normalized.promptTokens,\n      completion: normalized.completionTokens,\n      total: normalized.totalTokens,\n      ...(normalized.cachedPromptTokens !== undefined\n        ? { cached: normalized.cachedPromptTokens }\n        : {}),\n    },\n  };\n\n  runtime.emitEvent(EventType.MODEL_USED, payload);\n}\n",
     "import type {\n  IAgentRuntime,\n  ImageDescriptionParams,\n  ImageGenerationParams,\n  RecordLlmCallDetails,\n} from \"@elizaos/core\";\nimport { logger, ModelType, recordLlmCall } from \"@elizaos/core\";\nimport type {\n  ImageDescriptionResult,\n  ImageGenerationResult,\n  ImageQuality,\n  ImageSize,\n  ImageStyle,\n  OpenAIChatCompletionResponse,\n  OpenAIImageGenerationResponse,\n} from \"../types\";\nimport {\n  getAuthHeader,\n  getBaseURL,\n  getImageDescriptionAuthHeader,\n  getImageDescriptionBaseURL,\n  getImageDescriptionMaxTokens,\n  getImageDescriptionModel,\n  getImageModel,\n} from \"../utils/config\";\nimport { emitModelUsageEvent } from \"../utils/events\";\n\ninterface ExtendedImageGenerationParams extends ImageGenerationParams {\n  quality?: ImageQuality;\n  style?: ImageStyle;\n}\n\nconst DEFAULT_IMAGE_DESCRIPTION_PROMPT =\n  \"Please analyze this image and provide a title and detailed description.\";\n\nexport async function handleImageGeneration(\n  runtime: IAgentRuntime,\n  params: ImageGenerationParams\n): Promise<ImageGenerationResult[]> {\n  const modelName = getImageModel(runtime);\n  const count = params.count ?? 1;\n  const size: ImageSize = (params.size as ImageSize) ?? \"1024x1024\";\n  const extendedParams = params as ExtendedImageGenerationParams;\n\n  logger.debug(`[OpenAI] Using IMAGE model: ${modelName}`);\n\n  if (typeof params.prompt !== \"string\" || params.prompt.trim().length === 0) {\n    throw new Error(\"IMAGE generation requires a non-empty prompt\");\n  }\n\n  if (count < 1 || count > 10) {\n    throw new Error(\"IMAGE count must be between 1 and 10\");\n  }\n\n  const baseURL = getBaseURL(runtime);\n\n  const requestBody: Record<string, string | number> = {\n    model: modelName,\n    prompt: params.prompt,\n    n: count,\n    size,\n  };\n\n  if (extendedParams.quality) {\n    requestBody.quality = extendedParams.quality;\n  }\n  if (extendedParams.style) {\n    requestBody.style = extendedParams.style;\n  }\n\n  const details: RecordLlmCallDetails = {\n    model: modelName,\n    systemPrompt: \"\",\n    userPrompt: params.prompt,\n    temperature: 0,\n    maxTokens: 0,\n    purpose: \"external_llm\",\n    actionType: \"openai.images.generate\",\n  };\n  const data = await recordLlmCall(runtime, details, async () => {\n    const response = await fetch(`${baseURL}/images/generations`, {\n      method: \"POST\",\n      headers: {\n        ...getAuthHeader(runtime),\n        \"Content-Type\": \"application/json\",\n      },\n      body: JSON.stringify(requestBody),\n    });\n\n    if (!response.ok) {\n      const errorText = await response.text().catch(() => \"Unknown error\");\n      throw new Error(\n        `OpenAI image generation failed: ${response.status} ${response.statusText} - ${errorText}`\n      );\n    }\n\n    const responseData = (await response.json()) as OpenAIImageGenerationResponse;\n    details.response = JSON.stringify(responseData.data);\n    return responseData;\n  });\n\n  if (data.data.length === 0) {\n    throw new Error(\"OpenAI API returned no images\");\n  }\n\n  return data.data.map((item) => ({\n    url: item.url,\n    revisedPrompt: item.revised_prompt,\n  }));\n}\n\nfunction parseTitleFromResponse(content: string): string {\n  const titleMatch = content.match(/title[:\\s]+(.+?)(?:\\n|$)/i);\n  return titleMatch?.[1]?.trim() ?? \"Image Analysis\";\n}\n\nfunction parseDescriptionFromResponse(content: string): string {\n  return content.replace(/title[:\\s]+(.+?)(?:\\n|$)/i, \"\").trim();\n}\n\nexport async function handleImageDescription(\n  runtime: IAgentRuntime,\n  params: ImageDescriptionParams | string\n): Promise<ImageDescriptionResult> {\n  const modelName = getImageDescriptionModel(runtime);\n  const paramsWithMaxTokens = params as ImageDescriptionParams & { maxTokens?: number };\n  const maxTokens =\n    typeof params === \"object\" && typeof paramsWithMaxTokens.maxTokens === \"number\"\n      ? paramsWithMaxTokens.maxTokens\n      : getImageDescriptionMaxTokens(runtime);\n\n  logger.debug(`[OpenAI] Using IMAGE_DESCRIPTION model: ${modelName}`);\n\n  let imageUrl: string;\n  let promptText: string;\n\n  if (typeof params === \"string\") {\n    imageUrl = params;\n    promptText = DEFAULT_IMAGE_DESCRIPTION_PROMPT;\n  } else {\n    imageUrl = params.imageUrl;\n    promptText = params.prompt ?? DEFAULT_IMAGE_DESCRIPTION_PROMPT;\n  }\n\n  if (!imageUrl || imageUrl.trim().length === 0) {\n    throw new Error(\"IMAGE_DESCRIPTION requires a valid image URL\");\n  }\n\n  const baseURL = getImageDescriptionBaseURL(runtime);\n\n  const requestBody = {\n    model: modelName,\n    messages: [\n      {\n        role: \"user\",\n        content: [\n          { type: \"text\", text: promptText },\n          { type: \"image_url\", image_url: { url: imageUrl } },\n        ],\n      },\n    ],\n    max_tokens: maxTokens,\n  };\n\n  const details: RecordLlmCallDetails = {\n    model: modelName,\n    systemPrompt: \"\",\n    userPrompt: promptText,\n    temperature: 0,\n    maxTokens,\n    purpose: \"external_llm\",\n    actionType: \"openai.chat.completions.create\",\n  };\n  const data = await recordLlmCall(runtime, details, async () => {\n    const response = await fetch(`${baseURL}/chat/completions`, {\n      method: \"POST\",\n      headers: {\n        ...getImageDescriptionAuthHeader(runtime),\n        \"Content-Type\": \"application/json\",\n      },\n      body: JSON.stringify(requestBody),\n    });\n\n    if (!response.ok) {\n      const errorText = await response.text().catch(() => \"Unknown error\");\n      throw new Error(\n        `OpenAI image description failed: ${response.status} ${response.statusText} - ${errorText}`\n      );\n    }\n\n    const responseData = (await response.json()) as OpenAIChatCompletionResponse;\n    const responseContent = responseData.choices[0]?.message.content;\n    if (!responseContent) {\n      throw new Error(\"OpenAI API returned empty image description\");\n    }\n    details.response = responseContent;\n    if (responseData.usage) {\n      details.promptTokens = responseData.usage.prompt_tokens;\n      details.completionTokens = responseData.usage.completion_tokens;\n    }\n    return responseData;\n  });\n\n  if (data.usage) {\n    emitModelUsageEvent(\n      runtime,\n      ModelType.IMAGE_DESCRIPTION,\n      typeof params === \"string\" ? params : (params.prompt ?? \"\"),\n      {\n        promptTokens: data.usage.prompt_tokens,\n        completionTokens: data.usage.completion_tokens,\n        totalTokens: data.usage.total_tokens,\n      }\n    );\n  }\n\n  const firstChoice = data.choices[0];\n  const content = firstChoice?.message.content;\n\n  if (!content) {\n    throw new Error(\"OpenAI API returned empty image description\");\n  }\n\n  return {\n    title: parseTitleFromResponse(content),\n    description: parseDescriptionFromResponse(content),\n  };\n}\n",
     "/**\n * Deep Research model handler\n *\n * Provides deep research capabilities using OpenAI's o3-deep-research and o4-mini-deep-research models.\n * These models can find, analyze, and synthesize hundreds of sources to create comprehensive reports.\n *\n * @see https://platform.openai.com/docs/guides/deep-research\n */\n\nimport type {\n  IAgentRuntime,\n  JsonValue,\n  RecordLlmCallDetails,\n  ResearchAnnotation,\n  ResearchCodeInterpreterCall,\n  ResearchFileSearchCall,\n  ResearchMcpToolCall,\n  ResearchMessageOutput,\n  ResearchOutputItem,\n  ResearchParams,\n  ResearchResult,\n  ResearchTool,\n  ResearchWebSearchCall,\n} from \"@elizaos/core\";\nimport { logger, recordLlmCall } from \"@elizaos/core\";\nimport { getApiKey, getBaseURL, getResearchModel, getResearchTimeout } from \"../utils/config\";\n\n// ============================================================================\n// Types for OpenAI Responses API\n// ============================================================================\n\n/**\n * Tool configuration for the Responses API\n */\ninterface ResponsesApiTool {\n  type: \"web_search_preview\" | \"file_search\" | \"code_interpreter\" | \"mcp\";\n  vector_store_ids?: string[];\n  container?: { type: \"auto\" };\n  server_label?: string;\n  server_url?: string;\n  require_approval?: \"never\";\n}\n\n/**\n * Raw response from the OpenAI Responses API\n */\ninterface ResponsesApiResponse {\n  id: string;\n  object: string;\n  status?: \"queued\" | \"in_progress\" | \"completed\" | \"failed\";\n  output?: ResponsesApiOutputItem[];\n  output_text?: string;\n  error?: {\n    message: string;\n    code: string;\n  };\n}\n\n/**\n * Raw output item from the Responses API\n */\ninterface ResponsesApiOutputItem {\n  id?: string;\n  type: string;\n  status?: string;\n  action?: {\n    type: string;\n    query?: string;\n    url?: string;\n  };\n  query?: string;\n  results?: Array<{\n    file_id: string;\n    file_name: string;\n    score: number;\n  }>;\n  code?: string;\n  output?: string;\n  server_label?: string;\n  tool_name?: string;\n  arguments?: Record<string, unknown>;\n  result?: unknown;\n  content?: Array<{\n    type: string;\n    text: string;\n    annotations?: Array<{\n      url: string;\n      title: string;\n      start_index: number;\n      end_index: number;\n    }>;\n  }>;\n}\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\n/**\n * Converts ResearchTool params to Responses API tool format\n */\nfunction convertToolToApi(tool: ResearchTool): ResponsesApiTool {\n  switch (tool.type) {\n    case \"web_search_preview\":\n      return { type: \"web_search_preview\" };\n    case \"file_search\":\n      return {\n        type: \"file_search\",\n        vector_store_ids: tool.vectorStoreIds,\n      };\n    case \"code_interpreter\":\n      return {\n        type: \"code_interpreter\",\n        container: tool.container ?? { type: \"auto\" },\n      };\n    case \"mcp\":\n      return {\n        type: \"mcp\",\n        server_label: tool.serverLabel,\n        server_url: tool.serverUrl,\n        require_approval: tool.requireApproval ?? \"never\",\n      };\n    default:\n      throw new Error(`Unknown research tool type: ${(tool as ResearchTool).type}`);\n  }\n}\n\n/**\n * Converts raw API output items to typed ResearchOutputItem\n */\nfunction convertOutputItem(item: ResponsesApiOutputItem): ResearchOutputItem | null {\n  switch (item.type) {\n    case \"web_search_call\":\n      return {\n        id: item.id ?? \"\",\n        type: \"web_search_call\",\n        status: (item.status as \"completed\" | \"failed\") ?? \"completed\",\n        action: {\n          type: (item.action?.type as \"search\" | \"open_page\" | \"find_in_page\") ?? \"search\",\n          query: item.action?.query,\n          url: item.action?.url,\n        },\n      } satisfies ResearchWebSearchCall;\n\n    case \"file_search_call\":\n      return {\n        id: item.id ?? \"\",\n        type: \"file_search_call\",\n        status: (item.status as \"completed\" | \"failed\") ?? \"completed\",\n        query: item.query ?? \"\",\n        results: item.results?.map((r) => ({\n          fileId: r.file_id,\n          fileName: r.file_name,\n          score: r.score,\n        })),\n      } satisfies ResearchFileSearchCall;\n\n    case \"code_interpreter_call\":\n      return {\n        id: item.id ?? \"\",\n        type: \"code_interpreter_call\",\n        status: (item.status as \"completed\" | \"failed\") ?? \"completed\",\n        code: item.code ?? \"\",\n        output: item.output,\n      } satisfies ResearchCodeInterpreterCall;\n\n    case \"mcp_tool_call\":\n      return {\n        id: item.id ?? \"\",\n        type: \"mcp_tool_call\",\n        status: (item.status as \"completed\" | \"failed\") ?? \"completed\",\n        serverLabel: item.server_label ?? \"\",\n        toolName: item.tool_name ?? \"\",\n        arguments: (item.arguments ?? {}) as Record<string, JsonValue>,\n        result: item.result as JsonValue,\n      } satisfies ResearchMcpToolCall;\n\n    case \"message\":\n      return {\n        type: \"message\",\n        content:\n          item.content?.map((c) => ({\n            type: \"output_text\" as const,\n            text: c.text,\n            annotations:\n              c.annotations?.map((a) => ({\n                url: a.url,\n                title: a.title,\n                startIndex: a.start_index,\n                endIndex: a.end_index,\n              })) ?? [],\n          })) ?? [],\n      } satisfies ResearchMessageOutput;\n\n    default:\n      // Unknown output type, skip\n      return null;\n  }\n}\n\n/**\n * Extracts text and annotations from the response\n */\nfunction extractTextAndAnnotations(response: ResponsesApiResponse): {\n  text: string;\n  annotations: ResearchAnnotation[];\n} {\n  // Try output_text first (convenience field)\n  if (response.output_text) {\n    // Find annotations from message output items\n    const annotations: ResearchAnnotation[] = [];\n    if (response.output) {\n      for (const item of response.output) {\n        if (item.type === \"message\" && item.content) {\n          for (const content of item.content) {\n            if (content.annotations) {\n              for (const ann of content.annotations) {\n                annotations.push({\n                  url: ann.url,\n                  title: ann.title,\n                  startIndex: ann.start_index,\n                  endIndex: ann.end_index,\n                });\n              }\n            }\n          }\n        }\n      }\n    }\n    return { text: response.output_text, annotations };\n  }\n\n  // Fall back to extracting from message output items\n  let text = \"\";\n  const annotations: ResearchAnnotation[] = [];\n\n  if (response.output) {\n    for (const item of response.output) {\n      if (item.type === \"message\" && item.content) {\n        for (const content of item.content) {\n          text += content.text;\n          if (content.annotations) {\n            for (const ann of content.annotations) {\n              annotations.push({\n                url: ann.url,\n                title: ann.title,\n                startIndex: ann.start_index,\n                endIndex: ann.end_index,\n              });\n            }\n          }\n        }\n      }\n    }\n  }\n\n  return { text, annotations };\n}\n\n// ============================================================================\n// Main Handler\n// ============================================================================\n\n/**\n * Handles RESEARCH model requests using OpenAI's deep research models.\n *\n * Deep research models can take tens of minutes to complete tasks.\n * Use background mode for long-running tasks.\n *\n * @param runtime - The agent runtime\n * @param params - Research parameters\n * @returns Research result with text, annotations, and output items\n *\n * @example\n * ```typescript\n * const result = await handleResearch(runtime, {\n *   input: \"Research the economic impact of AI on global labor markets\",\n *   tools: [\n *     { type: \"web_search_preview\" },\n *     { type: \"code_interpreter\", container: { type: \"auto\" } }\n *   ],\n *   background: true,\n * });\n * console.log(result.text);\n * ```\n */\nexport async function handleResearch(\n  runtime: IAgentRuntime,\n  params: ResearchParams\n): Promise<ResearchResult> {\n  const apiKey = getApiKey(runtime);\n  if (!apiKey) {\n    throw new Error(\n      \"OPENAI_API_KEY is required for deep research. Set it in your environment variables or runtime settings.\"\n    );\n  }\n\n  const baseURL = getBaseURL(runtime);\n  const modelName = params.model ?? getResearchModel(runtime);\n  const timeout = getResearchTimeout(runtime);\n\n  logger.debug(`[OpenAI] Starting deep research with model: ${modelName}`);\n  logger.debug(`[OpenAI] Research input: ${params.input.substring(0, 100)}...`);\n\n  // Validate that at least one data source tool is provided\n  const dataSourceTools = params.tools?.filter(\n    (t) => t.type === \"web_search_preview\" || t.type === \"file_search\" || t.type === \"mcp\"\n  );\n\n  if (!dataSourceTools || dataSourceTools.length === 0) {\n    // Default to web search if no tools specified\n    logger.debug(\"[OpenAI] No data source tools specified, defaulting to web_search_preview\");\n    params.tools = [{ type: \"web_search_preview\" }, ...(params.tools ?? [])];\n  }\n\n  // Build the request body for the Responses API\n  const requestBody: Record<string, unknown> = {\n    model: modelName,\n    input: params.input,\n  };\n\n  if (params.instructions) {\n    requestBody.instructions = params.instructions;\n  }\n\n  if (params.background !== undefined) {\n    requestBody.background = params.background;\n  }\n\n  if (params.tools && params.tools.length > 0) {\n    requestBody.tools = params.tools.map(convertToolToApi);\n  }\n\n  if (params.maxToolCalls !== undefined) {\n    requestBody.max_tool_calls = params.maxToolCalls;\n  }\n\n  if (params.reasoningSummary) {\n    requestBody.reasoning = { summary: params.reasoningSummary };\n  }\n\n  logger.debug(`[OpenAI] Research request body: ${JSON.stringify(requestBody, null, 2)}`);\n\n  const details: RecordLlmCallDetails = {\n    model: modelName,\n    systemPrompt: params.instructions ?? \"\",\n    userPrompt: params.input,\n    temperature: 0,\n    maxTokens: 0,\n    purpose: \"external_llm\",\n    actionType: \"openai.responses.create\",\n  };\n  const data = await recordLlmCall(runtime, details, async () => {\n    // Make the API request\n    const response = await fetch(`${baseURL}/responses`, {\n      method: \"POST\",\n      headers: {\n        Authorization: `Bearer ${apiKey}`,\n        \"Content-Type\": \"application/json\",\n      },\n      body: JSON.stringify(requestBody),\n      signal: AbortSignal.timeout(timeout),\n    });\n\n    if (!response.ok) {\n      const errorText = await response.text();\n      logger.error(`[OpenAI] Research request failed: ${response.status} ${errorText}`);\n      throw new Error(`Deep research request failed: ${response.status} ${response.statusText}`);\n    }\n\n    const responseData = (await response.json()) as ResponsesApiResponse;\n    details.response = responseData.output_text ?? \"\";\n    return responseData;\n  });\n\n  if (data.error) {\n    logger.error(`[OpenAI] Research API error: ${data.error.message}`);\n    throw new Error(`Deep research error: ${data.error.message}`);\n  }\n\n  logger.debug(`[OpenAI] Research response received. Status: ${data.status ?? \"completed\"}`);\n\n  // Extract text and annotations\n  const { text, annotations } = extractTextAndAnnotations(data);\n\n  // Convert output items\n  const outputItems: ResearchOutputItem[] = [];\n  if (data.output) {\n    for (const item of data.output) {\n      const converted = convertOutputItem(item);\n      if (converted) {\n        outputItems.push(converted);\n      }\n    }\n  }\n\n  const result: ResearchResult = {\n    id: data.id,\n    text,\n    annotations,\n    outputItems,\n    status: data.status,\n  };\n\n  logger.info(\n    `[OpenAI] Research completed. Text length: ${text.length}, Annotations: ${annotations.length}, Output items: ${outputItems.length}`\n  );\n\n  return result;\n}\n",
-    "/**\n * Text generation model handlers\n *\n * Provides text generation using OpenAI's language models.\n */\n\nimport type {\n  GenerateTextParams,\n  IAgentRuntime,\n  JsonValue,\n  ModelTypeName,\n  RecordLlmCallDetails,\n} from \"@elizaos/core\";\nimport {\n  buildCanonicalSystemPrompt,\n  dropDuplicateLeadingSystemMessage,\n  logger,\n  ModelType,\n  normalizeSchemaForCerebras,\n  recordLlmCall,\n  resolveEffectiveSystemPrompt,\n  sanitizeFunctionNameForCerebras,\n} from \"@elizaos/core\";\nimport {\n  generateText,\n  type JSONSchema7,\n  jsonSchema,\n  type LanguageModelUsage,\n  type ModelMessage,\n  Output,\n  streamText,\n  type ToolChoice,\n  type ToolSet,\n  type UserContent,\n} from \"ai\";\nimport { createOpenAIClient } from \"../providers\";\nimport type { TextStreamResult, TokenUsage } from \"../types\";\nimport {\n  getActionPlannerModel,\n  getExperimentalTelemetry,\n  getLargeModel,\n  getMediumModel,\n  getMegaModel,\n  getNanoModel,\n  getResponseHandlerModel,\n  getSmallModel,\n  isCerebrasMode,\n} from \"../utils/config\";\nimport { emitModelUsageEvent } from \"../utils/events\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\n/**\n * Function to get model name from runtime\n */\ntype ModelNameGetter = (runtime: IAgentRuntime) => string;\n\ntype PromptCacheRetention = \"in_memory\" | \"24h\";\ntype ChatAttachment = {\n  data: string | Uint8Array | URL;\n  mediaType: string;\n  filename?: string;\n};\n\ninterface OpenAIPromptCacheOptions {\n  promptCacheKey?: string;\n  promptCacheRetention?: PromptCacheRetention;\n}\n\ninterface GenerateTextParamsWithOpenAIOptions\n  extends Omit<\n    GenerateTextParams,\n    \"messages\" | \"tools\" | \"toolChoice\" | \"responseSchema\" | \"providerOptions\"\n  > {\n  model?: string;\n  attachments?: ChatAttachment[];\n  messages?: unknown[];\n  tools?: unknown;\n  toolChoice?: unknown;\n  responseSchema?: unknown;\n  providerOptions?: Record<string, object | JsonValue> & {\n    agentName?: string;\n    openai?: OpenAIPromptCacheOptions;\n  };\n}\n\ntype NativeOutput = NonNullable<Parameters<typeof generateText<ToolSet>>[0][\"output\"]>;\ntype NativeGenerateTextParams = Parameters<typeof generateText<ToolSet, NativeOutput>>[0];\ntype NativeStreamTextParams = Parameters<typeof streamText<ToolSet, NativeOutput>>[0];\ntype NativePrompt =\n  | { prompt: string; messages?: never }\n  | { messages: ModelMessage[]; prompt?: never };\ntype NativeTextParams = Omit<NativeGenerateTextParams, \"messages\" | \"prompt\"> &\n  Omit<NativeStreamTextParams, \"messages\" | \"prompt\"> &\n  NativePrompt & {\n    // Re-declared explicitly: TypeScript's `Parameters<typeof generateText>`\n    // inference produces an overload-union that drops this field, but the\n    // ai SDK's runtime signature accepts it (see ai@6 `CallSettings & Prompt`).\n    allowSystemInMessages?: boolean;\n  };\ntype NativeProviderOptions = NativeTextParams[\"providerOptions\"];\ntype NativeTelemetrySettings = NativeTextParams[\"experimental_telemetry\"];\n\ntype LanguageModelUsageWithCache = Omit<LanguageModelUsage, \"inputTokenDetails\"> & {\n  inputTokenDetails?: LanguageModelUsage[\"inputTokenDetails\"] & {\n    cachedInputTokens?: number;\n    cacheCreationInputTokens?: number;\n    cacheCreationTokens?: number;\n  };\n  cachedInputTokens?: number;\n  cacheReadInputTokens?: number;\n  cacheCreationInputTokens?: number;\n  cacheWriteInputTokens?: number;\n  input_tokens_details?: {\n    cached_tokens?: number;\n    cache_read_input_tokens?: number;\n    cache_creation_input_tokens?: number;\n  };\n  prompt_tokens_details?: {\n    cached_tokens?: number;\n  };\n};\n\ninterface NativeGenerateTextResult {\n  text: string;\n  toolCalls?: unknown[];\n  finishReason?: string;\n  usage?: TokenUsage;\n  providerMetadata?: unknown;\n}\n\ntype NativeTextModelResult = string & NativeGenerateTextResult;\n\nconst TEXT_NANO_MODEL_TYPE = ModelType.TEXT_NANO as ModelTypeName;\nconst TEXT_MEDIUM_MODEL_TYPE = ModelType.TEXT_MEDIUM as ModelTypeName;\nconst TEXT_MEGA_MODEL_TYPE = ModelType.TEXT_MEGA as ModelTypeName;\nconst RESPONSE_HANDLER_MODEL_TYPE = ModelType.RESPONSE_HANDLER as ModelTypeName;\nconst ACTION_PLANNER_MODEL_TYPE = ModelType.ACTION_PLANNER as ModelTypeName;\n\nfunction resolveRequestedModelName(\n  params: GenerateTextParamsWithOpenAIOptions,\n  runtime: IAgentRuntime,\n  getModelFn: ModelNameGetter\n): string {\n  return typeof params.model === \"string\" && params.model.trim().length > 0\n    ? params.model.trim()\n    : getModelFn(runtime);\n}\n\nfunction buildUserContent(params: GenerateTextParamsWithOpenAIOptions): UserContent {\n  const content: Array<\n    | { type: \"text\"; text: string }\n    | {\n        type: \"file\";\n        data: string | Uint8Array | URL;\n        mediaType: string;\n        filename?: string;\n      }\n  > = [{ type: \"text\", text: params.prompt ?? \"\" }];\n\n  for (const attachment of params.attachments ?? []) {\n    content.push({\n      type: \"file\",\n      data: attachment.data,\n      mediaType: attachment.mediaType,\n      ...(attachment.filename ? { filename: attachment.filename } : {}),\n    });\n  }\n\n  return content;\n}\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\n/**\n * Converts AI SDK usage to our token usage format.\n *\n * Emits both the legacy `cachedPromptTokens` (kept for back-compat with\n * existing OpenAI consumers) and the canonical v5 `cacheReadInputTokens`\n * (consumed by the trajectory recorder + cost table). They always carry the\n * same value when the AI SDK reports cached input.\n */\nfunction convertUsage(usage: LanguageModelUsage | undefined): TokenUsage | undefined {\n  if (!usage) {\n    return undefined;\n  }\n\n  // The AI SDK uses inputTokens/outputTokens\n  const promptTokens = usage.inputTokens ?? 0;\n  const completionTokens = usage.outputTokens ?? 0;\n  const usageWithCache: LanguageModelUsageWithCache = usage;\n  const cachedInput =\n    firstNumber(\n      usageWithCache.cacheReadInputTokens,\n      usageWithCache.cachedInputTokens,\n      usageWithCache.inputTokenDetails?.cacheReadTokens,\n      usageWithCache.inputTokenDetails?.cachedInputTokens,\n      usageWithCache.input_tokens_details?.cache_read_input_tokens,\n      usageWithCache.input_tokens_details?.cached_tokens,\n      usageWithCache.prompt_tokens_details?.cached_tokens\n    ) ?? undefined;\n  const cacheCreationInput = firstNumber(\n    usageWithCache.cacheCreationInputTokens,\n    usageWithCache.cacheWriteInputTokens,\n    usageWithCache.inputTokenDetails?.cacheCreationInputTokens,\n    usageWithCache.inputTokenDetails?.cacheCreationTokens,\n    usageWithCache.inputTokenDetails?.cacheWriteTokens,\n    usageWithCache.input_tokens_details?.cache_creation_input_tokens\n  );\n\n  return {\n    promptTokens,\n    completionTokens,\n    totalTokens: promptTokens + completionTokens,\n    cachedPromptTokens: cachedInput,\n    cacheReadInputTokens: cachedInput,\n    cacheCreationInputTokens: cacheCreationInput,\n  };\n}\n\nfunction firstNumber(...values: unknown[]): number | undefined {\n  for (const value of values) {\n    if (typeof value === \"number\" && Number.isFinite(value)) {\n      return value;\n    }\n    if (typeof value === \"string\" && value.trim().length > 0) {\n      const parsed = Number(value);\n      if (Number.isFinite(parsed)) {\n        return parsed;\n      }\n    }\n  }\n  return undefined;\n}\n\nfunction resolvePromptCacheOptions(params: GenerateTextParams): OpenAIPromptCacheOptions {\n  const withOpenAIOptions = params as GenerateTextParamsWithOpenAIOptions;\n  return {\n    promptCacheKey: withOpenAIOptions.providerOptions?.openai?.promptCacheKey,\n    promptCacheRetention: withOpenAIOptions.providerOptions?.openai?.promptCacheRetention,\n  };\n}\n\n/**\n * Forward `OPENAI_REASONING_EFFORT` (runtime setting / process.env) as\n * `reasoning_effort` on the outbound chat completions request. This is\n * the OpenAI-spec knob for reasoning-capable models (`o1-*`, `o3-*`,\n * `gpt-oss-*`, `deepseek-r1`, `qwen-3-thinking`, etc.) — including\n * Cerebras and OpenRouter, which honor the same field. `\"low\"` keeps\n * reasoning short enough that visible content always fits inside\n * `max_tokens`, which is the failure mode on Cerebras gpt-oss-120b when\n * left unset.\n *\n * In Cerebras mode the field defaults to `\"low\"` when unset, but ONLY for\n * reasoning-capable models (e.g. gpt-oss-*, deepseek-r1, qwen-3-thinking):\n * gpt-oss-120b emits a separate reasoning channel and, left unbounded, spends\n * the whole token budget reasoning — returning empty visible content, which\n * makes the agent fall back to \"I don't have a reply for that\". `\"low\"` keeps\n * reasoning short so a reply always materializes. Non-reasoning Cerebras models\n * (Llama, etc.) reject `reasoning_effort`, so they must never receive the\n * default. For all other models an unset/invalid value yields `undefined`, so\n * they pay no overhead and the wire stays clean. An explicit valid\n * `OPENAI_REASONING_EFFORT` always wins.\n *\n * Valid values follow the OpenAI spec exactly: `minimal`, `low`,\n * `medium`, `high`. Anything else is logged and ignored.\n */\ntype ReasoningEffort = \"minimal\" | \"low\" | \"medium\" | \"high\";\n\nconst VALID_REASONING_EFFORTS: readonly ReasoningEffort[] = [\"minimal\", \"low\", \"medium\", \"high\"];\n\n/**\n * Reasoning-capable model families that emit a separate reasoning channel and\n * honor `reasoning_effort`. Used to gate the Cerebras `\"low\"` default so\n * non-reasoning models (Llama, etc.) are never sent the field.\n */\nfunction isReasoningModel(modelName: string | undefined): boolean {\n  if (!modelName) return false;\n  const m = modelName.toLowerCase();\n  return (\n    m.includes(\"gpt-oss\") ||\n    m.includes(\"o1\") ||\n    m.includes(\"o3\") ||\n    m.includes(\"o4\") ||\n    m.includes(\"deepseek-r1\") ||\n    m.includes(\"thinking\") ||\n    m.includes(\"reasoning\") ||\n    m.includes(\"qwq\")\n  );\n}\n\nfunction resolveReasoningEffort(\n  runtime: IAgentRuntime,\n  modelName?: string\n): ReasoningEffort | undefined {\n  const raw = runtime.getSetting(\"OPENAI_REASONING_EFFORT\");\n  const normalized = typeof raw === \"string\" ? raw.trim().toLowerCase() : \"\";\n  if (normalized) {\n    if ((VALID_REASONING_EFFORTS as readonly string[]).includes(normalized)) {\n      return normalized as ReasoningEffort;\n    }\n    logger.warn(\n      `[OpenAI] OPENAI_REASONING_EFFORT=${raw} is not a valid reasoning effort; ignoring. Expected one of: ${VALID_REASONING_EFFORTS.join(\", \")}.`\n    );\n  }\n  // gpt-oss-120b on Cerebras returns empty content when reasoning runs\n  // unbounded; default to \"low\" so a visible reply always fits — but only for\n  // reasoning-capable models. Non-reasoning Cerebras models (Llama, etc.)\n  // reject `reasoning_effort` and would break. An explicit valid value above\n  // wins over this default.\n  if (isCerebrasMode(runtime) && isReasoningModel(modelName)) {\n    return \"low\";\n  }\n  return undefined;\n}\n\nfunction resolveProviderOptions(\n  params: GenerateTextParams,\n  runtime: IAgentRuntime,\n  modelName?: string\n): Record<string, unknown> | undefined {\n  const withOpenAIOptions = params as GenerateTextParamsWithOpenAIOptions;\n  const rawProviderOptions = withOpenAIOptions.providerOptions;\n  const promptCacheOptions = resolvePromptCacheOptions(params);\n  const reasoningEffort = resolveReasoningEffort(runtime, modelName);\n\n  if (\n    !rawProviderOptions &&\n    !promptCacheOptions.promptCacheKey &&\n    !promptCacheOptions.promptCacheRetention &&\n    !reasoningEffort\n  ) {\n    return undefined;\n  }\n\n  // Cerebras supports prompt caching on gpt-oss-120b — 128-token blocks,\n  // default-on. The `prompt_cache_key` field IS accepted by Cerebras's\n  // OpenAI-compatible endpoint and surfaces hit counts via\n  // `usage.prompt_tokens_details.cached_tokens` (same shape as OpenAI), so\n  // we keep it in the request body. Only `prompt_cache_retention` is an\n  // OpenAI-direct-only field that Cerebras rejects with HTTP 400\n  // (`wrong_api_format`), so we strip just that one when in Cerebras mode.\n  const skipCacheRetention = isCerebrasMode(runtime);\n\n  const { agentName: _agentName, openai: rawOpenAIOptions, ...rest } = rawProviderOptions ?? {};\n  // When on Cerebras, scrub OpenAI-direct-only fields (e.g. `promptCacheRetention`)\n  // from `rawOpenAIOptions` before they're spread; otherwise they reach the wire\n  // and the Cerebras endpoint rejects with HTTP 400 `wrong_api_format`.\n  const sanitizedRawOpenAIOptions = (() => {\n    if (!rawOpenAIOptions || typeof rawOpenAIOptions !== \"object\") return rawOpenAIOptions;\n    if (!skipCacheRetention) return rawOpenAIOptions;\n    const { promptCacheRetention: _drop, ...rest2 } = rawOpenAIOptions as Record<string, unknown>;\n    return rest2;\n  })();\n  const openaiOptions = {\n    ...(sanitizedRawOpenAIOptions ?? {}),\n    ...(promptCacheOptions.promptCacheKey\n      ? { promptCacheKey: promptCacheOptions.promptCacheKey }\n      : {}),\n    ...(!skipCacheRetention && promptCacheOptions.promptCacheRetention\n      ? { promptCacheRetention: promptCacheOptions.promptCacheRetention }\n      : {}),\n    // The caller's explicit `reasoningEffort` wins over the resolved default\n    // (env var, or Cerebras \"low\") — same precedence pattern as promptCacheKey.\n    ...((sanitizedRawOpenAIOptions as { reasoningEffort?: unknown } | undefined)\n      ?.reasoningEffort === undefined && reasoningEffort\n      ? { reasoningEffort }\n      : {}),\n  };\n\n  const providerOptions = {\n    ...rest,\n    ...(Object.keys(openaiOptions).length > 0 ? { openai: openaiOptions } : {}),\n  };\n\n  return Object.keys(providerOptions).length > 0 ? providerOptions : undefined;\n}\n\nfunction buildStructuredOutput(responseSchema: unknown): NativeOutput {\n  if (\n    responseSchema &&\n    typeof responseSchema === \"object\" &&\n    \"responseFormat\" in responseSchema &&\n    \"parseCompleteOutput\" in responseSchema\n  ) {\n    return responseSchema as NativeOutput;\n  }\n\n  const schemaOptions =\n    responseSchema && typeof responseSchema === \"object\" && \"schema\" in responseSchema\n      ? (responseSchema as { schema: unknown; name?: string; description?: string })\n      : { schema: responseSchema };\n\n  return Output.object({\n    schema: jsonSchema(sanitizeJsonSchema(schemaOptions.schema, true)),\n    ...(schemaOptions.name ? { name: schemaOptions.name } : {}),\n    ...(schemaOptions.description ? { description: schemaOptions.description } : {}),\n  }) as NativeOutput;\n}\n\nfunction normalizeNativeTools(\n  tools: unknown,\n  options: { cerebrasMode?: boolean } = {}\n): ToolSet | undefined {\n  if (!tools) {\n    return undefined;\n  }\n\n  // Existing AI SDK callers already pass a ToolSet keyed by tool name. Keep it\n  // intact so custom tool instances, execute hooks, and dynamic tool metadata\n  // are preserved.\n  if (!Array.isArray(tools)) {\n    return tools as ToolSet;\n  }\n\n  const toolSet: Record<string, unknown> = {};\n\n  for (const rawTool of tools) {\n    const tool = asRecord(rawTool);\n    const functionTool = asRecord(tool.function);\n    const name = firstString(tool.name, functionTool.name);\n\n    if (!name) {\n      throw new Error(\"[OpenAI] Native tool definition is missing a name.\");\n    }\n\n    const description = firstString(tool.description, functionTool.description);\n    // Default to a permissive object schema. The empty-properties shape\n    // (`{ type: \"object\", properties: {}, additionalProperties: false }`) is\n    // accepted by OpenAI but rejected by strict-grammar providers like\n    // Cerebras with `Object fields require at least one of: 'properties' or\n    // 'anyOf' with a list of possible properties`.\n    const rawSchema =\n      tool.parameters ?? functionTool.parameters ?? ({ type: \"object\" } satisfies JSONSchema7);\n    let inputSchema = sanitizeJsonSchema(rawSchema, true);\n    if (options.cerebrasMode) {\n      // User-supplied schemas may still contain empty-properties subobjects\n      // even after sanitizeJsonSchema. Apply Cerebras-specific normalization\n      // recursively so deep schemas are accepted by the grammar compiler.\n      // Pass isRoot: true so the top-level invariant is enforced (must be\n      // type:\"object\" with no root oneOf/anyOf/enum/not).\n      inputSchema = normalizeSchemaForCerebras(inputSchema, true) as JSONSchema7;\n    }\n\n    // Cerebras's grammar compiler rejects function names containing characters\n    // outside `[a-zA-Z0-9_-]` (e.g. `math.factorial`). The AI SDK looks up\n    // tools by the registered key, so we register under the sanitized name AND\n    // surface it to the model under that name. Tool calls come back with the\n    // sanitized name, which the runtime resolves through its action registry —\n    // any caller relying on dotted action names should pre-sanitize.\n    const registeredName = options.cerebrasMode ? sanitizeFunctionNameForCerebras(name) : name;\n\n    toolSet[registeredName] = {\n      ...(description ? { description } : {}),\n      inputSchema: jsonSchema(inputSchema as JSONSchema7),\n    };\n  }\n\n  return Object.keys(toolSet).length > 0 ? (toolSet as ToolSet) : undefined;\n}\n\nfunction normalizeNativeMessages(messages: unknown): ModelMessage[] | undefined {\n  if (!Array.isArray(messages)) {\n    return undefined;\n  }\n\n  return messages.map((message) => normalizeNativeMessage(message));\n}\n\nfunction normalizeNativeMessage(message: unknown): ModelMessage {\n  const raw = asRecord(message);\n  const providerOptions = asOptionalRecord(raw.providerOptions);\n\n  if (raw.role === \"system\") {\n    return {\n      role: \"system\",\n      content: stringifyMessageContent(raw.content),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  if (raw.role === \"assistant\") {\n    return {\n      role: \"assistant\",\n      content: normalizeAssistantContent(raw),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  if (raw.role === \"tool\") {\n    return {\n      role: \"tool\",\n      content: normalizeToolContent(raw),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  return {\n    role: \"user\",\n    content: normalizeUserContent(raw.content),\n    ...(providerOptions ? { providerOptions } : {}),\n  } as ModelMessage;\n}\n\n/**\n * Strip reasoning-only parts from outbound assistant content.\n *\n * OpenAI-spec reasoning models (Cerebras gpt-oss-120b, OpenAI o1/o3,\n * DeepSeek R1, Qwen-3-thinking, etc.) return reasoning in the assistant\n * response — either as a separate `reasoning` / `reasoning_content`\n * field, or as content parts with `type: \"reasoning\"`. Echoing those\n * back to the next turn is wrong on both ends:\n *   - Cerebras returns HTTP 400 (`messages.X.assistant.reasoning_content:\n *     property is unsupported`).\n *   - OpenAI silently drops them, which wastes prompt tokens.\n *\n * The AI SDK upstream of this normalizer surfaces those reasoning blocks\n * as `{ type: \"reasoning\", ... }` content parts. We drop them here so\n * the wire stays spec-clean for the next turn. The reasoning itself\n * remains usable as a single-turn signal (still on the response object);\n * we only refuse to round-trip it.\n */\nfunction stripReasoningParts(content: unknown[]): unknown[] {\n  return content.filter((part) => {\n    if (!part || typeof part !== \"object\") return true;\n    const type = (part as { type?: unknown }).type;\n    return type !== \"reasoning\" && type !== \"thinking\";\n  });\n}\n\nfunction normalizeAssistantContent(message: Record<string, unknown>): unknown {\n  const toolCalls = Array.isArray(message.toolCalls) ? message.toolCalls : [];\n\n  if (toolCalls.length === 0) {\n    if (Array.isArray(message.content)) {\n      return stripReasoningParts(message.content);\n    }\n    if (typeof message.content === \"string\") {\n      return message.content;\n    }\n    return \"\";\n  }\n\n  const parts: unknown[] = [];\n  if (typeof message.content === \"string\" && message.content.length > 0) {\n    parts.push({ type: \"text\", text: message.content });\n  } else if (Array.isArray(message.content)) {\n    parts.push(...stripReasoningParts(message.content));\n  }\n\n  for (const toolCall of toolCalls) {\n    const rawCall = asRecord(toolCall);\n    const rawFunction = asRecord(rawCall.function);\n    const toolCallId = firstString(rawCall.toolCallId, rawCall.id);\n    const toolName = firstString(rawCall.toolName, rawCall.name, rawFunction.name);\n\n    if (!toolCallId || !toolName) {\n      continue;\n    }\n\n    parts.push({\n      type: \"tool-call\",\n      toolCallId,\n      toolName,\n      input: parseToolCallInput(rawCall, rawFunction),\n    });\n  }\n\n  return parts;\n}\n\nfunction normalizeToolContent(message: Record<string, unknown>): unknown[] {\n  if (Array.isArray(message.content)) {\n    return message.content;\n  }\n\n  const toolCallId = firstString(message.toolCallId, message.id) ?? \"tool-call\";\n  const toolName = firstString(message.toolName, message.name) ?? \"tool\";\n  const parsed = parseJsonIfPossible(message.content);\n\n  return [\n    {\n      type: \"tool-result\",\n      toolCallId,\n      toolName,\n      output:\n        typeof parsed === \"string\"\n          ? { type: \"text\", value: parsed }\n          : { type: \"json\", value: parsed },\n    },\n  ];\n}\n\nfunction normalizeUserContent(content: unknown): UserContent {\n  if (Array.isArray(content)) {\n    return content as UserContent;\n  }\n  return stringifyMessageContent(content);\n}\n\nfunction stringifyMessageContent(content: unknown): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n  if (content == null) {\n    return \"\";\n  }\n  return typeof content === \"object\" ? JSON.stringify(content) : String(content);\n}\n\nfunction parseToolCallInput(\n  rawCall: Record<string, unknown>,\n  rawFunction: Record<string, unknown>\n): unknown {\n  if (\"input\" in rawCall) {\n    return rawCall.input;\n  }\n  return parseJsonIfPossible(rawCall.arguments ?? rawFunction.arguments ?? {});\n}\n\nfunction parseJsonIfPossible(value: unknown): unknown {\n  if (typeof value !== \"string\") {\n    return value ?? \"\";\n  }\n  try {\n    return JSON.parse(value);\n  } catch {\n    return value;\n  }\n}\n\nfunction normalizeToolChoice(toolChoice: unknown): ToolChoice<ToolSet> | undefined {\n  if (!toolChoice) {\n    return undefined;\n  }\n\n  if (\n    typeof toolChoice === \"string\" &&\n    (toolChoice === \"auto\" || toolChoice === \"none\" || toolChoice === \"required\")\n  ) {\n    return toolChoice;\n  }\n\n  const choice = asRecord(toolChoice);\n  if (choice.type === \"tool\") {\n    if (typeof choice.toolName === \"string\" && choice.toolName.length > 0) {\n      return toolChoice as ToolChoice<ToolSet>;\n    }\n    const toolName = firstString(choice.toolName, choice.name);\n    if (toolName) {\n      return { type: \"tool\", toolName };\n    }\n  }\n\n  if (choice.type === \"function\") {\n    const fn = asRecord(choice.function);\n    const toolName = firstString(fn.name);\n    if (toolName) {\n      return { type: \"tool\", toolName };\n    }\n  }\n\n  const namedTool = firstString(choice.name);\n  if (namedTool) {\n    return { type: \"tool\", toolName: namedTool };\n  }\n\n  return toolChoice as ToolChoice<ToolSet>;\n}\n\nfunction hasIllegalStrictRoot(node: Record<string, unknown>): boolean {\n  // Strict-mode JSON schema validators on OpenAI-compatible providers (Groq,\n  // Cerebras, OpenAI strict tools) reject tool-parameters whose top level is\n  // not `type: \"object\"` or carries `oneOf`/`anyOf`/`enum`/`not` at the root.\n  // The error wording varies by provider but the constraint is uniform.\n  if (node.type !== \"object\") return true;\n  if (Array.isArray(node.oneOf) && node.oneOf.length > 0) return true;\n  if (Array.isArray(node.anyOf) && node.anyOf.length > 0) return true;\n  if (Array.isArray(node.enum)) return true;\n  if (node.not !== undefined) return true;\n  return false;\n}\n\nfunction sanitizeJsonSchema(schema: unknown, isRoot = false): JSONSchema7 {\n  if (!schema || typeof schema !== \"object\" || Array.isArray(schema)) {\n    // Permissive fallback: no `properties: {}`/`additionalProperties: false`\n    // pair, which strict-grammar providers reject. See `normalizeSchemaForCerebras`\n    // in @elizaos/core for the rationale.\n    return { type: \"object\" };\n  }\n\n  const record = schema as Record<string, unknown>;\n  let sanitized: Record<string, unknown> = { ...record };\n\n  if (typeof sanitized.type !== \"string\") {\n    const inferredType = inferJsonSchemaType(sanitized, isRoot);\n    if (inferredType) {\n      sanitized.type = inferredType;\n    }\n  }\n\n  if (isRoot && hasIllegalStrictRoot(sanitized)) {\n    // Wrap the original schema under properties.value. Strict-tool callers\n    // that unwrap arguments will see `{ value: <original> }`. The recursion\n    // below normalises the wrapped child like any other property.\n    sanitized = {\n      type: \"object\",\n      properties: { value: { ...record } },\n      required: [\"value\"],\n      additionalProperties: false,\n    };\n  }\n\n  if (\n    sanitized.properties &&\n    typeof sanitized.properties === \"object\" &&\n    !Array.isArray(sanitized.properties)\n  ) {\n    const properties: Record<string, unknown> = {};\n    for (const [key, value] of Object.entries(sanitized.properties as Record<string, unknown>)) {\n      properties[key] = sanitizeJsonSchema(value);\n    }\n    sanitized.properties = properties;\n\n    const propertyKeys = Object.keys(properties);\n    const existingRequired = Array.isArray(sanitized.required)\n      ? sanitized.required.filter((key): key is string => typeof key === \"string\")\n      : [];\n    sanitized.required = [...new Set([...existingRequired, ...propertyKeys])];\n  }\n\n  if (sanitized.type === \"object\" && sanitized.additionalProperties !== false) {\n    sanitized.additionalProperties = false;\n  }\n\n  if (sanitized.items) {\n    sanitized.items = Array.isArray(sanitized.items)\n      ? sanitized.items.map((item) => sanitizeJsonSchema(item))\n      : sanitizeJsonSchema(sanitized.items);\n  }\n\n  for (const unionKey of [\"anyOf\", \"oneOf\", \"allOf\"] as const) {\n    const value = sanitized[unionKey];\n    if (Array.isArray(value)) {\n      sanitized[unionKey] = value.map((item) => sanitizeJsonSchema(item));\n    }\n  }\n\n  return sanitized as JSONSchema7;\n}\n\nfunction inferJsonSchemaType(schema: Record<string, unknown>, isRoot: boolean): string | undefined {\n  if (\n    \"properties\" in schema ||\n    \"required\" in schema ||\n    \"additionalProperties\" in schema ||\n    isRoot\n  ) {\n    return \"object\";\n  }\n  if (\"items\" in schema) {\n    return \"array\";\n  }\n  if (Array.isArray(schema.enum) && schema.enum.length > 0) {\n    const types = new Set(schema.enum.map((value) => typeof value));\n    if (types.size === 1) {\n      const [type] = [...types];\n      if (type === \"string\" || type === \"number\" || type === \"boolean\") {\n        return type;\n      }\n    }\n  }\n  return undefined;\n}\n\nfunction asRecord(value: unknown): Record<string, unknown> {\n  return value && typeof value === \"object\" && !Array.isArray(value)\n    ? (value as Record<string, unknown>)\n    : {};\n}\n\nfunction asOptionalRecord(value: unknown): Record<string, unknown> | undefined {\n  return value && typeof value === \"object\" && !Array.isArray(value)\n    ? (value as Record<string, unknown>)\n    : undefined;\n}\n\nfunction firstString(...values: unknown[]): string | undefined {\n  for (const value of values) {\n    if (typeof value === \"string\" && value.length > 0) {\n      return value;\n    }\n  }\n  return undefined;\n}\n\nfunction usesNativeTextResult(params: GenerateTextParamsWithOpenAIOptions): boolean {\n  return Boolean(params.messages || params.tools || params.toolChoice || params.responseSchema);\n}\n\nfunction buildNativeTextResult(\n  result: {\n    text: string;\n    toolCalls?: unknown[];\n    finishReason?: string;\n    usage?: LanguageModelUsage;\n    providerMetadata?: unknown;\n  },\n  modelName?: string\n): NativeGenerateTextResult {\n  return {\n    text: result.text,\n    toolCalls: result.toolCalls ?? [],\n    finishReason: result.finishReason,\n    usage: convertUsage(result.usage),\n    providerMetadata: mergeProviderModelName(result.providerMetadata, modelName),\n  };\n}\n\nfunction handledPromise<T>(value: T | PromiseLike<T>): Promise<T> {\n  const promise = Promise.resolve(value);\n  promise.catch(() => {\n    // The streaming path primarily consumes `textStream`. AI SDK companion\n    // promises such as `text` can reject later on empty streams even when no\n    // caller requested them, which otherwise surfaces as an unhandled rejection.\n  });\n  return promise;\n}\n\nfunction handledMappedPromise<T, U>(\n  value: T | PromiseLike<T>,\n  mapper: (resolved: T) => U | PromiseLike<U>\n): Promise<U> {\n  return handledPromise(handledPromise(value).then(mapper));\n}\n\nfunction mergeProviderModelName(providerMetadata: unknown, modelName?: string): unknown {\n  if (!modelName) {\n    return providerMetadata;\n  }\n  if (\n    providerMetadata &&\n    typeof providerMetadata === \"object\" &&\n    !Array.isArray(providerMetadata)\n  ) {\n    return {\n      ...(providerMetadata as Record<string, unknown>),\n      modelName,\n    };\n  }\n  return { modelName };\n}\n\nfunction createLlmCallDetails(\n  modelName: string,\n  params: GenerateTextParams,\n  systemPrompt: string | undefined,\n  actionType: string,\n  modelType?: ModelTypeName,\n  providerOptions?: Record<string, unknown>,\n  generateParams?: NativeTextParams\n): RecordLlmCallDetails {\n  const originalParams = params as GenerateTextParamsWithOpenAIOptions;\n  const nativeParams = generateParams as\n    | (NativeTextParams & {\n        output?: unknown;\n        maxOutputTokens?: unknown;\n      })\n    | undefined;\n  const nativePrompt = nativeParams && \"prompt\" in nativeParams ? nativeParams.prompt : undefined;\n  const nativeMessages =\n    nativeParams && \"messages\" in nativeParams && Array.isArray(nativeParams.messages)\n      ? nativeParams.messages\n      : undefined;\n  const nativeSystem =\n    typeof nativeParams?.system === \"string\" ? nativeParams.system : systemPrompt;\n  return {\n    model: modelName,\n    modelType,\n    provider: \"vercel-ai-sdk\",\n    systemPrompt: nativeSystem ?? \"\",\n    userPrompt:\n      typeof nativePrompt === \"string\"\n        ? nativePrompt\n        : typeof params.prompt === \"string\"\n          ? params.prompt\n          : \"\",\n    prompt: typeof nativePrompt === \"string\" ? nativePrompt : undefined,\n    messages: nativeMessages,\n    tools: nativeParams?.tools ?? originalParams.tools,\n    toolChoice: nativeParams?.toolChoice ?? originalParams.toolChoice,\n    output:\n      nativeParams?.output !== undefined\n        ? buildTrajectoryOutputDescriptor(originalParams.responseSchema, nativeParams.output)\n        : undefined,\n    responseSchema: originalParams.responseSchema,\n    providerOptions:\n      providerOptions ?? nativeParams?.providerOptions ?? originalParams.providerOptions,\n    temperature: params.temperature ?? 0,\n    maxTokens:\n      typeof nativeParams?.maxOutputTokens === \"number\"\n        ? nativeParams.maxOutputTokens\n        : params.omitMaxTokens\n          ? 0\n          : (params.maxTokens ?? 8192),\n    maxTokensOmitted:\n      params.omitMaxTokens && typeof nativeParams?.maxOutputTokens !== \"number\" ? true : undefined,\n    purpose: \"external_llm\",\n    actionType,\n  };\n}\n\nfunction buildTrajectoryOutputDescriptor(responseSchema: unknown, output: unknown): unknown {\n  if (responseSchema !== undefined) {\n    return {\n      type: \"object\",\n      schema: responseSchema,\n    };\n  }\n  return toTrajectoryJsonSafe(output);\n}\n\nfunction toTrajectoryJsonSafe(value: unknown): unknown {\n  try {\n    return JSON.parse(\n      JSON.stringify(value, (_key, nested) => {\n        if (typeof nested === \"function\") return undefined;\n        if (typeof nested === \"bigint\") return nested.toString();\n        return nested;\n      })\n    ) as unknown;\n  } catch {\n    return String(value);\n  }\n}\n\nfunction applyUsageToDetails(\n  details: RecordLlmCallDetails,\n  usage: LanguageModelUsage | undefined\n): void {\n  if (!usage) {\n    return;\n  }\n  details.promptTokens = usage.inputTokens ?? 0;\n  details.completionTokens = usage.outputTokens ?? 0;\n}\n\n// ============================================================================\n// Core Generation Function\n// ============================================================================\n\n/**\n * Generates text using the specified model type.\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @param modelType - The type of model (TEXT_SMALL or TEXT_LARGE)\n * @param getModelFn - Function to get the model name\n * @returns Generated text or stream result\n */\nasync function generateTextByModelType(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams,\n  modelType: ModelTypeName,\n  getModelFn: ModelNameGetter\n): Promise<string | TextStreamResult> {\n  const paramsWithAttachments = params as GenerateTextParamsWithOpenAIOptions;\n  const openai = createOpenAIClient(runtime);\n  const modelName = resolveRequestedModelName(paramsWithAttachments, runtime, getModelFn);\n\n  logger.debug(`[OpenAI] Using ${modelType} model: ${modelName}`);\n  const providerOptions = resolveProviderOptions(params, runtime, modelName);\n  const hasAttachments = (paramsWithAttachments.attachments?.length ?? 0) > 0;\n  const userContent = hasAttachments ? buildUserContent(paramsWithAttachments) : undefined;\n  const shouldReturnNativeResult = usesNativeTextResult(paramsWithAttachments);\n\n  const systemPrompt = resolveEffectiveSystemPrompt({\n    params: paramsWithAttachments,\n    fallback: buildCanonicalSystemPrompt({ character: runtime.character }),\n  });\n  const agentName = paramsWithAttachments.providerOptions?.agentName;\n  const telemetryConfig: NativeTelemetrySettings = {\n    isEnabled: getExperimentalTelemetry(runtime),\n    functionId: agentName ? `agent:${agentName}` : undefined,\n    metadata: agentName ? { agentName } : undefined,\n  };\n\n  // Chat Completions is the default: broadest compatibility, and it works\n  // against every OpenAI-compatible endpoint (Cerebras, local servers, proxies).\n  // gpt-5 / gpt-5-mini reasoning models ignore temperature/penalty/stop params.\n  //\n  const model = openai.chat(modelName);\n  const cerebrasMode = isCerebrasMode(runtime);\n  const normalizedTools = normalizeNativeTools(paramsWithAttachments.tools, {\n    cerebrasMode,\n  });\n  const normalizedToolChoice = normalizeToolChoice(paramsWithAttachments.toolChoice);\n  const normalizedMessages = normalizeNativeMessages(paramsWithAttachments.messages);\n  const wireMessages = dropDuplicateLeadingSystemMessage(normalizedMessages, systemPrompt);\n  const effectiveMessages =\n    wireMessages && wireMessages.length > 0 ? wireMessages : normalizedMessages;\n  const promptText =\n    typeof params.prompt === \"string\" && params.prompt.length > 0 ? params.prompt : \"\";\n  const promptOrMessages: NativePrompt =\n    effectiveMessages && effectiveMessages.length > 0\n      ? { messages: effectiveMessages }\n      : userContent\n        ? { messages: [{ role: \"user\" as const, content: userContent }] }\n        : { prompt: promptText };\n  // elizaOS callers pass `responseFormat: { type: \"json_object\" | \"text\" }`\n  // (see `GenerateTextParams` in @elizaos/core). The AI SDK's equivalent\n  // is `responseFormat: { type: \"json\" }` (which translates to\n  // `response_format: { type: \"json_object\" }` at the OpenAI wire layer).\n  // Translate the shape so the param actually reaches the API call —\n  // before this, callers asking for json_object were silently ignored\n  // and Cerebras returned plain text, dropping us into the simple-reply\n  // fallback every turn.\n  const callerResponseFormat = (paramsWithAttachments as { responseFormat?: unknown })\n    .responseFormat;\n  const responseFormatType =\n    typeof callerResponseFormat === \"string\"\n      ? callerResponseFormat\n      : callerResponseFormat &&\n          typeof callerResponseFormat === \"object\" &&\n          \"type\" in callerResponseFormat\n        ? (callerResponseFormat as { type: string }).type\n        : undefined;\n  const wireResponseFormat: { type: \"json\" } | { type: \"text\" } | undefined =\n    responseFormatType === \"json_object\"\n      ? { type: \"json\" }\n      : responseFormatType === \"text\"\n        ? { type: \"text\" }\n        : undefined;\n\n  const generateParams: NativeTextParams = {\n    model,\n    ...promptOrMessages,\n    system: systemPrompt,\n    allowSystemInMessages: true,\n    // Omit the cap when the caller opted out (direct-channel Stage-1) so the\n    // model's own max applies — a hardcoded value 400s when it exceeds the\n    // model's limit. Other callers keep the 8192 default.\n    ...(params.omitMaxTokens ? {} : { maxOutputTokens: params.maxTokens ?? 8192 }),\n    experimental_telemetry: telemetryConfig,\n    ...(normalizedTools ? { tools: normalizedTools } : {}),\n    ...(normalizedToolChoice ? { toolChoice: normalizedToolChoice } : {}),\n    // Cerebras's OpenAI-compatible endpoint does not accept the\n    // `response_format: { type: \"json_schema\", ... }` payload that the AI SDK\n    // emits when `output: Output.object(...)` is set. Fall back to relying on\n    // `responseFormat: { type: \"json_object\" }` (already passed by callers)\n    // plus the schema embedded in the prompt body.\n    ...(paramsWithAttachments.responseSchema && !isCerebrasMode(runtime)\n      ? { output: buildStructuredOutput(paramsWithAttachments.responseSchema) }\n      : {}),\n    ...(wireResponseFormat ? { responseFormat: wireResponseFormat } : {}),\n    ...(providerOptions ? { providerOptions: providerOptions as NativeProviderOptions } : {}),\n  };\n\n  // Handle streaming mode\n  if (params.stream) {\n    const details = createLlmCallDetails(\n      modelName,\n      params,\n      systemPrompt,\n      \"ai.streamText\",\n      modelType,\n      providerOptions,\n      generateParams\n    );\n    details.response = \"\";\n    const result = await recordLlmCall(runtime, details, () => streamText(generateParams));\n\n    return {\n      textStream: (async function* textStreamWithCallback() {\n        for await (const chunk of result.textStream) {\n          params.onStreamChunk?.(chunk);\n          yield chunk;\n        }\n      })(),\n      text: handledPromise(result.text),\n      ...(shouldReturnNativeResult ? { toolCalls: handledPromise(result.toolCalls) } : {}),\n      usage: handledMappedPromise(result.usage, convertUsage),\n      finishReason: handledMappedPromise(result.finishReason, (r) => r as string | undefined),\n    };\n  }\n\n  // Non-streaming mode\n  const details = createLlmCallDetails(\n    modelName,\n    params,\n    systemPrompt,\n    \"ai.generateText\",\n    modelType,\n    providerOptions,\n    generateParams\n  );\n  const result = await recordLlmCall(runtime, details, async () => {\n    const result = await generateText(generateParams);\n    details.response = result.text;\n    details.toolCalls = result.toolCalls;\n    details.finishReason = result.finishReason as string | undefined;\n    details.providerMetadata = result.providerMetadata;\n    applyUsageToDetails(details, result.usage);\n    return result;\n  });\n\n  if (result.usage) {\n    emitModelUsageEvent(runtime, modelType, params.prompt ?? \"\", result.usage);\n  }\n\n  if (shouldReturnNativeResult) {\n    return buildNativeTextResult(result, modelName) as NativeTextModelResult;\n  }\n\n  return result.text;\n}\n\n// ============================================================================\n// Public Handlers\n// ============================================================================\n\n/**\n * Handles TEXT_SMALL model requests.\n *\n * Uses the configured small model (default: gpt-5-mini).\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @returns Generated text or stream result\n */\nexport async function handleTextSmall(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ModelType.TEXT_SMALL, getSmallModel);\n}\n\nexport async function handleTextNano(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_NANO_MODEL_TYPE, getNanoModel);\n}\n\nexport async function handleTextMedium(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_MEDIUM_MODEL_TYPE, getMediumModel);\n}\n\n/**\n * Handles TEXT_LARGE model requests.\n *\n * Uses the configured large model (default: gpt-5).\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @returns Generated text or stream result\n */\nexport async function handleTextLarge(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ModelType.TEXT_LARGE, getLargeModel);\n}\n\nexport async function handleTextMega(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_MEGA_MODEL_TYPE, getMegaModel);\n}\n\nexport async function handleResponseHandler(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(\n    runtime,\n    params,\n    RESPONSE_HANDLER_MODEL_TYPE,\n    getResponseHandlerModel\n  );\n}\n\nexport async function handleActionPlanner(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ACTION_PLANNER_MODEL_TYPE, getActionPlannerModel);\n}\n\n// ─── Test-only exports ──────────────────────────────────────────────────────\n// These are exported for the shape tests in `__tests__/reasoning-effort.shape.test.ts`.\n// Not part of the public API; do not import outside tests.\n\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_resolveProviderOptions = resolveProviderOptions;\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_normalizeNativeMessages = normalizeNativeMessages;\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_stripReasoningParts = stripReasoningParts;\n",
+    "/**\n * Text generation model handlers\n *\n * Provides text generation using OpenAI's language models.\n */\n\nimport type {\n  GenerateTextParams,\n  IAgentRuntime,\n  JsonValue,\n  ModelTypeName,\n  RecordLlmCallDetails,\n} from \"@elizaos/core\";\nimport {\n  buildCanonicalSystemPrompt,\n  dropDuplicateLeadingSystemMessage,\n  logger,\n  ModelType,\n  normalizeSchemaForCerebras,\n  recordLlmCall,\n  resolveEffectiveSystemPrompt,\n  sanitizeFunctionNameForCerebras,\n} from \"@elizaos/core\";\nimport {\n  generateText,\n  type JSONSchema7,\n  jsonSchema,\n  type LanguageModelUsage,\n  type ModelMessage,\n  Output,\n  streamText,\n  type ToolChoice,\n  type ToolSet,\n  type UserContent,\n} from \"ai\";\nimport { createOpenAIClient } from \"../providers\";\nimport type { TextStreamResult, TokenUsage } from \"../types\";\nimport {\n  getActionPlannerModel,\n  getExperimentalTelemetry,\n  getLargeModel,\n  getMediumModel,\n  getMegaModel,\n  getNanoModel,\n  getResponseHandlerModel,\n  getSmallModel,\n  isCerebrasMode,\n} from \"../utils/config\";\nimport { emitModelUsageEvent } from \"../utils/events\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\n/**\n * Function to get model name from runtime\n */\ntype ModelNameGetter = (runtime: IAgentRuntime) => string;\n\ntype PromptCacheRetention = \"in_memory\" | \"24h\";\ntype ChatAttachment = {\n  data: string | Uint8Array | URL;\n  mediaType: string;\n  filename?: string;\n};\n\ninterface OpenAIPromptCacheOptions {\n  promptCacheKey?: string;\n  promptCacheRetention?: PromptCacheRetention;\n}\n\ninterface GenerateTextParamsWithOpenAIOptions\n  extends Omit<\n    GenerateTextParams,\n    \"messages\" | \"tools\" | \"toolChoice\" | \"responseSchema\" | \"providerOptions\"\n  > {\n  model?: string;\n  attachments?: ChatAttachment[];\n  messages?: unknown[];\n  tools?: unknown;\n  toolChoice?: unknown;\n  responseSchema?: unknown;\n  providerOptions?: Record<string, object | JsonValue> & {\n    agentName?: string;\n    openai?: OpenAIPromptCacheOptions;\n  };\n}\n\ntype NativeOutput = NonNullable<Parameters<typeof generateText<ToolSet>>[0][\"output\"]>;\ntype NativeGenerateTextParams = Parameters<typeof generateText<ToolSet, NativeOutput>>[0];\ntype NativeStreamTextParams = Parameters<typeof streamText<ToolSet, NativeOutput>>[0];\ntype NativePrompt =\n  | { prompt: string; messages?: never }\n  | { messages: ModelMessage[]; prompt?: never };\ntype NativeTextParams = Omit<NativeGenerateTextParams, \"messages\" | \"prompt\"> &\n  Omit<NativeStreamTextParams, \"messages\" | \"prompt\"> &\n  NativePrompt & {\n    // Re-declared explicitly: TypeScript's `Parameters<typeof generateText>`\n    // inference produces an overload-union that drops this field, but the\n    // ai SDK's runtime signature accepts it (see ai@6 `CallSettings & Prompt`).\n    allowSystemInMessages?: boolean;\n  };\ntype NativeProviderOptions = NativeTextParams[\"providerOptions\"];\ntype NativeTelemetrySettings = NativeTextParams[\"experimental_telemetry\"];\n\ntype LanguageModelUsageWithCache = Omit<LanguageModelUsage, \"inputTokenDetails\"> & {\n  inputTokenDetails?: LanguageModelUsage[\"inputTokenDetails\"] & {\n    cachedInputTokens?: number;\n    cacheCreationInputTokens?: number;\n    cacheCreationTokens?: number;\n  };\n  cachedInputTokens?: number;\n  cacheReadInputTokens?: number;\n  cacheCreationInputTokens?: number;\n  cacheWriteInputTokens?: number;\n  input_tokens_details?: {\n    cached_tokens?: number;\n    cache_read_input_tokens?: number;\n    cache_creation_input_tokens?: number;\n  };\n  prompt_tokens_details?: {\n    cached_tokens?: number;\n  };\n};\n\ninterface NativeGenerateTextResult {\n  text: string;\n  toolCalls?: unknown[];\n  finishReason?: string;\n  usage?: TokenUsage;\n  providerMetadata?: unknown;\n}\n\ntype NativeTextModelResult = string & NativeGenerateTextResult;\n\nconst TEXT_NANO_MODEL_TYPE = ModelType.TEXT_NANO as ModelTypeName;\nconst TEXT_MEDIUM_MODEL_TYPE = ModelType.TEXT_MEDIUM as ModelTypeName;\nconst TEXT_MEGA_MODEL_TYPE = ModelType.TEXT_MEGA as ModelTypeName;\nconst RESPONSE_HANDLER_MODEL_TYPE = ModelType.RESPONSE_HANDLER as ModelTypeName;\nconst ACTION_PLANNER_MODEL_TYPE = ModelType.ACTION_PLANNER as ModelTypeName;\n\nfunction resolveRequestedModelName(\n  params: GenerateTextParamsWithOpenAIOptions,\n  runtime: IAgentRuntime,\n  getModelFn: ModelNameGetter\n): string {\n  return typeof params.model === \"string\" && params.model.trim().length > 0\n    ? params.model.trim()\n    : getModelFn(runtime);\n}\n\nfunction buildUserContent(params: GenerateTextParamsWithOpenAIOptions): UserContent {\n  const content: Array<\n    | { type: \"text\"; text: string }\n    | {\n        type: \"file\";\n        data: string | Uint8Array | URL;\n        mediaType: string;\n        filename?: string;\n      }\n  > = [{ type: \"text\", text: params.prompt ?? \"\" }];\n\n  for (const attachment of params.attachments ?? []) {\n    content.push({\n      type: \"file\",\n      data: attachment.data,\n      mediaType: attachment.mediaType,\n      ...(attachment.filename ? { filename: attachment.filename } : {}),\n    });\n  }\n\n  return content;\n}\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\n/**\n * Converts AI SDK usage to our token usage format.\n *\n * Emits both the legacy `cachedPromptTokens` (kept for back-compat with\n * existing OpenAI consumers) and the canonical v5 `cacheReadInputTokens`\n * (consumed by the trajectory recorder + cost table). They always carry the\n * same value when the AI SDK reports cached input.\n */\nfunction convertUsage(usage: LanguageModelUsage | undefined): TokenUsage | undefined {\n  if (!usage) {\n    return undefined;\n  }\n\n  // The AI SDK uses inputTokens/outputTokens\n  const promptTokens = usage.inputTokens ?? 0;\n  const completionTokens = usage.outputTokens ?? 0;\n  const usageWithCache: LanguageModelUsageWithCache = usage;\n  const cachedInput =\n    firstNumber(\n      usageWithCache.cacheReadInputTokens,\n      usageWithCache.cachedInputTokens,\n      usageWithCache.inputTokenDetails?.cacheReadTokens,\n      usageWithCache.inputTokenDetails?.cachedInputTokens,\n      usageWithCache.input_tokens_details?.cache_read_input_tokens,\n      usageWithCache.input_tokens_details?.cached_tokens,\n      usageWithCache.prompt_tokens_details?.cached_tokens\n    ) ?? undefined;\n  const cacheCreationInput = firstNumber(\n    usageWithCache.cacheCreationInputTokens,\n    usageWithCache.cacheWriteInputTokens,\n    usageWithCache.inputTokenDetails?.cacheCreationInputTokens,\n    usageWithCache.inputTokenDetails?.cacheCreationTokens,\n    usageWithCache.inputTokenDetails?.cacheWriteTokens,\n    usageWithCache.input_tokens_details?.cache_creation_input_tokens\n  );\n\n  return {\n    promptTokens,\n    completionTokens,\n    totalTokens: promptTokens + completionTokens,\n    cachedPromptTokens: cachedInput,\n    cacheReadInputTokens: cachedInput,\n    cacheCreationInputTokens: cacheCreationInput,\n  };\n}\n\nfunction firstNumber(...values: unknown[]): number | undefined {\n  for (const value of values) {\n    if (typeof value === \"number\" && Number.isFinite(value)) {\n      return value;\n    }\n    if (typeof value === \"string\" && value.trim().length > 0) {\n      const parsed = Number(value);\n      if (Number.isFinite(parsed)) {\n        return parsed;\n      }\n    }\n  }\n  return undefined;\n}\n\nfunction resolvePromptCacheOptions(params: GenerateTextParams): OpenAIPromptCacheOptions {\n  const withOpenAIOptions = params as GenerateTextParamsWithOpenAIOptions;\n  return {\n    promptCacheKey: withOpenAIOptions.providerOptions?.openai?.promptCacheKey,\n    promptCacheRetention: withOpenAIOptions.providerOptions?.openai?.promptCacheRetention,\n  };\n}\n\n/**\n * Forward `OPENAI_REASONING_EFFORT` (runtime setting / process.env) as\n * `reasoning_effort` on the outbound chat completions request. This is\n * the OpenAI-spec knob for reasoning-capable models (`o1-*`, `o3-*`,\n * `gpt-oss-*`, `deepseek-r1`, and similar families) — including\n * Cerebras and OpenRouter, which honor the same field. `\"low\"` keeps\n * reasoning short enough that visible content always fits inside\n * `max_tokens`, which is the failure mode on Cerebras gpt-oss-120b when\n * left unset.\n *\n * In Cerebras mode the field defaults to `\"low\"` when unset, but ONLY for\n * reasoning-capable models (e.g. gpt-oss-* and deepseek-r1):\n * gpt-oss-120b emits a separate reasoning channel and, left unbounded, spends\n * the whole token budget reasoning — returning empty visible content, which\n * makes the agent fall back to \"I don't have a reply for that\". `\"low\"` keeps\n * reasoning short so a reply always materializes. Non-reasoning Cerebras models\n * (Llama, etc.) reject `reasoning_effort`, so they must never receive the\n * default. For all other models an unset/invalid value yields `undefined`, so\n * they pay no overhead and the wire stays clean. An explicit valid\n * `OPENAI_REASONING_EFFORT` always wins.\n *\n * Valid values follow the OpenAI spec exactly: `minimal`, `low`,\n * `medium`, `high`. Anything else is logged and ignored.\n */\ntype ReasoningEffort = \"minimal\" | \"low\" | \"medium\" | \"high\";\n\nconst VALID_REASONING_EFFORTS: readonly ReasoningEffort[] = [\"minimal\", \"low\", \"medium\", \"high\"];\n\n/**\n * Reasoning-capable model families that emit a separate reasoning channel and\n * honor `reasoning_effort`. Used to gate the Cerebras `\"low\"` default so\n * non-reasoning models (Llama, etc.) are never sent the field.\n */\nfunction isReasoningModel(modelName: string | undefined): boolean {\n  if (!modelName) return false;\n  const m = modelName.toLowerCase();\n  return (\n    m.includes(\"gpt-oss\") ||\n    m.includes(\"o1\") ||\n    m.includes(\"o3\") ||\n    m.includes(\"o4\") ||\n    m.includes(\"deepseek-r1\") ||\n    m.includes(\"thinking\") ||\n    m.includes(\"reasoning\") ||\n    m.includes(\"qwq\")\n  );\n}\n\nfunction resolveReasoningEffort(\n  runtime: IAgentRuntime,\n  modelName?: string\n): ReasoningEffort | undefined {\n  const raw = runtime.getSetting(\"OPENAI_REASONING_EFFORT\");\n  const normalized = typeof raw === \"string\" ? raw.trim().toLowerCase() : \"\";\n  if (normalized) {\n    if ((VALID_REASONING_EFFORTS as readonly string[]).includes(normalized)) {\n      return normalized as ReasoningEffort;\n    }\n    logger.warn(\n      `[OpenAI] OPENAI_REASONING_EFFORT=${raw} is not a valid reasoning effort; ignoring. Expected one of: ${VALID_REASONING_EFFORTS.join(\", \")}.`\n    );\n  }\n  // gpt-oss-120b on Cerebras returns empty content when reasoning runs\n  // unbounded; default to \"low\" so a visible reply always fits — but only for\n  // reasoning-capable models. Non-reasoning Cerebras models (Llama, etc.)\n  // reject `reasoning_effort` and would break. An explicit valid value above\n  // wins over this default.\n  if (isCerebrasMode(runtime) && isReasoningModel(modelName)) {\n    return \"low\";\n  }\n  return undefined;\n}\n\nfunction resolveProviderOptions(\n  params: GenerateTextParams,\n  runtime: IAgentRuntime,\n  modelName?: string\n): Record<string, unknown> | undefined {\n  const withOpenAIOptions = params as GenerateTextParamsWithOpenAIOptions;\n  const rawProviderOptions = withOpenAIOptions.providerOptions;\n  const promptCacheOptions = resolvePromptCacheOptions(params);\n  const reasoningEffort = resolveReasoningEffort(runtime, modelName);\n\n  if (\n    !rawProviderOptions &&\n    !promptCacheOptions.promptCacheKey &&\n    !promptCacheOptions.promptCacheRetention &&\n    !reasoningEffort\n  ) {\n    return undefined;\n  }\n\n  // Cerebras supports prompt caching on gpt-oss-120b — 128-token blocks,\n  // default-on. The `prompt_cache_key` field IS accepted by Cerebras's\n  // OpenAI-compatible endpoint and surfaces hit counts via\n  // `usage.prompt_tokens_details.cached_tokens` (same shape as OpenAI), so\n  // we keep it in the request body. Only `prompt_cache_retention` is an\n  // OpenAI-direct-only field that Cerebras rejects with HTTP 400\n  // (`wrong_api_format`), so we strip just that one when in Cerebras mode.\n  const skipCacheRetention = isCerebrasMode(runtime);\n\n  const { agentName: _agentName, openai: rawOpenAIOptions, ...rest } = rawProviderOptions ?? {};\n  // When on Cerebras, scrub OpenAI-direct-only fields (e.g. `promptCacheRetention`)\n  // from `rawOpenAIOptions` before they're spread; otherwise they reach the wire\n  // and the Cerebras endpoint rejects with HTTP 400 `wrong_api_format`.\n  const sanitizedRawOpenAIOptions = (() => {\n    if (!rawOpenAIOptions || typeof rawOpenAIOptions !== \"object\") return rawOpenAIOptions;\n    if (!skipCacheRetention) return rawOpenAIOptions;\n    const { promptCacheRetention: _drop, ...rest2 } = rawOpenAIOptions as Record<string, unknown>;\n    return rest2;\n  })();\n  const openaiOptions = {\n    ...(sanitizedRawOpenAIOptions ?? {}),\n    ...(promptCacheOptions.promptCacheKey\n      ? { promptCacheKey: promptCacheOptions.promptCacheKey }\n      : {}),\n    ...(!skipCacheRetention && promptCacheOptions.promptCacheRetention\n      ? { promptCacheRetention: promptCacheOptions.promptCacheRetention }\n      : {}),\n    // The caller's explicit `reasoningEffort` wins over the resolved default\n    // (env var, or Cerebras \"low\") — same precedence pattern as promptCacheKey.\n    ...((sanitizedRawOpenAIOptions as { reasoningEffort?: unknown } | undefined)\n      ?.reasoningEffort === undefined && reasoningEffort\n      ? { reasoningEffort }\n      : {}),\n  };\n\n  const providerOptions = {\n    ...rest,\n    ...(Object.keys(openaiOptions).length > 0 ? { openai: openaiOptions } : {}),\n  };\n\n  return Object.keys(providerOptions).length > 0 ? providerOptions : undefined;\n}\n\nfunction buildStructuredOutput(responseSchema: unknown): NativeOutput {\n  if (\n    responseSchema &&\n    typeof responseSchema === \"object\" &&\n    \"responseFormat\" in responseSchema &&\n    \"parseCompleteOutput\" in responseSchema\n  ) {\n    return responseSchema as NativeOutput;\n  }\n\n  const schemaOptions =\n    responseSchema && typeof responseSchema === \"object\" && \"schema\" in responseSchema\n      ? (responseSchema as { schema: unknown; name?: string; description?: string })\n      : { schema: responseSchema };\n\n  return Output.object({\n    schema: jsonSchema(sanitizeJsonSchema(schemaOptions.schema, true)),\n    ...(schemaOptions.name ? { name: schemaOptions.name } : {}),\n    ...(schemaOptions.description ? { description: schemaOptions.description } : {}),\n  }) as NativeOutput;\n}\n\nfunction normalizeNativeTools(\n  tools: unknown,\n  options: { cerebrasMode?: boolean } = {}\n): ToolSet | undefined {\n  if (!tools) {\n    return undefined;\n  }\n\n  // Existing AI SDK callers already pass a ToolSet keyed by tool name. Keep it\n  // intact so custom tool instances, execute hooks, and dynamic tool metadata\n  // are preserved.\n  if (!Array.isArray(tools)) {\n    return tools as ToolSet;\n  }\n\n  const toolSet: Record<string, unknown> = {};\n\n  for (const rawTool of tools) {\n    const tool = asRecord(rawTool);\n    const functionTool = asRecord(tool.function);\n    const name = firstString(tool.name, functionTool.name);\n\n    if (!name) {\n      throw new Error(\"[OpenAI] Native tool definition is missing a name.\");\n    }\n\n    const description = firstString(tool.description, functionTool.description);\n    // Default to a permissive object schema. The empty-properties shape\n    // (`{ type: \"object\", properties: {}, additionalProperties: false }`) is\n    // accepted by OpenAI but rejected by strict-grammar providers like\n    // Cerebras with `Object fields require at least one of: 'properties' or\n    // 'anyOf' with a list of possible properties`.\n    const rawSchema =\n      tool.parameters ?? functionTool.parameters ?? ({ type: \"object\" } satisfies JSONSchema7);\n    let inputSchema = sanitizeJsonSchema(rawSchema, true);\n    if (options.cerebrasMode) {\n      // User-supplied schemas may still contain empty-properties subobjects\n      // even after sanitizeJsonSchema. Apply Cerebras-specific normalization\n      // recursively so deep schemas are accepted by the grammar compiler.\n      // Pass isRoot: true so the top-level invariant is enforced (must be\n      // type:\"object\" with no root oneOf/anyOf/enum/not).\n      inputSchema = normalizeSchemaForCerebras(inputSchema, true) as JSONSchema7;\n    }\n\n    // Cerebras's grammar compiler rejects function names containing characters\n    // outside `[a-zA-Z0-9_-]` (e.g. `math.factorial`). The AI SDK looks up\n    // tools by the registered key, so we register under the sanitized name AND\n    // surface it to the model under that name. Tool calls come back with the\n    // sanitized name, which the runtime resolves through its action registry —\n    // any caller relying on dotted action names should pre-sanitize.\n    const registeredName = options.cerebrasMode ? sanitizeFunctionNameForCerebras(name) : name;\n\n    toolSet[registeredName] = {\n      ...(description ? { description } : {}),\n      inputSchema: jsonSchema(inputSchema as JSONSchema7),\n    };\n  }\n\n  return Object.keys(toolSet).length > 0 ? (toolSet as ToolSet) : undefined;\n}\n\nfunction normalizeNativeMessages(messages: unknown): ModelMessage[] | undefined {\n  if (!Array.isArray(messages)) {\n    return undefined;\n  }\n\n  return messages.map((message) => normalizeNativeMessage(message));\n}\n\nfunction normalizeNativeMessage(message: unknown): ModelMessage {\n  const raw = asRecord(message);\n  const providerOptions = asOptionalRecord(raw.providerOptions);\n\n  if (raw.role === \"system\") {\n    return {\n      role: \"system\",\n      content: stringifyMessageContent(raw.content),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  if (raw.role === \"assistant\") {\n    return {\n      role: \"assistant\",\n      content: normalizeAssistantContent(raw),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  if (raw.role === \"tool\") {\n    return {\n      role: \"tool\",\n      content: normalizeToolContent(raw),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  return {\n    role: \"user\",\n    content: normalizeUserContent(raw.content),\n    ...(providerOptions ? { providerOptions } : {}),\n  } as ModelMessage;\n}\n\n/**\n * Strip reasoning-only parts from outbound assistant content.\n *\n * OpenAI-spec reasoning models (Cerebras gpt-oss-120b, OpenAI o1/o3,\n * DeepSeek R1, and similar families) return reasoning in the assistant\n * response — either as a separate `reasoning` / `reasoning_content`\n * field, or as content parts with `type: \"reasoning\"`. Echoing those\n * back to the next turn is wrong on both ends:\n *   - Cerebras returns HTTP 400 (`messages.X.assistant.reasoning_content:\n *     property is unsupported`).\n *   - OpenAI silently drops them, which wastes prompt tokens.\n *\n * The AI SDK upstream of this normalizer surfaces those reasoning blocks\n * as `{ type: \"reasoning\", ... }` content parts. We drop them here so\n * the wire stays spec-clean for the next turn. The reasoning itself\n * remains usable as a single-turn signal (still on the response object);\n * we only refuse to round-trip it.\n */\nfunction stripReasoningParts(content: unknown[]): unknown[] {\n  return content.filter((part) => {\n    if (!part || typeof part !== \"object\") return true;\n    const type = (part as { type?: unknown }).type;\n    return type !== \"reasoning\" && type !== \"thinking\";\n  });\n}\n\nfunction normalizeAssistantContent(message: Record<string, unknown>): unknown {\n  const toolCalls = Array.isArray(message.toolCalls) ? message.toolCalls : [];\n\n  if (toolCalls.length === 0) {\n    if (Array.isArray(message.content)) {\n      return stripReasoningParts(message.content);\n    }\n    if (typeof message.content === \"string\") {\n      return message.content;\n    }\n    return \"\";\n  }\n\n  const parts: unknown[] = [];\n  if (typeof message.content === \"string\" && message.content.length > 0) {\n    parts.push({ type: \"text\", text: message.content });\n  } else if (Array.isArray(message.content)) {\n    parts.push(...stripReasoningParts(message.content));\n  }\n\n  for (const toolCall of toolCalls) {\n    const rawCall = asRecord(toolCall);\n    const rawFunction = asRecord(rawCall.function);\n    const toolCallId = firstString(rawCall.toolCallId, rawCall.id);\n    const toolName = firstString(rawCall.toolName, rawCall.name, rawFunction.name);\n\n    if (!toolCallId || !toolName) {\n      continue;\n    }\n\n    parts.push({\n      type: \"tool-call\",\n      toolCallId,\n      toolName,\n      input: parseToolCallInput(rawCall, rawFunction),\n    });\n  }\n\n  return parts;\n}\n\nfunction normalizeToolContent(message: Record<string, unknown>): unknown[] {\n  if (Array.isArray(message.content)) {\n    return message.content;\n  }\n\n  const toolCallId = firstString(message.toolCallId, message.id) ?? \"tool-call\";\n  const toolName = firstString(message.toolName, message.name) ?? \"tool\";\n  const parsed = parseJsonIfPossible(message.content);\n\n  return [\n    {\n      type: \"tool-result\",\n      toolCallId,\n      toolName,\n      output:\n        typeof parsed === \"string\"\n          ? { type: \"text\", value: parsed }\n          : { type: \"json\", value: parsed },\n    },\n  ];\n}\n\nfunction normalizeUserContent(content: unknown): UserContent {\n  if (Array.isArray(content)) {\n    return content as UserContent;\n  }\n  return stringifyMessageContent(content);\n}\n\nfunction stringifyMessageContent(content: unknown): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n  if (content == null) {\n    return \"\";\n  }\n  return typeof content === \"object\" ? JSON.stringify(content) : String(content);\n}\n\nfunction parseToolCallInput(\n  rawCall: Record<string, unknown>,\n  rawFunction: Record<string, unknown>\n): unknown {\n  if (\"input\" in rawCall) {\n    return rawCall.input;\n  }\n  return parseJsonIfPossible(rawCall.arguments ?? rawFunction.arguments ?? {});\n}\n\nfunction parseJsonIfPossible(value: unknown): unknown {\n  if (typeof value !== \"string\") {\n    return value ?? \"\";\n  }\n  try {\n    return JSON.parse(value);\n  } catch {\n    return value;\n  }\n}\n\nfunction normalizeToolChoice(toolChoice: unknown): ToolChoice<ToolSet> | undefined {\n  if (!toolChoice) {\n    return undefined;\n  }\n\n  if (\n    typeof toolChoice === \"string\" &&\n    (toolChoice === \"auto\" || toolChoice === \"none\" || toolChoice === \"required\")\n  ) {\n    return toolChoice;\n  }\n\n  const choice = asRecord(toolChoice);\n  if (choice.type === \"tool\") {\n    if (typeof choice.toolName === \"string\" && choice.toolName.length > 0) {\n      return toolChoice as ToolChoice<ToolSet>;\n    }\n    const toolName = firstString(choice.toolName, choice.name);\n    if (toolName) {\n      return { type: \"tool\", toolName };\n    }\n  }\n\n  if (choice.type === \"function\") {\n    const fn = asRecord(choice.function);\n    const toolName = firstString(fn.name);\n    if (toolName) {\n      return { type: \"tool\", toolName };\n    }\n  }\n\n  const namedTool = firstString(choice.name);\n  if (namedTool) {\n    return { type: \"tool\", toolName: namedTool };\n  }\n\n  return toolChoice as ToolChoice<ToolSet>;\n}\n\nfunction hasIllegalStrictRoot(node: Record<string, unknown>): boolean {\n  // Strict-mode JSON schema validators on OpenAI-compatible providers (Groq,\n  // Cerebras, OpenAI strict tools) reject tool-parameters whose top level is\n  // not `type: \"object\"` or carries `oneOf`/`anyOf`/`enum`/`not` at the root.\n  // The error wording varies by provider but the constraint is uniform.\n  if (node.type !== \"object\") return true;\n  if (Array.isArray(node.oneOf) && node.oneOf.length > 0) return true;\n  if (Array.isArray(node.anyOf) && node.anyOf.length > 0) return true;\n  if (Array.isArray(node.enum)) return true;\n  if (node.not !== undefined) return true;\n  return false;\n}\n\nfunction sanitizeJsonSchema(schema: unknown, isRoot = false): JSONSchema7 {\n  if (!schema || typeof schema !== \"object\" || Array.isArray(schema)) {\n    // Permissive fallback: no `properties: {}`/`additionalProperties: false`\n    // pair, which strict-grammar providers reject. See `normalizeSchemaForCerebras`\n    // in @elizaos/core for the rationale.\n    return { type: \"object\" };\n  }\n\n  const record = schema as Record<string, unknown>;\n  let sanitized: Record<string, unknown> = { ...record };\n\n  if (typeof sanitized.type !== \"string\") {\n    const inferredType = inferJsonSchemaType(sanitized, isRoot);\n    if (inferredType) {\n      sanitized.type = inferredType;\n    }\n  }\n\n  if (isRoot && hasIllegalStrictRoot(sanitized)) {\n    // Wrap the original schema under properties.value. Strict-tool callers\n    // that unwrap arguments will see `{ value: <original> }`. The recursion\n    // below normalises the wrapped child like any other property.\n    sanitized = {\n      type: \"object\",\n      properties: { value: { ...record } },\n      required: [\"value\"],\n      additionalProperties: false,\n    };\n  }\n\n  if (\n    sanitized.properties &&\n    typeof sanitized.properties === \"object\" &&\n    !Array.isArray(sanitized.properties)\n  ) {\n    const properties: Record<string, unknown> = {};\n    for (const [key, value] of Object.entries(sanitized.properties as Record<string, unknown>)) {\n      properties[key] = sanitizeJsonSchema(value);\n    }\n    sanitized.properties = properties;\n\n    const propertyKeys = Object.keys(properties);\n    const existingRequired = Array.isArray(sanitized.required)\n      ? sanitized.required.filter((key): key is string => typeof key === \"string\")\n      : [];\n    sanitized.required = [...new Set([...existingRequired, ...propertyKeys])];\n  }\n\n  if (sanitized.type === \"object\" && sanitized.additionalProperties !== false) {\n    sanitized.additionalProperties = false;\n  }\n\n  if (sanitized.items) {\n    sanitized.items = Array.isArray(sanitized.items)\n      ? sanitized.items.map((item) => sanitizeJsonSchema(item))\n      : sanitizeJsonSchema(sanitized.items);\n  }\n\n  for (const unionKey of [\"anyOf\", \"oneOf\", \"allOf\"] as const) {\n    const value = sanitized[unionKey];\n    if (Array.isArray(value)) {\n      sanitized[unionKey] = value.map((item) => sanitizeJsonSchema(item));\n    }\n  }\n\n  return sanitized as JSONSchema7;\n}\n\nfunction inferJsonSchemaType(schema: Record<string, unknown>, isRoot: boolean): string | undefined {\n  if (\n    \"properties\" in schema ||\n    \"required\" in schema ||\n    \"additionalProperties\" in schema ||\n    isRoot\n  ) {\n    return \"object\";\n  }\n  if (\"items\" in schema) {\n    return \"array\";\n  }\n  if (Array.isArray(schema.enum) && schema.enum.length > 0) {\n    const types = new Set(schema.enum.map((value) => typeof value));\n    if (types.size === 1) {\n      const [type] = [...types];\n      if (type === \"string\" || type === \"number\" || type === \"boolean\") {\n        return type;\n      }\n    }\n  }\n  return undefined;\n}\n\nfunction asRecord(value: unknown): Record<string, unknown> {\n  return value && typeof value === \"object\" && !Array.isArray(value)\n    ? (value as Record<string, unknown>)\n    : {};\n}\n\nfunction asOptionalRecord(value: unknown): Record<string, unknown> | undefined {\n  return value && typeof value === \"object\" && !Array.isArray(value)\n    ? (value as Record<string, unknown>)\n    : undefined;\n}\n\nfunction firstString(...values: unknown[]): string | undefined {\n  for (const value of values) {\n    if (typeof value === \"string\" && value.length > 0) {\n      return value;\n    }\n  }\n  return undefined;\n}\n\nfunction usesNativeTextResult(params: GenerateTextParamsWithOpenAIOptions): boolean {\n  return Boolean(params.messages || params.tools || params.toolChoice || params.responseSchema);\n}\n\nfunction buildNativeTextResult(\n  result: {\n    text: string;\n    toolCalls?: unknown[];\n    finishReason?: string;\n    usage?: LanguageModelUsage;\n    providerMetadata?: unknown;\n  },\n  modelName?: string\n): NativeGenerateTextResult {\n  return {\n    text: result.text,\n    toolCalls: result.toolCalls ?? [],\n    finishReason: result.finishReason,\n    usage: convertUsage(result.usage),\n    providerMetadata: mergeProviderModelName(result.providerMetadata, modelName),\n  };\n}\n\nfunction handledPromise<T>(value: T | PromiseLike<T>): Promise<T> {\n  const promise = Promise.resolve(value);\n  promise.catch(() => {\n    // The streaming path primarily consumes `textStream`. AI SDK companion\n    // promises such as `text` can reject later on empty streams even when no\n    // caller requested them, which otherwise surfaces as an unhandled rejection.\n  });\n  return promise;\n}\n\nfunction handledMappedPromise<T, U>(\n  value: T | PromiseLike<T>,\n  mapper: (resolved: T) => U | PromiseLike<U>\n): Promise<U> {\n  return handledPromise(handledPromise(value).then(mapper));\n}\n\nfunction mergeProviderModelName(providerMetadata: unknown, modelName?: string): unknown {\n  if (!modelName) {\n    return providerMetadata;\n  }\n  if (\n    providerMetadata &&\n    typeof providerMetadata === \"object\" &&\n    !Array.isArray(providerMetadata)\n  ) {\n    return {\n      ...(providerMetadata as Record<string, unknown>),\n      modelName,\n    };\n  }\n  return { modelName };\n}\n\nfunction createLlmCallDetails(\n  modelName: string,\n  params: GenerateTextParams,\n  systemPrompt: string | undefined,\n  actionType: string,\n  modelType?: ModelTypeName,\n  providerOptions?: Record<string, unknown>,\n  generateParams?: NativeTextParams\n): RecordLlmCallDetails {\n  const originalParams = params as GenerateTextParamsWithOpenAIOptions;\n  const nativeParams = generateParams as\n    | (NativeTextParams & {\n        output?: unknown;\n        maxOutputTokens?: unknown;\n      })\n    | undefined;\n  const nativePrompt = nativeParams && \"prompt\" in nativeParams ? nativeParams.prompt : undefined;\n  const nativeMessages =\n    nativeParams && \"messages\" in nativeParams && Array.isArray(nativeParams.messages)\n      ? nativeParams.messages\n      : undefined;\n  const nativeSystem =\n    typeof nativeParams?.system === \"string\" ? nativeParams.system : systemPrompt;\n  return {\n    model: modelName,\n    modelType,\n    provider: \"vercel-ai-sdk\",\n    systemPrompt: nativeSystem ?? \"\",\n    userPrompt:\n      typeof nativePrompt === \"string\"\n        ? nativePrompt\n        : typeof params.prompt === \"string\"\n          ? params.prompt\n          : \"\",\n    prompt: typeof nativePrompt === \"string\" ? nativePrompt : undefined,\n    messages: nativeMessages,\n    tools: nativeParams?.tools ?? originalParams.tools,\n    toolChoice: nativeParams?.toolChoice ?? originalParams.toolChoice,\n    output:\n      nativeParams?.output !== undefined\n        ? buildTrajectoryOutputDescriptor(originalParams.responseSchema, nativeParams.output)\n        : undefined,\n    responseSchema: originalParams.responseSchema,\n    providerOptions:\n      providerOptions ?? nativeParams?.providerOptions ?? originalParams.providerOptions,\n    temperature: params.temperature ?? 0,\n    maxTokens:\n      typeof nativeParams?.maxOutputTokens === \"number\"\n        ? nativeParams.maxOutputTokens\n        : params.omitMaxTokens\n          ? 0\n          : (params.maxTokens ?? 8192),\n    maxTokensOmitted:\n      params.omitMaxTokens && typeof nativeParams?.maxOutputTokens !== \"number\" ? true : undefined,\n    purpose: \"external_llm\",\n    actionType,\n  };\n}\n\nfunction buildTrajectoryOutputDescriptor(responseSchema: unknown, output: unknown): unknown {\n  if (responseSchema !== undefined) {\n    return {\n      type: \"object\",\n      schema: responseSchema,\n    };\n  }\n  return toTrajectoryJsonSafe(output);\n}\n\nfunction toTrajectoryJsonSafe(value: unknown): unknown {\n  try {\n    return JSON.parse(\n      JSON.stringify(value, (_key, nested) => {\n        if (typeof nested === \"function\") return undefined;\n        if (typeof nested === \"bigint\") return nested.toString();\n        return nested;\n      })\n    ) as unknown;\n  } catch {\n    return String(value);\n  }\n}\n\nfunction applyUsageToDetails(\n  details: RecordLlmCallDetails,\n  usage: LanguageModelUsage | undefined\n): void {\n  if (!usage) {\n    return;\n  }\n  details.promptTokens = usage.inputTokens ?? 0;\n  details.completionTokens = usage.outputTokens ?? 0;\n}\n\n// ============================================================================\n// Core Generation Function\n// ============================================================================\n\n/**\n * Generates text using the specified model type.\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @param modelType - The type of model (TEXT_SMALL or TEXT_LARGE)\n * @param getModelFn - Function to get the model name\n * @returns Generated text or stream result\n */\nasync function generateTextByModelType(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams,\n  modelType: ModelTypeName,\n  getModelFn: ModelNameGetter\n): Promise<string | TextStreamResult> {\n  const paramsWithAttachments = params as GenerateTextParamsWithOpenAIOptions;\n  const openai = createOpenAIClient(runtime);\n  const modelName = resolveRequestedModelName(paramsWithAttachments, runtime, getModelFn);\n\n  logger.debug(`[OpenAI] Using ${modelType} model: ${modelName}`);\n  const providerOptions = resolveProviderOptions(params, runtime, modelName);\n  const hasAttachments = (paramsWithAttachments.attachments?.length ?? 0) > 0;\n  const userContent = hasAttachments ? buildUserContent(paramsWithAttachments) : undefined;\n  const shouldReturnNativeResult = usesNativeTextResult(paramsWithAttachments);\n\n  const systemPrompt = resolveEffectiveSystemPrompt({\n    params: paramsWithAttachments,\n    fallback: buildCanonicalSystemPrompt({ character: runtime.character }),\n  });\n  const agentName = paramsWithAttachments.providerOptions?.agentName;\n  const telemetryConfig: NativeTelemetrySettings = {\n    isEnabled: getExperimentalTelemetry(runtime),\n    functionId: agentName ? `agent:${agentName}` : undefined,\n    metadata: agentName ? { agentName } : undefined,\n  };\n\n  // Chat Completions is the default: broadest compatibility, and it works\n  // against every OpenAI-compatible endpoint (Cerebras, local servers, proxies).\n  // gpt-5 / gpt-5-mini reasoning models ignore temperature/penalty/stop params.\n  //\n  const model = openai.chat(modelName);\n  const cerebrasMode = isCerebrasMode(runtime);\n  const normalizedTools = normalizeNativeTools(paramsWithAttachments.tools, {\n    cerebrasMode,\n  });\n  const normalizedToolChoice = normalizeToolChoice(paramsWithAttachments.toolChoice);\n  const normalizedMessages = normalizeNativeMessages(paramsWithAttachments.messages);\n  const wireMessages = dropDuplicateLeadingSystemMessage(normalizedMessages, systemPrompt);\n  const effectiveMessages =\n    wireMessages && wireMessages.length > 0 ? wireMessages : normalizedMessages;\n  const promptText =\n    typeof params.prompt === \"string\" && params.prompt.length > 0 ? params.prompt : \"\";\n  const promptOrMessages: NativePrompt =\n    effectiveMessages && effectiveMessages.length > 0\n      ? { messages: effectiveMessages }\n      : userContent\n        ? { messages: [{ role: \"user\" as const, content: userContent }] }\n        : { prompt: promptText };\n  // elizaOS callers pass `responseFormat: { type: \"json_object\" | \"text\" }`\n  // (see `GenerateTextParams` in @elizaos/core). The AI SDK's equivalent\n  // is `responseFormat: { type: \"json\" }` (which translates to\n  // `response_format: { type: \"json_object\" }` at the OpenAI wire layer).\n  // Translate the shape so the param actually reaches the API call —\n  // before this, callers asking for json_object were silently ignored\n  // and Cerebras returned plain text, dropping us into the simple-reply\n  // fallback every turn.\n  const callerResponseFormat = (paramsWithAttachments as { responseFormat?: unknown })\n    .responseFormat;\n  const responseFormatType =\n    typeof callerResponseFormat === \"string\"\n      ? callerResponseFormat\n      : callerResponseFormat &&\n          typeof callerResponseFormat === \"object\" &&\n          \"type\" in callerResponseFormat\n        ? (callerResponseFormat as { type: string }).type\n        : undefined;\n  const wireResponseFormat: { type: \"json\" } | { type: \"text\" } | undefined =\n    responseFormatType === \"json_object\"\n      ? { type: \"json\" }\n      : responseFormatType === \"text\"\n        ? { type: \"text\" }\n        : undefined;\n\n  const generateParams: NativeTextParams = {\n    model,\n    ...promptOrMessages,\n    system: systemPrompt,\n    allowSystemInMessages: true,\n    // Omit the cap when the caller opted out (direct-channel Stage-1) so the\n    // model's own max applies — a hardcoded value 400s when it exceeds the\n    // model's limit. Other callers keep the 8192 default.\n    ...(params.omitMaxTokens ? {} : { maxOutputTokens: params.maxTokens ?? 8192 }),\n    experimental_telemetry: telemetryConfig,\n    ...(normalizedTools ? { tools: normalizedTools } : {}),\n    ...(normalizedToolChoice ? { toolChoice: normalizedToolChoice } : {}),\n    // Cerebras's OpenAI-compatible endpoint does not accept the\n    // `response_format: { type: \"json_schema\", ... }` payload that the AI SDK\n    // emits when `output: Output.object(...)` is set. Fall back to relying on\n    // `responseFormat: { type: \"json_object\" }` (already passed by callers)\n    // plus the schema embedded in the prompt body.\n    ...(paramsWithAttachments.responseSchema && !isCerebrasMode(runtime)\n      ? { output: buildStructuredOutput(paramsWithAttachments.responseSchema) }\n      : {}),\n    ...(wireResponseFormat ? { responseFormat: wireResponseFormat } : {}),\n    ...(providerOptions ? { providerOptions: providerOptions as NativeProviderOptions } : {}),\n  };\n\n  // Handle streaming mode\n  if (params.stream) {\n    const details = createLlmCallDetails(\n      modelName,\n      params,\n      systemPrompt,\n      \"ai.streamText\",\n      modelType,\n      providerOptions,\n      generateParams\n    );\n    details.response = \"\";\n    const result = await recordLlmCall(runtime, details, () => streamText(generateParams));\n\n    return {\n      textStream: (async function* textStreamWithCallback() {\n        for await (const chunk of result.textStream) {\n          params.onStreamChunk?.(chunk);\n          yield chunk;\n        }\n      })(),\n      text: handledPromise(result.text),\n      ...(shouldReturnNativeResult ? { toolCalls: handledPromise(result.toolCalls) } : {}),\n      usage: handledMappedPromise(result.usage, convertUsage),\n      finishReason: handledMappedPromise(result.finishReason, (r) => r as string | undefined),\n    };\n  }\n\n  // Non-streaming mode\n  const details = createLlmCallDetails(\n    modelName,\n    params,\n    systemPrompt,\n    \"ai.generateText\",\n    modelType,\n    providerOptions,\n    generateParams\n  );\n  const result = await recordLlmCall(runtime, details, async () => {\n    const result = await generateText(generateParams);\n    details.response = result.text;\n    details.toolCalls = result.toolCalls;\n    details.finishReason = result.finishReason as string | undefined;\n    details.providerMetadata = result.providerMetadata;\n    applyUsageToDetails(details, result.usage);\n    return result;\n  });\n\n  if (result.usage) {\n    emitModelUsageEvent(runtime, modelType, params.prompt ?? \"\", result.usage);\n  }\n\n  if (shouldReturnNativeResult) {\n    return buildNativeTextResult(result, modelName) as NativeTextModelResult;\n  }\n\n  return result.text;\n}\n\n// ============================================================================\n// Public Handlers\n// ============================================================================\n\n/**\n * Handles TEXT_SMALL model requests.\n *\n * Uses the configured small model (default: gpt-5-mini).\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @returns Generated text or stream result\n */\nexport async function handleTextSmall(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ModelType.TEXT_SMALL, getSmallModel);\n}\n\nexport async function handleTextNano(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_NANO_MODEL_TYPE, getNanoModel);\n}\n\nexport async function handleTextMedium(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_MEDIUM_MODEL_TYPE, getMediumModel);\n}\n\n/**\n * Handles TEXT_LARGE model requests.\n *\n * Uses the configured large model (default: gpt-5).\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @returns Generated text or stream result\n */\nexport async function handleTextLarge(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ModelType.TEXT_LARGE, getLargeModel);\n}\n\nexport async function handleTextMega(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_MEGA_MODEL_TYPE, getMegaModel);\n}\n\nexport async function handleResponseHandler(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(\n    runtime,\n    params,\n    RESPONSE_HANDLER_MODEL_TYPE,\n    getResponseHandlerModel\n  );\n}\n\nexport async function handleActionPlanner(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ACTION_PLANNER_MODEL_TYPE, getActionPlannerModel);\n}\n\n// ─── Test-only exports ──────────────────────────────────────────────────────\n// These are exported for the shape tests in `__tests__/reasoning-effort.shape.test.ts`.\n// Not part of the public API; do not import outside tests.\n\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_resolveProviderOptions = resolveProviderOptions;\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_normalizeNativeMessages = normalizeNativeMessages;\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_stripReasoningParts = stripReasoningParts;\n",
     "import { createOpenAI, type OpenAIProvider } from \"@ai-sdk/openai\";\nimport type { IAgentRuntime } from \"@elizaos/core\";\nimport { getApiKey, getBaseURL, isProxyMode } from \"../utils/config\";\n\nconst PROXY_API_KEY = \"sk-proxy\";\n\nexport function createOpenAIClient(runtime: IAgentRuntime): OpenAIProvider {\n  const baseURL = getBaseURL(runtime);\n  const apiKey = getApiKey(runtime);\n\n  if (!apiKey && isProxyMode(runtime)) {\n    return createOpenAI({\n      apiKey: PROXY_API_KEY,\n      baseURL,\n    });\n  }\n\n  if (!apiKey) {\n    throw new Error(\n      \"OPENAI_API_KEY is required. Set it in your environment variables or runtime settings.\"\n    );\n  }\n\n  return createOpenAI({\n    apiKey,\n    baseURL,\n  });\n}\n",
     "import type { IAgentRuntime, ModelTypeName } from \"@elizaos/core\";\nimport { ModelType } from \"@elizaos/core\";\nimport {\n  encodingForModel,\n  getEncoding,\n  type Tiktoken,\n  type TiktokenEncoding,\n  type TiktokenModel,\n} from \"js-tiktoken\";\nimport { getLargeModel, getSmallModel } from \"./config\";\n\ntype SupportedEncoding = \"cl100k_base\" | \"o200k_base\";\n\nfunction resolveTokenizerEncoding(modelName: string): Tiktoken {\n  const normalized = modelName.toLowerCase();\n  const fallbackEncoding: SupportedEncoding = normalized.includes(\"4o\")\n    ? \"o200k_base\"\n    : \"cl100k_base\";\n  try {\n    return encodingForModel(modelName as TiktokenModel);\n  } catch {\n    return getEncoding(fallbackEncoding as TiktokenEncoding);\n  }\n}\n\nfunction getModelName(runtime: IAgentRuntime, modelType: ModelTypeName): string {\n  if (modelType === ModelType.TEXT_SMALL) {\n    return getSmallModel(runtime);\n  }\n  return getLargeModel(runtime);\n}\n\nexport function tokenizeText(\n  runtime: IAgentRuntime,\n  modelType: ModelTypeName,\n  text: string\n): number[] {\n  const modelName = getModelName(runtime, modelType);\n  const encoder = resolveTokenizerEncoding(modelName);\n  return encoder.encode(text);\n}\n\nexport function detokenizeText(\n  runtime: IAgentRuntime,\n  modelType: ModelTypeName,\n  tokens: number[]\n): string {\n  const modelName = getModelName(runtime, modelType);\n  const encoder = resolveTokenizerEncoding(modelName);\n  return encoder.decode(tokens);\n}\n\nexport function countTokens(\n  runtime: IAgentRuntime,\n  modelType: ModelTypeName,\n  text: string\n): number {\n  const tokens = tokenizeText(runtime, modelType, text);\n  return tokens.length;\n}\n\nexport function truncateToTokenLimit(\n  runtime: IAgentRuntime,\n  modelType: ModelTypeName,\n  text: string,\n  maxTokens: number\n): string {\n  const tokens = tokenizeText(runtime, modelType, text);\n  if (tokens.length <= maxTokens) {\n    return text;\n  }\n  const truncatedTokens = tokens.slice(0, maxTokens);\n  return detokenizeText(runtime, modelType, truncatedTokens);\n}\n",
     "import type { DetokenizeTextParams, IAgentRuntime, TokenizeTextParams } from \"@elizaos/core\";\nimport { detokenizeText, tokenizeText } from \"../utils/tokenization\";\n\nexport async function handleTokenizerEncode(\n  runtime: IAgentRuntime,\n  params: TokenizeTextParams\n): Promise<number[]> {\n  if (!params.prompt) {\n    throw new Error(\"Tokenization requires a non-empty prompt\");\n  }\n  const modelType = params.modelType;\n  return tokenizeText(runtime, modelType, params.prompt);\n}\n\nexport async function handleTokenizerDecode(\n  runtime: IAgentRuntime,\n  params: DetokenizeTextParams\n): Promise<string> {\n  if (!params.tokens || !Array.isArray(params.tokens)) {\n    throw new Error(\"Detokenization requires a valid tokens array\");\n  }\n  if (params.tokens.length === 0) {\n    return \"\";\n  }\n  for (let i = 0; i < params.tokens.length; i++) {\n    const token = params.tokens[i];\n    if (typeof token !== \"number\" || !Number.isFinite(token)) {\n      throw new Error(`Invalid token at index ${i}: expected number`);\n    }\n  }\n  const modelType = params.modelType;\n  return detokenizeText(runtime, modelType, params.tokens);\n}\n",

package/dist/models/text.d.ts

@@ -12,7 +12,7 @@ declare function normalizeNativeMessages(messages: unknown): ModelMessage[] | un
  * Strip reasoning-only parts from outbound assistant content.
  *
  * OpenAI-spec reasoning models (Cerebras gpt-oss-120b, OpenAI o1/o3,
- * DeepSeek R1, Qwen-3-thinking, etc.) return reasoning in the assistant
+ * DeepSeek R1, and similar families) return reasoning in the assistant
  * response — either as a separate `reasoning` / `reasoning_content`
  * field, or as content parts with `type: "reasoning"`. Echoing those
  * back to the next turn is wrong on both ends:

package/dist/node/index.node.js.map

@@ -11,7 +11,7 @@
     "import type { IAgentRuntime, ModelTypeName } from \"@elizaos/core\";\nimport { EventType } from \"@elizaos/core\";\nimport type { TokenUsage } from \"../types\";\n\nconst MAX_PROMPT_LENGTH = 200;\n\ninterface ModelUsageEventPayload {\n  runtime: IAgentRuntime;\n  source: \"openai\";\n  provider: \"openai\";\n  type: ModelTypeName;\n  prompt: string;\n  tokens: {\n    prompt: number;\n    completion: number;\n    total: number;\n    cached?: number;\n  };\n}\n\ninterface AISDKUsage {\n  inputTokens?: number;\n  outputTokens?: number;\n  totalTokens?: number;\n  cachedInputTokens?: number;\n}\n\ninterface OpenAIAPIUsage {\n  promptTokens?: number;\n  completionTokens?: number;\n  totalTokens?: number;\n  cachedPromptTokens?: number;\n  promptTokensDetails?: {\n    cachedTokens?: number;\n  };\n}\n\ntype ModelUsage = TokenUsage | AISDKUsage | OpenAIAPIUsage;\n\nfunction truncatePrompt(prompt: string): string {\n  if (prompt.length <= MAX_PROMPT_LENGTH) {\n    return prompt;\n  }\n  return `${prompt.slice(0, MAX_PROMPT_LENGTH)}…`;\n}\n\nfunction normalizeUsage(usage: ModelUsage): TokenUsage {\n  if (\"promptTokens\" in usage) {\n    const promptTokensDetails =\n      \"promptTokensDetails\" in usage ? usage.promptTokensDetails : undefined;\n    const cachedPromptTokens = usage.cachedPromptTokens ?? promptTokensDetails?.cachedTokens;\n    return {\n      promptTokens: usage.promptTokens ?? 0,\n      completionTokens: usage.completionTokens ?? 0,\n      totalTokens: usage.totalTokens ?? (usage.promptTokens ?? 0) + (usage.completionTokens ?? 0),\n      cachedPromptTokens,\n    };\n  }\n  if (\"inputTokens\" in usage || \"outputTokens\" in usage) {\n    const input = (usage as AISDKUsage).inputTokens ?? 0;\n    const output = (usage as AISDKUsage).outputTokens ?? 0;\n    const total = (usage as AISDKUsage).totalTokens ?? input + output;\n    return {\n      promptTokens: input,\n      completionTokens: output,\n      totalTokens: total,\n      cachedPromptTokens: (usage as AISDKUsage).cachedInputTokens,\n    };\n  }\n  return {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n  };\n}\n\nexport function emitModelUsageEvent(\n  runtime: IAgentRuntime,\n  type: ModelTypeName,\n  prompt: string,\n  usage: ModelUsage\n): void {\n  const normalized = normalizeUsage(usage);\n\n  const payload: ModelUsageEventPayload = {\n    runtime,\n    source: \"openai\",\n    provider: \"openai\",\n    type,\n    prompt: truncatePrompt(prompt),\n    tokens: {\n      prompt: normalized.promptTokens,\n      completion: normalized.completionTokens,\n      total: normalized.totalTokens,\n      ...(normalized.cachedPromptTokens !== undefined\n        ? { cached: normalized.cachedPromptTokens }\n        : {}),\n    },\n  };\n\n  runtime.emitEvent(EventType.MODEL_USED, payload);\n}\n",
     "import type {\n  IAgentRuntime,\n  ImageDescriptionParams,\n  ImageGenerationParams,\n  RecordLlmCallDetails,\n} from \"@elizaos/core\";\nimport { logger, ModelType, recordLlmCall } from \"@elizaos/core\";\nimport type {\n  ImageDescriptionResult,\n  ImageGenerationResult,\n  ImageQuality,\n  ImageSize,\n  ImageStyle,\n  OpenAIChatCompletionResponse,\n  OpenAIImageGenerationResponse,\n} from \"../types\";\nimport {\n  getAuthHeader,\n  getBaseURL,\n  getImageDescriptionAuthHeader,\n  getImageDescriptionBaseURL,\n  getImageDescriptionMaxTokens,\n  getImageDescriptionModel,\n  getImageModel,\n} from \"../utils/config\";\nimport { emitModelUsageEvent } from \"../utils/events\";\n\ninterface ExtendedImageGenerationParams extends ImageGenerationParams {\n  quality?: ImageQuality;\n  style?: ImageStyle;\n}\n\nconst DEFAULT_IMAGE_DESCRIPTION_PROMPT =\n  \"Please analyze this image and provide a title and detailed description.\";\n\nexport async function handleImageGeneration(\n  runtime: IAgentRuntime,\n  params: ImageGenerationParams\n): Promise<ImageGenerationResult[]> {\n  const modelName = getImageModel(runtime);\n  const count = params.count ?? 1;\n  const size: ImageSize = (params.size as ImageSize) ?? \"1024x1024\";\n  const extendedParams = params as ExtendedImageGenerationParams;\n\n  logger.debug(`[OpenAI] Using IMAGE model: ${modelName}`);\n\n  if (typeof params.prompt !== \"string\" || params.prompt.trim().length === 0) {\n    throw new Error(\"IMAGE generation requires a non-empty prompt\");\n  }\n\n  if (count < 1 || count > 10) {\n    throw new Error(\"IMAGE count must be between 1 and 10\");\n  }\n\n  const baseURL = getBaseURL(runtime);\n\n  const requestBody: Record<string, string | number> = {\n    model: modelName,\n    prompt: params.prompt,\n    n: count,\n    size,\n  };\n\n  if (extendedParams.quality) {\n    requestBody.quality = extendedParams.quality;\n  }\n  if (extendedParams.style) {\n    requestBody.style = extendedParams.style;\n  }\n\n  const details: RecordLlmCallDetails = {\n    model: modelName,\n    systemPrompt: \"\",\n    userPrompt: params.prompt,\n    temperature: 0,\n    maxTokens: 0,\n    purpose: \"external_llm\",\n    actionType: \"openai.images.generate\",\n  };\n  const data = await recordLlmCall(runtime, details, async () => {\n    const response = await fetch(`${baseURL}/images/generations`, {\n      method: \"POST\",\n      headers: {\n        ...getAuthHeader(runtime),\n        \"Content-Type\": \"application/json\",\n      },\n      body: JSON.stringify(requestBody),\n    });\n\n    if (!response.ok) {\n      const errorText = await response.text().catch(() => \"Unknown error\");\n      throw new Error(\n        `OpenAI image generation failed: ${response.status} ${response.statusText} - ${errorText}`\n      );\n    }\n\n    const responseData = (await response.json()) as OpenAIImageGenerationResponse;\n    details.response = JSON.stringify(responseData.data);\n    return responseData;\n  });\n\n  if (data.data.length === 0) {\n    throw new Error(\"OpenAI API returned no images\");\n  }\n\n  return data.data.map((item) => ({\n    url: item.url,\n    revisedPrompt: item.revised_prompt,\n  }));\n}\n\nfunction parseTitleFromResponse(content: string): string {\n  const titleMatch = content.match(/title[:\\s]+(.+?)(?:\\n|$)/i);\n  return titleMatch?.[1]?.trim() ?? \"Image Analysis\";\n}\n\nfunction parseDescriptionFromResponse(content: string): string {\n  return content.replace(/title[:\\s]+(.+?)(?:\\n|$)/i, \"\").trim();\n}\n\nexport async function handleImageDescription(\n  runtime: IAgentRuntime,\n  params: ImageDescriptionParams | string\n): Promise<ImageDescriptionResult> {\n  const modelName = getImageDescriptionModel(runtime);\n  const paramsWithMaxTokens = params as ImageDescriptionParams & { maxTokens?: number };\n  const maxTokens =\n    typeof params === \"object\" && typeof paramsWithMaxTokens.maxTokens === \"number\"\n      ? paramsWithMaxTokens.maxTokens\n      : getImageDescriptionMaxTokens(runtime);\n\n  logger.debug(`[OpenAI] Using IMAGE_DESCRIPTION model: ${modelName}`);\n\n  let imageUrl: string;\n  let promptText: string;\n\n  if (typeof params === \"string\") {\n    imageUrl = params;\n    promptText = DEFAULT_IMAGE_DESCRIPTION_PROMPT;\n  } else {\n    imageUrl = params.imageUrl;\n    promptText = params.prompt ?? DEFAULT_IMAGE_DESCRIPTION_PROMPT;\n  }\n\n  if (!imageUrl || imageUrl.trim().length === 0) {\n    throw new Error(\"IMAGE_DESCRIPTION requires a valid image URL\");\n  }\n\n  const baseURL = getImageDescriptionBaseURL(runtime);\n\n  const requestBody = {\n    model: modelName,\n    messages: [\n      {\n        role: \"user\",\n        content: [\n          { type: \"text\", text: promptText },\n          { type: \"image_url\", image_url: { url: imageUrl } },\n        ],\n      },\n    ],\n    max_tokens: maxTokens,\n  };\n\n  const details: RecordLlmCallDetails = {\n    model: modelName,\n    systemPrompt: \"\",\n    userPrompt: promptText,\n    temperature: 0,\n    maxTokens,\n    purpose: \"external_llm\",\n    actionType: \"openai.chat.completions.create\",\n  };\n  const data = await recordLlmCall(runtime, details, async () => {\n    const response = await fetch(`${baseURL}/chat/completions`, {\n      method: \"POST\",\n      headers: {\n        ...getImageDescriptionAuthHeader(runtime),\n        \"Content-Type\": \"application/json\",\n      },\n      body: JSON.stringify(requestBody),\n    });\n\n    if (!response.ok) {\n      const errorText = await response.text().catch(() => \"Unknown error\");\n      throw new Error(\n        `OpenAI image description failed: ${response.status} ${response.statusText} - ${errorText}`\n      );\n    }\n\n    const responseData = (await response.json()) as OpenAIChatCompletionResponse;\n    const responseContent = responseData.choices[0]?.message.content;\n    if (!responseContent) {\n      throw new Error(\"OpenAI API returned empty image description\");\n    }\n    details.response = responseContent;\n    if (responseData.usage) {\n      details.promptTokens = responseData.usage.prompt_tokens;\n      details.completionTokens = responseData.usage.completion_tokens;\n    }\n    return responseData;\n  });\n\n  if (data.usage) {\n    emitModelUsageEvent(\n      runtime,\n      ModelType.IMAGE_DESCRIPTION,\n      typeof params === \"string\" ? params : (params.prompt ?? \"\"),\n      {\n        promptTokens: data.usage.prompt_tokens,\n        completionTokens: data.usage.completion_tokens,\n        totalTokens: data.usage.total_tokens,\n      }\n    );\n  }\n\n  const firstChoice = data.choices[0];\n  const content = firstChoice?.message.content;\n\n  if (!content) {\n    throw new Error(\"OpenAI API returned empty image description\");\n  }\n\n  return {\n    title: parseTitleFromResponse(content),\n    description: parseDescriptionFromResponse(content),\n  };\n}\n",
     "/**\n * Deep Research model handler\n *\n * Provides deep research capabilities using OpenAI's o3-deep-research and o4-mini-deep-research models.\n * These models can find, analyze, and synthesize hundreds of sources to create comprehensive reports.\n *\n * @see https://platform.openai.com/docs/guides/deep-research\n */\n\nimport type {\n  IAgentRuntime,\n  JsonValue,\n  RecordLlmCallDetails,\n  ResearchAnnotation,\n  ResearchCodeInterpreterCall,\n  ResearchFileSearchCall,\n  ResearchMcpToolCall,\n  ResearchMessageOutput,\n  ResearchOutputItem,\n  ResearchParams,\n  ResearchResult,\n  ResearchTool,\n  ResearchWebSearchCall,\n} from \"@elizaos/core\";\nimport { logger, recordLlmCall } from \"@elizaos/core\";\nimport { getApiKey, getBaseURL, getResearchModel, getResearchTimeout } from \"../utils/config\";\n\n// ============================================================================\n// Types for OpenAI Responses API\n// ============================================================================\n\n/**\n * Tool configuration for the Responses API\n */\ninterface ResponsesApiTool {\n  type: \"web_search_preview\" | \"file_search\" | \"code_interpreter\" | \"mcp\";\n  vector_store_ids?: string[];\n  container?: { type: \"auto\" };\n  server_label?: string;\n  server_url?: string;\n  require_approval?: \"never\";\n}\n\n/**\n * Raw response from the OpenAI Responses API\n */\ninterface ResponsesApiResponse {\n  id: string;\n  object: string;\n  status?: \"queued\" | \"in_progress\" | \"completed\" | \"failed\";\n  output?: ResponsesApiOutputItem[];\n  output_text?: string;\n  error?: {\n    message: string;\n    code: string;\n  };\n}\n\n/**\n * Raw output item from the Responses API\n */\ninterface ResponsesApiOutputItem {\n  id?: string;\n  type: string;\n  status?: string;\n  action?: {\n    type: string;\n    query?: string;\n    url?: string;\n  };\n  query?: string;\n  results?: Array<{\n    file_id: string;\n    file_name: string;\n    score: number;\n  }>;\n  code?: string;\n  output?: string;\n  server_label?: string;\n  tool_name?: string;\n  arguments?: Record<string, unknown>;\n  result?: unknown;\n  content?: Array<{\n    type: string;\n    text: string;\n    annotations?: Array<{\n      url: string;\n      title: string;\n      start_index: number;\n      end_index: number;\n    }>;\n  }>;\n}\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\n/**\n * Converts ResearchTool params to Responses API tool format\n */\nfunction convertToolToApi(tool: ResearchTool): ResponsesApiTool {\n  switch (tool.type) {\n    case \"web_search_preview\":\n      return { type: \"web_search_preview\" };\n    case \"file_search\":\n      return {\n        type: \"file_search\",\n        vector_store_ids: tool.vectorStoreIds,\n      };\n    case \"code_interpreter\":\n      return {\n        type: \"code_interpreter\",\n        container: tool.container ?? { type: \"auto\" },\n      };\n    case \"mcp\":\n      return {\n        type: \"mcp\",\n        server_label: tool.serverLabel,\n        server_url: tool.serverUrl,\n        require_approval: tool.requireApproval ?? \"never\",\n      };\n    default:\n      throw new Error(`Unknown research tool type: ${(tool as ResearchTool).type}`);\n  }\n}\n\n/**\n * Converts raw API output items to typed ResearchOutputItem\n */\nfunction convertOutputItem(item: ResponsesApiOutputItem): ResearchOutputItem | null {\n  switch (item.type) {\n    case \"web_search_call\":\n      return {\n        id: item.id ?? \"\",\n        type: \"web_search_call\",\n        status: (item.status as \"completed\" | \"failed\") ?? \"completed\",\n        action: {\n          type: (item.action?.type as \"search\" | \"open_page\" | \"find_in_page\") ?? \"search\",\n          query: item.action?.query,\n          url: item.action?.url,\n        },\n      } satisfies ResearchWebSearchCall;\n\n    case \"file_search_call\":\n      return {\n        id: item.id ?? \"\",\n        type: \"file_search_call\",\n        status: (item.status as \"completed\" | \"failed\") ?? \"completed\",\n        query: item.query ?? \"\",\n        results: item.results?.map((r) => ({\n          fileId: r.file_id,\n          fileName: r.file_name,\n          score: r.score,\n        })),\n      } satisfies ResearchFileSearchCall;\n\n    case \"code_interpreter_call\":\n      return {\n        id: item.id ?? \"\",\n        type: \"code_interpreter_call\",\n        status: (item.status as \"completed\" | \"failed\") ?? \"completed\",\n        code: item.code ?? \"\",\n        output: item.output,\n      } satisfies ResearchCodeInterpreterCall;\n\n    case \"mcp_tool_call\":\n      return {\n        id: item.id ?? \"\",\n        type: \"mcp_tool_call\",\n        status: (item.status as \"completed\" | \"failed\") ?? \"completed\",\n        serverLabel: item.server_label ?? \"\",\n        toolName: item.tool_name ?? \"\",\n        arguments: (item.arguments ?? {}) as Record<string, JsonValue>,\n        result: item.result as JsonValue,\n      } satisfies ResearchMcpToolCall;\n\n    case \"message\":\n      return {\n        type: \"message\",\n        content:\n          item.content?.map((c) => ({\n            type: \"output_text\" as const,\n            text: c.text,\n            annotations:\n              c.annotations?.map((a) => ({\n                url: a.url,\n                title: a.title,\n                startIndex: a.start_index,\n                endIndex: a.end_index,\n              })) ?? [],\n          })) ?? [],\n      } satisfies ResearchMessageOutput;\n\n    default:\n      // Unknown output type, skip\n      return null;\n  }\n}\n\n/**\n * Extracts text and annotations from the response\n */\nfunction extractTextAndAnnotations(response: ResponsesApiResponse): {\n  text: string;\n  annotations: ResearchAnnotation[];\n} {\n  // Try output_text first (convenience field)\n  if (response.output_text) {\n    // Find annotations from message output items\n    const annotations: ResearchAnnotation[] = [];\n    if (response.output) {\n      for (const item of response.output) {\n        if (item.type === \"message\" && item.content) {\n          for (const content of item.content) {\n            if (content.annotations) {\n              for (const ann of content.annotations) {\n                annotations.push({\n                  url: ann.url,\n                  title: ann.title,\n                  startIndex: ann.start_index,\n                  endIndex: ann.end_index,\n                });\n              }\n            }\n          }\n        }\n      }\n    }\n    return { text: response.output_text, annotations };\n  }\n\n  // Fall back to extracting from message output items\n  let text = \"\";\n  const annotations: ResearchAnnotation[] = [];\n\n  if (response.output) {\n    for (const item of response.output) {\n      if (item.type === \"message\" && item.content) {\n        for (const content of item.content) {\n          text += content.text;\n          if (content.annotations) {\n            for (const ann of content.annotations) {\n              annotations.push({\n                url: ann.url,\n                title: ann.title,\n                startIndex: ann.start_index,\n                endIndex: ann.end_index,\n              });\n            }\n          }\n        }\n      }\n    }\n  }\n\n  return { text, annotations };\n}\n\n// ============================================================================\n// Main Handler\n// ============================================================================\n\n/**\n * Handles RESEARCH model requests using OpenAI's deep research models.\n *\n * Deep research models can take tens of minutes to complete tasks.\n * Use background mode for long-running tasks.\n *\n * @param runtime - The agent runtime\n * @param params - Research parameters\n * @returns Research result with text, annotations, and output items\n *\n * @example\n * ```typescript\n * const result = await handleResearch(runtime, {\n *   input: \"Research the economic impact of AI on global labor markets\",\n *   tools: [\n *     { type: \"web_search_preview\" },\n *     { type: \"code_interpreter\", container: { type: \"auto\" } }\n *   ],\n *   background: true,\n * });\n * console.log(result.text);\n * ```\n */\nexport async function handleResearch(\n  runtime: IAgentRuntime,\n  params: ResearchParams\n): Promise<ResearchResult> {\n  const apiKey = getApiKey(runtime);\n  if (!apiKey) {\n    throw new Error(\n      \"OPENAI_API_KEY is required for deep research. Set it in your environment variables or runtime settings.\"\n    );\n  }\n\n  const baseURL = getBaseURL(runtime);\n  const modelName = params.model ?? getResearchModel(runtime);\n  const timeout = getResearchTimeout(runtime);\n\n  logger.debug(`[OpenAI] Starting deep research with model: ${modelName}`);\n  logger.debug(`[OpenAI] Research input: ${params.input.substring(0, 100)}...`);\n\n  // Validate that at least one data source tool is provided\n  const dataSourceTools = params.tools?.filter(\n    (t) => t.type === \"web_search_preview\" || t.type === \"file_search\" || t.type === \"mcp\"\n  );\n\n  if (!dataSourceTools || dataSourceTools.length === 0) {\n    // Default to web search if no tools specified\n    logger.debug(\"[OpenAI] No data source tools specified, defaulting to web_search_preview\");\n    params.tools = [{ type: \"web_search_preview\" }, ...(params.tools ?? [])];\n  }\n\n  // Build the request body for the Responses API\n  const requestBody: Record<string, unknown> = {\n    model: modelName,\n    input: params.input,\n  };\n\n  if (params.instructions) {\n    requestBody.instructions = params.instructions;\n  }\n\n  if (params.background !== undefined) {\n    requestBody.background = params.background;\n  }\n\n  if (params.tools && params.tools.length > 0) {\n    requestBody.tools = params.tools.map(convertToolToApi);\n  }\n\n  if (params.maxToolCalls !== undefined) {\n    requestBody.max_tool_calls = params.maxToolCalls;\n  }\n\n  if (params.reasoningSummary) {\n    requestBody.reasoning = { summary: params.reasoningSummary };\n  }\n\n  logger.debug(`[OpenAI] Research request body: ${JSON.stringify(requestBody, null, 2)}`);\n\n  const details: RecordLlmCallDetails = {\n    model: modelName,\n    systemPrompt: params.instructions ?? \"\",\n    userPrompt: params.input,\n    temperature: 0,\n    maxTokens: 0,\n    purpose: \"external_llm\",\n    actionType: \"openai.responses.create\",\n  };\n  const data = await recordLlmCall(runtime, details, async () => {\n    // Make the API request\n    const response = await fetch(`${baseURL}/responses`, {\n      method: \"POST\",\n      headers: {\n        Authorization: `Bearer ${apiKey}`,\n        \"Content-Type\": \"application/json\",\n      },\n      body: JSON.stringify(requestBody),\n      signal: AbortSignal.timeout(timeout),\n    });\n\n    if (!response.ok) {\n      const errorText = await response.text();\n      logger.error(`[OpenAI] Research request failed: ${response.status} ${errorText}`);\n      throw new Error(`Deep research request failed: ${response.status} ${response.statusText}`);\n    }\n\n    const responseData = (await response.json()) as ResponsesApiResponse;\n    details.response = responseData.output_text ?? \"\";\n    return responseData;\n  });\n\n  if (data.error) {\n    logger.error(`[OpenAI] Research API error: ${data.error.message}`);\n    throw new Error(`Deep research error: ${data.error.message}`);\n  }\n\n  logger.debug(`[OpenAI] Research response received. Status: ${data.status ?? \"completed\"}`);\n\n  // Extract text and annotations\n  const { text, annotations } = extractTextAndAnnotations(data);\n\n  // Convert output items\n  const outputItems: ResearchOutputItem[] = [];\n  if (data.output) {\n    for (const item of data.output) {\n      const converted = convertOutputItem(item);\n      if (converted) {\n        outputItems.push(converted);\n      }\n    }\n  }\n\n  const result: ResearchResult = {\n    id: data.id,\n    text,\n    annotations,\n    outputItems,\n    status: data.status,\n  };\n\n  logger.info(\n    `[OpenAI] Research completed. Text length: ${text.length}, Annotations: ${annotations.length}, Output items: ${outputItems.length}`\n  );\n\n  return result;\n}\n",
-    "/**\n * Text generation model handlers\n *\n * Provides text generation using OpenAI's language models.\n */\n\nimport type {\n  GenerateTextParams,\n  IAgentRuntime,\n  JsonValue,\n  ModelTypeName,\n  RecordLlmCallDetails,\n} from \"@elizaos/core\";\nimport {\n  buildCanonicalSystemPrompt,\n  dropDuplicateLeadingSystemMessage,\n  logger,\n  ModelType,\n  normalizeSchemaForCerebras,\n  recordLlmCall,\n  resolveEffectiveSystemPrompt,\n  sanitizeFunctionNameForCerebras,\n} from \"@elizaos/core\";\nimport {\n  generateText,\n  type JSONSchema7,\n  jsonSchema,\n  type LanguageModelUsage,\n  type ModelMessage,\n  Output,\n  streamText,\n  type ToolChoice,\n  type ToolSet,\n  type UserContent,\n} from \"ai\";\nimport { createOpenAIClient } from \"../providers\";\nimport type { TextStreamResult, TokenUsage } from \"../types\";\nimport {\n  getActionPlannerModel,\n  getExperimentalTelemetry,\n  getLargeModel,\n  getMediumModel,\n  getMegaModel,\n  getNanoModel,\n  getResponseHandlerModel,\n  getSmallModel,\n  isCerebrasMode,\n} from \"../utils/config\";\nimport { emitModelUsageEvent } from \"../utils/events\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\n/**\n * Function to get model name from runtime\n */\ntype ModelNameGetter = (runtime: IAgentRuntime) => string;\n\ntype PromptCacheRetention = \"in_memory\" | \"24h\";\ntype ChatAttachment = {\n  data: string | Uint8Array | URL;\n  mediaType: string;\n  filename?: string;\n};\n\ninterface OpenAIPromptCacheOptions {\n  promptCacheKey?: string;\n  promptCacheRetention?: PromptCacheRetention;\n}\n\ninterface GenerateTextParamsWithOpenAIOptions\n  extends Omit<\n    GenerateTextParams,\n    \"messages\" | \"tools\" | \"toolChoice\" | \"responseSchema\" | \"providerOptions\"\n  > {\n  model?: string;\n  attachments?: ChatAttachment[];\n  messages?: unknown[];\n  tools?: unknown;\n  toolChoice?: unknown;\n  responseSchema?: unknown;\n  providerOptions?: Record<string, object | JsonValue> & {\n    agentName?: string;\n    openai?: OpenAIPromptCacheOptions;\n  };\n}\n\ntype NativeOutput = NonNullable<Parameters<typeof generateText<ToolSet>>[0][\"output\"]>;\ntype NativeGenerateTextParams = Parameters<typeof generateText<ToolSet, NativeOutput>>[0];\ntype NativeStreamTextParams = Parameters<typeof streamText<ToolSet, NativeOutput>>[0];\ntype NativePrompt =\n  | { prompt: string; messages?: never }\n  | { messages: ModelMessage[]; prompt?: never };\ntype NativeTextParams = Omit<NativeGenerateTextParams, \"messages\" | \"prompt\"> &\n  Omit<NativeStreamTextParams, \"messages\" | \"prompt\"> &\n  NativePrompt & {\n    // Re-declared explicitly: TypeScript's `Parameters<typeof generateText>`\n    // inference produces an overload-union that drops this field, but the\n    // ai SDK's runtime signature accepts it (see ai@6 `CallSettings & Prompt`).\n    allowSystemInMessages?: boolean;\n  };\ntype NativeProviderOptions = NativeTextParams[\"providerOptions\"];\ntype NativeTelemetrySettings = NativeTextParams[\"experimental_telemetry\"];\n\ntype LanguageModelUsageWithCache = Omit<LanguageModelUsage, \"inputTokenDetails\"> & {\n  inputTokenDetails?: LanguageModelUsage[\"inputTokenDetails\"] & {\n    cachedInputTokens?: number;\n    cacheCreationInputTokens?: number;\n    cacheCreationTokens?: number;\n  };\n  cachedInputTokens?: number;\n  cacheReadInputTokens?: number;\n  cacheCreationInputTokens?: number;\n  cacheWriteInputTokens?: number;\n  input_tokens_details?: {\n    cached_tokens?: number;\n    cache_read_input_tokens?: number;\n    cache_creation_input_tokens?: number;\n  };\n  prompt_tokens_details?: {\n    cached_tokens?: number;\n  };\n};\n\ninterface NativeGenerateTextResult {\n  text: string;\n  toolCalls?: unknown[];\n  finishReason?: string;\n  usage?: TokenUsage;\n  providerMetadata?: unknown;\n}\n\ntype NativeTextModelResult = string & NativeGenerateTextResult;\n\nconst TEXT_NANO_MODEL_TYPE = ModelType.TEXT_NANO as ModelTypeName;\nconst TEXT_MEDIUM_MODEL_TYPE = ModelType.TEXT_MEDIUM as ModelTypeName;\nconst TEXT_MEGA_MODEL_TYPE = ModelType.TEXT_MEGA as ModelTypeName;\nconst RESPONSE_HANDLER_MODEL_TYPE = ModelType.RESPONSE_HANDLER as ModelTypeName;\nconst ACTION_PLANNER_MODEL_TYPE = ModelType.ACTION_PLANNER as ModelTypeName;\n\nfunction resolveRequestedModelName(\n  params: GenerateTextParamsWithOpenAIOptions,\n  runtime: IAgentRuntime,\n  getModelFn: ModelNameGetter\n): string {\n  return typeof params.model === \"string\" && params.model.trim().length > 0\n    ? params.model.trim()\n    : getModelFn(runtime);\n}\n\nfunction buildUserContent(params: GenerateTextParamsWithOpenAIOptions): UserContent {\n  const content: Array<\n    | { type: \"text\"; text: string }\n    | {\n        type: \"file\";\n        data: string | Uint8Array | URL;\n        mediaType: string;\n        filename?: string;\n      }\n  > = [{ type: \"text\", text: params.prompt ?? \"\" }];\n\n  for (const attachment of params.attachments ?? []) {\n    content.push({\n      type: \"file\",\n      data: attachment.data,\n      mediaType: attachment.mediaType,\n      ...(attachment.filename ? { filename: attachment.filename } : {}),\n    });\n  }\n\n  return content;\n}\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\n/**\n * Converts AI SDK usage to our token usage format.\n *\n * Emits both the legacy `cachedPromptTokens` (kept for back-compat with\n * existing OpenAI consumers) and the canonical v5 `cacheReadInputTokens`\n * (consumed by the trajectory recorder + cost table). They always carry the\n * same value when the AI SDK reports cached input.\n */\nfunction convertUsage(usage: LanguageModelUsage | undefined): TokenUsage | undefined {\n  if (!usage) {\n    return undefined;\n  }\n\n  // The AI SDK uses inputTokens/outputTokens\n  const promptTokens = usage.inputTokens ?? 0;\n  const completionTokens = usage.outputTokens ?? 0;\n  const usageWithCache: LanguageModelUsageWithCache = usage;\n  const cachedInput =\n    firstNumber(\n      usageWithCache.cacheReadInputTokens,\n      usageWithCache.cachedInputTokens,\n      usageWithCache.inputTokenDetails?.cacheReadTokens,\n      usageWithCache.inputTokenDetails?.cachedInputTokens,\n      usageWithCache.input_tokens_details?.cache_read_input_tokens,\n      usageWithCache.input_tokens_details?.cached_tokens,\n      usageWithCache.prompt_tokens_details?.cached_tokens\n    ) ?? undefined;\n  const cacheCreationInput = firstNumber(\n    usageWithCache.cacheCreationInputTokens,\n    usageWithCache.cacheWriteInputTokens,\n    usageWithCache.inputTokenDetails?.cacheCreationInputTokens,\n    usageWithCache.inputTokenDetails?.cacheCreationTokens,\n    usageWithCache.inputTokenDetails?.cacheWriteTokens,\n    usageWithCache.input_tokens_details?.cache_creation_input_tokens\n  );\n\n  return {\n    promptTokens,\n    completionTokens,\n    totalTokens: promptTokens + completionTokens,\n    cachedPromptTokens: cachedInput,\n    cacheReadInputTokens: cachedInput,\n    cacheCreationInputTokens: cacheCreationInput,\n  };\n}\n\nfunction firstNumber(...values: unknown[]): number | undefined {\n  for (const value of values) {\n    if (typeof value === \"number\" && Number.isFinite(value)) {\n      return value;\n    }\n    if (typeof value === \"string\" && value.trim().length > 0) {\n      const parsed = Number(value);\n      if (Number.isFinite(parsed)) {\n        return parsed;\n      }\n    }\n  }\n  return undefined;\n}\n\nfunction resolvePromptCacheOptions(params: GenerateTextParams): OpenAIPromptCacheOptions {\n  const withOpenAIOptions = params as GenerateTextParamsWithOpenAIOptions;\n  return {\n    promptCacheKey: withOpenAIOptions.providerOptions?.openai?.promptCacheKey,\n    promptCacheRetention: withOpenAIOptions.providerOptions?.openai?.promptCacheRetention,\n  };\n}\n\n/**\n * Forward `OPENAI_REASONING_EFFORT` (runtime setting / process.env) as\n * `reasoning_effort` on the outbound chat completions request. This is\n * the OpenAI-spec knob for reasoning-capable models (`o1-*`, `o3-*`,\n * `gpt-oss-*`, `deepseek-r1`, `qwen-3-thinking`, etc.) — including\n * Cerebras and OpenRouter, which honor the same field. `\"low\"` keeps\n * reasoning short enough that visible content always fits inside\n * `max_tokens`, which is the failure mode on Cerebras gpt-oss-120b when\n * left unset.\n *\n * In Cerebras mode the field defaults to `\"low\"` when unset, but ONLY for\n * reasoning-capable models (e.g. gpt-oss-*, deepseek-r1, qwen-3-thinking):\n * gpt-oss-120b emits a separate reasoning channel and, left unbounded, spends\n * the whole token budget reasoning — returning empty visible content, which\n * makes the agent fall back to \"I don't have a reply for that\". `\"low\"` keeps\n * reasoning short so a reply always materializes. Non-reasoning Cerebras models\n * (Llama, etc.) reject `reasoning_effort`, so they must never receive the\n * default. For all other models an unset/invalid value yields `undefined`, so\n * they pay no overhead and the wire stays clean. An explicit valid\n * `OPENAI_REASONING_EFFORT` always wins.\n *\n * Valid values follow the OpenAI spec exactly: `minimal`, `low`,\n * `medium`, `high`. Anything else is logged and ignored.\n */\ntype ReasoningEffort = \"minimal\" | \"low\" | \"medium\" | \"high\";\n\nconst VALID_REASONING_EFFORTS: readonly ReasoningEffort[] = [\"minimal\", \"low\", \"medium\", \"high\"];\n\n/**\n * Reasoning-capable model families that emit a separate reasoning channel and\n * honor `reasoning_effort`. Used to gate the Cerebras `\"low\"` default so\n * non-reasoning models (Llama, etc.) are never sent the field.\n */\nfunction isReasoningModel(modelName: string | undefined): boolean {\n  if (!modelName) return false;\n  const m = modelName.toLowerCase();\n  return (\n    m.includes(\"gpt-oss\") ||\n    m.includes(\"o1\") ||\n    m.includes(\"o3\") ||\n    m.includes(\"o4\") ||\n    m.includes(\"deepseek-r1\") ||\n    m.includes(\"thinking\") ||\n    m.includes(\"reasoning\") ||\n    m.includes(\"qwq\")\n  );\n}\n\nfunction resolveReasoningEffort(\n  runtime: IAgentRuntime,\n  modelName?: string\n): ReasoningEffort | undefined {\n  const raw = runtime.getSetting(\"OPENAI_REASONING_EFFORT\");\n  const normalized = typeof raw === \"string\" ? raw.trim().toLowerCase() : \"\";\n  if (normalized) {\n    if ((VALID_REASONING_EFFORTS as readonly string[]).includes(normalized)) {\n      return normalized as ReasoningEffort;\n    }\n    logger.warn(\n      `[OpenAI] OPENAI_REASONING_EFFORT=${raw} is not a valid reasoning effort; ignoring. Expected one of: ${VALID_REASONING_EFFORTS.join(\", \")}.`\n    );\n  }\n  // gpt-oss-120b on Cerebras returns empty content when reasoning runs\n  // unbounded; default to \"low\" so a visible reply always fits — but only for\n  // reasoning-capable models. Non-reasoning Cerebras models (Llama, etc.)\n  // reject `reasoning_effort` and would break. An explicit valid value above\n  // wins over this default.\n  if (isCerebrasMode(runtime) && isReasoningModel(modelName)) {\n    return \"low\";\n  }\n  return undefined;\n}\n\nfunction resolveProviderOptions(\n  params: GenerateTextParams,\n  runtime: IAgentRuntime,\n  modelName?: string\n): Record<string, unknown> | undefined {\n  const withOpenAIOptions = params as GenerateTextParamsWithOpenAIOptions;\n  const rawProviderOptions = withOpenAIOptions.providerOptions;\n  const promptCacheOptions = resolvePromptCacheOptions(params);\n  const reasoningEffort = resolveReasoningEffort(runtime, modelName);\n\n  if (\n    !rawProviderOptions &&\n    !promptCacheOptions.promptCacheKey &&\n    !promptCacheOptions.promptCacheRetention &&\n    !reasoningEffort\n  ) {\n    return undefined;\n  }\n\n  // Cerebras supports prompt caching on gpt-oss-120b — 128-token blocks,\n  // default-on. The `prompt_cache_key` field IS accepted by Cerebras's\n  // OpenAI-compatible endpoint and surfaces hit counts via\n  // `usage.prompt_tokens_details.cached_tokens` (same shape as OpenAI), so\n  // we keep it in the request body. Only `prompt_cache_retention` is an\n  // OpenAI-direct-only field that Cerebras rejects with HTTP 400\n  // (`wrong_api_format`), so we strip just that one when in Cerebras mode.\n  const skipCacheRetention = isCerebrasMode(runtime);\n\n  const { agentName: _agentName, openai: rawOpenAIOptions, ...rest } = rawProviderOptions ?? {};\n  // When on Cerebras, scrub OpenAI-direct-only fields (e.g. `promptCacheRetention`)\n  // from `rawOpenAIOptions` before they're spread; otherwise they reach the wire\n  // and the Cerebras endpoint rejects with HTTP 400 `wrong_api_format`.\n  const sanitizedRawOpenAIOptions = (() => {\n    if (!rawOpenAIOptions || typeof rawOpenAIOptions !== \"object\") return rawOpenAIOptions;\n    if (!skipCacheRetention) return rawOpenAIOptions;\n    const { promptCacheRetention: _drop, ...rest2 } = rawOpenAIOptions as Record<string, unknown>;\n    return rest2;\n  })();\n  const openaiOptions = {\n    ...(sanitizedRawOpenAIOptions ?? {}),\n    ...(promptCacheOptions.promptCacheKey\n      ? { promptCacheKey: promptCacheOptions.promptCacheKey }\n      : {}),\n    ...(!skipCacheRetention && promptCacheOptions.promptCacheRetention\n      ? { promptCacheRetention: promptCacheOptions.promptCacheRetention }\n      : {}),\n    // The caller's explicit `reasoningEffort` wins over the resolved default\n    // (env var, or Cerebras \"low\") — same precedence pattern as promptCacheKey.\n    ...((sanitizedRawOpenAIOptions as { reasoningEffort?: unknown } | undefined)\n      ?.reasoningEffort === undefined && reasoningEffort\n      ? { reasoningEffort }\n      : {}),\n  };\n\n  const providerOptions = {\n    ...rest,\n    ...(Object.keys(openaiOptions).length > 0 ? { openai: openaiOptions } : {}),\n  };\n\n  return Object.keys(providerOptions).length > 0 ? providerOptions : undefined;\n}\n\nfunction buildStructuredOutput(responseSchema: unknown): NativeOutput {\n  if (\n    responseSchema &&\n    typeof responseSchema === \"object\" &&\n    \"responseFormat\" in responseSchema &&\n    \"parseCompleteOutput\" in responseSchema\n  ) {\n    return responseSchema as NativeOutput;\n  }\n\n  const schemaOptions =\n    responseSchema && typeof responseSchema === \"object\" && \"schema\" in responseSchema\n      ? (responseSchema as { schema: unknown; name?: string; description?: string })\n      : { schema: responseSchema };\n\n  return Output.object({\n    schema: jsonSchema(sanitizeJsonSchema(schemaOptions.schema, true)),\n    ...(schemaOptions.name ? { name: schemaOptions.name } : {}),\n    ...(schemaOptions.description ? { description: schemaOptions.description } : {}),\n  }) as NativeOutput;\n}\n\nfunction normalizeNativeTools(\n  tools: unknown,\n  options: { cerebrasMode?: boolean } = {}\n): ToolSet | undefined {\n  if (!tools) {\n    return undefined;\n  }\n\n  // Existing AI SDK callers already pass a ToolSet keyed by tool name. Keep it\n  // intact so custom tool instances, execute hooks, and dynamic tool metadata\n  // are preserved.\n  if (!Array.isArray(tools)) {\n    return tools as ToolSet;\n  }\n\n  const toolSet: Record<string, unknown> = {};\n\n  for (const rawTool of tools) {\n    const tool = asRecord(rawTool);\n    const functionTool = asRecord(tool.function);\n    const name = firstString(tool.name, functionTool.name);\n\n    if (!name) {\n      throw new Error(\"[OpenAI] Native tool definition is missing a name.\");\n    }\n\n    const description = firstString(tool.description, functionTool.description);\n    // Default to a permissive object schema. The empty-properties shape\n    // (`{ type: \"object\", properties: {}, additionalProperties: false }`) is\n    // accepted by OpenAI but rejected by strict-grammar providers like\n    // Cerebras with `Object fields require at least one of: 'properties' or\n    // 'anyOf' with a list of possible properties`.\n    const rawSchema =\n      tool.parameters ?? functionTool.parameters ?? ({ type: \"object\" } satisfies JSONSchema7);\n    let inputSchema = sanitizeJsonSchema(rawSchema, true);\n    if (options.cerebrasMode) {\n      // User-supplied schemas may still contain empty-properties subobjects\n      // even after sanitizeJsonSchema. Apply Cerebras-specific normalization\n      // recursively so deep schemas are accepted by the grammar compiler.\n      // Pass isRoot: true so the top-level invariant is enforced (must be\n      // type:\"object\" with no root oneOf/anyOf/enum/not).\n      inputSchema = normalizeSchemaForCerebras(inputSchema, true) as JSONSchema7;\n    }\n\n    // Cerebras's grammar compiler rejects function names containing characters\n    // outside `[a-zA-Z0-9_-]` (e.g. `math.factorial`). The AI SDK looks up\n    // tools by the registered key, so we register under the sanitized name AND\n    // surface it to the model under that name. Tool calls come back with the\n    // sanitized name, which the runtime resolves through its action registry —\n    // any caller relying on dotted action names should pre-sanitize.\n    const registeredName = options.cerebrasMode ? sanitizeFunctionNameForCerebras(name) : name;\n\n    toolSet[registeredName] = {\n      ...(description ? { description } : {}),\n      inputSchema: jsonSchema(inputSchema as JSONSchema7),\n    };\n  }\n\n  return Object.keys(toolSet).length > 0 ? (toolSet as ToolSet) : undefined;\n}\n\nfunction normalizeNativeMessages(messages: unknown): ModelMessage[] | undefined {\n  if (!Array.isArray(messages)) {\n    return undefined;\n  }\n\n  return messages.map((message) => normalizeNativeMessage(message));\n}\n\nfunction normalizeNativeMessage(message: unknown): ModelMessage {\n  const raw = asRecord(message);\n  const providerOptions = asOptionalRecord(raw.providerOptions);\n\n  if (raw.role === \"system\") {\n    return {\n      role: \"system\",\n      content: stringifyMessageContent(raw.content),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  if (raw.role === \"assistant\") {\n    return {\n      role: \"assistant\",\n      content: normalizeAssistantContent(raw),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  if (raw.role === \"tool\") {\n    return {\n      role: \"tool\",\n      content: normalizeToolContent(raw),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  return {\n    role: \"user\",\n    content: normalizeUserContent(raw.content),\n    ...(providerOptions ? { providerOptions } : {}),\n  } as ModelMessage;\n}\n\n/**\n * Strip reasoning-only parts from outbound assistant content.\n *\n * OpenAI-spec reasoning models (Cerebras gpt-oss-120b, OpenAI o1/o3,\n * DeepSeek R1, Qwen-3-thinking, etc.) return reasoning in the assistant\n * response — either as a separate `reasoning` / `reasoning_content`\n * field, or as content parts with `type: \"reasoning\"`. Echoing those\n * back to the next turn is wrong on both ends:\n *   - Cerebras returns HTTP 400 (`messages.X.assistant.reasoning_content:\n *     property is unsupported`).\n *   - OpenAI silently drops them, which wastes prompt tokens.\n *\n * The AI SDK upstream of this normalizer surfaces those reasoning blocks\n * as `{ type: \"reasoning\", ... }` content parts. We drop them here so\n * the wire stays spec-clean for the next turn. The reasoning itself\n * remains usable as a single-turn signal (still on the response object);\n * we only refuse to round-trip it.\n */\nfunction stripReasoningParts(content: unknown[]): unknown[] {\n  return content.filter((part) => {\n    if (!part || typeof part !== \"object\") return true;\n    const type = (part as { type?: unknown }).type;\n    return type !== \"reasoning\" && type !== \"thinking\";\n  });\n}\n\nfunction normalizeAssistantContent(message: Record<string, unknown>): unknown {\n  const toolCalls = Array.isArray(message.toolCalls) ? message.toolCalls : [];\n\n  if (toolCalls.length === 0) {\n    if (Array.isArray(message.content)) {\n      return stripReasoningParts(message.content);\n    }\n    if (typeof message.content === \"string\") {\n      return message.content;\n    }\n    return \"\";\n  }\n\n  const parts: unknown[] = [];\n  if (typeof message.content === \"string\" && message.content.length > 0) {\n    parts.push({ type: \"text\", text: message.content });\n  } else if (Array.isArray(message.content)) {\n    parts.push(...stripReasoningParts(message.content));\n  }\n\n  for (const toolCall of toolCalls) {\n    const rawCall = asRecord(toolCall);\n    const rawFunction = asRecord(rawCall.function);\n    const toolCallId = firstString(rawCall.toolCallId, rawCall.id);\n    const toolName = firstString(rawCall.toolName, rawCall.name, rawFunction.name);\n\n    if (!toolCallId || !toolName) {\n      continue;\n    }\n\n    parts.push({\n      type: \"tool-call\",\n      toolCallId,\n      toolName,\n      input: parseToolCallInput(rawCall, rawFunction),\n    });\n  }\n\n  return parts;\n}\n\nfunction normalizeToolContent(message: Record<string, unknown>): unknown[] {\n  if (Array.isArray(message.content)) {\n    return message.content;\n  }\n\n  const toolCallId = firstString(message.toolCallId, message.id) ?? \"tool-call\";\n  const toolName = firstString(message.toolName, message.name) ?? \"tool\";\n  const parsed = parseJsonIfPossible(message.content);\n\n  return [\n    {\n      type: \"tool-result\",\n      toolCallId,\n      toolName,\n      output:\n        typeof parsed === \"string\"\n          ? { type: \"text\", value: parsed }\n          : { type: \"json\", value: parsed },\n    },\n  ];\n}\n\nfunction normalizeUserContent(content: unknown): UserContent {\n  if (Array.isArray(content)) {\n    return content as UserContent;\n  }\n  return stringifyMessageContent(content);\n}\n\nfunction stringifyMessageContent(content: unknown): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n  if (content == null) {\n    return \"\";\n  }\n  return typeof content === \"object\" ? JSON.stringify(content) : String(content);\n}\n\nfunction parseToolCallInput(\n  rawCall: Record<string, unknown>,\n  rawFunction: Record<string, unknown>\n): unknown {\n  if (\"input\" in rawCall) {\n    return rawCall.input;\n  }\n  return parseJsonIfPossible(rawCall.arguments ?? rawFunction.arguments ?? {});\n}\n\nfunction parseJsonIfPossible(value: unknown): unknown {\n  if (typeof value !== \"string\") {\n    return value ?? \"\";\n  }\n  try {\n    return JSON.parse(value);\n  } catch {\n    return value;\n  }\n}\n\nfunction normalizeToolChoice(toolChoice: unknown): ToolChoice<ToolSet> | undefined {\n  if (!toolChoice) {\n    return undefined;\n  }\n\n  if (\n    typeof toolChoice === \"string\" &&\n    (toolChoice === \"auto\" || toolChoice === \"none\" || toolChoice === \"required\")\n  ) {\n    return toolChoice;\n  }\n\n  const choice = asRecord(toolChoice);\n  if (choice.type === \"tool\") {\n    if (typeof choice.toolName === \"string\" && choice.toolName.length > 0) {\n      return toolChoice as ToolChoice<ToolSet>;\n    }\n    const toolName = firstString(choice.toolName, choice.name);\n    if (toolName) {\n      return { type: \"tool\", toolName };\n    }\n  }\n\n  if (choice.type === \"function\") {\n    const fn = asRecord(choice.function);\n    const toolName = firstString(fn.name);\n    if (toolName) {\n      return { type: \"tool\", toolName };\n    }\n  }\n\n  const namedTool = firstString(choice.name);\n  if (namedTool) {\n    return { type: \"tool\", toolName: namedTool };\n  }\n\n  return toolChoice as ToolChoice<ToolSet>;\n}\n\nfunction hasIllegalStrictRoot(node: Record<string, unknown>): boolean {\n  // Strict-mode JSON schema validators on OpenAI-compatible providers (Groq,\n  // Cerebras, OpenAI strict tools) reject tool-parameters whose top level is\n  // not `type: \"object\"` or carries `oneOf`/`anyOf`/`enum`/`not` at the root.\n  // The error wording varies by provider but the constraint is uniform.\n  if (node.type !== \"object\") return true;\n  if (Array.isArray(node.oneOf) && node.oneOf.length > 0) return true;\n  if (Array.isArray(node.anyOf) && node.anyOf.length > 0) return true;\n  if (Array.isArray(node.enum)) return true;\n  if (node.not !== undefined) return true;\n  return false;\n}\n\nfunction sanitizeJsonSchema(schema: unknown, isRoot = false): JSONSchema7 {\n  if (!schema || typeof schema !== \"object\" || Array.isArray(schema)) {\n    // Permissive fallback: no `properties: {}`/`additionalProperties: false`\n    // pair, which strict-grammar providers reject. See `normalizeSchemaForCerebras`\n    // in @elizaos/core for the rationale.\n    return { type: \"object\" };\n  }\n\n  const record = schema as Record<string, unknown>;\n  let sanitized: Record<string, unknown> = { ...record };\n\n  if (typeof sanitized.type !== \"string\") {\n    const inferredType = inferJsonSchemaType(sanitized, isRoot);\n    if (inferredType) {\n      sanitized.type = inferredType;\n    }\n  }\n\n  if (isRoot && hasIllegalStrictRoot(sanitized)) {\n    // Wrap the original schema under properties.value. Strict-tool callers\n    // that unwrap arguments will see `{ value: <original> }`. The recursion\n    // below normalises the wrapped child like any other property.\n    sanitized = {\n      type: \"object\",\n      properties: { value: { ...record } },\n      required: [\"value\"],\n      additionalProperties: false,\n    };\n  }\n\n  if (\n    sanitized.properties &&\n    typeof sanitized.properties === \"object\" &&\n    !Array.isArray(sanitized.properties)\n  ) {\n    const properties: Record<string, unknown> = {};\n    for (const [key, value] of Object.entries(sanitized.properties as Record<string, unknown>)) {\n      properties[key] = sanitizeJsonSchema(value);\n    }\n    sanitized.properties = properties;\n\n    const propertyKeys = Object.keys(properties);\n    const existingRequired = Array.isArray(sanitized.required)\n      ? sanitized.required.filter((key): key is string => typeof key === \"string\")\n      : [];\n    sanitized.required = [...new Set([...existingRequired, ...propertyKeys])];\n  }\n\n  if (sanitized.type === \"object\" && sanitized.additionalProperties !== false) {\n    sanitized.additionalProperties = false;\n  }\n\n  if (sanitized.items) {\n    sanitized.items = Array.isArray(sanitized.items)\n      ? sanitized.items.map((item) => sanitizeJsonSchema(item))\n      : sanitizeJsonSchema(sanitized.items);\n  }\n\n  for (const unionKey of [\"anyOf\", \"oneOf\", \"allOf\"] as const) {\n    const value = sanitized[unionKey];\n    if (Array.isArray(value)) {\n      sanitized[unionKey] = value.map((item) => sanitizeJsonSchema(item));\n    }\n  }\n\n  return sanitized as JSONSchema7;\n}\n\nfunction inferJsonSchemaType(schema: Record<string, unknown>, isRoot: boolean): string | undefined {\n  if (\n    \"properties\" in schema ||\n    \"required\" in schema ||\n    \"additionalProperties\" in schema ||\n    isRoot\n  ) {\n    return \"object\";\n  }\n  if (\"items\" in schema) {\n    return \"array\";\n  }\n  if (Array.isArray(schema.enum) && schema.enum.length > 0) {\n    const types = new Set(schema.enum.map((value) => typeof value));\n    if (types.size === 1) {\n      const [type] = [...types];\n      if (type === \"string\" || type === \"number\" || type === \"boolean\") {\n        return type;\n      }\n    }\n  }\n  return undefined;\n}\n\nfunction asRecord(value: unknown): Record<string, unknown> {\n  return value && typeof value === \"object\" && !Array.isArray(value)\n    ? (value as Record<string, unknown>)\n    : {};\n}\n\nfunction asOptionalRecord(value: unknown): Record<string, unknown> | undefined {\n  return value && typeof value === \"object\" && !Array.isArray(value)\n    ? (value as Record<string, unknown>)\n    : undefined;\n}\n\nfunction firstString(...values: unknown[]): string | undefined {\n  for (const value of values) {\n    if (typeof value === \"string\" && value.length > 0) {\n      return value;\n    }\n  }\n  return undefined;\n}\n\nfunction usesNativeTextResult(params: GenerateTextParamsWithOpenAIOptions): boolean {\n  return Boolean(params.messages || params.tools || params.toolChoice || params.responseSchema);\n}\n\nfunction buildNativeTextResult(\n  result: {\n    text: string;\n    toolCalls?: unknown[];\n    finishReason?: string;\n    usage?: LanguageModelUsage;\n    providerMetadata?: unknown;\n  },\n  modelName?: string\n): NativeGenerateTextResult {\n  return {\n    text: result.text,\n    toolCalls: result.toolCalls ?? [],\n    finishReason: result.finishReason,\n    usage: convertUsage(result.usage),\n    providerMetadata: mergeProviderModelName(result.providerMetadata, modelName),\n  };\n}\n\nfunction handledPromise<T>(value: T | PromiseLike<T>): Promise<T> {\n  const promise = Promise.resolve(value);\n  promise.catch(() => {\n    // The streaming path primarily consumes `textStream`. AI SDK companion\n    // promises such as `text` can reject later on empty streams even when no\n    // caller requested them, which otherwise surfaces as an unhandled rejection.\n  });\n  return promise;\n}\n\nfunction handledMappedPromise<T, U>(\n  value: T | PromiseLike<T>,\n  mapper: (resolved: T) => U | PromiseLike<U>\n): Promise<U> {\n  return handledPromise(handledPromise(value).then(mapper));\n}\n\nfunction mergeProviderModelName(providerMetadata: unknown, modelName?: string): unknown {\n  if (!modelName) {\n    return providerMetadata;\n  }\n  if (\n    providerMetadata &&\n    typeof providerMetadata === \"object\" &&\n    !Array.isArray(providerMetadata)\n  ) {\n    return {\n      ...(providerMetadata as Record<string, unknown>),\n      modelName,\n    };\n  }\n  return { modelName };\n}\n\nfunction createLlmCallDetails(\n  modelName: string,\n  params: GenerateTextParams,\n  systemPrompt: string | undefined,\n  actionType: string,\n  modelType?: ModelTypeName,\n  providerOptions?: Record<string, unknown>,\n  generateParams?: NativeTextParams\n): RecordLlmCallDetails {\n  const originalParams = params as GenerateTextParamsWithOpenAIOptions;\n  const nativeParams = generateParams as\n    | (NativeTextParams & {\n        output?: unknown;\n        maxOutputTokens?: unknown;\n      })\n    | undefined;\n  const nativePrompt = nativeParams && \"prompt\" in nativeParams ? nativeParams.prompt : undefined;\n  const nativeMessages =\n    nativeParams && \"messages\" in nativeParams && Array.isArray(nativeParams.messages)\n      ? nativeParams.messages\n      : undefined;\n  const nativeSystem =\n    typeof nativeParams?.system === \"string\" ? nativeParams.system : systemPrompt;\n  return {\n    model: modelName,\n    modelType,\n    provider: \"vercel-ai-sdk\",\n    systemPrompt: nativeSystem ?? \"\",\n    userPrompt:\n      typeof nativePrompt === \"string\"\n        ? nativePrompt\n        : typeof params.prompt === \"string\"\n          ? params.prompt\n          : \"\",\n    prompt: typeof nativePrompt === \"string\" ? nativePrompt : undefined,\n    messages: nativeMessages,\n    tools: nativeParams?.tools ?? originalParams.tools,\n    toolChoice: nativeParams?.toolChoice ?? originalParams.toolChoice,\n    output:\n      nativeParams?.output !== undefined\n        ? buildTrajectoryOutputDescriptor(originalParams.responseSchema, nativeParams.output)\n        : undefined,\n    responseSchema: originalParams.responseSchema,\n    providerOptions:\n      providerOptions ?? nativeParams?.providerOptions ?? originalParams.providerOptions,\n    temperature: params.temperature ?? 0,\n    maxTokens:\n      typeof nativeParams?.maxOutputTokens === \"number\"\n        ? nativeParams.maxOutputTokens\n        : params.omitMaxTokens\n          ? 0\n          : (params.maxTokens ?? 8192),\n    maxTokensOmitted:\n      params.omitMaxTokens && typeof nativeParams?.maxOutputTokens !== \"number\" ? true : undefined,\n    purpose: \"external_llm\",\n    actionType,\n  };\n}\n\nfunction buildTrajectoryOutputDescriptor(responseSchema: unknown, output: unknown): unknown {\n  if (responseSchema !== undefined) {\n    return {\n      type: \"object\",\n      schema: responseSchema,\n    };\n  }\n  return toTrajectoryJsonSafe(output);\n}\n\nfunction toTrajectoryJsonSafe(value: unknown): unknown {\n  try {\n    return JSON.parse(\n      JSON.stringify(value, (_key, nested) => {\n        if (typeof nested === \"function\") return undefined;\n        if (typeof nested === \"bigint\") return nested.toString();\n        return nested;\n      })\n    ) as unknown;\n  } catch {\n    return String(value);\n  }\n}\n\nfunction applyUsageToDetails(\n  details: RecordLlmCallDetails,\n  usage: LanguageModelUsage | undefined\n): void {\n  if (!usage) {\n    return;\n  }\n  details.promptTokens = usage.inputTokens ?? 0;\n  details.completionTokens = usage.outputTokens ?? 0;\n}\n\n// ============================================================================\n// Core Generation Function\n// ============================================================================\n\n/**\n * Generates text using the specified model type.\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @param modelType - The type of model (TEXT_SMALL or TEXT_LARGE)\n * @param getModelFn - Function to get the model name\n * @returns Generated text or stream result\n */\nasync function generateTextByModelType(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams,\n  modelType: ModelTypeName,\n  getModelFn: ModelNameGetter\n): Promise<string | TextStreamResult> {\n  const paramsWithAttachments = params as GenerateTextParamsWithOpenAIOptions;\n  const openai = createOpenAIClient(runtime);\n  const modelName = resolveRequestedModelName(paramsWithAttachments, runtime, getModelFn);\n\n  logger.debug(`[OpenAI] Using ${modelType} model: ${modelName}`);\n  const providerOptions = resolveProviderOptions(params, runtime, modelName);\n  const hasAttachments = (paramsWithAttachments.attachments?.length ?? 0) > 0;\n  const userContent = hasAttachments ? buildUserContent(paramsWithAttachments) : undefined;\n  const shouldReturnNativeResult = usesNativeTextResult(paramsWithAttachments);\n\n  const systemPrompt = resolveEffectiveSystemPrompt({\n    params: paramsWithAttachments,\n    fallback: buildCanonicalSystemPrompt({ character: runtime.character }),\n  });\n  const agentName = paramsWithAttachments.providerOptions?.agentName;\n  const telemetryConfig: NativeTelemetrySettings = {\n    isEnabled: getExperimentalTelemetry(runtime),\n    functionId: agentName ? `agent:${agentName}` : undefined,\n    metadata: agentName ? { agentName } : undefined,\n  };\n\n  // Chat Completions is the default: broadest compatibility, and it works\n  // against every OpenAI-compatible endpoint (Cerebras, local servers, proxies).\n  // gpt-5 / gpt-5-mini reasoning models ignore temperature/penalty/stop params.\n  //\n  const model = openai.chat(modelName);\n  const cerebrasMode = isCerebrasMode(runtime);\n  const normalizedTools = normalizeNativeTools(paramsWithAttachments.tools, {\n    cerebrasMode,\n  });\n  const normalizedToolChoice = normalizeToolChoice(paramsWithAttachments.toolChoice);\n  const normalizedMessages = normalizeNativeMessages(paramsWithAttachments.messages);\n  const wireMessages = dropDuplicateLeadingSystemMessage(normalizedMessages, systemPrompt);\n  const effectiveMessages =\n    wireMessages && wireMessages.length > 0 ? wireMessages : normalizedMessages;\n  const promptText =\n    typeof params.prompt === \"string\" && params.prompt.length > 0 ? params.prompt : \"\";\n  const promptOrMessages: NativePrompt =\n    effectiveMessages && effectiveMessages.length > 0\n      ? { messages: effectiveMessages }\n      : userContent\n        ? { messages: [{ role: \"user\" as const, content: userContent }] }\n        : { prompt: promptText };\n  // elizaOS callers pass `responseFormat: { type: \"json_object\" | \"text\" }`\n  // (see `GenerateTextParams` in @elizaos/core). The AI SDK's equivalent\n  // is `responseFormat: { type: \"json\" }` (which translates to\n  // `response_format: { type: \"json_object\" }` at the OpenAI wire layer).\n  // Translate the shape so the param actually reaches the API call —\n  // before this, callers asking for json_object were silently ignored\n  // and Cerebras returned plain text, dropping us into the simple-reply\n  // fallback every turn.\n  const callerResponseFormat = (paramsWithAttachments as { responseFormat?: unknown })\n    .responseFormat;\n  const responseFormatType =\n    typeof callerResponseFormat === \"string\"\n      ? callerResponseFormat\n      : callerResponseFormat &&\n          typeof callerResponseFormat === \"object\" &&\n          \"type\" in callerResponseFormat\n        ? (callerResponseFormat as { type: string }).type\n        : undefined;\n  const wireResponseFormat: { type: \"json\" } | { type: \"text\" } | undefined =\n    responseFormatType === \"json_object\"\n      ? { type: \"json\" }\n      : responseFormatType === \"text\"\n        ? { type: \"text\" }\n        : undefined;\n\n  const generateParams: NativeTextParams = {\n    model,\n    ...promptOrMessages,\n    system: systemPrompt,\n    allowSystemInMessages: true,\n    // Omit the cap when the caller opted out (direct-channel Stage-1) so the\n    // model's own max applies — a hardcoded value 400s when it exceeds the\n    // model's limit. Other callers keep the 8192 default.\n    ...(params.omitMaxTokens ? {} : { maxOutputTokens: params.maxTokens ?? 8192 }),\n    experimental_telemetry: telemetryConfig,\n    ...(normalizedTools ? { tools: normalizedTools } : {}),\n    ...(normalizedToolChoice ? { toolChoice: normalizedToolChoice } : {}),\n    // Cerebras's OpenAI-compatible endpoint does not accept the\n    // `response_format: { type: \"json_schema\", ... }` payload that the AI SDK\n    // emits when `output: Output.object(...)` is set. Fall back to relying on\n    // `responseFormat: { type: \"json_object\" }` (already passed by callers)\n    // plus the schema embedded in the prompt body.\n    ...(paramsWithAttachments.responseSchema && !isCerebrasMode(runtime)\n      ? { output: buildStructuredOutput(paramsWithAttachments.responseSchema) }\n      : {}),\n    ...(wireResponseFormat ? { responseFormat: wireResponseFormat } : {}),\n    ...(providerOptions ? { providerOptions: providerOptions as NativeProviderOptions } : {}),\n  };\n\n  // Handle streaming mode\n  if (params.stream) {\n    const details = createLlmCallDetails(\n      modelName,\n      params,\n      systemPrompt,\n      \"ai.streamText\",\n      modelType,\n      providerOptions,\n      generateParams\n    );\n    details.response = \"\";\n    const result = await recordLlmCall(runtime, details, () => streamText(generateParams));\n\n    return {\n      textStream: (async function* textStreamWithCallback() {\n        for await (const chunk of result.textStream) {\n          params.onStreamChunk?.(chunk);\n          yield chunk;\n        }\n      })(),\n      text: handledPromise(result.text),\n      ...(shouldReturnNativeResult ? { toolCalls: handledPromise(result.toolCalls) } : {}),\n      usage: handledMappedPromise(result.usage, convertUsage),\n      finishReason: handledMappedPromise(result.finishReason, (r) => r as string | undefined),\n    };\n  }\n\n  // Non-streaming mode\n  const details = createLlmCallDetails(\n    modelName,\n    params,\n    systemPrompt,\n    \"ai.generateText\",\n    modelType,\n    providerOptions,\n    generateParams\n  );\n  const result = await recordLlmCall(runtime, details, async () => {\n    const result = await generateText(generateParams);\n    details.response = result.text;\n    details.toolCalls = result.toolCalls;\n    details.finishReason = result.finishReason as string | undefined;\n    details.providerMetadata = result.providerMetadata;\n    applyUsageToDetails(details, result.usage);\n    return result;\n  });\n\n  if (result.usage) {\n    emitModelUsageEvent(runtime, modelType, params.prompt ?? \"\", result.usage);\n  }\n\n  if (shouldReturnNativeResult) {\n    return buildNativeTextResult(result, modelName) as NativeTextModelResult;\n  }\n\n  return result.text;\n}\n\n// ============================================================================\n// Public Handlers\n// ============================================================================\n\n/**\n * Handles TEXT_SMALL model requests.\n *\n * Uses the configured small model (default: gpt-5-mini).\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @returns Generated text or stream result\n */\nexport async function handleTextSmall(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ModelType.TEXT_SMALL, getSmallModel);\n}\n\nexport async function handleTextNano(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_NANO_MODEL_TYPE, getNanoModel);\n}\n\nexport async function handleTextMedium(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_MEDIUM_MODEL_TYPE, getMediumModel);\n}\n\n/**\n * Handles TEXT_LARGE model requests.\n *\n * Uses the configured large model (default: gpt-5).\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @returns Generated text or stream result\n */\nexport async function handleTextLarge(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ModelType.TEXT_LARGE, getLargeModel);\n}\n\nexport async function handleTextMega(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_MEGA_MODEL_TYPE, getMegaModel);\n}\n\nexport async function handleResponseHandler(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(\n    runtime,\n    params,\n    RESPONSE_HANDLER_MODEL_TYPE,\n    getResponseHandlerModel\n  );\n}\n\nexport async function handleActionPlanner(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ACTION_PLANNER_MODEL_TYPE, getActionPlannerModel);\n}\n\n// ─── Test-only exports ──────────────────────────────────────────────────────\n// These are exported for the shape tests in `__tests__/reasoning-effort.shape.test.ts`.\n// Not part of the public API; do not import outside tests.\n\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_resolveProviderOptions = resolveProviderOptions;\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_normalizeNativeMessages = normalizeNativeMessages;\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_stripReasoningParts = stripReasoningParts;\n",
+    "/**\n * Text generation model handlers\n *\n * Provides text generation using OpenAI's language models.\n */\n\nimport type {\n  GenerateTextParams,\n  IAgentRuntime,\n  JsonValue,\n  ModelTypeName,\n  RecordLlmCallDetails,\n} from \"@elizaos/core\";\nimport {\n  buildCanonicalSystemPrompt,\n  dropDuplicateLeadingSystemMessage,\n  logger,\n  ModelType,\n  normalizeSchemaForCerebras,\n  recordLlmCall,\n  resolveEffectiveSystemPrompt,\n  sanitizeFunctionNameForCerebras,\n} from \"@elizaos/core\";\nimport {\n  generateText,\n  type JSONSchema7,\n  jsonSchema,\n  type LanguageModelUsage,\n  type ModelMessage,\n  Output,\n  streamText,\n  type ToolChoice,\n  type ToolSet,\n  type UserContent,\n} from \"ai\";\nimport { createOpenAIClient } from \"../providers\";\nimport type { TextStreamResult, TokenUsage } from \"../types\";\nimport {\n  getActionPlannerModel,\n  getExperimentalTelemetry,\n  getLargeModel,\n  getMediumModel,\n  getMegaModel,\n  getNanoModel,\n  getResponseHandlerModel,\n  getSmallModel,\n  isCerebrasMode,\n} from \"../utils/config\";\nimport { emitModelUsageEvent } from \"../utils/events\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\n/**\n * Function to get model name from runtime\n */\ntype ModelNameGetter = (runtime: IAgentRuntime) => string;\n\ntype PromptCacheRetention = \"in_memory\" | \"24h\";\ntype ChatAttachment = {\n  data: string | Uint8Array | URL;\n  mediaType: string;\n  filename?: string;\n};\n\ninterface OpenAIPromptCacheOptions {\n  promptCacheKey?: string;\n  promptCacheRetention?: PromptCacheRetention;\n}\n\ninterface GenerateTextParamsWithOpenAIOptions\n  extends Omit<\n    GenerateTextParams,\n    \"messages\" | \"tools\" | \"toolChoice\" | \"responseSchema\" | \"providerOptions\"\n  > {\n  model?: string;\n  attachments?: ChatAttachment[];\n  messages?: unknown[];\n  tools?: unknown;\n  toolChoice?: unknown;\n  responseSchema?: unknown;\n  providerOptions?: Record<string, object | JsonValue> & {\n    agentName?: string;\n    openai?: OpenAIPromptCacheOptions;\n  };\n}\n\ntype NativeOutput = NonNullable<Parameters<typeof generateText<ToolSet>>[0][\"output\"]>;\ntype NativeGenerateTextParams = Parameters<typeof generateText<ToolSet, NativeOutput>>[0];\ntype NativeStreamTextParams = Parameters<typeof streamText<ToolSet, NativeOutput>>[0];\ntype NativePrompt =\n  | { prompt: string; messages?: never }\n  | { messages: ModelMessage[]; prompt?: never };\ntype NativeTextParams = Omit<NativeGenerateTextParams, \"messages\" | \"prompt\"> &\n  Omit<NativeStreamTextParams, \"messages\" | \"prompt\"> &\n  NativePrompt & {\n    // Re-declared explicitly: TypeScript's `Parameters<typeof generateText>`\n    // inference produces an overload-union that drops this field, but the\n    // ai SDK's runtime signature accepts it (see ai@6 `CallSettings & Prompt`).\n    allowSystemInMessages?: boolean;\n  };\ntype NativeProviderOptions = NativeTextParams[\"providerOptions\"];\ntype NativeTelemetrySettings = NativeTextParams[\"experimental_telemetry\"];\n\ntype LanguageModelUsageWithCache = Omit<LanguageModelUsage, \"inputTokenDetails\"> & {\n  inputTokenDetails?: LanguageModelUsage[\"inputTokenDetails\"] & {\n    cachedInputTokens?: number;\n    cacheCreationInputTokens?: number;\n    cacheCreationTokens?: number;\n  };\n  cachedInputTokens?: number;\n  cacheReadInputTokens?: number;\n  cacheCreationInputTokens?: number;\n  cacheWriteInputTokens?: number;\n  input_tokens_details?: {\n    cached_tokens?: number;\n    cache_read_input_tokens?: number;\n    cache_creation_input_tokens?: number;\n  };\n  prompt_tokens_details?: {\n    cached_tokens?: number;\n  };\n};\n\ninterface NativeGenerateTextResult {\n  text: string;\n  toolCalls?: unknown[];\n  finishReason?: string;\n  usage?: TokenUsage;\n  providerMetadata?: unknown;\n}\n\ntype NativeTextModelResult = string & NativeGenerateTextResult;\n\nconst TEXT_NANO_MODEL_TYPE = ModelType.TEXT_NANO as ModelTypeName;\nconst TEXT_MEDIUM_MODEL_TYPE = ModelType.TEXT_MEDIUM as ModelTypeName;\nconst TEXT_MEGA_MODEL_TYPE = ModelType.TEXT_MEGA as ModelTypeName;\nconst RESPONSE_HANDLER_MODEL_TYPE = ModelType.RESPONSE_HANDLER as ModelTypeName;\nconst ACTION_PLANNER_MODEL_TYPE = ModelType.ACTION_PLANNER as ModelTypeName;\n\nfunction resolveRequestedModelName(\n  params: GenerateTextParamsWithOpenAIOptions,\n  runtime: IAgentRuntime,\n  getModelFn: ModelNameGetter\n): string {\n  return typeof params.model === \"string\" && params.model.trim().length > 0\n    ? params.model.trim()\n    : getModelFn(runtime);\n}\n\nfunction buildUserContent(params: GenerateTextParamsWithOpenAIOptions): UserContent {\n  const content: Array<\n    | { type: \"text\"; text: string }\n    | {\n        type: \"file\";\n        data: string | Uint8Array | URL;\n        mediaType: string;\n        filename?: string;\n      }\n  > = [{ type: \"text\", text: params.prompt ?? \"\" }];\n\n  for (const attachment of params.attachments ?? []) {\n    content.push({\n      type: \"file\",\n      data: attachment.data,\n      mediaType: attachment.mediaType,\n      ...(attachment.filename ? { filename: attachment.filename } : {}),\n    });\n  }\n\n  return content;\n}\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\n/**\n * Converts AI SDK usage to our token usage format.\n *\n * Emits both the legacy `cachedPromptTokens` (kept for back-compat with\n * existing OpenAI consumers) and the canonical v5 `cacheReadInputTokens`\n * (consumed by the trajectory recorder + cost table). They always carry the\n * same value when the AI SDK reports cached input.\n */\nfunction convertUsage(usage: LanguageModelUsage | undefined): TokenUsage | undefined {\n  if (!usage) {\n    return undefined;\n  }\n\n  // The AI SDK uses inputTokens/outputTokens\n  const promptTokens = usage.inputTokens ?? 0;\n  const completionTokens = usage.outputTokens ?? 0;\n  const usageWithCache: LanguageModelUsageWithCache = usage;\n  const cachedInput =\n    firstNumber(\n      usageWithCache.cacheReadInputTokens,\n      usageWithCache.cachedInputTokens,\n      usageWithCache.inputTokenDetails?.cacheReadTokens,\n      usageWithCache.inputTokenDetails?.cachedInputTokens,\n      usageWithCache.input_tokens_details?.cache_read_input_tokens,\n      usageWithCache.input_tokens_details?.cached_tokens,\n      usageWithCache.prompt_tokens_details?.cached_tokens\n    ) ?? undefined;\n  const cacheCreationInput = firstNumber(\n    usageWithCache.cacheCreationInputTokens,\n    usageWithCache.cacheWriteInputTokens,\n    usageWithCache.inputTokenDetails?.cacheCreationInputTokens,\n    usageWithCache.inputTokenDetails?.cacheCreationTokens,\n    usageWithCache.inputTokenDetails?.cacheWriteTokens,\n    usageWithCache.input_tokens_details?.cache_creation_input_tokens\n  );\n\n  return {\n    promptTokens,\n    completionTokens,\n    totalTokens: promptTokens + completionTokens,\n    cachedPromptTokens: cachedInput,\n    cacheReadInputTokens: cachedInput,\n    cacheCreationInputTokens: cacheCreationInput,\n  };\n}\n\nfunction firstNumber(...values: unknown[]): number | undefined {\n  for (const value of values) {\n    if (typeof value === \"number\" && Number.isFinite(value)) {\n      return value;\n    }\n    if (typeof value === \"string\" && value.trim().length > 0) {\n      const parsed = Number(value);\n      if (Number.isFinite(parsed)) {\n        return parsed;\n      }\n    }\n  }\n  return undefined;\n}\n\nfunction resolvePromptCacheOptions(params: GenerateTextParams): OpenAIPromptCacheOptions {\n  const withOpenAIOptions = params as GenerateTextParamsWithOpenAIOptions;\n  return {\n    promptCacheKey: withOpenAIOptions.providerOptions?.openai?.promptCacheKey,\n    promptCacheRetention: withOpenAIOptions.providerOptions?.openai?.promptCacheRetention,\n  };\n}\n\n/**\n * Forward `OPENAI_REASONING_EFFORT` (runtime setting / process.env) as\n * `reasoning_effort` on the outbound chat completions request. This is\n * the OpenAI-spec knob for reasoning-capable models (`o1-*`, `o3-*`,\n * `gpt-oss-*`, `deepseek-r1`, and similar families) — including\n * Cerebras and OpenRouter, which honor the same field. `\"low\"` keeps\n * reasoning short enough that visible content always fits inside\n * `max_tokens`, which is the failure mode on Cerebras gpt-oss-120b when\n * left unset.\n *\n * In Cerebras mode the field defaults to `\"low\"` when unset, but ONLY for\n * reasoning-capable models (e.g. gpt-oss-* and deepseek-r1):\n * gpt-oss-120b emits a separate reasoning channel and, left unbounded, spends\n * the whole token budget reasoning — returning empty visible content, which\n * makes the agent fall back to \"I don't have a reply for that\". `\"low\"` keeps\n * reasoning short so a reply always materializes. Non-reasoning Cerebras models\n * (Llama, etc.) reject `reasoning_effort`, so they must never receive the\n * default. For all other models an unset/invalid value yields `undefined`, so\n * they pay no overhead and the wire stays clean. An explicit valid\n * `OPENAI_REASONING_EFFORT` always wins.\n *\n * Valid values follow the OpenAI spec exactly: `minimal`, `low`,\n * `medium`, `high`. Anything else is logged and ignored.\n */\ntype ReasoningEffort = \"minimal\" | \"low\" | \"medium\" | \"high\";\n\nconst VALID_REASONING_EFFORTS: readonly ReasoningEffort[] = [\"minimal\", \"low\", \"medium\", \"high\"];\n\n/**\n * Reasoning-capable model families that emit a separate reasoning channel and\n * honor `reasoning_effort`. Used to gate the Cerebras `\"low\"` default so\n * non-reasoning models (Llama, etc.) are never sent the field.\n */\nfunction isReasoningModel(modelName: string | undefined): boolean {\n  if (!modelName) return false;\n  const m = modelName.toLowerCase();\n  return (\n    m.includes(\"gpt-oss\") ||\n    m.includes(\"o1\") ||\n    m.includes(\"o3\") ||\n    m.includes(\"o4\") ||\n    m.includes(\"deepseek-r1\") ||\n    m.includes(\"thinking\") ||\n    m.includes(\"reasoning\") ||\n    m.includes(\"qwq\")\n  );\n}\n\nfunction resolveReasoningEffort(\n  runtime: IAgentRuntime,\n  modelName?: string\n): ReasoningEffort | undefined {\n  const raw = runtime.getSetting(\"OPENAI_REASONING_EFFORT\");\n  const normalized = typeof raw === \"string\" ? raw.trim().toLowerCase() : \"\";\n  if (normalized) {\n    if ((VALID_REASONING_EFFORTS as readonly string[]).includes(normalized)) {\n      return normalized as ReasoningEffort;\n    }\n    logger.warn(\n      `[OpenAI] OPENAI_REASONING_EFFORT=${raw} is not a valid reasoning effort; ignoring. Expected one of: ${VALID_REASONING_EFFORTS.join(\", \")}.`\n    );\n  }\n  // gpt-oss-120b on Cerebras returns empty content when reasoning runs\n  // unbounded; default to \"low\" so a visible reply always fits — but only for\n  // reasoning-capable models. Non-reasoning Cerebras models (Llama, etc.)\n  // reject `reasoning_effort` and would break. An explicit valid value above\n  // wins over this default.\n  if (isCerebrasMode(runtime) && isReasoningModel(modelName)) {\n    return \"low\";\n  }\n  return undefined;\n}\n\nfunction resolveProviderOptions(\n  params: GenerateTextParams,\n  runtime: IAgentRuntime,\n  modelName?: string\n): Record<string, unknown> | undefined {\n  const withOpenAIOptions = params as GenerateTextParamsWithOpenAIOptions;\n  const rawProviderOptions = withOpenAIOptions.providerOptions;\n  const promptCacheOptions = resolvePromptCacheOptions(params);\n  const reasoningEffort = resolveReasoningEffort(runtime, modelName);\n\n  if (\n    !rawProviderOptions &&\n    !promptCacheOptions.promptCacheKey &&\n    !promptCacheOptions.promptCacheRetention &&\n    !reasoningEffort\n  ) {\n    return undefined;\n  }\n\n  // Cerebras supports prompt caching on gpt-oss-120b — 128-token blocks,\n  // default-on. The `prompt_cache_key` field IS accepted by Cerebras's\n  // OpenAI-compatible endpoint and surfaces hit counts via\n  // `usage.prompt_tokens_details.cached_tokens` (same shape as OpenAI), so\n  // we keep it in the request body. Only `prompt_cache_retention` is an\n  // OpenAI-direct-only field that Cerebras rejects with HTTP 400\n  // (`wrong_api_format`), so we strip just that one when in Cerebras mode.\n  const skipCacheRetention = isCerebrasMode(runtime);\n\n  const { agentName: _agentName, openai: rawOpenAIOptions, ...rest } = rawProviderOptions ?? {};\n  // When on Cerebras, scrub OpenAI-direct-only fields (e.g. `promptCacheRetention`)\n  // from `rawOpenAIOptions` before they're spread; otherwise they reach the wire\n  // and the Cerebras endpoint rejects with HTTP 400 `wrong_api_format`.\n  const sanitizedRawOpenAIOptions = (() => {\n    if (!rawOpenAIOptions || typeof rawOpenAIOptions !== \"object\") return rawOpenAIOptions;\n    if (!skipCacheRetention) return rawOpenAIOptions;\n    const { promptCacheRetention: _drop, ...rest2 } = rawOpenAIOptions as Record<string, unknown>;\n    return rest2;\n  })();\n  const openaiOptions = {\n    ...(sanitizedRawOpenAIOptions ?? {}),\n    ...(promptCacheOptions.promptCacheKey\n      ? { promptCacheKey: promptCacheOptions.promptCacheKey }\n      : {}),\n    ...(!skipCacheRetention && promptCacheOptions.promptCacheRetention\n      ? { promptCacheRetention: promptCacheOptions.promptCacheRetention }\n      : {}),\n    // The caller's explicit `reasoningEffort` wins over the resolved default\n    // (env var, or Cerebras \"low\") — same precedence pattern as promptCacheKey.\n    ...((sanitizedRawOpenAIOptions as { reasoningEffort?: unknown } | undefined)\n      ?.reasoningEffort === undefined && reasoningEffort\n      ? { reasoningEffort }\n      : {}),\n  };\n\n  const providerOptions = {\n    ...rest,\n    ...(Object.keys(openaiOptions).length > 0 ? { openai: openaiOptions } : {}),\n  };\n\n  return Object.keys(providerOptions).length > 0 ? providerOptions : undefined;\n}\n\nfunction buildStructuredOutput(responseSchema: unknown): NativeOutput {\n  if (\n    responseSchema &&\n    typeof responseSchema === \"object\" &&\n    \"responseFormat\" in responseSchema &&\n    \"parseCompleteOutput\" in responseSchema\n  ) {\n    return responseSchema as NativeOutput;\n  }\n\n  const schemaOptions =\n    responseSchema && typeof responseSchema === \"object\" && \"schema\" in responseSchema\n      ? (responseSchema as { schema: unknown; name?: string; description?: string })\n      : { schema: responseSchema };\n\n  return Output.object({\n    schema: jsonSchema(sanitizeJsonSchema(schemaOptions.schema, true)),\n    ...(schemaOptions.name ? { name: schemaOptions.name } : {}),\n    ...(schemaOptions.description ? { description: schemaOptions.description } : {}),\n  }) as NativeOutput;\n}\n\nfunction normalizeNativeTools(\n  tools: unknown,\n  options: { cerebrasMode?: boolean } = {}\n): ToolSet | undefined {\n  if (!tools) {\n    return undefined;\n  }\n\n  // Existing AI SDK callers already pass a ToolSet keyed by tool name. Keep it\n  // intact so custom tool instances, execute hooks, and dynamic tool metadata\n  // are preserved.\n  if (!Array.isArray(tools)) {\n    return tools as ToolSet;\n  }\n\n  const toolSet: Record<string, unknown> = {};\n\n  for (const rawTool of tools) {\n    const tool = asRecord(rawTool);\n    const functionTool = asRecord(tool.function);\n    const name = firstString(tool.name, functionTool.name);\n\n    if (!name) {\n      throw new Error(\"[OpenAI] Native tool definition is missing a name.\");\n    }\n\n    const description = firstString(tool.description, functionTool.description);\n    // Default to a permissive object schema. The empty-properties shape\n    // (`{ type: \"object\", properties: {}, additionalProperties: false }`) is\n    // accepted by OpenAI but rejected by strict-grammar providers like\n    // Cerebras with `Object fields require at least one of: 'properties' or\n    // 'anyOf' with a list of possible properties`.\n    const rawSchema =\n      tool.parameters ?? functionTool.parameters ?? ({ type: \"object\" } satisfies JSONSchema7);\n    let inputSchema = sanitizeJsonSchema(rawSchema, true);\n    if (options.cerebrasMode) {\n      // User-supplied schemas may still contain empty-properties subobjects\n      // even after sanitizeJsonSchema. Apply Cerebras-specific normalization\n      // recursively so deep schemas are accepted by the grammar compiler.\n      // Pass isRoot: true so the top-level invariant is enforced (must be\n      // type:\"object\" with no root oneOf/anyOf/enum/not).\n      inputSchema = normalizeSchemaForCerebras(inputSchema, true) as JSONSchema7;\n    }\n\n    // Cerebras's grammar compiler rejects function names containing characters\n    // outside `[a-zA-Z0-9_-]` (e.g. `math.factorial`). The AI SDK looks up\n    // tools by the registered key, so we register under the sanitized name AND\n    // surface it to the model under that name. Tool calls come back with the\n    // sanitized name, which the runtime resolves through its action registry —\n    // any caller relying on dotted action names should pre-sanitize.\n    const registeredName = options.cerebrasMode ? sanitizeFunctionNameForCerebras(name) : name;\n\n    toolSet[registeredName] = {\n      ...(description ? { description } : {}),\n      inputSchema: jsonSchema(inputSchema as JSONSchema7),\n    };\n  }\n\n  return Object.keys(toolSet).length > 0 ? (toolSet as ToolSet) : undefined;\n}\n\nfunction normalizeNativeMessages(messages: unknown): ModelMessage[] | undefined {\n  if (!Array.isArray(messages)) {\n    return undefined;\n  }\n\n  return messages.map((message) => normalizeNativeMessage(message));\n}\n\nfunction normalizeNativeMessage(message: unknown): ModelMessage {\n  const raw = asRecord(message);\n  const providerOptions = asOptionalRecord(raw.providerOptions);\n\n  if (raw.role === \"system\") {\n    return {\n      role: \"system\",\n      content: stringifyMessageContent(raw.content),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  if (raw.role === \"assistant\") {\n    return {\n      role: \"assistant\",\n      content: normalizeAssistantContent(raw),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  if (raw.role === \"tool\") {\n    return {\n      role: \"tool\",\n      content: normalizeToolContent(raw),\n      ...(providerOptions ? { providerOptions } : {}),\n    } as ModelMessage;\n  }\n\n  return {\n    role: \"user\",\n    content: normalizeUserContent(raw.content),\n    ...(providerOptions ? { providerOptions } : {}),\n  } as ModelMessage;\n}\n\n/**\n * Strip reasoning-only parts from outbound assistant content.\n *\n * OpenAI-spec reasoning models (Cerebras gpt-oss-120b, OpenAI o1/o3,\n * DeepSeek R1, and similar families) return reasoning in the assistant\n * response — either as a separate `reasoning` / `reasoning_content`\n * field, or as content parts with `type: \"reasoning\"`. Echoing those\n * back to the next turn is wrong on both ends:\n *   - Cerebras returns HTTP 400 (`messages.X.assistant.reasoning_content:\n *     property is unsupported`).\n *   - OpenAI silently drops them, which wastes prompt tokens.\n *\n * The AI SDK upstream of this normalizer surfaces those reasoning blocks\n * as `{ type: \"reasoning\", ... }` content parts. We drop them here so\n * the wire stays spec-clean for the next turn. The reasoning itself\n * remains usable as a single-turn signal (still on the response object);\n * we only refuse to round-trip it.\n */\nfunction stripReasoningParts(content: unknown[]): unknown[] {\n  return content.filter((part) => {\n    if (!part || typeof part !== \"object\") return true;\n    const type = (part as { type?: unknown }).type;\n    return type !== \"reasoning\" && type !== \"thinking\";\n  });\n}\n\nfunction normalizeAssistantContent(message: Record<string, unknown>): unknown {\n  const toolCalls = Array.isArray(message.toolCalls) ? message.toolCalls : [];\n\n  if (toolCalls.length === 0) {\n    if (Array.isArray(message.content)) {\n      return stripReasoningParts(message.content);\n    }\n    if (typeof message.content === \"string\") {\n      return message.content;\n    }\n    return \"\";\n  }\n\n  const parts: unknown[] = [];\n  if (typeof message.content === \"string\" && message.content.length > 0) {\n    parts.push({ type: \"text\", text: message.content });\n  } else if (Array.isArray(message.content)) {\n    parts.push(...stripReasoningParts(message.content));\n  }\n\n  for (const toolCall of toolCalls) {\n    const rawCall = asRecord(toolCall);\n    const rawFunction = asRecord(rawCall.function);\n    const toolCallId = firstString(rawCall.toolCallId, rawCall.id);\n    const toolName = firstString(rawCall.toolName, rawCall.name, rawFunction.name);\n\n    if (!toolCallId || !toolName) {\n      continue;\n    }\n\n    parts.push({\n      type: \"tool-call\",\n      toolCallId,\n      toolName,\n      input: parseToolCallInput(rawCall, rawFunction),\n    });\n  }\n\n  return parts;\n}\n\nfunction normalizeToolContent(message: Record<string, unknown>): unknown[] {\n  if (Array.isArray(message.content)) {\n    return message.content;\n  }\n\n  const toolCallId = firstString(message.toolCallId, message.id) ?? \"tool-call\";\n  const toolName = firstString(message.toolName, message.name) ?? \"tool\";\n  const parsed = parseJsonIfPossible(message.content);\n\n  return [\n    {\n      type: \"tool-result\",\n      toolCallId,\n      toolName,\n      output:\n        typeof parsed === \"string\"\n          ? { type: \"text\", value: parsed }\n          : { type: \"json\", value: parsed },\n    },\n  ];\n}\n\nfunction normalizeUserContent(content: unknown): UserContent {\n  if (Array.isArray(content)) {\n    return content as UserContent;\n  }\n  return stringifyMessageContent(content);\n}\n\nfunction stringifyMessageContent(content: unknown): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n  if (content == null) {\n    return \"\";\n  }\n  return typeof content === \"object\" ? JSON.stringify(content) : String(content);\n}\n\nfunction parseToolCallInput(\n  rawCall: Record<string, unknown>,\n  rawFunction: Record<string, unknown>\n): unknown {\n  if (\"input\" in rawCall) {\n    return rawCall.input;\n  }\n  return parseJsonIfPossible(rawCall.arguments ?? rawFunction.arguments ?? {});\n}\n\nfunction parseJsonIfPossible(value: unknown): unknown {\n  if (typeof value !== \"string\") {\n    return value ?? \"\";\n  }\n  try {\n    return JSON.parse(value);\n  } catch {\n    return value;\n  }\n}\n\nfunction normalizeToolChoice(toolChoice: unknown): ToolChoice<ToolSet> | undefined {\n  if (!toolChoice) {\n    return undefined;\n  }\n\n  if (\n    typeof toolChoice === \"string\" &&\n    (toolChoice === \"auto\" || toolChoice === \"none\" || toolChoice === \"required\")\n  ) {\n    return toolChoice;\n  }\n\n  const choice = asRecord(toolChoice);\n  if (choice.type === \"tool\") {\n    if (typeof choice.toolName === \"string\" && choice.toolName.length > 0) {\n      return toolChoice as ToolChoice<ToolSet>;\n    }\n    const toolName = firstString(choice.toolName, choice.name);\n    if (toolName) {\n      return { type: \"tool\", toolName };\n    }\n  }\n\n  if (choice.type === \"function\") {\n    const fn = asRecord(choice.function);\n    const toolName = firstString(fn.name);\n    if (toolName) {\n      return { type: \"tool\", toolName };\n    }\n  }\n\n  const namedTool = firstString(choice.name);\n  if (namedTool) {\n    return { type: \"tool\", toolName: namedTool };\n  }\n\n  return toolChoice as ToolChoice<ToolSet>;\n}\n\nfunction hasIllegalStrictRoot(node: Record<string, unknown>): boolean {\n  // Strict-mode JSON schema validators on OpenAI-compatible providers (Groq,\n  // Cerebras, OpenAI strict tools) reject tool-parameters whose top level is\n  // not `type: \"object\"` or carries `oneOf`/`anyOf`/`enum`/`not` at the root.\n  // The error wording varies by provider but the constraint is uniform.\n  if (node.type !== \"object\") return true;\n  if (Array.isArray(node.oneOf) && node.oneOf.length > 0) return true;\n  if (Array.isArray(node.anyOf) && node.anyOf.length > 0) return true;\n  if (Array.isArray(node.enum)) return true;\n  if (node.not !== undefined) return true;\n  return false;\n}\n\nfunction sanitizeJsonSchema(schema: unknown, isRoot = false): JSONSchema7 {\n  if (!schema || typeof schema !== \"object\" || Array.isArray(schema)) {\n    // Permissive fallback: no `properties: {}`/`additionalProperties: false`\n    // pair, which strict-grammar providers reject. See `normalizeSchemaForCerebras`\n    // in @elizaos/core for the rationale.\n    return { type: \"object\" };\n  }\n\n  const record = schema as Record<string, unknown>;\n  let sanitized: Record<string, unknown> = { ...record };\n\n  if (typeof sanitized.type !== \"string\") {\n    const inferredType = inferJsonSchemaType(sanitized, isRoot);\n    if (inferredType) {\n      sanitized.type = inferredType;\n    }\n  }\n\n  if (isRoot && hasIllegalStrictRoot(sanitized)) {\n    // Wrap the original schema under properties.value. Strict-tool callers\n    // that unwrap arguments will see `{ value: <original> }`. The recursion\n    // below normalises the wrapped child like any other property.\n    sanitized = {\n      type: \"object\",\n      properties: { value: { ...record } },\n      required: [\"value\"],\n      additionalProperties: false,\n    };\n  }\n\n  if (\n    sanitized.properties &&\n    typeof sanitized.properties === \"object\" &&\n    !Array.isArray(sanitized.properties)\n  ) {\n    const properties: Record<string, unknown> = {};\n    for (const [key, value] of Object.entries(sanitized.properties as Record<string, unknown>)) {\n      properties[key] = sanitizeJsonSchema(value);\n    }\n    sanitized.properties = properties;\n\n    const propertyKeys = Object.keys(properties);\n    const existingRequired = Array.isArray(sanitized.required)\n      ? sanitized.required.filter((key): key is string => typeof key === \"string\")\n      : [];\n    sanitized.required = [...new Set([...existingRequired, ...propertyKeys])];\n  }\n\n  if (sanitized.type === \"object\" && sanitized.additionalProperties !== false) {\n    sanitized.additionalProperties = false;\n  }\n\n  if (sanitized.items) {\n    sanitized.items = Array.isArray(sanitized.items)\n      ? sanitized.items.map((item) => sanitizeJsonSchema(item))\n      : sanitizeJsonSchema(sanitized.items);\n  }\n\n  for (const unionKey of [\"anyOf\", \"oneOf\", \"allOf\"] as const) {\n    const value = sanitized[unionKey];\n    if (Array.isArray(value)) {\n      sanitized[unionKey] = value.map((item) => sanitizeJsonSchema(item));\n    }\n  }\n\n  return sanitized as JSONSchema7;\n}\n\nfunction inferJsonSchemaType(schema: Record<string, unknown>, isRoot: boolean): string | undefined {\n  if (\n    \"properties\" in schema ||\n    \"required\" in schema ||\n    \"additionalProperties\" in schema ||\n    isRoot\n  ) {\n    return \"object\";\n  }\n  if (\"items\" in schema) {\n    return \"array\";\n  }\n  if (Array.isArray(schema.enum) && schema.enum.length > 0) {\n    const types = new Set(schema.enum.map((value) => typeof value));\n    if (types.size === 1) {\n      const [type] = [...types];\n      if (type === \"string\" || type === \"number\" || type === \"boolean\") {\n        return type;\n      }\n    }\n  }\n  return undefined;\n}\n\nfunction asRecord(value: unknown): Record<string, unknown> {\n  return value && typeof value === \"object\" && !Array.isArray(value)\n    ? (value as Record<string, unknown>)\n    : {};\n}\n\nfunction asOptionalRecord(value: unknown): Record<string, unknown> | undefined {\n  return value && typeof value === \"object\" && !Array.isArray(value)\n    ? (value as Record<string, unknown>)\n    : undefined;\n}\n\nfunction firstString(...values: unknown[]): string | undefined {\n  for (const value of values) {\n    if (typeof value === \"string\" && value.length > 0) {\n      return value;\n    }\n  }\n  return undefined;\n}\n\nfunction usesNativeTextResult(params: GenerateTextParamsWithOpenAIOptions): boolean {\n  return Boolean(params.messages || params.tools || params.toolChoice || params.responseSchema);\n}\n\nfunction buildNativeTextResult(\n  result: {\n    text: string;\n    toolCalls?: unknown[];\n    finishReason?: string;\n    usage?: LanguageModelUsage;\n    providerMetadata?: unknown;\n  },\n  modelName?: string\n): NativeGenerateTextResult {\n  return {\n    text: result.text,\n    toolCalls: result.toolCalls ?? [],\n    finishReason: result.finishReason,\n    usage: convertUsage(result.usage),\n    providerMetadata: mergeProviderModelName(result.providerMetadata, modelName),\n  };\n}\n\nfunction handledPromise<T>(value: T | PromiseLike<T>): Promise<T> {\n  const promise = Promise.resolve(value);\n  promise.catch(() => {\n    // The streaming path primarily consumes `textStream`. AI SDK companion\n    // promises such as `text` can reject later on empty streams even when no\n    // caller requested them, which otherwise surfaces as an unhandled rejection.\n  });\n  return promise;\n}\n\nfunction handledMappedPromise<T, U>(\n  value: T | PromiseLike<T>,\n  mapper: (resolved: T) => U | PromiseLike<U>\n): Promise<U> {\n  return handledPromise(handledPromise(value).then(mapper));\n}\n\nfunction mergeProviderModelName(providerMetadata: unknown, modelName?: string): unknown {\n  if (!modelName) {\n    return providerMetadata;\n  }\n  if (\n    providerMetadata &&\n    typeof providerMetadata === \"object\" &&\n    !Array.isArray(providerMetadata)\n  ) {\n    return {\n      ...(providerMetadata as Record<string, unknown>),\n      modelName,\n    };\n  }\n  return { modelName };\n}\n\nfunction createLlmCallDetails(\n  modelName: string,\n  params: GenerateTextParams,\n  systemPrompt: string | undefined,\n  actionType: string,\n  modelType?: ModelTypeName,\n  providerOptions?: Record<string, unknown>,\n  generateParams?: NativeTextParams\n): RecordLlmCallDetails {\n  const originalParams = params as GenerateTextParamsWithOpenAIOptions;\n  const nativeParams = generateParams as\n    | (NativeTextParams & {\n        output?: unknown;\n        maxOutputTokens?: unknown;\n      })\n    | undefined;\n  const nativePrompt = nativeParams && \"prompt\" in nativeParams ? nativeParams.prompt : undefined;\n  const nativeMessages =\n    nativeParams && \"messages\" in nativeParams && Array.isArray(nativeParams.messages)\n      ? nativeParams.messages\n      : undefined;\n  const nativeSystem =\n    typeof nativeParams?.system === \"string\" ? nativeParams.system : systemPrompt;\n  return {\n    model: modelName,\n    modelType,\n    provider: \"vercel-ai-sdk\",\n    systemPrompt: nativeSystem ?? \"\",\n    userPrompt:\n      typeof nativePrompt === \"string\"\n        ? nativePrompt\n        : typeof params.prompt === \"string\"\n          ? params.prompt\n          : \"\",\n    prompt: typeof nativePrompt === \"string\" ? nativePrompt : undefined,\n    messages: nativeMessages,\n    tools: nativeParams?.tools ?? originalParams.tools,\n    toolChoice: nativeParams?.toolChoice ?? originalParams.toolChoice,\n    output:\n      nativeParams?.output !== undefined\n        ? buildTrajectoryOutputDescriptor(originalParams.responseSchema, nativeParams.output)\n        : undefined,\n    responseSchema: originalParams.responseSchema,\n    providerOptions:\n      providerOptions ?? nativeParams?.providerOptions ?? originalParams.providerOptions,\n    temperature: params.temperature ?? 0,\n    maxTokens:\n      typeof nativeParams?.maxOutputTokens === \"number\"\n        ? nativeParams.maxOutputTokens\n        : params.omitMaxTokens\n          ? 0\n          : (params.maxTokens ?? 8192),\n    maxTokensOmitted:\n      params.omitMaxTokens && typeof nativeParams?.maxOutputTokens !== \"number\" ? true : undefined,\n    purpose: \"external_llm\",\n    actionType,\n  };\n}\n\nfunction buildTrajectoryOutputDescriptor(responseSchema: unknown, output: unknown): unknown {\n  if (responseSchema !== undefined) {\n    return {\n      type: \"object\",\n      schema: responseSchema,\n    };\n  }\n  return toTrajectoryJsonSafe(output);\n}\n\nfunction toTrajectoryJsonSafe(value: unknown): unknown {\n  try {\n    return JSON.parse(\n      JSON.stringify(value, (_key, nested) => {\n        if (typeof nested === \"function\") return undefined;\n        if (typeof nested === \"bigint\") return nested.toString();\n        return nested;\n      })\n    ) as unknown;\n  } catch {\n    return String(value);\n  }\n}\n\nfunction applyUsageToDetails(\n  details: RecordLlmCallDetails,\n  usage: LanguageModelUsage | undefined\n): void {\n  if (!usage) {\n    return;\n  }\n  details.promptTokens = usage.inputTokens ?? 0;\n  details.completionTokens = usage.outputTokens ?? 0;\n}\n\n// ============================================================================\n// Core Generation Function\n// ============================================================================\n\n/**\n * Generates text using the specified model type.\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @param modelType - The type of model (TEXT_SMALL or TEXT_LARGE)\n * @param getModelFn - Function to get the model name\n * @returns Generated text or stream result\n */\nasync function generateTextByModelType(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams,\n  modelType: ModelTypeName,\n  getModelFn: ModelNameGetter\n): Promise<string | TextStreamResult> {\n  const paramsWithAttachments = params as GenerateTextParamsWithOpenAIOptions;\n  const openai = createOpenAIClient(runtime);\n  const modelName = resolveRequestedModelName(paramsWithAttachments, runtime, getModelFn);\n\n  logger.debug(`[OpenAI] Using ${modelType} model: ${modelName}`);\n  const providerOptions = resolveProviderOptions(params, runtime, modelName);\n  const hasAttachments = (paramsWithAttachments.attachments?.length ?? 0) > 0;\n  const userContent = hasAttachments ? buildUserContent(paramsWithAttachments) : undefined;\n  const shouldReturnNativeResult = usesNativeTextResult(paramsWithAttachments);\n\n  const systemPrompt = resolveEffectiveSystemPrompt({\n    params: paramsWithAttachments,\n    fallback: buildCanonicalSystemPrompt({ character: runtime.character }),\n  });\n  const agentName = paramsWithAttachments.providerOptions?.agentName;\n  const telemetryConfig: NativeTelemetrySettings = {\n    isEnabled: getExperimentalTelemetry(runtime),\n    functionId: agentName ? `agent:${agentName}` : undefined,\n    metadata: agentName ? { agentName } : undefined,\n  };\n\n  // Chat Completions is the default: broadest compatibility, and it works\n  // against every OpenAI-compatible endpoint (Cerebras, local servers, proxies).\n  // gpt-5 / gpt-5-mini reasoning models ignore temperature/penalty/stop params.\n  //\n  const model = openai.chat(modelName);\n  const cerebrasMode = isCerebrasMode(runtime);\n  const normalizedTools = normalizeNativeTools(paramsWithAttachments.tools, {\n    cerebrasMode,\n  });\n  const normalizedToolChoice = normalizeToolChoice(paramsWithAttachments.toolChoice);\n  const normalizedMessages = normalizeNativeMessages(paramsWithAttachments.messages);\n  const wireMessages = dropDuplicateLeadingSystemMessage(normalizedMessages, systemPrompt);\n  const effectiveMessages =\n    wireMessages && wireMessages.length > 0 ? wireMessages : normalizedMessages;\n  const promptText =\n    typeof params.prompt === \"string\" && params.prompt.length > 0 ? params.prompt : \"\";\n  const promptOrMessages: NativePrompt =\n    effectiveMessages && effectiveMessages.length > 0\n      ? { messages: effectiveMessages }\n      : userContent\n        ? { messages: [{ role: \"user\" as const, content: userContent }] }\n        : { prompt: promptText };\n  // elizaOS callers pass `responseFormat: { type: \"json_object\" | \"text\" }`\n  // (see `GenerateTextParams` in @elizaos/core). The AI SDK's equivalent\n  // is `responseFormat: { type: \"json\" }` (which translates to\n  // `response_format: { type: \"json_object\" }` at the OpenAI wire layer).\n  // Translate the shape so the param actually reaches the API call —\n  // before this, callers asking for json_object were silently ignored\n  // and Cerebras returned plain text, dropping us into the simple-reply\n  // fallback every turn.\n  const callerResponseFormat = (paramsWithAttachments as { responseFormat?: unknown })\n    .responseFormat;\n  const responseFormatType =\n    typeof callerResponseFormat === \"string\"\n      ? callerResponseFormat\n      : callerResponseFormat &&\n          typeof callerResponseFormat === \"object\" &&\n          \"type\" in callerResponseFormat\n        ? (callerResponseFormat as { type: string }).type\n        : undefined;\n  const wireResponseFormat: { type: \"json\" } | { type: \"text\" } | undefined =\n    responseFormatType === \"json_object\"\n      ? { type: \"json\" }\n      : responseFormatType === \"text\"\n        ? { type: \"text\" }\n        : undefined;\n\n  const generateParams: NativeTextParams = {\n    model,\n    ...promptOrMessages,\n    system: systemPrompt,\n    allowSystemInMessages: true,\n    // Omit the cap when the caller opted out (direct-channel Stage-1) so the\n    // model's own max applies — a hardcoded value 400s when it exceeds the\n    // model's limit. Other callers keep the 8192 default.\n    ...(params.omitMaxTokens ? {} : { maxOutputTokens: params.maxTokens ?? 8192 }),\n    experimental_telemetry: telemetryConfig,\n    ...(normalizedTools ? { tools: normalizedTools } : {}),\n    ...(normalizedToolChoice ? { toolChoice: normalizedToolChoice } : {}),\n    // Cerebras's OpenAI-compatible endpoint does not accept the\n    // `response_format: { type: \"json_schema\", ... }` payload that the AI SDK\n    // emits when `output: Output.object(...)` is set. Fall back to relying on\n    // `responseFormat: { type: \"json_object\" }` (already passed by callers)\n    // plus the schema embedded in the prompt body.\n    ...(paramsWithAttachments.responseSchema && !isCerebrasMode(runtime)\n      ? { output: buildStructuredOutput(paramsWithAttachments.responseSchema) }\n      : {}),\n    ...(wireResponseFormat ? { responseFormat: wireResponseFormat } : {}),\n    ...(providerOptions ? { providerOptions: providerOptions as NativeProviderOptions } : {}),\n  };\n\n  // Handle streaming mode\n  if (params.stream) {\n    const details = createLlmCallDetails(\n      modelName,\n      params,\n      systemPrompt,\n      \"ai.streamText\",\n      modelType,\n      providerOptions,\n      generateParams\n    );\n    details.response = \"\";\n    const result = await recordLlmCall(runtime, details, () => streamText(generateParams));\n\n    return {\n      textStream: (async function* textStreamWithCallback() {\n        for await (const chunk of result.textStream) {\n          params.onStreamChunk?.(chunk);\n          yield chunk;\n        }\n      })(),\n      text: handledPromise(result.text),\n      ...(shouldReturnNativeResult ? { toolCalls: handledPromise(result.toolCalls) } : {}),\n      usage: handledMappedPromise(result.usage, convertUsage),\n      finishReason: handledMappedPromise(result.finishReason, (r) => r as string | undefined),\n    };\n  }\n\n  // Non-streaming mode\n  const details = createLlmCallDetails(\n    modelName,\n    params,\n    systemPrompt,\n    \"ai.generateText\",\n    modelType,\n    providerOptions,\n    generateParams\n  );\n  const result = await recordLlmCall(runtime, details, async () => {\n    const result = await generateText(generateParams);\n    details.response = result.text;\n    details.toolCalls = result.toolCalls;\n    details.finishReason = result.finishReason as string | undefined;\n    details.providerMetadata = result.providerMetadata;\n    applyUsageToDetails(details, result.usage);\n    return result;\n  });\n\n  if (result.usage) {\n    emitModelUsageEvent(runtime, modelType, params.prompt ?? \"\", result.usage);\n  }\n\n  if (shouldReturnNativeResult) {\n    return buildNativeTextResult(result, modelName) as NativeTextModelResult;\n  }\n\n  return result.text;\n}\n\n// ============================================================================\n// Public Handlers\n// ============================================================================\n\n/**\n * Handles TEXT_SMALL model requests.\n *\n * Uses the configured small model (default: gpt-5-mini).\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @returns Generated text or stream result\n */\nexport async function handleTextSmall(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ModelType.TEXT_SMALL, getSmallModel);\n}\n\nexport async function handleTextNano(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_NANO_MODEL_TYPE, getNanoModel);\n}\n\nexport async function handleTextMedium(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_MEDIUM_MODEL_TYPE, getMediumModel);\n}\n\n/**\n * Handles TEXT_LARGE model requests.\n *\n * Uses the configured large model (default: gpt-5).\n *\n * @param runtime - The agent runtime\n * @param params - Generation parameters\n * @returns Generated text or stream result\n */\nexport async function handleTextLarge(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ModelType.TEXT_LARGE, getLargeModel);\n}\n\nexport async function handleTextMega(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, TEXT_MEGA_MODEL_TYPE, getMegaModel);\n}\n\nexport async function handleResponseHandler(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(\n    runtime,\n    params,\n    RESPONSE_HANDLER_MODEL_TYPE,\n    getResponseHandlerModel\n  );\n}\n\nexport async function handleActionPlanner(\n  runtime: IAgentRuntime,\n  params: GenerateTextParams\n): Promise<string | TextStreamResult> {\n  return generateTextByModelType(runtime, params, ACTION_PLANNER_MODEL_TYPE, getActionPlannerModel);\n}\n\n// ─── Test-only exports ──────────────────────────────────────────────────────\n// These are exported for the shape tests in `__tests__/reasoning-effort.shape.test.ts`.\n// Not part of the public API; do not import outside tests.\n\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_resolveProviderOptions = resolveProviderOptions;\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_normalizeNativeMessages = normalizeNativeMessages;\n/** @internal — exported for unit tests only. */\nexport const __INTERNAL_stripReasoningParts = stripReasoningParts;\n",
     "import { createOpenAI, type OpenAIProvider } from \"@ai-sdk/openai\";\nimport type { IAgentRuntime } from \"@elizaos/core\";\nimport { getApiKey, getBaseURL, isProxyMode } from \"../utils/config\";\n\nconst PROXY_API_KEY = \"sk-proxy\";\n\nexport function createOpenAIClient(runtime: IAgentRuntime): OpenAIProvider {\n  const baseURL = getBaseURL(runtime);\n  const apiKey = getApiKey(runtime);\n\n  if (!apiKey && isProxyMode(runtime)) {\n    return createOpenAI({\n      apiKey: PROXY_API_KEY,\n      baseURL,\n    });\n  }\n\n  if (!apiKey) {\n    throw new Error(\n      \"OPENAI_API_KEY is required. Set it in your environment variables or runtime settings.\"\n    );\n  }\n\n  return createOpenAI({\n    apiKey,\n    baseURL,\n  });\n}\n",
     "import type { IAgentRuntime, ModelTypeName } from \"@elizaos/core\";\nimport { ModelType } from \"@elizaos/core\";\nimport {\n  encodingForModel,\n  getEncoding,\n  type Tiktoken,\n  type TiktokenEncoding,\n  type TiktokenModel,\n} from \"js-tiktoken\";\nimport { getLargeModel, getSmallModel } from \"./config\";\n\ntype SupportedEncoding = \"cl100k_base\" | \"o200k_base\";\n\nfunction resolveTokenizerEncoding(modelName: string): Tiktoken {\n  const normalized = modelName.toLowerCase();\n  const fallbackEncoding: SupportedEncoding = normalized.includes(\"4o\")\n    ? \"o200k_base\"\n    : \"cl100k_base\";\n  try {\n    return encodingForModel(modelName as TiktokenModel);\n  } catch {\n    return getEncoding(fallbackEncoding as TiktokenEncoding);\n  }\n}\n\nfunction getModelName(runtime: IAgentRuntime, modelType: ModelTypeName): string {\n  if (modelType === ModelType.TEXT_SMALL) {\n    return getSmallModel(runtime);\n  }\n  return getLargeModel(runtime);\n}\n\nexport function tokenizeText(\n  runtime: IAgentRuntime,\n  modelType: ModelTypeName,\n  text: string\n): number[] {\n  const modelName = getModelName(runtime, modelType);\n  const encoder = resolveTokenizerEncoding(modelName);\n  return encoder.encode(text);\n}\n\nexport function detokenizeText(\n  runtime: IAgentRuntime,\n  modelType: ModelTypeName,\n  tokens: number[]\n): string {\n  const modelName = getModelName(runtime, modelType);\n  const encoder = resolveTokenizerEncoding(modelName);\n  return encoder.decode(tokens);\n}\n\nexport function countTokens(\n  runtime: IAgentRuntime,\n  modelType: ModelTypeName,\n  text: string\n): number {\n  const tokens = tokenizeText(runtime, modelType, text);\n  return tokens.length;\n}\n\nexport function truncateToTokenLimit(\n  runtime: IAgentRuntime,\n  modelType: ModelTypeName,\n  text: string,\n  maxTokens: number\n): string {\n  const tokens = tokenizeText(runtime, modelType, text);\n  if (tokens.length <= maxTokens) {\n    return text;\n  }\n  const truncatedTokens = tokens.slice(0, maxTokens);\n  return detokenizeText(runtime, modelType, truncatedTokens);\n}\n",
     "import type { DetokenizeTextParams, IAgentRuntime, TokenizeTextParams } from \"@elizaos/core\";\nimport { detokenizeText, tokenizeText } from \"../utils/tokenization\";\n\nexport async function handleTokenizerEncode(\n  runtime: IAgentRuntime,\n  params: TokenizeTextParams\n): Promise<number[]> {\n  if (!params.prompt) {\n    throw new Error(\"Tokenization requires a non-empty prompt\");\n  }\n  const modelType = params.modelType;\n  return tokenizeText(runtime, modelType, params.prompt);\n}\n\nexport async function handleTokenizerDecode(\n  runtime: IAgentRuntime,\n  params: DetokenizeTextParams\n): Promise<string> {\n  if (!params.tokens || !Array.isArray(params.tokens)) {\n    throw new Error(\"Detokenization requires a valid tokens array\");\n  }\n  if (params.tokens.length === 0) {\n    return \"\";\n  }\n  for (let i = 0; i < params.tokens.length; i++) {\n    const token = params.tokens[i];\n    if (typeof token !== \"number\" || !Number.isFinite(token)) {\n      throw new Error(`Invalid token at index ${i}: expected number`);\n    }\n  }\n  const modelType = params.modelType;\n  return detokenizeText(runtime, modelType, params.tokens);\n}\n",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@elizaos/plugin-openai",
-  "version": "2.0.3-beta.4",
+  "version": "2.0.3-beta.5",
   "type": "module",
   "main": "dist/node/index.node.js",
   "module": "dist/node/index.node.js",
@@ -60,8 +60,8 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.14",
-    "@elizaos/core": "2.0.3-beta.4",
-    "@elizaos/test-harness": "2.0.3-beta.4",
+    "@elizaos/core": "2.0.3-beta.5",
+    "@elizaos/test-harness": "2.0.3-beta.5",
     "@types/bun": "^1.3.8",
     "@types/json-schema": "^7.0.15",
     "@types/node": "^25.0.3",
@@ -70,7 +70,7 @@
     "vitest": "^4.0.0"
   },
   "peerDependencies": {
-    "@elizaos/core": "2.0.3-beta.4",
+    "@elizaos/core": "2.0.3-beta.5",
     "zod": "^4.4.3"
   },
   "scripts": {
@@ -78,7 +78,7 @@
     "dev": "bun --hot build.ts",
     "lint": "bunx @biomejs/biome check --write --unsafe .",
     "lint:check": "bunx @biomejs/biome check .",
-    "clean": "rm -rf dist .turbo",
+    "clean": "node ../../packages/scripts/rm-path-recursive.mjs dist .turbo",
     "format": "bunx @biomejs/biome format --write .",
     "format:check": "bunx @biomejs/biome format .",
     "typecheck": "tsgo --noEmit",
@@ -245,7 +245,7 @@
       }
     }
   },
-  "gitHead": "f76f55793a0fb8d6b869dd8ddce03970de061e34",
+  "gitHead": "ff6157011c9459670021cc28a6797592a78b8817",
   "eliza": {
     "platforms": [
       "browser",