@langchain/openai 1.2.4 → 1.2.6

package/dist/tools/computerUse.cjs

@@ -60,7 +60,7 @@ const ComputerUseWaitActionSchema = zod_v4.z.object({
 	type: zod_v4.z.literal("wait"),
 	duration: zod_v4.z.number().optional()
 });
-const ComputerUseActionUnionSchema = zod_v4.z.discriminatedUnion("type", [
+const ComputerUseActionUnionSchema = zod_v4.z.union([
 	ComputerUseScreenshotActionSchema,
 	ComputerUseClickActionSchema,
 	ComputerUseDoubleClickActionSchema,

package/dist/tools/computerUse.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"computerUse.cjs","names":["z","options: ComputerUseOptions","input: ComputerUseInput","runtime: ToolRuntime<{ messages: BaseMessage[] }>","ToolMessage"],"sources":["../../src/tools/computerUse.ts"],"sourcesContent":["/* eslint-disable @typescript-eslint/no-explicit-any */\nimport { z } from \"zod/v4\";\nimport { OpenAI as OpenAIClient } from \"openai\";\nimport { tool, type DynamicStructuredTool } from \"@langchain/core/tools\";\nimport { type ToolRuntime } from \"@langchain/core/tools\";\nimport {\n  ToolMessage,\n  type AIMessage,\n  type BaseMessage,\n} from \"@langchain/core/messages\";\n\n/**\n * The type of computer environment to control.\n */\nexport type ComputerUseEnvironment =\n  | \"browser\"\n  | \"mac\"\n  | \"windows\"\n  | \"linux\"\n  | \"ubuntu\";\n\n/**\n * Re-export action types from OpenAI SDK for convenience.\n */\nexport type ComputerUseClickAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Click;\nexport type ComputerUseDoubleClickAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.DoubleClick;\nexport type ComputerUseDragAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Drag;\nexport type ComputerUseKeypressAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Keypress;\nexport type ComputerUseMoveAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Move;\nexport type ComputerUseScreenshotAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Screenshot;\nexport type ComputerUseScrollAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Scroll;\nexport type ComputerUseTypeAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Type;\nexport type ComputerUseWaitAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Wait;\n\n/**\n * Union type of all computer use actions from OpenAI SDK.\n */\nexport type ComputerUseAction =\n  OpenAIClient.Responses.ResponseComputerToolCall[\"action\"];\n\n// Zod schemas for computer use actions\nconst ComputerUseScreenshotActionSchema = z.object({\n  type: z.literal(\"screenshot\"),\n});\n\nconst ComputerUseClickActionSchema = z.object({\n  type: z.literal(\"click\"),\n  x: z.number(),\n  y: z.number(),\n  button: z.enum([\"left\", \"right\", \"wheel\", \"back\", \"forward\"]).default(\"left\"),\n});\n\nconst ComputerUseDoubleClickActionSchema = z.object({\n  type: z.literal(\"double_click\"),\n  x: z.number(),\n  y: z.number(),\n  button: z.enum([\"left\", \"right\", \"wheel\", \"back\", \"forward\"]).default(\"left\"),\n});\n\nconst ComputerUseDragActionSchema = z.object({\n  type: z.literal(\"drag\"),\n  path: z.array(z.object({ x: z.number(), y: z.number() })),\n});\n\nconst ComputerUseKeypressActionSchema = z.object({\n  type: z.literal(\"keypress\"),\n  keys: z.array(z.string()),\n});\n\nconst ComputerUseMoveActionSchema = z.object({\n  type: z.literal(\"move\"),\n  x: z.number(),\n  y: z.number(),\n});\n\nconst ComputerUseScrollActionSchema = z.object({\n  type: z.literal(\"scroll\"),\n  x: z.number(),\n  y: z.number(),\n  scroll_x: z.number(),\n  scroll_y: z.number(),\n});\n\nconst ComputerUseTypeActionSchema = z.object({\n  type: z.literal(\"type\"),\n  text: z.string(),\n});\n\nconst ComputerUseWaitActionSchema = z.object({\n  type: z.literal(\"wait\"),\n  duration: z.number().optional(),\n});\n\n// Discriminated union schema for individual action types\nconst ComputerUseActionUnionSchema = z.discriminatedUnion(\"type\", [\n  ComputerUseScreenshotActionSchema,\n  ComputerUseClickActionSchema,\n  ComputerUseDoubleClickActionSchema,\n  ComputerUseDragActionSchema,\n  ComputerUseKeypressActionSchema,\n  ComputerUseMoveActionSchema,\n  ComputerUseScrollActionSchema,\n  ComputerUseTypeActionSchema,\n  ComputerUseWaitActionSchema,\n]);\n\n// Schema for the input structure received from parseComputerCall\n// The action is wrapped in an `action` property: { action: { type: 'screenshot' } }\nexport const ComputerUseActionSchema = z.object({\n  action: ComputerUseActionUnionSchema,\n});\n\n// TypeScript types derived from Zod schemas\nexport type ComputerUseScreenshotActionType = z.infer<\n  typeof ComputerUseScreenshotActionSchema\n>;\nexport type ComputerUseClickActionType = z.infer<\n  typeof ComputerUseClickActionSchema\n>;\nexport type ComputerUseDoubleClickActionType = z.infer<\n  typeof ComputerUseDoubleClickActionSchema\n>;\nexport type ComputerUseDragActionType = z.infer<\n  typeof ComputerUseDragActionSchema\n>;\nexport type ComputerUseKeypressActionType = z.infer<\n  typeof ComputerUseKeypressActionSchema\n>;\nexport type ComputerUseMoveActionType = z.infer<\n  typeof ComputerUseMoveActionSchema\n>;\nexport type ComputerUseScrollActionType = z.infer<\n  typeof ComputerUseScrollActionSchema\n>;\nexport type ComputerUseTypeActionType = z.infer<\n  typeof ComputerUseTypeActionSchema\n>;\nexport type ComputerUseWaitActionType = z.infer<\n  typeof ComputerUseWaitActionSchema\n>;\n\n/**\n * Input structure for the Computer Use tool.\n * The action is wrapped in an `action` property.\n */\nexport interface ComputerUseInput {\n  action: ComputerUseAction;\n}\n\nexport type ComputerUseReturnType =\n  | string\n  | Promise<string>\n  | ToolMessage<any>\n  | Promise<ToolMessage<any>>;\n\n/**\n * Options for the Computer Use tool.\n */\nexport interface ComputerUseOptions {\n  /**\n   * The width of the computer display in pixels.\n   */\n  displayWidth: number;\n\n  /**\n   * The height of the computer display in pixels.\n   */\n  displayHeight: number;\n\n  /**\n   * The type of computer environment to control.\n   * - `browser`: Browser automation (recommended for most use cases)\n   * - `mac`: macOS environment\n   * - `windows`: Windows environment\n   * - `linux`: Linux environment\n   * - `ubuntu`: Ubuntu environment\n   */\n  environment: ComputerUseEnvironment;\n\n  /**\n   * Execute function that handles computer action execution.\n   * This function receives the action input and should return a base64-encoded\n   * screenshot of the result.\n   */\n  execute: (\n    action: ComputerUseAction,\n    runtime: ToolRuntime<any, any>\n  ) => ComputerUseReturnType;\n}\n\n/**\n * OpenAI Computer Use tool type for the Responses API.\n */\nexport type ComputerUseTool = OpenAIClient.Responses.ComputerTool;\n\nconst TOOL_NAME = \"computer_use\";\n\n/**\n * Creates a Computer Use tool that allows models to control computer interfaces\n * and perform tasks by simulating mouse clicks, keyboard input, scrolling, and more.\n *\n * **Computer Use** is a practical application of OpenAI's Computer-Using Agent (CUA)\n * model (`computer-use-preview`), which combines vision capabilities with advanced\n * reasoning to simulate controlling computer interfaces.\n *\n * **How it works**:\n * The tool operates in a continuous loop:\n * 1. Model sends computer actions (click, type, scroll, etc.)\n * 2. Your code executes these actions in a controlled environment\n * 3. You capture a screenshot of the result\n * 4. Send the screenshot back to the model\n * 5. Repeat until the task is complete\n *\n * **Important**: Computer use is in beta and requires careful consideration:\n * - Use in sandboxed environments only\n * - Do not use for high-stakes or authenticated tasks\n * - Always implement human-in-the-loop for important decisions\n * - Handle safety checks appropriately\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-computer-use | OpenAI Computer Use Documentation}\n *\n * @param options - Configuration options for the Computer Use tool\n * @returns A Computer Use tool that can be passed to `bindTools`\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"computer-use-preview\" });\n *\n * // With execute callback for automatic action handling\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n *   execute: async (action) => {\n *     if (action.type === \"screenshot\") {\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"click\") {\n *       await page.mouse.click(action.x, action.y, { button: action.button });\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"type\") {\n *       await page.keyboard.type(action.text);\n *       return captureScreenshot();\n *     }\n *     // Handle other actions...\n *     return captureScreenshot();\n *   },\n * });\n *\n * const llmWithComputer = model.bindTools([computer]);\n * const response = await llmWithComputer.invoke(\n *   \"Check the latest news on bing.com\"\n * );\n * ```\n *\n * @example\n * ```typescript\n * // Without execute callback (manual action handling)\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n * });\n *\n * const response = await model.invoke(\"Check the news\", {\n *   tools: [computer],\n * });\n *\n * // Access the computer call from the response\n * const computerCall = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"computer_call\"\n * );\n * if (computerCall) {\n *   console.log(\"Action to execute:\", computerCall.action);\n *   // Execute the action manually, then send back a screenshot\n * }\n * ```\n *\n * @example\n * ```typescript\n * // For macOS desktop automation with Docker\n * const computer = tools.computerUse({\n *   displayWidth: 1920,\n *   displayHeight: 1080,\n *   environment: \"mac\",\n *   execute: async (action) => {\n *     if (action.type === \"click\") {\n *       await dockerExec(\n *         `DISPLAY=:99 xdotool mousemove ${action.x} ${action.y} click 1`,\n *         containerName\n *       );\n *     }\n *     // Capture screenshot from container\n *     return await getDockerScreenshot(containerName);\n *   },\n * });\n * ```\n *\n * @remarks\n * - Only available through the Responses API (not Chat Completions)\n * - Requires `computer-use-preview` model\n * - Actions include: click, double_click, drag, keypress, move, screenshot, scroll, type, wait\n * - Safety checks may be returned that require acknowledgment before proceeding\n * - Use `truncation: \"auto\"` parameter when making requests\n * - Recommended to use with `reasoning.summary` for debugging\n */\nexport function computerUse(options: ComputerUseOptions) {\n  const computerTool = tool(\n    async (\n      input: ComputerUseInput,\n      runtime: ToolRuntime<{ messages: BaseMessage[] }>\n    ) => {\n      /**\n       * get computer_use call id from runtime\n       */\n      const aiMessage = runtime.state?.messages.at(-1) as AIMessage | undefined;\n      const computerToolCall = aiMessage?.tool_calls?.find(\n        (tc) => tc.name === \"computer_use\"\n      );\n      const computerToolCallId = computerToolCall?.id;\n      if (!computerToolCallId) {\n        throw new Error(\"Computer use call id not found\");\n      }\n\n      const result = await options.execute(input.action, runtime);\n\n      /**\n       * make sure {@link ToolMessage} is returned with the correct additional kwargs\n       */\n      if (typeof result === \"string\") {\n        return new ToolMessage({\n          content: result,\n          tool_call_id: computerToolCallId,\n          additional_kwargs: {\n            type: \"computer_call_output\",\n          },\n        });\n      }\n\n      /**\n       * make sure {@link ToolMessage} is returned with the correct additional kwargs\n       */\n      return new ToolMessage({\n        ...result,\n        tool_call_id: computerToolCallId,\n        additional_kwargs: {\n          type: \"computer_call_output\",\n          ...result.additional_kwargs,\n        },\n      });\n    },\n    {\n      name: TOOL_NAME,\n      description:\n        \"Control a computer interface by executing mouse clicks, keyboard input, scrolling, and other actions.\",\n      schema: ComputerUseActionSchema,\n    }\n  );\n\n  computerTool.extras = {\n    ...(computerTool.extras ?? {}),\n    providerToolDefinition: {\n      type: \"computer_use_preview\",\n      display_width: options.displayWidth,\n      display_height: options.displayHeight,\n      environment: options.environment,\n    } satisfies ComputerUseTool,\n  };\n\n  /**\n   * return as typed {@link DynamicStructuredTool} so we don't get any type\n   * errors like \"can't export tool without reference\"\n   */\n  return computerTool as DynamicStructuredTool<\n    typeof ComputerUseActionSchema,\n    ComputerUseInput,\n    unknown,\n    ComputerUseReturnType\n  >;\n}\n"],"mappings":";;;;;;AAkDA,MAAM,oCAAoCA,SAAE,OAAO,EACjD,MAAMA,SAAE,QAAQ,aAAa,CAC9B,EAAC;AAEF,MAAM,+BAA+BA,SAAE,OAAO;CAC5C,MAAMA,SAAE,QAAQ,QAAQ;CACxB,GAAGA,SAAE,QAAQ;CACb,GAAGA,SAAE,QAAQ;CACb,QAAQA,SAAE,KAAK;EAAC;EAAQ;EAAS;EAAS;EAAQ;CAAU,EAAC,CAAC,QAAQ,OAAO;AAC9E,EAAC;AAEF,MAAM,qCAAqCA,SAAE,OAAO;CAClD,MAAMA,SAAE,QAAQ,eAAe;CAC/B,GAAGA,SAAE,QAAQ;CACb,GAAGA,SAAE,QAAQ;CACb,QAAQA,SAAE,KAAK;EAAC;EAAQ;EAAS;EAAS;EAAQ;CAAU,EAAC,CAAC,QAAQ,OAAO;AAC9E,EAAC;AAEF,MAAM,8BAA8BA,SAAE,OAAO;CAC3C,MAAMA,SAAE,QAAQ,OAAO;CACvB,MAAMA,SAAE,MAAMA,SAAE,OAAO;EAAE,GAAGA,SAAE,QAAQ;EAAE,GAAGA,SAAE,QAAQ;CAAE,EAAC,CAAC;AAC1D,EAAC;AAEF,MAAM,kCAAkCA,SAAE,OAAO;CAC/C,MAAMA,SAAE,QAAQ,WAAW;CAC3B,MAAMA,SAAE,MAAMA,SAAE,QAAQ,CAAC;AAC1B,EAAC;AAEF,MAAM,8BAA8BA,SAAE,OAAO;CAC3C,MAAMA,SAAE,QAAQ,OAAO;CACvB,GAAGA,SAAE,QAAQ;CACb,GAAGA,SAAE,QAAQ;AACd,EAAC;AAEF,MAAM,gCAAgCA,SAAE,OAAO;CAC7C,MAAMA,SAAE,QAAQ,SAAS;CACzB,GAAGA,SAAE,QAAQ;CACb,GAAGA,SAAE,QAAQ;CACb,UAAUA,SAAE,QAAQ;CACpB,UAAUA,SAAE,QAAQ;AACrB,EAAC;AAEF,MAAM,8BAA8BA,SAAE,OAAO;CAC3C,MAAMA,SAAE,QAAQ,OAAO;CACvB,MAAMA,SAAE,QAAQ;AACjB,EAAC;AAEF,MAAM,8BAA8BA,SAAE,OAAO;CAC3C,MAAMA,SAAE,QAAQ,OAAO;CACvB,UAAUA,SAAE,QAAQ,CAAC,UAAU;AAChC,EAAC;AAGF,MAAM,+BAA+BA,SAAE,mBAAmB,QAAQ;CAChE;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACD,EAAC;AAIF,MAAa,0BAA0BA,SAAE,OAAO,EAC9C,QAAQ,6BACT,EAAC;AAqFF,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkHlB,SAAgB,YAAYC,SAA6B;CACvD,MAAM,gDACJ,OACEC,OACAC,YACG;;;;EAIH,MAAM,YAAY,QAAQ,OAAO,SAAS,GAAG,GAAG;EAChD,MAAM,mBAAmB,WAAW,YAAY,KAC9C,CAAC,OAAO,GAAG,SAAS,eACrB;EACD,MAAM,qBAAqB,kBAAkB;AAC7C,MAAI,CAAC,mBACH,OAAM,IAAI,MAAM;EAGlB,MAAM,SAAS,MAAM,QAAQ,QAAQ,MAAM,QAAQ,QAAQ;;;;AAK3D,MAAI,OAAO,WAAW,SACpB,QAAO,IAAIC,sCAAY;GACrB,SAAS;GACT,cAAc;GACd,mBAAmB,EACjB,MAAM,uBACP;EACF;;;;AAMH,SAAO,IAAIA,sCAAY;GACrB,GAAG;GACH,cAAc;GACd,mBAAmB;IACjB,MAAM;IACN,GAAG,OAAO;GACX;EACF;CACF,GACD;EACE,MAAM;EACN,aACE;EACF,QAAQ;CACT,EACF;CAED,aAAa,SAAS;EACpB,GAAI,aAAa,UAAU,CAAE;EAC7B,wBAAwB;GACtB,MAAM;GACN,eAAe,QAAQ;GACvB,gBAAgB,QAAQ;GACxB,aAAa,QAAQ;EACtB;CACF;;;;;AAMD,QAAO;AAMR"}
+{"version":3,"file":"computerUse.cjs","names":["z","options: ComputerUseOptions","input: ComputerUseInput","runtime: ToolRuntime<{ messages: BaseMessage[] }>","ToolMessage"],"sources":["../../src/tools/computerUse.ts"],"sourcesContent":["/* eslint-disable @typescript-eslint/no-explicit-any */\nimport { z } from \"zod/v4\";\nimport { OpenAI as OpenAIClient } from \"openai\";\nimport { tool, type DynamicStructuredTool } from \"@langchain/core/tools\";\nimport { type ToolRuntime } from \"@langchain/core/tools\";\nimport {\n  ToolMessage,\n  type AIMessage,\n  type BaseMessage,\n} from \"@langchain/core/messages\";\n\n/**\n * The type of computer environment to control.\n */\nexport type ComputerUseEnvironment =\n  | \"browser\"\n  | \"mac\"\n  | \"windows\"\n  | \"linux\"\n  | \"ubuntu\";\n\n/**\n * Re-export action types from OpenAI SDK for convenience.\n */\nexport type ComputerUseClickAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Click;\nexport type ComputerUseDoubleClickAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.DoubleClick;\nexport type ComputerUseDragAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Drag;\nexport type ComputerUseKeypressAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Keypress;\nexport type ComputerUseMoveAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Move;\nexport type ComputerUseScreenshotAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Screenshot;\nexport type ComputerUseScrollAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Scroll;\nexport type ComputerUseTypeAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Type;\nexport type ComputerUseWaitAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Wait;\n\n/**\n * Union type of all computer use actions from OpenAI SDK.\n */\nexport type ComputerUseAction =\n  OpenAIClient.Responses.ResponseComputerToolCall[\"action\"];\n\n// Zod schemas for computer use actions\nconst ComputerUseScreenshotActionSchema = z.object({\n  type: z.literal(\"screenshot\"),\n});\n\nconst ComputerUseClickActionSchema = z.object({\n  type: z.literal(\"click\"),\n  x: z.number(),\n  y: z.number(),\n  button: z.enum([\"left\", \"right\", \"wheel\", \"back\", \"forward\"]).default(\"left\"),\n});\n\nconst ComputerUseDoubleClickActionSchema = z.object({\n  type: z.literal(\"double_click\"),\n  x: z.number(),\n  y: z.number(),\n  button: z.enum([\"left\", \"right\", \"wheel\", \"back\", \"forward\"]).default(\"left\"),\n});\n\nconst ComputerUseDragActionSchema = z.object({\n  type: z.literal(\"drag\"),\n  path: z.array(z.object({ x: z.number(), y: z.number() })),\n});\n\nconst ComputerUseKeypressActionSchema = z.object({\n  type: z.literal(\"keypress\"),\n  keys: z.array(z.string()),\n});\n\nconst ComputerUseMoveActionSchema = z.object({\n  type: z.literal(\"move\"),\n  x: z.number(),\n  y: z.number(),\n});\n\nconst ComputerUseScrollActionSchema = z.object({\n  type: z.literal(\"scroll\"),\n  x: z.number(),\n  y: z.number(),\n  scroll_x: z.number(),\n  scroll_y: z.number(),\n});\n\nconst ComputerUseTypeActionSchema = z.object({\n  type: z.literal(\"type\"),\n  text: z.string(),\n});\n\nconst ComputerUseWaitActionSchema = z.object({\n  type: z.literal(\"wait\"),\n  duration: z.number().optional(),\n});\n\n// Union schema for individual action types\nconst ComputerUseActionUnionSchema = z.union([\n  ComputerUseScreenshotActionSchema,\n  ComputerUseClickActionSchema,\n  ComputerUseDoubleClickActionSchema,\n  ComputerUseDragActionSchema,\n  ComputerUseKeypressActionSchema,\n  ComputerUseMoveActionSchema,\n  ComputerUseScrollActionSchema,\n  ComputerUseTypeActionSchema,\n  ComputerUseWaitActionSchema,\n]);\n\n// Schema for the input structure received from parseComputerCall\n// The action is wrapped in an `action` property: { action: { type: 'screenshot' } }\nexport const ComputerUseActionSchema = z.object({\n  action: ComputerUseActionUnionSchema,\n});\n\n// TypeScript types derived from Zod schemas\nexport type ComputerUseScreenshotActionType = z.infer<\n  typeof ComputerUseScreenshotActionSchema\n>;\nexport type ComputerUseClickActionType = z.infer<\n  typeof ComputerUseClickActionSchema\n>;\nexport type ComputerUseDoubleClickActionType = z.infer<\n  typeof ComputerUseDoubleClickActionSchema\n>;\nexport type ComputerUseDragActionType = z.infer<\n  typeof ComputerUseDragActionSchema\n>;\nexport type ComputerUseKeypressActionType = z.infer<\n  typeof ComputerUseKeypressActionSchema\n>;\nexport type ComputerUseMoveActionType = z.infer<\n  typeof ComputerUseMoveActionSchema\n>;\nexport type ComputerUseScrollActionType = z.infer<\n  typeof ComputerUseScrollActionSchema\n>;\nexport type ComputerUseTypeActionType = z.infer<\n  typeof ComputerUseTypeActionSchema\n>;\nexport type ComputerUseWaitActionType = z.infer<\n  typeof ComputerUseWaitActionSchema\n>;\n\n/**\n * Input structure for the Computer Use tool.\n * The action is wrapped in an `action` property.\n */\nexport interface ComputerUseInput {\n  action: ComputerUseAction;\n}\n\nexport type ComputerUseReturnType =\n  | string\n  | Promise<string>\n  | ToolMessage<any>\n  | Promise<ToolMessage<any>>;\n\n/**\n * Options for the Computer Use tool.\n */\nexport interface ComputerUseOptions {\n  /**\n   * The width of the computer display in pixels.\n   */\n  displayWidth: number;\n\n  /**\n   * The height of the computer display in pixels.\n   */\n  displayHeight: number;\n\n  /**\n   * The type of computer environment to control.\n   * - `browser`: Browser automation (recommended for most use cases)\n   * - `mac`: macOS environment\n   * - `windows`: Windows environment\n   * - `linux`: Linux environment\n   * - `ubuntu`: Ubuntu environment\n   */\n  environment: ComputerUseEnvironment;\n\n  /**\n   * Execute function that handles computer action execution.\n   * This function receives the action input and should return a base64-encoded\n   * screenshot of the result.\n   */\n  execute: (\n    action: ComputerUseAction,\n    runtime: ToolRuntime<any, any>\n  ) => ComputerUseReturnType;\n}\n\n/**\n * OpenAI Computer Use tool type for the Responses API.\n */\nexport type ComputerUseTool = OpenAIClient.Responses.ComputerTool;\n\nconst TOOL_NAME = \"computer_use\";\n\n/**\n * Creates a Computer Use tool that allows models to control computer interfaces\n * and perform tasks by simulating mouse clicks, keyboard input, scrolling, and more.\n *\n * **Computer Use** is a practical application of OpenAI's Computer-Using Agent (CUA)\n * model (`computer-use-preview`), which combines vision capabilities with advanced\n * reasoning to simulate controlling computer interfaces.\n *\n * **How it works**:\n * The tool operates in a continuous loop:\n * 1. Model sends computer actions (click, type, scroll, etc.)\n * 2. Your code executes these actions in a controlled environment\n * 3. You capture a screenshot of the result\n * 4. Send the screenshot back to the model\n * 5. Repeat until the task is complete\n *\n * **Important**: Computer use is in beta and requires careful consideration:\n * - Use in sandboxed environments only\n * - Do not use for high-stakes or authenticated tasks\n * - Always implement human-in-the-loop for important decisions\n * - Handle safety checks appropriately\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-computer-use | OpenAI Computer Use Documentation}\n *\n * @param options - Configuration options for the Computer Use tool\n * @returns A Computer Use tool that can be passed to `bindTools`\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"computer-use-preview\" });\n *\n * // With execute callback for automatic action handling\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n *   execute: async (action) => {\n *     if (action.type === \"screenshot\") {\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"click\") {\n *       await page.mouse.click(action.x, action.y, { button: action.button });\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"type\") {\n *       await page.keyboard.type(action.text);\n *       return captureScreenshot();\n *     }\n *     // Handle other actions...\n *     return captureScreenshot();\n *   },\n * });\n *\n * const llmWithComputer = model.bindTools([computer]);\n * const response = await llmWithComputer.invoke(\n *   \"Check the latest news on bing.com\"\n * );\n * ```\n *\n * @example\n * ```typescript\n * // Without execute callback (manual action handling)\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n * });\n *\n * const response = await model.invoke(\"Check the news\", {\n *   tools: [computer],\n * });\n *\n * // Access the computer call from the response\n * const computerCall = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"computer_call\"\n * );\n * if (computerCall) {\n *   console.log(\"Action to execute:\", computerCall.action);\n *   // Execute the action manually, then send back a screenshot\n * }\n * ```\n *\n * @example\n * ```typescript\n * // For macOS desktop automation with Docker\n * const computer = tools.computerUse({\n *   displayWidth: 1920,\n *   displayHeight: 1080,\n *   environment: \"mac\",\n *   execute: async (action) => {\n *     if (action.type === \"click\") {\n *       await dockerExec(\n *         `DISPLAY=:99 xdotool mousemove ${action.x} ${action.y} click 1`,\n *         containerName\n *       );\n *     }\n *     // Capture screenshot from container\n *     return await getDockerScreenshot(containerName);\n *   },\n * });\n * ```\n *\n * @remarks\n * - Only available through the Responses API (not Chat Completions)\n * - Requires `computer-use-preview` model\n * - Actions include: click, double_click, drag, keypress, move, screenshot, scroll, type, wait\n * - Safety checks may be returned that require acknowledgment before proceeding\n * - Use `truncation: \"auto\"` parameter when making requests\n * - Recommended to use with `reasoning.summary` for debugging\n */\nexport function computerUse(options: ComputerUseOptions) {\n  const computerTool = tool(\n    async (\n      input: ComputerUseInput,\n      runtime: ToolRuntime<{ messages: BaseMessage[] }>\n    ) => {\n      /**\n       * get computer_use call id from runtime\n       */\n      const aiMessage = runtime.state?.messages.at(-1) as AIMessage | undefined;\n      const computerToolCall = aiMessage?.tool_calls?.find(\n        (tc) => tc.name === \"computer_use\"\n      );\n      const computerToolCallId = computerToolCall?.id;\n      if (!computerToolCallId) {\n        throw new Error(\"Computer use call id not found\");\n      }\n\n      const result = await options.execute(input.action, runtime);\n\n      /**\n       * make sure {@link ToolMessage} is returned with the correct additional kwargs\n       */\n      if (typeof result === \"string\") {\n        return new ToolMessage({\n          content: result,\n          tool_call_id: computerToolCallId,\n          additional_kwargs: {\n            type: \"computer_call_output\",\n          },\n        });\n      }\n\n      /**\n       * make sure {@link ToolMessage} is returned with the correct additional kwargs\n       */\n      return new ToolMessage({\n        ...result,\n        tool_call_id: computerToolCallId,\n        additional_kwargs: {\n          type: \"computer_call_output\",\n          ...result.additional_kwargs,\n        },\n      });\n    },\n    {\n      name: TOOL_NAME,\n      description:\n        \"Control a computer interface by executing mouse clicks, keyboard input, scrolling, and other actions.\",\n      schema: ComputerUseActionSchema,\n    }\n  );\n\n  computerTool.extras = {\n    ...(computerTool.extras ?? {}),\n    providerToolDefinition: {\n      type: \"computer_use_preview\",\n      display_width: options.displayWidth,\n      display_height: options.displayHeight,\n      environment: options.environment,\n    } satisfies ComputerUseTool,\n  };\n\n  /**\n   * return as typed {@link DynamicStructuredTool} so we don't get any type\n   * errors like \"can't export tool without reference\"\n   */\n  return computerTool as DynamicStructuredTool<\n    typeof ComputerUseActionSchema,\n    ComputerUseInput,\n    unknown,\n    ComputerUseReturnType\n  >;\n}\n"],"mappings":";;;;;;AAkDA,MAAM,oCAAoCA,SAAE,OAAO,EACjD,MAAMA,SAAE,QAAQ,aAAa,CAC9B,EAAC;AAEF,MAAM,+BAA+BA,SAAE,OAAO;CAC5C,MAAMA,SAAE,QAAQ,QAAQ;CACxB,GAAGA,SAAE,QAAQ;CACb,GAAGA,SAAE,QAAQ;CACb,QAAQA,SAAE,KAAK;EAAC;EAAQ;EAAS;EAAS;EAAQ;CAAU,EAAC,CAAC,QAAQ,OAAO;AAC9E,EAAC;AAEF,MAAM,qCAAqCA,SAAE,OAAO;CAClD,MAAMA,SAAE,QAAQ,eAAe;CAC/B,GAAGA,SAAE,QAAQ;CACb,GAAGA,SAAE,QAAQ;CACb,QAAQA,SAAE,KAAK;EAAC;EAAQ;EAAS;EAAS;EAAQ;CAAU,EAAC,CAAC,QAAQ,OAAO;AAC9E,EAAC;AAEF,MAAM,8BAA8BA,SAAE,OAAO;CAC3C,MAAMA,SAAE,QAAQ,OAAO;CACvB,MAAMA,SAAE,MAAMA,SAAE,OAAO;EAAE,GAAGA,SAAE,QAAQ;EAAE,GAAGA,SAAE,QAAQ;CAAE,EAAC,CAAC;AAC1D,EAAC;AAEF,MAAM,kCAAkCA,SAAE,OAAO;CAC/C,MAAMA,SAAE,QAAQ,WAAW;CAC3B,MAAMA,SAAE,MAAMA,SAAE,QAAQ,CAAC;AAC1B,EAAC;AAEF,MAAM,8BAA8BA,SAAE,OAAO;CAC3C,MAAMA,SAAE,QAAQ,OAAO;CACvB,GAAGA,SAAE,QAAQ;CACb,GAAGA,SAAE,QAAQ;AACd,EAAC;AAEF,MAAM,gCAAgCA,SAAE,OAAO;CAC7C,MAAMA,SAAE,QAAQ,SAAS;CACzB,GAAGA,SAAE,QAAQ;CACb,GAAGA,SAAE,QAAQ;CACb,UAAUA,SAAE,QAAQ;CACpB,UAAUA,SAAE,QAAQ;AACrB,EAAC;AAEF,MAAM,8BAA8BA,SAAE,OAAO;CAC3C,MAAMA,SAAE,QAAQ,OAAO;CACvB,MAAMA,SAAE,QAAQ;AACjB,EAAC;AAEF,MAAM,8BAA8BA,SAAE,OAAO;CAC3C,MAAMA,SAAE,QAAQ,OAAO;CACvB,UAAUA,SAAE,QAAQ,CAAC,UAAU;AAChC,EAAC;AAGF,MAAM,+BAA+BA,SAAE,MAAM;CAC3C;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACD,EAAC;AAIF,MAAa,0BAA0BA,SAAE,OAAO,EAC9C,QAAQ,6BACT,EAAC;AAqFF,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkHlB,SAAgB,YAAYC,SAA6B;CACvD,MAAM,gDACJ,OACEC,OACAC,YACG;;;;EAIH,MAAM,YAAY,QAAQ,OAAO,SAAS,GAAG,GAAG;EAChD,MAAM,mBAAmB,WAAW,YAAY,KAC9C,CAAC,OAAO,GAAG,SAAS,eACrB;EACD,MAAM,qBAAqB,kBAAkB;AAC7C,MAAI,CAAC,mBACH,OAAM,IAAI,MAAM;EAGlB,MAAM,SAAS,MAAM,QAAQ,QAAQ,MAAM,QAAQ,QAAQ;;;;AAK3D,MAAI,OAAO,WAAW,SACpB,QAAO,IAAIC,sCAAY;GACrB,SAAS;GACT,cAAc;GACd,mBAAmB,EACjB,MAAM,uBACP;EACF;;;;AAMH,SAAO,IAAIA,sCAAY;GACrB,GAAG;GACH,cAAc;GACd,mBAAmB;IACjB,MAAM;IACN,GAAG,OAAO;GACX;EACF;CACF,GACD;EACE,MAAM;EACN,aACE;EACF,QAAQ;CACT,EACF;CAED,aAAa,SAAS;EACpB,GAAI,aAAa,UAAU,CAAE;EAC7B,wBAAwB;GACtB,MAAM;GACN,eAAe,QAAQ;GACvB,gBAAgB,QAAQ;GACxB,aAAa,QAAQ;EACtB;CACF;;;;;AAMD,QAAO;AAMR"}

package/dist/tools/computerUse.d.cts

@@ -178,7 +178,7 @@ type ComputerUseTool = OpenAI$1.Responses.ComputerTool;
  * - Recommended to use with `reasoning.summary` for debugging
  */
 declare function computerUse(options: ComputerUseOptions): DynamicStructuredTool<z.ZodObject<{
-  action: z.ZodDiscriminatedUnion<[z.ZodObject<{
+  action: z.ZodUnion<readonly [z.ZodObject<{
     type: z.ZodLiteral<"screenshot">;
   }, z.core.$strip>, z.ZodObject<{
     type: z.ZodLiteral<"click">;

package/dist/tools/computerUse.d.cts.map

@@ -1 +1 @@
-{"version":3,"file":"computerUse.d.cts","names":["z","OpenAI","OpenAIClient","DynamicStructuredTool","ToolRuntime","ToolMessage","ComputerUseEnvironment","ComputerUseClickAction","Responses","ResponseComputerToolCall","Click","ComputerUseDoubleClickAction","DoubleClick","ComputerUseDragAction","Drag","ComputerUseKeypressAction","Keypress","ComputerUseMoveAction","Move","ComputerUseScreenshotAction","Screenshot","ComputerUseScrollAction","Scroll","ComputerUseTypeAction","Type","ComputerUseWaitAction","Wait","ComputerUseAction","ComputerUseScreenshotActionSchema","ZodLiteral","core","$strip","ZodObject","ComputerUseClickActionSchema","ZodNumber","ZodEnum","ZodDefault","ComputerUseDoubleClickActionSchema","ComputerUseDragActionSchema","ZodArray","ComputerUseKeypressActionSchema","ZodString","ComputerUseMoveActionSchema","ComputerUseScrollActionSchema","ComputerUseTypeActionSchema","ComputerUseWaitActionSchema","ZodOptional","ComputerUseActionSchema","ZodDiscriminatedUnion","ComputerUseScreenshotActionType","infer","ComputerUseClickActionType","ComputerUseDoubleClickActionType","ComputerUseDragActionType","ComputerUseKeypressActionType","ComputerUseMoveActionType","ComputerUseScrollActionType","ComputerUseTypeActionType","ComputerUseWaitActionType","ComputerUseInput","ComputerUseReturnType","Promise","ComputerUseOptions","ComputerUseTool","ComputerTool","computerUse"],"sources":["../../src/tools/computerUse.d.ts"],"sourcesContent":["import { z } from \"zod/v4\";\nimport { OpenAI as OpenAIClient } from \"openai\";\nimport { type DynamicStructuredTool } from \"@langchain/core/tools\";\nimport { type ToolRuntime } from \"@langchain/core/tools\";\nimport { ToolMessage } from \"@langchain/core/messages\";\n/**\n * The type of computer environment to control.\n */\nexport type ComputerUseEnvironment = \"browser\" | \"mac\" | \"windows\" | \"linux\" | \"ubuntu\";\n/**\n * Re-export action types from OpenAI SDK for convenience.\n */\nexport type ComputerUseClickAction = OpenAIClient.Responses.ResponseComputerToolCall.Click;\nexport type ComputerUseDoubleClickAction = OpenAIClient.Responses.ResponseComputerToolCall.DoubleClick;\nexport type ComputerUseDragAction = OpenAIClient.Responses.ResponseComputerToolCall.Drag;\nexport type ComputerUseKeypressAction = OpenAIClient.Responses.ResponseComputerToolCall.Keypress;\nexport type ComputerUseMoveAction = OpenAIClient.Responses.ResponseComputerToolCall.Move;\nexport type ComputerUseScreenshotAction = OpenAIClient.Responses.ResponseComputerToolCall.Screenshot;\nexport type ComputerUseScrollAction = OpenAIClient.Responses.ResponseComputerToolCall.Scroll;\nexport type ComputerUseTypeAction = OpenAIClient.Responses.ResponseComputerToolCall.Type;\nexport type ComputerUseWaitAction = OpenAIClient.Responses.ResponseComputerToolCall.Wait;\n/**\n * Union type of all computer use actions from OpenAI SDK.\n */\nexport type ComputerUseAction = OpenAIClient.Responses.ResponseComputerToolCall[\"action\"];\ndeclare const ComputerUseScreenshotActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"screenshot\">;\n}, z.core.$strip>;\ndeclare const ComputerUseClickActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"click\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    button: z.ZodDefault<z.ZodEnum<{\n        back: \"back\";\n        forward: \"forward\";\n        left: \"left\";\n        right: \"right\";\n        wheel: \"wheel\";\n    }>>;\n}, z.core.$strip>;\ndeclare const ComputerUseDoubleClickActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"double_click\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    button: z.ZodDefault<z.ZodEnum<{\n        back: \"back\";\n        forward: \"forward\";\n        left: \"left\";\n        right: \"right\";\n        wheel: \"wheel\";\n    }>>;\n}, z.core.$strip>;\ndeclare const ComputerUseDragActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"drag\">;\n    path: z.ZodArray<z.ZodObject<{\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>>;\n}, z.core.$strip>;\ndeclare const ComputerUseKeypressActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"keypress\">;\n    keys: z.ZodArray<z.ZodString>;\n}, z.core.$strip>;\ndeclare const ComputerUseMoveActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"move\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n}, z.core.$strip>;\ndeclare const ComputerUseScrollActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"scroll\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    scroll_x: z.ZodNumber;\n    scroll_y: z.ZodNumber;\n}, z.core.$strip>;\ndeclare const ComputerUseTypeActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"type\">;\n    text: z.ZodString;\n}, z.core.$strip>;\ndeclare const ComputerUseWaitActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"wait\">;\n    duration: z.ZodOptional<z.ZodNumber>;\n}, z.core.$strip>;\nexport declare const ComputerUseActionSchema: z.ZodObject<{\n    action: z.ZodDiscriminatedUnion<[z.ZodObject<{\n        type: z.ZodLiteral<\"screenshot\">;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"double_click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"drag\">;\n        path: z.ZodArray<z.ZodObject<{\n            x: z.ZodNumber;\n            y: z.ZodNumber;\n        }, z.core.$strip>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"keypress\">;\n        keys: z.ZodArray<z.ZodString>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"move\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"scroll\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        scroll_x: z.ZodNumber;\n        scroll_y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"type\">;\n        text: z.ZodString;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"wait\">;\n        duration: z.ZodOptional<z.ZodNumber>;\n    }, z.core.$strip>]>;\n}, z.core.$strip>;\nexport type ComputerUseScreenshotActionType = z.infer<typeof ComputerUseScreenshotActionSchema>;\nexport type ComputerUseClickActionType = z.infer<typeof ComputerUseClickActionSchema>;\nexport type ComputerUseDoubleClickActionType = z.infer<typeof ComputerUseDoubleClickActionSchema>;\nexport type ComputerUseDragActionType = z.infer<typeof ComputerUseDragActionSchema>;\nexport type ComputerUseKeypressActionType = z.infer<typeof ComputerUseKeypressActionSchema>;\nexport type ComputerUseMoveActionType = z.infer<typeof ComputerUseMoveActionSchema>;\nexport type ComputerUseScrollActionType = z.infer<typeof ComputerUseScrollActionSchema>;\nexport type ComputerUseTypeActionType = z.infer<typeof ComputerUseTypeActionSchema>;\nexport type ComputerUseWaitActionType = z.infer<typeof ComputerUseWaitActionSchema>;\n/**\n * Input structure for the Computer Use tool.\n * The action is wrapped in an `action` property.\n */\nexport interface ComputerUseInput {\n    action: ComputerUseAction;\n}\nexport type ComputerUseReturnType = string | Promise<string> | ToolMessage<any> | Promise<ToolMessage<any>>;\n/**\n * Options for the Computer Use tool.\n */\nexport interface ComputerUseOptions {\n    /**\n     * The width of the computer display in pixels.\n     */\n    displayWidth: number;\n    /**\n     * The height of the computer display in pixels.\n     */\n    displayHeight: number;\n    /**\n     * The type of computer environment to control.\n     * - `browser`: Browser automation (recommended for most use cases)\n     * - `mac`: macOS environment\n     * - `windows`: Windows environment\n     * - `linux`: Linux environment\n     * - `ubuntu`: Ubuntu environment\n     */\n    environment: ComputerUseEnvironment;\n    /**\n     * Execute function that handles computer action execution.\n     * This function receives the action input and should return a base64-encoded\n     * screenshot of the result.\n     */\n    execute: (action: ComputerUseAction, runtime: ToolRuntime<any, any>) => ComputerUseReturnType;\n}\n/**\n * OpenAI Computer Use tool type for the Responses API.\n */\nexport type ComputerUseTool = OpenAIClient.Responses.ComputerTool;\n/**\n * Creates a Computer Use tool that allows models to control computer interfaces\n * and perform tasks by simulating mouse clicks, keyboard input, scrolling, and more.\n *\n * **Computer Use** is a practical application of OpenAI's Computer-Using Agent (CUA)\n * model (`computer-use-preview`), which combines vision capabilities with advanced\n * reasoning to simulate controlling computer interfaces.\n *\n * **How it works**:\n * The tool operates in a continuous loop:\n * 1. Model sends computer actions (click, type, scroll, etc.)\n * 2. Your code executes these actions in a controlled environment\n * 3. You capture a screenshot of the result\n * 4. Send the screenshot back to the model\n * 5. Repeat until the task is complete\n *\n * **Important**: Computer use is in beta and requires careful consideration:\n * - Use in sandboxed environments only\n * - Do not use for high-stakes or authenticated tasks\n * - Always implement human-in-the-loop for important decisions\n * - Handle safety checks appropriately\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-computer-use | OpenAI Computer Use Documentation}\n *\n * @param options - Configuration options for the Computer Use tool\n * @returns A Computer Use tool that can be passed to `bindTools`\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"computer-use-preview\" });\n *\n * // With execute callback for automatic action handling\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n *   execute: async (action) => {\n *     if (action.type === \"screenshot\") {\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"click\") {\n *       await page.mouse.click(action.x, action.y, { button: action.button });\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"type\") {\n *       await page.keyboard.type(action.text);\n *       return captureScreenshot();\n *     }\n *     // Handle other actions...\n *     return captureScreenshot();\n *   },\n * });\n *\n * const llmWithComputer = model.bindTools([computer]);\n * const response = await llmWithComputer.invoke(\n *   \"Check the latest news on bing.com\"\n * );\n * ```\n *\n * @example\n * ```typescript\n * // Without execute callback (manual action handling)\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n * });\n *\n * const response = await model.invoke(\"Check the news\", {\n *   tools: [computer],\n * });\n *\n * // Access the computer call from the response\n * const computerCall = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"computer_call\"\n * );\n * if (computerCall) {\n *   console.log(\"Action to execute:\", computerCall.action);\n *   // Execute the action manually, then send back a screenshot\n * }\n * ```\n *\n * @example\n * ```typescript\n * // For macOS desktop automation with Docker\n * const computer = tools.computerUse({\n *   displayWidth: 1920,\n *   displayHeight: 1080,\n *   environment: \"mac\",\n *   execute: async (action) => {\n *     if (action.type === \"click\") {\n *       await dockerExec(\n *         `DISPLAY=:99 xdotool mousemove ${action.x} ${action.y} click 1`,\n *         containerName\n *       );\n *     }\n *     // Capture screenshot from container\n *     return await getDockerScreenshot(containerName);\n *   },\n * });\n * ```\n *\n * @remarks\n * - Only available through the Responses API (not Chat Completions)\n * - Requires `computer-use-preview` model\n * - Actions include: click, double_click, drag, keypress, move, screenshot, scroll, type, wait\n * - Safety checks may be returned that require acknowledgment before proceeding\n * - Use `truncation: \"auto\"` parameter when making requests\n * - Recommended to use with `reasoning.summary` for debugging\n */\nexport declare function computerUse(options: ComputerUseOptions): DynamicStructuredTool<z.ZodObject<{\n    action: z.ZodDiscriminatedUnion<[z.ZodObject<{\n        type: z.ZodLiteral<\"screenshot\">;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"double_click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"drag\">;\n        path: z.ZodArray<z.ZodObject<{\n            x: z.ZodNumber;\n            y: z.ZodNumber;\n        }, z.core.$strip>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"keypress\">;\n        keys: z.ZodArray<z.ZodString>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"move\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"scroll\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        scroll_x: z.ZodNumber;\n        scroll_y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"type\">;\n        text: z.ZodString;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"wait\">;\n        duration: z.ZodOptional<z.ZodNumber>;\n    }, z.core.$strip>]>;\n}, z.core.$strip>, ComputerUseInput, unknown, ComputerUseReturnType, string>;\nexport {};\n//# sourceMappingURL=computerUse.d.ts.map"],"mappings":";;;;;;;;AAQA;AAIA;AACYW,KALAL,sBAAAA,GAK4B,SAAA,GAAGJ,KAAAA,GAAaM,SAAAA,GAAUC,OAAAA,GAAAA,QAAAA;AAClE;AACA;AACA;AACYU,KALAZ,sBAAAA,GAAyBL,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBC,KAKpBD;AACrDY,KALAV,4BAAAA,GAA+BT,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBG,WAKC;AAChFW,KALAV,qBAAAA,GAAwBX,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBK,IAKAU;AACxEC,KALAV,yBAAAA,GAA4Bb,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBO,QAKA;AAI5EW,KARAV,qBAAAA,GAAwBf,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBS,IAQL;AA4H9DyC,KAnILxC,2BAAAA,GAA8BjB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBW,UAoI7D;AAEjBwC,KArIAvC,uBAAAA,GAA0BnB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBa,MAqIrD;AAAYuC,KApIjCtC,qBAAAA,GAAwBrB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBe,IAoIvCqC;AAAkBxD,KAnInDoB,qBAAAA,GAAwBvB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBiB,IAmIrBrB;;;AAA0B;AAIxEyD,KAnILnC,iBAAAA,GAAoBzB,QAAAA,CAAaM,SAAAA,CAAUC,wBAmIpB,CAAA,QAAA,CAAA;;;;;UAPlBkD,gBAAAA;UACLhC;;KAEAiC,qBAAAA,YAAiCC,kBAAkBxD,mBAAmBwD,QAAQxD;;;;UAIzEyD,kBAAAA;;;;;;;;;;;;;;;;;eAiBAxD;;;;;;oBAMKqB,4BAA4BvB,0BAA0BwD;;;;;KAKhEG,eAAAA,GAAkB7D,QAAAA,CAAaM,SAAAA,CAAUwD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAiH7BC,WAAAA,UAAqBH,qBAAqB3D,sBAAsBH,CAAAA,CAAEgC;UAC9EhC,CAAAA,CAAEgD,uBAAuBhD,CAAAA,CAAEgC;UACzBhC,CAAAA,CAAE6B;KACT7B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;YACGlC,CAAAA,CAAEoC,WAAWpC,CAAAA,CAAEmC;;;;;;;KAOxBnC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;YACGlC,CAAAA,CAAEoC,WAAWpC,CAAAA,CAAEmC;;;;;;;KAOxBnC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEuC,SAASvC,CAAAA,CAAEgC;SACZhC,CAAAA,CAAEkC;SACFlC,CAAAA,CAAEkC;OACNlC,CAAAA,CAAE8B,IAAAA,CAAKC;KACX/B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEuC,SAASvC,CAAAA,CAAEyC;KACpBzC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;KACNlC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;cACKlC,CAAAA,CAAEkC;cACFlC,CAAAA,CAAEkC;KACblC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEyC;KACTzC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;cACE7B,CAAAA,CAAE8C,YAAY9C,CAAAA,CAAEkC;KAC3BlC,CAAAA,CAAE8B,IAAAA,CAAKC;GACX/B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS4B,2BAA2BC"}
+{"version":3,"file":"computerUse.d.cts","names":["z","OpenAI","OpenAIClient","DynamicStructuredTool","ToolRuntime","ToolMessage","ComputerUseEnvironment","ComputerUseClickAction","Responses","ResponseComputerToolCall","Click","ComputerUseDoubleClickAction","DoubleClick","ComputerUseDragAction","Drag","ComputerUseKeypressAction","Keypress","ComputerUseMoveAction","Move","ComputerUseScreenshotAction","Screenshot","ComputerUseScrollAction","Scroll","ComputerUseTypeAction","Type","ComputerUseWaitAction","Wait","ComputerUseAction","ComputerUseScreenshotActionSchema","ZodLiteral","core","$strip","ZodObject","ComputerUseClickActionSchema","ZodNumber","ZodEnum","ZodDefault","ComputerUseDoubleClickActionSchema","ComputerUseDragActionSchema","ZodArray","ComputerUseKeypressActionSchema","ZodString","ComputerUseMoveActionSchema","ComputerUseScrollActionSchema","ComputerUseTypeActionSchema","ComputerUseWaitActionSchema","ZodOptional","ComputerUseActionSchema","ZodUnion","ComputerUseScreenshotActionType","infer","ComputerUseClickActionType","ComputerUseDoubleClickActionType","ComputerUseDragActionType","ComputerUseKeypressActionType","ComputerUseMoveActionType","ComputerUseScrollActionType","ComputerUseTypeActionType","ComputerUseWaitActionType","ComputerUseInput","ComputerUseReturnType","Promise","ComputerUseOptions","ComputerUseTool","ComputerTool","computerUse"],"sources":["../../src/tools/computerUse.d.ts"],"sourcesContent":["import { z } from \"zod/v4\";\nimport { OpenAI as OpenAIClient } from \"openai\";\nimport { type DynamicStructuredTool } from \"@langchain/core/tools\";\nimport { type ToolRuntime } from \"@langchain/core/tools\";\nimport { ToolMessage } from \"@langchain/core/messages\";\n/**\n * The type of computer environment to control.\n */\nexport type ComputerUseEnvironment = \"browser\" | \"mac\" | \"windows\" | \"linux\" | \"ubuntu\";\n/**\n * Re-export action types from OpenAI SDK for convenience.\n */\nexport type ComputerUseClickAction = OpenAIClient.Responses.ResponseComputerToolCall.Click;\nexport type ComputerUseDoubleClickAction = OpenAIClient.Responses.ResponseComputerToolCall.DoubleClick;\nexport type ComputerUseDragAction = OpenAIClient.Responses.ResponseComputerToolCall.Drag;\nexport type ComputerUseKeypressAction = OpenAIClient.Responses.ResponseComputerToolCall.Keypress;\nexport type ComputerUseMoveAction = OpenAIClient.Responses.ResponseComputerToolCall.Move;\nexport type ComputerUseScreenshotAction = OpenAIClient.Responses.ResponseComputerToolCall.Screenshot;\nexport type ComputerUseScrollAction = OpenAIClient.Responses.ResponseComputerToolCall.Scroll;\nexport type ComputerUseTypeAction = OpenAIClient.Responses.ResponseComputerToolCall.Type;\nexport type ComputerUseWaitAction = OpenAIClient.Responses.ResponseComputerToolCall.Wait;\n/**\n * Union type of all computer use actions from OpenAI SDK.\n */\nexport type ComputerUseAction = OpenAIClient.Responses.ResponseComputerToolCall[\"action\"];\ndeclare const ComputerUseScreenshotActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"screenshot\">;\n}, z.core.$strip>;\ndeclare const ComputerUseClickActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"click\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    button: z.ZodDefault<z.ZodEnum<{\n        back: \"back\";\n        forward: \"forward\";\n        left: \"left\";\n        right: \"right\";\n        wheel: \"wheel\";\n    }>>;\n}, z.core.$strip>;\ndeclare const ComputerUseDoubleClickActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"double_click\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    button: z.ZodDefault<z.ZodEnum<{\n        back: \"back\";\n        forward: \"forward\";\n        left: \"left\";\n        right: \"right\";\n        wheel: \"wheel\";\n    }>>;\n}, z.core.$strip>;\ndeclare const ComputerUseDragActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"drag\">;\n    path: z.ZodArray<z.ZodObject<{\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>>;\n}, z.core.$strip>;\ndeclare const ComputerUseKeypressActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"keypress\">;\n    keys: z.ZodArray<z.ZodString>;\n}, z.core.$strip>;\ndeclare const ComputerUseMoveActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"move\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n}, z.core.$strip>;\ndeclare const ComputerUseScrollActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"scroll\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    scroll_x: z.ZodNumber;\n    scroll_y: z.ZodNumber;\n}, z.core.$strip>;\ndeclare const ComputerUseTypeActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"type\">;\n    text: z.ZodString;\n}, z.core.$strip>;\ndeclare const ComputerUseWaitActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"wait\">;\n    duration: z.ZodOptional<z.ZodNumber>;\n}, z.core.$strip>;\nexport declare const ComputerUseActionSchema: z.ZodObject<{\n    action: z.ZodUnion<readonly [z.ZodObject<{\n        type: z.ZodLiteral<\"screenshot\">;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"double_click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"drag\">;\n        path: z.ZodArray<z.ZodObject<{\n            x: z.ZodNumber;\n            y: z.ZodNumber;\n        }, z.core.$strip>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"keypress\">;\n        keys: z.ZodArray<z.ZodString>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"move\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"scroll\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        scroll_x: z.ZodNumber;\n        scroll_y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"type\">;\n        text: z.ZodString;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"wait\">;\n        duration: z.ZodOptional<z.ZodNumber>;\n    }, z.core.$strip>]>;\n}, z.core.$strip>;\nexport type ComputerUseScreenshotActionType = z.infer<typeof ComputerUseScreenshotActionSchema>;\nexport type ComputerUseClickActionType = z.infer<typeof ComputerUseClickActionSchema>;\nexport type ComputerUseDoubleClickActionType = z.infer<typeof ComputerUseDoubleClickActionSchema>;\nexport type ComputerUseDragActionType = z.infer<typeof ComputerUseDragActionSchema>;\nexport type ComputerUseKeypressActionType = z.infer<typeof ComputerUseKeypressActionSchema>;\nexport type ComputerUseMoveActionType = z.infer<typeof ComputerUseMoveActionSchema>;\nexport type ComputerUseScrollActionType = z.infer<typeof ComputerUseScrollActionSchema>;\nexport type ComputerUseTypeActionType = z.infer<typeof ComputerUseTypeActionSchema>;\nexport type ComputerUseWaitActionType = z.infer<typeof ComputerUseWaitActionSchema>;\n/**\n * Input structure for the Computer Use tool.\n * The action is wrapped in an `action` property.\n */\nexport interface ComputerUseInput {\n    action: ComputerUseAction;\n}\nexport type ComputerUseReturnType = string | Promise<string> | ToolMessage<any> | Promise<ToolMessage<any>>;\n/**\n * Options for the Computer Use tool.\n */\nexport interface ComputerUseOptions {\n    /**\n     * The width of the computer display in pixels.\n     */\n    displayWidth: number;\n    /**\n     * The height of the computer display in pixels.\n     */\n    displayHeight: number;\n    /**\n     * The type of computer environment to control.\n     * - `browser`: Browser automation (recommended for most use cases)\n     * - `mac`: macOS environment\n     * - `windows`: Windows environment\n     * - `linux`: Linux environment\n     * - `ubuntu`: Ubuntu environment\n     */\n    environment: ComputerUseEnvironment;\n    /**\n     * Execute function that handles computer action execution.\n     * This function receives the action input and should return a base64-encoded\n     * screenshot of the result.\n     */\n    execute: (action: ComputerUseAction, runtime: ToolRuntime<any, any>) => ComputerUseReturnType;\n}\n/**\n * OpenAI Computer Use tool type for the Responses API.\n */\nexport type ComputerUseTool = OpenAIClient.Responses.ComputerTool;\n/**\n * Creates a Computer Use tool that allows models to control computer interfaces\n * and perform tasks by simulating mouse clicks, keyboard input, scrolling, and more.\n *\n * **Computer Use** is a practical application of OpenAI's Computer-Using Agent (CUA)\n * model (`computer-use-preview`), which combines vision capabilities with advanced\n * reasoning to simulate controlling computer interfaces.\n *\n * **How it works**:\n * The tool operates in a continuous loop:\n * 1. Model sends computer actions (click, type, scroll, etc.)\n * 2. Your code executes these actions in a controlled environment\n * 3. You capture a screenshot of the result\n * 4. Send the screenshot back to the model\n * 5. Repeat until the task is complete\n *\n * **Important**: Computer use is in beta and requires careful consideration:\n * - Use in sandboxed environments only\n * - Do not use for high-stakes or authenticated tasks\n * - Always implement human-in-the-loop for important decisions\n * - Handle safety checks appropriately\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-computer-use | OpenAI Computer Use Documentation}\n *\n * @param options - Configuration options for the Computer Use tool\n * @returns A Computer Use tool that can be passed to `bindTools`\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"computer-use-preview\" });\n *\n * // With execute callback for automatic action handling\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n *   execute: async (action) => {\n *     if (action.type === \"screenshot\") {\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"click\") {\n *       await page.mouse.click(action.x, action.y, { button: action.button });\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"type\") {\n *       await page.keyboard.type(action.text);\n *       return captureScreenshot();\n *     }\n *     // Handle other actions...\n *     return captureScreenshot();\n *   },\n * });\n *\n * const llmWithComputer = model.bindTools([computer]);\n * const response = await llmWithComputer.invoke(\n *   \"Check the latest news on bing.com\"\n * );\n * ```\n *\n * @example\n * ```typescript\n * // Without execute callback (manual action handling)\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n * });\n *\n * const response = await model.invoke(\"Check the news\", {\n *   tools: [computer],\n * });\n *\n * // Access the computer call from the response\n * const computerCall = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"computer_call\"\n * );\n * if (computerCall) {\n *   console.log(\"Action to execute:\", computerCall.action);\n *   // Execute the action manually, then send back a screenshot\n * }\n * ```\n *\n * @example\n * ```typescript\n * // For macOS desktop automation with Docker\n * const computer = tools.computerUse({\n *   displayWidth: 1920,\n *   displayHeight: 1080,\n *   environment: \"mac\",\n *   execute: async (action) => {\n *     if (action.type === \"click\") {\n *       await dockerExec(\n *         `DISPLAY=:99 xdotool mousemove ${action.x} ${action.y} click 1`,\n *         containerName\n *       );\n *     }\n *     // Capture screenshot from container\n *     return await getDockerScreenshot(containerName);\n *   },\n * });\n * ```\n *\n * @remarks\n * - Only available through the Responses API (not Chat Completions)\n * - Requires `computer-use-preview` model\n * - Actions include: click, double_click, drag, keypress, move, screenshot, scroll, type, wait\n * - Safety checks may be returned that require acknowledgment before proceeding\n * - Use `truncation: \"auto\"` parameter when making requests\n * - Recommended to use with `reasoning.summary` for debugging\n */\nexport declare function computerUse(options: ComputerUseOptions): DynamicStructuredTool<z.ZodObject<{\n    action: z.ZodUnion<readonly [z.ZodObject<{\n        type: z.ZodLiteral<\"screenshot\">;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"double_click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"drag\">;\n        path: z.ZodArray<z.ZodObject<{\n            x: z.ZodNumber;\n            y: z.ZodNumber;\n        }, z.core.$strip>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"keypress\">;\n        keys: z.ZodArray<z.ZodString>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"move\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"scroll\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        scroll_x: z.ZodNumber;\n        scroll_y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"type\">;\n        text: z.ZodString;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"wait\">;\n        duration: z.ZodOptional<z.ZodNumber>;\n    }, z.core.$strip>]>;\n}, z.core.$strip>, ComputerUseInput, unknown, ComputerUseReturnType, string>;\nexport {};\n//# sourceMappingURL=computerUse.d.ts.map"],"mappings":";;;;;;;;AAQA;AAIA;AACYW,KALAL,sBAAAA,GAK4B,SAAA,GAAGJ,KAAAA,GAAaM,SAAAA,GAAUC,OAAAA,GAAAA,QAAAA;AAClE;AACA;AACA;AACYU,KALAZ,sBAAAA,GAAyBL,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBC,KAKpBD;AACrDY,KALAV,4BAAAA,GAA+BT,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBG,WAKC;AAChFW,KALAV,qBAAAA,GAAwBX,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBK,IAKAU;AACxEC,KALAV,yBAAAA,GAA4Bb,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBO,QAKA;AAI5EW,KARAV,qBAAAA,GAAwBf,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBS,IAQL;AA4H9DyC,KAnILxC,2BAAAA,GAA8BjB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBW,UAoI7D;AAEjBwC,KArIAvC,uBAAAA,GAA0BnB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBa,MAqIrD;AAAYuC,KApIjCtC,qBAAAA,GAAwBrB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBe,IAoIvCqC;AAAkBxD,KAnInDoB,qBAAAA,GAAwBvB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBiB,IAmIrBrB;;;AAA0B;AAIxEyD,KAnILnC,iBAAAA,GAAoBzB,QAAAA,CAAaM,SAAAA,CAAUC,wBAmIpB,CAAA,QAAA,CAAA;;;;;UAPlBkD,gBAAAA;UACLhC;;KAEAiC,qBAAAA,YAAiCC,kBAAkBxD,mBAAmBwD,QAAQxD;;;;UAIzEyD,kBAAAA;;;;;;;;;;;;;;;;;eAiBAxD;;;;;;oBAMKqB,4BAA4BvB,0BAA0BwD;;;;;KAKhEG,eAAAA,GAAkB7D,QAAAA,CAAaM,SAAAA,CAAUwD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAiH7BC,WAAAA,UAAqBH,qBAAqB3D,sBAAsBH,CAAAA,CAAEgC;UAC9EhC,CAAAA,CAAEgD,mBAAmBhD,CAAAA,CAAEgC;UACrBhC,CAAAA,CAAE6B;KACT7B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;YACGlC,CAAAA,CAAEoC,WAAWpC,CAAAA,CAAEmC;;;;;;;KAOxBnC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;YACGlC,CAAAA,CAAEoC,WAAWpC,CAAAA,CAAEmC;;;;;;;KAOxBnC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEuC,SAASvC,CAAAA,CAAEgC;SACZhC,CAAAA,CAAEkC;SACFlC,CAAAA,CAAEkC;OACNlC,CAAAA,CAAE8B,IAAAA,CAAKC;KACX/B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEuC,SAASvC,CAAAA,CAAEyC;KACpBzC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;KACNlC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;cACKlC,CAAAA,CAAEkC;cACFlC,CAAAA,CAAEkC;KACblC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEyC;KACTzC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;cACE7B,CAAAA,CAAE8C,YAAY9C,CAAAA,CAAEkC;KAC3BlC,CAAAA,CAAE8B,IAAAA,CAAKC;GACX/B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS4B,2BAA2BC"}

package/dist/tools/computerUse.d.ts

@@ -178,7 +178,7 @@ type ComputerUseTool = OpenAI$1.Responses.ComputerTool;
  * - Recommended to use with `reasoning.summary` for debugging
  */
 declare function computerUse(options: ComputerUseOptions): DynamicStructuredTool<z.ZodObject<{
-  action: z.ZodDiscriminatedUnion<[z.ZodObject<{
+  action: z.ZodUnion<readonly [z.ZodObject<{
     type: z.ZodLiteral<"screenshot">;
   }, z.core.$strip>, z.ZodObject<{
     type: z.ZodLiteral<"click">;

package/dist/tools/computerUse.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"computerUse.d.ts","names":["z","OpenAI","OpenAIClient","DynamicStructuredTool","ToolRuntime","ToolMessage","ComputerUseEnvironment","ComputerUseClickAction","Responses","ResponseComputerToolCall","Click","ComputerUseDoubleClickAction","DoubleClick","ComputerUseDragAction","Drag","ComputerUseKeypressAction","Keypress","ComputerUseMoveAction","Move","ComputerUseScreenshotAction","Screenshot","ComputerUseScrollAction","Scroll","ComputerUseTypeAction","Type","ComputerUseWaitAction","Wait","ComputerUseAction","ComputerUseScreenshotActionSchema","ZodLiteral","core","$strip","ZodObject","ComputerUseClickActionSchema","ZodNumber","ZodEnum","ZodDefault","ComputerUseDoubleClickActionSchema","ComputerUseDragActionSchema","ZodArray","ComputerUseKeypressActionSchema","ZodString","ComputerUseMoveActionSchema","ComputerUseScrollActionSchema","ComputerUseTypeActionSchema","ComputerUseWaitActionSchema","ZodOptional","ComputerUseActionSchema","ZodDiscriminatedUnion","ComputerUseScreenshotActionType","infer","ComputerUseClickActionType","ComputerUseDoubleClickActionType","ComputerUseDragActionType","ComputerUseKeypressActionType","ComputerUseMoveActionType","ComputerUseScrollActionType","ComputerUseTypeActionType","ComputerUseWaitActionType","ComputerUseInput","ComputerUseReturnType","Promise","ComputerUseOptions","ComputerUseTool","ComputerTool","computerUse"],"sources":["../../src/tools/computerUse.d.ts"],"sourcesContent":["import { z } from \"zod/v4\";\nimport { OpenAI as OpenAIClient } from \"openai\";\nimport { type DynamicStructuredTool } from \"@langchain/core/tools\";\nimport { type ToolRuntime } from \"@langchain/core/tools\";\nimport { ToolMessage } from \"@langchain/core/messages\";\n/**\n * The type of computer environment to control.\n */\nexport type ComputerUseEnvironment = \"browser\" | \"mac\" | \"windows\" | \"linux\" | \"ubuntu\";\n/**\n * Re-export action types from OpenAI SDK for convenience.\n */\nexport type ComputerUseClickAction = OpenAIClient.Responses.ResponseComputerToolCall.Click;\nexport type ComputerUseDoubleClickAction = OpenAIClient.Responses.ResponseComputerToolCall.DoubleClick;\nexport type ComputerUseDragAction = OpenAIClient.Responses.ResponseComputerToolCall.Drag;\nexport type ComputerUseKeypressAction = OpenAIClient.Responses.ResponseComputerToolCall.Keypress;\nexport type ComputerUseMoveAction = OpenAIClient.Responses.ResponseComputerToolCall.Move;\nexport type ComputerUseScreenshotAction = OpenAIClient.Responses.ResponseComputerToolCall.Screenshot;\nexport type ComputerUseScrollAction = OpenAIClient.Responses.ResponseComputerToolCall.Scroll;\nexport type ComputerUseTypeAction = OpenAIClient.Responses.ResponseComputerToolCall.Type;\nexport type ComputerUseWaitAction = OpenAIClient.Responses.ResponseComputerToolCall.Wait;\n/**\n * Union type of all computer use actions from OpenAI SDK.\n */\nexport type ComputerUseAction = OpenAIClient.Responses.ResponseComputerToolCall[\"action\"];\ndeclare const ComputerUseScreenshotActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"screenshot\">;\n}, z.core.$strip>;\ndeclare const ComputerUseClickActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"click\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    button: z.ZodDefault<z.ZodEnum<{\n        back: \"back\";\n        forward: \"forward\";\n        left: \"left\";\n        right: \"right\";\n        wheel: \"wheel\";\n    }>>;\n}, z.core.$strip>;\ndeclare const ComputerUseDoubleClickActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"double_click\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    button: z.ZodDefault<z.ZodEnum<{\n        back: \"back\";\n        forward: \"forward\";\n        left: \"left\";\n        right: \"right\";\n        wheel: \"wheel\";\n    }>>;\n}, z.core.$strip>;\ndeclare const ComputerUseDragActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"drag\">;\n    path: z.ZodArray<z.ZodObject<{\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>>;\n}, z.core.$strip>;\ndeclare const ComputerUseKeypressActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"keypress\">;\n    keys: z.ZodArray<z.ZodString>;\n}, z.core.$strip>;\ndeclare const ComputerUseMoveActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"move\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n}, z.core.$strip>;\ndeclare const ComputerUseScrollActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"scroll\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    scroll_x: z.ZodNumber;\n    scroll_y: z.ZodNumber;\n}, z.core.$strip>;\ndeclare const ComputerUseTypeActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"type\">;\n    text: z.ZodString;\n}, z.core.$strip>;\ndeclare const ComputerUseWaitActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"wait\">;\n    duration: z.ZodOptional<z.ZodNumber>;\n}, z.core.$strip>;\nexport declare const ComputerUseActionSchema: z.ZodObject<{\n    action: z.ZodDiscriminatedUnion<[z.ZodObject<{\n        type: z.ZodLiteral<\"screenshot\">;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"double_click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"drag\">;\n        path: z.ZodArray<z.ZodObject<{\n            x: z.ZodNumber;\n            y: z.ZodNumber;\n        }, z.core.$strip>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"keypress\">;\n        keys: z.ZodArray<z.ZodString>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"move\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"scroll\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        scroll_x: z.ZodNumber;\n        scroll_y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"type\">;\n        text: z.ZodString;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"wait\">;\n        duration: z.ZodOptional<z.ZodNumber>;\n    }, z.core.$strip>]>;\n}, z.core.$strip>;\nexport type ComputerUseScreenshotActionType = z.infer<typeof ComputerUseScreenshotActionSchema>;\nexport type ComputerUseClickActionType = z.infer<typeof ComputerUseClickActionSchema>;\nexport type ComputerUseDoubleClickActionType = z.infer<typeof ComputerUseDoubleClickActionSchema>;\nexport type ComputerUseDragActionType = z.infer<typeof ComputerUseDragActionSchema>;\nexport type ComputerUseKeypressActionType = z.infer<typeof ComputerUseKeypressActionSchema>;\nexport type ComputerUseMoveActionType = z.infer<typeof ComputerUseMoveActionSchema>;\nexport type ComputerUseScrollActionType = z.infer<typeof ComputerUseScrollActionSchema>;\nexport type ComputerUseTypeActionType = z.infer<typeof ComputerUseTypeActionSchema>;\nexport type ComputerUseWaitActionType = z.infer<typeof ComputerUseWaitActionSchema>;\n/**\n * Input structure for the Computer Use tool.\n * The action is wrapped in an `action` property.\n */\nexport interface ComputerUseInput {\n    action: ComputerUseAction;\n}\nexport type ComputerUseReturnType = string | Promise<string> | ToolMessage<any> | Promise<ToolMessage<any>>;\n/**\n * Options for the Computer Use tool.\n */\nexport interface ComputerUseOptions {\n    /**\n     * The width of the computer display in pixels.\n     */\n    displayWidth: number;\n    /**\n     * The height of the computer display in pixels.\n     */\n    displayHeight: number;\n    /**\n     * The type of computer environment to control.\n     * - `browser`: Browser automation (recommended for most use cases)\n     * - `mac`: macOS environment\n     * - `windows`: Windows environment\n     * - `linux`: Linux environment\n     * - `ubuntu`: Ubuntu environment\n     */\n    environment: ComputerUseEnvironment;\n    /**\n     * Execute function that handles computer action execution.\n     * This function receives the action input and should return a base64-encoded\n     * screenshot of the result.\n     */\n    execute: (action: ComputerUseAction, runtime: ToolRuntime<any, any>) => ComputerUseReturnType;\n}\n/**\n * OpenAI Computer Use tool type for the Responses API.\n */\nexport type ComputerUseTool = OpenAIClient.Responses.ComputerTool;\n/**\n * Creates a Computer Use tool that allows models to control computer interfaces\n * and perform tasks by simulating mouse clicks, keyboard input, scrolling, and more.\n *\n * **Computer Use** is a practical application of OpenAI's Computer-Using Agent (CUA)\n * model (`computer-use-preview`), which combines vision capabilities with advanced\n * reasoning to simulate controlling computer interfaces.\n *\n * **How it works**:\n * The tool operates in a continuous loop:\n * 1. Model sends computer actions (click, type, scroll, etc.)\n * 2. Your code executes these actions in a controlled environment\n * 3. You capture a screenshot of the result\n * 4. Send the screenshot back to the model\n * 5. Repeat until the task is complete\n *\n * **Important**: Computer use is in beta and requires careful consideration:\n * - Use in sandboxed environments only\n * - Do not use for high-stakes or authenticated tasks\n * - Always implement human-in-the-loop for important decisions\n * - Handle safety checks appropriately\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-computer-use | OpenAI Computer Use Documentation}\n *\n * @param options - Configuration options for the Computer Use tool\n * @returns A Computer Use tool that can be passed to `bindTools`\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"computer-use-preview\" });\n *\n * // With execute callback for automatic action handling\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n *   execute: async (action) => {\n *     if (action.type === \"screenshot\") {\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"click\") {\n *       await page.mouse.click(action.x, action.y, { button: action.button });\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"type\") {\n *       await page.keyboard.type(action.text);\n *       return captureScreenshot();\n *     }\n *     // Handle other actions...\n *     return captureScreenshot();\n *   },\n * });\n *\n * const llmWithComputer = model.bindTools([computer]);\n * const response = await llmWithComputer.invoke(\n *   \"Check the latest news on bing.com\"\n * );\n * ```\n *\n * @example\n * ```typescript\n * // Without execute callback (manual action handling)\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n * });\n *\n * const response = await model.invoke(\"Check the news\", {\n *   tools: [computer],\n * });\n *\n * // Access the computer call from the response\n * const computerCall = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"computer_call\"\n * );\n * if (computerCall) {\n *   console.log(\"Action to execute:\", computerCall.action);\n *   // Execute the action manually, then send back a screenshot\n * }\n * ```\n *\n * @example\n * ```typescript\n * // For macOS desktop automation with Docker\n * const computer = tools.computerUse({\n *   displayWidth: 1920,\n *   displayHeight: 1080,\n *   environment: \"mac\",\n *   execute: async (action) => {\n *     if (action.type === \"click\") {\n *       await dockerExec(\n *         `DISPLAY=:99 xdotool mousemove ${action.x} ${action.y} click 1`,\n *         containerName\n *       );\n *     }\n *     // Capture screenshot from container\n *     return await getDockerScreenshot(containerName);\n *   },\n * });\n * ```\n *\n * @remarks\n * - Only available through the Responses API (not Chat Completions)\n * - Requires `computer-use-preview` model\n * - Actions include: click, double_click, drag, keypress, move, screenshot, scroll, type, wait\n * - Safety checks may be returned that require acknowledgment before proceeding\n * - Use `truncation: \"auto\"` parameter when making requests\n * - Recommended to use with `reasoning.summary` for debugging\n */\nexport declare function computerUse(options: ComputerUseOptions): DynamicStructuredTool<z.ZodObject<{\n    action: z.ZodDiscriminatedUnion<[z.ZodObject<{\n        type: z.ZodLiteral<\"screenshot\">;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"double_click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"drag\">;\n        path: z.ZodArray<z.ZodObject<{\n            x: z.ZodNumber;\n            y: z.ZodNumber;\n        }, z.core.$strip>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"keypress\">;\n        keys: z.ZodArray<z.ZodString>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"move\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"scroll\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        scroll_x: z.ZodNumber;\n        scroll_y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"type\">;\n        text: z.ZodString;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"wait\">;\n        duration: z.ZodOptional<z.ZodNumber>;\n    }, z.core.$strip>]>;\n}, z.core.$strip>, ComputerUseInput, unknown, ComputerUseReturnType, string>;\nexport {};\n//# sourceMappingURL=computerUse.d.ts.map"],"mappings":";;;;;;;;AAQA;AAIA;AACYW,KALAL,sBAAAA,GAK4B,SAAA,GAAGJ,KAAAA,GAAaM,SAAAA,GAAUC,OAAAA,GAAAA,QAAAA;AAClE;AACA;AACA;AACYU,KALAZ,sBAAAA,GAAyBL,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBC,KAKpBD;AACrDY,KALAV,4BAAAA,GAA+BT,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBG,WAKC;AAChFW,KALAV,qBAAAA,GAAwBX,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBK,IAKAU;AACxEC,KALAV,yBAAAA,GAA4Bb,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBO,QAKA;AAI5EW,KARAV,qBAAAA,GAAwBf,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBS,IAQL;AA4H9DyC,KAnILxC,2BAAAA,GAA8BjB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBW,UAoI7D;AAEjBwC,KArIAvC,uBAAAA,GAA0BnB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBa,MAqIrD;AAAYuC,KApIjCtC,qBAAAA,GAAwBrB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBe,IAoIvCqC;AAAkBxD,KAnInDoB,qBAAAA,GAAwBvB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBiB,IAmIrBrB;;;AAA0B;AAIxEyD,KAnILnC,iBAAAA,GAAoBzB,QAAAA,CAAaM,SAAAA,CAAUC,wBAmIpB,CAAA,QAAA,CAAA;;;;;UAPlBkD,gBAAAA;UACLhC;;KAEAiC,qBAAAA,YAAiCC,kBAAkBxD,mBAAmBwD,QAAQxD;;;;UAIzEyD,kBAAAA;;;;;;;;;;;;;;;;;eAiBAxD;;;;;;oBAMKqB,4BAA4BvB,0BAA0BwD;;;;;KAKhEG,eAAAA,GAAkB7D,QAAAA,CAAaM,SAAAA,CAAUwD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAiH7BC,WAAAA,UAAqBH,qBAAqB3D,sBAAsBH,CAAAA,CAAEgC;UAC9EhC,CAAAA,CAAEgD,uBAAuBhD,CAAAA,CAAEgC;UACzBhC,CAAAA,CAAE6B;KACT7B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;YACGlC,CAAAA,CAAEoC,WAAWpC,CAAAA,CAAEmC;;;;;;;KAOxBnC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;YACGlC,CAAAA,CAAEoC,WAAWpC,CAAAA,CAAEmC;;;;;;;KAOxBnC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEuC,SAASvC,CAAAA,CAAEgC;SACZhC,CAAAA,CAAEkC;SACFlC,CAAAA,CAAEkC;OACNlC,CAAAA,CAAE8B,IAAAA,CAAKC;KACX/B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEuC,SAASvC,CAAAA,CAAEyC;KACpBzC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;KACNlC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;cACKlC,CAAAA,CAAEkC;cACFlC,CAAAA,CAAEkC;KACblC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEyC;KACTzC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;cACE7B,CAAAA,CAAE8C,YAAY9C,CAAAA,CAAEkC;KAC3BlC,CAAAA,CAAE8B,IAAAA,CAAKC;GACX/B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS4B,2BAA2BC"}
+{"version":3,"file":"computerUse.d.ts","names":["z","OpenAI","OpenAIClient","DynamicStructuredTool","ToolRuntime","ToolMessage","ComputerUseEnvironment","ComputerUseClickAction","Responses","ResponseComputerToolCall","Click","ComputerUseDoubleClickAction","DoubleClick","ComputerUseDragAction","Drag","ComputerUseKeypressAction","Keypress","ComputerUseMoveAction","Move","ComputerUseScreenshotAction","Screenshot","ComputerUseScrollAction","Scroll","ComputerUseTypeAction","Type","ComputerUseWaitAction","Wait","ComputerUseAction","ComputerUseScreenshotActionSchema","ZodLiteral","core","$strip","ZodObject","ComputerUseClickActionSchema","ZodNumber","ZodEnum","ZodDefault","ComputerUseDoubleClickActionSchema","ComputerUseDragActionSchema","ZodArray","ComputerUseKeypressActionSchema","ZodString","ComputerUseMoveActionSchema","ComputerUseScrollActionSchema","ComputerUseTypeActionSchema","ComputerUseWaitActionSchema","ZodOptional","ComputerUseActionSchema","ZodUnion","ComputerUseScreenshotActionType","infer","ComputerUseClickActionType","ComputerUseDoubleClickActionType","ComputerUseDragActionType","ComputerUseKeypressActionType","ComputerUseMoveActionType","ComputerUseScrollActionType","ComputerUseTypeActionType","ComputerUseWaitActionType","ComputerUseInput","ComputerUseReturnType","Promise","ComputerUseOptions","ComputerUseTool","ComputerTool","computerUse"],"sources":["../../src/tools/computerUse.d.ts"],"sourcesContent":["import { z } from \"zod/v4\";\nimport { OpenAI as OpenAIClient } from \"openai\";\nimport { type DynamicStructuredTool } from \"@langchain/core/tools\";\nimport { type ToolRuntime } from \"@langchain/core/tools\";\nimport { ToolMessage } from \"@langchain/core/messages\";\n/**\n * The type of computer environment to control.\n */\nexport type ComputerUseEnvironment = \"browser\" | \"mac\" | \"windows\" | \"linux\" | \"ubuntu\";\n/**\n * Re-export action types from OpenAI SDK for convenience.\n */\nexport type ComputerUseClickAction = OpenAIClient.Responses.ResponseComputerToolCall.Click;\nexport type ComputerUseDoubleClickAction = OpenAIClient.Responses.ResponseComputerToolCall.DoubleClick;\nexport type ComputerUseDragAction = OpenAIClient.Responses.ResponseComputerToolCall.Drag;\nexport type ComputerUseKeypressAction = OpenAIClient.Responses.ResponseComputerToolCall.Keypress;\nexport type ComputerUseMoveAction = OpenAIClient.Responses.ResponseComputerToolCall.Move;\nexport type ComputerUseScreenshotAction = OpenAIClient.Responses.ResponseComputerToolCall.Screenshot;\nexport type ComputerUseScrollAction = OpenAIClient.Responses.ResponseComputerToolCall.Scroll;\nexport type ComputerUseTypeAction = OpenAIClient.Responses.ResponseComputerToolCall.Type;\nexport type ComputerUseWaitAction = OpenAIClient.Responses.ResponseComputerToolCall.Wait;\n/**\n * Union type of all computer use actions from OpenAI SDK.\n */\nexport type ComputerUseAction = OpenAIClient.Responses.ResponseComputerToolCall[\"action\"];\ndeclare const ComputerUseScreenshotActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"screenshot\">;\n}, z.core.$strip>;\ndeclare const ComputerUseClickActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"click\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    button: z.ZodDefault<z.ZodEnum<{\n        back: \"back\";\n        forward: \"forward\";\n        left: \"left\";\n        right: \"right\";\n        wheel: \"wheel\";\n    }>>;\n}, z.core.$strip>;\ndeclare const ComputerUseDoubleClickActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"double_click\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    button: z.ZodDefault<z.ZodEnum<{\n        back: \"back\";\n        forward: \"forward\";\n        left: \"left\";\n        right: \"right\";\n        wheel: \"wheel\";\n    }>>;\n}, z.core.$strip>;\ndeclare const ComputerUseDragActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"drag\">;\n    path: z.ZodArray<z.ZodObject<{\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>>;\n}, z.core.$strip>;\ndeclare const ComputerUseKeypressActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"keypress\">;\n    keys: z.ZodArray<z.ZodString>;\n}, z.core.$strip>;\ndeclare const ComputerUseMoveActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"move\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n}, z.core.$strip>;\ndeclare const ComputerUseScrollActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"scroll\">;\n    x: z.ZodNumber;\n    y: z.ZodNumber;\n    scroll_x: z.ZodNumber;\n    scroll_y: z.ZodNumber;\n}, z.core.$strip>;\ndeclare const ComputerUseTypeActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"type\">;\n    text: z.ZodString;\n}, z.core.$strip>;\ndeclare const ComputerUseWaitActionSchema: z.ZodObject<{\n    type: z.ZodLiteral<\"wait\">;\n    duration: z.ZodOptional<z.ZodNumber>;\n}, z.core.$strip>;\nexport declare const ComputerUseActionSchema: z.ZodObject<{\n    action: z.ZodUnion<readonly [z.ZodObject<{\n        type: z.ZodLiteral<\"screenshot\">;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"double_click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"drag\">;\n        path: z.ZodArray<z.ZodObject<{\n            x: z.ZodNumber;\n            y: z.ZodNumber;\n        }, z.core.$strip>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"keypress\">;\n        keys: z.ZodArray<z.ZodString>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"move\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"scroll\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        scroll_x: z.ZodNumber;\n        scroll_y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"type\">;\n        text: z.ZodString;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"wait\">;\n        duration: z.ZodOptional<z.ZodNumber>;\n    }, z.core.$strip>]>;\n}, z.core.$strip>;\nexport type ComputerUseScreenshotActionType = z.infer<typeof ComputerUseScreenshotActionSchema>;\nexport type ComputerUseClickActionType = z.infer<typeof ComputerUseClickActionSchema>;\nexport type ComputerUseDoubleClickActionType = z.infer<typeof ComputerUseDoubleClickActionSchema>;\nexport type ComputerUseDragActionType = z.infer<typeof ComputerUseDragActionSchema>;\nexport type ComputerUseKeypressActionType = z.infer<typeof ComputerUseKeypressActionSchema>;\nexport type ComputerUseMoveActionType = z.infer<typeof ComputerUseMoveActionSchema>;\nexport type ComputerUseScrollActionType = z.infer<typeof ComputerUseScrollActionSchema>;\nexport type ComputerUseTypeActionType = z.infer<typeof ComputerUseTypeActionSchema>;\nexport type ComputerUseWaitActionType = z.infer<typeof ComputerUseWaitActionSchema>;\n/**\n * Input structure for the Computer Use tool.\n * The action is wrapped in an `action` property.\n */\nexport interface ComputerUseInput {\n    action: ComputerUseAction;\n}\nexport type ComputerUseReturnType = string | Promise<string> | ToolMessage<any> | Promise<ToolMessage<any>>;\n/**\n * Options for the Computer Use tool.\n */\nexport interface ComputerUseOptions {\n    /**\n     * The width of the computer display in pixels.\n     */\n    displayWidth: number;\n    /**\n     * The height of the computer display in pixels.\n     */\n    displayHeight: number;\n    /**\n     * The type of computer environment to control.\n     * - `browser`: Browser automation (recommended for most use cases)\n     * - `mac`: macOS environment\n     * - `windows`: Windows environment\n     * - `linux`: Linux environment\n     * - `ubuntu`: Ubuntu environment\n     */\n    environment: ComputerUseEnvironment;\n    /**\n     * Execute function that handles computer action execution.\n     * This function receives the action input and should return a base64-encoded\n     * screenshot of the result.\n     */\n    execute: (action: ComputerUseAction, runtime: ToolRuntime<any, any>) => ComputerUseReturnType;\n}\n/**\n * OpenAI Computer Use tool type for the Responses API.\n */\nexport type ComputerUseTool = OpenAIClient.Responses.ComputerTool;\n/**\n * Creates a Computer Use tool that allows models to control computer interfaces\n * and perform tasks by simulating mouse clicks, keyboard input, scrolling, and more.\n *\n * **Computer Use** is a practical application of OpenAI's Computer-Using Agent (CUA)\n * model (`computer-use-preview`), which combines vision capabilities with advanced\n * reasoning to simulate controlling computer interfaces.\n *\n * **How it works**:\n * The tool operates in a continuous loop:\n * 1. Model sends computer actions (click, type, scroll, etc.)\n * 2. Your code executes these actions in a controlled environment\n * 3. You capture a screenshot of the result\n * 4. Send the screenshot back to the model\n * 5. Repeat until the task is complete\n *\n * **Important**: Computer use is in beta and requires careful consideration:\n * - Use in sandboxed environments only\n * - Do not use for high-stakes or authenticated tasks\n * - Always implement human-in-the-loop for important decisions\n * - Handle safety checks appropriately\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-computer-use | OpenAI Computer Use Documentation}\n *\n * @param options - Configuration options for the Computer Use tool\n * @returns A Computer Use tool that can be passed to `bindTools`\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"computer-use-preview\" });\n *\n * // With execute callback for automatic action handling\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n *   execute: async (action) => {\n *     if (action.type === \"screenshot\") {\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"click\") {\n *       await page.mouse.click(action.x, action.y, { button: action.button });\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"type\") {\n *       await page.keyboard.type(action.text);\n *       return captureScreenshot();\n *     }\n *     // Handle other actions...\n *     return captureScreenshot();\n *   },\n * });\n *\n * const llmWithComputer = model.bindTools([computer]);\n * const response = await llmWithComputer.invoke(\n *   \"Check the latest news on bing.com\"\n * );\n * ```\n *\n * @example\n * ```typescript\n * // Without execute callback (manual action handling)\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n * });\n *\n * const response = await model.invoke(\"Check the news\", {\n *   tools: [computer],\n * });\n *\n * // Access the computer call from the response\n * const computerCall = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"computer_call\"\n * );\n * if (computerCall) {\n *   console.log(\"Action to execute:\", computerCall.action);\n *   // Execute the action manually, then send back a screenshot\n * }\n * ```\n *\n * @example\n * ```typescript\n * // For macOS desktop automation with Docker\n * const computer = tools.computerUse({\n *   displayWidth: 1920,\n *   displayHeight: 1080,\n *   environment: \"mac\",\n *   execute: async (action) => {\n *     if (action.type === \"click\") {\n *       await dockerExec(\n *         `DISPLAY=:99 xdotool mousemove ${action.x} ${action.y} click 1`,\n *         containerName\n *       );\n *     }\n *     // Capture screenshot from container\n *     return await getDockerScreenshot(containerName);\n *   },\n * });\n * ```\n *\n * @remarks\n * - Only available through the Responses API (not Chat Completions)\n * - Requires `computer-use-preview` model\n * - Actions include: click, double_click, drag, keypress, move, screenshot, scroll, type, wait\n * - Safety checks may be returned that require acknowledgment before proceeding\n * - Use `truncation: \"auto\"` parameter when making requests\n * - Recommended to use with `reasoning.summary` for debugging\n */\nexport declare function computerUse(options: ComputerUseOptions): DynamicStructuredTool<z.ZodObject<{\n    action: z.ZodUnion<readonly [z.ZodObject<{\n        type: z.ZodLiteral<\"screenshot\">;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"double_click\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        button: z.ZodDefault<z.ZodEnum<{\n            back: \"back\";\n            forward: \"forward\";\n            left: \"left\";\n            right: \"right\";\n            wheel: \"wheel\";\n        }>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"drag\">;\n        path: z.ZodArray<z.ZodObject<{\n            x: z.ZodNumber;\n            y: z.ZodNumber;\n        }, z.core.$strip>>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"keypress\">;\n        keys: z.ZodArray<z.ZodString>;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"move\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"scroll\">;\n        x: z.ZodNumber;\n        y: z.ZodNumber;\n        scroll_x: z.ZodNumber;\n        scroll_y: z.ZodNumber;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"type\">;\n        text: z.ZodString;\n    }, z.core.$strip>, z.ZodObject<{\n        type: z.ZodLiteral<\"wait\">;\n        duration: z.ZodOptional<z.ZodNumber>;\n    }, z.core.$strip>]>;\n}, z.core.$strip>, ComputerUseInput, unknown, ComputerUseReturnType, string>;\nexport {};\n//# sourceMappingURL=computerUse.d.ts.map"],"mappings":";;;;;;;;AAQA;AAIA;AACYW,KALAL,sBAAAA,GAK4B,SAAA,GAAGJ,KAAAA,GAAaM,SAAAA,GAAUC,OAAAA,GAAAA,QAAAA;AAClE;AACA;AACA;AACYU,KALAZ,sBAAAA,GAAyBL,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBC,KAKpBD;AACrDY,KALAV,4BAAAA,GAA+BT,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBG,WAKC;AAChFW,KALAV,qBAAAA,GAAwBX,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBK,IAKAU;AACxEC,KALAV,yBAAAA,GAA4Bb,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBO,QAKA;AAI5EW,KARAV,qBAAAA,GAAwBf,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBS,IAQL;AA4H9DyC,KAnILxC,2BAAAA,GAA8BjB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBW,UAoI7D;AAEjBwC,KArIAvC,uBAAAA,GAA0BnB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBa,MAqIrD;AAAYuC,KApIjCtC,qBAAAA,GAAwBrB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBe,IAoIvCqC;AAAkBxD,KAnInDoB,qBAAAA,GAAwBvB,QAAAA,CAAaM,SAAAA,CAAUC,wBAAAA,CAAyBiB,IAmIrBrB;;;AAA0B;AAIxEyD,KAnILnC,iBAAAA,GAAoBzB,QAAAA,CAAaM,SAAAA,CAAUC,wBAmIpB,CAAA,QAAA,CAAA;;;;;UAPlBkD,gBAAAA;UACLhC;;KAEAiC,qBAAAA,YAAiCC,kBAAkBxD,mBAAmBwD,QAAQxD;;;;UAIzEyD,kBAAAA;;;;;;;;;;;;;;;;;eAiBAxD;;;;;;oBAMKqB,4BAA4BvB,0BAA0BwD;;;;;KAKhEG,eAAAA,GAAkB7D,QAAAA,CAAaM,SAAAA,CAAUwD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAiH7BC,WAAAA,UAAqBH,qBAAqB3D,sBAAsBH,CAAAA,CAAEgC;UAC9EhC,CAAAA,CAAEgD,mBAAmBhD,CAAAA,CAAEgC;UACrBhC,CAAAA,CAAE6B;KACT7B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;YACGlC,CAAAA,CAAEoC,WAAWpC,CAAAA,CAAEmC;;;;;;;KAOxBnC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;YACGlC,CAAAA,CAAEoC,WAAWpC,CAAAA,CAAEmC;;;;;;;KAOxBnC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEuC,SAASvC,CAAAA,CAAEgC;SACZhC,CAAAA,CAAEkC;SACFlC,CAAAA,CAAEkC;OACNlC,CAAAA,CAAE8B,IAAAA,CAAKC;KACX/B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEuC,SAASvC,CAAAA,CAAEyC;KACpBzC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;KACNlC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;OACL7B,CAAAA,CAAEkC;OACFlC,CAAAA,CAAEkC;cACKlC,CAAAA,CAAEkC;cACFlC,CAAAA,CAAEkC;KACblC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;UACF7B,CAAAA,CAAEyC;KACTzC,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS/B,CAAAA,CAAEgC;UACXhC,CAAAA,CAAE6B;cACE7B,CAAAA,CAAE8C,YAAY9C,CAAAA,CAAEkC;KAC3BlC,CAAAA,CAAE8B,IAAAA,CAAKC;GACX/B,CAAAA,CAAE8B,IAAAA,CAAKC,SAAS4B,2BAA2BC"}

package/dist/tools/computerUse.js

@@ -59,7 +59,7 @@ const ComputerUseWaitActionSchema = z.object({
 	type: z.literal("wait"),
 	duration: z.number().optional()
 });
-const ComputerUseActionUnionSchema = z.discriminatedUnion("type", [
+const ComputerUseActionUnionSchema = z.union([
 	ComputerUseScreenshotActionSchema,
 	ComputerUseClickActionSchema,
 	ComputerUseDoubleClickActionSchema,

package/dist/tools/computerUse.js.map

@@ -1 +1 @@
-{"version":3,"file":"computerUse.js","names":["options: ComputerUseOptions","input: ComputerUseInput","runtime: ToolRuntime<{ messages: BaseMessage[] }>"],"sources":["../../src/tools/computerUse.ts"],"sourcesContent":["/* eslint-disable @typescript-eslint/no-explicit-any */\nimport { z } from \"zod/v4\";\nimport { OpenAI as OpenAIClient } from \"openai\";\nimport { tool, type DynamicStructuredTool } from \"@langchain/core/tools\";\nimport { type ToolRuntime } from \"@langchain/core/tools\";\nimport {\n  ToolMessage,\n  type AIMessage,\n  type BaseMessage,\n} from \"@langchain/core/messages\";\n\n/**\n * The type of computer environment to control.\n */\nexport type ComputerUseEnvironment =\n  | \"browser\"\n  | \"mac\"\n  | \"windows\"\n  | \"linux\"\n  | \"ubuntu\";\n\n/**\n * Re-export action types from OpenAI SDK for convenience.\n */\nexport type ComputerUseClickAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Click;\nexport type ComputerUseDoubleClickAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.DoubleClick;\nexport type ComputerUseDragAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Drag;\nexport type ComputerUseKeypressAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Keypress;\nexport type ComputerUseMoveAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Move;\nexport type ComputerUseScreenshotAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Screenshot;\nexport type ComputerUseScrollAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Scroll;\nexport type ComputerUseTypeAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Type;\nexport type ComputerUseWaitAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Wait;\n\n/**\n * Union type of all computer use actions from OpenAI SDK.\n */\nexport type ComputerUseAction =\n  OpenAIClient.Responses.ResponseComputerToolCall[\"action\"];\n\n// Zod schemas for computer use actions\nconst ComputerUseScreenshotActionSchema = z.object({\n  type: z.literal(\"screenshot\"),\n});\n\nconst ComputerUseClickActionSchema = z.object({\n  type: z.literal(\"click\"),\n  x: z.number(),\n  y: z.number(),\n  button: z.enum([\"left\", \"right\", \"wheel\", \"back\", \"forward\"]).default(\"left\"),\n});\n\nconst ComputerUseDoubleClickActionSchema = z.object({\n  type: z.literal(\"double_click\"),\n  x: z.number(),\n  y: z.number(),\n  button: z.enum([\"left\", \"right\", \"wheel\", \"back\", \"forward\"]).default(\"left\"),\n});\n\nconst ComputerUseDragActionSchema = z.object({\n  type: z.literal(\"drag\"),\n  path: z.array(z.object({ x: z.number(), y: z.number() })),\n});\n\nconst ComputerUseKeypressActionSchema = z.object({\n  type: z.literal(\"keypress\"),\n  keys: z.array(z.string()),\n});\n\nconst ComputerUseMoveActionSchema = z.object({\n  type: z.literal(\"move\"),\n  x: z.number(),\n  y: z.number(),\n});\n\nconst ComputerUseScrollActionSchema = z.object({\n  type: z.literal(\"scroll\"),\n  x: z.number(),\n  y: z.number(),\n  scroll_x: z.number(),\n  scroll_y: z.number(),\n});\n\nconst ComputerUseTypeActionSchema = z.object({\n  type: z.literal(\"type\"),\n  text: z.string(),\n});\n\nconst ComputerUseWaitActionSchema = z.object({\n  type: z.literal(\"wait\"),\n  duration: z.number().optional(),\n});\n\n// Discriminated union schema for individual action types\nconst ComputerUseActionUnionSchema = z.discriminatedUnion(\"type\", [\n  ComputerUseScreenshotActionSchema,\n  ComputerUseClickActionSchema,\n  ComputerUseDoubleClickActionSchema,\n  ComputerUseDragActionSchema,\n  ComputerUseKeypressActionSchema,\n  ComputerUseMoveActionSchema,\n  ComputerUseScrollActionSchema,\n  ComputerUseTypeActionSchema,\n  ComputerUseWaitActionSchema,\n]);\n\n// Schema for the input structure received from parseComputerCall\n// The action is wrapped in an `action` property: { action: { type: 'screenshot' } }\nexport const ComputerUseActionSchema = z.object({\n  action: ComputerUseActionUnionSchema,\n});\n\n// TypeScript types derived from Zod schemas\nexport type ComputerUseScreenshotActionType = z.infer<\n  typeof ComputerUseScreenshotActionSchema\n>;\nexport type ComputerUseClickActionType = z.infer<\n  typeof ComputerUseClickActionSchema\n>;\nexport type ComputerUseDoubleClickActionType = z.infer<\n  typeof ComputerUseDoubleClickActionSchema\n>;\nexport type ComputerUseDragActionType = z.infer<\n  typeof ComputerUseDragActionSchema\n>;\nexport type ComputerUseKeypressActionType = z.infer<\n  typeof ComputerUseKeypressActionSchema\n>;\nexport type ComputerUseMoveActionType = z.infer<\n  typeof ComputerUseMoveActionSchema\n>;\nexport type ComputerUseScrollActionType = z.infer<\n  typeof ComputerUseScrollActionSchema\n>;\nexport type ComputerUseTypeActionType = z.infer<\n  typeof ComputerUseTypeActionSchema\n>;\nexport type ComputerUseWaitActionType = z.infer<\n  typeof ComputerUseWaitActionSchema\n>;\n\n/**\n * Input structure for the Computer Use tool.\n * The action is wrapped in an `action` property.\n */\nexport interface ComputerUseInput {\n  action: ComputerUseAction;\n}\n\nexport type ComputerUseReturnType =\n  | string\n  | Promise<string>\n  | ToolMessage<any>\n  | Promise<ToolMessage<any>>;\n\n/**\n * Options for the Computer Use tool.\n */\nexport interface ComputerUseOptions {\n  /**\n   * The width of the computer display in pixels.\n   */\n  displayWidth: number;\n\n  /**\n   * The height of the computer display in pixels.\n   */\n  displayHeight: number;\n\n  /**\n   * The type of computer environment to control.\n   * - `browser`: Browser automation (recommended for most use cases)\n   * - `mac`: macOS environment\n   * - `windows`: Windows environment\n   * - `linux`: Linux environment\n   * - `ubuntu`: Ubuntu environment\n   */\n  environment: ComputerUseEnvironment;\n\n  /**\n   * Execute function that handles computer action execution.\n   * This function receives the action input and should return a base64-encoded\n   * screenshot of the result.\n   */\n  execute: (\n    action: ComputerUseAction,\n    runtime: ToolRuntime<any, any>\n  ) => ComputerUseReturnType;\n}\n\n/**\n * OpenAI Computer Use tool type for the Responses API.\n */\nexport type ComputerUseTool = OpenAIClient.Responses.ComputerTool;\n\nconst TOOL_NAME = \"computer_use\";\n\n/**\n * Creates a Computer Use tool that allows models to control computer interfaces\n * and perform tasks by simulating mouse clicks, keyboard input, scrolling, and more.\n *\n * **Computer Use** is a practical application of OpenAI's Computer-Using Agent (CUA)\n * model (`computer-use-preview`), which combines vision capabilities with advanced\n * reasoning to simulate controlling computer interfaces.\n *\n * **How it works**:\n * The tool operates in a continuous loop:\n * 1. Model sends computer actions (click, type, scroll, etc.)\n * 2. Your code executes these actions in a controlled environment\n * 3. You capture a screenshot of the result\n * 4. Send the screenshot back to the model\n * 5. Repeat until the task is complete\n *\n * **Important**: Computer use is in beta and requires careful consideration:\n * - Use in sandboxed environments only\n * - Do not use for high-stakes or authenticated tasks\n * - Always implement human-in-the-loop for important decisions\n * - Handle safety checks appropriately\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-computer-use | OpenAI Computer Use Documentation}\n *\n * @param options - Configuration options for the Computer Use tool\n * @returns A Computer Use tool that can be passed to `bindTools`\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"computer-use-preview\" });\n *\n * // With execute callback for automatic action handling\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n *   execute: async (action) => {\n *     if (action.type === \"screenshot\") {\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"click\") {\n *       await page.mouse.click(action.x, action.y, { button: action.button });\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"type\") {\n *       await page.keyboard.type(action.text);\n *       return captureScreenshot();\n *     }\n *     // Handle other actions...\n *     return captureScreenshot();\n *   },\n * });\n *\n * const llmWithComputer = model.bindTools([computer]);\n * const response = await llmWithComputer.invoke(\n *   \"Check the latest news on bing.com\"\n * );\n * ```\n *\n * @example\n * ```typescript\n * // Without execute callback (manual action handling)\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n * });\n *\n * const response = await model.invoke(\"Check the news\", {\n *   tools: [computer],\n * });\n *\n * // Access the computer call from the response\n * const computerCall = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"computer_call\"\n * );\n * if (computerCall) {\n *   console.log(\"Action to execute:\", computerCall.action);\n *   // Execute the action manually, then send back a screenshot\n * }\n * ```\n *\n * @example\n * ```typescript\n * // For macOS desktop automation with Docker\n * const computer = tools.computerUse({\n *   displayWidth: 1920,\n *   displayHeight: 1080,\n *   environment: \"mac\",\n *   execute: async (action) => {\n *     if (action.type === \"click\") {\n *       await dockerExec(\n *         `DISPLAY=:99 xdotool mousemove ${action.x} ${action.y} click 1`,\n *         containerName\n *       );\n *     }\n *     // Capture screenshot from container\n *     return await getDockerScreenshot(containerName);\n *   },\n * });\n * ```\n *\n * @remarks\n * - Only available through the Responses API (not Chat Completions)\n * - Requires `computer-use-preview` model\n * - Actions include: click, double_click, drag, keypress, move, screenshot, scroll, type, wait\n * - Safety checks may be returned that require acknowledgment before proceeding\n * - Use `truncation: \"auto\"` parameter when making requests\n * - Recommended to use with `reasoning.summary` for debugging\n */\nexport function computerUse(options: ComputerUseOptions) {\n  const computerTool = tool(\n    async (\n      input: ComputerUseInput,\n      runtime: ToolRuntime<{ messages: BaseMessage[] }>\n    ) => {\n      /**\n       * get computer_use call id from runtime\n       */\n      const aiMessage = runtime.state?.messages.at(-1) as AIMessage | undefined;\n      const computerToolCall = aiMessage?.tool_calls?.find(\n        (tc) => tc.name === \"computer_use\"\n      );\n      const computerToolCallId = computerToolCall?.id;\n      if (!computerToolCallId) {\n        throw new Error(\"Computer use call id not found\");\n      }\n\n      const result = await options.execute(input.action, runtime);\n\n      /**\n       * make sure {@link ToolMessage} is returned with the correct additional kwargs\n       */\n      if (typeof result === \"string\") {\n        return new ToolMessage({\n          content: result,\n          tool_call_id: computerToolCallId,\n          additional_kwargs: {\n            type: \"computer_call_output\",\n          },\n        });\n      }\n\n      /**\n       * make sure {@link ToolMessage} is returned with the correct additional kwargs\n       */\n      return new ToolMessage({\n        ...result,\n        tool_call_id: computerToolCallId,\n        additional_kwargs: {\n          type: \"computer_call_output\",\n          ...result.additional_kwargs,\n        },\n      });\n    },\n    {\n      name: TOOL_NAME,\n      description:\n        \"Control a computer interface by executing mouse clicks, keyboard input, scrolling, and other actions.\",\n      schema: ComputerUseActionSchema,\n    }\n  );\n\n  computerTool.extras = {\n    ...(computerTool.extras ?? {}),\n    providerToolDefinition: {\n      type: \"computer_use_preview\",\n      display_width: options.displayWidth,\n      display_height: options.displayHeight,\n      environment: options.environment,\n    } satisfies ComputerUseTool,\n  };\n\n  /**\n   * return as typed {@link DynamicStructuredTool} so we don't get any type\n   * errors like \"can't export tool without reference\"\n   */\n  return computerTool as DynamicStructuredTool<\n    typeof ComputerUseActionSchema,\n    ComputerUseInput,\n    unknown,\n    ComputerUseReturnType\n  >;\n}\n"],"mappings":";;;;;AAkDA,MAAM,oCAAoC,EAAE,OAAO,EACjD,MAAM,EAAE,QAAQ,aAAa,CAC9B,EAAC;AAEF,MAAM,+BAA+B,EAAE,OAAO;CAC5C,MAAM,EAAE,QAAQ,QAAQ;CACxB,GAAG,EAAE,QAAQ;CACb,GAAG,EAAE,QAAQ;CACb,QAAQ,EAAE,KAAK;EAAC;EAAQ;EAAS;EAAS;EAAQ;CAAU,EAAC,CAAC,QAAQ,OAAO;AAC9E,EAAC;AAEF,MAAM,qCAAqC,EAAE,OAAO;CAClD,MAAM,EAAE,QAAQ,eAAe;CAC/B,GAAG,EAAE,QAAQ;CACb,GAAG,EAAE,QAAQ;CACb,QAAQ,EAAE,KAAK;EAAC;EAAQ;EAAS;EAAS;EAAQ;CAAU,EAAC,CAAC,QAAQ,OAAO;AAC9E,EAAC;AAEF,MAAM,8BAA8B,EAAE,OAAO;CAC3C,MAAM,EAAE,QAAQ,OAAO;CACvB,MAAM,EAAE,MAAM,EAAE,OAAO;EAAE,GAAG,EAAE,QAAQ;EAAE,GAAG,EAAE,QAAQ;CAAE,EAAC,CAAC;AAC1D,EAAC;AAEF,MAAM,kCAAkC,EAAE,OAAO;CAC/C,MAAM,EAAE,QAAQ,WAAW;CAC3B,MAAM,EAAE,MAAM,EAAE,QAAQ,CAAC;AAC1B,EAAC;AAEF,MAAM,8BAA8B,EAAE,OAAO;CAC3C,MAAM,EAAE,QAAQ,OAAO;CACvB,GAAG,EAAE,QAAQ;CACb,GAAG,EAAE,QAAQ;AACd,EAAC;AAEF,MAAM,gCAAgC,EAAE,OAAO;CAC7C,MAAM,EAAE,QAAQ,SAAS;CACzB,GAAG,EAAE,QAAQ;CACb,GAAG,EAAE,QAAQ;CACb,UAAU,EAAE,QAAQ;CACpB,UAAU,EAAE,QAAQ;AACrB,EAAC;AAEF,MAAM,8BAA8B,EAAE,OAAO;CAC3C,MAAM,EAAE,QAAQ,OAAO;CACvB,MAAM,EAAE,QAAQ;AACjB,EAAC;AAEF,MAAM,8BAA8B,EAAE,OAAO;CAC3C,MAAM,EAAE,QAAQ,OAAO;CACvB,UAAU,EAAE,QAAQ,CAAC,UAAU;AAChC,EAAC;AAGF,MAAM,+BAA+B,EAAE,mBAAmB,QAAQ;CAChE;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACD,EAAC;AAIF,MAAa,0BAA0B,EAAE,OAAO,EAC9C,QAAQ,6BACT,EAAC;AAqFF,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkHlB,SAAgB,YAAYA,SAA6B;CACvD,MAAM,eAAe,KACnB,OACEC,OACAC,YACG;;;;EAIH,MAAM,YAAY,QAAQ,OAAO,SAAS,GAAG,GAAG;EAChD,MAAM,mBAAmB,WAAW,YAAY,KAC9C,CAAC,OAAO,GAAG,SAAS,eACrB;EACD,MAAM,qBAAqB,kBAAkB;AAC7C,MAAI,CAAC,mBACH,OAAM,IAAI,MAAM;EAGlB,MAAM,SAAS,MAAM,QAAQ,QAAQ,MAAM,QAAQ,QAAQ;;;;AAK3D,MAAI,OAAO,WAAW,SACpB,QAAO,IAAI,YAAY;GACrB,SAAS;GACT,cAAc;GACd,mBAAmB,EACjB,MAAM,uBACP;EACF;;;;AAMH,SAAO,IAAI,YAAY;GACrB,GAAG;GACH,cAAc;GACd,mBAAmB;IACjB,MAAM;IACN,GAAG,OAAO;GACX;EACF;CACF,GACD;EACE,MAAM;EACN,aACE;EACF,QAAQ;CACT,EACF;CAED,aAAa,SAAS;EACpB,GAAI,aAAa,UAAU,CAAE;EAC7B,wBAAwB;GACtB,MAAM;GACN,eAAe,QAAQ;GACvB,gBAAgB,QAAQ;GACxB,aAAa,QAAQ;EACtB;CACF;;;;;AAMD,QAAO;AAMR"}
+{"version":3,"file":"computerUse.js","names":["options: ComputerUseOptions","input: ComputerUseInput","runtime: ToolRuntime<{ messages: BaseMessage[] }>"],"sources":["../../src/tools/computerUse.ts"],"sourcesContent":["/* eslint-disable @typescript-eslint/no-explicit-any */\nimport { z } from \"zod/v4\";\nimport { OpenAI as OpenAIClient } from \"openai\";\nimport { tool, type DynamicStructuredTool } from \"@langchain/core/tools\";\nimport { type ToolRuntime } from \"@langchain/core/tools\";\nimport {\n  ToolMessage,\n  type AIMessage,\n  type BaseMessage,\n} from \"@langchain/core/messages\";\n\n/**\n * The type of computer environment to control.\n */\nexport type ComputerUseEnvironment =\n  | \"browser\"\n  | \"mac\"\n  | \"windows\"\n  | \"linux\"\n  | \"ubuntu\";\n\n/**\n * Re-export action types from OpenAI SDK for convenience.\n */\nexport type ComputerUseClickAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Click;\nexport type ComputerUseDoubleClickAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.DoubleClick;\nexport type ComputerUseDragAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Drag;\nexport type ComputerUseKeypressAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Keypress;\nexport type ComputerUseMoveAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Move;\nexport type ComputerUseScreenshotAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Screenshot;\nexport type ComputerUseScrollAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Scroll;\nexport type ComputerUseTypeAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Type;\nexport type ComputerUseWaitAction =\n  OpenAIClient.Responses.ResponseComputerToolCall.Wait;\n\n/**\n * Union type of all computer use actions from OpenAI SDK.\n */\nexport type ComputerUseAction =\n  OpenAIClient.Responses.ResponseComputerToolCall[\"action\"];\n\n// Zod schemas for computer use actions\nconst ComputerUseScreenshotActionSchema = z.object({\n  type: z.literal(\"screenshot\"),\n});\n\nconst ComputerUseClickActionSchema = z.object({\n  type: z.literal(\"click\"),\n  x: z.number(),\n  y: z.number(),\n  button: z.enum([\"left\", \"right\", \"wheel\", \"back\", \"forward\"]).default(\"left\"),\n});\n\nconst ComputerUseDoubleClickActionSchema = z.object({\n  type: z.literal(\"double_click\"),\n  x: z.number(),\n  y: z.number(),\n  button: z.enum([\"left\", \"right\", \"wheel\", \"back\", \"forward\"]).default(\"left\"),\n});\n\nconst ComputerUseDragActionSchema = z.object({\n  type: z.literal(\"drag\"),\n  path: z.array(z.object({ x: z.number(), y: z.number() })),\n});\n\nconst ComputerUseKeypressActionSchema = z.object({\n  type: z.literal(\"keypress\"),\n  keys: z.array(z.string()),\n});\n\nconst ComputerUseMoveActionSchema = z.object({\n  type: z.literal(\"move\"),\n  x: z.number(),\n  y: z.number(),\n});\n\nconst ComputerUseScrollActionSchema = z.object({\n  type: z.literal(\"scroll\"),\n  x: z.number(),\n  y: z.number(),\n  scroll_x: z.number(),\n  scroll_y: z.number(),\n});\n\nconst ComputerUseTypeActionSchema = z.object({\n  type: z.literal(\"type\"),\n  text: z.string(),\n});\n\nconst ComputerUseWaitActionSchema = z.object({\n  type: z.literal(\"wait\"),\n  duration: z.number().optional(),\n});\n\n// Union schema for individual action types\nconst ComputerUseActionUnionSchema = z.union([\n  ComputerUseScreenshotActionSchema,\n  ComputerUseClickActionSchema,\n  ComputerUseDoubleClickActionSchema,\n  ComputerUseDragActionSchema,\n  ComputerUseKeypressActionSchema,\n  ComputerUseMoveActionSchema,\n  ComputerUseScrollActionSchema,\n  ComputerUseTypeActionSchema,\n  ComputerUseWaitActionSchema,\n]);\n\n// Schema for the input structure received from parseComputerCall\n// The action is wrapped in an `action` property: { action: { type: 'screenshot' } }\nexport const ComputerUseActionSchema = z.object({\n  action: ComputerUseActionUnionSchema,\n});\n\n// TypeScript types derived from Zod schemas\nexport type ComputerUseScreenshotActionType = z.infer<\n  typeof ComputerUseScreenshotActionSchema\n>;\nexport type ComputerUseClickActionType = z.infer<\n  typeof ComputerUseClickActionSchema\n>;\nexport type ComputerUseDoubleClickActionType = z.infer<\n  typeof ComputerUseDoubleClickActionSchema\n>;\nexport type ComputerUseDragActionType = z.infer<\n  typeof ComputerUseDragActionSchema\n>;\nexport type ComputerUseKeypressActionType = z.infer<\n  typeof ComputerUseKeypressActionSchema\n>;\nexport type ComputerUseMoveActionType = z.infer<\n  typeof ComputerUseMoveActionSchema\n>;\nexport type ComputerUseScrollActionType = z.infer<\n  typeof ComputerUseScrollActionSchema\n>;\nexport type ComputerUseTypeActionType = z.infer<\n  typeof ComputerUseTypeActionSchema\n>;\nexport type ComputerUseWaitActionType = z.infer<\n  typeof ComputerUseWaitActionSchema\n>;\n\n/**\n * Input structure for the Computer Use tool.\n * The action is wrapped in an `action` property.\n */\nexport interface ComputerUseInput {\n  action: ComputerUseAction;\n}\n\nexport type ComputerUseReturnType =\n  | string\n  | Promise<string>\n  | ToolMessage<any>\n  | Promise<ToolMessage<any>>;\n\n/**\n * Options for the Computer Use tool.\n */\nexport interface ComputerUseOptions {\n  /**\n   * The width of the computer display in pixels.\n   */\n  displayWidth: number;\n\n  /**\n   * The height of the computer display in pixels.\n   */\n  displayHeight: number;\n\n  /**\n   * The type of computer environment to control.\n   * - `browser`: Browser automation (recommended for most use cases)\n   * - `mac`: macOS environment\n   * - `windows`: Windows environment\n   * - `linux`: Linux environment\n   * - `ubuntu`: Ubuntu environment\n   */\n  environment: ComputerUseEnvironment;\n\n  /**\n   * Execute function that handles computer action execution.\n   * This function receives the action input and should return a base64-encoded\n   * screenshot of the result.\n   */\n  execute: (\n    action: ComputerUseAction,\n    runtime: ToolRuntime<any, any>\n  ) => ComputerUseReturnType;\n}\n\n/**\n * OpenAI Computer Use tool type for the Responses API.\n */\nexport type ComputerUseTool = OpenAIClient.Responses.ComputerTool;\n\nconst TOOL_NAME = \"computer_use\";\n\n/**\n * Creates a Computer Use tool that allows models to control computer interfaces\n * and perform tasks by simulating mouse clicks, keyboard input, scrolling, and more.\n *\n * **Computer Use** is a practical application of OpenAI's Computer-Using Agent (CUA)\n * model (`computer-use-preview`), which combines vision capabilities with advanced\n * reasoning to simulate controlling computer interfaces.\n *\n * **How it works**:\n * The tool operates in a continuous loop:\n * 1. Model sends computer actions (click, type, scroll, etc.)\n * 2. Your code executes these actions in a controlled environment\n * 3. You capture a screenshot of the result\n * 4. Send the screenshot back to the model\n * 5. Repeat until the task is complete\n *\n * **Important**: Computer use is in beta and requires careful consideration:\n * - Use in sandboxed environments only\n * - Do not use for high-stakes or authenticated tasks\n * - Always implement human-in-the-loop for important decisions\n * - Handle safety checks appropriately\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-computer-use | OpenAI Computer Use Documentation}\n *\n * @param options - Configuration options for the Computer Use tool\n * @returns A Computer Use tool that can be passed to `bindTools`\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"computer-use-preview\" });\n *\n * // With execute callback for automatic action handling\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n *   execute: async (action) => {\n *     if (action.type === \"screenshot\") {\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"click\") {\n *       await page.mouse.click(action.x, action.y, { button: action.button });\n *       return captureScreenshot();\n *     }\n *     if (action.type === \"type\") {\n *       await page.keyboard.type(action.text);\n *       return captureScreenshot();\n *     }\n *     // Handle other actions...\n *     return captureScreenshot();\n *   },\n * });\n *\n * const llmWithComputer = model.bindTools([computer]);\n * const response = await llmWithComputer.invoke(\n *   \"Check the latest news on bing.com\"\n * );\n * ```\n *\n * @example\n * ```typescript\n * // Without execute callback (manual action handling)\n * const computer = tools.computerUse({\n *   displayWidth: 1024,\n *   displayHeight: 768,\n *   environment: \"browser\",\n * });\n *\n * const response = await model.invoke(\"Check the news\", {\n *   tools: [computer],\n * });\n *\n * // Access the computer call from the response\n * const computerCall = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"computer_call\"\n * );\n * if (computerCall) {\n *   console.log(\"Action to execute:\", computerCall.action);\n *   // Execute the action manually, then send back a screenshot\n * }\n * ```\n *\n * @example\n * ```typescript\n * // For macOS desktop automation with Docker\n * const computer = tools.computerUse({\n *   displayWidth: 1920,\n *   displayHeight: 1080,\n *   environment: \"mac\",\n *   execute: async (action) => {\n *     if (action.type === \"click\") {\n *       await dockerExec(\n *         `DISPLAY=:99 xdotool mousemove ${action.x} ${action.y} click 1`,\n *         containerName\n *       );\n *     }\n *     // Capture screenshot from container\n *     return await getDockerScreenshot(containerName);\n *   },\n * });\n * ```\n *\n * @remarks\n * - Only available through the Responses API (not Chat Completions)\n * - Requires `computer-use-preview` model\n * - Actions include: click, double_click, drag, keypress, move, screenshot, scroll, type, wait\n * - Safety checks may be returned that require acknowledgment before proceeding\n * - Use `truncation: \"auto\"` parameter when making requests\n * - Recommended to use with `reasoning.summary` for debugging\n */\nexport function computerUse(options: ComputerUseOptions) {\n  const computerTool = tool(\n    async (\n      input: ComputerUseInput,\n      runtime: ToolRuntime<{ messages: BaseMessage[] }>\n    ) => {\n      /**\n       * get computer_use call id from runtime\n       */\n      const aiMessage = runtime.state?.messages.at(-1) as AIMessage | undefined;\n      const computerToolCall = aiMessage?.tool_calls?.find(\n        (tc) => tc.name === \"computer_use\"\n      );\n      const computerToolCallId = computerToolCall?.id;\n      if (!computerToolCallId) {\n        throw new Error(\"Computer use call id not found\");\n      }\n\n      const result = await options.execute(input.action, runtime);\n\n      /**\n       * make sure {@link ToolMessage} is returned with the correct additional kwargs\n       */\n      if (typeof result === \"string\") {\n        return new ToolMessage({\n          content: result,\n          tool_call_id: computerToolCallId,\n          additional_kwargs: {\n            type: \"computer_call_output\",\n          },\n        });\n      }\n\n      /**\n       * make sure {@link ToolMessage} is returned with the correct additional kwargs\n       */\n      return new ToolMessage({\n        ...result,\n        tool_call_id: computerToolCallId,\n        additional_kwargs: {\n          type: \"computer_call_output\",\n          ...result.additional_kwargs,\n        },\n      });\n    },\n    {\n      name: TOOL_NAME,\n      description:\n        \"Control a computer interface by executing mouse clicks, keyboard input, scrolling, and other actions.\",\n      schema: ComputerUseActionSchema,\n    }\n  );\n\n  computerTool.extras = {\n    ...(computerTool.extras ?? {}),\n    providerToolDefinition: {\n      type: \"computer_use_preview\",\n      display_width: options.displayWidth,\n      display_height: options.displayHeight,\n      environment: options.environment,\n    } satisfies ComputerUseTool,\n  };\n\n  /**\n   * return as typed {@link DynamicStructuredTool} so we don't get any type\n   * errors like \"can't export tool without reference\"\n   */\n  return computerTool as DynamicStructuredTool<\n    typeof ComputerUseActionSchema,\n    ComputerUseInput,\n    unknown,\n    ComputerUseReturnType\n  >;\n}\n"],"mappings":";;;;;AAkDA,MAAM,oCAAoC,EAAE,OAAO,EACjD,MAAM,EAAE,QAAQ,aAAa,CAC9B,EAAC;AAEF,MAAM,+BAA+B,EAAE,OAAO;CAC5C,MAAM,EAAE,QAAQ,QAAQ;CACxB,GAAG,EAAE,QAAQ;CACb,GAAG,EAAE,QAAQ;CACb,QAAQ,EAAE,KAAK;EAAC;EAAQ;EAAS;EAAS;EAAQ;CAAU,EAAC,CAAC,QAAQ,OAAO;AAC9E,EAAC;AAEF,MAAM,qCAAqC,EAAE,OAAO;CAClD,MAAM,EAAE,QAAQ,eAAe;CAC/B,GAAG,EAAE,QAAQ;CACb,GAAG,EAAE,QAAQ;CACb,QAAQ,EAAE,KAAK;EAAC;EAAQ;EAAS;EAAS;EAAQ;CAAU,EAAC,CAAC,QAAQ,OAAO;AAC9E,EAAC;AAEF,MAAM,8BAA8B,EAAE,OAAO;CAC3C,MAAM,EAAE,QAAQ,OAAO;CACvB,MAAM,EAAE,MAAM,EAAE,OAAO;EAAE,GAAG,EAAE,QAAQ;EAAE,GAAG,EAAE,QAAQ;CAAE,EAAC,CAAC;AAC1D,EAAC;AAEF,MAAM,kCAAkC,EAAE,OAAO;CAC/C,MAAM,EAAE,QAAQ,WAAW;CAC3B,MAAM,EAAE,MAAM,EAAE,QAAQ,CAAC;AAC1B,EAAC;AAEF,MAAM,8BAA8B,EAAE,OAAO;CAC3C,MAAM,EAAE,QAAQ,OAAO;CACvB,GAAG,EAAE,QAAQ;CACb,GAAG,EAAE,QAAQ;AACd,EAAC;AAEF,MAAM,gCAAgC,EAAE,OAAO;CAC7C,MAAM,EAAE,QAAQ,SAAS;CACzB,GAAG,EAAE,QAAQ;CACb,GAAG,EAAE,QAAQ;CACb,UAAU,EAAE,QAAQ;CACpB,UAAU,EAAE,QAAQ;AACrB,EAAC;AAEF,MAAM,8BAA8B,EAAE,OAAO;CAC3C,MAAM,EAAE,QAAQ,OAAO;CACvB,MAAM,EAAE,QAAQ;AACjB,EAAC;AAEF,MAAM,8BAA8B,EAAE,OAAO;CAC3C,MAAM,EAAE,QAAQ,OAAO;CACvB,UAAU,EAAE,QAAQ,CAAC,UAAU;AAChC,EAAC;AAGF,MAAM,+BAA+B,EAAE,MAAM;CAC3C;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACD,EAAC;AAIF,MAAa,0BAA0B,EAAE,OAAO,EAC9C,QAAQ,6BACT,EAAC;AAqFF,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAkHlB,SAAgB,YAAYA,SAA6B;CACvD,MAAM,eAAe,KACnB,OACEC,OACAC,YACG;;;;EAIH,MAAM,YAAY,QAAQ,OAAO,SAAS,GAAG,GAAG;EAChD,MAAM,mBAAmB,WAAW,YAAY,KAC9C,CAAC,OAAO,GAAG,SAAS,eACrB;EACD,MAAM,qBAAqB,kBAAkB;AAC7C,MAAI,CAAC,mBACH,OAAM,IAAI,MAAM;EAGlB,MAAM,SAAS,MAAM,QAAQ,QAAQ,MAAM,QAAQ,QAAQ;;;;AAK3D,MAAI,OAAO,WAAW,SACpB,QAAO,IAAI,YAAY;GACrB,SAAS;GACT,cAAc;GACd,mBAAmB,EACjB,MAAM,uBACP;EACF;;;;AAMH,SAAO,IAAI,YAAY;GACrB,GAAG;GACH,cAAc;GACd,mBAAmB;IACjB,MAAM;IACN,GAAG,OAAO;GACX;EACF;CACF,GACD;EACE,MAAM;EACN,aACE;EACF,QAAQ;CACT,EACF;CAED,aAAa,SAAS;EACpB,GAAI,aAAa,UAAU,CAAE;EAC7B,wBAAwB;GACtB,MAAM;GACN,eAAe,QAAQ;GACvB,gBAAgB,QAAQ;GACxB,aAAa,QAAQ;EACtB;CACF;;;;;AAMD,QAAO;AAMR"}

package/dist/tools/imageGeneration.cjs

@@ -110,6 +110,7 @@ function convertInputImageMask(mask) {
 function imageGeneration(options) {
 	return {
 		type: "image_generation",
+		action: options?.action,
 		background: options?.background,
 		input_fidelity: options?.inputFidelity,
 		input_image_mask: convertInputImageMask(options?.inputImageMask),

package/dist/tools/imageGeneration.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"imageGeneration.cjs","names":["mask: ImageGenerationInputMask | undefined","options?: ImageGenerationOptions"],"sources":["../../src/tools/imageGeneration.ts"],"sourcesContent":["import { OpenAI as OpenAIClient } from \"openai\";\nimport type { ServerTool } from \"@langchain/core/tools\";\n\n/**\n * Optional mask for inpainting. Allows you to specify areas of the image\n * that should be regenerated.\n */\nexport interface ImageGenerationInputMask {\n  /**\n   * Base64-encoded mask image URL.\n   */\n  imageUrl?: string;\n  /**\n   * File ID for the mask image (uploaded via OpenAI File API).\n   */\n  fileId?: string;\n}\n\n/**\n * Options for the Image Generation tool.\n */\nexport interface ImageGenerationOptions {\n  /**\n   * Background type for the generated image.\n   * - `transparent`: Generate image with transparent background\n   * - `opaque`: Generate image with opaque background\n   * - `auto`: Let the model decide based on the prompt\n   * @default \"auto\"\n   */\n  background?: \"transparent\" | \"opaque\" | \"auto\";\n\n  /**\n   * Control how much effort the model will exert to match the style and features,\n   * especially facial features, of input images. This parameter is only supported\n   * for `gpt-image-1`. Unsupported for `gpt-image-1-mini`.\n   * - `high`: Higher fidelity to input images\n   * - `low`: Lower fidelity to input images\n   * @default \"low\"\n   */\n  inputFidelity?: \"high\" | \"low\";\n\n  /**\n   * Optional mask for inpainting. Use this to specify areas of an image\n   * that should be regenerated.\n   */\n  inputImageMask?: ImageGenerationInputMask;\n\n  /**\n   * The image generation model to use.\n   * @default \"gpt-image-1\"\n   */\n  model?: \"gpt-image-1\" | \"gpt-image-1-mini\" | \"gpt-image-1.5\";\n\n  /**\n   * Moderation level for the generated image.\n   * - `auto`: Standard moderation\n   * - `low`: Less restrictive moderation\n   * @default \"auto\"\n   */\n  moderation?: \"auto\" | \"low\";\n\n  /**\n   * Compression level for the output image (0-100).\n   * Only applies to JPEG and WebP formats.\n   * @default 100\n   */\n  outputCompression?: number;\n\n  /**\n   * The output format of the generated image.\n   * @default \"png\"\n   */\n  outputFormat?: \"png\" | \"webp\" | \"jpeg\";\n\n  /**\n   * Number of partial images to generate in streaming mode (0-3).\n   * When set, the model will return partial images as they are generated,\n   * providing faster visual feedback.\n   * @default 0\n   */\n  partialImages?: number;\n\n  /**\n   * The quality of the generated image.\n   * - `low`: Faster generation, lower quality\n   * - `medium`: Balanced generation time and quality\n   * - `high`: Slower generation, higher quality\n   * - `auto`: Let the model decide based on the prompt\n   * @default \"auto\"\n   */\n  quality?: \"low\" | \"medium\" | \"high\" | \"auto\";\n\n  /**\n   * The size of the generated image.\n   * - `1024x1024`: Square format\n   * - `1024x1536`: Portrait format\n   * - `1536x1024`: Landscape format\n   * - `auto`: Let the model decide based on the prompt\n   * @default \"auto\"\n   */\n  size?: \"1024x1024\" | \"1024x1536\" | \"1536x1024\" | \"auto\";\n}\n\n/**\n * OpenAI Image Generation tool type for the Responses API.\n */\nexport type ImageGenerationTool = OpenAIClient.Responses.Tool.ImageGeneration;\n\n/**\n * Converts input mask options to the API format.\n */\nfunction convertInputImageMask(\n  mask: ImageGenerationInputMask | undefined\n): ImageGenerationTool[\"input_image_mask\"] {\n  if (!mask) return undefined;\n  return {\n    image_url: mask.imageUrl,\n    file_id: mask.fileId,\n  };\n}\n\n/**\n * Creates an Image Generation tool that allows models to generate or edit images\n * using text prompts and optional image inputs.\n *\n * The image generation tool leverages the GPT Image model and automatically\n * optimizes text inputs for improved performance. When included in a request,\n * the model can decide when and how to generate images as part of the conversation.\n *\n * **Key Features**:\n * - Generate images from text descriptions\n * - Edit existing images with text instructions\n * - Multi-turn image editing by referencing previous responses\n * - Configurable output options (size, quality, format)\n * - Streaming support for partial image generation\n *\n * **Prompting Tips**:\n * - Use terms like \"draw\" or \"edit\" in your prompt for best results\n * - For combining images, say \"edit the first image by adding this element\" instead of \"combine\"\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-image-generation | OpenAI Image Generation Documentation}\n *\n * @param options - Configuration options for the Image Generation tool\n * @returns An Image Generation tool definition to be passed to the OpenAI Responses API\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"gpt-4o\" });\n *\n * // Basic usage - generate an image\n * const response = await model.invoke(\n *   \"Generate an image of a gray tabby cat hugging an otter with an orange scarf\",\n *   { tools: [tools.imageGeneration()] }\n * );\n *\n * // Access the generated image\n * const imageData = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"image_generation_call\"\n * );\n * if (imageData?.result) {\n *   // imageData.result contains the base64-encoded image\n *   const fs = await import(\"fs\");\n *   fs.writeFileSync(\"output.png\", Buffer.from(imageData.result, \"base64\"));\n * }\n *\n * // With custom options\n * const response = await model.invoke(\n *   \"Draw a beautiful sunset over mountains\",\n *   {\n *     tools: [tools.imageGeneration({\n *       size: \"1536x1024\",      // Landscape format\n *       quality: \"high\",        // Higher quality output\n *       outputFormat: \"jpeg\",   // JPEG format\n *       outputCompression: 90,  // 90% compression\n *     })]\n *   }\n * );\n *\n * // With transparent background\n * const response = await model.invoke(\n *   \"Create a logo with a transparent background\",\n *   {\n *     tools: [tools.imageGeneration({\n *       background: \"transparent\",\n *       outputFormat: \"png\",\n *     })]\n *   }\n * );\n *\n * // Force the model to use image generation\n * const response = await model.invoke(\n *   \"A serene lake at dawn\",\n *   {\n *     tools: [tools.imageGeneration()],\n *     tool_choice: { type: \"image_generation\" },\n *   }\n * );\n *\n * // Enable streaming with partial images\n * const response = await model.invoke(\n *   \"Draw a detailed fantasy castle\",\n *   {\n *     tools: [tools.imageGeneration({\n *       partialImages: 2,  // Get 2 partial images during generation\n *     })]\n *   }\n * );\n * ```\n *\n * @remarks\n * - Supported models: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o3\n * - The image generation process always uses `gpt-image-1` model internally\n * - The model will automatically revise prompts for improved performance\n * - Access the revised prompt via `revised_prompt` field in the output\n * - Multi-turn editing is supported by passing previous response messages\n */\nexport function imageGeneration(options?: ImageGenerationOptions): ServerTool {\n  return {\n    type: \"image_generation\",\n    background: options?.background,\n    input_fidelity: options?.inputFidelity,\n    input_image_mask: convertInputImageMask(options?.inputImageMask),\n    // @ts-expect-error - type issue with openai not supporting gpt-image-1.5\n    model: options?.model,\n    moderation: options?.moderation,\n    output_compression: options?.outputCompression,\n    output_format: options?.outputFormat,\n    partial_images: options?.partialImages,\n    quality: options?.quality,\n    size: options?.size,\n  } satisfies ImageGenerationTool;\n}\n"],"mappings":";;;;;AA+GA,SAAS,sBACPA,MACyC;AACzC,KAAI,CAAC,KAAM,QAAO;AAClB,QAAO;EACL,WAAW,KAAK;EAChB,SAAS,KAAK;CACf;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmGD,SAAgB,gBAAgBC,SAA8C;AAC5E,QAAO;EACL,MAAM;EACN,YAAY,SAAS;EACrB,gBAAgB,SAAS;EACzB,kBAAkB,sBAAsB,SAAS,eAAe;EAEhE,OAAO,SAAS;EAChB,YAAY,SAAS;EACrB,oBAAoB,SAAS;EAC7B,eAAe,SAAS;EACxB,gBAAgB,SAAS;EACzB,SAAS,SAAS;EAClB,MAAM,SAAS;CAChB;AACF"}
+{"version":3,"file":"imageGeneration.cjs","names":["mask: ImageGenerationInputMask | undefined","options?: ImageGenerationOptions"],"sources":["../../src/tools/imageGeneration.ts"],"sourcesContent":["import { OpenAI as OpenAIClient } from \"openai\";\nimport type { ServerTool } from \"@langchain/core/tools\";\n\n/**\n * Optional mask for inpainting. Allows you to specify areas of the image\n * that should be regenerated.\n */\nexport interface ImageGenerationInputMask {\n  /**\n   * Base64-encoded mask image URL.\n   */\n  imageUrl?: string;\n  /**\n   * File ID for the mask image (uploaded via OpenAI File API).\n   */\n  fileId?: string;\n}\n\n/**\n * Options for the Image Generation tool.\n */\nexport interface ImageGenerationOptions {\n  /**\n   * Whether to generate a new image or edit an existing image.\n   * - `generate`: Generate a new image from scratch\n   * - `edit`: Edit an existing image\n   * - `auto`: Let the model decide based on the input\n   * @default \"auto\"\n   */\n  action?: \"generate\" | \"edit\" | \"auto\";\n\n  /**\n   * Background type for the generated image.\n   * - `transparent`: Generate image with transparent background\n   * - `opaque`: Generate image with opaque background\n   * - `auto`: Let the model decide based on the prompt\n   * @default \"auto\"\n   */\n  background?: \"transparent\" | \"opaque\" | \"auto\";\n\n  /**\n   * Control how much effort the model will exert to match the style and features,\n   * especially facial features, of input images. This parameter is only supported\n   * for `gpt-image-1`. Unsupported for `gpt-image-1-mini`.\n   * - `high`: Higher fidelity to input images\n   * - `low`: Lower fidelity to input images\n   * @default \"low\"\n   */\n  inputFidelity?: \"high\" | \"low\";\n\n  /**\n   * Optional mask for inpainting. Use this to specify areas of an image\n   * that should be regenerated.\n   */\n  inputImageMask?: ImageGenerationInputMask;\n\n  /**\n   * The image generation model to use.\n   * @default \"gpt-image-1\"\n   */\n  model?: \"gpt-image-1\" | \"gpt-image-1-mini\" | \"gpt-image-1.5\";\n\n  /**\n   * Moderation level for the generated image.\n   * - `auto`: Standard moderation\n   * - `low`: Less restrictive moderation\n   * @default \"auto\"\n   */\n  moderation?: \"auto\" | \"low\";\n\n  /**\n   * Compression level for the output image (0-100).\n   * Only applies to JPEG and WebP formats.\n   * @default 100\n   */\n  outputCompression?: number;\n\n  /**\n   * The output format of the generated image.\n   * @default \"png\"\n   */\n  outputFormat?: \"png\" | \"webp\" | \"jpeg\";\n\n  /**\n   * Number of partial images to generate in streaming mode (0-3).\n   * When set, the model will return partial images as they are generated,\n   * providing faster visual feedback.\n   * @default 0\n   */\n  partialImages?: number;\n\n  /**\n   * The quality of the generated image.\n   * - `low`: Faster generation, lower quality\n   * - `medium`: Balanced generation time and quality\n   * - `high`: Slower generation, higher quality\n   * - `auto`: Let the model decide based on the prompt\n   * @default \"auto\"\n   */\n  quality?: \"low\" | \"medium\" | \"high\" | \"auto\";\n\n  /**\n   * The size of the generated image.\n   * - `1024x1024`: Square format\n   * - `1024x1536`: Portrait format\n   * - `1536x1024`: Landscape format\n   * - `auto`: Let the model decide based on the prompt\n   * @default \"auto\"\n   */\n  size?: \"1024x1024\" | \"1024x1536\" | \"1536x1024\" | \"auto\";\n}\n\n/**\n * OpenAI Image Generation tool type for the Responses API.\n */\nexport type ImageGenerationTool = OpenAIClient.Responses.Tool.ImageGeneration;\n\n/**\n * Converts input mask options to the API format.\n */\nfunction convertInputImageMask(\n  mask: ImageGenerationInputMask | undefined\n): ImageGenerationTool[\"input_image_mask\"] {\n  if (!mask) return undefined;\n  return {\n    image_url: mask.imageUrl,\n    file_id: mask.fileId,\n  };\n}\n\n/**\n * Creates an Image Generation tool that allows models to generate or edit images\n * using text prompts and optional image inputs.\n *\n * The image generation tool leverages the GPT Image model and automatically\n * optimizes text inputs for improved performance. When included in a request,\n * the model can decide when and how to generate images as part of the conversation.\n *\n * **Key Features**:\n * - Generate images from text descriptions\n * - Edit existing images with text instructions\n * - Multi-turn image editing by referencing previous responses\n * - Configurable output options (size, quality, format)\n * - Streaming support for partial image generation\n *\n * **Prompting Tips**:\n * - Use terms like \"draw\" or \"edit\" in your prompt for best results\n * - For combining images, say \"edit the first image by adding this element\" instead of \"combine\"\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-image-generation | OpenAI Image Generation Documentation}\n *\n * @param options - Configuration options for the Image Generation tool\n * @returns An Image Generation tool definition to be passed to the OpenAI Responses API\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"gpt-4o\" });\n *\n * // Basic usage - generate an image\n * const response = await model.invoke(\n *   \"Generate an image of a gray tabby cat hugging an otter with an orange scarf\",\n *   { tools: [tools.imageGeneration()] }\n * );\n *\n * // Access the generated image\n * const imageData = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"image_generation_call\"\n * );\n * if (imageData?.result) {\n *   // imageData.result contains the base64-encoded image\n *   const fs = await import(\"fs\");\n *   fs.writeFileSync(\"output.png\", Buffer.from(imageData.result, \"base64\"));\n * }\n *\n * // With custom options\n * const response = await model.invoke(\n *   \"Draw a beautiful sunset over mountains\",\n *   {\n *     tools: [tools.imageGeneration({\n *       size: \"1536x1024\",      // Landscape format\n *       quality: \"high\",        // Higher quality output\n *       outputFormat: \"jpeg\",   // JPEG format\n *       outputCompression: 90,  // 90% compression\n *     })]\n *   }\n * );\n *\n * // With transparent background\n * const response = await model.invoke(\n *   \"Create a logo with a transparent background\",\n *   {\n *     tools: [tools.imageGeneration({\n *       background: \"transparent\",\n *       outputFormat: \"png\",\n *     })]\n *   }\n * );\n *\n * // Force the model to use image generation\n * const response = await model.invoke(\n *   \"A serene lake at dawn\",\n *   {\n *     tools: [tools.imageGeneration()],\n *     tool_choice: { type: \"image_generation\" },\n *   }\n * );\n *\n * // Enable streaming with partial images\n * const response = await model.invoke(\n *   \"Draw a detailed fantasy castle\",\n *   {\n *     tools: [tools.imageGeneration({\n *       partialImages: 2,  // Get 2 partial images during generation\n *     })]\n *   }\n * );\n * ```\n *\n * @remarks\n * - Supported models: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o3\n * - The image generation process always uses `gpt-image-1` model internally\n * - The model will automatically revise prompts for improved performance\n * - Access the revised prompt via `revised_prompt` field in the output\n * - Multi-turn editing is supported by passing previous response messages\n */\nexport function imageGeneration(options?: ImageGenerationOptions): ServerTool {\n  return {\n    type: \"image_generation\",\n    action: options?.action,\n    background: options?.background,\n    input_fidelity: options?.inputFidelity,\n    input_image_mask: convertInputImageMask(options?.inputImageMask),\n    model: options?.model,\n    moderation: options?.moderation,\n    output_compression: options?.outputCompression,\n    output_format: options?.outputFormat,\n    partial_images: options?.partialImages,\n    quality: options?.quality,\n    size: options?.size,\n  } satisfies ImageGenerationTool;\n}\n"],"mappings":";;;;;AAwHA,SAAS,sBACPA,MACyC;AACzC,KAAI,CAAC,KAAM,QAAO;AAClB,QAAO;EACL,WAAW,KAAK;EAChB,SAAS,KAAK;CACf;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmGD,SAAgB,gBAAgBC,SAA8C;AAC5E,QAAO;EACL,MAAM;EACN,QAAQ,SAAS;EACjB,YAAY,SAAS;EACrB,gBAAgB,SAAS;EACzB,kBAAkB,sBAAsB,SAAS,eAAe;EAChE,OAAO,SAAS;EAChB,YAAY,SAAS;EACrB,oBAAoB,SAAS;EAC7B,eAAe,SAAS;EACxB,gBAAgB,SAAS;EACzB,SAAS,SAAS;EAClB,MAAM,SAAS;CAChB;AACF"}

package/dist/tools/imageGeneration.d.cts

@@ -21,6 +21,14 @@ interface ImageGenerationInputMask {
  * Options for the Image Generation tool.
  */
 interface ImageGenerationOptions {
+  /**
+   * Whether to generate a new image or edit an existing image.
+   * - `generate`: Generate a new image from scratch
+   * - `edit`: Edit an existing image
+   * - `auto`: Let the model decide based on the input
+   * @default "auto"
+   */
+  action?: "generate" | "edit" | "auto";
   /**
    * Background type for the generated image.
    * - `transparent`: Generate image with transparent background

package/dist/tools/imageGeneration.d.cts.map

@@ -1 +1 @@
-{"version":3,"file":"imageGeneration.d.cts","names":["OpenAI","OpenAIClient","ServerTool","ImageGenerationInputMask","ImageGenerationOptions","ImageGenerationTool","Responses","Tool","ImageGeneration","imageGeneration"],"sources":["../../src/tools/imageGeneration.d.ts"],"sourcesContent":["import { OpenAI as OpenAIClient } from \"openai\";\nimport type { ServerTool } from \"@langchain/core/tools\";\n/**\n * Optional mask for inpainting. Allows you to specify areas of the image\n * that should be regenerated.\n */\nexport interface ImageGenerationInputMask {\n    /**\n     * Base64-encoded mask image URL.\n     */\n    imageUrl?: string;\n    /**\n     * File ID for the mask image (uploaded via OpenAI File API).\n     */\n    fileId?: string;\n}\n/**\n * Options for the Image Generation tool.\n */\nexport interface ImageGenerationOptions {\n    /**\n     * Background type for the generated image.\n     * - `transparent`: Generate image with transparent background\n     * - `opaque`: Generate image with opaque background\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    background?: \"transparent\" | \"opaque\" | \"auto\";\n    /**\n     * Control how much effort the model will exert to match the style and features,\n     * especially facial features, of input images. This parameter is only supported\n     * for `gpt-image-1`. Unsupported for `gpt-image-1-mini`.\n     * - `high`: Higher fidelity to input images\n     * - `low`: Lower fidelity to input images\n     * @default \"low\"\n     */\n    inputFidelity?: \"high\" | \"low\";\n    /**\n     * Optional mask for inpainting. Use this to specify areas of an image\n     * that should be regenerated.\n     */\n    inputImageMask?: ImageGenerationInputMask;\n    /**\n     * The image generation model to use.\n     * @default \"gpt-image-1\"\n     */\n    model?: \"gpt-image-1\" | \"gpt-image-1-mini\" | \"gpt-image-1.5\";\n    /**\n     * Moderation level for the generated image.\n     * - `auto`: Standard moderation\n     * - `low`: Less restrictive moderation\n     * @default \"auto\"\n     */\n    moderation?: \"auto\" | \"low\";\n    /**\n     * Compression level for the output image (0-100).\n     * Only applies to JPEG and WebP formats.\n     * @default 100\n     */\n    outputCompression?: number;\n    /**\n     * The output format of the generated image.\n     * @default \"png\"\n     */\n    outputFormat?: \"png\" | \"webp\" | \"jpeg\";\n    /**\n     * Number of partial images to generate in streaming mode (0-3).\n     * When set, the model will return partial images as they are generated,\n     * providing faster visual feedback.\n     * @default 0\n     */\n    partialImages?: number;\n    /**\n     * The quality of the generated image.\n     * - `low`: Faster generation, lower quality\n     * - `medium`: Balanced generation time and quality\n     * - `high`: Slower generation, higher quality\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    quality?: \"low\" | \"medium\" | \"high\" | \"auto\";\n    /**\n     * The size of the generated image.\n     * - `1024x1024`: Square format\n     * - `1024x1536`: Portrait format\n     * - `1536x1024`: Landscape format\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    size?: \"1024x1024\" | \"1024x1536\" | \"1536x1024\" | \"auto\";\n}\n/**\n * OpenAI Image Generation tool type for the Responses API.\n */\nexport type ImageGenerationTool = OpenAIClient.Responses.Tool.ImageGeneration;\n/**\n * Creates an Image Generation tool that allows models to generate or edit images\n * using text prompts and optional image inputs.\n *\n * The image generation tool leverages the GPT Image model and automatically\n * optimizes text inputs for improved performance. When included in a request,\n * the model can decide when and how to generate images as part of the conversation.\n *\n * **Key Features**:\n * - Generate images from text descriptions\n * - Edit existing images with text instructions\n * - Multi-turn image editing by referencing previous responses\n * - Configurable output options (size, quality, format)\n * - Streaming support for partial image generation\n *\n * **Prompting Tips**:\n * - Use terms like \"draw\" or \"edit\" in your prompt for best results\n * - For combining images, say \"edit the first image by adding this element\" instead of \"combine\"\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-image-generation | OpenAI Image Generation Documentation}\n *\n * @param options - Configuration options for the Image Generation tool\n * @returns An Image Generation tool definition to be passed to the OpenAI Responses API\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"gpt-4o\" });\n *\n * // Basic usage - generate an image\n * const response = await model.invoke(\n *   \"Generate an image of a gray tabby cat hugging an otter with an orange scarf\",\n *   { tools: [tools.imageGeneration()] }\n * );\n *\n * // Access the generated image\n * const imageData = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"image_generation_call\"\n * );\n * if (imageData?.result) {\n *   // imageData.result contains the base64-encoded image\n *   const fs = await import(\"fs\");\n *   fs.writeFileSync(\"output.png\", Buffer.from(imageData.result, \"base64\"));\n * }\n *\n * // With custom options\n * const response = await model.invoke(\n *   \"Draw a beautiful sunset over mountains\",\n *   {\n *     tools: [tools.imageGeneration({\n *       size: \"1536x1024\",      // Landscape format\n *       quality: \"high\",        // Higher quality output\n *       outputFormat: \"jpeg\",   // JPEG format\n *       outputCompression: 90,  // 90% compression\n *     })]\n *   }\n * );\n *\n * // With transparent background\n * const response = await model.invoke(\n *   \"Create a logo with a transparent background\",\n *   {\n *     tools: [tools.imageGeneration({\n *       background: \"transparent\",\n *       outputFormat: \"png\",\n *     })]\n *   }\n * );\n *\n * // Force the model to use image generation\n * const response = await model.invoke(\n *   \"A serene lake at dawn\",\n *   {\n *     tools: [tools.imageGeneration()],\n *     tool_choice: { type: \"image_generation\" },\n *   }\n * );\n *\n * // Enable streaming with partial images\n * const response = await model.invoke(\n *   \"Draw a detailed fantasy castle\",\n *   {\n *     tools: [tools.imageGeneration({\n *       partialImages: 2,  // Get 2 partial images during generation\n *     })]\n *   }\n * );\n * ```\n *\n * @remarks\n * - Supported models: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o3\n * - The image generation process always uses `gpt-image-1` model internally\n * - The model will automatically revise prompts for improved performance\n * - Access the revised prompt via `revised_prompt` field in the output\n * - Multi-turn editing is supported by passing previous response messages\n */\nexport declare function imageGeneration(options?: ImageGenerationOptions): ServerTool;\n//# sourceMappingURL=imageGeneration.d.ts.map"],"mappings":";;;;;;;AAMA;AAaA;AA2EYK,UAxFKF,wBAAAA,CAwFiBF;EAkGVQ;;;;;;;;;;;;UA7KPL,sBAAAA;;;;;;;;;;;;;;;;;;;;;;mBAsBID;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAqDTE,mBAAAA,GAAsBJ,QAAAA,CAAaK,SAAAA,CAAUC,IAAAA,CAAKC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAkGtCC,eAAAA,WAA0BL,yBAAyBF"}
+{"version":3,"file":"imageGeneration.d.cts","names":["OpenAI","OpenAIClient","ServerTool","ImageGenerationInputMask","ImageGenerationOptions","ImageGenerationTool","Responses","Tool","ImageGeneration","imageGeneration"],"sources":["../../src/tools/imageGeneration.d.ts"],"sourcesContent":["import { OpenAI as OpenAIClient } from \"openai\";\nimport type { ServerTool } from \"@langchain/core/tools\";\n/**\n * Optional mask for inpainting. Allows you to specify areas of the image\n * that should be regenerated.\n */\nexport interface ImageGenerationInputMask {\n    /**\n     * Base64-encoded mask image URL.\n     */\n    imageUrl?: string;\n    /**\n     * File ID for the mask image (uploaded via OpenAI File API).\n     */\n    fileId?: string;\n}\n/**\n * Options for the Image Generation tool.\n */\nexport interface ImageGenerationOptions {\n    /**\n     * Whether to generate a new image or edit an existing image.\n     * - `generate`: Generate a new image from scratch\n     * - `edit`: Edit an existing image\n     * - `auto`: Let the model decide based on the input\n     * @default \"auto\"\n     */\n    action?: \"generate\" | \"edit\" | \"auto\";\n    /**\n     * Background type for the generated image.\n     * - `transparent`: Generate image with transparent background\n     * - `opaque`: Generate image with opaque background\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    background?: \"transparent\" | \"opaque\" | \"auto\";\n    /**\n     * Control how much effort the model will exert to match the style and features,\n     * especially facial features, of input images. This parameter is only supported\n     * for `gpt-image-1`. Unsupported for `gpt-image-1-mini`.\n     * - `high`: Higher fidelity to input images\n     * - `low`: Lower fidelity to input images\n     * @default \"low\"\n     */\n    inputFidelity?: \"high\" | \"low\";\n    /**\n     * Optional mask for inpainting. Use this to specify areas of an image\n     * that should be regenerated.\n     */\n    inputImageMask?: ImageGenerationInputMask;\n    /**\n     * The image generation model to use.\n     * @default \"gpt-image-1\"\n     */\n    model?: \"gpt-image-1\" | \"gpt-image-1-mini\" | \"gpt-image-1.5\";\n    /**\n     * Moderation level for the generated image.\n     * - `auto`: Standard moderation\n     * - `low`: Less restrictive moderation\n     * @default \"auto\"\n     */\n    moderation?: \"auto\" | \"low\";\n    /**\n     * Compression level for the output image (0-100).\n     * Only applies to JPEG and WebP formats.\n     * @default 100\n     */\n    outputCompression?: number;\n    /**\n     * The output format of the generated image.\n     * @default \"png\"\n     */\n    outputFormat?: \"png\" | \"webp\" | \"jpeg\";\n    /**\n     * Number of partial images to generate in streaming mode (0-3).\n     * When set, the model will return partial images as they are generated,\n     * providing faster visual feedback.\n     * @default 0\n     */\n    partialImages?: number;\n    /**\n     * The quality of the generated image.\n     * - `low`: Faster generation, lower quality\n     * - `medium`: Balanced generation time and quality\n     * - `high`: Slower generation, higher quality\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    quality?: \"low\" | \"medium\" | \"high\" | \"auto\";\n    /**\n     * The size of the generated image.\n     * - `1024x1024`: Square format\n     * - `1024x1536`: Portrait format\n     * - `1536x1024`: Landscape format\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    size?: \"1024x1024\" | \"1024x1536\" | \"1536x1024\" | \"auto\";\n}\n/**\n * OpenAI Image Generation tool type for the Responses API.\n */\nexport type ImageGenerationTool = OpenAIClient.Responses.Tool.ImageGeneration;\n/**\n * Creates an Image Generation tool that allows models to generate or edit images\n * using text prompts and optional image inputs.\n *\n * The image generation tool leverages the GPT Image model and automatically\n * optimizes text inputs for improved performance. When included in a request,\n * the model can decide when and how to generate images as part of the conversation.\n *\n * **Key Features**:\n * - Generate images from text descriptions\n * - Edit existing images with text instructions\n * - Multi-turn image editing by referencing previous responses\n * - Configurable output options (size, quality, format)\n * - Streaming support for partial image generation\n *\n * **Prompting Tips**:\n * - Use terms like \"draw\" or \"edit\" in your prompt for best results\n * - For combining images, say \"edit the first image by adding this element\" instead of \"combine\"\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-image-generation | OpenAI Image Generation Documentation}\n *\n * @param options - Configuration options for the Image Generation tool\n * @returns An Image Generation tool definition to be passed to the OpenAI Responses API\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"gpt-4o\" });\n *\n * // Basic usage - generate an image\n * const response = await model.invoke(\n *   \"Generate an image of a gray tabby cat hugging an otter with an orange scarf\",\n *   { tools: [tools.imageGeneration()] }\n * );\n *\n * // Access the generated image\n * const imageData = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"image_generation_call\"\n * );\n * if (imageData?.result) {\n *   // imageData.result contains the base64-encoded image\n *   const fs = await import(\"fs\");\n *   fs.writeFileSync(\"output.png\", Buffer.from(imageData.result, \"base64\"));\n * }\n *\n * // With custom options\n * const response = await model.invoke(\n *   \"Draw a beautiful sunset over mountains\",\n *   {\n *     tools: [tools.imageGeneration({\n *       size: \"1536x1024\",      // Landscape format\n *       quality: \"high\",        // Higher quality output\n *       outputFormat: \"jpeg\",   // JPEG format\n *       outputCompression: 90,  // 90% compression\n *     })]\n *   }\n * );\n *\n * // With transparent background\n * const response = await model.invoke(\n *   \"Create a logo with a transparent background\",\n *   {\n *     tools: [tools.imageGeneration({\n *       background: \"transparent\",\n *       outputFormat: \"png\",\n *     })]\n *   }\n * );\n *\n * // Force the model to use image generation\n * const response = await model.invoke(\n *   \"A serene lake at dawn\",\n *   {\n *     tools: [tools.imageGeneration()],\n *     tool_choice: { type: \"image_generation\" },\n *   }\n * );\n *\n * // Enable streaming with partial images\n * const response = await model.invoke(\n *   \"Draw a detailed fantasy castle\",\n *   {\n *     tools: [tools.imageGeneration({\n *       partialImages: 2,  // Get 2 partial images during generation\n *     })]\n *   }\n * );\n * ```\n *\n * @remarks\n * - Supported models: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o3\n * - The image generation process always uses `gpt-image-1` model internally\n * - The model will automatically revise prompts for improved performance\n * - Access the revised prompt via `revised_prompt` field in the output\n * - Multi-turn editing is supported by passing previous response messages\n */\nexport declare function imageGeneration(options?: ImageGenerationOptions): ServerTool;\n//# sourceMappingURL=imageGeneration.d.ts.map"],"mappings":";;;;;;;AAMA;AAaA;AAmFYK,UAhGKF,wBAAAA,CAgGiBF;EAkGVQ;;;;;;;;;;;;UArLPL,sBAAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;mBA8BID;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAqDTE,mBAAAA,GAAsBJ,QAAAA,CAAaK,SAAAA,CAAUC,IAAAA,CAAKC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAkGtCC,eAAAA,WAA0BL,yBAAyBF"}

package/dist/tools/imageGeneration.d.ts

@@ -21,6 +21,14 @@ interface ImageGenerationInputMask {
  * Options for the Image Generation tool.
  */
 interface ImageGenerationOptions {
+  /**
+   * Whether to generate a new image or edit an existing image.
+   * - `generate`: Generate a new image from scratch
+   * - `edit`: Edit an existing image
+   * - `auto`: Let the model decide based on the input
+   * @default "auto"
+   */
+  action?: "generate" | "edit" | "auto";
   /**
    * Background type for the generated image.
    * - `transparent`: Generate image with transparent background

package/dist/tools/imageGeneration.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"imageGeneration.d.ts","names":["OpenAI","OpenAIClient","ServerTool","ImageGenerationInputMask","ImageGenerationOptions","ImageGenerationTool","Responses","Tool","ImageGeneration","imageGeneration"],"sources":["../../src/tools/imageGeneration.d.ts"],"sourcesContent":["import { OpenAI as OpenAIClient } from \"openai\";\nimport type { ServerTool } from \"@langchain/core/tools\";\n/**\n * Optional mask for inpainting. Allows you to specify areas of the image\n * that should be regenerated.\n */\nexport interface ImageGenerationInputMask {\n    /**\n     * Base64-encoded mask image URL.\n     */\n    imageUrl?: string;\n    /**\n     * File ID for the mask image (uploaded via OpenAI File API).\n     */\n    fileId?: string;\n}\n/**\n * Options for the Image Generation tool.\n */\nexport interface ImageGenerationOptions {\n    /**\n     * Background type for the generated image.\n     * - `transparent`: Generate image with transparent background\n     * - `opaque`: Generate image with opaque background\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    background?: \"transparent\" | \"opaque\" | \"auto\";\n    /**\n     * Control how much effort the model will exert to match the style and features,\n     * especially facial features, of input images. This parameter is only supported\n     * for `gpt-image-1`. Unsupported for `gpt-image-1-mini`.\n     * - `high`: Higher fidelity to input images\n     * - `low`: Lower fidelity to input images\n     * @default \"low\"\n     */\n    inputFidelity?: \"high\" | \"low\";\n    /**\n     * Optional mask for inpainting. Use this to specify areas of an image\n     * that should be regenerated.\n     */\n    inputImageMask?: ImageGenerationInputMask;\n    /**\n     * The image generation model to use.\n     * @default \"gpt-image-1\"\n     */\n    model?: \"gpt-image-1\" | \"gpt-image-1-mini\" | \"gpt-image-1.5\";\n    /**\n     * Moderation level for the generated image.\n     * - `auto`: Standard moderation\n     * - `low`: Less restrictive moderation\n     * @default \"auto\"\n     */\n    moderation?: \"auto\" | \"low\";\n    /**\n     * Compression level for the output image (0-100).\n     * Only applies to JPEG and WebP formats.\n     * @default 100\n     */\n    outputCompression?: number;\n    /**\n     * The output format of the generated image.\n     * @default \"png\"\n     */\n    outputFormat?: \"png\" | \"webp\" | \"jpeg\";\n    /**\n     * Number of partial images to generate in streaming mode (0-3).\n     * When set, the model will return partial images as they are generated,\n     * providing faster visual feedback.\n     * @default 0\n     */\n    partialImages?: number;\n    /**\n     * The quality of the generated image.\n     * - `low`: Faster generation, lower quality\n     * - `medium`: Balanced generation time and quality\n     * - `high`: Slower generation, higher quality\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    quality?: \"low\" | \"medium\" | \"high\" | \"auto\";\n    /**\n     * The size of the generated image.\n     * - `1024x1024`: Square format\n     * - `1024x1536`: Portrait format\n     * - `1536x1024`: Landscape format\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    size?: \"1024x1024\" | \"1024x1536\" | \"1536x1024\" | \"auto\";\n}\n/**\n * OpenAI Image Generation tool type for the Responses API.\n */\nexport type ImageGenerationTool = OpenAIClient.Responses.Tool.ImageGeneration;\n/**\n * Creates an Image Generation tool that allows models to generate or edit images\n * using text prompts and optional image inputs.\n *\n * The image generation tool leverages the GPT Image model and automatically\n * optimizes text inputs for improved performance. When included in a request,\n * the model can decide when and how to generate images as part of the conversation.\n *\n * **Key Features**:\n * - Generate images from text descriptions\n * - Edit existing images with text instructions\n * - Multi-turn image editing by referencing previous responses\n * - Configurable output options (size, quality, format)\n * - Streaming support for partial image generation\n *\n * **Prompting Tips**:\n * - Use terms like \"draw\" or \"edit\" in your prompt for best results\n * - For combining images, say \"edit the first image by adding this element\" instead of \"combine\"\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-image-generation | OpenAI Image Generation Documentation}\n *\n * @param options - Configuration options for the Image Generation tool\n * @returns An Image Generation tool definition to be passed to the OpenAI Responses API\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"gpt-4o\" });\n *\n * // Basic usage - generate an image\n * const response = await model.invoke(\n *   \"Generate an image of a gray tabby cat hugging an otter with an orange scarf\",\n *   { tools: [tools.imageGeneration()] }\n * );\n *\n * // Access the generated image\n * const imageData = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"image_generation_call\"\n * );\n * if (imageData?.result) {\n *   // imageData.result contains the base64-encoded image\n *   const fs = await import(\"fs\");\n *   fs.writeFileSync(\"output.png\", Buffer.from(imageData.result, \"base64\"));\n * }\n *\n * // With custom options\n * const response = await model.invoke(\n *   \"Draw a beautiful sunset over mountains\",\n *   {\n *     tools: [tools.imageGeneration({\n *       size: \"1536x1024\",      // Landscape format\n *       quality: \"high\",        // Higher quality output\n *       outputFormat: \"jpeg\",   // JPEG format\n *       outputCompression: 90,  // 90% compression\n *     })]\n *   }\n * );\n *\n * // With transparent background\n * const response = await model.invoke(\n *   \"Create a logo with a transparent background\",\n *   {\n *     tools: [tools.imageGeneration({\n *       background: \"transparent\",\n *       outputFormat: \"png\",\n *     })]\n *   }\n * );\n *\n * // Force the model to use image generation\n * const response = await model.invoke(\n *   \"A serene lake at dawn\",\n *   {\n *     tools: [tools.imageGeneration()],\n *     tool_choice: { type: \"image_generation\" },\n *   }\n * );\n *\n * // Enable streaming with partial images\n * const response = await model.invoke(\n *   \"Draw a detailed fantasy castle\",\n *   {\n *     tools: [tools.imageGeneration({\n *       partialImages: 2,  // Get 2 partial images during generation\n *     })]\n *   }\n * );\n * ```\n *\n * @remarks\n * - Supported models: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o3\n * - The image generation process always uses `gpt-image-1` model internally\n * - The model will automatically revise prompts for improved performance\n * - Access the revised prompt via `revised_prompt` field in the output\n * - Multi-turn editing is supported by passing previous response messages\n */\nexport declare function imageGeneration(options?: ImageGenerationOptions): ServerTool;\n//# sourceMappingURL=imageGeneration.d.ts.map"],"mappings":";;;;;;;AAMA;AAaA;AA2EYK,UAxFKF,wBAAAA,CAwFiBF;EAkGVQ;;;;;;;;;;;;UA7KPL,sBAAAA;;;;;;;;;;;;;;;;;;;;;;mBAsBID;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAqDTE,mBAAAA,GAAsBJ,QAAAA,CAAaK,SAAAA,CAAUC,IAAAA,CAAKC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAkGtCC,eAAAA,WAA0BL,yBAAyBF"}
+{"version":3,"file":"imageGeneration.d.ts","names":["OpenAI","OpenAIClient","ServerTool","ImageGenerationInputMask","ImageGenerationOptions","ImageGenerationTool","Responses","Tool","ImageGeneration","imageGeneration"],"sources":["../../src/tools/imageGeneration.d.ts"],"sourcesContent":["import { OpenAI as OpenAIClient } from \"openai\";\nimport type { ServerTool } from \"@langchain/core/tools\";\n/**\n * Optional mask for inpainting. Allows you to specify areas of the image\n * that should be regenerated.\n */\nexport interface ImageGenerationInputMask {\n    /**\n     * Base64-encoded mask image URL.\n     */\n    imageUrl?: string;\n    /**\n     * File ID for the mask image (uploaded via OpenAI File API).\n     */\n    fileId?: string;\n}\n/**\n * Options for the Image Generation tool.\n */\nexport interface ImageGenerationOptions {\n    /**\n     * Whether to generate a new image or edit an existing image.\n     * - `generate`: Generate a new image from scratch\n     * - `edit`: Edit an existing image\n     * - `auto`: Let the model decide based on the input\n     * @default \"auto\"\n     */\n    action?: \"generate\" | \"edit\" | \"auto\";\n    /**\n     * Background type for the generated image.\n     * - `transparent`: Generate image with transparent background\n     * - `opaque`: Generate image with opaque background\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    background?: \"transparent\" | \"opaque\" | \"auto\";\n    /**\n     * Control how much effort the model will exert to match the style and features,\n     * especially facial features, of input images. This parameter is only supported\n     * for `gpt-image-1`. Unsupported for `gpt-image-1-mini`.\n     * - `high`: Higher fidelity to input images\n     * - `low`: Lower fidelity to input images\n     * @default \"low\"\n     */\n    inputFidelity?: \"high\" | \"low\";\n    /**\n     * Optional mask for inpainting. Use this to specify areas of an image\n     * that should be regenerated.\n     */\n    inputImageMask?: ImageGenerationInputMask;\n    /**\n     * The image generation model to use.\n     * @default \"gpt-image-1\"\n     */\n    model?: \"gpt-image-1\" | \"gpt-image-1-mini\" | \"gpt-image-1.5\";\n    /**\n     * Moderation level for the generated image.\n     * - `auto`: Standard moderation\n     * - `low`: Less restrictive moderation\n     * @default \"auto\"\n     */\n    moderation?: \"auto\" | \"low\";\n    /**\n     * Compression level for the output image (0-100).\n     * Only applies to JPEG and WebP formats.\n     * @default 100\n     */\n    outputCompression?: number;\n    /**\n     * The output format of the generated image.\n     * @default \"png\"\n     */\n    outputFormat?: \"png\" | \"webp\" | \"jpeg\";\n    /**\n     * Number of partial images to generate in streaming mode (0-3).\n     * When set, the model will return partial images as they are generated,\n     * providing faster visual feedback.\n     * @default 0\n     */\n    partialImages?: number;\n    /**\n     * The quality of the generated image.\n     * - `low`: Faster generation, lower quality\n     * - `medium`: Balanced generation time and quality\n     * - `high`: Slower generation, higher quality\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    quality?: \"low\" | \"medium\" | \"high\" | \"auto\";\n    /**\n     * The size of the generated image.\n     * - `1024x1024`: Square format\n     * - `1024x1536`: Portrait format\n     * - `1536x1024`: Landscape format\n     * - `auto`: Let the model decide based on the prompt\n     * @default \"auto\"\n     */\n    size?: \"1024x1024\" | \"1024x1536\" | \"1536x1024\" | \"auto\";\n}\n/**\n * OpenAI Image Generation tool type for the Responses API.\n */\nexport type ImageGenerationTool = OpenAIClient.Responses.Tool.ImageGeneration;\n/**\n * Creates an Image Generation tool that allows models to generate or edit images\n * using text prompts and optional image inputs.\n *\n * The image generation tool leverages the GPT Image model and automatically\n * optimizes text inputs for improved performance. When included in a request,\n * the model can decide when and how to generate images as part of the conversation.\n *\n * **Key Features**:\n * - Generate images from text descriptions\n * - Edit existing images with text instructions\n * - Multi-turn image editing by referencing previous responses\n * - Configurable output options (size, quality, format)\n * - Streaming support for partial image generation\n *\n * **Prompting Tips**:\n * - Use terms like \"draw\" or \"edit\" in your prompt for best results\n * - For combining images, say \"edit the first image by adding this element\" instead of \"combine\"\n *\n * @see {@link https://platform.openai.com/docs/guides/tools-image-generation | OpenAI Image Generation Documentation}\n *\n * @param options - Configuration options for the Image Generation tool\n * @returns An Image Generation tool definition to be passed to the OpenAI Responses API\n *\n * @example\n * ```typescript\n * import { ChatOpenAI, tools } from \"@langchain/openai\";\n *\n * const model = new ChatOpenAI({ model: \"gpt-4o\" });\n *\n * // Basic usage - generate an image\n * const response = await model.invoke(\n *   \"Generate an image of a gray tabby cat hugging an otter with an orange scarf\",\n *   { tools: [tools.imageGeneration()] }\n * );\n *\n * // Access the generated image\n * const imageData = response.additional_kwargs.tool_outputs?.find(\n *   (output) => output.type === \"image_generation_call\"\n * );\n * if (imageData?.result) {\n *   // imageData.result contains the base64-encoded image\n *   const fs = await import(\"fs\");\n *   fs.writeFileSync(\"output.png\", Buffer.from(imageData.result, \"base64\"));\n * }\n *\n * // With custom options\n * const response = await model.invoke(\n *   \"Draw a beautiful sunset over mountains\",\n *   {\n *     tools: [tools.imageGeneration({\n *       size: \"1536x1024\",      // Landscape format\n *       quality: \"high\",        // Higher quality output\n *       outputFormat: \"jpeg\",   // JPEG format\n *       outputCompression: 90,  // 90% compression\n *     })]\n *   }\n * );\n *\n * // With transparent background\n * const response = await model.invoke(\n *   \"Create a logo with a transparent background\",\n *   {\n *     tools: [tools.imageGeneration({\n *       background: \"transparent\",\n *       outputFormat: \"png\",\n *     })]\n *   }\n * );\n *\n * // Force the model to use image generation\n * const response = await model.invoke(\n *   \"A serene lake at dawn\",\n *   {\n *     tools: [tools.imageGeneration()],\n *     tool_choice: { type: \"image_generation\" },\n *   }\n * );\n *\n * // Enable streaming with partial images\n * const response = await model.invoke(\n *   \"Draw a detailed fantasy castle\",\n *   {\n *     tools: [tools.imageGeneration({\n *       partialImages: 2,  // Get 2 partial images during generation\n *     })]\n *   }\n * );\n * ```\n *\n * @remarks\n * - Supported models: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o3\n * - The image generation process always uses `gpt-image-1` model internally\n * - The model will automatically revise prompts for improved performance\n * - Access the revised prompt via `revised_prompt` field in the output\n * - Multi-turn editing is supported by passing previous response messages\n */\nexport declare function imageGeneration(options?: ImageGenerationOptions): ServerTool;\n//# sourceMappingURL=imageGeneration.d.ts.map"],"mappings":";;;;;;;AAMA;AAaA;AAmFYK,UAhGKF,wBAAAA,CAgGiBF;EAkGVQ;;;;;;;;;;;;UArLPL,sBAAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;mBA8BID;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAqDTE,mBAAAA,GAAsBJ,QAAAA,CAAaK,SAAAA,CAAUC,IAAAA,CAAKC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAkGtCC,eAAAA,WAA0BL,yBAAyBF"}

package/dist/tools/imageGeneration.js

@@ -109,6 +109,7 @@ function convertInputImageMask(mask) {
 function imageGeneration(options) {
 	return {
 		type: "image_generation",
+		action: options?.action,
 		background: options?.background,
 		input_fidelity: options?.inputFidelity,
 		input_image_mask: convertInputImageMask(options?.inputImageMask),