@midscene/core 1.3.7 → 1.3.8-beta-20260206024209.0

package/dist/es/ai-model/inspect.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"ai-model/inspect.mjs","sources":["../../../src/ai-model/inspect.ts"],"sourcesContent":["import type {\n  AIDataExtractionResponse,\n  AIElementResponse,\n  AISectionLocatorResponse,\n  AIUsageInfo,\n  Rect,\n  ServiceExtractOption,\n  UIContext,\n} from '@/types';\nimport type { IModelConfig } from '@midscene/shared/env';\nimport {\n  generateElementByPoint,\n  generateElementByRect,\n} from '@midscene/shared/extractor/dom-util';\nimport {\n  cropByRect,\n  paddingToMatchBlockByBase64,\n  preProcessImageUrl,\n  scaleImage,\n} from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport type { LocateResultElement } from '@midscene/shared/types';\nimport { assert } from '@midscene/shared/utils';\nimport type {\n  ChatCompletionSystemMessageParam,\n  ChatCompletionUserMessageParam,\n} from 'openai/resources/index';\nimport type { TMultimodalPrompt, TUserPrompt } from '../common';\nimport { adaptBboxToRect, expandSearchArea, mergeRects } from '../common';\nimport { parseAutoGLMLocateResponse } from './auto-glm/parser';\nimport { getAutoGLMLocatePrompt } from './auto-glm/prompt';\nimport { isAutoGLM } from './auto-glm/util';\nimport {\n  extractDataQueryPrompt,\n  parseXMLExtractionResponse,\n  systemPromptToExtract,\n} from './prompt/extraction';\nimport {\n  findElementPrompt,\n  systemPromptToLocateElement,\n} from './prompt/llm-locator';\nimport {\n  sectionLocatorInstruction,\n  systemPromptToLocateSection,\n} from './prompt/llm-section-locator';\nimport {\n  orderSensitiveJudgePrompt,\n  systemPromptToJudgeOrderSensitive,\n} from './prompt/order-sensitive-judge';\nimport {\n  AIResponseParseError,\n  callAI,\n  callAIWithObjectResponse,\n  callAIWithStringResponse,\n} from './service-caller/index';\n\nexport type AIArgs = [\n  ChatCompletionSystemMessageParam,\n  ...ChatCompletionUserMessageParam[],\n];\n\nconst debugInspect = getDebug('ai:inspect');\nconst debugSection = getDebug('ai:section');\n\nconst extraTextFromUserPrompt = (prompt: TUserPrompt): string => {\n  if (typeof prompt === 'string') {\n    return prompt;\n  } else {\n    return prompt.prompt;\n  }\n};\n\nconst promptsToChatParam = async (\n  multimodalPrompt: TMultimodalPrompt,\n): Promise<ChatCompletionUserMessageParam[]> => {\n  const msgs: ChatCompletionUserMessageParam[] = [];\n  if (multimodalPrompt?.images?.length) {\n    msgs.push({\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: 'Next, I will provide all the reference images.',\n        },\n      ],\n    });\n\n    for (const item of multimodalPrompt.images) {\n      const base64 = await preProcessImageUrl(\n        item.url,\n        !!multimodalPrompt.convertHttpImage2Base64,\n      );\n\n      msgs.push({\n        role: 'user',\n        content: [\n          {\n            type: 'text',\n            text: `this is the reference image named '${item.name}':`,\n          },\n        ],\n      });\n\n      msgs.push({\n        role: 'user',\n        content: [\n          {\n            type: 'image_url',\n            image_url: {\n              url: base64,\n              detail: 'high',\n            },\n          },\n        ],\n      });\n    }\n  }\n  return msgs;\n};\n\nexport async function AiLocateElement(options: {\n  context: UIContext;\n  targetElementDescription: TUserPrompt;\n  callAIFn: typeof callAIWithObjectResponse<\n    AIElementResponse | [number, number]\n  >;\n  searchConfig?: Awaited<ReturnType<typeof AiLocateSection>>;\n  modelConfig: IModelConfig;\n}): Promise<{\n  parseResult: {\n    elements: LocateResultElement[];\n    errors?: string[];\n  };\n  rect?: Rect;\n  rawResponse: string;\n  usage?: AIUsageInfo;\n  reasoning_content?: string;\n}> {\n  const { context, targetElementDescription, callAIFn, modelConfig } = options;\n  const { modelFamily } = modelConfig;\n  const screenshotBase64 = context.screenshot.base64;\n\n  assert(\n    targetElementDescription,\n    'cannot find the target element description',\n  );\n  const targetElementDescriptionText = extraTextFromUserPrompt(\n    targetElementDescription,\n  );\n  const userInstructionPrompt = findElementPrompt(targetElementDescriptionText);\n  const systemPrompt = isAutoGLM(modelFamily)\n    ? getAutoGLMLocatePrompt(modelFamily)\n    : systemPromptToLocateElement(modelFamily);\n\n  let imagePayload = screenshotBase64;\n  let imageWidth = context.size.width;\n  let imageHeight = context.size.height;\n  let originalImageWidth = imageWidth;\n  let originalImageHeight = imageHeight;\n\n  if (options.searchConfig) {\n    assert(\n      options.searchConfig.rect,\n      'searchArea is provided but its rect cannot be found. Failed to locate element',\n    );\n    assert(\n      options.searchConfig.imageBase64,\n      'searchArea is provided but its imageBase64 cannot be found. Failed to locate element',\n    );\n\n    imagePayload = options.searchConfig.imageBase64;\n    imageWidth = options.searchConfig.rect?.width;\n    imageHeight = options.searchConfig.rect?.height;\n    originalImageWidth = imageWidth;\n    originalImageHeight = imageHeight;\n  } else if (modelFamily === 'qwen2.5-vl') {\n    const paddedResult = await paddingToMatchBlockByBase64(imagePayload);\n    imageWidth = paddedResult.width;\n    imageHeight = paddedResult.height;\n    imagePayload = paddedResult.imageBase64;\n  }\n\n  const msgs: AIArgs = [\n    { role: 'system', content: systemPrompt },\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n        {\n          type: 'text',\n          text: isAutoGLM(modelFamily)\n            ? `Tap: ${userInstructionPrompt}`\n            : userInstructionPrompt,\n        },\n      ],\n    },\n  ];\n\n  if (typeof targetElementDescription !== 'string') {\n    const addOns = await promptsToChatParam({\n      images: targetElementDescription.images,\n      convertHttpImage2Base64: targetElementDescription.convertHttpImage2Base64,\n    });\n    msgs.push(...addOns);\n  }\n\n  if (isAutoGLM(modelFamily)) {\n    const { content: rawResponseContent, usage } =\n      await callAIWithStringResponse(msgs, modelConfig);\n\n    debugInspect('auto-glm rawResponse:', rawResponseContent);\n\n    const parsed = parseAutoGLMLocateResponse(rawResponseContent);\n\n    debugInspect('auto-glm thinking:', parsed.think);\n    debugInspect('auto-glm coordinates:', parsed.coordinates);\n\n    let resRect: Rect | undefined;\n    let matchedElements: LocateResultElement[] = [];\n    let errors: string[] = [];\n\n    if (parsed.error || !parsed.coordinates) {\n      errors = [parsed.error || 'Failed to parse auto-glm response'];\n      debugInspect('auto-glm parse error:', errors[0]);\n    } else {\n      const { x, y } = parsed.coordinates;\n\n      debugInspect('auto-glm coordinates [0-999]:', { x, y });\n\n      // Convert auto-glm coordinates [0,999] to pixel bbox\n      // Map from [0,999] to pixel coordinates\n      const pixelX = Math.round((x * imageWidth) / 1000);\n      const pixelY = Math.round((y * imageHeight) / 1000);\n\n      debugInspect('auto-glm pixel coordinates:', { pixelX, pixelY });\n\n      // Apply offset if searching in a cropped area\n      let finalX = pixelX;\n      let finalY = pixelY;\n      if (options.searchConfig?.rect) {\n        finalX += options.searchConfig.rect.left;\n        finalY += options.searchConfig.rect.top;\n      }\n\n      const element: LocateResultElement = generateElementByPoint(\n        [finalX, finalY],\n        targetElementDescriptionText as string,\n      );\n\n      resRect = element.rect;\n      debugInspect('auto-glm resRect:', resRect);\n\n      if (element) {\n        matchedElements = [element];\n      }\n    }\n\n    return {\n      rect: resRect,\n      parseResult: {\n        elements: matchedElements,\n        errors,\n      },\n      rawResponse: rawResponseContent,\n      usage,\n      reasoning_content: parsed.think,\n    };\n  }\n\n  let res: Awaited<ReturnType<typeof callAIFn>>;\n  try {\n    res = await callAIFn(msgs, modelConfig);\n  } catch (callError) {\n    // Return error with usage and rawResponse if available\n    const errorMessage =\n      callError instanceof Error ? callError.message : String(callError);\n    const rawResponse =\n      callError instanceof AIResponseParseError\n        ? callError.rawResponse\n        : errorMessage;\n    const usage =\n      callError instanceof AIResponseParseError ? callError.usage : undefined;\n    return {\n      rect: undefined,\n      parseResult: {\n        elements: [],\n        errors: [`AI call error: ${errorMessage}`],\n      },\n      rawResponse,\n      usage,\n      reasoning_content: undefined,\n    };\n  }\n\n  const rawResponse = JSON.stringify(res.content);\n\n  let resRect: Rect | undefined;\n  let matchedElements: LocateResultElement[] = [];\n  let errors: string[] | undefined =\n    'errors' in res.content ? res.content.errors : [];\n  try {\n    if (\n      'bbox' in res.content &&\n      Array.isArray(res.content.bbox) &&\n      res.content.bbox.length >= 1\n    ) {\n      resRect = adaptBboxToRect(\n        res.content.bbox,\n        imageWidth,\n        imageHeight,\n        options.searchConfig?.rect?.left,\n        options.searchConfig?.rect?.top,\n        originalImageWidth,\n        originalImageHeight,\n        modelFamily,\n        options.searchConfig?.scale,\n      );\n\n      debugInspect('resRect', resRect);\n\n      const element: LocateResultElement = generateElementByRect(\n        resRect,\n        targetElementDescriptionText as string,\n      );\n      errors = [];\n\n      if (element) {\n        matchedElements = [element];\n      }\n    }\n  } catch (e) {\n    const msg =\n      e instanceof Error\n        ? `Failed to parse bbox: ${e.message}`\n        : 'unknown error in locate';\n    if (!errors || errors?.length === 0) {\n      errors = [msg];\n    } else {\n      errors.push(`(${msg})`);\n    }\n  }\n\n  return {\n    rect: resRect,\n    parseResult: {\n      elements: matchedElements as LocateResultElement[],\n      errors: errors as string[],\n    },\n    rawResponse,\n    usage: res.usage,\n    reasoning_content: res.reasoning_content,\n  };\n}\n\nexport async function AiLocateSection(options: {\n  context: UIContext;\n  sectionDescription: TUserPrompt;\n  modelConfig: IModelConfig;\n}): Promise<{\n  rect?: Rect;\n  imageBase64?: string;\n  scale?: number;\n  error?: string;\n  rawResponse: string;\n  usage?: AIUsageInfo;\n}> {\n  const { context, sectionDescription, modelConfig } = options;\n  const { modelFamily } = modelConfig;\n  const screenshotBase64 = context.screenshot.base64;\n\n  const systemPrompt = systemPromptToLocateSection(modelFamily);\n  const sectionLocatorInstructionText = sectionLocatorInstruction(\n    extraTextFromUserPrompt(sectionDescription),\n  );\n  const msgs: AIArgs = [\n    { role: 'system', content: systemPrompt },\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'image_url',\n          image_url: {\n            url: screenshotBase64,\n            detail: 'high',\n          },\n        },\n        {\n          type: 'text',\n          text: sectionLocatorInstructionText,\n        },\n      ],\n    },\n  ];\n\n  if (typeof sectionDescription !== 'string') {\n    const addOns = await promptsToChatParam({\n      images: sectionDescription.images,\n      convertHttpImage2Base64: sectionDescription.convertHttpImage2Base64,\n    });\n    msgs.push(...addOns);\n  }\n\n  let result: Awaited<\n    ReturnType<typeof callAIWithObjectResponse<AISectionLocatorResponse>>\n  >;\n  try {\n    result = await callAIWithObjectResponse<AISectionLocatorResponse>(\n      msgs,\n      modelConfig,\n    );\n  } catch (callError) {\n    // Return error with usage and rawResponse if available\n    const errorMessage =\n      callError instanceof Error ? callError.message : String(callError);\n    const rawResponse =\n      callError instanceof AIResponseParseError\n        ? callError.rawResponse\n        : errorMessage;\n    const usage =\n      callError instanceof AIResponseParseError ? callError.usage : undefined;\n    return {\n      rect: undefined,\n      imageBase64: undefined,\n      error: `AI call error: ${errorMessage}`,\n      rawResponse,\n      usage,\n    };\n  }\n\n  let sectionRect: Rect | undefined;\n  const sectionBbox = result.content.bbox;\n  if (sectionBbox) {\n    const targetRect = adaptBboxToRect(\n      sectionBbox,\n      context.size.width,\n      context.size.height,\n      0,\n      0,\n      context.size.width,\n      context.size.height,\n      modelFamily,\n    );\n    debugSection('original targetRect %j', targetRect);\n\n    const referenceBboxList = result.content.references_bbox || [];\n    debugSection('referenceBboxList %j', referenceBboxList);\n\n    const referenceRects = referenceBboxList\n      .filter((bbox) => Array.isArray(bbox))\n      .map((bbox) => {\n        return adaptBboxToRect(\n          bbox,\n          context.size.width,\n          context.size.height,\n          0,\n          0,\n          context.size.width,\n          context.size.height,\n          modelFamily,\n        );\n      });\n    debugSection('referenceRects %j', referenceRects);\n\n    // merge the sectionRect and referenceRects\n    const mergedRect = mergeRects([targetRect, ...referenceRects]);\n    debugSection('mergedRect %j', mergedRect);\n\n    sectionRect = expandSearchArea(mergedRect, context.size);\n    debugSection('expanded sectionRect %j', sectionRect);\n  }\n\n  let imageBase64 = screenshotBase64;\n  let scale: number | undefined;\n\n  if (sectionRect) {\n    const originalWidth = sectionRect.width;\n    const originalHeight = sectionRect.height;\n\n    const croppedResult = await cropByRect(\n      screenshotBase64,\n      sectionRect,\n      modelFamily === 'qwen2.5-vl',\n    );\n\n    const scaleRatio = 2;\n    const scaledResult = await scaleImage(\n      croppedResult.imageBase64,\n      scaleRatio,\n    );\n\n    imageBase64 = scaledResult.imageBase64;\n    scale = scaleRatio;\n    sectionRect.width = scaledResult.width;\n    sectionRect.height = scaledResult.height;\n\n    debugSection(\n      'scaled sectionRect from %dx%d to %dx%d (scale=%d)',\n      originalWidth,\n      originalHeight,\n      sectionRect.width,\n      sectionRect.height,\n      scale,\n    );\n  }\n\n  return {\n    rect: sectionRect,\n    imageBase64,\n    scale,\n    error: result.content.error,\n    rawResponse: JSON.stringify(result.content),\n    usage: result.usage,\n  };\n}\n\nexport async function AiExtractElementInfo<T>(options: {\n  dataQuery: string | Record<string, string>;\n  multimodalPrompt?: TMultimodalPrompt;\n  context: UIContext;\n  pageDescription?: string;\n  extractOption?: ServiceExtractOption;\n  modelConfig: IModelConfig;\n}) {\n  const { dataQuery, context, extractOption, multimodalPrompt, modelConfig } =\n    options;\n  const systemPrompt = systemPromptToExtract();\n  const screenshotBase64 = context.screenshot.base64;\n\n  const extractDataPromptText = extractDataQueryPrompt(\n    options.pageDescription || '',\n    dataQuery,\n  );\n\n  const userContent: ChatCompletionUserMessageParam['content'] = [];\n\n  if (extractOption?.screenshotIncluded !== false) {\n    userContent.push({\n      type: 'image_url',\n      image_url: {\n        url: screenshotBase64,\n        detail: 'high',\n      },\n    });\n  }\n\n  userContent.push({\n    type: 'text',\n    text: extractDataPromptText,\n  });\n\n  const msgs: AIArgs = [\n    { role: 'system', content: systemPrompt },\n    {\n      role: 'user',\n      content: userContent,\n    },\n  ];\n\n  if (multimodalPrompt) {\n    const addOns = await promptsToChatParam({\n      images: multimodalPrompt.images,\n      convertHttpImage2Base64: multimodalPrompt.convertHttpImage2Base64,\n    });\n    msgs.push(...addOns);\n  }\n\n  const {\n    content: rawResponse,\n    usage,\n    reasoning_content,\n  } = await callAI(msgs, modelConfig);\n\n  // Parse XML response to JSON object\n  let parseResult: AIDataExtractionResponse<T>;\n  try {\n    parseResult = parseXMLExtractionResponse<T>(rawResponse);\n  } catch (parseError) {\n    // Throw AIResponseParseError with usage and rawResponse preserved\n    const errorMessage =\n      parseError instanceof Error ? parseError.message : String(parseError);\n    throw new AIResponseParseError(\n      `XML parse error: ${errorMessage}`,\n      rawResponse,\n      usage,\n    );\n  }\n\n  return {\n    parseResult,\n    rawResponse,\n    usage,\n    reasoning_content,\n  };\n}\n\nexport async function AiJudgeOrderSensitive(\n  description: string,\n  callAIFn: typeof callAIWithObjectResponse<{ isOrderSensitive: boolean }>,\n  modelConfig: IModelConfig,\n): Promise<{\n  isOrderSensitive: boolean;\n  usage?: AIUsageInfo;\n}> {\n  const systemPrompt = systemPromptToJudgeOrderSensitive();\n  const userPrompt = orderSensitiveJudgePrompt(description);\n\n  const msgs: AIArgs = [\n    { role: 'system', content: systemPrompt },\n    {\n      role: 'user',\n      content: userPrompt,\n    },\n  ];\n\n  const result = await callAIFn(msgs, modelConfig);\n\n  return {\n    isOrderSensitive: result.content.isOrderSensitive ?? false,\n    usage: result.usage,\n  };\n}\n"],"names":["debugInspect","getDebug","debugSection","extraTextFromUserPrompt","prompt","promptsToChatParam","multimodalPrompt","msgs","item","base64","preProcessImageUrl","AiLocateElement","options","context","targetElementDescription","callAIFn","modelConfig","modelFamily","screenshotBase64","assert","targetElementDescriptionText","userInstructionPrompt","findElementPrompt","systemPrompt","isAutoGLM","getAutoGLMLocatePrompt","systemPromptToLocateElement","imagePayload","imageWidth","imageHeight","originalImageWidth","originalImageHeight","paddedResult","paddingToMatchBlockByBase64","addOns","rawResponseContent","usage","callAIWithStringResponse","parsed","parseAutoGLMLocateResponse","resRect","matchedElements","errors","x","y","pixelX","Math","pixelY","finalX","finalY","element","generateElementByPoint","res","callError","errorMessage","Error","String","rawResponse","AIResponseParseError","undefined","JSON","Array","adaptBboxToRect","generateElementByRect","e","msg","AiLocateSection","sectionDescription","systemPromptToLocateSection","sectionLocatorInstructionText","sectionLocatorInstruction","result","callAIWithObjectResponse","sectionRect","sectionBbox","targetRect","referenceBboxList","referenceRects","bbox","mergedRect","mergeRects","expandSearchArea","imageBase64","scale","originalWidth","originalHeight","croppedResult","cropByRect","scaleRatio","scaledResult","scaleImage","AiExtractElementInfo","dataQuery","extractOption","systemPromptToExtract","extractDataPromptText","extractDataQueryPrompt","userContent","reasoning_content","callAI","parseResult","parseXMLExtractionResponse","parseError","AiJudgeOrderSensitive","description","systemPromptToJudgeOrderSensitive","userPrompt","orderSensitiveJudgePrompt"],"mappings":";;;;;;;;;;;;;AA6DA,MAAMA,eAAeC,SAAS;AAC9B,MAAMC,eAAeD,SAAS;AAE9B,MAAME,0BAA0B,CAACC;IAC/B,IAAI,AAAkB,YAAlB,OAAOA,QACT,OAAOA;IAEP,OAAOA,OAAO,MAAM;AAExB;AAEA,MAAMC,qBAAqB,OACzBC;IAEA,MAAMC,OAAyC,EAAE;IACjD,IAAID,kBAAkB,QAAQ,QAAQ;QACpCC,KAAK,IAAI,CAAC;YACR,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM;gBACR;aACD;QACH;QAEA,KAAK,MAAMC,QAAQF,iBAAiB,MAAM,CAAE;YAC1C,MAAMG,SAAS,MAAMC,mBACnBF,KAAK,GAAG,EACR,CAAC,CAACF,iBAAiB,uBAAuB;YAG5CC,KAAK,IAAI,CAAC;gBACR,MAAM;gBACN,SAAS;oBACP;wBACE,MAAM;wBACN,MAAM,CAAC,mCAAmC,EAAEC,KAAK,IAAI,CAAC,EAAE,CAAC;oBAC3D;iBACD;YACH;YAEAD,KAAK,IAAI,CAAC;gBACR,MAAM;gBACN,SAAS;oBACP;wBACE,MAAM;wBACN,WAAW;4BACT,KAAKE;4BACL,QAAQ;wBACV;oBACF;iBACD;YACH;QACF;IACF;IACA,OAAOF;AACT;AAEO,eAAeI,gBAAgBC,OAQrC;IAUC,MAAM,EAAEC,OAAO,EAAEC,wBAAwB,EAAEC,QAAQ,EAAEC,WAAW,EAAE,GAAGJ;IACrE,MAAM,EAAEK,WAAW,EAAE,GAAGD;IACxB,MAAME,mBAAmBL,QAAQ,UAAU,CAAC,MAAM;IAElDM,OACEL,0BACA;IAEF,MAAMM,+BAA+BjB,wBACnCW;IAEF,MAAMO,wBAAwBC,kBAAkBF;IAChD,MAAMG,eAAeC,UAAUP,eAC3BQ,uBAAuBR,eACvBS,4BAA4BT;IAEhC,IAAIU,eAAeT;IACnB,IAAIU,aAAaf,QAAQ,IAAI,CAAC,KAAK;IACnC,IAAIgB,cAAchB,QAAQ,IAAI,CAAC,MAAM;IACrC,IAAIiB,qBAAqBF;IACzB,IAAIG,sBAAsBF;IAE1B,IAAIjB,QAAQ,YAAY,EAAE;QACxBO,OACEP,QAAQ,YAAY,CAAC,IAAI,EACzB;QAEFO,OACEP,QAAQ,YAAY,CAAC,WAAW,EAChC;QAGFe,eAAef,QAAQ,YAAY,CAAC,WAAW;QAC/CgB,aAAahB,QAAQ,YAAY,CAAC,IAAI,EAAE;QACxCiB,cAAcjB,QAAQ,YAAY,CAAC,IAAI,EAAE;QACzCkB,qBAAqBF;QACrBG,sBAAsBF;IACxB,OAAO,IAAIZ,AAAgB,iBAAhBA,aAA8B;QACvC,MAAMe,eAAe,MAAMC,4BAA4BN;QACvDC,aAAaI,aAAa,KAAK;QAC/BH,cAAcG,aAAa,MAAM;QACjCL,eAAeK,aAAa,WAAW;IACzC;IAEA,MAAMzB,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKI;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAMH,UAAUP,eACZ,CAAC,KAAK,EAAEI,uBAAuB,GAC/BA;gBACN;aACD;QACH;KACD;IAED,IAAI,AAAoC,YAApC,OAAOP,0BAAuC;QAChD,MAAMoB,SAAS,MAAM7B,mBAAmB;YACtC,QAAQS,yBAAyB,MAAM;YACvC,yBAAyBA,yBAAyB,uBAAuB;QAC3E;QACAP,KAAK,IAAI,IAAI2B;IACf;IAEA,IAAIV,UAAUP,cAAc;QAC1B,MAAM,EAAE,SAASkB,kBAAkB,EAAEC,KAAK,EAAE,GAC1C,MAAMC,yBAAyB9B,MAAMS;QAEvChB,aAAa,yBAAyBmC;QAEtC,MAAMG,SAASC,2BAA2BJ;QAE1CnC,aAAa,sBAAsBsC,OAAO,KAAK;QAC/CtC,aAAa,yBAAyBsC,OAAO,WAAW;QAExD,IAAIE;QACJ,IAAIC,kBAAyC,EAAE;QAC/C,IAAIC,SAAmB,EAAE;QAEzB,IAAIJ,OAAO,KAAK,IAAI,CAACA,OAAO,WAAW,EAAE;YACvCI,SAAS;gBAACJ,OAAO,KAAK,IAAI;aAAoC;YAC9DtC,aAAa,yBAAyB0C,MAAM,CAAC,EAAE;QACjD,OAAO;YACL,MAAM,EAAEC,CAAC,EAAEC,CAAC,EAAE,GAAGN,OAAO,WAAW;YAEnCtC,aAAa,iCAAiC;gBAAE2C;gBAAGC;YAAE;YAIrD,MAAMC,SAASC,KAAK,KAAK,CAAEH,IAAIf,aAAc;YAC7C,MAAMmB,SAASD,KAAK,KAAK,CAAEF,IAAIf,cAAe;YAE9C7B,aAAa,+BAA+B;gBAAE6C;gBAAQE;YAAO;YAG7D,IAAIC,SAASH;YACb,IAAII,SAASF;YACb,IAAInC,QAAQ,YAAY,EAAE,MAAM;gBAC9BoC,UAAUpC,QAAQ,YAAY,CAAC,IAAI,CAAC,IAAI;gBACxCqC,UAAUrC,QAAQ,YAAY,CAAC,IAAI,CAAC,GAAG;YACzC;YAEA,MAAMsC,UAA+BC,uBACnC;gBAACH;gBAAQC;aAAO,EAChB7B;YAGFoB,UAAUU,QAAQ,IAAI;YACtBlD,aAAa,qBAAqBwC;YAElC,IAAIU,SACFT,kBAAkB;gBAACS;aAAQ;QAE/B;QAEA,OAAO;YACL,MAAMV;YACN,aAAa;gBACX,UAAUC;gBACVC;YACF;YACA,aAAaP;YACbC;YACA,mBAAmBE,OAAO,KAAK;QACjC;IACF;IAEA,IAAIc;IACJ,IAAI;QACFA,MAAM,MAAMrC,SAASR,MAAMS;IAC7B,EAAE,OAAOqC,WAAW;QAElB,MAAMC,eACJD,qBAAqBE,QAAQF,UAAU,OAAO,GAAGG,OAAOH;QAC1D,MAAMI,cACJJ,qBAAqBK,uBACjBL,UAAU,WAAW,GACrBC;QACN,MAAMlB,QACJiB,qBAAqBK,uBAAuBL,UAAU,KAAK,GAAGM;QAChE,OAAO;YACL,MAAMA;YACN,aAAa;gBACX,UAAU,EAAE;gBACZ,QAAQ;oBAAC,CAAC,eAAe,EAAEL,cAAc;iBAAC;YAC5C;YACAG;YACArB;YACA,mBAAmBuB;QACrB;IACF;IAEA,MAAMF,cAAcG,KAAK,SAAS,CAACR,IAAI,OAAO;IAE9C,IAAIZ;IACJ,IAAIC,kBAAyC,EAAE;IAC/C,IAAIC,SACF,YAAYU,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,MAAM,GAAG,EAAE;IACnD,IAAI;QACF,IACE,UAAUA,IAAI,OAAO,IACrBS,MAAM,OAAO,CAACT,IAAI,OAAO,CAAC,IAAI,KAC9BA,IAAI,OAAO,CAAC,IAAI,CAAC,MAAM,IAAI,GAC3B;YACAZ,UAAUsB,gBACRV,IAAI,OAAO,CAAC,IAAI,EAChBxB,YACAC,aACAjB,QAAQ,YAAY,EAAE,MAAM,MAC5BA,QAAQ,YAAY,EAAE,MAAM,KAC5BkB,oBACAC,qBACAd,aACAL,QAAQ,YAAY,EAAE;YAGxBZ,aAAa,WAAWwC;YAExB,MAAMU,UAA+Ba,sBACnCvB,SACApB;YAEFsB,SAAS,EAAE;YAEX,IAAIQ,SACFT,kBAAkB;gBAACS;aAAQ;QAE/B;IACF,EAAE,OAAOc,GAAG;QACV,MAAMC,MACJD,aAAaT,QACT,CAAC,sBAAsB,EAAES,EAAE,OAAO,EAAE,GACpC;QACN,IAAI,AAACtB,UAAUA,QAAQ,WAAW,GAGhCA,OAAO,IAAI,CAAC,CAAC,CAAC,EAAEuB,IAAI,CAAC,CAAC;aAFtBvB,SAAS;YAACuB;SAAI;IAIlB;IAEA,OAAO;QACL,MAAMzB;QACN,aAAa;YACX,UAAUC;YACV,QAAQC;QACV;QACAe;QACA,OAAOL,IAAI,KAAK;QAChB,mBAAmBA,IAAI,iBAAiB;IAC1C;AACF;AAEO,eAAec,gBAAgBtD,OAIrC;IAQC,MAAM,EAAEC,OAAO,EAAEsD,kBAAkB,EAAEnD,WAAW,EAAE,GAAGJ;IACrD,MAAM,EAAEK,WAAW,EAAE,GAAGD;IACxB,MAAME,mBAAmBL,QAAQ,UAAU,CAAC,MAAM;IAElD,MAAMU,eAAe6C,4BAA4BnD;IACjD,MAAMoD,gCAAgCC,0BACpCnE,wBAAwBgE;IAE1B,MAAM5D,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKL;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAMmD;gBACR;aACD;QACH;KACD;IAED,IAAI,AAA8B,YAA9B,OAAOF,oBAAiC;QAC1C,MAAMjC,SAAS,MAAM7B,mBAAmB;YACtC,QAAQ8D,mBAAmB,MAAM;YACjC,yBAAyBA,mBAAmB,uBAAuB;QACrE;QACA5D,KAAK,IAAI,IAAI2B;IACf;IAEA,IAAIqC;IAGJ,IAAI;QACFA,SAAS,MAAMC,yBACbjE,MACAS;IAEJ,EAAE,OAAOqC,WAAW;QAElB,MAAMC,eACJD,qBAAqBE,QAAQF,UAAU,OAAO,GAAGG,OAAOH;QAC1D,MAAMI,cACJJ,qBAAqBK,uBACjBL,UAAU,WAAW,GACrBC;QACN,MAAMlB,QACJiB,qBAAqBK,uBAAuBL,UAAU,KAAK,GAAGM;QAChE,OAAO;YACL,MAAMA;YACN,aAAaA;YACb,OAAO,CAAC,eAAe,EAAEL,cAAc;YACvCG;YACArB;QACF;IACF;IAEA,IAAIqC;IACJ,MAAMC,cAAcH,OAAO,OAAO,CAAC,IAAI;IACvC,IAAIG,aAAa;QACf,MAAMC,aAAab,gBACjBY,aACA7D,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAEFf,aAAa,0BAA0ByE;QAEvC,MAAMC,oBAAoBL,OAAO,OAAO,CAAC,eAAe,IAAI,EAAE;QAC9DrE,aAAa,wBAAwB0E;QAErC,MAAMC,iBAAiBD,kBACpB,MAAM,CAAC,CAACE,OAASjB,MAAM,OAAO,CAACiB,OAC/B,GAAG,CAAC,CAACA,OACGhB,gBACLgB,MACAjE,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAGNf,aAAa,qBAAqB2E;QAGlC,MAAME,aAAaC,WAAW;YAACL;eAAeE;SAAe;QAC7D3E,aAAa,iBAAiB6E;QAE9BN,cAAcQ,iBAAiBF,YAAYlE,QAAQ,IAAI;QACvDX,aAAa,2BAA2BuE;IAC1C;IAEA,IAAIS,cAAchE;IAClB,IAAIiE;IAEJ,IAAIV,aAAa;QACf,MAAMW,gBAAgBX,YAAY,KAAK;QACvC,MAAMY,iBAAiBZ,YAAY,MAAM;QAEzC,MAAMa,gBAAgB,MAAMC,WAC1BrE,kBACAuD,aACAxD,AAAgB,iBAAhBA;QAGF,MAAMuE,aAAa;QACnB,MAAMC,eAAe,MAAMC,WACzBJ,cAAc,WAAW,EACzBE;QAGFN,cAAcO,aAAa,WAAW;QACtCN,QAAQK;QACRf,YAAY,KAAK,GAAGgB,aAAa,KAAK;QACtChB,YAAY,MAAM,GAAGgB,aAAa,MAAM;QAExCvF,aACE,qDACAkF,eACAC,gBACAZ,YAAY,KAAK,EACjBA,YAAY,MAAM,EAClBU;IAEJ;IAEA,OAAO;QACL,MAAMV;QACNS;QACAC;QACA,OAAOZ,OAAO,OAAO,CAAC,KAAK;QAC3B,aAAaX,KAAK,SAAS,CAACW,OAAO,OAAO;QAC1C,OAAOA,OAAO,KAAK;IACrB;AACF;AAEO,eAAeoB,qBAAwB/E,OAO7C;IACC,MAAM,EAAEgF,SAAS,EAAE/E,OAAO,EAAEgF,aAAa,EAAEvF,gBAAgB,EAAEU,WAAW,EAAE,GACxEJ;IACF,MAAMW,eAAeuE;IACrB,MAAM5E,mBAAmBL,QAAQ,UAAU,CAAC,MAAM;IAElD,MAAMkF,wBAAwBC,uBAC5BpF,QAAQ,eAAe,IAAI,IAC3BgF;IAGF,MAAMK,cAAyD,EAAE;IAEjE,IAAIJ,eAAe,uBAAuB,OACxCI,YAAY,IAAI,CAAC;QACf,MAAM;QACN,WAAW;YACT,KAAK/E;YACL,QAAQ;QACV;IACF;IAGF+E,YAAY,IAAI,CAAC;QACf,MAAM;QACN,MAAMF;IACR;IAEA,MAAMxF,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAAS0E;QACX;KACD;IAED,IAAI3F,kBAAkB;QACpB,MAAM4B,SAAS,MAAM7B,mBAAmB;YACtC,QAAQC,iBAAiB,MAAM;YAC/B,yBAAyBA,iBAAiB,uBAAuB;QACnE;QACAC,KAAK,IAAI,IAAI2B;IACf;IAEA,MAAM,EACJ,SAASuB,WAAW,EACpBrB,KAAK,EACL8D,iBAAiB,EAClB,GAAG,MAAMC,OAAO5F,MAAMS;IAGvB,IAAIoF;IACJ,IAAI;QACFA,cAAcC,2BAA8B5C;IAC9C,EAAE,OAAO6C,YAAY;QAEnB,MAAMhD,eACJgD,sBAAsB/C,QAAQ+C,WAAW,OAAO,GAAG9C,OAAO8C;QAC5D,MAAM,IAAI5C,qBACR,CAAC,iBAAiB,EAAEJ,cAAc,EAClCG,aACArB;IAEJ;IAEA,OAAO;QACLgE;QACA3C;QACArB;QACA8D;IACF;AACF;AAEO,eAAeK,sBACpBC,WAAmB,EACnBzF,QAAwE,EACxEC,WAAyB;IAKzB,MAAMO,eAAekF;IACrB,MAAMC,aAAaC,0BAA0BH;IAE7C,MAAMjG,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAASmF;QACX;KACD;IAED,MAAMnC,SAAS,MAAMxD,SAASR,MAAMS;IAEpC,OAAO;QACL,kBAAkBuD,OAAO,OAAO,CAAC,gBAAgB,IAAI;QACrD,OAAOA,OAAO,KAAK;IACrB;AACF"}
+{"version":3,"file":"ai-model/inspect.mjs","sources":["../../../src/ai-model/inspect.ts"],"sourcesContent":["import type {\n  AIDataExtractionResponse,\n  AIElementResponse,\n  AISectionLocatorResponse,\n  AIUsageInfo,\n  Rect,\n  ServiceExtractOption,\n  UIContext,\n} from '@/types';\nimport type { IModelConfig } from '@midscene/shared/env';\nimport {\n  generateElementByPoint,\n  generateElementByRect,\n} from '@midscene/shared/extractor/dom-util';\nimport {\n  cropByRect,\n  paddingToMatchBlockByBase64,\n  preProcessImageUrl,\n  scaleImage,\n} from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport type { LocateResultElement } from '@midscene/shared/types';\nimport { assert } from '@midscene/shared/utils';\nimport type {\n  ChatCompletionSystemMessageParam,\n  ChatCompletionUserMessageParam,\n} from 'openai/resources/index';\nimport type { TMultimodalPrompt, TUserPrompt } from '../common';\nimport { adaptBboxToRect, expandSearchArea, mergeRects } from '../common';\nimport { parseAutoGLMLocateResponse } from './auto-glm/parser';\nimport { getAutoGLMLocatePrompt } from './auto-glm/prompt';\nimport { isAutoGLM } from './auto-glm/util';\nimport {\n  extractDataQueryPrompt,\n  parseXMLExtractionResponse,\n  systemPromptToExtract,\n} from './prompt/extraction';\nimport {\n  findElementPrompt,\n  systemPromptToLocateElement,\n} from './prompt/llm-locator';\nimport {\n  sectionLocatorInstruction,\n  systemPromptToLocateSection,\n} from './prompt/llm-section-locator';\nimport {\n  orderSensitiveJudgePrompt,\n  systemPromptToJudgeOrderSensitive,\n} from './prompt/order-sensitive-judge';\nimport {\n  AIResponseParseError,\n  callAI,\n  callAIWithObjectResponse,\n  callAIWithStringResponse,\n} from './service-caller/index';\n\nexport type AIArgs = [\n  ChatCompletionSystemMessageParam,\n  ...ChatCompletionUserMessageParam[],\n];\n\nconst debugInspect = getDebug('ai:inspect');\nconst debugSection = getDebug('ai:section');\n\nconst extraTextFromUserPrompt = (prompt: TUserPrompt): string => {\n  if (typeof prompt === 'string') {\n    return prompt;\n  } else {\n    return prompt.prompt;\n  }\n};\n\nconst promptsToChatParam = async (\n  multimodalPrompt: TMultimodalPrompt,\n): Promise<ChatCompletionUserMessageParam[]> => {\n  const msgs: ChatCompletionUserMessageParam[] = [];\n  if (multimodalPrompt?.images?.length) {\n    msgs.push({\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: 'Next, I will provide all the reference images.',\n        },\n      ],\n    });\n\n    for (const item of multimodalPrompt.images) {\n      const base64 = await preProcessImageUrl(\n        item.url,\n        !!multimodalPrompt.convertHttpImage2Base64,\n      );\n\n      msgs.push({\n        role: 'user',\n        content: [\n          {\n            type: 'text',\n            text: `this is the reference image named '${item.name}':`,\n          },\n        ],\n      });\n\n      msgs.push({\n        role: 'user',\n        content: [\n          {\n            type: 'image_url',\n            image_url: {\n              url: base64,\n              detail: 'high',\n            },\n          },\n        ],\n      });\n    }\n  }\n  return msgs;\n};\n\nexport async function AiLocateElement(options: {\n  context: UIContext;\n  targetElementDescription: TUserPrompt;\n  searchConfig?: Awaited<ReturnType<typeof AiLocateSection>>;\n  modelConfig: IModelConfig;\n}): Promise<{\n  parseResult: {\n    elements: LocateResultElement[];\n    errors?: string[];\n  };\n  rect?: Rect;\n  rawResponse: string;\n  usage?: AIUsageInfo;\n  reasoning_content?: string;\n}> {\n  const { context, targetElementDescription, modelConfig } = options;\n  const { modelFamily } = modelConfig;\n  const screenshotBase64 = context.screenshot.base64;\n\n  assert(\n    targetElementDescription,\n    'cannot find the target element description',\n  );\n  const targetElementDescriptionText = extraTextFromUserPrompt(\n    targetElementDescription,\n  );\n  const userInstructionPrompt = findElementPrompt(targetElementDescriptionText);\n  const systemPrompt = isAutoGLM(modelFamily)\n    ? getAutoGLMLocatePrompt(modelFamily)\n    : systemPromptToLocateElement(modelFamily);\n\n  let imagePayload = screenshotBase64;\n  let imageWidth = context.size.width;\n  let imageHeight = context.size.height;\n  let originalImageWidth = imageWidth;\n  let originalImageHeight = imageHeight;\n\n  if (options.searchConfig) {\n    assert(\n      options.searchConfig.rect,\n      'searchArea is provided but its rect cannot be found. Failed to locate element',\n    );\n    assert(\n      options.searchConfig.imageBase64,\n      'searchArea is provided but its imageBase64 cannot be found. Failed to locate element',\n    );\n\n    imagePayload = options.searchConfig.imageBase64;\n    imageWidth = options.searchConfig.rect?.width;\n    imageHeight = options.searchConfig.rect?.height;\n    originalImageWidth = imageWidth;\n    originalImageHeight = imageHeight;\n  } else if (modelFamily === 'qwen2.5-vl') {\n    const paddedResult = await paddingToMatchBlockByBase64(imagePayload);\n    imageWidth = paddedResult.width;\n    imageHeight = paddedResult.height;\n    imagePayload = paddedResult.imageBase64;\n  }\n\n  const msgs: AIArgs = [\n    { role: 'system', content: systemPrompt },\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n        {\n          type: 'text',\n          text: isAutoGLM(modelFamily)\n            ? `Tap: ${userInstructionPrompt}`\n            : userInstructionPrompt,\n        },\n      ],\n    },\n  ];\n\n  if (typeof targetElementDescription !== 'string') {\n    const addOns = await promptsToChatParam({\n      images: targetElementDescription.images,\n      convertHttpImage2Base64: targetElementDescription.convertHttpImage2Base64,\n    });\n    msgs.push(...addOns);\n  }\n\n  if (isAutoGLM(modelFamily)) {\n    const { content: rawResponseContent, usage } =\n      await callAIWithStringResponse(msgs, modelConfig);\n\n    debugInspect('auto-glm rawResponse:', rawResponseContent);\n\n    const parsed = parseAutoGLMLocateResponse(rawResponseContent);\n\n    debugInspect('auto-glm thinking:', parsed.think);\n    debugInspect('auto-glm coordinates:', parsed.coordinates);\n\n    let resRect: Rect | undefined;\n    let matchedElements: LocateResultElement[] = [];\n    let errors: string[] = [];\n\n    if (parsed.error || !parsed.coordinates) {\n      errors = [parsed.error || 'Failed to parse auto-glm response'];\n      debugInspect('auto-glm parse error:', errors[0]);\n    } else {\n      const { x, y } = parsed.coordinates;\n\n      debugInspect('auto-glm coordinates [0-999]:', { x, y });\n\n      // Convert auto-glm coordinates [0,999] to pixel bbox\n      // Map from [0,999] to pixel coordinates\n      const pixelX = Math.round((x * imageWidth) / 1000);\n      const pixelY = Math.round((y * imageHeight) / 1000);\n\n      debugInspect('auto-glm pixel coordinates:', { pixelX, pixelY });\n\n      // Apply offset if searching in a cropped area\n      let finalX = pixelX;\n      let finalY = pixelY;\n      if (options.searchConfig?.rect) {\n        finalX += options.searchConfig.rect.left;\n        finalY += options.searchConfig.rect.top;\n      }\n\n      const element: LocateResultElement = generateElementByPoint(\n        [finalX, finalY],\n        targetElementDescriptionText as string,\n      );\n\n      resRect = element.rect;\n      debugInspect('auto-glm resRect:', resRect);\n\n      if (element) {\n        matchedElements = [element];\n      }\n    }\n\n    return {\n      rect: resRect,\n      parseResult: {\n        elements: matchedElements,\n        errors,\n      },\n      rawResponse: rawResponseContent,\n      usage,\n      reasoning_content: parsed.think,\n    };\n  }\n\n  let res: Awaited<\n    ReturnType<\n      typeof callAIWithObjectResponse<AIElementResponse | [number, number]>\n    >\n  >;\n  try {\n    res = await callAIWithObjectResponse<AIElementResponse | [number, number]>(\n      msgs,\n      modelConfig,\n    );\n  } catch (callError) {\n    // Return error with usage and rawResponse if available\n    const errorMessage =\n      callError instanceof Error ? callError.message : String(callError);\n    const rawResponse =\n      callError instanceof AIResponseParseError\n        ? callError.rawResponse\n        : errorMessage;\n    const usage =\n      callError instanceof AIResponseParseError ? callError.usage : undefined;\n    return {\n      rect: undefined,\n      parseResult: {\n        elements: [],\n        errors: [`AI call error: ${errorMessage}`],\n      },\n      rawResponse,\n      usage,\n      reasoning_content: undefined,\n    };\n  }\n\n  const rawResponse = JSON.stringify(res.content);\n\n  let resRect: Rect | undefined;\n  let matchedElements: LocateResultElement[] = [];\n  let errors: string[] | undefined =\n    'errors' in res.content ? res.content.errors : [];\n  try {\n    if (\n      'bbox' in res.content &&\n      Array.isArray(res.content.bbox) &&\n      res.content.bbox.length >= 1\n    ) {\n      resRect = adaptBboxToRect(\n        res.content.bbox,\n        imageWidth,\n        imageHeight,\n        options.searchConfig?.rect?.left,\n        options.searchConfig?.rect?.top,\n        originalImageWidth,\n        originalImageHeight,\n        modelFamily,\n        options.searchConfig?.scale,\n      );\n\n      debugInspect('resRect', resRect);\n\n      const element: LocateResultElement = generateElementByRect(\n        resRect,\n        targetElementDescriptionText as string,\n      );\n      errors = [];\n\n      if (element) {\n        matchedElements = [element];\n      }\n    }\n  } catch (e) {\n    const msg =\n      e instanceof Error\n        ? `Failed to parse bbox: ${e.message}`\n        : 'unknown error in locate';\n    if (!errors || errors?.length === 0) {\n      errors = [msg];\n    } else {\n      errors.push(`(${msg})`);\n    }\n  }\n\n  return {\n    rect: resRect,\n    parseResult: {\n      elements: matchedElements as LocateResultElement[],\n      errors: errors as string[],\n    },\n    rawResponse,\n    usage: res.usage,\n    reasoning_content: res.reasoning_content,\n  };\n}\n\nexport async function AiLocateSection(options: {\n  context: UIContext;\n  sectionDescription: TUserPrompt;\n  modelConfig: IModelConfig;\n}): Promise<{\n  rect?: Rect;\n  imageBase64?: string;\n  scale?: number;\n  error?: string;\n  rawResponse: string;\n  usage?: AIUsageInfo;\n}> {\n  const { context, sectionDescription, modelConfig } = options;\n  const { modelFamily } = modelConfig;\n  const screenshotBase64 = context.screenshot.base64;\n\n  const systemPrompt = systemPromptToLocateSection(modelFamily);\n  const sectionLocatorInstructionText = sectionLocatorInstruction(\n    extraTextFromUserPrompt(sectionDescription),\n  );\n  const msgs: AIArgs = [\n    { role: 'system', content: systemPrompt },\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'image_url',\n          image_url: {\n            url: screenshotBase64,\n            detail: 'high',\n          },\n        },\n        {\n          type: 'text',\n          text: sectionLocatorInstructionText,\n        },\n      ],\n    },\n  ];\n\n  if (typeof sectionDescription !== 'string') {\n    const addOns = await promptsToChatParam({\n      images: sectionDescription.images,\n      convertHttpImage2Base64: sectionDescription.convertHttpImage2Base64,\n    });\n    msgs.push(...addOns);\n  }\n\n  let result: Awaited<\n    ReturnType<typeof callAIWithObjectResponse<AISectionLocatorResponse>>\n  >;\n  try {\n    result = await callAIWithObjectResponse<AISectionLocatorResponse>(\n      msgs,\n      modelConfig,\n    );\n  } catch (callError) {\n    // Return error with usage and rawResponse if available\n    const errorMessage =\n      callError instanceof Error ? callError.message : String(callError);\n    const rawResponse =\n      callError instanceof AIResponseParseError\n        ? callError.rawResponse\n        : errorMessage;\n    const usage =\n      callError instanceof AIResponseParseError ? callError.usage : undefined;\n    return {\n      rect: undefined,\n      imageBase64: undefined,\n      error: `AI call error: ${errorMessage}`,\n      rawResponse,\n      usage,\n    };\n  }\n\n  let sectionRect: Rect | undefined;\n  const sectionBbox = result.content.bbox;\n  if (sectionBbox) {\n    const targetRect = adaptBboxToRect(\n      sectionBbox,\n      context.size.width,\n      context.size.height,\n      0,\n      0,\n      context.size.width,\n      context.size.height,\n      modelFamily,\n    );\n    debugSection('original targetRect %j', targetRect);\n\n    const referenceBboxList = result.content.references_bbox || [];\n    debugSection('referenceBboxList %j', referenceBboxList);\n\n    const referenceRects = referenceBboxList\n      .filter((bbox) => Array.isArray(bbox))\n      .map((bbox) => {\n        return adaptBboxToRect(\n          bbox,\n          context.size.width,\n          context.size.height,\n          0,\n          0,\n          context.size.width,\n          context.size.height,\n          modelFamily,\n        );\n      });\n    debugSection('referenceRects %j', referenceRects);\n\n    // merge the sectionRect and referenceRects\n    const mergedRect = mergeRects([targetRect, ...referenceRects]);\n    debugSection('mergedRect %j', mergedRect);\n\n    sectionRect = expandSearchArea(mergedRect, context.size);\n    debugSection('expanded sectionRect %j', sectionRect);\n  }\n\n  let imageBase64 = screenshotBase64;\n  let scale: number | undefined;\n\n  if (sectionRect) {\n    const originalWidth = sectionRect.width;\n    const originalHeight = sectionRect.height;\n\n    const croppedResult = await cropByRect(\n      screenshotBase64,\n      sectionRect,\n      modelFamily === 'qwen2.5-vl',\n    );\n\n    const scaleRatio = 2;\n    const scaledResult = await scaleImage(\n      croppedResult.imageBase64,\n      scaleRatio,\n    );\n\n    imageBase64 = scaledResult.imageBase64;\n    scale = scaleRatio;\n    sectionRect.width = scaledResult.width;\n    sectionRect.height = scaledResult.height;\n\n    debugSection(\n      'scaled sectionRect from %dx%d to %dx%d (scale=%d)',\n      originalWidth,\n      originalHeight,\n      sectionRect.width,\n      sectionRect.height,\n      scale,\n    );\n  }\n\n  return {\n    rect: sectionRect,\n    imageBase64,\n    scale,\n    error: result.content.error,\n    rawResponse: JSON.stringify(result.content),\n    usage: result.usage,\n  };\n}\n\nexport async function AiExtractElementInfo<T>(options: {\n  dataQuery: string | Record<string, string>;\n  multimodalPrompt?: TMultimodalPrompt;\n  context: UIContext;\n  pageDescription?: string;\n  extractOption?: ServiceExtractOption;\n  modelConfig: IModelConfig;\n}) {\n  const { dataQuery, context, extractOption, multimodalPrompt, modelConfig } =\n    options;\n  const systemPrompt = systemPromptToExtract();\n  const screenshotBase64 = context.screenshot.base64;\n\n  const extractDataPromptText = extractDataQueryPrompt(\n    options.pageDescription || '',\n    dataQuery,\n  );\n\n  const userContent: ChatCompletionUserMessageParam['content'] = [];\n\n  if (extractOption?.screenshotIncluded !== false) {\n    userContent.push({\n      type: 'image_url',\n      image_url: {\n        url: screenshotBase64,\n        detail: 'high',\n      },\n    });\n  }\n\n  userContent.push({\n    type: 'text',\n    text: extractDataPromptText,\n  });\n\n  const msgs: AIArgs = [\n    { role: 'system', content: systemPrompt },\n    {\n      role: 'user',\n      content: userContent,\n    },\n  ];\n\n  if (multimodalPrompt) {\n    const addOns = await promptsToChatParam({\n      images: multimodalPrompt.images,\n      convertHttpImage2Base64: multimodalPrompt.convertHttpImage2Base64,\n    });\n    msgs.push(...addOns);\n  }\n\n  const {\n    content: rawResponse,\n    usage,\n    reasoning_content,\n  } = await callAI(msgs, modelConfig);\n\n  // Parse XML response to JSON object\n  let parseResult: AIDataExtractionResponse<T>;\n  try {\n    parseResult = parseXMLExtractionResponse<T>(rawResponse);\n  } catch (parseError) {\n    // Throw AIResponseParseError with usage and rawResponse preserved\n    const errorMessage =\n      parseError instanceof Error ? parseError.message : String(parseError);\n    throw new AIResponseParseError(\n      `XML parse error: ${errorMessage}`,\n      rawResponse,\n      usage,\n    );\n  }\n\n  return {\n    parseResult,\n    rawResponse,\n    usage,\n    reasoning_content,\n  };\n}\n\nexport async function AiJudgeOrderSensitive(\n  description: string,\n  callAIFn: typeof callAIWithObjectResponse<{ isOrderSensitive: boolean }>,\n  modelConfig: IModelConfig,\n): Promise<{\n  isOrderSensitive: boolean;\n  usage?: AIUsageInfo;\n}> {\n  const systemPrompt = systemPromptToJudgeOrderSensitive();\n  const userPrompt = orderSensitiveJudgePrompt(description);\n\n  const msgs: AIArgs = [\n    { role: 'system', content: systemPrompt },\n    {\n      role: 'user',\n      content: userPrompt,\n    },\n  ];\n\n  const result = await callAIFn(msgs, modelConfig);\n\n  return {\n    isOrderSensitive: result.content.isOrderSensitive ?? false,\n    usage: result.usage,\n  };\n}\n"],"names":["debugInspect","getDebug","debugSection","extraTextFromUserPrompt","prompt","promptsToChatParam","multimodalPrompt","msgs","item","base64","preProcessImageUrl","AiLocateElement","options","context","targetElementDescription","modelConfig","modelFamily","screenshotBase64","assert","targetElementDescriptionText","userInstructionPrompt","findElementPrompt","systemPrompt","isAutoGLM","getAutoGLMLocatePrompt","systemPromptToLocateElement","imagePayload","imageWidth","imageHeight","originalImageWidth","originalImageHeight","paddedResult","paddingToMatchBlockByBase64","addOns","rawResponseContent","usage","callAIWithStringResponse","parsed","parseAutoGLMLocateResponse","resRect","matchedElements","errors","x","y","pixelX","Math","pixelY","finalX","finalY","element","generateElementByPoint","res","callAIWithObjectResponse","callError","errorMessage","Error","String","rawResponse","AIResponseParseError","undefined","JSON","Array","adaptBboxToRect","generateElementByRect","e","msg","AiLocateSection","sectionDescription","systemPromptToLocateSection","sectionLocatorInstructionText","sectionLocatorInstruction","result","sectionRect","sectionBbox","targetRect","referenceBboxList","referenceRects","bbox","mergedRect","mergeRects","expandSearchArea","imageBase64","scale","originalWidth","originalHeight","croppedResult","cropByRect","scaleRatio","scaledResult","scaleImage","AiExtractElementInfo","dataQuery","extractOption","systemPromptToExtract","extractDataPromptText","extractDataQueryPrompt","userContent","reasoning_content","callAI","parseResult","parseXMLExtractionResponse","parseError","AiJudgeOrderSensitive","description","callAIFn","systemPromptToJudgeOrderSensitive","userPrompt","orderSensitiveJudgePrompt"],"mappings":";;;;;;;;;;;;;AA6DA,MAAMA,eAAeC,SAAS;AAC9B,MAAMC,eAAeD,SAAS;AAE9B,MAAME,0BAA0B,CAACC;IAC/B,IAAI,AAAkB,YAAlB,OAAOA,QACT,OAAOA;IAEP,OAAOA,OAAO,MAAM;AAExB;AAEA,MAAMC,qBAAqB,OACzBC;IAEA,MAAMC,OAAyC,EAAE;IACjD,IAAID,kBAAkB,QAAQ,QAAQ;QACpCC,KAAK,IAAI,CAAC;YACR,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM;gBACR;aACD;QACH;QAEA,KAAK,MAAMC,QAAQF,iBAAiB,MAAM,CAAE;YAC1C,MAAMG,SAAS,MAAMC,mBACnBF,KAAK,GAAG,EACR,CAAC,CAACF,iBAAiB,uBAAuB;YAG5CC,KAAK,IAAI,CAAC;gBACR,MAAM;gBACN,SAAS;oBACP;wBACE,MAAM;wBACN,MAAM,CAAC,mCAAmC,EAAEC,KAAK,IAAI,CAAC,EAAE,CAAC;oBAC3D;iBACD;YACH;YAEAD,KAAK,IAAI,CAAC;gBACR,MAAM;gBACN,SAAS;oBACP;wBACE,MAAM;wBACN,WAAW;4BACT,KAAKE;4BACL,QAAQ;wBACV;oBACF;iBACD;YACH;QACF;IACF;IACA,OAAOF;AACT;AAEO,eAAeI,gBAAgBC,OAKrC;IAUC,MAAM,EAAEC,OAAO,EAAEC,wBAAwB,EAAEC,WAAW,EAAE,GAAGH;IAC3D,MAAM,EAAEI,WAAW,EAAE,GAAGD;IACxB,MAAME,mBAAmBJ,QAAQ,UAAU,CAAC,MAAM;IAElDK,OACEJ,0BACA;IAEF,MAAMK,+BAA+BhB,wBACnCW;IAEF,MAAMM,wBAAwBC,kBAAkBF;IAChD,MAAMG,eAAeC,UAAUP,eAC3BQ,uBAAuBR,eACvBS,4BAA4BT;IAEhC,IAAIU,eAAeT;IACnB,IAAIU,aAAad,QAAQ,IAAI,CAAC,KAAK;IACnC,IAAIe,cAAcf,QAAQ,IAAI,CAAC,MAAM;IACrC,IAAIgB,qBAAqBF;IACzB,IAAIG,sBAAsBF;IAE1B,IAAIhB,QAAQ,YAAY,EAAE;QACxBM,OACEN,QAAQ,YAAY,CAAC,IAAI,EACzB;QAEFM,OACEN,QAAQ,YAAY,CAAC,WAAW,EAChC;QAGFc,eAAed,QAAQ,YAAY,CAAC,WAAW;QAC/Ce,aAAaf,QAAQ,YAAY,CAAC,IAAI,EAAE;QACxCgB,cAAchB,QAAQ,YAAY,CAAC,IAAI,EAAE;QACzCiB,qBAAqBF;QACrBG,sBAAsBF;IACxB,OAAO,IAAIZ,AAAgB,iBAAhBA,aAA8B;QACvC,MAAMe,eAAe,MAAMC,4BAA4BN;QACvDC,aAAaI,aAAa,KAAK;QAC/BH,cAAcG,aAAa,MAAM;QACjCL,eAAeK,aAAa,WAAW;IACzC;IAEA,MAAMxB,OAAe;QACnB;YAAE,MAAM;YAAU,SAASe;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKI;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAMH,UAAUP,eACZ,CAAC,KAAK,EAAEI,uBAAuB,GAC/BA;gBACN;aACD;QACH;KACD;IAED,IAAI,AAAoC,YAApC,OAAON,0BAAuC;QAChD,MAAMmB,SAAS,MAAM5B,mBAAmB;YACtC,QAAQS,yBAAyB,MAAM;YACvC,yBAAyBA,yBAAyB,uBAAuB;QAC3E;QACAP,KAAK,IAAI,IAAI0B;IACf;IAEA,IAAIV,UAAUP,cAAc;QAC1B,MAAM,EAAE,SAASkB,kBAAkB,EAAEC,KAAK,EAAE,GAC1C,MAAMC,yBAAyB7B,MAAMQ;QAEvCf,aAAa,yBAAyBkC;QAEtC,MAAMG,SAASC,2BAA2BJ;QAE1ClC,aAAa,sBAAsBqC,OAAO,KAAK;QAC/CrC,aAAa,yBAAyBqC,OAAO,WAAW;QAExD,IAAIE;QACJ,IAAIC,kBAAyC,EAAE;QAC/C,IAAIC,SAAmB,EAAE;QAEzB,IAAIJ,OAAO,KAAK,IAAI,CAACA,OAAO,WAAW,EAAE;YACvCI,SAAS;gBAACJ,OAAO,KAAK,IAAI;aAAoC;YAC9DrC,aAAa,yBAAyByC,MAAM,CAAC,EAAE;QACjD,OAAO;YACL,MAAM,EAAEC,CAAC,EAAEC,CAAC,EAAE,GAAGN,OAAO,WAAW;YAEnCrC,aAAa,iCAAiC;gBAAE0C;gBAAGC;YAAE;YAIrD,MAAMC,SAASC,KAAK,KAAK,CAAEH,IAAIf,aAAc;YAC7C,MAAMmB,SAASD,KAAK,KAAK,CAAEF,IAAIf,cAAe;YAE9C5B,aAAa,+BAA+B;gBAAE4C;gBAAQE;YAAO;YAG7D,IAAIC,SAASH;YACb,IAAII,SAASF;YACb,IAAIlC,QAAQ,YAAY,EAAE,MAAM;gBAC9BmC,UAAUnC,QAAQ,YAAY,CAAC,IAAI,CAAC,IAAI;gBACxCoC,UAAUpC,QAAQ,YAAY,CAAC,IAAI,CAAC,GAAG;YACzC;YAEA,MAAMqC,UAA+BC,uBACnC;gBAACH;gBAAQC;aAAO,EAChB7B;YAGFoB,UAAUU,QAAQ,IAAI;YACtBjD,aAAa,qBAAqBuC;YAElC,IAAIU,SACFT,kBAAkB;gBAACS;aAAQ;QAE/B;QAEA,OAAO;YACL,MAAMV;YACN,aAAa;gBACX,UAAUC;gBACVC;YACF;YACA,aAAaP;YACbC;YACA,mBAAmBE,OAAO,KAAK;QACjC;IACF;IAEA,IAAIc;IAKJ,IAAI;QACFA,MAAM,MAAMC,yBACV7C,MACAQ;IAEJ,EAAE,OAAOsC,WAAW;QAElB,MAAMC,eACJD,qBAAqBE,QAAQF,UAAU,OAAO,GAAGG,OAAOH;QAC1D,MAAMI,cACJJ,qBAAqBK,uBACjBL,UAAU,WAAW,GACrBC;QACN,MAAMnB,QACJkB,qBAAqBK,uBAAuBL,UAAU,KAAK,GAAGM;QAChE,OAAO;YACL,MAAMA;YACN,aAAa;gBACX,UAAU,EAAE;gBACZ,QAAQ;oBAAC,CAAC,eAAe,EAAEL,cAAc;iBAAC;YAC5C;YACAG;YACAtB;YACA,mBAAmBwB;QACrB;IACF;IAEA,MAAMF,cAAcG,KAAK,SAAS,CAACT,IAAI,OAAO;IAE9C,IAAIZ;IACJ,IAAIC,kBAAyC,EAAE;IAC/C,IAAIC,SACF,YAAYU,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,MAAM,GAAG,EAAE;IACnD,IAAI;QACF,IACE,UAAUA,IAAI,OAAO,IACrBU,MAAM,OAAO,CAACV,IAAI,OAAO,CAAC,IAAI,KAC9BA,IAAI,OAAO,CAAC,IAAI,CAAC,MAAM,IAAI,GAC3B;YACAZ,UAAUuB,gBACRX,IAAI,OAAO,CAAC,IAAI,EAChBxB,YACAC,aACAhB,QAAQ,YAAY,EAAE,MAAM,MAC5BA,QAAQ,YAAY,EAAE,MAAM,KAC5BiB,oBACAC,qBACAd,aACAJ,QAAQ,YAAY,EAAE;YAGxBZ,aAAa,WAAWuC;YAExB,MAAMU,UAA+Bc,sBACnCxB,SACApB;YAEFsB,SAAS,EAAE;YAEX,IAAIQ,SACFT,kBAAkB;gBAACS;aAAQ;QAE/B;IACF,EAAE,OAAOe,GAAG;QACV,MAAMC,MACJD,aAAaT,QACT,CAAC,sBAAsB,EAAES,EAAE,OAAO,EAAE,GACpC;QACN,IAAI,AAACvB,UAAUA,QAAQ,WAAW,GAGhCA,OAAO,IAAI,CAAC,CAAC,CAAC,EAAEwB,IAAI,CAAC,CAAC;aAFtBxB,SAAS;YAACwB;SAAI;IAIlB;IAEA,OAAO;QACL,MAAM1B;QACN,aAAa;YACX,UAAUC;YACV,QAAQC;QACV;QACAgB;QACA,OAAON,IAAI,KAAK;QAChB,mBAAmBA,IAAI,iBAAiB;IAC1C;AACF;AAEO,eAAee,gBAAgBtD,OAIrC;IAQC,MAAM,EAAEC,OAAO,EAAEsD,kBAAkB,EAAEpD,WAAW,EAAE,GAAGH;IACrD,MAAM,EAAEI,WAAW,EAAE,GAAGD;IACxB,MAAME,mBAAmBJ,QAAQ,UAAU,CAAC,MAAM;IAElD,MAAMS,eAAe8C,4BAA4BpD;IACjD,MAAMqD,gCAAgCC,0BACpCnE,wBAAwBgE;IAE1B,MAAM5D,OAAe;QACnB;YAAE,MAAM;YAAU,SAASe;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKL;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAMoD;gBACR;aACD;QACH;KACD;IAED,IAAI,AAA8B,YAA9B,OAAOF,oBAAiC;QAC1C,MAAMlC,SAAS,MAAM5B,mBAAmB;YACtC,QAAQ8D,mBAAmB,MAAM;YACjC,yBAAyBA,mBAAmB,uBAAuB;QACrE;QACA5D,KAAK,IAAI,IAAI0B;IACf;IAEA,IAAIsC;IAGJ,IAAI;QACFA,SAAS,MAAMnB,yBACb7C,MACAQ;IAEJ,EAAE,OAAOsC,WAAW;QAElB,MAAMC,eACJD,qBAAqBE,QAAQF,UAAU,OAAO,GAAGG,OAAOH;QAC1D,MAAMI,cACJJ,qBAAqBK,uBACjBL,UAAU,WAAW,GACrBC;QACN,MAAMnB,QACJkB,qBAAqBK,uBAAuBL,UAAU,KAAK,GAAGM;QAChE,OAAO;YACL,MAAMA;YACN,aAAaA;YACb,OAAO,CAAC,eAAe,EAAEL,cAAc;YACvCG;YACAtB;QACF;IACF;IAEA,IAAIqC;IACJ,MAAMC,cAAcF,OAAO,OAAO,CAAC,IAAI;IACvC,IAAIE,aAAa;QACf,MAAMC,aAAaZ,gBACjBW,aACA5D,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBG;QAEFd,aAAa,0BAA0BwE;QAEvC,MAAMC,oBAAoBJ,OAAO,OAAO,CAAC,eAAe,IAAI,EAAE;QAC9DrE,aAAa,wBAAwByE;QAErC,MAAMC,iBAAiBD,kBACpB,MAAM,CAAC,CAACE,OAAShB,MAAM,OAAO,CAACgB,OAC/B,GAAG,CAAC,CAACA,OACGf,gBACLe,MACAhE,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBG;QAGNd,aAAa,qBAAqB0E;QAGlC,MAAME,aAAaC,WAAW;YAACL;eAAeE;SAAe;QAC7D1E,aAAa,iBAAiB4E;QAE9BN,cAAcQ,iBAAiBF,YAAYjE,QAAQ,IAAI;QACvDX,aAAa,2BAA2BsE;IAC1C;IAEA,IAAIS,cAAchE;IAClB,IAAIiE;IAEJ,IAAIV,aAAa;QACf,MAAMW,gBAAgBX,YAAY,KAAK;QACvC,MAAMY,iBAAiBZ,YAAY,MAAM;QAEzC,MAAMa,gBAAgB,MAAMC,WAC1BrE,kBACAuD,aACAxD,AAAgB,iBAAhBA;QAGF,MAAMuE,aAAa;QACnB,MAAMC,eAAe,MAAMC,WACzBJ,cAAc,WAAW,EACzBE;QAGFN,cAAcO,aAAa,WAAW;QACtCN,QAAQK;QACRf,YAAY,KAAK,GAAGgB,aAAa,KAAK;QACtChB,YAAY,MAAM,GAAGgB,aAAa,MAAM;QAExCtF,aACE,qDACAiF,eACAC,gBACAZ,YAAY,KAAK,EACjBA,YAAY,MAAM,EAClBU;IAEJ;IAEA,OAAO;QACL,MAAMV;QACNS;QACAC;QACA,OAAOX,OAAO,OAAO,CAAC,KAAK;QAC3B,aAAaX,KAAK,SAAS,CAACW,OAAO,OAAO;QAC1C,OAAOA,OAAO,KAAK;IACrB;AACF;AAEO,eAAemB,qBAAwB9E,OAO7C;IACC,MAAM,EAAE+E,SAAS,EAAE9E,OAAO,EAAE+E,aAAa,EAAEtF,gBAAgB,EAAES,WAAW,EAAE,GACxEH;IACF,MAAMU,eAAeuE;IACrB,MAAM5E,mBAAmBJ,QAAQ,UAAU,CAAC,MAAM;IAElD,MAAMiF,wBAAwBC,uBAC5BnF,QAAQ,eAAe,IAAI,IAC3B+E;IAGF,MAAMK,cAAyD,EAAE;IAEjE,IAAIJ,eAAe,uBAAuB,OACxCI,YAAY,IAAI,CAAC;QACf,MAAM;QACN,WAAW;YACT,KAAK/E;YACL,QAAQ;QACV;IACF;IAGF+E,YAAY,IAAI,CAAC;QACf,MAAM;QACN,MAAMF;IACR;IAEA,MAAMvF,OAAe;QACnB;YAAE,MAAM;YAAU,SAASe;QAAa;QACxC;YACE,MAAM;YACN,SAAS0E;QACX;KACD;IAED,IAAI1F,kBAAkB;QACpB,MAAM2B,SAAS,MAAM5B,mBAAmB;YACtC,QAAQC,iBAAiB,MAAM;YAC/B,yBAAyBA,iBAAiB,uBAAuB;QACnE;QACAC,KAAK,IAAI,IAAI0B;IACf;IAEA,MAAM,EACJ,SAASwB,WAAW,EACpBtB,KAAK,EACL8D,iBAAiB,EAClB,GAAG,MAAMC,OAAO3F,MAAMQ;IAGvB,IAAIoF;IACJ,IAAI;QACFA,cAAcC,2BAA8B3C;IAC9C,EAAE,OAAO4C,YAAY;QAEnB,MAAM/C,eACJ+C,sBAAsB9C,QAAQ8C,WAAW,OAAO,GAAG7C,OAAO6C;QAC5D,MAAM,IAAI3C,qBACR,CAAC,iBAAiB,EAAEJ,cAAc,EAClCG,aACAtB;IAEJ;IAEA,OAAO;QACLgE;QACA1C;QACAtB;QACA8D;IACF;AACF;AAEO,eAAeK,sBACpBC,WAAmB,EACnBC,QAAwE,EACxEzF,WAAyB;IAKzB,MAAMO,eAAemF;IACrB,MAAMC,aAAaC,0BAA0BJ;IAE7C,MAAMhG,OAAe;QACnB;YAAE,MAAM;YAAU,SAASe;QAAa;QACxC;YACE,MAAM;YACN,SAASoF;QACX;KACD;IAED,MAAMnC,SAAS,MAAMiC,SAASjG,MAAMQ;IAEpC,OAAO;QACL,kBAAkBwD,OAAO,OAAO,CAAC,gBAAgB,IAAI;QACrD,OAAOA,OAAO,KAAK;IACrB;AACF"}

package/dist/es/ai-model/llm-planning.mjs

@@ -199,6 +199,7 @@ async function plan(userInstruction, opts) {
     if (includeSubGoals) {
         if (planFromAI.updateSubGoals?.length) conversationHistory.setSubGoals(planFromAI.updateSubGoals);
         if (planFromAI.markFinishedIndexes?.length) for (const index of planFromAI.markFinishedIndexes)conversationHistory.markSubGoalFinished(index);
+        if (planFromAI.log) conversationHistory.appendSubGoalLog(planFromAI.log);
     }
     if (planFromAI.memory) conversationHistory.appendMemory(planFromAI.memory);
     conversationHistory.append({

package/dist/es/ai-model/llm-planning.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"ai-model/llm-planning.mjs","sources":["../../../src/ai-model/llm-planning.ts"],"sourcesContent":["import type {\n  DeepThinkOption,\n  DeviceAction,\n  InterfaceType,\n  PlanningAIResponse,\n  RawResponsePlanningAIResponse,\n  UIContext,\n} from '@/types';\nimport type { IModelConfig, TModelFamily } from '@midscene/shared/env';\nimport { paddingToMatchBlockByBase64 } from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport type { ChatCompletionMessageParam } from 'openai/resources/index';\nimport {\n  buildYamlFlowFromPlans,\n  fillBboxParam,\n  findAllMidsceneLocatorField,\n} from '../common';\nimport type { ConversationHistory } from './conversation-history';\nimport { systemPromptToTaskPlanning } from './prompt/llm-planning';\nimport {\n  extractXMLTag,\n  parseMarkFinishedIndexes,\n  parseSubGoalsFromXML,\n} from './prompt/util';\nimport {\n  AIResponseParseError,\n  callAI,\n  safeParseJson,\n} from './service-caller/index';\n\nconst debug = getDebug('planning');\n\n/**\n * Parse XML response from LLM and convert to RawResponsePlanningAIResponse\n */\nexport function parseXMLPlanningResponse(\n  xmlString: string,\n  modelFamily: TModelFamily | undefined,\n): RawResponsePlanningAIResponse {\n  const thought = extractXMLTag(xmlString, 'thought');\n  const memory = extractXMLTag(xmlString, 'memory');\n  const log = extractXMLTag(xmlString, 'log') || '';\n  const error = extractXMLTag(xmlString, 'error');\n  const actionType = extractXMLTag(xmlString, 'action-type');\n  const actionParamStr = extractXMLTag(xmlString, 'action-param-json');\n\n  // Parse complete-goal tag with success attribute\n  const completeGoalRegex =\n    /<complete-goal\\s+success=\"(true|false)\">([\\s\\S]*?)<\\/complete-goal>/i;\n  const completeGoalMatch = xmlString.match(completeGoalRegex);\n  let finalizeMessage: string | undefined;\n  let finalizeSuccess: boolean | undefined;\n\n  if (completeGoalMatch) {\n    finalizeSuccess = completeGoalMatch[1] === 'true';\n    finalizeMessage = completeGoalMatch[2]?.trim() || undefined;\n  }\n\n  // Parse sub-goal related tags\n  const updatePlanContent = extractXMLTag(xmlString, 'update-plan-content');\n  const markSubGoalDone = extractXMLTag(xmlString, 'mark-sub-goal-done');\n\n  const updateSubGoals = updatePlanContent\n    ? parseSubGoalsFromXML(updatePlanContent)\n    : undefined;\n  const markFinishedIndexes = markSubGoalDone\n    ? parseMarkFinishedIndexes(markSubGoalDone)\n    : undefined;\n\n  // Parse action\n  let action: any = null;\n  if (actionType && actionType.toLowerCase() !== 'null') {\n    const type = actionType.trim();\n    let param: any = undefined;\n\n    if (actionParamStr) {\n      try {\n        // Parse the JSON string in action-param-json\n        param = safeParseJson(actionParamStr, modelFamily);\n      } catch (e) {\n        throw new Error(`Failed to parse action-param-json: ${e}`);\n      }\n    }\n\n    action = {\n      type,\n      ...(param !== undefined ? { param } : {}),\n    };\n  }\n\n  return {\n    ...(thought ? { thought } : {}),\n    ...(memory ? { memory } : {}),\n    log,\n    ...(error ? { error } : {}),\n    action,\n    ...(finalizeMessage !== undefined ? { finalizeMessage } : {}),\n    ...(finalizeSuccess !== undefined ? { finalizeSuccess } : {}),\n    ...(updateSubGoals?.length ? { updateSubGoals } : {}),\n    ...(markFinishedIndexes?.length ? { markFinishedIndexes } : {}),\n  };\n}\n\nexport async function plan(\n  userInstruction: string,\n  opts: {\n    context: UIContext;\n    interfaceType: InterfaceType;\n    actionSpace: DeviceAction<any>[];\n    actionContext?: string;\n    modelConfig: IModelConfig;\n    conversationHistory: ConversationHistory;\n    includeBbox: boolean;\n    imagesIncludeCount?: number;\n    deepThink?: DeepThinkOption;\n  },\n): Promise<PlanningAIResponse> {\n  const { context, modelConfig, conversationHistory } = opts;\n  const { size } = context;\n  const screenshotBase64 = context.screenshot.base64;\n\n  const { modelFamily } = modelConfig;\n\n  // Only enable sub-goals when deepThink is true\n  const includeSubGoals = opts.deepThink === true;\n\n  const systemPrompt = await systemPromptToTaskPlanning({\n    actionSpace: opts.actionSpace,\n    modelFamily,\n    includeBbox: opts.includeBbox,\n    includeThought: true, // always include thought\n    includeSubGoals,\n  });\n\n  let imagePayload = screenshotBase64;\n  let imageWidth = size.width;\n  let imageHeight = size.height;\n  const rightLimit = imageWidth;\n  const bottomLimit = imageHeight;\n\n  // Process image based on VL mode requirements\n  if (modelFamily === 'qwen2.5-vl') {\n    const paddedResult = await paddingToMatchBlockByBase64(imagePayload);\n    imageWidth = paddedResult.width;\n    imageHeight = paddedResult.height;\n    imagePayload = paddedResult.imageBase64;\n  }\n\n  const actionContext = opts.actionContext\n    ? `<high_priority_knowledge>${opts.actionContext}</high_priority_knowledge>\\n`\n    : '';\n\n  const instruction: ChatCompletionMessageParam[] = [\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `${actionContext}<user_instruction>${userInstruction}</user_instruction>`,\n        },\n      ],\n    },\n  ];\n\n  let latestFeedbackMessage: ChatCompletionMessageParam;\n\n  // Build sub-goal status text to include in the message (only when deepThink is enabled)\n  const subGoalsText = includeSubGoals\n    ? conversationHistory.subGoalsToText()\n    : '';\n  const subGoalsSection = subGoalsText ? `\\n\\n${subGoalsText}` : '';\n\n  // Build memories text to include in the message\n  const memoriesText = conversationHistory.memoriesToText();\n  const memoriesSection = memoriesText ? `\\n\\n${memoriesText}` : '';\n\n  if (conversationHistory.pendingFeedbackMessage) {\n    latestFeedbackMessage = {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `${conversationHistory.pendingFeedbackMessage}. The previous action has been executed, here is the latest screenshot. Please continue according to the instruction.${memoriesSection}${subGoalsSection}`,\n        },\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n      ],\n    };\n\n    conversationHistory.resetPendingFeedbackMessageIfExists();\n  } else {\n    latestFeedbackMessage = {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `this is the latest screenshot${memoriesSection}${subGoalsSection}`,\n        },\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n      ],\n    };\n  }\n  conversationHistory.append(latestFeedbackMessage);\n\n  // Compress history if it exceeds the threshold to avoid context overflow\n  conversationHistory.compressHistory(50, 20);\n\n  const historyLog = conversationHistory.snapshot(opts.imagesIncludeCount);\n\n  const msgs: ChatCompletionMessageParam[] = [\n    { role: 'system', content: systemPrompt },\n    ...instruction,\n    ...historyLog,\n  ];\n\n  const {\n    content: rawResponse,\n    usage,\n    reasoning_content,\n  } = await callAI(msgs, modelConfig, {\n    deepThink: opts.deepThink === 'unset' ? undefined : opts.deepThink,\n  });\n\n  // Parse XML response to JSON object, capture parsing errors\n  let planFromAI: RawResponsePlanningAIResponse;\n  try {\n    planFromAI = parseXMLPlanningResponse(rawResponse, modelFamily);\n  } catch (parseError) {\n    // Throw AIResponseParseError with usage and rawResponse preserved\n    const errorMessage =\n      parseError instanceof Error ? parseError.message : String(parseError);\n    throw new AIResponseParseError(\n      `XML parse error: ${errorMessage}`,\n      rawResponse,\n      usage,\n    );\n  }\n\n  if (planFromAI.action && planFromAI.finalizeSuccess !== undefined) {\n    console.warn(\n      'Planning response included both an action and complete-goal; ignoring complete-goal output.',\n    );\n    planFromAI.finalizeMessage = undefined;\n    planFromAI.finalizeSuccess = undefined;\n  }\n\n  const actions = planFromAI.action ? [planFromAI.action] : [];\n  let shouldContinuePlanning = true;\n\n  // Check if goal is completed via complete-goal tag\n  if (planFromAI.finalizeSuccess !== undefined) {\n    debug('goal completed via complete-goal tag, stop planning');\n    shouldContinuePlanning = false;\n    // Mark all sub-goals as finished when goal is completed (only when deepThink is enabled)\n    if (includeSubGoals) {\n      conversationHistory.markAllSubGoalsFinished();\n    }\n  }\n\n  const returnValue: PlanningAIResponse = {\n    ...planFromAI,\n    actions,\n    rawResponse,\n    usage,\n    reasoning_content,\n    yamlFlow: buildYamlFlowFromPlans(actions, opts.actionSpace),\n    shouldContinuePlanning,\n  };\n\n  assert(planFromAI, \"can't get plans from AI\");\n\n  actions.forEach((action) => {\n    const type = action.type;\n    const actionInActionSpace = opts.actionSpace.find(\n      (action) => action.name === type,\n    );\n\n    debug('actionInActionSpace matched', actionInActionSpace);\n    const locateFields = actionInActionSpace\n      ? findAllMidsceneLocatorField(actionInActionSpace.paramSchema)\n      : [];\n\n    debug('locateFields', locateFields);\n\n    locateFields.forEach((field) => {\n      const locateResult = action.param[field];\n      if (locateResult && modelFamily !== undefined) {\n        // Always use model family to fill bbox parameters\n        action.param[field] = fillBboxParam(\n          locateResult,\n          imageWidth,\n          imageHeight,\n          modelFamily,\n        );\n      }\n    });\n  });\n\n  // Update sub-goals in conversation history based on response (only when deepThink is enabled)\n  if (includeSubGoals) {\n    if (planFromAI.updateSubGoals?.length) {\n      conversationHistory.setSubGoals(planFromAI.updateSubGoals);\n    }\n    if (planFromAI.markFinishedIndexes?.length) {\n      for (const index of planFromAI.markFinishedIndexes) {\n        conversationHistory.markSubGoalFinished(index);\n      }\n    }\n  }\n\n  // Append memory to conversation history if present\n  if (planFromAI.memory) {\n    conversationHistory.appendMemory(planFromAI.memory);\n  }\n\n  conversationHistory.append({\n    role: 'assistant',\n    content: [\n      {\n        type: 'text',\n        text: rawResponse,\n      },\n    ],\n  });\n\n  return returnValue;\n}\n"],"names":["debug","getDebug","parseXMLPlanningResponse","xmlString","modelFamily","thought","extractXMLTag","memory","log","error","actionType","actionParamStr","completeGoalRegex","completeGoalMatch","finalizeMessage","finalizeSuccess","undefined","updatePlanContent","markSubGoalDone","updateSubGoals","parseSubGoalsFromXML","markFinishedIndexes","parseMarkFinishedIndexes","action","type","param","safeParseJson","e","Error","plan","userInstruction","opts","context","modelConfig","conversationHistory","size","screenshotBase64","includeSubGoals","systemPrompt","systemPromptToTaskPlanning","imagePayload","imageWidth","imageHeight","paddedResult","paddingToMatchBlockByBase64","actionContext","instruction","latestFeedbackMessage","subGoalsText","subGoalsSection","memoriesText","memoriesSection","historyLog","msgs","rawResponse","usage","reasoning_content","callAI","planFromAI","parseError","errorMessage","String","AIResponseParseError","console","actions","shouldContinuePlanning","returnValue","buildYamlFlowFromPlans","assert","actionInActionSpace","locateFields","findAllMidsceneLocatorField","field","locateResult","fillBboxParam","index"],"mappings":";;;;;;;AA+BA,MAAMA,QAAQC,SAAS;AAKhB,SAASC,yBACdC,SAAiB,EACjBC,WAAqC;IAErC,MAAMC,UAAUC,cAAcH,WAAW;IACzC,MAAMI,SAASD,cAAcH,WAAW;IACxC,MAAMK,MAAMF,cAAcH,WAAW,UAAU;IAC/C,MAAMM,QAAQH,cAAcH,WAAW;IACvC,MAAMO,aAAaJ,cAAcH,WAAW;IAC5C,MAAMQ,iBAAiBL,cAAcH,WAAW;IAGhD,MAAMS,oBACJ;IACF,MAAMC,oBAAoBV,UAAU,KAAK,CAACS;IAC1C,IAAIE;IACJ,IAAIC;IAEJ,IAAIF,mBAAmB;QACrBE,kBAAkBF,AAAyB,WAAzBA,iBAAiB,CAAC,EAAE;QACtCC,kBAAkBD,iBAAiB,CAAC,EAAE,EAAE,UAAUG;IACpD;IAGA,MAAMC,oBAAoBX,cAAcH,WAAW;IACnD,MAAMe,kBAAkBZ,cAAcH,WAAW;IAEjD,MAAMgB,iBAAiBF,oBACnBG,qBAAqBH,qBACrBD;IACJ,MAAMK,sBAAsBH,kBACxBI,yBAAyBJ,mBACzBF;IAGJ,IAAIO,SAAc;IAClB,IAAIb,cAAcA,AAA6B,WAA7BA,WAAW,WAAW,IAAe;QACrD,MAAMc,OAAOd,WAAW,IAAI;QAC5B,IAAIe;QAEJ,IAAId,gBACF,IAAI;YAEFc,QAAQC,cAAcf,gBAAgBP;QACxC,EAAE,OAAOuB,GAAG;YACV,MAAM,IAAIC,MAAM,CAAC,mCAAmC,EAAED,GAAG;QAC3D;QAGFJ,SAAS;YACPC;YACA,GAAIC,AAAUT,WAAVS,QAAsB;gBAAEA;YAAM,IAAI,CAAC,CAAC;QAC1C;IACF;IAEA,OAAO;QACL,GAAIpB,UAAU;YAAEA;QAAQ,IAAI,CAAC,CAAC;QAC9B,GAAIE,SAAS;YAAEA;QAAO,IAAI,CAAC,CAAC;QAC5BC;QACA,GAAIC,QAAQ;YAAEA;QAAM,IAAI,CAAC,CAAC;QAC1Bc;QACA,GAAIT,AAAoBE,WAApBF,kBAAgC;YAAEA;QAAgB,IAAI,CAAC,CAAC;QAC5D,GAAIC,AAAoBC,WAApBD,kBAAgC;YAAEA;QAAgB,IAAI,CAAC,CAAC;QAC5D,GAAII,gBAAgB,SAAS;YAAEA;QAAe,IAAI,CAAC,CAAC;QACpD,GAAIE,qBAAqB,SAAS;YAAEA;QAAoB,IAAI,CAAC,CAAC;IAChE;AACF;AAEO,eAAeQ,KACpBC,eAAuB,EACvBC,IAUC;IAED,MAAM,EAAEC,OAAO,EAAEC,WAAW,EAAEC,mBAAmB,EAAE,GAAGH;IACtD,MAAM,EAAEI,IAAI,EAAE,GAAGH;IACjB,MAAMI,mBAAmBJ,QAAQ,UAAU,CAAC,MAAM;IAElD,MAAM,EAAE5B,WAAW,EAAE,GAAG6B;IAGxB,MAAMI,kBAAkBN,AAAmB,SAAnBA,KAAK,SAAS;IAEtC,MAAMO,eAAe,MAAMC,2BAA2B;QACpD,aAAaR,KAAK,WAAW;QAC7B3B;QACA,aAAa2B,KAAK,WAAW;QAC7B,gBAAgB;QAChBM;IACF;IAEA,IAAIG,eAAeJ;IACnB,IAAIK,aAAaN,KAAK,KAAK;IAC3B,IAAIO,cAAcP,KAAK,MAAM;IAK7B,IAAI/B,AAAgB,iBAAhBA,aAA8B;QAChC,MAAMuC,eAAe,MAAMC,4BAA4BJ;QACvDC,aAAaE,aAAa,KAAK;QAC/BD,cAAcC,aAAa,MAAM;QACjCH,eAAeG,aAAa,WAAW;IACzC;IAEA,MAAME,gBAAgBd,KAAK,aAAa,GACpC,CAAC,yBAAyB,EAAEA,KAAK,aAAa,CAAC,4BAA4B,CAAC,GAC5E;IAEJ,MAAMe,cAA4C;QAChD;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,GAAGD,cAAc,kBAAkB,EAAEf,gBAAgB,mBAAmB,CAAC;gBACjF;aACD;QACH;KACD;IAED,IAAIiB;IAGJ,MAAMC,eAAeX,kBACjBH,oBAAoB,cAAc,KAClC;IACJ,MAAMe,kBAAkBD,eAAe,CAAC,IAAI,EAAEA,cAAc,GAAG;IAG/D,MAAME,eAAehB,oBAAoB,cAAc;IACvD,MAAMiB,kBAAkBD,eAAe,CAAC,IAAI,EAAEA,cAAc,GAAG;IAE/D,IAAIhB,oBAAoB,sBAAsB,EAAE;QAC9Ca,wBAAwB;YACtB,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,GAAGb,oBAAoB,sBAAsB,CAAC,qHAAqH,EAAEiB,kBAAkBF,iBAAiB;gBAChN;gBACA;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKT;wBACL,QAAQ;oBACV;gBACF;aACD;QACH;QAEAN,oBAAoB,mCAAmC;IACzD,OACEa,wBAAwB;QACtB,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAM,CAAC,6BAA6B,EAAEI,kBAAkBF,iBAAiB;YAC3E;YACA;gBACE,MAAM;gBACN,WAAW;oBACT,KAAKT;oBACL,QAAQ;gBACV;YACF;SACD;IACH;IAEFN,oBAAoB,MAAM,CAACa;IAG3Bb,oBAAoB,eAAe,CAAC,IAAI;IAExC,MAAMkB,aAAalB,oBAAoB,QAAQ,CAACH,KAAK,kBAAkB;IAEvE,MAAMsB,OAAqC;QACzC;YAAE,MAAM;YAAU,SAASf;QAAa;WACrCQ;WACAM;KACJ;IAED,MAAM,EACJ,SAASE,WAAW,EACpBC,KAAK,EACLC,iBAAiB,EAClB,GAAG,MAAMC,OAAOJ,MAAMpB,aAAa;QAClC,WAAWF,AAAmB,YAAnBA,KAAK,SAAS,GAAef,SAAYe,KAAK,SAAS;IACpE;IAGA,IAAI2B;IACJ,IAAI;QACFA,aAAaxD,yBAAyBoD,aAAalD;IACrD,EAAE,OAAOuD,YAAY;QAEnB,MAAMC,eACJD,sBAAsB/B,QAAQ+B,WAAW,OAAO,GAAGE,OAAOF;QAC5D,MAAM,IAAIG,qBACR,CAAC,iBAAiB,EAAEF,cAAc,EAClCN,aACAC;IAEJ;IAEA,IAAIG,WAAW,MAAM,IAAIA,AAA+B1C,WAA/B0C,WAAW,eAAe,EAAgB;QACjEK,QAAQ,IAAI,CACV;QAEFL,WAAW,eAAe,GAAG1C;QAC7B0C,WAAW,eAAe,GAAG1C;IAC/B;IAEA,MAAMgD,UAAUN,WAAW,MAAM,GAAG;QAACA,WAAW,MAAM;KAAC,GAAG,EAAE;IAC5D,IAAIO,yBAAyB;IAG7B,IAAIP,AAA+B1C,WAA/B0C,WAAW,eAAe,EAAgB;QAC5C1D,MAAM;QACNiE,yBAAyB;QAEzB,IAAI5B,iBACFH,oBAAoB,uBAAuB;IAE/C;IAEA,MAAMgC,cAAkC;QACtC,GAAGR,UAAU;QACbM;QACAV;QACAC;QACAC;QACA,UAAUW,uBAAuBH,SAASjC,KAAK,WAAW;QAC1DkC;IACF;IAEAG,OAAOV,YAAY;IAEnBM,QAAQ,OAAO,CAAC,CAACzC;QACf,MAAMC,OAAOD,OAAO,IAAI;QACxB,MAAM8C,sBAAsBtC,KAAK,WAAW,CAAC,IAAI,CAC/C,CAACR,SAAWA,OAAO,IAAI,KAAKC;QAG9BxB,MAAM,+BAA+BqE;QACrC,MAAMC,eAAeD,sBACjBE,4BAA4BF,oBAAoB,WAAW,IAC3D,EAAE;QAENrE,MAAM,gBAAgBsE;QAEtBA,aAAa,OAAO,CAAC,CAACE;YACpB,MAAMC,eAAelD,OAAO,KAAK,CAACiD,MAAM;YACxC,IAAIC,gBAAgBrE,AAAgBY,WAAhBZ,aAElBmB,OAAO,KAAK,CAACiD,MAAM,GAAGE,cACpBD,cACAhC,YACAC,aACAtC;QAGN;IACF;IAGA,IAAIiC,iBAAiB;QACnB,IAAIqB,WAAW,cAAc,EAAE,QAC7BxB,oBAAoB,WAAW,CAACwB,WAAW,cAAc;QAE3D,IAAIA,WAAW,mBAAmB,EAAE,QAClC,KAAK,MAAMiB,SAASjB,WAAW,mBAAmB,CAChDxB,oBAAoB,mBAAmB,CAACyC;IAG9C;IAGA,IAAIjB,WAAW,MAAM,EACnBxB,oBAAoB,YAAY,CAACwB,WAAW,MAAM;IAGpDxB,oBAAoB,MAAM,CAAC;QACzB,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAMoB;YACR;SACD;IACH;IAEA,OAAOY;AACT"}
+{"version":3,"file":"ai-model/llm-planning.mjs","sources":["../../../src/ai-model/llm-planning.ts"],"sourcesContent":["import type {\n  DeepThinkOption,\n  DeviceAction,\n  InterfaceType,\n  PlanningAIResponse,\n  RawResponsePlanningAIResponse,\n  UIContext,\n} from '@/types';\nimport type { IModelConfig, TModelFamily } from '@midscene/shared/env';\nimport { paddingToMatchBlockByBase64 } from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport type { ChatCompletionMessageParam } from 'openai/resources/index';\nimport {\n  buildYamlFlowFromPlans,\n  fillBboxParam,\n  findAllMidsceneLocatorField,\n} from '../common';\nimport type { ConversationHistory } from './conversation-history';\nimport { systemPromptToTaskPlanning } from './prompt/llm-planning';\nimport {\n  extractXMLTag,\n  parseMarkFinishedIndexes,\n  parseSubGoalsFromXML,\n} from './prompt/util';\nimport {\n  AIResponseParseError,\n  callAI,\n  safeParseJson,\n} from './service-caller/index';\n\nconst debug = getDebug('planning');\n\n/**\n * Parse XML response from LLM and convert to RawResponsePlanningAIResponse\n */\nexport function parseXMLPlanningResponse(\n  xmlString: string,\n  modelFamily: TModelFamily | undefined,\n): RawResponsePlanningAIResponse {\n  const thought = extractXMLTag(xmlString, 'thought');\n  const memory = extractXMLTag(xmlString, 'memory');\n  const log = extractXMLTag(xmlString, 'log') || '';\n  const error = extractXMLTag(xmlString, 'error');\n  const actionType = extractXMLTag(xmlString, 'action-type');\n  const actionParamStr = extractXMLTag(xmlString, 'action-param-json');\n\n  // Parse complete-goal tag with success attribute\n  const completeGoalRegex =\n    /<complete-goal\\s+success=\"(true|false)\">([\\s\\S]*?)<\\/complete-goal>/i;\n  const completeGoalMatch = xmlString.match(completeGoalRegex);\n  let finalizeMessage: string | undefined;\n  let finalizeSuccess: boolean | undefined;\n\n  if (completeGoalMatch) {\n    finalizeSuccess = completeGoalMatch[1] === 'true';\n    finalizeMessage = completeGoalMatch[2]?.trim() || undefined;\n  }\n\n  // Parse sub-goal related tags\n  const updatePlanContent = extractXMLTag(xmlString, 'update-plan-content');\n  const markSubGoalDone = extractXMLTag(xmlString, 'mark-sub-goal-done');\n\n  const updateSubGoals = updatePlanContent\n    ? parseSubGoalsFromXML(updatePlanContent)\n    : undefined;\n  const markFinishedIndexes = markSubGoalDone\n    ? parseMarkFinishedIndexes(markSubGoalDone)\n    : undefined;\n\n  // Parse action\n  let action: any = null;\n  if (actionType && actionType.toLowerCase() !== 'null') {\n    const type = actionType.trim();\n    let param: any = undefined;\n\n    if (actionParamStr) {\n      try {\n        // Parse the JSON string in action-param-json\n        param = safeParseJson(actionParamStr, modelFamily);\n      } catch (e) {\n        throw new Error(`Failed to parse action-param-json: ${e}`);\n      }\n    }\n\n    action = {\n      type,\n      ...(param !== undefined ? { param } : {}),\n    };\n  }\n\n  return {\n    ...(thought ? { thought } : {}),\n    ...(memory ? { memory } : {}),\n    log,\n    ...(error ? { error } : {}),\n    action,\n    ...(finalizeMessage !== undefined ? { finalizeMessage } : {}),\n    ...(finalizeSuccess !== undefined ? { finalizeSuccess } : {}),\n    ...(updateSubGoals?.length ? { updateSubGoals } : {}),\n    ...(markFinishedIndexes?.length ? { markFinishedIndexes } : {}),\n  };\n}\n\nexport async function plan(\n  userInstruction: string,\n  opts: {\n    context: UIContext;\n    interfaceType: InterfaceType;\n    actionSpace: DeviceAction<any>[];\n    actionContext?: string;\n    modelConfig: IModelConfig;\n    conversationHistory: ConversationHistory;\n    includeBbox: boolean;\n    imagesIncludeCount?: number;\n    deepThink?: DeepThinkOption;\n  },\n): Promise<PlanningAIResponse> {\n  const { context, modelConfig, conversationHistory } = opts;\n  const { size } = context;\n  const screenshotBase64 = context.screenshot.base64;\n\n  const { modelFamily } = modelConfig;\n\n  // Only enable sub-goals when deepThink is true\n  const includeSubGoals = opts.deepThink === true;\n\n  const systemPrompt = await systemPromptToTaskPlanning({\n    actionSpace: opts.actionSpace,\n    modelFamily,\n    includeBbox: opts.includeBbox,\n    includeThought: true, // always include thought\n    includeSubGoals,\n  });\n\n  let imagePayload = screenshotBase64;\n  let imageWidth = size.width;\n  let imageHeight = size.height;\n  const rightLimit = imageWidth;\n  const bottomLimit = imageHeight;\n\n  // Process image based on VL mode requirements\n  if (modelFamily === 'qwen2.5-vl') {\n    const paddedResult = await paddingToMatchBlockByBase64(imagePayload);\n    imageWidth = paddedResult.width;\n    imageHeight = paddedResult.height;\n    imagePayload = paddedResult.imageBase64;\n  }\n\n  const actionContext = opts.actionContext\n    ? `<high_priority_knowledge>${opts.actionContext}</high_priority_knowledge>\\n`\n    : '';\n\n  const instruction: ChatCompletionMessageParam[] = [\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `${actionContext}<user_instruction>${userInstruction}</user_instruction>`,\n        },\n      ],\n    },\n  ];\n\n  let latestFeedbackMessage: ChatCompletionMessageParam;\n\n  // Build sub-goal status text to include in the message (only when deepThink is enabled)\n  const subGoalsText = includeSubGoals\n    ? conversationHistory.subGoalsToText()\n    : '';\n  const subGoalsSection = subGoalsText ? `\\n\\n${subGoalsText}` : '';\n\n  // Build memories text to include in the message\n  const memoriesText = conversationHistory.memoriesToText();\n  const memoriesSection = memoriesText ? `\\n\\n${memoriesText}` : '';\n\n  if (conversationHistory.pendingFeedbackMessage) {\n    latestFeedbackMessage = {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `${conversationHistory.pendingFeedbackMessage}. The previous action has been executed, here is the latest screenshot. Please continue according to the instruction.${memoriesSection}${subGoalsSection}`,\n        },\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n      ],\n    };\n\n    conversationHistory.resetPendingFeedbackMessageIfExists();\n  } else {\n    latestFeedbackMessage = {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `this is the latest screenshot${memoriesSection}${subGoalsSection}`,\n        },\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n      ],\n    };\n  }\n  conversationHistory.append(latestFeedbackMessage);\n\n  // Compress history if it exceeds the threshold to avoid context overflow\n  conversationHistory.compressHistory(50, 20);\n\n  const historyLog = conversationHistory.snapshot(opts.imagesIncludeCount);\n\n  const msgs: ChatCompletionMessageParam[] = [\n    { role: 'system', content: systemPrompt },\n    ...instruction,\n    ...historyLog,\n  ];\n\n  const {\n    content: rawResponse,\n    usage,\n    reasoning_content,\n  } = await callAI(msgs, modelConfig, {\n    deepThink: opts.deepThink === 'unset' ? undefined : opts.deepThink,\n  });\n\n  // Parse XML response to JSON object, capture parsing errors\n  let planFromAI: RawResponsePlanningAIResponse;\n  try {\n    planFromAI = parseXMLPlanningResponse(rawResponse, modelFamily);\n  } catch (parseError) {\n    // Throw AIResponseParseError with usage and rawResponse preserved\n    const errorMessage =\n      parseError instanceof Error ? parseError.message : String(parseError);\n    throw new AIResponseParseError(\n      `XML parse error: ${errorMessage}`,\n      rawResponse,\n      usage,\n    );\n  }\n\n  if (planFromAI.action && planFromAI.finalizeSuccess !== undefined) {\n    console.warn(\n      'Planning response included both an action and complete-goal; ignoring complete-goal output.',\n    );\n    planFromAI.finalizeMessage = undefined;\n    planFromAI.finalizeSuccess = undefined;\n  }\n\n  const actions = planFromAI.action ? [planFromAI.action] : [];\n  let shouldContinuePlanning = true;\n\n  // Check if goal is completed via complete-goal tag\n  if (planFromAI.finalizeSuccess !== undefined) {\n    debug('goal completed via complete-goal tag, stop planning');\n    shouldContinuePlanning = false;\n    // Mark all sub-goals as finished when goal is completed (only when deepThink is enabled)\n    if (includeSubGoals) {\n      conversationHistory.markAllSubGoalsFinished();\n    }\n  }\n\n  const returnValue: PlanningAIResponse = {\n    ...planFromAI,\n    actions,\n    rawResponse,\n    usage,\n    reasoning_content,\n    yamlFlow: buildYamlFlowFromPlans(actions, opts.actionSpace),\n    shouldContinuePlanning,\n  };\n\n  assert(planFromAI, \"can't get plans from AI\");\n\n  actions.forEach((action) => {\n    const type = action.type;\n    const actionInActionSpace = opts.actionSpace.find(\n      (action) => action.name === type,\n    );\n\n    debug('actionInActionSpace matched', actionInActionSpace);\n    const locateFields = actionInActionSpace\n      ? findAllMidsceneLocatorField(actionInActionSpace.paramSchema)\n      : [];\n\n    debug('locateFields', locateFields);\n\n    locateFields.forEach((field) => {\n      const locateResult = action.param[field];\n      if (locateResult && modelFamily !== undefined) {\n        // Always use model family to fill bbox parameters\n        action.param[field] = fillBboxParam(\n          locateResult,\n          imageWidth,\n          imageHeight,\n          modelFamily,\n        );\n      }\n    });\n  });\n\n  // Update sub-goals in conversation history based on response (only when deepThink is enabled)\n  if (includeSubGoals) {\n    if (planFromAI.updateSubGoals?.length) {\n      conversationHistory.setSubGoals(planFromAI.updateSubGoals);\n    }\n    if (planFromAI.markFinishedIndexes?.length) {\n      for (const index of planFromAI.markFinishedIndexes) {\n        conversationHistory.markSubGoalFinished(index);\n      }\n    }\n    // Append the planning log to the currently running sub-goal\n    if (planFromAI.log) {\n      conversationHistory.appendSubGoalLog(planFromAI.log);\n    }\n  }\n\n  // Append memory to conversation history if present\n  if (planFromAI.memory) {\n    conversationHistory.appendMemory(planFromAI.memory);\n  }\n\n  conversationHistory.append({\n    role: 'assistant',\n    content: [\n      {\n        type: 'text',\n        text: rawResponse,\n      },\n    ],\n  });\n\n  return returnValue;\n}\n"],"names":["debug","getDebug","parseXMLPlanningResponse","xmlString","modelFamily","thought","extractXMLTag","memory","log","error","actionType","actionParamStr","completeGoalRegex","completeGoalMatch","finalizeMessage","finalizeSuccess","undefined","updatePlanContent","markSubGoalDone","updateSubGoals","parseSubGoalsFromXML","markFinishedIndexes","parseMarkFinishedIndexes","action","type","param","safeParseJson","e","Error","plan","userInstruction","opts","context","modelConfig","conversationHistory","size","screenshotBase64","includeSubGoals","systemPrompt","systemPromptToTaskPlanning","imagePayload","imageWidth","imageHeight","paddedResult","paddingToMatchBlockByBase64","actionContext","instruction","latestFeedbackMessage","subGoalsText","subGoalsSection","memoriesText","memoriesSection","historyLog","msgs","rawResponse","usage","reasoning_content","callAI","planFromAI","parseError","errorMessage","String","AIResponseParseError","console","actions","shouldContinuePlanning","returnValue","buildYamlFlowFromPlans","assert","actionInActionSpace","locateFields","findAllMidsceneLocatorField","field","locateResult","fillBboxParam","index"],"mappings":";;;;;;;AA+BA,MAAMA,QAAQC,SAAS;AAKhB,SAASC,yBACdC,SAAiB,EACjBC,WAAqC;IAErC,MAAMC,UAAUC,cAAcH,WAAW;IACzC,MAAMI,SAASD,cAAcH,WAAW;IACxC,MAAMK,MAAMF,cAAcH,WAAW,UAAU;IAC/C,MAAMM,QAAQH,cAAcH,WAAW;IACvC,MAAMO,aAAaJ,cAAcH,WAAW;IAC5C,MAAMQ,iBAAiBL,cAAcH,WAAW;IAGhD,MAAMS,oBACJ;IACF,MAAMC,oBAAoBV,UAAU,KAAK,CAACS;IAC1C,IAAIE;IACJ,IAAIC;IAEJ,IAAIF,mBAAmB;QACrBE,kBAAkBF,AAAyB,WAAzBA,iBAAiB,CAAC,EAAE;QACtCC,kBAAkBD,iBAAiB,CAAC,EAAE,EAAE,UAAUG;IACpD;IAGA,MAAMC,oBAAoBX,cAAcH,WAAW;IACnD,MAAMe,kBAAkBZ,cAAcH,WAAW;IAEjD,MAAMgB,iBAAiBF,oBACnBG,qBAAqBH,qBACrBD;IACJ,MAAMK,sBAAsBH,kBACxBI,yBAAyBJ,mBACzBF;IAGJ,IAAIO,SAAc;IAClB,IAAIb,cAAcA,AAA6B,WAA7BA,WAAW,WAAW,IAAe;QACrD,MAAMc,OAAOd,WAAW,IAAI;QAC5B,IAAIe;QAEJ,IAAId,gBACF,IAAI;YAEFc,QAAQC,cAAcf,gBAAgBP;QACxC,EAAE,OAAOuB,GAAG;YACV,MAAM,IAAIC,MAAM,CAAC,mCAAmC,EAAED,GAAG;QAC3D;QAGFJ,SAAS;YACPC;YACA,GAAIC,AAAUT,WAAVS,QAAsB;gBAAEA;YAAM,IAAI,CAAC,CAAC;QAC1C;IACF;IAEA,OAAO;QACL,GAAIpB,UAAU;YAAEA;QAAQ,IAAI,CAAC,CAAC;QAC9B,GAAIE,SAAS;YAAEA;QAAO,IAAI,CAAC,CAAC;QAC5BC;QACA,GAAIC,QAAQ;YAAEA;QAAM,IAAI,CAAC,CAAC;QAC1Bc;QACA,GAAIT,AAAoBE,WAApBF,kBAAgC;YAAEA;QAAgB,IAAI,CAAC,CAAC;QAC5D,GAAIC,AAAoBC,WAApBD,kBAAgC;YAAEA;QAAgB,IAAI,CAAC,CAAC;QAC5D,GAAII,gBAAgB,SAAS;YAAEA;QAAe,IAAI,CAAC,CAAC;QACpD,GAAIE,qBAAqB,SAAS;YAAEA;QAAoB,IAAI,CAAC,CAAC;IAChE;AACF;AAEO,eAAeQ,KACpBC,eAAuB,EACvBC,IAUC;IAED,MAAM,EAAEC,OAAO,EAAEC,WAAW,EAAEC,mBAAmB,EAAE,GAAGH;IACtD,MAAM,EAAEI,IAAI,EAAE,GAAGH;IACjB,MAAMI,mBAAmBJ,QAAQ,UAAU,CAAC,MAAM;IAElD,MAAM,EAAE5B,WAAW,EAAE,GAAG6B;IAGxB,MAAMI,kBAAkBN,AAAmB,SAAnBA,KAAK,SAAS;IAEtC,MAAMO,eAAe,MAAMC,2BAA2B;QACpD,aAAaR,KAAK,WAAW;QAC7B3B;QACA,aAAa2B,KAAK,WAAW;QAC7B,gBAAgB;QAChBM;IACF;IAEA,IAAIG,eAAeJ;IACnB,IAAIK,aAAaN,KAAK,KAAK;IAC3B,IAAIO,cAAcP,KAAK,MAAM;IAK7B,IAAI/B,AAAgB,iBAAhBA,aAA8B;QAChC,MAAMuC,eAAe,MAAMC,4BAA4BJ;QACvDC,aAAaE,aAAa,KAAK;QAC/BD,cAAcC,aAAa,MAAM;QACjCH,eAAeG,aAAa,WAAW;IACzC;IAEA,MAAME,gBAAgBd,KAAK,aAAa,GACpC,CAAC,yBAAyB,EAAEA,KAAK,aAAa,CAAC,4BAA4B,CAAC,GAC5E;IAEJ,MAAMe,cAA4C;QAChD;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,GAAGD,cAAc,kBAAkB,EAAEf,gBAAgB,mBAAmB,CAAC;gBACjF;aACD;QACH;KACD;IAED,IAAIiB;IAGJ,MAAMC,eAAeX,kBACjBH,oBAAoB,cAAc,KAClC;IACJ,MAAMe,kBAAkBD,eAAe,CAAC,IAAI,EAAEA,cAAc,GAAG;IAG/D,MAAME,eAAehB,oBAAoB,cAAc;IACvD,MAAMiB,kBAAkBD,eAAe,CAAC,IAAI,EAAEA,cAAc,GAAG;IAE/D,IAAIhB,oBAAoB,sBAAsB,EAAE;QAC9Ca,wBAAwB;YACtB,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,GAAGb,oBAAoB,sBAAsB,CAAC,qHAAqH,EAAEiB,kBAAkBF,iBAAiB;gBAChN;gBACA;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKT;wBACL,QAAQ;oBACV;gBACF;aACD;QACH;QAEAN,oBAAoB,mCAAmC;IACzD,OACEa,wBAAwB;QACtB,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAM,CAAC,6BAA6B,EAAEI,kBAAkBF,iBAAiB;YAC3E;YACA;gBACE,MAAM;gBACN,WAAW;oBACT,KAAKT;oBACL,QAAQ;gBACV;YACF;SACD;IACH;IAEFN,oBAAoB,MAAM,CAACa;IAG3Bb,oBAAoB,eAAe,CAAC,IAAI;IAExC,MAAMkB,aAAalB,oBAAoB,QAAQ,CAACH,KAAK,kBAAkB;IAEvE,MAAMsB,OAAqC;QACzC;YAAE,MAAM;YAAU,SAASf;QAAa;WACrCQ;WACAM;KACJ;IAED,MAAM,EACJ,SAASE,WAAW,EACpBC,KAAK,EACLC,iBAAiB,EAClB,GAAG,MAAMC,OAAOJ,MAAMpB,aAAa;QAClC,WAAWF,AAAmB,YAAnBA,KAAK,SAAS,GAAef,SAAYe,KAAK,SAAS;IACpE;IAGA,IAAI2B;IACJ,IAAI;QACFA,aAAaxD,yBAAyBoD,aAAalD;IACrD,EAAE,OAAOuD,YAAY;QAEnB,MAAMC,eACJD,sBAAsB/B,QAAQ+B,WAAW,OAAO,GAAGE,OAAOF;QAC5D,MAAM,IAAIG,qBACR,CAAC,iBAAiB,EAAEF,cAAc,EAClCN,aACAC;IAEJ;IAEA,IAAIG,WAAW,MAAM,IAAIA,AAA+B1C,WAA/B0C,WAAW,eAAe,EAAgB;QACjEK,QAAQ,IAAI,CACV;QAEFL,WAAW,eAAe,GAAG1C;QAC7B0C,WAAW,eAAe,GAAG1C;IAC/B;IAEA,MAAMgD,UAAUN,WAAW,MAAM,GAAG;QAACA,WAAW,MAAM;KAAC,GAAG,EAAE;IAC5D,IAAIO,yBAAyB;IAG7B,IAAIP,AAA+B1C,WAA/B0C,WAAW,eAAe,EAAgB;QAC5C1D,MAAM;QACNiE,yBAAyB;QAEzB,IAAI5B,iBACFH,oBAAoB,uBAAuB;IAE/C;IAEA,MAAMgC,cAAkC;QACtC,GAAGR,UAAU;QACbM;QACAV;QACAC;QACAC;QACA,UAAUW,uBAAuBH,SAASjC,KAAK,WAAW;QAC1DkC;IACF;IAEAG,OAAOV,YAAY;IAEnBM,QAAQ,OAAO,CAAC,CAACzC;QACf,MAAMC,OAAOD,OAAO,IAAI;QACxB,MAAM8C,sBAAsBtC,KAAK,WAAW,CAAC,IAAI,CAC/C,CAACR,SAAWA,OAAO,IAAI,KAAKC;QAG9BxB,MAAM,+BAA+BqE;QACrC,MAAMC,eAAeD,sBACjBE,4BAA4BF,oBAAoB,WAAW,IAC3D,EAAE;QAENrE,MAAM,gBAAgBsE;QAEtBA,aAAa,OAAO,CAAC,CAACE;YACpB,MAAMC,eAAelD,OAAO,KAAK,CAACiD,MAAM;YACxC,IAAIC,gBAAgBrE,AAAgBY,WAAhBZ,aAElBmB,OAAO,KAAK,CAACiD,MAAM,GAAGE,cACpBD,cACAhC,YACAC,aACAtC;QAGN;IACF;IAGA,IAAIiC,iBAAiB;QACnB,IAAIqB,WAAW,cAAc,EAAE,QAC7BxB,oBAAoB,WAAW,CAACwB,WAAW,cAAc;QAE3D,IAAIA,WAAW,mBAAmB,EAAE,QAClC,KAAK,MAAMiB,SAASjB,WAAW,mBAAmB,CAChDxB,oBAAoB,mBAAmB,CAACyC;QAI5C,IAAIjB,WAAW,GAAG,EAChBxB,oBAAoB,gBAAgB,CAACwB,WAAW,GAAG;IAEvD;IAGA,IAAIA,WAAW,MAAM,EACnBxB,oBAAoB,YAAY,CAACwB,WAAW,MAAM;IAGpDxB,oBAAoB,MAAM,CAAC;QACzB,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAMoB;YACR;SACD;IACH;IAEA,OAAOY;AACT"}

package/dist/es/ai-model/prompt/llm-planning.mjs

@@ -89,7 +89,7 @@ async function systemPromptToTaskPlanning({ actionSpace, modelFamily, includeBbo
   }`;
     const step1Title = shouldIncludeSubGoals ? '## Step 1: Observe and Plan (related tags: <thought>, <update-plan-content>, <mark-sub-goal-done>)' : '## Step 1: Observe (related tags: <thought>)';
     const step1Description = shouldIncludeSubGoals ? "First, observe the current screenshot and previous logs, then break down the user's instruction into multiple high-level sub-goals. Update the status of sub-goals based on what you see in the current screenshot." : 'First, observe the current screenshot and previous logs to understand the current state.';
-    const explicitInstructionRule = 'CRITICAL - Following Explicit Instructions: When the user gives you specific operation steps (not high-level goals), you MUST execute ONLY those exact steps - nothing more, nothing less. Do NOT add extra actions even if they seem logical. For example: "fill out the form" means only fill fields, do NOT submit; "click the button" means only click, do NOT wait for page load or verify results; "type \'hello\'" means only type, do NOT press Enter.';
+    const explicitInstructionRule = shouldIncludeSubGoals ? 'CRITICAL - Following Explicit Instructions: When the user gives you specific operation steps (not high-level goals), you MUST execute ONLY those exact steps - nothing more, nothing less. Do NOT add extra actions even if they seem logical. For example: "fill out the form" means only fill fields, do NOT submit; "click the button" means only click, do NOT wait for page load or verify results; "type \'hello\'" means only type, do NOT press Enter.' : 'CRITICAL - Following Instructions: You MUST execute ONLY the exact steps the user gives you - nothing more, nothing less. Do NOT add extra actions even if they seem logical. For example: "fill out the form" means only fill fields, do NOT submit; "click the button" means only click, do NOT wait for page load or verify results; "type \'hello\'" means only type, do NOT press Enter.';
     const thoughtTagDescription = shouldIncludeSubGoals ? `REQUIRED: You MUST always output the <thought> tag. Never skip it.
 
 Include your thought process in the <thought> tag. It should answer: What is the user's requirement? What is the current state based on the screenshot? Are all sub-goals completed? If not, what should be the next action? Write your thoughts naturally without numbering or section headers.
@@ -184,13 +184,16 @@ ${shouldIncludeSubGoals ? 'Based on the current screenshot and the status of all
 
 The user's instruction defines the EXACT scope of what you must accomplish. You MUST follow it precisely - nothing more, nothing less. Violating this rule may cause severe consequences such as data loss, unintended operations, or system failures.
 
-**Explicit instructions vs. High-level goals:**
+${shouldIncludeSubGoals ? `**Explicit instructions vs. High-level goals:**
 - If the user gives you **explicit operation steps** (e.g., "click X", "type Y", "fill out the form"), treat them as exact commands. Execute ONLY those steps, nothing more.
 - If the user gives you a **high-level goal** (e.g., "log in to the system", "complete the purchase"), you may determine the necessary steps to achieve it.
 
 **What "goal accomplished" means:**
 - The goal is accomplished when you have done EXACTLY what the user asked - no extra steps, no assumptions.
-- Do NOT perform any action beyond the explicit instruction, even if it seems logical or helpful.
+- Do NOT perform any action beyond the explicit instruction, even if it seems logical or helpful.` : `**Follow the instruction exactly:**
+- Treat the user's instruction as exact commands. Execute ONLY those steps, nothing more.
+- The goal is accomplished when you have done EXACTLY what the user asked - no extra steps, no assumptions.
+- Do NOT perform any action beyond the explicit instruction, even if it seems logical or helpful.`}
 
 **Examples - Explicit instructions (execute exactly, no extra steps):**
 - "fill out the form" → Goal accomplished when all fields are filled. Do NOT submit the form.
@@ -328,7 +331,10 @@ The previous action has been executed, here is the latest screenshot. Please con
 Sub-goals:
 1. Fill in the Name field with 'John' (running)
 2. Fill in the Email field with 'john@example.com' (pending)
+3. Return the filled email address (pending)
 Current sub-goal is: Fill in the Name field with 'John'
+Actions performed for current sub-goal:
+- Click on the Name field to start filling the form
 
 **Screenshot:** [Shows the form with Name field now focused/active]
 
@@ -350,7 +356,11 @@ The previous action has been executed, here is the latest screenshot. Please con
 Sub-goals:
 1. Fill in the Name field with 'John' (running)
 2. Fill in the Email field with 'john@example.com' (pending)
+3. Return the filled email address (pending)
 Current sub-goal is: Fill in the Name field with 'John'
+Actions performed for current sub-goal:
+- Click on the Name field to start filling the form
+- Typing 'John' into the Name field
 
 **Screenshot:** [Shows the form with Name field containing 'John']
 
@@ -376,7 +386,10 @@ The previous action has been executed, here is the latest screenshot. Please con
 Sub-goals:
 1. Fill in the Name field with 'John' (finished)
 2. Fill in the Email field with 'john@example.com' (running)
+3. Return the filled email address (pending)
 Current sub-goal is: Fill in the Email field with 'john@example.com'
+Actions performed for current sub-goal:
+- Moving to the Email field
 
 **Screenshot:** [Shows the form with Name='John' and Email field focused]
 
@@ -400,6 +413,9 @@ Sub-goals:
 2. Fill in the Email field with 'john@example.com' (running)
 3. Return the filled email address (pending)
 Current sub-goal is: Fill in the Email field with 'john@example.com'
+Actions performed for current sub-goal:
+- Moving to the Email field
+- Typing email address into the Email field
 
 **Screenshot:** [Shows the form with Name='John' and Email='john@example.com']
 

package/dist/es/ai-model/prompt/llm-planning.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"ai-model/prompt/llm-planning.mjs","sources":["../../../../src/ai-model/prompt/llm-planning.ts"],"sourcesContent":["import type { DeviceAction } from '@/types';\nimport type { TModelFamily } from '@midscene/shared/env';\nimport { getPreferredLanguage } from '@midscene/shared/env';\nimport {\n  getZodDescription,\n  getZodTypeName,\n} from '@midscene/shared/zod-schema-utils';\nimport type { z } from 'zod';\nimport { bboxDescription } from './common';\n\nconst vlLocateParam = (modelFamily: TModelFamily | undefined) => {\n  if (modelFamily) {\n    return `{bbox: [number, number, number, number], prompt: string } // ${bboxDescription(modelFamily)}`;\n  }\n  return '{ prompt: string /* description of the target element */ }';\n};\n\n/**\n * Find ZodDefault in the wrapper chain and return its default value\n */\nconst findDefaultValue = (field: unknown): any | undefined => {\n  let current = field;\n  const visited = new Set<unknown>();\n\n  while (current && !visited.has(current)) {\n    visited.add(current);\n    const currentWithDef = current as {\n      _def?: {\n        typeName?: string;\n        defaultValue?: () => any;\n        innerType?: unknown;\n      };\n    };\n\n    if (!currentWithDef._def?.typeName) break;\n\n    if (currentWithDef._def.typeName === 'ZodDefault') {\n      return currentWithDef._def.defaultValue?.();\n    }\n\n    // Continue unwrapping if it's a wrapper type\n    if (\n      currentWithDef._def.typeName === 'ZodOptional' ||\n      currentWithDef._def.typeName === 'ZodNullable'\n    ) {\n      current = currentWithDef._def.innerType;\n    } else {\n      break;\n    }\n  }\n\n  return undefined;\n};\n\nexport const descriptionForAction = (\n  action: DeviceAction<any>,\n  locatorSchemaTypeDescription: string,\n) => {\n  const tab = '  ';\n  const fields: string[] = [];\n\n  // Add the action type field\n  fields.push(`- type: \"${action.name}\"`);\n\n  // Handle paramSchema if it exists\n  if (action.paramSchema) {\n    const paramLines: string[] = [];\n\n    // Check if paramSchema is a ZodObject with shape\n    const schema = action.paramSchema as {\n      _def?: { typeName?: string };\n      shape?: Record<string, unknown>;\n    };\n    const isZodObject = schema._def?.typeName === 'ZodObject';\n\n    if (isZodObject && schema.shape) {\n      // Original logic for ZodObject schemas\n      const shape = schema.shape;\n\n      for (const [key, field] of Object.entries(shape)) {\n        if (field && typeof field === 'object') {\n          // Check if field is optional\n          const isOptional =\n            typeof (field as { isOptional?: () => boolean }).isOptional ===\n              'function' &&\n            (field as { isOptional: () => boolean }).isOptional();\n          const keyWithOptional = isOptional ? `${key}?` : key;\n\n          // Get the type name using extracted helper\n          const typeName = getZodTypeName(field, locatorSchemaTypeDescription);\n\n          // Get description using extracted helper\n          const description = getZodDescription(field as z.ZodTypeAny);\n\n          // Check if field has a default value by searching the wrapper chain\n          const defaultValue = findDefaultValue(field);\n          const hasDefault = defaultValue !== undefined;\n\n          // Build param line for this field\n          let paramLine = `${keyWithOptional}: ${typeName}`;\n          const comments: string[] = [];\n          if (description) {\n            comments.push(description);\n          }\n          if (hasDefault) {\n            const defaultStr =\n              typeof defaultValue === 'string'\n                ? `\"${defaultValue}\"`\n                : JSON.stringify(defaultValue);\n            comments.push(`default: ${defaultStr}`);\n          }\n          if (comments.length > 0) {\n            paramLine += ` // ${comments.join(', ')}`;\n          }\n\n          paramLines.push(paramLine);\n        }\n      }\n\n      // Add the param section to fields if there are paramLines\n      if (paramLines.length > 0) {\n        fields.push('- param:');\n        paramLines.forEach((line) => {\n          fields.push(`  - ${line}`);\n        });\n      }\n    } else {\n      // Handle non-object schemas (string, number, etc.)\n      const typeName = getZodTypeName(schema);\n      const description = getZodDescription(schema as z.ZodTypeAny);\n\n      // For simple types, indicate that param should be the direct value, not an object\n      let paramDescription = `- param: ${typeName}`;\n      if (description) {\n        paramDescription += ` // ${description}`;\n      }\n      paramDescription += ' (pass the value directly, not as an object)';\n\n      fields.push(paramDescription);\n    }\n  }\n\n  return `- ${action.name}, ${action.description || 'No description provided'}\n${tab}${fields.join(`\\n${tab}`)}\n`.trim();\n};\n\nexport async function systemPromptToTaskPlanning({\n  actionSpace,\n  modelFamily,\n  includeBbox,\n  includeThought,\n  includeSubGoals,\n}: {\n  actionSpace: DeviceAction<any>[];\n  modelFamily: TModelFamily | undefined;\n  includeBbox: boolean;\n  includeThought?: boolean;\n  includeSubGoals?: boolean;\n}) {\n  const preferredLanguage = getPreferredLanguage();\n\n  // Validate parameters: if includeBbox is true, modelFamily must be defined\n  if (includeBbox && !modelFamily) {\n    throw new Error(\n      'modelFamily cannot be undefined when includeBbox is true. A valid modelFamily is required for bbox-based location.',\n    );\n  }\n\n  const actionDescriptionList = actionSpace.map((action) => {\n    return descriptionForAction(\n      action,\n      vlLocateParam(includeBbox ? modelFamily : undefined),\n    );\n  });\n  const actionList = actionDescriptionList.join('\\n');\n\n  const shouldIncludeThought = includeThought ?? true;\n  const shouldIncludeSubGoals = includeSubGoals ?? false;\n\n  // Generate locate object examples based on includeBbox\n  const locateExample1 = includeBbox\n    ? `{\n    \"prompt\": \"Add to cart button for Sauce Labs Backpack\",\n    \"bbox\": [345, 442, 458, 483]\n  }`\n    : `{\n    \"prompt\": \"Add to cart button for Sauce Labs Backpack\"\n  }`;\n\n  // Locate examples for multi-turn conversation\n  const locateNameField = includeBbox\n    ? `{\n    \"prompt\": \"Name input field in the registration form\",\n    \"bbox\": [120, 180, 380, 210]\n  }`\n    : `{\n    \"prompt\": \"Name input field in the registration form\"\n  }`;\n\n  const locateEmailField = includeBbox\n    ? `{\n    \"prompt\": \"Email input field in the registration form\",\n    \"bbox\": [120, 240, 380, 270]\n  }`\n    : `{\n    \"prompt\": \"Email input field in the registration form\"\n  }`;\n\n  const thoughtTag = (content: string) =>\n    shouldIncludeThought ? `<thought>${content}</thought>\\n` : '';\n\n  // Sub-goals related content - only included when shouldIncludeSubGoals is true\n  const step1Title = shouldIncludeSubGoals\n    ? '## Step 1: Observe and Plan (related tags: <thought>, <update-plan-content>, <mark-sub-goal-done>)'\n    : '## Step 1: Observe (related tags: <thought>)';\n\n  const step1Description = shouldIncludeSubGoals\n    ? \"First, observe the current screenshot and previous logs, then break down the user's instruction into multiple high-level sub-goals. Update the status of sub-goals based on what you see in the current screenshot.\"\n    : 'First, observe the current screenshot and previous logs to understand the current state.';\n\n  const explicitInstructionRule = `CRITICAL - Following Explicit Instructions: When the user gives you specific operation steps (not high-level goals), you MUST execute ONLY those exact steps - nothing more, nothing less. Do NOT add extra actions even if they seem logical. For example: \"fill out the form\" means only fill fields, do NOT submit; \"click the button\" means only click, do NOT wait for page load or verify results; \"type 'hello'\" means only type, do NOT press Enter.`;\n\n  const thoughtTagDescription = shouldIncludeSubGoals\n    ? `REQUIRED: You MUST always output the <thought> tag. Never skip it.\n\nInclude your thought process in the <thought> tag. It should answer: What is the user's requirement? What is the current state based on the screenshot? Are all sub-goals completed? If not, what should be the next action? Write your thoughts naturally without numbering or section headers.\n\n${explicitInstructionRule}`\n    : `REQUIRED: You MUST always output the <thought> tag. Never skip it.\n\nInclude your thought process in the <thought> tag. It should answer: What is the current state based on the screenshot? What should be the next action? Write your thoughts naturally without numbering or section headers.\n\n${explicitInstructionRule}`;\n\n  const subGoalTags = shouldIncludeSubGoals\n    ? `\n\n* <update-plan-content> tag\n\nUse this structure to give or update your plan:\n\n<update-plan-content>\n  <sub-goal index=\"1\" status=\"finished|pending\">sub goal description</sub-goal>\n  <sub-goal index=\"2\" status=\"finished|pending\">sub goal description</sub-goal>\n  ...\n</update-plan-content>\n\n* <mark-sub-goal-done> tag\n\nUse this structure to mark a sub-goal as done:\n\n<mark-sub-goal-done>\n  <sub-goal index=\"1\" status=\"finished\" />\n</mark-sub-goal-done>\n\nIMPORTANT: You MUST only mark a sub-goal as \"finished\" AFTER you have confirmed the task is actually completed by observing the result in the screenshot. Do NOT mark a sub-goal as done just because you expect the next action will complete it. Wait until you see visual confirmation in the screenshot that the sub-goal has been achieved.\n\n* Note\n\nDuring execution, you can call <update-plan-content> at any time to update the plan based on the latest screenshot and completed sub-goals.\n\n### Example\n\nIf the user wants to \"log in to a system using username and password, complete all to-do items, and submit a registration form\", you can break it down into the following sub-goals:\n\n<thought>...</thought>\n<update-plan-content>\n  <sub-goal index=\"1\" status=\"pending\">Log in to the system</sub-goal>\n  <sub-goal index=\"2\" status=\"pending\">Complete all to-do items</sub-goal>\n  <sub-goal index=\"3\" status=\"pending\">Submit the registration form</sub-goal>\n</update-plan-content>\n\nAfter logging in and seeing the to-do items, you can mark the sub-goal as done:\n\n<mark-sub-goal-done>\n  <sub-goal index=\"1\" status=\"finished\" />\n</mark-sub-goal-done>\n\nAt this point, the status of all sub-goals is:\n\n<update-plan-content>\n  <sub-goal index=\"1\" status=\"finished\" />\n  <sub-goal index=\"2\" status=\"pending\" />\n  <sub-goal index=\"3\" status=\"pending\" />\n</update-plan-content>\n\nAfter some time, when the last sub-goal is also completed, you can mark it as done as well:\n\n<mark-sub-goal-done>\n  <sub-goal index=\"3\" status=\"finished\" />\n</mark-sub-goal-done>`\n    : '';\n\n  // Step numbering adjusts based on whether sub-goals are included\n  // When includeSubGoals=false, memory step is skipped\n  const memoryStepNumber = 2; // Only used when shouldIncludeSubGoals is true\n  const checkGoalStepNumber = shouldIncludeSubGoals ? 3 : 2;\n  const actionStepNumber = shouldIncludeSubGoals ? 4 : 3;\n\n  return `\nTarget: You are an expert to manipulate the UI to accomplish the user's instruction. User will give you an instruction, some screenshots, background knowledge and previous logs indicating what have been done. Your task is to accomplish the instruction by thinking through the path to complete the task and give the next action to execute.\n\n${step1Title}\n\n${step1Description}\n\n* <thought> tag (REQUIRED)\n\n${thoughtTagDescription}\n${subGoalTags}\n${\n  shouldIncludeSubGoals\n    ? `\n## Step ${memoryStepNumber}: Memory Data from Current Screenshot (related tags: <memory>)\n\nWhile observing the current screenshot, if you notice any information that might be needed in follow-up actions, record it here. The current screenshot will NOT be available in subsequent steps, so this memory is your only way to preserve essential information. Examples: extracted data, element states, content that needs to be referenced.\n\nDon't use this tag if no information needs to be preserved.\n`\n    : ''\n}\n## Step ${checkGoalStepNumber}: Check if Goal is Accomplished (related tags: <complete-goal>)\n\n${shouldIncludeSubGoals ? 'Based on the current screenshot and the status of all sub-goals, determine' : 'Determine'} if the entire task is completed.\n\n### CRITICAL: The User's Instruction is the Supreme Authority\n\nThe user's instruction defines the EXACT scope of what you must accomplish. You MUST follow it precisely - nothing more, nothing less. Violating this rule may cause severe consequences such as data loss, unintended operations, or system failures.\n\n**Explicit instructions vs. High-level goals:**\n- If the user gives you **explicit operation steps** (e.g., \"click X\", \"type Y\", \"fill out the form\"), treat them as exact commands. Execute ONLY those steps, nothing more.\n- If the user gives you a **high-level goal** (e.g., \"log in to the system\", \"complete the purchase\"), you may determine the necessary steps to achieve it.\n\n**What \"goal accomplished\" means:**\n- The goal is accomplished when you have done EXACTLY what the user asked - no extra steps, no assumptions.\n- Do NOT perform any action beyond the explicit instruction, even if it seems logical or helpful.\n\n**Examples - Explicit instructions (execute exactly, no extra steps):**\n- \"fill out the form\" → Goal accomplished when all fields are filled. Do NOT submit the form.\n- \"click the login button\" → Goal accomplished once clicked. Do NOT wait for page load or verify login success.\n- \"type 'hello' in the search box\" → Goal accomplished when 'hello' is typed. Do NOT press Enter or trigger search.\n- \"select the first item\" → Goal accomplished when selected. Do NOT proceed to checkout.\n\n**Special case - Assertion instructions:**\n- If the user's instruction includes an assertion (e.g., \"verify that...\", \"check that...\", \"assert...\"), and you observe from the screenshot that the assertion condition is NOT satisfied and cannot be satisfied, mark the goal as failed (success=\"false\").\n\n### Output Rules\n\n- If the task is NOT complete, skip this section and continue to Step ${actionStepNumber}.\n- Use the <complete-goal success=\"true|false\">message</complete-goal> tag to output the result if the goal is accomplished or failed.\n  - the 'success' attribute is required. ${shouldIncludeSubGoals ? 'It means whether the expected goal is accomplished based on what you observe in the current screenshot. ' : ''}No matter what actions were executed or what errors occurred during execution, if the expected goal is accomplished, set success=\"true\". If the expected goal is not accomplished and cannot be accomplished, set success=\"false\".\n  - the 'message' is the information that will be provided to the user. If the user asks for a specific format, strictly follow that.\n- If you output <complete-goal>, do NOT output <action-type> or <action-param-json>. The task ends here.\n\n## Step ${actionStepNumber}: Determine Next Action (related tags: <log>, <action-type>, <action-param-json>, <error>)\n\nONLY if the task is not complete: Think what the next action is according to the current screenshot${shouldIncludeSubGoals ? ' and the plan' : ''}.\n\n- Don't give extra actions or plans beyond the instruction or the plan. For example, don't try to submit the form if the instruction is only to fill something.\n- Consider the current screenshot and give the action that is most likely to accomplish the instruction. For example, if the next step is to click a button but it's not visible in the screenshot, you should try to find it first instead of give a click action.\n- Make sure the previous actions are completed successfully. Otherwise, retry or do something else to recover.\n- Give just the next ONE action you should do (if any)\n- If there are some error messages reported by the previous actions, don't give up, try parse a new action to recover. If the error persists for more than 3 times, you should think this is an error and set the \"error\" field to the error message.\n\n### Supporting actions list\n\n${actionList}\n\n### Log to give user feedback (preamble message)\n\nThe <log> tag is a brief preamble message to the user explaining what you're about to do. It should follow these principles and examples:\n\n- **Use ${preferredLanguage}**\n- **Keep it concise**: be no more than 1-2 sentences, focused on immediate, tangible next steps. (8–12 words or Chinese characters for quick updates).\n- **Build on prior context**: if this is not the first action to be done, use the preamble message to connect the dots with what's been done so far and create a sense of momentum and clarity for the user to understand your next actions.\n- **Keep your tone light, friendly and curious**: add small touches of personality in preambles feel collaborative and engaging.\n\n**Examples:**\n- <log>Click the login button</log>\n- <log>Scroll to find the 'Yes' button in popup</log>\n- <log>Previous actions failed to find the 'Yes' button, i will try again</log>\n- <log>Go back to find the login button</log>\n\n### If there is some action to do ...\n\n- Use the <action-type> and <action-param-json> tags to output the action to be executed.\n- The <action-type> MUST be one of the supporting actions. 'complete-goal' is NOT a valid action-type.\nFor example:\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateExample1}\n}\n</action-param-json>\n\n### If you think there is an error ...\n\n- Use the <error> tag to output the error message.\n\nFor example:\n<error>Unable to find the required element on the page</error>\n\n### If there is no action to do ...\n\n- Don't output <action-type> or <action-param-json> if there is no action to do.\n\n## Return Format\n\nReturn in XML format following this decision flow:\n\n**Always include (REQUIRED):**\n<!-- Step 1: Observe${shouldIncludeSubGoals ? ' and Plan' : ''} -->\n<thought>Your thought process here. NEVER skip this tag.</thought>\n${\n  shouldIncludeSubGoals\n    ? `\n<!-- required when no update-plan-content is provided in the previous response -->\n<update-plan-content>...</update-plan-content>\n\n<!-- required when any sub-goal is completed -->\n<mark-sub-goal-done>\n  <sub-goal index=\"1\" status=\"finished\" />\n</mark-sub-goal-done>\n`\n    : ''\n}${\n  shouldIncludeSubGoals\n    ? `\n<!-- Step ${memoryStepNumber}: Memory data from current screenshot if needed -->\n<memory>...</memory>\n`\n    : ''\n}\n**Then choose ONE of the following paths:**\n\n**Path A: If the goal is accomplished or failed (Step ${checkGoalStepNumber})**\n<complete-goal success=\"true|false\">...</complete-goal>\n\n**Path B: If the goal is NOT complete yet (Step ${actionStepNumber})**\n<!-- Determine next action -->\n<log>...</log>\n<action-type>...</action-type>\n<action-param-json>...</action-param-json>\n\n<!-- OR if there's an error -->\n<error>...</error>\n${\n  shouldIncludeSubGoals\n    ? `\n## Multi-turn Conversation Example\n\nBelow is an example of a multi-turn conversation for \"fill out the registration form with name 'John' and email 'john@example.com', then return the filled email address\":\n\n### Turn 1 - Initial instruction\n\n**User message:**\n<user_instruction>fill out the registration form with name 'John' and email 'john@example.com', then return the filled email address</user_instruction>\n\n**Screenshot:** [Shows a registration form with empty Name and Email fields]\n\n**Your response:**\n<thought>The user wants me to fill out the registration form with specific values and return the email address. I can see the form has two fields: Name and Email. Both are currently empty. I'll break this down into sub-goals and start with the Name field. Note: The instruction is to fill the form only (not submit), and return the email at the end.</thought>\n<update-plan-content>\n  <sub-goal index=\"1\" status=\"pending\">Fill in the Name field with 'John'</sub-goal>\n  <sub-goal index=\"2\" status=\"pending\">Fill in the Email field with 'john@example.com'</sub-goal>\n  <sub-goal index=\"3\" status=\"pending\">Return the filled email address</sub-goal>\n</update-plan-content>\n<log>Click on the Name field to start filling the form</log>\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateNameField}\n}\n</action-param-json>\n\n### Turn 2 - After clicking Name field\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\nSub-goals:\n1. Fill in the Name field with 'John' (running)\n2. Fill in the Email field with 'john@example.com' (pending)\nCurrent sub-goal is: Fill in the Name field with 'John'\n\n**Screenshot:** [Shows the form with Name field now focused/active]\n\n**Your response:**\n<thought>The Name field is now focused. I need to type 'John' into this field. Current sub-goal is running, will be completed after input.</thought>\n<log>Typing 'John' into the Name field</log>\n<action-type>Input</action-type>\n<action-param-json>\n{\n  \"value\": \"John\"\n}\n</action-param-json>\n\n### Turn 3 - After entering name\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\nSub-goals:\n1. Fill in the Name field with 'John' (running)\n2. Fill in the Email field with 'john@example.com' (pending)\nCurrent sub-goal is: Fill in the Name field with 'John'\n\n**Screenshot:** [Shows the form with Name field containing 'John']\n\n**Your response:**\n<thought>The Name field now shows 'John'. Sub-goal 1 is complete. Next, I need to fill the Email field for sub-goal 2.</thought>\n<mark-sub-goal-done>\n  <sub-goal index=\"1\" status=\"finished\" />\n</mark-sub-goal-done>\n<memory>Name field has been filled with 'John'</memory>\n<log>Moving to the Email field</log>\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateEmailField}\n}\n</action-param-json>\n\n### Turn 4 - After clicking Email field\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\nSub-goals:\n1. Fill in the Name field with 'John' (finished)\n2. Fill in the Email field with 'john@example.com' (running)\nCurrent sub-goal is: Fill in the Email field with 'john@example.com'\n\n**Screenshot:** [Shows the form with Name='John' and Email field focused]\n\n**Your response:**\n<thought>The Email field is now focused. I'll enter 'john@example.com'. After this, sub-goal 2 will be complete and my task will be done.</thought>\n<log>Typing email address into the Email field</log>\n<action-type>Input</action-type>\n<action-param-json>\n{\n  \"value\": \"john@example.com\"\n}\n</action-param-json>\n\n### Turn 5 - After entering email (Goal accomplished)\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\nSub-goals:\n1. Fill in the Name field with 'John' (finished)\n2. Fill in the Email field with 'john@example.com' (running)\n3. Return the filled email address (pending)\nCurrent sub-goal is: Fill in the Email field with 'john@example.com'\n\n**Screenshot:** [Shows the form with Name='John' and Email='john@example.com']\n\n**Your response:**\n<thought>Both fields are now filled: Name shows 'John' and Email shows 'john@example.com'. Sub-goal 2 is complete. The user asked me to return the filled email address, so I need to include 'john@example.com' in my response. All sub-goals are now finished.</thought>\n<mark-sub-goal-done>\n  <sub-goal index=\"2\" status=\"finished\" />\n  <sub-goal index=\"3\" status=\"finished\" />\n</mark-sub-goal-done>\n<complete-goal success=\"true\">john@example.com</complete-goal>\n`\n    : `\n## Multi-turn Conversation Example\n\nBelow is an example of a multi-turn conversation for \"fill out the registration form with name 'John' and email 'john@example.com', then return the filled email address\":\n\n### Turn 1 - Initial instruction\n\n**User message:**\n<user_instruction>fill out the registration form with name 'John' and email 'john@example.com', then return the filled email address</user_instruction>\n\n**Screenshot:** [Shows a registration form with empty Name and Email fields]\n\n**Your response:**\n<thought>The user wants me to fill out the registration form with specific values and return the email address. I can see the form has two fields: Name and Email. Both are currently empty. I should start by clicking on the Name field. Note: The instruction is to fill the form only (not submit), and return the email at the end.</thought>\n<log>Click on the Name field to start filling the form</log>\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateNameField}\n}\n</action-param-json>\n\n### Turn 2 - After clicking Name field\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\n**Screenshot:** [Shows the form with Name field now focused/active]\n\n**Your response:**\n<thought>The Name field is now focused. I need to type 'John' into this field. After this, I'll need to fill the Email field as well.</thought>\n<log>Typing 'John' into the Name field</log>\n<action-type>Input</action-type>\n<action-param-json>\n{\n  \"value\": \"John\"\n}\n</action-param-json>\n\n### Turn 3 - After entering name\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\n**Screenshot:** [Shows the form with Name field containing 'John']\n\n**Your response:**\n<thought>The Name field now shows 'John'. Next, I need to fill the Email field with 'john@example.com'. I'll click on the Email field first.</thought>\n<log>Moving to the Email field</log>\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateEmailField}\n}\n</action-param-json>\n\n### Turn 4 - After clicking Email field\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\n**Screenshot:** [Shows the form with Name='John' and Email field focused]\n\n**Your response:**\n<thought>The Email field is now focused. I'll enter 'john@example.com'. After this, both required fields will be filled and my task will be complete.</thought>\n<log>Typing email address into the Email field</log>\n<action-type>Input</action-type>\n<action-param-json>\n{\n  \"value\": \"john@example.com\"\n}\n</action-param-json>\n\n### Turn 5 - After entering email (Goal accomplished)\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\n**Screenshot:** [Shows the form with Name='John' and Email='john@example.com']\n\n**Your response:**\n<thought>Both fields are now filled: Name shows 'John' and Email shows 'john@example.com'. The user asked me to return the filled email address, so I should include 'john@example.com' in my response. Goal accomplished.</thought>\n<complete-goal success=\"true\">john@example.com</complete-goal>\n`\n}`;\n}\n"],"names":["vlLocateParam","modelFamily","bboxDescription","findDefaultValue","field","current","visited","Set","currentWithDef","descriptionForAction","action","locatorSchemaTypeDescription","tab","fields","paramLines","schema","isZodObject","shape","key","Object","isOptional","keyWithOptional","typeName","getZodTypeName","description","getZodDescription","defaultValue","hasDefault","undefined","paramLine","comments","defaultStr","JSON","line","paramDescription","systemPromptToTaskPlanning","actionSpace","includeBbox","includeThought","includeSubGoals","preferredLanguage","getPreferredLanguage","Error","actionDescriptionList","actionList","shouldIncludeSubGoals","locateExample1","locateNameField","locateEmailField","step1Title","step1Description","explicitInstructionRule","thoughtTagDescription","subGoalTags","memoryStepNumber","checkGoalStepNumber","actionStepNumber"],"mappings":";;;AAUA,MAAMA,gBAAgB,CAACC;IACrB,IAAIA,aACF,OAAO,CAAC,6DAA6D,EAAEC,gBAAgBD,cAAc;IAEvG,OAAO;AACT;AAKA,MAAME,mBAAmB,CAACC;IACxB,IAAIC,UAAUD;IACd,MAAME,UAAU,IAAIC;IAEpB,MAAOF,WAAW,CAACC,QAAQ,GAAG,CAACD,SAAU;QACvCC,QAAQ,GAAG,CAACD;QACZ,MAAMG,iBAAiBH;QAQvB,IAAI,CAACG,eAAe,IAAI,EAAE,UAAU;QAEpC,IAAIA,AAAiC,iBAAjCA,eAAe,IAAI,CAAC,QAAQ,EAC9B,OAAOA,eAAe,IAAI,CAAC,YAAY;QAIzC,IACEA,AAAiC,kBAAjCA,eAAe,IAAI,CAAC,QAAQ,IAC5BA,AAAiC,kBAAjCA,eAAe,IAAI,CAAC,QAAQ,EAE5BH,UAAUG,eAAe,IAAI,CAAC,SAAS;aAEvC;IAEJ;AAGF;AAEO,MAAMC,uBAAuB,CAClCC,QACAC;IAEA,MAAMC,MAAM;IACZ,MAAMC,SAAmB,EAAE;IAG3BA,OAAO,IAAI,CAAC,CAAC,SAAS,EAAEH,OAAO,IAAI,CAAC,CAAC,CAAC;IAGtC,IAAIA,OAAO,WAAW,EAAE;QACtB,MAAMI,aAAuB,EAAE;QAG/B,MAAMC,SAASL,OAAO,WAAW;QAIjC,MAAMM,cAAcD,OAAO,IAAI,EAAE,aAAa;QAE9C,IAAIC,eAAeD,OAAO,KAAK,EAAE;YAE/B,MAAME,QAAQF,OAAO,KAAK;YAE1B,KAAK,MAAM,CAACG,KAAKd,MAAM,IAAIe,OAAO,OAAO,CAACF,OACxC,IAAIb,SAAS,AAAiB,YAAjB,OAAOA,OAAoB;gBAEtC,MAAMgB,aACJ,AACE,cADF,OAAQhB,MAAyC,UAAU,IAE1DA,MAAwC,UAAU;gBACrD,MAAMiB,kBAAkBD,aAAa,GAAGF,IAAI,CAAC,CAAC,GAAGA;gBAGjD,MAAMI,WAAWC,eAAenB,OAAOO;gBAGvC,MAAMa,cAAcC,kBAAkBrB;gBAGtC,MAAMsB,eAAevB,iBAAiBC;gBACtC,MAAMuB,aAAaD,AAAiBE,WAAjBF;gBAGnB,IAAIG,YAAY,GAAGR,gBAAgB,EAAE,EAAEC,UAAU;gBACjD,MAAMQ,WAAqB,EAAE;gBAC7B,IAAIN,aACFM,SAAS,IAAI,CAACN;gBAEhB,IAAIG,YAAY;oBACd,MAAMI,aACJ,AAAwB,YAAxB,OAAOL,eACH,CAAC,CAAC,EAAEA,aAAa,CAAC,CAAC,GACnBM,KAAK,SAAS,CAACN;oBACrBI,SAAS,IAAI,CAAC,CAAC,SAAS,EAAEC,YAAY;gBACxC;gBACA,IAAID,SAAS,MAAM,GAAG,GACpBD,aAAa,CAAC,IAAI,EAAEC,SAAS,IAAI,CAAC,OAAO;gBAG3ChB,WAAW,IAAI,CAACe;YAClB;YAIF,IAAIf,WAAW,MAAM,GAAG,GAAG;gBACzBD,OAAO,IAAI,CAAC;gBACZC,WAAW,OAAO,CAAC,CAACmB;oBAClBpB,OAAO,IAAI,CAAC,CAAC,IAAI,EAAEoB,MAAM;gBAC3B;YACF;QACF,OAAO;YAEL,MAAMX,WAAWC,eAAeR;YAChC,MAAMS,cAAcC,kBAAkBV;YAGtC,IAAImB,mBAAmB,CAAC,SAAS,EAAEZ,UAAU;YAC7C,IAAIE,aACFU,oBAAoB,CAAC,IAAI,EAAEV,aAAa;YAE1CU,oBAAoB;YAEpBrB,OAAO,IAAI,CAACqB;QACd;IACF;IAEA,OAAO,CAAC,EAAE,EAAExB,OAAO,IAAI,CAAC,EAAE,EAAEA,OAAO,WAAW,IAAI,0BAA0B;AAC9E,EAAEE,MAAMC,OAAO,IAAI,CAAC,CAAC,EAAE,EAAED,KAAK,EAAE;AAChC,CAAC,CAAC,IAAI;AACN;AAEO,eAAeuB,2BAA2B,EAC/CC,WAAW,EACXnC,WAAW,EACXoC,WAAW,EACXC,cAAc,EACdC,eAAe,EAOhB;IACC,MAAMC,oBAAoBC;IAG1B,IAAIJ,eAAe,CAACpC,aAClB,MAAM,IAAIyC,MACR;IAIJ,MAAMC,wBAAwBP,YAAY,GAAG,CAAC,CAAC1B,SACtCD,qBACLC,QACAV,cAAcqC,cAAcpC,cAAc2B;IAG9C,MAAMgB,aAAaD,sBAAsB,IAAI,CAAC;IAG9C,MAAME,wBAAwBN,mBAAmB;IAGjD,MAAMO,iBAAiBT,cACnB,CAAC;;;GAGJ,CAAC,GACE,CAAC;;GAEJ,CAAC;IAGF,MAAMU,kBAAkBV,cACpB,CAAC;;;GAGJ,CAAC,GACE,CAAC;;GAEJ,CAAC;IAEF,MAAMW,mBAAmBX,cACrB,CAAC;;;GAGJ,CAAC,GACE,CAAC;;GAEJ,CAAC;IAMF,MAAMY,aAAaJ,wBACf,uGACA;IAEJ,MAAMK,mBAAmBL,wBACrB,wNACA;IAEJ,MAAMM,0BAA0B;IAEhC,MAAMC,wBAAwBP,wBAC1B,CAAC;;;;AAIP,EAAEM,yBAAyB,GACrB,CAAC;;;;AAIP,EAAEA,yBAAyB;IAEzB,MAAME,cAAcR,wBAChB,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBAuDc,CAAC,GAChB;IAIJ,MAAMS,mBAAmB;IACzB,MAAMC,sBAAsBV,wBAAwB,IAAI;IACxD,MAAMW,mBAAmBX,wBAAwB,IAAI;IAErD,OAAO,CAAC;;;AAGV,EAAEI,WAAW;;AAEb,EAAEC,iBAAiB;;;;AAInB,EAAEE,sBAAsB;AACxB,EAAEC,YAAY;AACd,EACER,wBACI,CAAC;QACC,EAAES,iBAAiB;;;;;AAK3B,CAAC,GACK,GACL;QACO,EAAEC,oBAAoB;;AAE9B,EAAEV,wBAAwB,+EAA+E,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;sEAyB/C,EAAEW,iBAAiB;;yCAEhD,EAAEX,wBAAwB,6GAA6G,GAAG;;;;QAI3K,EAAEW,iBAAiB;;mGAEwE,EAAEX,wBAAwB,kBAAkB,GAAG;;;;;;;;;;AAUlJ,EAAED,WAAW;;;;;;QAML,EAAEJ,kBAAkB;;;;;;;;;;;;;;;;;;;YAmBhB,EAAEM,eAAe;;;;;;;;;;;;;;;;;;;;oBAoBT,EAAED,wBAAwB,cAAc,GAAG;;AAE/D,EACEA,wBACI,CAAC;;;;;;;;AAQP,CAAC,GACK,KAEJA,wBACI,CAAC;UACG,EAAES,iBAAiB;;AAE7B,CAAC,GACK,GACL;;;sDAGqD,EAAEC,oBAAoB;;;gDAG5B,EAAEC,iBAAiB;;;;;;;;AAQnE,EACEX,wBACI,CAAC;;;;;;;;;;;;;;;;;;;;;;;YAuBK,EAAEE,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;YAgDlB,EAAEC,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8C/B,CAAC,GACK,CAAC;;;;;;;;;;;;;;;;;;YAkBK,EAAED,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;YAkClB,EAAEC,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+B/B,CAAC,EACC;AACF"}
+{"version":3,"file":"ai-model/prompt/llm-planning.mjs","sources":["../../../../src/ai-model/prompt/llm-planning.ts"],"sourcesContent":["import type { DeviceAction } from '@/types';\nimport type { TModelFamily } from '@midscene/shared/env';\nimport { getPreferredLanguage } from '@midscene/shared/env';\nimport {\n  getZodDescription,\n  getZodTypeName,\n} from '@midscene/shared/zod-schema-utils';\nimport type { z } from 'zod';\nimport { bboxDescription } from './common';\n\nconst vlLocateParam = (modelFamily: TModelFamily | undefined) => {\n  if (modelFamily) {\n    return `{bbox: [number, number, number, number], prompt: string } // ${bboxDescription(modelFamily)}`;\n  }\n  return '{ prompt: string /* description of the target element */ }';\n};\n\n/**\n * Find ZodDefault in the wrapper chain and return its default value\n */\nconst findDefaultValue = (field: unknown): any | undefined => {\n  let current = field;\n  const visited = new Set<unknown>();\n\n  while (current && !visited.has(current)) {\n    visited.add(current);\n    const currentWithDef = current as {\n      _def?: {\n        typeName?: string;\n        defaultValue?: () => any;\n        innerType?: unknown;\n      };\n    };\n\n    if (!currentWithDef._def?.typeName) break;\n\n    if (currentWithDef._def.typeName === 'ZodDefault') {\n      return currentWithDef._def.defaultValue?.();\n    }\n\n    // Continue unwrapping if it's a wrapper type\n    if (\n      currentWithDef._def.typeName === 'ZodOptional' ||\n      currentWithDef._def.typeName === 'ZodNullable'\n    ) {\n      current = currentWithDef._def.innerType;\n    } else {\n      break;\n    }\n  }\n\n  return undefined;\n};\n\nexport const descriptionForAction = (\n  action: DeviceAction<any>,\n  locatorSchemaTypeDescription: string,\n) => {\n  const tab = '  ';\n  const fields: string[] = [];\n\n  // Add the action type field\n  fields.push(`- type: \"${action.name}\"`);\n\n  // Handle paramSchema if it exists\n  if (action.paramSchema) {\n    const paramLines: string[] = [];\n\n    // Check if paramSchema is a ZodObject with shape\n    const schema = action.paramSchema as {\n      _def?: { typeName?: string };\n      shape?: Record<string, unknown>;\n    };\n    const isZodObject = schema._def?.typeName === 'ZodObject';\n\n    if (isZodObject && schema.shape) {\n      // Original logic for ZodObject schemas\n      const shape = schema.shape;\n\n      for (const [key, field] of Object.entries(shape)) {\n        if (field && typeof field === 'object') {\n          // Check if field is optional\n          const isOptional =\n            typeof (field as { isOptional?: () => boolean }).isOptional ===\n              'function' &&\n            (field as { isOptional: () => boolean }).isOptional();\n          const keyWithOptional = isOptional ? `${key}?` : key;\n\n          // Get the type name using extracted helper\n          const typeName = getZodTypeName(field, locatorSchemaTypeDescription);\n\n          // Get description using extracted helper\n          const description = getZodDescription(field as z.ZodTypeAny);\n\n          // Check if field has a default value by searching the wrapper chain\n          const defaultValue = findDefaultValue(field);\n          const hasDefault = defaultValue !== undefined;\n\n          // Build param line for this field\n          let paramLine = `${keyWithOptional}: ${typeName}`;\n          const comments: string[] = [];\n          if (description) {\n            comments.push(description);\n          }\n          if (hasDefault) {\n            const defaultStr =\n              typeof defaultValue === 'string'\n                ? `\"${defaultValue}\"`\n                : JSON.stringify(defaultValue);\n            comments.push(`default: ${defaultStr}`);\n          }\n          if (comments.length > 0) {\n            paramLine += ` // ${comments.join(', ')}`;\n          }\n\n          paramLines.push(paramLine);\n        }\n      }\n\n      // Add the param section to fields if there are paramLines\n      if (paramLines.length > 0) {\n        fields.push('- param:');\n        paramLines.forEach((line) => {\n          fields.push(`  - ${line}`);\n        });\n      }\n    } else {\n      // Handle non-object schemas (string, number, etc.)\n      const typeName = getZodTypeName(schema);\n      const description = getZodDescription(schema as z.ZodTypeAny);\n\n      // For simple types, indicate that param should be the direct value, not an object\n      let paramDescription = `- param: ${typeName}`;\n      if (description) {\n        paramDescription += ` // ${description}`;\n      }\n      paramDescription += ' (pass the value directly, not as an object)';\n\n      fields.push(paramDescription);\n    }\n  }\n\n  return `- ${action.name}, ${action.description || 'No description provided'}\n${tab}${fields.join(`\\n${tab}`)}\n`.trim();\n};\n\nexport async function systemPromptToTaskPlanning({\n  actionSpace,\n  modelFamily,\n  includeBbox,\n  includeThought,\n  includeSubGoals,\n}: {\n  actionSpace: DeviceAction<any>[];\n  modelFamily: TModelFamily | undefined;\n  includeBbox: boolean;\n  includeThought?: boolean;\n  includeSubGoals?: boolean;\n}) {\n  const preferredLanguage = getPreferredLanguage();\n\n  // Validate parameters: if includeBbox is true, modelFamily must be defined\n  if (includeBbox && !modelFamily) {\n    throw new Error(\n      'modelFamily cannot be undefined when includeBbox is true. A valid modelFamily is required for bbox-based location.',\n    );\n  }\n\n  const actionDescriptionList = actionSpace.map((action) => {\n    return descriptionForAction(\n      action,\n      vlLocateParam(includeBbox ? modelFamily : undefined),\n    );\n  });\n  const actionList = actionDescriptionList.join('\\n');\n\n  const shouldIncludeThought = includeThought ?? true;\n  const shouldIncludeSubGoals = includeSubGoals ?? false;\n\n  // Generate locate object examples based on includeBbox\n  const locateExample1 = includeBbox\n    ? `{\n    \"prompt\": \"Add to cart button for Sauce Labs Backpack\",\n    \"bbox\": [345, 442, 458, 483]\n  }`\n    : `{\n    \"prompt\": \"Add to cart button for Sauce Labs Backpack\"\n  }`;\n\n  // Locate examples for multi-turn conversation\n  const locateNameField = includeBbox\n    ? `{\n    \"prompt\": \"Name input field in the registration form\",\n    \"bbox\": [120, 180, 380, 210]\n  }`\n    : `{\n    \"prompt\": \"Name input field in the registration form\"\n  }`;\n\n  const locateEmailField = includeBbox\n    ? `{\n    \"prompt\": \"Email input field in the registration form\",\n    \"bbox\": [120, 240, 380, 270]\n  }`\n    : `{\n    \"prompt\": \"Email input field in the registration form\"\n  }`;\n\n  const thoughtTag = (content: string) =>\n    shouldIncludeThought ? `<thought>${content}</thought>\\n` : '';\n\n  // Sub-goals related content - only included when shouldIncludeSubGoals is true\n  const step1Title = shouldIncludeSubGoals\n    ? '## Step 1: Observe and Plan (related tags: <thought>, <update-plan-content>, <mark-sub-goal-done>)'\n    : '## Step 1: Observe (related tags: <thought>)';\n\n  const step1Description = shouldIncludeSubGoals\n    ? \"First, observe the current screenshot and previous logs, then break down the user's instruction into multiple high-level sub-goals. Update the status of sub-goals based on what you see in the current screenshot.\"\n    : 'First, observe the current screenshot and previous logs to understand the current state.';\n\n  const explicitInstructionRule = shouldIncludeSubGoals\n    ? `CRITICAL - Following Explicit Instructions: When the user gives you specific operation steps (not high-level goals), you MUST execute ONLY those exact steps - nothing more, nothing less. Do NOT add extra actions even if they seem logical. For example: \"fill out the form\" means only fill fields, do NOT submit; \"click the button\" means only click, do NOT wait for page load or verify results; \"type 'hello'\" means only type, do NOT press Enter.`\n    : `CRITICAL - Following Instructions: You MUST execute ONLY the exact steps the user gives you - nothing more, nothing less. Do NOT add extra actions even if they seem logical. For example: \"fill out the form\" means only fill fields, do NOT submit; \"click the button\" means only click, do NOT wait for page load or verify results; \"type 'hello'\" means only type, do NOT press Enter.`;\n\n  const thoughtTagDescription = shouldIncludeSubGoals\n    ? `REQUIRED: You MUST always output the <thought> tag. Never skip it.\n\nInclude your thought process in the <thought> tag. It should answer: What is the user's requirement? What is the current state based on the screenshot? Are all sub-goals completed? If not, what should be the next action? Write your thoughts naturally without numbering or section headers.\n\n${explicitInstructionRule}`\n    : `REQUIRED: You MUST always output the <thought> tag. Never skip it.\n\nInclude your thought process in the <thought> tag. It should answer: What is the current state based on the screenshot? What should be the next action? Write your thoughts naturally without numbering or section headers.\n\n${explicitInstructionRule}`;\n\n  const subGoalTags = shouldIncludeSubGoals\n    ? `\n\n* <update-plan-content> tag\n\nUse this structure to give or update your plan:\n\n<update-plan-content>\n  <sub-goal index=\"1\" status=\"finished|pending\">sub goal description</sub-goal>\n  <sub-goal index=\"2\" status=\"finished|pending\">sub goal description</sub-goal>\n  ...\n</update-plan-content>\n\n* <mark-sub-goal-done> tag\n\nUse this structure to mark a sub-goal as done:\n\n<mark-sub-goal-done>\n  <sub-goal index=\"1\" status=\"finished\" />\n</mark-sub-goal-done>\n\nIMPORTANT: You MUST only mark a sub-goal as \"finished\" AFTER you have confirmed the task is actually completed by observing the result in the screenshot. Do NOT mark a sub-goal as done just because you expect the next action will complete it. Wait until you see visual confirmation in the screenshot that the sub-goal has been achieved.\n\n* Note\n\nDuring execution, you can call <update-plan-content> at any time to update the plan based on the latest screenshot and completed sub-goals.\n\n### Example\n\nIf the user wants to \"log in to a system using username and password, complete all to-do items, and submit a registration form\", you can break it down into the following sub-goals:\n\n<thought>...</thought>\n<update-plan-content>\n  <sub-goal index=\"1\" status=\"pending\">Log in to the system</sub-goal>\n  <sub-goal index=\"2\" status=\"pending\">Complete all to-do items</sub-goal>\n  <sub-goal index=\"3\" status=\"pending\">Submit the registration form</sub-goal>\n</update-plan-content>\n\nAfter logging in and seeing the to-do items, you can mark the sub-goal as done:\n\n<mark-sub-goal-done>\n  <sub-goal index=\"1\" status=\"finished\" />\n</mark-sub-goal-done>\n\nAt this point, the status of all sub-goals is:\n\n<update-plan-content>\n  <sub-goal index=\"1\" status=\"finished\" />\n  <sub-goal index=\"2\" status=\"pending\" />\n  <sub-goal index=\"3\" status=\"pending\" />\n</update-plan-content>\n\nAfter some time, when the last sub-goal is also completed, you can mark it as done as well:\n\n<mark-sub-goal-done>\n  <sub-goal index=\"3\" status=\"finished\" />\n</mark-sub-goal-done>`\n    : '';\n\n  // Step numbering adjusts based on whether sub-goals are included\n  // When includeSubGoals=false, memory step is skipped\n  const memoryStepNumber = 2; // Only used when shouldIncludeSubGoals is true\n  const checkGoalStepNumber = shouldIncludeSubGoals ? 3 : 2;\n  const actionStepNumber = shouldIncludeSubGoals ? 4 : 3;\n\n  return `\nTarget: You are an expert to manipulate the UI to accomplish the user's instruction. User will give you an instruction, some screenshots, background knowledge and previous logs indicating what have been done. Your task is to accomplish the instruction by thinking through the path to complete the task and give the next action to execute.\n\n${step1Title}\n\n${step1Description}\n\n* <thought> tag (REQUIRED)\n\n${thoughtTagDescription}\n${subGoalTags}\n${\n  shouldIncludeSubGoals\n    ? `\n## Step ${memoryStepNumber}: Memory Data from Current Screenshot (related tags: <memory>)\n\nWhile observing the current screenshot, if you notice any information that might be needed in follow-up actions, record it here. The current screenshot will NOT be available in subsequent steps, so this memory is your only way to preserve essential information. Examples: extracted data, element states, content that needs to be referenced.\n\nDon't use this tag if no information needs to be preserved.\n`\n    : ''\n}\n## Step ${checkGoalStepNumber}: Check if Goal is Accomplished (related tags: <complete-goal>)\n\n${shouldIncludeSubGoals ? 'Based on the current screenshot and the status of all sub-goals, determine' : 'Determine'} if the entire task is completed.\n\n### CRITICAL: The User's Instruction is the Supreme Authority\n\nThe user's instruction defines the EXACT scope of what you must accomplish. You MUST follow it precisely - nothing more, nothing less. Violating this rule may cause severe consequences such as data loss, unintended operations, or system failures.\n\n${shouldIncludeSubGoals ? `**Explicit instructions vs. High-level goals:**\n- If the user gives you **explicit operation steps** (e.g., \"click X\", \"type Y\", \"fill out the form\"), treat them as exact commands. Execute ONLY those steps, nothing more.\n- If the user gives you a **high-level goal** (e.g., \"log in to the system\", \"complete the purchase\"), you may determine the necessary steps to achieve it.\n\n**What \"goal accomplished\" means:**\n- The goal is accomplished when you have done EXACTLY what the user asked - no extra steps, no assumptions.\n- Do NOT perform any action beyond the explicit instruction, even if it seems logical or helpful.` : `**Follow the instruction exactly:**\n- Treat the user's instruction as exact commands. Execute ONLY those steps, nothing more.\n- The goal is accomplished when you have done EXACTLY what the user asked - no extra steps, no assumptions.\n- Do NOT perform any action beyond the explicit instruction, even if it seems logical or helpful.`}\n\n**Examples - Explicit instructions (execute exactly, no extra steps):**\n- \"fill out the form\" → Goal accomplished when all fields are filled. Do NOT submit the form.\n- \"click the login button\" → Goal accomplished once clicked. Do NOT wait for page load or verify login success.\n- \"type 'hello' in the search box\" → Goal accomplished when 'hello' is typed. Do NOT press Enter or trigger search.\n- \"select the first item\" → Goal accomplished when selected. Do NOT proceed to checkout.\n\n**Special case - Assertion instructions:**\n- If the user's instruction includes an assertion (e.g., \"verify that...\", \"check that...\", \"assert...\"), and you observe from the screenshot that the assertion condition is NOT satisfied and cannot be satisfied, mark the goal as failed (success=\"false\").\n\n### Output Rules\n\n- If the task is NOT complete, skip this section and continue to Step ${actionStepNumber}.\n- Use the <complete-goal success=\"true|false\">message</complete-goal> tag to output the result if the goal is accomplished or failed.\n  - the 'success' attribute is required. ${shouldIncludeSubGoals ? 'It means whether the expected goal is accomplished based on what you observe in the current screenshot. ' : ''}No matter what actions were executed or what errors occurred during execution, if the expected goal is accomplished, set success=\"true\". If the expected goal is not accomplished and cannot be accomplished, set success=\"false\".\n  - the 'message' is the information that will be provided to the user. If the user asks for a specific format, strictly follow that.\n- If you output <complete-goal>, do NOT output <action-type> or <action-param-json>. The task ends here.\n\n## Step ${actionStepNumber}: Determine Next Action (related tags: <log>, <action-type>, <action-param-json>, <error>)\n\nONLY if the task is not complete: Think what the next action is according to the current screenshot${shouldIncludeSubGoals ? ' and the plan' : ''}.\n\n- Don't give extra actions or plans beyond the instruction or the plan. For example, don't try to submit the form if the instruction is only to fill something.\n- Consider the current screenshot and give the action that is most likely to accomplish the instruction. For example, if the next step is to click a button but it's not visible in the screenshot, you should try to find it first instead of give a click action.\n- Make sure the previous actions are completed successfully. Otherwise, retry or do something else to recover.\n- Give just the next ONE action you should do (if any)\n- If there are some error messages reported by the previous actions, don't give up, try parse a new action to recover. If the error persists for more than 3 times, you should think this is an error and set the \"error\" field to the error message.\n\n### Supporting actions list\n\n${actionList}\n\n### Log to give user feedback (preamble message)\n\nThe <log> tag is a brief preamble message to the user explaining what you're about to do. It should follow these principles and examples:\n\n- **Use ${preferredLanguage}**\n- **Keep it concise**: be no more than 1-2 sentences, focused on immediate, tangible next steps. (8–12 words or Chinese characters for quick updates).\n- **Build on prior context**: if this is not the first action to be done, use the preamble message to connect the dots with what's been done so far and create a sense of momentum and clarity for the user to understand your next actions.\n- **Keep your tone light, friendly and curious**: add small touches of personality in preambles feel collaborative and engaging.\n\n**Examples:**\n- <log>Click the login button</log>\n- <log>Scroll to find the 'Yes' button in popup</log>\n- <log>Previous actions failed to find the 'Yes' button, i will try again</log>\n- <log>Go back to find the login button</log>\n\n### If there is some action to do ...\n\n- Use the <action-type> and <action-param-json> tags to output the action to be executed.\n- The <action-type> MUST be one of the supporting actions. 'complete-goal' is NOT a valid action-type.\nFor example:\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateExample1}\n}\n</action-param-json>\n\n### If you think there is an error ...\n\n- Use the <error> tag to output the error message.\n\nFor example:\n<error>Unable to find the required element on the page</error>\n\n### If there is no action to do ...\n\n- Don't output <action-type> or <action-param-json> if there is no action to do.\n\n## Return Format\n\nReturn in XML format following this decision flow:\n\n**Always include (REQUIRED):**\n<!-- Step 1: Observe${shouldIncludeSubGoals ? ' and Plan' : ''} -->\n<thought>Your thought process here. NEVER skip this tag.</thought>\n${\n  shouldIncludeSubGoals\n    ? `\n<!-- required when no update-plan-content is provided in the previous response -->\n<update-plan-content>...</update-plan-content>\n\n<!-- required when any sub-goal is completed -->\n<mark-sub-goal-done>\n  <sub-goal index=\"1\" status=\"finished\" />\n</mark-sub-goal-done>\n`\n    : ''\n}${\n  shouldIncludeSubGoals\n    ? `\n<!-- Step ${memoryStepNumber}: Memory data from current screenshot if needed -->\n<memory>...</memory>\n`\n    : ''\n}\n**Then choose ONE of the following paths:**\n\n**Path A: If the goal is accomplished or failed (Step ${checkGoalStepNumber})**\n<complete-goal success=\"true|false\">...</complete-goal>\n\n**Path B: If the goal is NOT complete yet (Step ${actionStepNumber})**\n<!-- Determine next action -->\n<log>...</log>\n<action-type>...</action-type>\n<action-param-json>...</action-param-json>\n\n<!-- OR if there's an error -->\n<error>...</error>\n${\n  shouldIncludeSubGoals\n    ? `\n## Multi-turn Conversation Example\n\nBelow is an example of a multi-turn conversation for \"fill out the registration form with name 'John' and email 'john@example.com', then return the filled email address\":\n\n### Turn 1 - Initial instruction\n\n**User message:**\n<user_instruction>fill out the registration form with name 'John' and email 'john@example.com', then return the filled email address</user_instruction>\n\n**Screenshot:** [Shows a registration form with empty Name and Email fields]\n\n**Your response:**\n<thought>The user wants me to fill out the registration form with specific values and return the email address. I can see the form has two fields: Name and Email. Both are currently empty. I'll break this down into sub-goals and start with the Name field. Note: The instruction is to fill the form only (not submit), and return the email at the end.</thought>\n<update-plan-content>\n  <sub-goal index=\"1\" status=\"pending\">Fill in the Name field with 'John'</sub-goal>\n  <sub-goal index=\"2\" status=\"pending\">Fill in the Email field with 'john@example.com'</sub-goal>\n  <sub-goal index=\"3\" status=\"pending\">Return the filled email address</sub-goal>\n</update-plan-content>\n<log>Click on the Name field to start filling the form</log>\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateNameField}\n}\n</action-param-json>\n\n### Turn 2 - After clicking Name field\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\nSub-goals:\n1. Fill in the Name field with 'John' (running)\n2. Fill in the Email field with 'john@example.com' (pending)\n3. Return the filled email address (pending)\nCurrent sub-goal is: Fill in the Name field with 'John'\nActions performed for current sub-goal:\n- Click on the Name field to start filling the form\n\n**Screenshot:** [Shows the form with Name field now focused/active]\n\n**Your response:**\n<thought>The Name field is now focused. I need to type 'John' into this field. Current sub-goal is running, will be completed after input.</thought>\n<log>Typing 'John' into the Name field</log>\n<action-type>Input</action-type>\n<action-param-json>\n{\n  \"value\": \"John\"\n}\n</action-param-json>\n\n### Turn 3 - After entering name\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\nSub-goals:\n1. Fill in the Name field with 'John' (running)\n2. Fill in the Email field with 'john@example.com' (pending)\n3. Return the filled email address (pending)\nCurrent sub-goal is: Fill in the Name field with 'John'\nActions performed for current sub-goal:\n- Click on the Name field to start filling the form\n- Typing 'John' into the Name field\n\n**Screenshot:** [Shows the form with Name field containing 'John']\n\n**Your response:**\n<thought>The Name field now shows 'John'. Sub-goal 1 is complete. Next, I need to fill the Email field for sub-goal 2.</thought>\n<mark-sub-goal-done>\n  <sub-goal index=\"1\" status=\"finished\" />\n</mark-sub-goal-done>\n<memory>Name field has been filled with 'John'</memory>\n<log>Moving to the Email field</log>\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateEmailField}\n}\n</action-param-json>\n\n### Turn 4 - After clicking Email field\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\nSub-goals:\n1. Fill in the Name field with 'John' (finished)\n2. Fill in the Email field with 'john@example.com' (running)\n3. Return the filled email address (pending)\nCurrent sub-goal is: Fill in the Email field with 'john@example.com'\nActions performed for current sub-goal:\n- Moving to the Email field\n\n**Screenshot:** [Shows the form with Name='John' and Email field focused]\n\n**Your response:**\n<thought>The Email field is now focused. I'll enter 'john@example.com'. After this, sub-goal 2 will be complete and my task will be done.</thought>\n<log>Typing email address into the Email field</log>\n<action-type>Input</action-type>\n<action-param-json>\n{\n  \"value\": \"john@example.com\"\n}\n</action-param-json>\n\n### Turn 5 - After entering email (Goal accomplished)\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\nSub-goals:\n1. Fill in the Name field with 'John' (finished)\n2. Fill in the Email field with 'john@example.com' (running)\n3. Return the filled email address (pending)\nCurrent sub-goal is: Fill in the Email field with 'john@example.com'\nActions performed for current sub-goal:\n- Moving to the Email field\n- Typing email address into the Email field\n\n**Screenshot:** [Shows the form with Name='John' and Email='john@example.com']\n\n**Your response:**\n<thought>Both fields are now filled: Name shows 'John' and Email shows 'john@example.com'. Sub-goal 2 is complete. The user asked me to return the filled email address, so I need to include 'john@example.com' in my response. All sub-goals are now finished.</thought>\n<mark-sub-goal-done>\n  <sub-goal index=\"2\" status=\"finished\" />\n  <sub-goal index=\"3\" status=\"finished\" />\n</mark-sub-goal-done>\n<complete-goal success=\"true\">john@example.com</complete-goal>\n`\n    : `\n## Multi-turn Conversation Example\n\nBelow is an example of a multi-turn conversation for \"fill out the registration form with name 'John' and email 'john@example.com', then return the filled email address\":\n\n### Turn 1 - Initial instruction\n\n**User message:**\n<user_instruction>fill out the registration form with name 'John' and email 'john@example.com', then return the filled email address</user_instruction>\n\n**Screenshot:** [Shows a registration form with empty Name and Email fields]\n\n**Your response:**\n<thought>The user wants me to fill out the registration form with specific values and return the email address. I can see the form has two fields: Name and Email. Both are currently empty. I should start by clicking on the Name field. Note: The instruction is to fill the form only (not submit), and return the email at the end.</thought>\n<log>Click on the Name field to start filling the form</log>\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateNameField}\n}\n</action-param-json>\n\n### Turn 2 - After clicking Name field\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\n**Screenshot:** [Shows the form with Name field now focused/active]\n\n**Your response:**\n<thought>The Name field is now focused. I need to type 'John' into this field. After this, I'll need to fill the Email field as well.</thought>\n<log>Typing 'John' into the Name field</log>\n<action-type>Input</action-type>\n<action-param-json>\n{\n  \"value\": \"John\"\n}\n</action-param-json>\n\n### Turn 3 - After entering name\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\n**Screenshot:** [Shows the form with Name field containing 'John']\n\n**Your response:**\n<thought>The Name field now shows 'John'. Next, I need to fill the Email field with 'john@example.com'. I'll click on the Email field first.</thought>\n<log>Moving to the Email field</log>\n<action-type>Tap</action-type>\n<action-param-json>\n{\n  \"locate\": ${locateEmailField}\n}\n</action-param-json>\n\n### Turn 4 - After clicking Email field\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\n**Screenshot:** [Shows the form with Name='John' and Email field focused]\n\n**Your response:**\n<thought>The Email field is now focused. I'll enter 'john@example.com'. After this, both required fields will be filled and my task will be complete.</thought>\n<log>Typing email address into the Email field</log>\n<action-type>Input</action-type>\n<action-param-json>\n{\n  \"value\": \"john@example.com\"\n}\n</action-param-json>\n\n### Turn 5 - After entering email (Goal accomplished)\n\n**User message:**\nThe previous action has been executed, here is the latest screenshot. Please continue according to the instruction.\n\n**Screenshot:** [Shows the form with Name='John' and Email='john@example.com']\n\n**Your response:**\n<thought>Both fields are now filled: Name shows 'John' and Email shows 'john@example.com'. The user asked me to return the filled email address, so I should include 'john@example.com' in my response. Goal accomplished.</thought>\n<complete-goal success=\"true\">john@example.com</complete-goal>\n`\n}`;\n}\n"],"names":["vlLocateParam","modelFamily","bboxDescription","findDefaultValue","field","current","visited","Set","currentWithDef","descriptionForAction","action","locatorSchemaTypeDescription","tab","fields","paramLines","schema","isZodObject","shape","key","Object","isOptional","keyWithOptional","typeName","getZodTypeName","description","getZodDescription","defaultValue","hasDefault","undefined","paramLine","comments","defaultStr","JSON","line","paramDescription","systemPromptToTaskPlanning","actionSpace","includeBbox","includeThought","includeSubGoals","preferredLanguage","getPreferredLanguage","Error","actionDescriptionList","actionList","shouldIncludeSubGoals","locateExample1","locateNameField","locateEmailField","step1Title","step1Description","explicitInstructionRule","thoughtTagDescription","subGoalTags","memoryStepNumber","checkGoalStepNumber","actionStepNumber"],"mappings":";;;AAUA,MAAMA,gBAAgB,CAACC;IACrB,IAAIA,aACF,OAAO,CAAC,6DAA6D,EAAEC,gBAAgBD,cAAc;IAEvG,OAAO;AACT;AAKA,MAAME,mBAAmB,CAACC;IACxB,IAAIC,UAAUD;IACd,MAAME,UAAU,IAAIC;IAEpB,MAAOF,WAAW,CAACC,QAAQ,GAAG,CAACD,SAAU;QACvCC,QAAQ,GAAG,CAACD;QACZ,MAAMG,iBAAiBH;QAQvB,IAAI,CAACG,eAAe,IAAI,EAAE,UAAU;QAEpC,IAAIA,AAAiC,iBAAjCA,eAAe,IAAI,CAAC,QAAQ,EAC9B,OAAOA,eAAe,IAAI,CAAC,YAAY;QAIzC,IACEA,AAAiC,kBAAjCA,eAAe,IAAI,CAAC,QAAQ,IAC5BA,AAAiC,kBAAjCA,eAAe,IAAI,CAAC,QAAQ,EAE5BH,UAAUG,eAAe,IAAI,CAAC,SAAS;aAEvC;IAEJ;AAGF;AAEO,MAAMC,uBAAuB,CAClCC,QACAC;IAEA,MAAMC,MAAM;IACZ,MAAMC,SAAmB,EAAE;IAG3BA,OAAO,IAAI,CAAC,CAAC,SAAS,EAAEH,OAAO,IAAI,CAAC,CAAC,CAAC;IAGtC,IAAIA,OAAO,WAAW,EAAE;QACtB,MAAMI,aAAuB,EAAE;QAG/B,MAAMC,SAASL,OAAO,WAAW;QAIjC,MAAMM,cAAcD,OAAO,IAAI,EAAE,aAAa;QAE9C,IAAIC,eAAeD,OAAO,KAAK,EAAE;YAE/B,MAAME,QAAQF,OAAO,KAAK;YAE1B,KAAK,MAAM,CAACG,KAAKd,MAAM,IAAIe,OAAO,OAAO,CAACF,OACxC,IAAIb,SAAS,AAAiB,YAAjB,OAAOA,OAAoB;gBAEtC,MAAMgB,aACJ,AACE,cADF,OAAQhB,MAAyC,UAAU,IAE1DA,MAAwC,UAAU;gBACrD,MAAMiB,kBAAkBD,aAAa,GAAGF,IAAI,CAAC,CAAC,GAAGA;gBAGjD,MAAMI,WAAWC,eAAenB,OAAOO;gBAGvC,MAAMa,cAAcC,kBAAkBrB;gBAGtC,MAAMsB,eAAevB,iBAAiBC;gBACtC,MAAMuB,aAAaD,AAAiBE,WAAjBF;gBAGnB,IAAIG,YAAY,GAAGR,gBAAgB,EAAE,EAAEC,UAAU;gBACjD,MAAMQ,WAAqB,EAAE;gBAC7B,IAAIN,aACFM,SAAS,IAAI,CAACN;gBAEhB,IAAIG,YAAY;oBACd,MAAMI,aACJ,AAAwB,YAAxB,OAAOL,eACH,CAAC,CAAC,EAAEA,aAAa,CAAC,CAAC,GACnBM,KAAK,SAAS,CAACN;oBACrBI,SAAS,IAAI,CAAC,CAAC,SAAS,EAAEC,YAAY;gBACxC;gBACA,IAAID,SAAS,MAAM,GAAG,GACpBD,aAAa,CAAC,IAAI,EAAEC,SAAS,IAAI,CAAC,OAAO;gBAG3ChB,WAAW,IAAI,CAACe;YAClB;YAIF,IAAIf,WAAW,MAAM,GAAG,GAAG;gBACzBD,OAAO,IAAI,CAAC;gBACZC,WAAW,OAAO,CAAC,CAACmB;oBAClBpB,OAAO,IAAI,CAAC,CAAC,IAAI,EAAEoB,MAAM;gBAC3B;YACF;QACF,OAAO;YAEL,MAAMX,WAAWC,eAAeR;YAChC,MAAMS,cAAcC,kBAAkBV;YAGtC,IAAImB,mBAAmB,CAAC,SAAS,EAAEZ,UAAU;YAC7C,IAAIE,aACFU,oBAAoB,CAAC,IAAI,EAAEV,aAAa;YAE1CU,oBAAoB;YAEpBrB,OAAO,IAAI,CAACqB;QACd;IACF;IAEA,OAAO,CAAC,EAAE,EAAExB,OAAO,IAAI,CAAC,EAAE,EAAEA,OAAO,WAAW,IAAI,0BAA0B;AAC9E,EAAEE,MAAMC,OAAO,IAAI,CAAC,CAAC,EAAE,EAAED,KAAK,EAAE;AAChC,CAAC,CAAC,IAAI;AACN;AAEO,eAAeuB,2BAA2B,EAC/CC,WAAW,EACXnC,WAAW,EACXoC,WAAW,EACXC,cAAc,EACdC,eAAe,EAOhB;IACC,MAAMC,oBAAoBC;IAG1B,IAAIJ,eAAe,CAACpC,aAClB,MAAM,IAAIyC,MACR;IAIJ,MAAMC,wBAAwBP,YAAY,GAAG,CAAC,CAAC1B,SACtCD,qBACLC,QACAV,cAAcqC,cAAcpC,cAAc2B;IAG9C,MAAMgB,aAAaD,sBAAsB,IAAI,CAAC;IAG9C,MAAME,wBAAwBN,mBAAmB;IAGjD,MAAMO,iBAAiBT,cACnB,CAAC;;;GAGJ,CAAC,GACE,CAAC;;GAEJ,CAAC;IAGF,MAAMU,kBAAkBV,cACpB,CAAC;;;GAGJ,CAAC,GACE,CAAC;;GAEJ,CAAC;IAEF,MAAMW,mBAAmBX,cACrB,CAAC;;;GAGJ,CAAC,GACE,CAAC;;GAEJ,CAAC;IAMF,MAAMY,aAAaJ,wBACf,uGACA;IAEJ,MAAMK,mBAAmBL,wBACrB,wNACA;IAEJ,MAAMM,0BAA0BN,wBAC5B,mcACA;IAEJ,MAAMO,wBAAwBP,wBAC1B,CAAC;;;;AAIP,EAAEM,yBAAyB,GACrB,CAAC;;;;AAIP,EAAEA,yBAAyB;IAEzB,MAAME,cAAcR,wBAChB,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBAuDc,CAAC,GAChB;IAIJ,MAAMS,mBAAmB;IACzB,MAAMC,sBAAsBV,wBAAwB,IAAI;IACxD,MAAMW,mBAAmBX,wBAAwB,IAAI;IAErD,OAAO,CAAC;;;AAGV,EAAEI,WAAW;;AAEb,EAAEC,iBAAiB;;;;AAInB,EAAEE,sBAAsB;AACxB,EAAEC,YAAY;AACd,EACER,wBACI,CAAC;QACC,EAAES,iBAAiB;;;;;AAK3B,CAAC,GACK,GACL;QACO,EAAEC,oBAAoB;;AAE9B,EAAEV,wBAAwB,+EAA+E,YAAY;;;;;;AAMrH,EAAEA,wBAAwB,CAAC;;;;;;iGAMsE,CAAC,GAAG,CAAC;;;iGAGL,CAAC,CAAC;;;;;;;;;;;;;sEAa7B,EAAEW,iBAAiB;;yCAEhD,EAAEX,wBAAwB,6GAA6G,GAAG;;;;QAI3K,EAAEW,iBAAiB;;mGAEwE,EAAEX,wBAAwB,kBAAkB,GAAG;;;;;;;;;;AAUlJ,EAAED,WAAW;;;;;;QAML,EAAEJ,kBAAkB;;;;;;;;;;;;;;;;;;;YAmBhB,EAAEM,eAAe;;;;;;;;;;;;;;;;;;;;oBAoBT,EAAED,wBAAwB,cAAc,GAAG;;AAE/D,EACEA,wBACI,CAAC;;;;;;;;AAQP,CAAC,GACK,KAEJA,wBACI,CAAC;UACG,EAAES,iBAAiB;;AAE7B,CAAC,GACK,GACL;;;sDAGqD,EAAEC,oBAAoB;;;gDAG5B,EAAEC,iBAAiB;;;;;;;;AAQnE,EACEX,wBACI,CAAC;;;;;;;;;;;;;;;;;;;;;;;YAuBK,EAAEE,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;YAuDlB,EAAEC,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoD/B,CAAC,GACK,CAAC;;;;;;;;;;;;;;;;;;YAkBK,EAAED,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;YAkClB,EAAEC,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+B/B,CAAC,EACC;AACF"}

package/dist/es/ai-model/service-caller/index.mjs

@@ -242,10 +242,14 @@ async function callAI(messages, modelConfig, options) {
                 debugProfileDetail(`model usage detail: ${JSON.stringify(result.usage)}`);
                 if (!result.choices) throw new Error(`invalid response from LLM service: ${JSON.stringify(result)}`);
                 content = result.choices[0].message.content;
-                if (!content) throw new Error('empty content from AI model');
                 accumulatedReasoning = result.choices[0].message?.reasoning_content || '';
                 usage = result.usage;
                 requestId = result._request_id;
+                if (!content && accumulatedReasoning && 'doubao-vision' === modelFamily) {
+                    console.warn('empty content from AI model, using reasoning content');
+                    content = accumulatedReasoning;
+                }
+                if (!content) throw new Error('empty content from AI model');
                 break;
             } catch (error) {
                 lastError = error;
@@ -273,7 +277,7 @@ async function callAI(messages, modelConfig, options) {
             isStreamed: !!isStreaming
         };
     } catch (e) {
-        console.error(' call AI error', e);
+        console.error('call AI error', e);
         const newError = new Error(`failed to call ${isStreaming ? 'streaming ' : ''}AI model service (${modelName}): ${e.message}\nTrouble shooting: https://midscenejs.com/model-provider.html`, {
             cause: e
         });