@midscene/core 1.0.1-beta-20251120071640.0 → 1.0.1-beta-20251124014653.0

package/dist/lib/ai-model/inspect.js.map

@@ -1 +1 @@
-{"version":3,"file":"ai-model/inspect.js","sources":["webpack://@midscene/core/webpack/runtime/define_property_getters","webpack://@midscene/core/webpack/runtime/has_own_property","webpack://@midscene/core/webpack/runtime/make_namespace_object","webpack://@midscene/core/./src/ai-model/inspect.ts"],"sourcesContent":["__webpack_require__.d = (exports, definition) => {\n\tfor(var key in definition) {\n        if(__webpack_require__.o(definition, key) && !__webpack_require__.o(exports, key)) {\n            Object.defineProperty(exports, key, { enumerable: true, get: definition[key] });\n        }\n    }\n};","__webpack_require__.o = (obj, prop) => (Object.prototype.hasOwnProperty.call(obj, prop))","// define __esModule on exports\n__webpack_require__.r = (exports) => {\n\tif(typeof Symbol !== 'undefined' && Symbol.toStringTag) {\n\t\tObject.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });\n\t}\n\tObject.defineProperty(exports, '__esModule', { value: true });\n};","import type {\n  AIDataExtractionResponse,\n  AIElementResponse,\n  AISectionLocatorResponse,\n  AIUsageInfo,\n  Rect,\n  ReferenceImage,\n  ServiceExtractOption,\n  UIContext,\n} from '@/types';\nimport type { IModelConfig } from '@midscene/shared/env';\nimport { generateElementByPosition } from '@midscene/shared/extractor/dom-util';\nimport {\n  cropByRect,\n  paddingToMatchBlockByBase64,\n  preProcessImageUrl,\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 {\n  AIActionType,\n  adaptBboxToRect,\n  expandSearchArea,\n  mergeRects,\n} from './common';\nimport {\n  extractDataQueryPrompt,\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 { callAIWithObjectResponse } 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: `reference image ${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  referenceImage?: ReferenceImage;\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}> {\n  const { context, targetElementDescription, callAIFn, modelConfig } = options;\n  const { vlMode } = modelConfig;\n  const { screenshotBase64 } = context;\n\n  assert(\n    targetElementDescription,\n    'cannot find the target element description',\n  );\n  const targetElementDescriptionText = extraTextFromUserPrompt(\n    targetElementDescription,\n  );\n  const userInstructionPrompt = await findElementPrompt.format({\n    targetElementDescription: targetElementDescriptionText,\n  });\n  const systemPrompt = systemPromptToLocateElement(vlMode);\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 (vlMode === '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: 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  const res = await callAIFn(msgs, AIActionType.INSPECT_ELEMENT, modelConfig);\n\n  const rawResponse = JSON.stringify(res.content);\n\n  let resRect: Rect | undefined;\n  let matchedElements = 'elements' in res.content ? res.content.elements : [];\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        vlMode,\n      );\n\n      debugInspect('resRect', resRect);\n\n      const rectCenter = {\n        x: resRect.left + resRect.width / 2,\n        y: resRect.top + resRect.height / 2,\n      };\n\n      const element: LocateResultElement = generateElementByPosition(\n        rectCenter,\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  };\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  error?: string;\n  rawResponse: string;\n  usage?: AIUsageInfo;\n}> {\n  const { context, sectionDescription, modelConfig } = options;\n  const { vlMode } = modelConfig;\n  const { screenshotBase64 } = context;\n\n  const systemPrompt = systemPromptToLocateSection(vlMode);\n  const sectionLocatorInstructionText = await sectionLocatorInstruction.format({\n    sectionDescription: 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  const result = await callAIWithObjectResponse<AISectionLocatorResponse>(\n    msgs,\n    AIActionType.EXTRACT_DATA,\n    modelConfig,\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      vlMode,\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          vlMode,\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    // expand search area to at least 200 x 200\n    sectionRect = expandSearchArea(mergedRect, context.size, vlMode);\n    debugSection('expanded sectionRect %j', sectionRect);\n  }\n\n  let imageBase64 = screenshotBase64;\n  if (sectionRect) {\n    const croppedResult = await cropByRect(\n      screenshotBase64,\n      sectionRect,\n      vlMode === 'qwen2.5-vl',\n    );\n    imageBase64 = croppedResult.imageBase64;\n    sectionRect.width = croppedResult.width;\n    sectionRect.height = croppedResult.height;\n  }\n\n  return {\n    rect: sectionRect,\n    imageBase64,\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;\n\n  const extractDataPromptText = await 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 result = await callAIWithObjectResponse<AIDataExtractionResponse<T>>(\n    msgs,\n    AIActionType.EXTRACT_DATA,\n    modelConfig,\n  );\n  return {\n    parseResult: result.content,\n    usage: result.usage,\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(\n    msgs,\n    AIActionType.INSPECT_ELEMENT, // Reuse existing action type for now\n    modelConfig,\n  );\n\n  return {\n    isOrderSensitive: result.content.isOrderSensitive ?? false,\n    usage: result.usage,\n  };\n}\n"],"names":["__webpack_require__","definition","key","Object","obj","prop","Symbol","debugInspect","getDebug","debugSection","extraTextFromUserPrompt","prompt","promptsToChatParam","multimodalPrompt","_multimodalPrompt_images","msgs","item","base64","preProcessImageUrl","AiLocateElement","options","context","targetElementDescription","callAIFn","modelConfig","vlMode","screenshotBase64","assert","targetElementDescriptionText","userInstructionPrompt","findElementPrompt","systemPrompt","systemPromptToLocateElement","imagePayload","imageWidth","imageHeight","originalImageWidth","originalImageHeight","_options_searchConfig_rect","_options_searchConfig_rect1","paddedResult","paddingToMatchBlockByBase64","addOns","res","AIActionType","rawResponse","JSON","resRect","matchedElements","errors","Array","_options_searchConfig_rect2","_options_searchConfig_rect3","adaptBboxToRect","rectCenter","element","generateElementByPosition","e","msg","Error","AiLocateSection","sectionDescription","systemPromptToLocateSection","sectionLocatorInstructionText","sectionLocatorInstruction","result","callAIWithObjectResponse","sectionRect","sectionBbox","targetRect","referenceBboxList","referenceRects","bbox","mergedRect","mergeRects","expandSearchArea","imageBase64","croppedResult","cropByRect","AiExtractElementInfo","dataQuery","extractOption","systemPromptToExtract","extractDataPromptText","extractDataQueryPrompt","userContent","AiJudgeOrderSensitive","description","systemPromptToJudgeOrderSensitive","userPrompt","orderSensitiveJudgePrompt"],"mappings":";;;IAAAA,oBAAoB,CAAC,GAAG,CAAC,UAASC;QACjC,IAAI,IAAIC,OAAOD,WACR,IAAGD,oBAAoB,CAAC,CAACC,YAAYC,QAAQ,CAACF,oBAAoB,CAAC,CAAC,UAASE,MACzEC,OAAO,cAAc,CAAC,UAASD,KAAK;YAAE,YAAY;YAAM,KAAKD,UAAU,CAACC,IAAI;QAAC;IAGzF;;;ICNAF,oBAAoB,CAAC,GAAG,CAACI,KAAKC,OAAUF,OAAO,SAAS,CAAC,cAAc,CAAC,IAAI,CAACC,KAAKC;;;ICClFL,oBAAoB,CAAC,GAAG,CAAC;QACxB,IAAG,AAAkB,eAAlB,OAAOM,UAA0BA,OAAO,WAAW,EACrDH,OAAO,cAAc,CAAC,UAASG,OAAO,WAAW,EAAE;YAAE,OAAO;QAAS;QAEtEH,OAAO,cAAc,CAAC,UAAS,cAAc;YAAE,OAAO;QAAK;IAC5D;;;;;;;;;;;;;;;;;;;;ACgDA,MAAMI,eAAeC,AAAAA,IAAAA,uBAAAA,QAAAA,AAAAA,EAAS;AAC9B,MAAMC,eAAeD,AAAAA,IAAAA,uBAAAA,QAAAA,AAAAA,EAAS;AAE9B,MAAME,0BAA0B,CAACC;IAC/B,IAAI,AAAkB,YAAlB,OAAOA,QACT,OAAOA;IAEP,OAAOA,OAAO,MAAM;AAExB;AAEA,MAAMC,qBAAqB,OACzBC;QAGIC;IADJ,MAAMC,OAAyC,EAAE;IACjD,IAAID,QAAAA,mBAAAA,KAAAA,IAAAA,QAAAA,CAAAA,2BAAAA,iBAAkB,MAAM,AAAD,IAAvBA,KAAAA,IAAAA,yBAA0B,MAAM,EAAE;QACpCC,KAAK,IAAI,CAAC;YACR,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM;gBACR;aACD;QACH;QAEA,KAAK,MAAMC,QAAQH,iBAAiB,MAAM,CAAE;YAC1C,MAAMI,SAAS,MAAMC,AAAAA,IAAAA,oBAAAA,kBAAAA,AAAAA,EACnBF,KAAK,GAAG,EACR,CAAC,CAACH,iBAAiB,uBAAuB;YAG5CE,KAAK,IAAI,CAAC;gBACR,MAAM;gBACN,SAAS;oBACP;wBACE,MAAM;wBACN,MAAM,CAAC,gBAAgB,EAAEC,KAAK,IAAI,CAAC,CAAC,CAAC;oBACvC;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,OASrC;IASC,MAAM,EAAEC,OAAO,EAAEC,wBAAwB,EAAEC,QAAQ,EAAEC,WAAW,EAAE,GAAGJ;IACrE,MAAM,EAAEK,MAAM,EAAE,GAAGD;IACnB,MAAM,EAAEE,gBAAgB,EAAE,GAAGL;IAE7BM,IAAAA,sBAAAA,MAAAA,AAAAA,EACEL,0BACA;IAEF,MAAMM,+BAA+BlB,wBACnCY;IAEF,MAAMO,wBAAwB,MAAMC,+BAAAA,iBAAAA,CAAAA,MAAwB,CAAC;QAC3D,0BAA0BF;IAC5B;IACA,MAAMG,eAAeC,AAAAA,IAAAA,+BAAAA,2BAAAA,AAAAA,EAA4BP;IAEjD,IAAIQ,eAAeP;IACnB,IAAIQ,aAAab,QAAQ,IAAI,CAAC,KAAK;IACnC,IAAIc,cAAcd,QAAQ,IAAI,CAAC,MAAM;IACrC,IAAIe,qBAAqBF;IACzB,IAAIG,sBAAsBF;IAE1B,IAAIf,QAAQ,YAAY,EAAE;YAWXkB,4BACCC;QAXdZ,IAAAA,sBAAAA,MAAAA,AAAAA,EACEP,QAAQ,YAAY,CAAC,IAAI,EACzB;QAEFO,IAAAA,sBAAAA,MAAAA,AAAAA,EACEP,QAAQ,YAAY,CAAC,WAAW,EAChC;QAGFa,eAAeb,QAAQ,YAAY,CAAC,WAAW;QAC/Cc,aAAa,QAAAI,CAAAA,6BAAAA,QAAQ,YAAY,CAAC,IAAI,AAAD,IAAxBA,KAAAA,IAAAA,2BAA2B,KAAK;QAC7CH,cAAc,QAAAI,CAAAA,8BAAAA,QAAQ,YAAY,CAAC,IAAI,AAAD,IAAxBA,KAAAA,IAAAA,4BAA2B,MAAM;QAC/CH,qBAAqBF;QACrBG,sBAAsBF;IACxB,OAAO,IAAIV,AAAW,iBAAXA,QAAyB;QAClC,MAAMe,eAAe,MAAMC,AAAAA,IAAAA,oBAAAA,2BAAAA,AAAAA,EAA4BR;QACvDC,aAAaM,aAAa,KAAK;QAC/BL,cAAcK,aAAa,MAAM;QACjCP,eAAeO,aAAa,WAAW;IACzC;IAEA,MAAMzB,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKE;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAMJ;gBACR;aACD;QACH;KACD;IAED,IAAI,AAAoC,YAApC,OAAOP,0BAAuC;QAChD,MAAMoB,SAAS,MAAM9B,mBAAmB;YACtC,QAAQU,yBAAyB,MAAM;YACvC,yBAAyBA,yBAAyB,uBAAuB;QAC3E;QACAP,KAAK,IAAI,IAAI2B;IACf;IAEA,MAAMC,MAAM,MAAMpB,SAASR,MAAM6B,mCAAAA,YAAAA,CAAAA,eAA4B,EAAEpB;IAE/D,MAAMqB,cAAcC,KAAK,SAAS,CAACH,IAAI,OAAO;IAE9C,IAAII;IACJ,IAAIC,kBAAkB,cAAcL,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,QAAQ,GAAG,EAAE;IAC3E,IAAIM,SACF,YAAYN,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,MAAM,GAAG,EAAE;IACnD,IAAI;QACF,IACE,UAAUA,IAAI,OAAO,IACrBO,MAAM,OAAO,CAACP,IAAI,OAAO,CAAC,IAAI,KAC9BA,IAAI,OAAO,CAAC,IAAI,CAAC,MAAM,IAAI,GAC3B;gBAKEQ,6BAAAA,uBACAC,6BAAAA;YALFL,UAAUM,AAAAA,IAAAA,mCAAAA,eAAAA,AAAAA,EACRV,IAAI,OAAO,CAAC,IAAI,EAChBT,YACAC,aAAAA,QACAgB,CAAAA,wBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,sBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,IAAI,UAChCC,CAAAA,yBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,uBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,GAAG,EAC/BhB,oBACAC,qBACAZ;YAGFlB,aAAa,WAAWwC;YAExB,MAAMO,aAAa;gBACjB,GAAGP,QAAQ,IAAI,GAAGA,QAAQ,KAAK,GAAG;gBAClC,GAAGA,QAAQ,GAAG,GAAGA,QAAQ,MAAM,GAAG;YACpC;YAEA,MAAMQ,UAA+BC,AAAAA,IAAAA,yBAAAA,yBAAAA,AAAAA,EACnCF,YACA1B;YAEFqB,SAAS,EAAE;YAEX,IAAIM,SACFP,kBAAkB;gBAACO;aAAQ;QAE/B;IACF,EAAE,OAAOE,GAAG;QACV,MAAMC,MACJD,aAAaE,QACT,CAAC,sBAAsB,EAAEF,EAAE,OAAO,EAAE,GACpC;QACN,IAAI,AAACR,UAAUA,AAAAA,CAAAA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,MAAM,AAAD,MAAM,GAGhCA,OAAO,IAAI,CAAC,CAAC,CAAC,EAAES,IAAI,CAAC,CAAC;aAFtBT,SAAS;YAACS;SAAI;IAIlB;IAEA,OAAO;QACL,MAAMX;QACN,aAAa;YACX,UAAUC;YACV,QAAQC;QACV;QACAJ;QACA,OAAOF,IAAI,KAAK;IAClB;AACF;AAEO,eAAeiB,gBAAgBxC,OAIrC;IAOC,MAAM,EAAEC,OAAO,EAAEwC,kBAAkB,EAAErC,WAAW,EAAE,GAAGJ;IACrD,MAAM,EAAEK,MAAM,EAAE,GAAGD;IACnB,MAAM,EAAEE,gBAAgB,EAAE,GAAGL;IAE7B,MAAMU,eAAe+B,AAAAA,IAAAA,uCAAAA,2BAAAA,AAAAA,EAA4BrC;IACjD,MAAMsC,gCAAgC,MAAMC,uCAAAA,yBAAAA,CAAAA,MAAgC,CAAC;QAC3E,oBAAoBtD,wBAAwBmD;IAC9C;IACA,MAAM9C,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,MAAMqC;gBACR;aACD;QACH;KACD;IAED,IAAI,AAA8B,YAA9B,OAAOF,oBAAiC;QAC1C,MAAMnB,SAAS,MAAM9B,mBAAmB;YACtC,QAAQiD,mBAAmB,MAAM;YACjC,yBAAyBA,mBAAmB,uBAAuB;QACrE;QACA9C,KAAK,IAAI,IAAI2B;IACf;IAEA,MAAMuB,SAAS,MAAMC,AAAAA,IAAAA,yBAAAA,wBAAAA,AAAAA,EACnBnD,MACA6B,mCAAAA,YAAAA,CAAAA,YAAyB,EACzBpB;IAGF,IAAI2C;IACJ,MAAMC,cAAcH,OAAO,OAAO,CAAC,IAAI;IACvC,IAAIG,aAAa;QACf,MAAMC,aAAahB,AAAAA,IAAAA,mCAAAA,eAAAA,AAAAA,EACjBe,aACA/C,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAEFhB,aAAa,0BAA0B4D;QAEvC,MAAMC,oBAAoBL,OAAO,OAAO,CAAC,eAAe,IAAI,EAAE;QAC9DxD,aAAa,wBAAwB6D;QAErC,MAAMC,iBAAiBD,kBACpB,MAAM,CAAC,CAACE,OAAStB,MAAM,OAAO,CAACsB,OAC/B,GAAG,CAAC,CAACA,OACGnB,AAAAA,IAAAA,mCAAAA,eAAAA,AAAAA,EACLmB,MACAnD,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAGNhB,aAAa,qBAAqB8D;QAGlC,MAAME,aAAaC,AAAAA,IAAAA,mCAAAA,UAAAA,AAAAA,EAAW;YAACL;eAAeE;SAAe;QAC7D9D,aAAa,iBAAiBgE;QAG9BN,cAAcQ,AAAAA,IAAAA,mCAAAA,gBAAAA,AAAAA,EAAiBF,YAAYpD,QAAQ,IAAI,EAAEI;QACzDhB,aAAa,2BAA2B0D;IAC1C;IAEA,IAAIS,cAAclD;IAClB,IAAIyC,aAAa;QACf,MAAMU,gBAAgB,MAAMC,AAAAA,IAAAA,oBAAAA,UAAAA,AAAAA,EAC1BpD,kBACAyC,aACA1C,AAAW,iBAAXA;QAEFmD,cAAcC,cAAc,WAAW;QACvCV,YAAY,KAAK,GAAGU,cAAc,KAAK;QACvCV,YAAY,MAAM,GAAGU,cAAc,MAAM;IAC3C;IAEA,OAAO;QACL,MAAMV;QACNS;QACA,OAAOX,OAAO,OAAO,CAAC,KAAK;QAC3B,aAAanB,KAAK,SAAS,CAACmB,OAAO,OAAO;QAC1C,OAAOA,OAAO,KAAK;IACrB;AACF;AAEO,eAAec,qBAAwB3D,OAO7C;IACC,MAAM,EAAE4D,SAAS,EAAE3D,OAAO,EAAE4D,aAAa,EAAEpE,gBAAgB,EAAEW,WAAW,EAAE,GACxEJ;IACF,MAAMW,eAAemD,AAAAA,IAAAA,8BAAAA,qBAAAA,AAAAA;IACrB,MAAM,EAAExD,gBAAgB,EAAE,GAAGL;IAE7B,MAAM8D,wBAAwB,MAAMC,AAAAA,IAAAA,8BAAAA,sBAAAA,AAAAA,EAClChE,QAAQ,eAAe,IAAI,IAC3B4D;IAGF,MAAMK,cAAyD,EAAE;IAEjE,IAAIJ,AAAAA,CAAAA,QAAAA,gBAAAA,KAAAA,IAAAA,cAAe,kBAAkB,AAAD,MAAM,OACxCI,YAAY,IAAI,CAAC;QACf,MAAM;QACN,WAAW;YACT,KAAK3D;YACL,QAAQ;QACV;IACF;IAGF2D,YAAY,IAAI,CAAC;QACf,MAAM;QACN,MAAMF;IACR;IAEA,MAAMpE,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAASsD;QACX;KACD;IAED,IAAIxE,kBAAkB;QACpB,MAAM6B,SAAS,MAAM9B,mBAAmB;YACtC,QAAQC,iBAAiB,MAAM;YAC/B,yBAAyBA,iBAAiB,uBAAuB;QACnE;QACAE,KAAK,IAAI,IAAI2B;IACf;IAEA,MAAMuB,SAAS,MAAMC,AAAAA,IAAAA,yBAAAA,wBAAAA,AAAAA,EACnBnD,MACA6B,mCAAAA,YAAAA,CAAAA,YAAyB,EACzBpB;IAEF,OAAO;QACL,aAAayC,OAAO,OAAO;QAC3B,OAAOA,OAAO,KAAK;IACrB;AACF;AAEO,eAAeqB,sBACpBC,WAAmB,EACnBhE,QAAwE,EACxEC,WAAyB;IAKzB,MAAMO,eAAeyD,AAAAA,IAAAA,yCAAAA,iCAAAA,AAAAA;IACrB,MAAMC,aAAaC,AAAAA,IAAAA,yCAAAA,yBAAAA,AAAAA,EAA0BH;IAE7C,MAAMxE,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAAS0D;QACX;KACD;IAED,MAAMxB,SAAS,MAAM1C,SACnBR,MACA6B,mCAAAA,YAAAA,CAAAA,eAA4B,EAC5BpB;IAGF,OAAO;QACL,kBAAkByC,OAAO,OAAO,CAAC,gBAAgB,IAAI;QACrD,OAAOA,OAAO,KAAK;IACrB;AACF"}
+{"version":3,"file":"ai-model/inspect.js","sources":["webpack://@midscene/core/webpack/runtime/define_property_getters","webpack://@midscene/core/webpack/runtime/has_own_property","webpack://@midscene/core/webpack/runtime/make_namespace_object","webpack://@midscene/core/./src/ai-model/inspect.ts"],"sourcesContent":["__webpack_require__.d = (exports, definition) => {\n\tfor(var key in definition) {\n        if(__webpack_require__.o(definition, key) && !__webpack_require__.o(exports, key)) {\n            Object.defineProperty(exports, key, { enumerable: true, get: definition[key] });\n        }\n    }\n};","__webpack_require__.o = (obj, prop) => (Object.prototype.hasOwnProperty.call(obj, prop))","// define __esModule on exports\n__webpack_require__.r = (exports) => {\n\tif(typeof Symbol !== 'undefined' && Symbol.toStringTag) {\n\t\tObject.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });\n\t}\n\tObject.defineProperty(exports, '__esModule', { value: true });\n};","import type {\n  AIDataExtractionResponse,\n  AIElementResponse,\n  AISectionLocatorResponse,\n  AIUsageInfo,\n  Rect,\n  ReferenceImage,\n  ServiceExtractOption,\n  UIContext,\n} from '@/types';\nimport type { IModelConfig } from '@midscene/shared/env';\nimport { generateElementByPosition } from '@midscene/shared/extractor/dom-util';\nimport {\n  cropByRect,\n  paddingToMatchBlockByBase64,\n  preProcessImageUrl,\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 {\n  AIActionType,\n  adaptBboxToRect,\n  expandSearchArea,\n  mergeRects,\n} from '../common';\nimport {\n  extractDataQueryPrompt,\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 { callAIWithObjectResponse } 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  referenceImage?: ReferenceImage;\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}> {\n  const { context, targetElementDescription, callAIFn, modelConfig } = options;\n  const { vlMode } = modelConfig;\n  const { screenshotBase64 } = context;\n\n  assert(\n    targetElementDescription,\n    'cannot find the target element description',\n  );\n  const targetElementDescriptionText = extraTextFromUserPrompt(\n    targetElementDescription,\n  );\n  const userInstructionPrompt = await findElementPrompt.format({\n    targetElementDescription: targetElementDescriptionText,\n  });\n  const systemPrompt = systemPromptToLocateElement(vlMode);\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 (vlMode === '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: 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  const res = await callAIFn(msgs, AIActionType.INSPECT_ELEMENT, modelConfig);\n\n  const rawResponse = JSON.stringify(res.content);\n\n  let resRect: Rect | undefined;\n  let matchedElements = 'elements' in res.content ? res.content.elements : [];\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        vlMode,\n      );\n\n      debugInspect('resRect', resRect);\n\n      const rectCenter = {\n        x: resRect.left + resRect.width / 2,\n        y: resRect.top + resRect.height / 2,\n      };\n\n      const element: LocateResultElement = generateElementByPosition(\n        rectCenter,\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  };\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  error?: string;\n  rawResponse: string;\n  usage?: AIUsageInfo;\n}> {\n  const { context, sectionDescription, modelConfig } = options;\n  const { vlMode } = modelConfig;\n  const { screenshotBase64 } = context;\n\n  const systemPrompt = systemPromptToLocateSection(vlMode);\n  const sectionLocatorInstructionText = await sectionLocatorInstruction.format({\n    sectionDescription: 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  const result = await callAIWithObjectResponse<AISectionLocatorResponse>(\n    msgs,\n    AIActionType.EXTRACT_DATA,\n    modelConfig,\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      vlMode,\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          vlMode,\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    // expand search area to at least 200 x 200\n    sectionRect = expandSearchArea(mergedRect, context.size, vlMode);\n    debugSection('expanded sectionRect %j', sectionRect);\n  }\n\n  let imageBase64 = screenshotBase64;\n  if (sectionRect) {\n    const croppedResult = await cropByRect(\n      screenshotBase64,\n      sectionRect,\n      vlMode === 'qwen2.5-vl',\n    );\n    imageBase64 = croppedResult.imageBase64;\n    sectionRect.width = croppedResult.width;\n    sectionRect.height = croppedResult.height;\n  }\n\n  return {\n    rect: sectionRect,\n    imageBase64,\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;\n\n  const extractDataPromptText = await 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 result = await callAIWithObjectResponse<AIDataExtractionResponse<T>>(\n    msgs,\n    AIActionType.EXTRACT_DATA,\n    modelConfig,\n  );\n  return {\n    parseResult: result.content,\n    usage: result.usage,\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(\n    msgs,\n    AIActionType.INSPECT_ELEMENT, // Reuse existing action type for now\n    modelConfig,\n  );\n\n  return {\n    isOrderSensitive: result.content.isOrderSensitive ?? false,\n    usage: result.usage,\n  };\n}\n"],"names":["__webpack_require__","definition","key","Object","obj","prop","Symbol","debugInspect","getDebug","debugSection","extraTextFromUserPrompt","prompt","promptsToChatParam","multimodalPrompt","_multimodalPrompt_images","msgs","item","base64","preProcessImageUrl","AiLocateElement","options","context","targetElementDescription","callAIFn","modelConfig","vlMode","screenshotBase64","assert","targetElementDescriptionText","userInstructionPrompt","findElementPrompt","systemPrompt","systemPromptToLocateElement","imagePayload","imageWidth","imageHeight","originalImageWidth","originalImageHeight","_options_searchConfig_rect","_options_searchConfig_rect1","paddedResult","paddingToMatchBlockByBase64","addOns","res","AIActionType","rawResponse","JSON","resRect","matchedElements","errors","Array","_options_searchConfig_rect2","_options_searchConfig_rect3","adaptBboxToRect","rectCenter","element","generateElementByPosition","e","msg","Error","AiLocateSection","sectionDescription","systemPromptToLocateSection","sectionLocatorInstructionText","sectionLocatorInstruction","result","callAIWithObjectResponse","sectionRect","sectionBbox","targetRect","referenceBboxList","referenceRects","bbox","mergedRect","mergeRects","expandSearchArea","imageBase64","croppedResult","cropByRect","AiExtractElementInfo","dataQuery","extractOption","systemPromptToExtract","extractDataPromptText","extractDataQueryPrompt","userContent","AiJudgeOrderSensitive","description","systemPromptToJudgeOrderSensitive","userPrompt","orderSensitiveJudgePrompt"],"mappings":";;;IAAAA,oBAAoB,CAAC,GAAG,CAAC,UAASC;QACjC,IAAI,IAAIC,OAAOD,WACR,IAAGD,oBAAoB,CAAC,CAACC,YAAYC,QAAQ,CAACF,oBAAoB,CAAC,CAAC,UAASE,MACzEC,OAAO,cAAc,CAAC,UAASD,KAAK;YAAE,YAAY;YAAM,KAAKD,UAAU,CAACC,IAAI;QAAC;IAGzF;;;ICNAF,oBAAoB,CAAC,GAAG,CAACI,KAAKC,OAAUF,OAAO,SAAS,CAAC,cAAc,CAAC,IAAI,CAACC,KAAKC;;;ICClFL,oBAAoB,CAAC,GAAG,CAAC;QACxB,IAAG,AAAkB,eAAlB,OAAOM,UAA0BA,OAAO,WAAW,EACrDH,OAAO,cAAc,CAAC,UAASG,OAAO,WAAW,EAAE;YAAE,OAAO;QAAS;QAEtEH,OAAO,cAAc,CAAC,UAAS,cAAc;YAAE,OAAO;QAAK;IAC5D;;;;;;;;;;;;;;;;;;;;ACgDA,MAAMI,eAAeC,AAAAA,IAAAA,uBAAAA,QAAAA,AAAAA,EAAS;AAC9B,MAAMC,eAAeD,AAAAA,IAAAA,uBAAAA,QAAAA,AAAAA,EAAS;AAE9B,MAAME,0BAA0B,CAACC;IAC/B,IAAI,AAAkB,YAAlB,OAAOA,QACT,OAAOA;IAEP,OAAOA,OAAO,MAAM;AAExB;AAEA,MAAMC,qBAAqB,OACzBC;QAGIC;IADJ,MAAMC,OAAyC,EAAE;IACjD,IAAID,QAAAA,mBAAAA,KAAAA,IAAAA,QAAAA,CAAAA,2BAAAA,iBAAkB,MAAM,AAAD,IAAvBA,KAAAA,IAAAA,yBAA0B,MAAM,EAAE;QACpCC,KAAK,IAAI,CAAC;YACR,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM;gBACR;aACD;QACH;QAEA,KAAK,MAAMC,QAAQH,iBAAiB,MAAM,CAAE;YAC1C,MAAMI,SAAS,MAAMC,AAAAA,IAAAA,oBAAAA,kBAAAA,AAAAA,EACnBF,KAAK,GAAG,EACR,CAAC,CAACH,iBAAiB,uBAAuB;YAG5CE,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,OASrC;IASC,MAAM,EAAEC,OAAO,EAAEC,wBAAwB,EAAEC,QAAQ,EAAEC,WAAW,EAAE,GAAGJ;IACrE,MAAM,EAAEK,MAAM,EAAE,GAAGD;IACnB,MAAM,EAAEE,gBAAgB,EAAE,GAAGL;IAE7BM,IAAAA,sBAAAA,MAAAA,AAAAA,EACEL,0BACA;IAEF,MAAMM,+BAA+BlB,wBACnCY;IAEF,MAAMO,wBAAwB,MAAMC,+BAAAA,iBAAAA,CAAAA,MAAwB,CAAC;QAC3D,0BAA0BF;IAC5B;IACA,MAAMG,eAAeC,AAAAA,IAAAA,+BAAAA,2BAAAA,AAAAA,EAA4BP;IAEjD,IAAIQ,eAAeP;IACnB,IAAIQ,aAAab,QAAQ,IAAI,CAAC,KAAK;IACnC,IAAIc,cAAcd,QAAQ,IAAI,CAAC,MAAM;IACrC,IAAIe,qBAAqBF;IACzB,IAAIG,sBAAsBF;IAE1B,IAAIf,QAAQ,YAAY,EAAE;YAWXkB,4BACCC;QAXdZ,IAAAA,sBAAAA,MAAAA,AAAAA,EACEP,QAAQ,YAAY,CAAC,IAAI,EACzB;QAEFO,IAAAA,sBAAAA,MAAAA,AAAAA,EACEP,QAAQ,YAAY,CAAC,WAAW,EAChC;QAGFa,eAAeb,QAAQ,YAAY,CAAC,WAAW;QAC/Cc,aAAa,QAAAI,CAAAA,6BAAAA,QAAQ,YAAY,CAAC,IAAI,AAAD,IAAxBA,KAAAA,IAAAA,2BAA2B,KAAK;QAC7CH,cAAc,QAAAI,CAAAA,8BAAAA,QAAQ,YAAY,CAAC,IAAI,AAAD,IAAxBA,KAAAA,IAAAA,4BAA2B,MAAM;QAC/CH,qBAAqBF;QACrBG,sBAAsBF;IACxB,OAAO,IAAIV,AAAW,iBAAXA,QAAyB;QAClC,MAAMe,eAAe,MAAMC,AAAAA,IAAAA,oBAAAA,2BAAAA,AAAAA,EAA4BR;QACvDC,aAAaM,aAAa,KAAK;QAC/BL,cAAcK,aAAa,MAAM;QACjCP,eAAeO,aAAa,WAAW;IACzC;IAEA,MAAMzB,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKE;wBACL,QAAQ;oBACV;gBACF;gBACA;oBACE,MAAM;oBACN,MAAMJ;gBACR;aACD;QACH;KACD;IAED,IAAI,AAAoC,YAApC,OAAOP,0BAAuC;QAChD,MAAMoB,SAAS,MAAM9B,mBAAmB;YACtC,QAAQU,yBAAyB,MAAM;YACvC,yBAAyBA,yBAAyB,uBAAuB;QAC3E;QACAP,KAAK,IAAI,IAAI2B;IACf;IAEA,MAAMC,MAAM,MAAMpB,SAASR,MAAM6B,mCAAAA,YAAAA,CAAAA,eAA4B,EAAEpB;IAE/D,MAAMqB,cAAcC,KAAK,SAAS,CAACH,IAAI,OAAO;IAE9C,IAAII;IACJ,IAAIC,kBAAkB,cAAcL,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,QAAQ,GAAG,EAAE;IAC3E,IAAIM,SACF,YAAYN,IAAI,OAAO,GAAGA,IAAI,OAAO,CAAC,MAAM,GAAG,EAAE;IACnD,IAAI;QACF,IACE,UAAUA,IAAI,OAAO,IACrBO,MAAM,OAAO,CAACP,IAAI,OAAO,CAAC,IAAI,KAC9BA,IAAI,OAAO,CAAC,IAAI,CAAC,MAAM,IAAI,GAC3B;gBAKEQ,6BAAAA,uBACAC,6BAAAA;YALFL,UAAUM,AAAAA,IAAAA,mCAAAA,eAAAA,AAAAA,EACRV,IAAI,OAAO,CAAC,IAAI,EAChBT,YACAC,aAAAA,QACAgB,CAAAA,wBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,sBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,IAAI,UAChCC,CAAAA,yBAAAA,QAAQ,YAAY,AAAD,IAAnBA,KAAAA,IAAAA,QAAAA,CAAAA,8BAAAA,uBAAsB,IAAI,AAAD,IAAzBA,KAAAA,IAAAA,4BAA4B,GAAG,EAC/BhB,oBACAC,qBACAZ;YAGFlB,aAAa,WAAWwC;YAExB,MAAMO,aAAa;gBACjB,GAAGP,QAAQ,IAAI,GAAGA,QAAQ,KAAK,GAAG;gBAClC,GAAGA,QAAQ,GAAG,GAAGA,QAAQ,MAAM,GAAG;YACpC;YAEA,MAAMQ,UAA+BC,AAAAA,IAAAA,yBAAAA,yBAAAA,AAAAA,EACnCF,YACA1B;YAEFqB,SAAS,EAAE;YAEX,IAAIM,SACFP,kBAAkB;gBAACO;aAAQ;QAE/B;IACF,EAAE,OAAOE,GAAG;QACV,MAAMC,MACJD,aAAaE,QACT,CAAC,sBAAsB,EAAEF,EAAE,OAAO,EAAE,GACpC;QACN,IAAI,AAACR,UAAUA,AAAAA,CAAAA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,MAAM,AAAD,MAAM,GAGhCA,OAAO,IAAI,CAAC,CAAC,CAAC,EAAES,IAAI,CAAC,CAAC;aAFtBT,SAAS;YAACS;SAAI;IAIlB;IAEA,OAAO;QACL,MAAMX;QACN,aAAa;YACX,UAAUC;YACV,QAAQC;QACV;QACAJ;QACA,OAAOF,IAAI,KAAK;IAClB;AACF;AAEO,eAAeiB,gBAAgBxC,OAIrC;IAOC,MAAM,EAAEC,OAAO,EAAEwC,kBAAkB,EAAErC,WAAW,EAAE,GAAGJ;IACrD,MAAM,EAAEK,MAAM,EAAE,GAAGD;IACnB,MAAM,EAAEE,gBAAgB,EAAE,GAAGL;IAE7B,MAAMU,eAAe+B,AAAAA,IAAAA,uCAAAA,2BAAAA,AAAAA,EAA4BrC;IACjD,MAAMsC,gCAAgC,MAAMC,uCAAAA,yBAAAA,CAAAA,MAAgC,CAAC;QAC3E,oBAAoBtD,wBAAwBmD;IAC9C;IACA,MAAM9C,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,MAAMqC;gBACR;aACD;QACH;KACD;IAED,IAAI,AAA8B,YAA9B,OAAOF,oBAAiC;QAC1C,MAAMnB,SAAS,MAAM9B,mBAAmB;YACtC,QAAQiD,mBAAmB,MAAM;YACjC,yBAAyBA,mBAAmB,uBAAuB;QACrE;QACA9C,KAAK,IAAI,IAAI2B;IACf;IAEA,MAAMuB,SAAS,MAAMC,AAAAA,IAAAA,yBAAAA,wBAAAA,AAAAA,EACnBnD,MACA6B,mCAAAA,YAAAA,CAAAA,YAAyB,EACzBpB;IAGF,IAAI2C;IACJ,MAAMC,cAAcH,OAAO,OAAO,CAAC,IAAI;IACvC,IAAIG,aAAa;QACf,MAAMC,aAAahB,AAAAA,IAAAA,mCAAAA,eAAAA,AAAAA,EACjBe,aACA/C,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAEFhB,aAAa,0BAA0B4D;QAEvC,MAAMC,oBAAoBL,OAAO,OAAO,CAAC,eAAe,IAAI,EAAE;QAC9DxD,aAAa,wBAAwB6D;QAErC,MAAMC,iBAAiBD,kBACpB,MAAM,CAAC,CAACE,OAAStB,MAAM,OAAO,CAACsB,OAC/B,GAAG,CAAC,CAACA,OACGnB,AAAAA,IAAAA,mCAAAA,eAAAA,AAAAA,EACLmB,MACAnD,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnB,GACA,GACAA,QAAQ,IAAI,CAAC,KAAK,EAClBA,QAAQ,IAAI,CAAC,MAAM,EACnBI;QAGNhB,aAAa,qBAAqB8D;QAGlC,MAAME,aAAaC,AAAAA,IAAAA,mCAAAA,UAAAA,AAAAA,EAAW;YAACL;eAAeE;SAAe;QAC7D9D,aAAa,iBAAiBgE;QAG9BN,cAAcQ,AAAAA,IAAAA,mCAAAA,gBAAAA,AAAAA,EAAiBF,YAAYpD,QAAQ,IAAI,EAAEI;QACzDhB,aAAa,2BAA2B0D;IAC1C;IAEA,IAAIS,cAAclD;IAClB,IAAIyC,aAAa;QACf,MAAMU,gBAAgB,MAAMC,AAAAA,IAAAA,oBAAAA,UAAAA,AAAAA,EAC1BpD,kBACAyC,aACA1C,AAAW,iBAAXA;QAEFmD,cAAcC,cAAc,WAAW;QACvCV,YAAY,KAAK,GAAGU,cAAc,KAAK;QACvCV,YAAY,MAAM,GAAGU,cAAc,MAAM;IAC3C;IAEA,OAAO;QACL,MAAMV;QACNS;QACA,OAAOX,OAAO,OAAO,CAAC,KAAK;QAC3B,aAAanB,KAAK,SAAS,CAACmB,OAAO,OAAO;QAC1C,OAAOA,OAAO,KAAK;IACrB;AACF;AAEO,eAAec,qBAAwB3D,OAO7C;IACC,MAAM,EAAE4D,SAAS,EAAE3D,OAAO,EAAE4D,aAAa,EAAEpE,gBAAgB,EAAEW,WAAW,EAAE,GACxEJ;IACF,MAAMW,eAAemD,AAAAA,IAAAA,8BAAAA,qBAAAA,AAAAA;IACrB,MAAM,EAAExD,gBAAgB,EAAE,GAAGL;IAE7B,MAAM8D,wBAAwB,MAAMC,AAAAA,IAAAA,8BAAAA,sBAAAA,AAAAA,EAClChE,QAAQ,eAAe,IAAI,IAC3B4D;IAGF,MAAMK,cAAyD,EAAE;IAEjE,IAAIJ,AAAAA,CAAAA,QAAAA,gBAAAA,KAAAA,IAAAA,cAAe,kBAAkB,AAAD,MAAM,OACxCI,YAAY,IAAI,CAAC;QACf,MAAM;QACN,WAAW;YACT,KAAK3D;YACL,QAAQ;QACV;IACF;IAGF2D,YAAY,IAAI,CAAC;QACf,MAAM;QACN,MAAMF;IACR;IAEA,MAAMpE,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAASsD;QACX;KACD;IAED,IAAIxE,kBAAkB;QACpB,MAAM6B,SAAS,MAAM9B,mBAAmB;YACtC,QAAQC,iBAAiB,MAAM;YAC/B,yBAAyBA,iBAAiB,uBAAuB;QACnE;QACAE,KAAK,IAAI,IAAI2B;IACf;IAEA,MAAMuB,SAAS,MAAMC,AAAAA,IAAAA,yBAAAA,wBAAAA,AAAAA,EACnBnD,MACA6B,mCAAAA,YAAAA,CAAAA,YAAyB,EACzBpB;IAEF,OAAO;QACL,aAAayC,OAAO,OAAO;QAC3B,OAAOA,OAAO,KAAK;IACrB;AACF;AAEO,eAAeqB,sBACpBC,WAAmB,EACnBhE,QAAwE,EACxEC,WAAyB;IAKzB,MAAMO,eAAeyD,AAAAA,IAAAA,yCAAAA,iCAAAA,AAAAA;IACrB,MAAMC,aAAaC,AAAAA,IAAAA,yCAAAA,yBAAAA,AAAAA,EAA0BH;IAE7C,MAAMxE,OAAe;QACnB;YAAE,MAAM;YAAU,SAASgB;QAAa;QACxC;YACE,MAAM;YACN,SAAS0D;QACX;KACD;IAED,MAAMxB,SAAS,MAAM1C,SACnBR,MACA6B,mCAAAA,YAAAA,CAAAA,eAA4B,EAC5BpB;IAGF,OAAO;QACL,kBAAkByC,OAAO,OAAO,CAAC,gBAAgB,IAAI;QACrD,OAAOA,OAAO,KAAK;IACrB;AACF"}

package/dist/lib/ai-model/llm-planning.js

@@ -29,7 +29,7 @@ __webpack_require__.d(__webpack_exports__, {
 const img_namespaceObject = require("@midscene/shared/img");
 const logger_namespaceObject = require("@midscene/shared/logger");
 const utils_namespaceObject = require("@midscene/shared/utils");
-const external_common_js_namespaceObject = require("./common.js");
+const external_common_js_namespaceObject = require("../common.js");
 const llm_planning_js_namespaceObject = require("./prompt/llm-planning.js");
 const index_js_namespaceObject = require("./service-caller/index.js");
 const debug = (0, logger_namespaceObject.getDebug)('planning');
@@ -38,10 +38,10 @@ async function plan(userInstruction, opts) {
     const { context, modelConfig, conversationHistory } = opts;
     const { screenshotBase64, size } = context;
     const { vlMode } = modelConfig;
-    (0, utils_namespaceObject.assert)(vlMode, 'Planning requires vlMode to be configured.');
     const systemPrompt = await (0, llm_planning_js_namespaceObject.systemPromptToTaskPlanning)({
         actionSpace: opts.actionSpace,
-        vlMode: vlMode
+        vlMode,
+        includeBbox: opts.includeBbox
     });
     let imagePayload = screenshotBase64;
     let imageWidth = size.width;
@@ -119,7 +119,7 @@ async function plan(userInstruction, opts) {
         debug('locateFields', locateFields);
         locateFields.forEach((field)=>{
             const locateResult = action.param[field];
-            if (locateResult) action.param[field] = (0, external_common_js_namespaceObject.fillBboxParam)(locateResult, imageWidth, imageHeight, rightLimit, bottomLimit, vlMode);
+            if (locateResult && void 0 !== vlMode) action.param[field] = (0, external_common_js_namespaceObject.fillBboxParam)(locateResult, imageWidth, imageHeight, rightLimit, bottomLimit, vlMode);
         });
     });
     (0, utils_namespaceObject.assert)(!planFromAI.error, `Failed to plan actions: ${planFromAI.error}`);

package/dist/lib/ai-model/llm-planning.js.map

@@ -1 +1 @@
-{"version":3,"file":"ai-model/llm-planning.js","sources":["webpack://@midscene/core/webpack/runtime/define_property_getters","webpack://@midscene/core/webpack/runtime/has_own_property","webpack://@midscene/core/webpack/runtime/make_namespace_object","webpack://@midscene/core/./src/ai-model/llm-planning.ts"],"sourcesContent":["__webpack_require__.d = (exports, definition) => {\n\tfor(var key in definition) {\n        if(__webpack_require__.o(definition, key) && !__webpack_require__.o(exports, key)) {\n            Object.defineProperty(exports, key, { enumerable: true, get: definition[key] });\n        }\n    }\n};","__webpack_require__.o = (obj, prop) => (Object.prototype.hasOwnProperty.call(obj, prop))","// define __esModule on exports\n__webpack_require__.r = (exports) => {\n\tif(typeof Symbol !== 'undefined' && Symbol.toStringTag) {\n\t\tObject.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });\n\t}\n\tObject.defineProperty(exports, '__esModule', { value: true });\n};","import type {\n  DeviceAction,\n  InterfaceType,\n  PlanningAIResponse,\n  RawResponsePlanningAIResponse,\n  UIContext,\n} from '@/types';\nimport type { IModelConfig } 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  AIActionType,\n  buildYamlFlowFromPlans,\n  fillBboxParam,\n  findAllMidsceneLocatorField,\n} from './common';\nimport type { ConversationHistory } from './conversation-history';\nimport { systemPromptToTaskPlanning } from './prompt/llm-planning';\nimport { callAIWithObjectResponse } from './service-caller/index';\n\nconst debug = getDebug('planning');\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  },\n): Promise<PlanningAIResponse> {\n  const { context, modelConfig, conversationHistory } = opts;\n  const { screenshotBase64, size } = context;\n\n  const { vlMode } = modelConfig;\n\n  // Planning requires VL mode (validated by ModelConfigManager.getModelConfig)\n  assert(vlMode, 'Planning requires vlMode to be configured.');\n\n  const systemPrompt = await systemPromptToTaskPlanning({\n    actionSpace: opts.actionSpace,\n    vlMode: vlMode,\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 (vlMode === '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 historyLog = opts.conversationHistory?.snapshot() || [];\n  // .filter((item) => item.role === 'assistant') || [];\n\n  const knowledgeContext: ChatCompletionMessageParam[] = opts.actionContext\n    ? [\n        {\n          role: 'user',\n          content: [\n            {\n              type: 'text',\n              text: `<high_priority_knowledge>${opts.actionContext}</high_priority_knowledge>`,\n            },\n          ],\n        },\n      ]\n    : [];\n\n  const instruction: ChatCompletionMessageParam[] = [\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `<user_instruction>${userInstruction}</user_instruction>`,\n        },\n      ],\n    },\n  ];\n\n  const msgs: ChatCompletionMessageParam[] = [\n    { role: 'system', content: systemPrompt },\n    ...knowledgeContext,\n    ...instruction,\n    ...historyLog,\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n        // Planning uses pure vision mode, no DOM description needed\n      ],\n    },\n  ];\n\n  const { content: planFromAI, usage } =\n    await callAIWithObjectResponse<RawResponsePlanningAIResponse>(\n      msgs,\n      AIActionType.PLAN,\n      modelConfig,\n    );\n  const rawResponse = JSON.stringify(planFromAI, undefined, 2);\n\n  const actions = planFromAI.action ? [planFromAI.action] : [];\n  const returnValue: PlanningAIResponse = {\n    ...planFromAI,\n    actions,\n    rawResponse,\n    usage,\n    yamlFlow: buildYamlFlowFromPlans(\n      actions,\n      opts.actionSpace,\n      planFromAI.sleep,\n    ),\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) {\n        // Always use VL mode to fill bbox parameters\n        action.param[field] = fillBboxParam(\n          locateResult,\n          imageWidth,\n          imageHeight,\n          rightLimit,\n          bottomLimit,\n          vlMode,\n        );\n      }\n    });\n  });\n  // in Qwen-VL, error means error. In GPT-4o, error may mean more actions are needed.\n  assert(!planFromAI.error, `Failed to plan actions: ${planFromAI.error}`);\n\n  if (\n    actions.length === 0 &&\n    returnValue.more_actions_needed_by_instruction &&\n    !returnValue.sleep\n  ) {\n    console.warn(\n      'No actions planned for the prompt, but model said more actions are needed:',\n      userInstruction,\n    );\n  }\n\n  conversationHistory?.append({\n    role: 'assistant',\n    content: [\n      {\n        type: 'text',\n        text: rawResponse,\n      },\n    ],\n  });\n  conversationHistory?.append({\n    role: 'user',\n    content: [\n      {\n        type: 'text',\n        text: 'I have finished the action previously planned',\n      },\n    ],\n  });\n\n  return returnValue;\n}\n"],"names":["__webpack_require__","definition","key","Object","obj","prop","Symbol","debug","getDebug","plan","userInstruction","opts","_opts_conversationHistory","context","modelConfig","conversationHistory","screenshotBase64","size","vlMode","assert","systemPrompt","systemPromptToTaskPlanning","imagePayload","imageWidth","imageHeight","rightLimit","bottomLimit","paddedResult","paddingToMatchBlockByBase64","historyLog","knowledgeContext","instruction","msgs","planFromAI","usage","callAIWithObjectResponse","AIActionType","rawResponse","JSON","undefined","actions","returnValue","buildYamlFlowFromPlans","action","type","actionInActionSpace","locateFields","findAllMidsceneLocatorField","field","locateResult","fillBboxParam","console"],"mappings":";;;IAAAA,oBAAoB,CAAC,GAAG,CAAC,UAASC;QACjC,IAAI,IAAIC,OAAOD,WACR,IAAGD,oBAAoB,CAAC,CAACC,YAAYC,QAAQ,CAACF,oBAAoB,CAAC,CAAC,UAASE,MACzEC,OAAO,cAAc,CAAC,UAASD,KAAK;YAAE,YAAY;YAAM,KAAKD,UAAU,CAACC,IAAI;QAAC;IAGzF;;;ICNAF,oBAAoB,CAAC,GAAG,CAACI,KAAKC,OAAUF,OAAO,SAAS,CAAC,cAAc,CAAC,IAAI,CAACC,KAAKC;;;ICClFL,oBAAoB,CAAC,GAAG,CAAC;QACxB,IAAG,AAAkB,eAAlB,OAAOM,UAA0BA,OAAO,WAAW,EACrDH,OAAO,cAAc,CAAC,UAASG,OAAO,WAAW,EAAE;YAAE,OAAO;QAAS;QAEtEH,OAAO,cAAc,CAAC,UAAS,cAAc;YAAE,OAAO;QAAK;IAC5D;;;;;;;;;;;;;ACgBA,MAAMI,QAAQC,AAAAA,IAAAA,uBAAAA,QAAAA,AAAAA,EAAS;AAEhB,eAAeC,KACpBC,eAAuB,EACvBC,IAOC;QA6BkBC;IA3BnB,MAAM,EAAEC,OAAO,EAAEC,WAAW,EAAEC,mBAAmB,EAAE,GAAGJ;IACtD,MAAM,EAAEK,gBAAgB,EAAEC,IAAI,EAAE,GAAGJ;IAEnC,MAAM,EAAEK,MAAM,EAAE,GAAGJ;IAGnBK,IAAAA,sBAAAA,MAAAA,AAAAA,EAAOD,QAAQ;IAEf,MAAME,eAAe,MAAMC,AAAAA,IAAAA,gCAAAA,0BAAAA,AAAAA,EAA2B;QACpD,aAAaV,KAAK,WAAW;QAC7B,QAAQO;IACV;IAEA,IAAII,eAAeN;IACnB,IAAIO,aAAaN,KAAK,KAAK;IAC3B,IAAIO,cAAcP,KAAK,MAAM;IAC7B,MAAMQ,aAAaF;IACnB,MAAMG,cAAcF;IAGpB,IAAIN,AAAW,iBAAXA,QAAyB;QAC3B,MAAMS,eAAe,MAAMC,AAAAA,IAAAA,oBAAAA,2BAAAA,AAAAA,EAA4BN;QACvDC,aAAaI,aAAa,KAAK;QAC/BH,cAAcG,aAAa,MAAM;QACjCL,eAAeK,aAAa,WAAW;IACzC;IAEA,MAAME,aAAajB,AAAAA,SAAAA,CAAAA,4BAAAA,KAAK,mBAAmB,AAAD,IAAvBA,KAAAA,IAAAA,0BAA0B,QAAQ,EAAC,KAAK,EAAE;IAG7D,MAAMkB,mBAAiDnB,KAAK,aAAa,GACrE;QACE;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,CAAC,yBAAyB,EAAEA,KAAK,aAAa,CAAC,0BAA0B,CAAC;gBAClF;aACD;QACH;KACD,GACD,EAAE;IAEN,MAAMoB,cAA4C;QAChD;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,CAAC,kBAAkB,EAAErB,gBAAgB,mBAAmB,CAAC;gBACjE;aACD;QACH;KACD;IAED,MAAMsB,OAAqC;QACzC;YAAE,MAAM;YAAU,SAASZ;QAAa;WACrCU;WACAC;WACAF;QACH;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKP;wBACL,QAAQ;oBACV;gBACF;aAED;QACH;KACD;IAED,MAAM,EAAE,SAASW,UAAU,EAAEC,KAAK,EAAE,GAClC,MAAMC,AAAAA,IAAAA,yBAAAA,wBAAAA,AAAAA,EACJH,MACAI,mCAAAA,YAAAA,CAAAA,IAAiB,EACjBtB;IAEJ,MAAMuB,cAAcC,KAAK,SAAS,CAACL,YAAYM,QAAW;IAE1D,MAAMC,UAAUP,WAAW,MAAM,GAAG;QAACA,WAAW,MAAM;KAAC,GAAG,EAAE;IAC5D,MAAMQ,cAAkC;QACtC,GAAGR,UAAU;QACbO;QACAH;QACAH;QACA,UAAUQ,AAAAA,IAAAA,mCAAAA,sBAAAA,AAAAA,EACRF,SACA7B,KAAK,WAAW,EAChBsB,WAAW,KAAK;IAEpB;IAEAd,IAAAA,sBAAAA,MAAAA,AAAAA,EAAOc,YAAY;IAEnBO,QAAQ,OAAO,CAAC,CAACG;QACf,MAAMC,OAAOD,OAAO,IAAI;QACxB,MAAME,sBAAsBlC,KAAK,WAAW,CAAC,IAAI,CAC/C,CAACgC,SAAWA,OAAO,IAAI,KAAKC;QAG9BrC,MAAM,+BAA+BsC;QACrC,MAAMC,eAAeD,sBACjBE,AAAAA,IAAAA,mCAAAA,2BAAAA,AAAAA,EAA4BF,oBAAoB,WAAW,IAC3D,EAAE;QAENtC,MAAM,gBAAgBuC;QAEtBA,aAAa,OAAO,CAAC,CAACE;YACpB,MAAMC,eAAeN,OAAO,KAAK,CAACK,MAAM;YACxC,IAAIC,cAEFN,OAAO,KAAK,CAACK,MAAM,GAAGE,AAAAA,IAAAA,mCAAAA,aAAAA,AAAAA,EACpBD,cACA1B,YACAC,aACAC,YACAC,aACAR;QAGN;IACF;IAEAC,IAAAA,sBAAAA,MAAAA,AAAAA,EAAO,CAACc,WAAW,KAAK,EAAE,CAAC,wBAAwB,EAAEA,WAAW,KAAK,EAAE;IAEvE,IACEO,AAAmB,MAAnBA,QAAQ,MAAM,IACdC,YAAY,kCAAkC,IAC9C,CAACA,YAAY,KAAK,EAElBU,QAAQ,IAAI,CACV,8EACAzC;IAIJK,QAAAA,uBAAAA,oBAAqB,MAAM,CAAC;QAC1B,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAMsB;YACR;SACD;IACH;IACAtB,QAAAA,uBAAAA,oBAAqB,MAAM,CAAC;QAC1B,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAM;YACR;SACD;IACH;IAEA,OAAO0B;AACT"}
+{"version":3,"file":"ai-model/llm-planning.js","sources":["webpack://@midscene/core/webpack/runtime/define_property_getters","webpack://@midscene/core/webpack/runtime/has_own_property","webpack://@midscene/core/webpack/runtime/make_namespace_object","webpack://@midscene/core/./src/ai-model/llm-planning.ts"],"sourcesContent":["__webpack_require__.d = (exports, definition) => {\n\tfor(var key in definition) {\n        if(__webpack_require__.o(definition, key) && !__webpack_require__.o(exports, key)) {\n            Object.defineProperty(exports, key, { enumerable: true, get: definition[key] });\n        }\n    }\n};","__webpack_require__.o = (obj, prop) => (Object.prototype.hasOwnProperty.call(obj, prop))","// define __esModule on exports\n__webpack_require__.r = (exports) => {\n\tif(typeof Symbol !== 'undefined' && Symbol.toStringTag) {\n\t\tObject.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });\n\t}\n\tObject.defineProperty(exports, '__esModule', { value: true });\n};","import type {\n  DeviceAction,\n  InterfaceType,\n  PlanningAIResponse,\n  RawResponsePlanningAIResponse,\n  UIContext,\n} from '@/types';\nimport type { IModelConfig } 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  AIActionType,\n  buildYamlFlowFromPlans,\n  fillBboxParam,\n  findAllMidsceneLocatorField,\n} from '../common';\nimport type { ConversationHistory } from './conversation-history';\nimport { systemPromptToTaskPlanning } from './prompt/llm-planning';\nimport { callAIWithObjectResponse } from './service-caller/index';\n\nconst debug = getDebug('planning');\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  },\n): Promise<PlanningAIResponse> {\n  const { context, modelConfig, conversationHistory } = opts;\n  const { screenshotBase64, size } = context;\n\n  const { vlMode } = modelConfig;\n\n  const systemPrompt = await systemPromptToTaskPlanning({\n    actionSpace: opts.actionSpace,\n    vlMode,\n    includeBbox: opts.includeBbox,\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 (vlMode === '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 historyLog = opts.conversationHistory?.snapshot() || [];\n  // .filter((item) => item.role === 'assistant') || [];\n\n  const knowledgeContext: ChatCompletionMessageParam[] = opts.actionContext\n    ? [\n        {\n          role: 'user',\n          content: [\n            {\n              type: 'text',\n              text: `<high_priority_knowledge>${opts.actionContext}</high_priority_knowledge>`,\n            },\n          ],\n        },\n      ]\n    : [];\n\n  const instruction: ChatCompletionMessageParam[] = [\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `<user_instruction>${userInstruction}</user_instruction>`,\n        },\n      ],\n    },\n  ];\n\n  const msgs: ChatCompletionMessageParam[] = [\n    { role: 'system', content: systemPrompt },\n    ...knowledgeContext,\n    ...instruction,\n    ...historyLog,\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n        // Planning uses pure vision mode, no DOM description needed\n      ],\n    },\n  ];\n\n  const { content: planFromAI, usage } =\n    await callAIWithObjectResponse<RawResponsePlanningAIResponse>(\n      msgs,\n      AIActionType.PLAN,\n      modelConfig,\n    );\n  const rawResponse = JSON.stringify(planFromAI, undefined, 2);\n\n  const actions = planFromAI.action ? [planFromAI.action] : [];\n  const returnValue: PlanningAIResponse = {\n    ...planFromAI,\n    actions,\n    rawResponse,\n    usage,\n    yamlFlow: buildYamlFlowFromPlans(\n      actions,\n      opts.actionSpace,\n      planFromAI.sleep,\n    ),\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 && vlMode !== undefined) {\n        // Always use VL mode to fill bbox parameters\n        action.param[field] = fillBboxParam(\n          locateResult,\n          imageWidth,\n          imageHeight,\n          rightLimit,\n          bottomLimit,\n          vlMode,\n        );\n      }\n    });\n  });\n  // in Qwen-VL, error means error. In GPT-4o, error may mean more actions are needed.\n  assert(!planFromAI.error, `Failed to plan actions: ${planFromAI.error}`);\n\n  if (\n    actions.length === 0 &&\n    returnValue.more_actions_needed_by_instruction &&\n    !returnValue.sleep\n  ) {\n    console.warn(\n      'No actions planned for the prompt, but model said more actions are needed:',\n      userInstruction,\n    );\n  }\n\n  conversationHistory?.append({\n    role: 'assistant',\n    content: [\n      {\n        type: 'text',\n        text: rawResponse,\n      },\n    ],\n  });\n  conversationHistory?.append({\n    role: 'user',\n    content: [\n      {\n        type: 'text',\n        text: 'I have finished the action previously planned',\n      },\n    ],\n  });\n\n  return returnValue;\n}\n"],"names":["__webpack_require__","definition","key","Object","obj","prop","Symbol","debug","getDebug","plan","userInstruction","opts","_opts_conversationHistory","context","modelConfig","conversationHistory","screenshotBase64","size","vlMode","systemPrompt","systemPromptToTaskPlanning","imagePayload","imageWidth","imageHeight","rightLimit","bottomLimit","paddedResult","paddingToMatchBlockByBase64","historyLog","knowledgeContext","instruction","msgs","planFromAI","usage","callAIWithObjectResponse","AIActionType","rawResponse","JSON","undefined","actions","returnValue","buildYamlFlowFromPlans","assert","action","type","actionInActionSpace","locateFields","findAllMidsceneLocatorField","field","locateResult","fillBboxParam","console"],"mappings":";;;IAAAA,oBAAoB,CAAC,GAAG,CAAC,UAASC;QACjC,IAAI,IAAIC,OAAOD,WACR,IAAGD,oBAAoB,CAAC,CAACC,YAAYC,QAAQ,CAACF,oBAAoB,CAAC,CAAC,UAASE,MACzEC,OAAO,cAAc,CAAC,UAASD,KAAK;YAAE,YAAY;YAAM,KAAKD,UAAU,CAACC,IAAI;QAAC;IAGzF;;;ICNAF,oBAAoB,CAAC,GAAG,CAACI,KAAKC,OAAUF,OAAO,SAAS,CAAC,cAAc,CAAC,IAAI,CAACC,KAAKC;;;ICClFL,oBAAoB,CAAC,GAAG,CAAC;QACxB,IAAG,AAAkB,eAAlB,OAAOM,UAA0BA,OAAO,WAAW,EACrDH,OAAO,cAAc,CAAC,UAASG,OAAO,WAAW,EAAE;YAAE,OAAO;QAAS;QAEtEH,OAAO,cAAc,CAAC,UAAS,cAAc;YAAE,OAAO;QAAK;IAC5D;;;;;;;;;;;;;ACgBA,MAAMI,QAAQC,AAAAA,IAAAA,uBAAAA,QAAAA,AAAAA,EAAS;AAEhB,eAAeC,KACpBC,eAAuB,EACvBC,IAQC;QA2BkBC;IAzBnB,MAAM,EAAEC,OAAO,EAAEC,WAAW,EAAEC,mBAAmB,EAAE,GAAGJ;IACtD,MAAM,EAAEK,gBAAgB,EAAEC,IAAI,EAAE,GAAGJ;IAEnC,MAAM,EAAEK,MAAM,EAAE,GAAGJ;IAEnB,MAAMK,eAAe,MAAMC,AAAAA,IAAAA,gCAAAA,0BAAAA,AAAAA,EAA2B;QACpD,aAAaT,KAAK,WAAW;QAC7BO;QACA,aAAaP,KAAK,WAAW;IAC/B;IAEA,IAAIU,eAAeL;IACnB,IAAIM,aAAaL,KAAK,KAAK;IAC3B,IAAIM,cAAcN,KAAK,MAAM;IAC7B,MAAMO,aAAaF;IACnB,MAAMG,cAAcF;IAGpB,IAAIL,AAAW,iBAAXA,QAAyB;QAC3B,MAAMQ,eAAe,MAAMC,AAAAA,IAAAA,oBAAAA,2BAAAA,AAAAA,EAA4BN;QACvDC,aAAaI,aAAa,KAAK;QAC/BH,cAAcG,aAAa,MAAM;QACjCL,eAAeK,aAAa,WAAW;IACzC;IAEA,MAAME,aAAahB,AAAAA,SAAAA,CAAAA,4BAAAA,KAAK,mBAAmB,AAAD,IAAvBA,KAAAA,IAAAA,0BAA0B,QAAQ,EAAC,KAAK,EAAE;IAG7D,MAAMiB,mBAAiDlB,KAAK,aAAa,GACrE;QACE;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,CAAC,yBAAyB,EAAEA,KAAK,aAAa,CAAC,0BAA0B,CAAC;gBAClF;aACD;QACH;KACD,GACD,EAAE;IAEN,MAAMmB,cAA4C;QAChD;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,CAAC,kBAAkB,EAAEpB,gBAAgB,mBAAmB,CAAC;gBACjE;aACD;QACH;KACD;IAED,MAAMqB,OAAqC;QACzC;YAAE,MAAM;YAAU,SAASZ;QAAa;WACrCU;WACAC;WACAF;QACH;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKP;wBACL,QAAQ;oBACV;gBACF;aAED;QACH;KACD;IAED,MAAM,EAAE,SAASW,UAAU,EAAEC,KAAK,EAAE,GAClC,MAAMC,AAAAA,IAAAA,yBAAAA,wBAAAA,AAAAA,EACJH,MACAI,mCAAAA,YAAAA,CAAAA,IAAiB,EACjBrB;IAEJ,MAAMsB,cAAcC,KAAK,SAAS,CAACL,YAAYM,QAAW;IAE1D,MAAMC,UAAUP,WAAW,MAAM,GAAG;QAACA,WAAW,MAAM;KAAC,GAAG,EAAE;IAC5D,MAAMQ,cAAkC;QACtC,GAAGR,UAAU;QACbO;QACAH;QACAH;QACA,UAAUQ,AAAAA,IAAAA,mCAAAA,sBAAAA,AAAAA,EACRF,SACA5B,KAAK,WAAW,EAChBqB,WAAW,KAAK;IAEpB;IAEAU,IAAAA,sBAAAA,MAAAA,AAAAA,EAAOV,YAAY;IAEnBO,QAAQ,OAAO,CAAC,CAACI;QACf,MAAMC,OAAOD,OAAO,IAAI;QACxB,MAAME,sBAAsBlC,KAAK,WAAW,CAAC,IAAI,CAC/C,CAACgC,SAAWA,OAAO,IAAI,KAAKC;QAG9BrC,MAAM,+BAA+BsC;QACrC,MAAMC,eAAeD,sBACjBE,AAAAA,IAAAA,mCAAAA,2BAAAA,AAAAA,EAA4BF,oBAAoB,WAAW,IAC3D,EAAE;QAENtC,MAAM,gBAAgBuC;QAEtBA,aAAa,OAAO,CAAC,CAACE;YACpB,MAAMC,eAAeN,OAAO,KAAK,CAACK,MAAM;YACxC,IAAIC,gBAAgB/B,AAAWoB,WAAXpB,QAElByB,OAAO,KAAK,CAACK,MAAM,GAAGE,AAAAA,IAAAA,mCAAAA,aAAAA,AAAAA,EACpBD,cACA3B,YACAC,aACAC,YACAC,aACAP;QAGN;IACF;IAEAwB,IAAAA,sBAAAA,MAAAA,AAAAA,EAAO,CAACV,WAAW,KAAK,EAAE,CAAC,wBAAwB,EAAEA,WAAW,KAAK,EAAE;IAEvE,IACEO,AAAmB,MAAnBA,QAAQ,MAAM,IACdC,YAAY,kCAAkC,IAC9C,CAACA,YAAY,KAAK,EAElBW,QAAQ,IAAI,CACV,8EACAzC;IAIJK,QAAAA,uBAAAA,oBAAqB,MAAM,CAAC;QAC1B,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAMqB;YACR;SACD;IACH;IACArB,QAAAA,uBAAAA,oBAAqB,MAAM,CAAC;QAC1B,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAM;YACR;SACD;IACH;IAEA,OAAOyB;AACT"}

package/dist/lib/ai-model/prompt/llm-planning.js

@@ -28,14 +28,14 @@ __webpack_require__.d(__webpack_exports__, {
     descriptionForAction: ()=>descriptionForAction,
     planSchema: ()=>planSchema
 });
-const external_common_js_namespaceObject = require("../common.js");
+const external_common_js_namespaceObject = require("../../common.js");
 const external_common_js_namespaceObject_1 = require("./common.js");
-const vlCurrentLog = '"log": string, // Log your thoughts and what the next one action (ONLY ONE!) you can do according to the screenshot and the instruction. The log should contain the following information: "The user wants to do ... . According to the instruction and the previous logs, next step is to .... Now i am going to compose an action \'{ action-type }\' to do ....". If no action should be done, log the reason. Use the same language as the user\'s instruction.';
-const llmCurrentLog = '"log": string, // Log what the next actions you can do according to the screenshot and the instruction. The typical log looks like "Now i want to use action \'{ action-type }\' to do ..". If no action should be done, log the reason. ". Use the same language as the user\'s instruction.';
 const commonOutputFields = `"error"?: string, // Error messages about unexpected situations, if any. Only think it is an error when the situation is not foreseeable according to the instruction. Use the same language as the user's instruction.
   "more_actions_needed_by_instruction": boolean, // Consider if there is still more action(s) to do after the action in "Log" is done, according to the instruction. If so, set this field to true. Otherwise, set it to false.`;
-const vlLocateParam = ()=>'{bbox: [number, number, number, number], prompt: string }';
-const llmLocateParam = ()=>'{"id": string, "prompt": string}';
+const vlLocateParam = (vlMode)=>{
+    if (vlMode) return `{bbox: [number, number, number, number], prompt: string } // ${(0, external_common_js_namespaceObject_1.bboxDescription)(vlMode)}`;
+    return "{ prompt: string /* description of the target element */ }";
+};
 const descriptionForAction = (action, locatorSchemaTypeDescription)=>{
     const tab = '  ';
     const fields = [];
@@ -137,28 +137,25 @@ const descriptionForAction = (action, locatorSchemaTypeDescription)=>{
 ${tab}${fields.join(`\n${tab}`)}
 `.trim();
 };
-const systemTemplateOfVLPlanning = ({ actionSpace, vlMode })=>{
-    const actionNameList = actionSpace.map((action)=>action.name).join(', ');
-    const actionDescriptionList = actionSpace.map((action)=>descriptionForAction(action, vlLocateParam()));
+async function systemPromptToTaskPlanning({ actionSpace, vlMode, includeBbox }) {
+    if (includeBbox && !vlMode) throw new Error('vlMode cannot be undefined when includeBbox is true. A valid vlMode is required for bbox-based location.');
+    const actionDescriptionList = actionSpace.map((action)=>descriptionForAction(action, vlLocateParam(includeBbox ? vlMode : void 0)));
     const actionList = actionDescriptionList.join('\n');
     return `
-Target: User will give you a latest screenshot, an instruction and some previous logs indicating what have been done. Please tell what the next one action is (or null if no action should be done) to do the tasks the instruction requires. 
+Target: User will give you an instruction, some screenshots and previous logs indicating what have been done. Your task is to plan the next one action to accomplish the instruction.
+
+Please tell what the next one action is (or null if no action should be done) to do the tasks the instruction requires. 
 
 Restriction:
 - Don't give extra actions or plans beyond the instruction. ONLY plan for what the instruction requires. For example, don't try to submit the form if the instruction is only to fill something.
-- Always give ONLY ONE action in \`log\` field (or null if no action should be done), instead of multiple actions. Supported actions are ${actionNameList}.
-- Don't repeat actions in the previous logs.
-- Bbox is the bounding box of the element to be located. It's an array of 4 numbers, representing ${(0, external_common_js_namespaceObject_1.bboxDescription)(vlMode)}.
+- Give just the next ONE action you should do
 
 Supporting actions:
 ${actionList}
 
-Field description:
-* The \`prompt\` field inside the \`locate\` field is a short description that could be used to locate the element.
-
 Return in JSON format:
 {
-  ${vlCurrentLog}
+  "log": string, // Log your thoughts and what the next one action (ONLY ONE!) you can do according to the screenshot and the instruction. The log should contain the following information: "The user wants to do ... . According to the instruction and the previous logs, next step is to .... Now i am going to compose an action '{ action-type }' to do this". If no action should be done, log the reason. Use the same language as the user's instruction.
   ${commonOutputFields}
   "action": 
     {
@@ -178,7 +175,7 @@ this and output the JSON:
     "type": "Tap",
     "param": {
       "locate": {
-        "bbox": [100, 100, 200, 200],
+        ${vlMode ? '"bbox": [100, 100, 200, 200],' : ''}
         "prompt": "The 'Yes' button in popup"
       }
     }
@@ -186,137 +183,6 @@ this and output the JSON:
   "more_actions_needed_by_instruction": false,
 }
 `;
-};
-const systemTemplateOfLLM = ({ actionSpace })=>{
-    const actionNameList = actionSpace.map((action)=>action.name).join(' / ');
-    const actionDescriptionList = actionSpace.map((action)=>descriptionForAction(action, llmLocateParam()));
-    const actionList = actionDescriptionList.join('\n');
-    return `
-## Role
-
-You are a versatile professional in software UI automation. Your outstanding contributions will impact the user experience of billions of users.
-
-## Objective
-
-- Decompose the instruction user asked into a series of actions
-- Locate the target element if possible
-- If the instruction cannot be accomplished, give a further plan.
-
-## Workflow
-
-1. Receive the screenshot, element description of screenshot(if any), user's instruction and previous logs.
-2. Decompose the user's task into a sequence of feasible actions, and place it in the \`actions\` field. There are different types of actions (${actionNameList}). The "About the action" section below will give you more details.
-3. Consider whether the user's instruction will be accomplished after the actions you composed.
-- If the instruction is accomplished, set \`more_actions_needed_by_instruction\` to false.
-- If more actions are needed, set \`more_actions_needed_by_instruction\` to true. Get ready to hand over to the next talent people like you. Carefully log what have been done in the \`log\` field, he or she will continue the task according to your logs.
-4. If the task is not feasible on this page, set \`error\` field to the reason.
-
-## Constraints
-
-- All the actions you composed MUST be feasible, which means all the action fields can be filled with the page context information you get. If not, don't plan this action.
-- Trust the "What have been done" field about the task (if any), don't repeat actions in it.
-- Respond only with valid JSON. Do not write an introduction or summary or markdown prefix like \`\`\`json\`\`\`.
-- If the screenshot and the instruction are totally irrelevant, set reason in the \`error\` field.
-
-## About the \`actions\` field
-
-The \`locate\` param is commonly used in the \`param\` field of the action, means to locate the target element to perform the action, it conforms to the following scheme:
-
-type LocateParam = {
-  "id": string, // the id of the element found. It should either be the id marked with a rectangle in the screenshot or the id described in the description.
-  "prompt"?: string // the description of the element to find. It can only be omitted when locate is null.
-} | null // If it's not on the page, the LocateParam should be null
-
-## Supported actions
-
-Each action has a \`type\` and corresponding \`param\`. To be detailed:
-${actionList}
-
-`.trim();
-};
-const outputTemplate = `
-## Output JSON Format:
-
-The JSON format is as follows:
-
-{
-  "actions": [
-    // ... some actions
-  ],
-  ${llmCurrentLog}
-  ${commonOutputFields}
-}
-
-## Examples
-
-### Example: Decompose a task
-
-When you received the following information:
-
-* Instruction: 'Click the language switch button, wait 1s, click "English"'
-* Logs: null
-* Page Context (screenshot and description) shows: There is a language switch button, and the "English" option is not shown in the screenshot now.
-
-By viewing the page screenshot and description, you should consider this and output the JSON:
-
-* The user intent is: tap the switch button, sleep, and tap the 'English' option
-* The language switch button is shown in the screenshot, and can be located by the page description or the id marked with a rectangle. So we can plan a Tap action to do this.
-* Plan a Sleep action to wait for 1 second to ensure the language options are displayed.
-* The "English" option button is not shown in the screenshot now, it means it may only show after the previous actions are finished. So don't plan any action to do this.
-* Compose the log: The user wants to do click the language switch button, wait 1s, click "English". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.
-* The task cannot be accomplished (because the last tapping action is not finished yet), so the \`more_actions_needed_by_instruction\` field is true. The \`error\` field is null.
-
-{
-  "actions":[
-    {
-      "thought": "Click the language switch button to open the language options.",
-      "type": "Tap", 
-      "param": {
-        "locate": { id: "c81c4e9a33", prompt: "The language switch button" }
-      }
-    },
-    {
-      "thought": "Wait for 1 second to ensure the language options are displayed.",
-      "type": "Sleep",
-      "param": { "timeMs": 1000 },
-    }
-  ],
-  "error": null,
-  "more_actions_needed_by_instruction": true,
-  "log": "The user wants to do click the language switch button, wait 1s, click \"English\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.",
-}
-
-### Example: What NOT to do
-Wrong output:
-{
-  "actions":[
-    {
-      "thought": "Click the language switch button to open the language options.",
-      "type": "Tap",
-      "param": {
-        "locate": { "id": "c81c4e9a33" } // WRONG: prompt is missing, this is not a valid LocateParam
-      }
-    },
-    {
-      "thought": "Click the English option",
-      "type": "Tap", 
-      "param": {
-        "locate": null // WRONG: if the element is not on the page, you should not plan this action
-      }
-    }
-  ],
-  "more_actions_needed_by_instruction": false, // WRONG: should be true
-  "log": "The user wants to do click the language switch button, wait 1s, click \"English\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.",
-}
-`;
-async function systemPromptToTaskPlanning({ actionSpace, vlMode }) {
-    if (vlMode) return systemTemplateOfVLPlanning({
-        actionSpace,
-        vlMode
-    });
-    return `${systemTemplateOfLLM({
-        actionSpace
-    })}\n\n${outputTemplate}`;
 }
 const planSchema = {
     type: 'json_schema',

package/dist/lib/ai-model/prompt/llm-planning.js.map

@@ -1 +1 @@
-{"version":3,"file":"ai-model/prompt/llm-planning.js","sources":["webpack://@midscene/core/webpack/runtime/define_property_getters","webpack://@midscene/core/webpack/runtime/has_own_property","webpack://@midscene/core/webpack/runtime/make_namespace_object","webpack://@midscene/core/./src/ai-model/prompt/llm-planning.ts"],"sourcesContent":["__webpack_require__.d = (exports, definition) => {\n\tfor(var key in definition) {\n        if(__webpack_require__.o(definition, key) && !__webpack_require__.o(exports, key)) {\n            Object.defineProperty(exports, key, { enumerable: true, get: definition[key] });\n        }\n    }\n};","__webpack_require__.o = (obj, prop) => (Object.prototype.hasOwnProperty.call(obj, prop))","// define __esModule on exports\n__webpack_require__.r = (exports) => {\n\tif(typeof Symbol !== 'undefined' && Symbol.toStringTag) {\n\t\tObject.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });\n\t}\n\tObject.defineProperty(exports, '__esModule', { value: true });\n};","import type { DeviceAction } from '@/types';\nimport type { TVlModeTypes } from '@midscene/shared/env';\nimport type { ResponseFormatJSONSchema } from 'openai/resources/index';\nimport type { z } from 'zod';\nimport { ifMidsceneLocatorField } from '../common';\nimport { bboxDescription } from './common';\n\n// Note: put the log field first to trigger the CoT\n\nconst vlCurrentLog = `\"log\": string, // Log your thoughts and what the next one action (ONLY ONE!) you can do according to the screenshot and the instruction. The log should contain the following information: \"The user wants to do ... . According to the instruction and the previous logs, next step is to .... Now i am going to compose an action '{ action-type }' to do ....\". If no action should be done, log the reason. Use the same language as the user's instruction.`;\nconst llmCurrentLog = `\"log\": string, // Log what the next actions you can do according to the screenshot and the instruction. The typical log looks like \"Now i want to use action '{ action-type }' to do ..\". If no action should be done, log the reason. \". Use the same language as the user's instruction.`;\n\nconst commonOutputFields = `\"error\"?: string, // Error messages about unexpected situations, if any. Only think it is an error when the situation is not foreseeable according to the instruction. Use the same language as the user's instruction.\n  \"more_actions_needed_by_instruction\": boolean, // Consider if there is still more action(s) to do after the action in \"Log\" is done, according to the instruction. If so, set this field to true. Otherwise, set it to false.`;\nconst vlLocateParam = () =>\n  '{bbox: [number, number, number, number], prompt: string }';\nconst llmLocateParam = () => '{\"id\": string, \"prompt\": string}';\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 any;\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      // Helper function to get type name from zod schema\n      const getTypeName = (field: any): string => {\n        // Recursively unwrap optional, nullable, and other wrapper types to get the actual inner type\n        const unwrapField = (f: any): any => {\n          if (!f._def) return f;\n\n          const typeName = f._def.typeName;\n\n          // Handle wrapper types that have innerType\n          if (\n            typeName === 'ZodOptional' ||\n            typeName === 'ZodNullable' ||\n            typeName === 'ZodDefault'\n          ) {\n            return unwrapField(f._def.innerType);\n          }\n\n          // Handle ZodEffects (transformations, refinements, preprocessors)\n          if (typeName === 'ZodEffects') {\n            // For ZodEffects, unwrap the schema field which contains the underlying type\n            if (f._def.schema) {\n              return unwrapField(f._def.schema);\n            }\n          }\n\n          return f;\n        };\n\n        const actualField = unwrapField(field);\n        const fieldTypeName = actualField._def?.typeName;\n\n        if (fieldTypeName === 'ZodString') return 'string';\n        if (fieldTypeName === 'ZodNumber') return 'number';\n        if (fieldTypeName === 'ZodBoolean') return 'boolean';\n        if (fieldTypeName === 'ZodArray') return 'array';\n        if (fieldTypeName === 'ZodObject') {\n          // Check if this is a passthrough object (like MidsceneLocation)\n          if (ifMidsceneLocatorField(actualField)) {\n            return locatorSchemaTypeDescription;\n          }\n          return 'object';\n        }\n        if (fieldTypeName === 'ZodEnum') {\n          const values =\n            (actualField._def?.values as unknown[] | undefined)\n              ?.map((option: unknown) => String(`'${option}'`))\n              .join(', ') ?? 'enum';\n\n          return `enum(${values})`;\n        }\n        // Handle ZodUnion by taking the first option (for display purposes)\n        if (fieldTypeName === 'ZodUnion') {\n          const options = actualField._def?.options as any[] | undefined;\n          if (options && options.length > 0) {\n            // For unions, list all types\n            const types = options.map((opt: any) => getTypeName(opt));\n            return types.join(' | ');\n          }\n          return 'union';\n        }\n\n        console.warn(\n          'failed to parse Zod type. This may lead to wrong params from the LLM.\\n',\n          actualField._def,\n        );\n        return actualField.toString();\n      };\n\n      // Helper function to get description from zod schema\n      const getDescription = (field: z.ZodTypeAny): string | null => {\n        // Recursively unwrap optional, nullable, and other wrapper types to get the actual inner type\n        const unwrapField = (f: any): any => {\n          if (!f._def) return f;\n\n          const typeName = f._def.typeName;\n\n          // Handle wrapper types that have innerType\n          if (\n            typeName === 'ZodOptional' ||\n            typeName === 'ZodNullable' ||\n            typeName === 'ZodDefault'\n          ) {\n            return unwrapField(f._def.innerType);\n          }\n\n          // Handle ZodEffects (transformations, refinements, preprocessors)\n          if (typeName === 'ZodEffects') {\n            // For ZodEffects, unwrap the schema field which contains the underlying type\n            if (f._def.schema) {\n              return unwrapField(f._def.schema);\n            }\n          }\n\n          return f;\n        };\n\n        // Check for direct description on the original field (wrapper may have description)\n        if ('description' in field) {\n          return field.description || null;\n        }\n\n        const actualField = unwrapField(field);\n\n        // Check for description on the unwrapped field\n        if ('description' in actualField) {\n          return actualField.description || null;\n        }\n\n        // Check for MidsceneLocation fields and add description\n        if (actualField._def?.typeName === 'ZodObject') {\n          if ('midscene_location_field_flag' in actualField._def.shape()) {\n            return 'Location information for the target element';\n          }\n        }\n\n        return null;\n      };\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 any).isOptional === 'function' &&\n            (field as any).isOptional();\n          const keyWithOptional = isOptional ? `${key}?` : key;\n\n          // Get the type name\n          const typeName = getTypeName(field);\n\n          // Get description\n          const description = getDescription(field as z.ZodTypeAny);\n\n          // Build param line for this field\n          let paramLine = `${keyWithOptional}: ${typeName}`;\n          if (description) {\n            paramLine += ` // ${description}`;\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      // For simple primitive types, the param should be passed directly as the value\n      const schemaTypeName = schema._def?.typeName;\n      let typeName = 'unknown';\n\n      if (schemaTypeName === 'ZodString') typeName = 'string';\n      else if (schemaTypeName === 'ZodNumber') typeName = 'number';\n      else if (schemaTypeName === 'ZodBoolean') typeName = 'boolean';\n\n      // Get description if available\n      const description = 'description' in schema ? schema.description : null;\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\nconst systemTemplateOfVLPlanning = ({\n  actionSpace,\n  vlMode,\n}: {\n  actionSpace: DeviceAction<any>[];\n  vlMode: TVlModeTypes | undefined;\n}) => {\n  const actionNameList = actionSpace.map((action) => action.name).join(', ');\n  const actionDescriptionList = actionSpace.map((action) => {\n    return descriptionForAction(action, vlLocateParam());\n  });\n  const actionList = actionDescriptionList.join('\\n');\n\n  return `\nTarget: User will give you a latest screenshot, an instruction and some previous logs indicating what have been done. Please tell what the next one action is (or null if no action should be done) to do the tasks the instruction requires. \n\nRestriction:\n- Don't give extra actions or plans beyond the instruction. ONLY plan for what the instruction requires. For example, don't try to submit the form if the instruction is only to fill something.\n- Always give ONLY ONE action in \\`log\\` field (or null if no action should be done), instead of multiple actions. Supported actions are ${actionNameList}.\n- Don't repeat actions in the previous logs.\n- Bbox is the bounding box of the element to be located. It's an array of 4 numbers, representing ${bboxDescription(vlMode)}.\n\nSupporting actions:\n${actionList}\n\nField description:\n* The \\`prompt\\` field inside the \\`locate\\` field is a short description that could be used to locate the element.\n\nReturn in JSON format:\n{\n  ${vlCurrentLog}\n  ${commonOutputFields}\n  \"action\": \n    {\n      // one of the supporting actions\n    } | null,\n  ,\n  \"sleep\"?: number, // The sleep time after the action, in milliseconds.\n}\n\nFor example, when the instruction is \"click 'Confirm' button, and click 'Yes' in popup\" and the previous log shows \"The 'Confirm' button has been clicked\", by viewing the screenshot and previous logs, you should consider: We have already clicked the 'Confirm' button, so next we should find and click 'Yes' in popup.\n\nthis and output the JSON:\n\n{\n  \"log\": \"The user wants to do click 'Confirm' button, and click 'Yes' in popup. According to the instruction and the previous logs, next step is to tap the 'Yes' button in the popup. Now i am going to compose an action 'Tap' to click 'Yes' in popup.\",\n  \"action\": {\n    \"type\": \"Tap\",\n    \"param\": {\n      \"locate\": {\n        \"bbox\": [100, 100, 200, 200],\n        \"prompt\": \"The 'Yes' button in popup\"\n      }\n    }\n  },\n  \"more_actions_needed_by_instruction\": false,\n}\n`;\n};\n\nconst systemTemplateOfLLM = ({\n  actionSpace,\n}: { actionSpace: DeviceAction<any>[] }) => {\n  const actionNameList = actionSpace.map((action) => action.name).join(' / ');\n  const actionDescriptionList = actionSpace.map((action) => {\n    return descriptionForAction(action, llmLocateParam());\n  });\n  const actionList = actionDescriptionList.join('\\n');\n\n  return `\n## Role\n\nYou are a versatile professional in software UI automation. Your outstanding contributions will impact the user experience of billions of users.\n\n## Objective\n\n- Decompose the instruction user asked into a series of actions\n- Locate the target element if possible\n- If the instruction cannot be accomplished, give a further plan.\n\n## Workflow\n\n1. Receive the screenshot, element description of screenshot(if any), user's instruction and previous logs.\n2. Decompose the user's task into a sequence of feasible actions, and place it in the \\`actions\\` field. There are different types of actions (${actionNameList}). The \"About the action\" section below will give you more details.\n3. Consider whether the user's instruction will be accomplished after the actions you composed.\n- If the instruction is accomplished, set \\`more_actions_needed_by_instruction\\` to false.\n- If more actions are needed, set \\`more_actions_needed_by_instruction\\` to true. Get ready to hand over to the next talent people like you. Carefully log what have been done in the \\`log\\` field, he or she will continue the task according to your logs.\n4. If the task is not feasible on this page, set \\`error\\` field to the reason.\n\n## Constraints\n\n- All the actions you composed MUST be feasible, which means all the action fields can be filled with the page context information you get. If not, don't plan this action.\n- Trust the \"What have been done\" field about the task (if any), don't repeat actions in it.\n- Respond only with valid JSON. Do not write an introduction or summary or markdown prefix like \\`\\`\\`json\\`\\`\\`.\n- If the screenshot and the instruction are totally irrelevant, set reason in the \\`error\\` field.\n\n## About the \\`actions\\` field\n\nThe \\`locate\\` param is commonly used in the \\`param\\` field of the action, means to locate the target element to perform the action, it conforms to the following scheme:\n\ntype LocateParam = {\n  \"id\": string, // the id of the element found. It should either be the id marked with a rectangle in the screenshot or the id described in the description.\n  \"prompt\"?: string // the description of the element to find. It can only be omitted when locate is null.\n} | null // If it's not on the page, the LocateParam should be null\n\n## Supported actions\n\nEach action has a \\`type\\` and corresponding \\`param\\`. To be detailed:\n${actionList}\n\n`.trim();\n};\n\nconst outputTemplate = `\n## Output JSON Format:\n\nThe JSON format is as follows:\n\n{\n  \"actions\": [\n    // ... some actions\n  ],\n  ${llmCurrentLog}\n  ${commonOutputFields}\n}\n\n## Examples\n\n### Example: Decompose a task\n\nWhen you received the following information:\n\n* Instruction: 'Click the language switch button, wait 1s, click \"English\"'\n* Logs: null\n* Page Context (screenshot and description) shows: There is a language switch button, and the \"English\" option is not shown in the screenshot now.\n\nBy viewing the page screenshot and description, you should consider this and output the JSON:\n\n* The user intent is: tap the switch button, sleep, and tap the 'English' option\n* The language switch button is shown in the screenshot, and can be located by the page description or the id marked with a rectangle. So we can plan a Tap action to do this.\n* Plan a Sleep action to wait for 1 second to ensure the language options are displayed.\n* The \"English\" option button is not shown in the screenshot now, it means it may only show after the previous actions are finished. So don't plan any action to do this.\n* Compose the log: The user wants to do click the language switch button, wait 1s, click \"English\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.\n* The task cannot be accomplished (because the last tapping action is not finished yet), so the \\`more_actions_needed_by_instruction\\` field is true. The \\`error\\` field is null.\n\n{\n  \"actions\":[\n    {\n      \"thought\": \"Click the language switch button to open the language options.\",\n      \"type\": \"Tap\", \n      \"param\": {\n        \"locate\": { id: \"c81c4e9a33\", prompt: \"The language switch button\" }\n      }\n    },\n    {\n      \"thought\": \"Wait for 1 second to ensure the language options are displayed.\",\n      \"type\": \"Sleep\",\n      \"param\": { \"timeMs\": 1000 },\n    }\n  ],\n  \"error\": null,\n  \"more_actions_needed_by_instruction\": true,\n  \"log\": \"The user wants to do click the language switch button, wait 1s, click \\\"English\\\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.\",\n}\n\n### Example: What NOT to do\nWrong output:\n{\n  \"actions\":[\n    {\n      \"thought\": \"Click the language switch button to open the language options.\",\n      \"type\": \"Tap\",\n      \"param\": {\n        \"locate\": { \"id\": \"c81c4e9a33\" } // WRONG: prompt is missing, this is not a valid LocateParam\n      }\n    },\n    {\n      \"thought\": \"Click the English option\",\n      \"type\": \"Tap\", \n      \"param\": {\n        \"locate\": null // WRONG: if the element is not on the page, you should not plan this action\n      }\n    }\n  ],\n  \"more_actions_needed_by_instruction\": false, // WRONG: should be true\n  \"log\": \"The user wants to do click the language switch button, wait 1s, click \\\"English\\\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.\",\n}\n`;\n\nexport async function systemPromptToTaskPlanning({\n  actionSpace,\n  vlMode,\n}: {\n  actionSpace: DeviceAction<any>[];\n  vlMode: TVlModeTypes | undefined;\n}) {\n  if (vlMode) {\n    return systemTemplateOfVLPlanning({ actionSpace, vlMode });\n  }\n\n  return `${systemTemplateOfLLM({ actionSpace })}\\n\\n${outputTemplate}`;\n}\n\nexport const planSchema: ResponseFormatJSONSchema = {\n  type: 'json_schema',\n  json_schema: {\n    name: 'action_items',\n    strict: false,\n    schema: {\n      type: 'object',\n      strict: false,\n      properties: {\n        actions: {\n          type: 'array',\n          items: {\n            type: 'object',\n            strict: false,\n            properties: {\n              thought: {\n                type: 'string',\n                description:\n                  'Reasons for generating this task, and why this task is feasible on this page',\n              },\n              type: {\n                type: 'string',\n                description: 'Type of action',\n              },\n              param: {\n                anyOf: [\n                  { type: 'null' },\n                  {\n                    type: 'object',\n                    additionalProperties: true,\n                  },\n                ],\n                description: 'Parameter of the action',\n              },\n              locate: {\n                type: ['object', 'null'],\n                properties: {\n                  id: { type: 'string' },\n                  prompt: { type: 'string' },\n                },\n                required: ['id', 'prompt'],\n                additionalProperties: false,\n                description: 'Location information for the target element',\n              },\n            },\n            required: ['thought', 'type', 'param', 'locate'],\n            additionalProperties: false,\n          },\n          description: 'List of actions to be performed',\n        },\n        more_actions_needed_by_instruction: {\n          type: 'boolean',\n          description:\n            'If all the actions described in the instruction have been covered by this action and logs, set this field to false.',\n        },\n        log: {\n          type: 'string',\n          description:\n            'Log what these planned actions do. Do not include further actions that have not been planned.',\n        },\n        error: {\n          type: ['string', 'null'],\n          description: 'Error messages about unexpected situations',\n        },\n      },\n      required: [\n        'actions',\n        'more_actions_needed_by_instruction',\n        'log',\n        'error',\n      ],\n      additionalProperties: false,\n    },\n  },\n};\n"],"names":["__webpack_require__","definition","key","Object","obj","prop","Symbol","vlCurrentLog","llmCurrentLog","commonOutputFields","vlLocateParam","llmLocateParam","descriptionForAction","action","locatorSchemaTypeDescription","tab","fields","_schema__def","paramLines","schema","isZodObject","shape","getTypeName","field","_actualField__def","unwrapField","f","typeName","actualField","fieldTypeName","ifMidsceneLocatorField","_actualField__def_values","values","option","String","_actualField__def2","options","types","opt","console","getDescription","isOptional","keyWithOptional","description","paramLine","line","_schema__def1","schemaTypeName","paramDescription","systemTemplateOfVLPlanning","actionSpace","vlMode","actionNameList","actionDescriptionList","actionList","bboxDescription","systemTemplateOfLLM","outputTemplate","systemPromptToTaskPlanning","planSchema"],"mappings":";;;IAAAA,oBAAoB,CAAC,GAAG,CAAC,UAASC;QACjC,IAAI,IAAIC,OAAOD,WACR,IAAGD,oBAAoB,CAAC,CAACC,YAAYC,QAAQ,CAACF,oBAAoB,CAAC,CAAC,UAASE,MACzEC,OAAO,cAAc,CAAC,UAASD,KAAK;YAAE,YAAY;YAAM,KAAKD,UAAU,CAACC,IAAI;QAAC;IAGzF;;;ICNAF,oBAAoB,CAAC,GAAG,CAACI,KAAKC,OAAUF,OAAO,SAAS,CAAC,cAAc,CAAC,IAAI,CAACC,KAAKC;;;ICClFL,oBAAoB,CAAC,GAAG,CAAC;QACxB,IAAG,AAAkB,eAAlB,OAAOM,UAA0BA,OAAO,WAAW,EACrDH,OAAO,cAAc,CAAC,UAASG,OAAO,WAAW,EAAE;YAAE,OAAO;QAAS;QAEtEH,OAAO,cAAc,CAAC,UAAS,cAAc;YAAE,OAAO;QAAK;IAC5D;;;;;;;;;;;ACGA,MAAMI,eAAe;AACrB,MAAMC,gBAAgB;AAEtB,MAAMC,qBAAqB,CAAC;+NACmM,CAAC;AAChO,MAAMC,gBAAgB,IACpB;AACF,MAAMC,iBAAiB,IAAM;AAEtB,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;YAKFI;QAJpB,MAAMC,aAAuB,EAAE;QAG/B,MAAMC,SAASN,OAAO,WAAW;QACjC,MAAMO,cAAcH,AAAAA,SAAAA,CAAAA,eAAAA,OAAO,IAAI,AAAD,IAAVA,KAAAA,IAAAA,aAAa,QAAQ,AAAD,MAAM;QAE9C,IAAIG,eAAeD,OAAO,KAAK,EAAE;YAE/B,MAAME,QAAQF,OAAO,KAAK;YAG1B,MAAMG,cAAc,CAACC;oBA4BGC;gBA1BtB,MAAMC,cAAc,CAACC;oBACnB,IAAI,CAACA,EAAE,IAAI,EAAE,OAAOA;oBAEpB,MAAMC,WAAWD,EAAE,IAAI,CAAC,QAAQ;oBAGhC,IACEC,AAAa,kBAAbA,YACAA,AAAa,kBAAbA,YACAA,AAAa,iBAAbA,UAEA,OAAOF,YAAYC,EAAE,IAAI,CAAC,SAAS;oBAIrC,IAAIC,AAAa,iBAAbA,UAEF;wBAAA,IAAID,EAAE,IAAI,CAAC,MAAM,EACf,OAAOD,YAAYC,EAAE,IAAI,CAAC,MAAM;oBAClC;oBAGF,OAAOA;gBACT;gBAEA,MAAME,cAAcH,YAAYF;gBAChC,MAAMM,gBAAgB,QAAAL,CAAAA,oBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,kBAAkB,QAAQ;gBAEhD,IAAIK,AAAkB,gBAAlBA,eAA+B,OAAO;gBAC1C,IAAIA,AAAkB,gBAAlBA,eAA+B,OAAO;gBAC1C,IAAIA,AAAkB,iBAAlBA,eAAgC,OAAO;gBAC3C,IAAIA,AAAkB,eAAlBA,eAA8B,OAAO;gBACzC,IAAIA,AAAkB,gBAAlBA,eAA+B;oBAEjC,IAAIC,AAAAA,IAAAA,mCAAAA,sBAAAA,AAAAA,EAAuBF,cACzB,OAAOd;oBAET,OAAO;gBACT;gBACA,IAAIe,AAAkB,cAAlBA,eAA6B;wBAE5BE,0BAAAA;oBADH,MAAMC,SACJ,AAAC,SAAAD,CAAAA,qBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,QAAAA,CAAAA,2BAAAA,mBAAkB,MAAM,AAAD,IAAvBA,KAAAA,IAAAA,yBACG,GAAG,CAAC,CAACE,SAAoBC,OAAO,CAAC,CAAC,EAAED,OAAO,CAAC,CAAC,GAC9C,IAAI,CAAC,KAAI,KAAK;oBAEnB,OAAO,CAAC,KAAK,EAAED,OAAO,CAAC,CAAC;gBAC1B;gBAEA,IAAIH,AAAkB,eAAlBA,eAA8B;wBAChBM;oBAAhB,MAAMC,UAAU,QAAAD,CAAAA,qBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,mBAAkB,OAAO;oBACzC,IAAIC,WAAWA,QAAQ,MAAM,GAAG,GAAG;wBAEjC,MAAMC,QAAQD,QAAQ,GAAG,CAAC,CAACE,MAAahB,YAAYgB;wBACpD,OAAOD,MAAM,IAAI,CAAC;oBACpB;oBACA,OAAO;gBACT;gBAEAE,QAAQ,IAAI,CACV,2EACAX,YAAY,IAAI;gBAElB,OAAOA,YAAY,QAAQ;YAC7B;YAGA,MAAMY,iBAAiB,CAACjB;oBAwClBC;gBAtCJ,MAAMC,cAAc,CAACC;oBACnB,IAAI,CAACA,EAAE,IAAI,EAAE,OAAOA;oBAEpB,MAAMC,WAAWD,EAAE,IAAI,CAAC,QAAQ;oBAGhC,IACEC,AAAa,kBAAbA,YACAA,AAAa,kBAAbA,YACAA,AAAa,iBAAbA,UAEA,OAAOF,YAAYC,EAAE,IAAI,CAAC,SAAS;oBAIrC,IAAIC,AAAa,iBAAbA,UAEF;wBAAA,IAAID,EAAE,IAAI,CAAC,MAAM,EACf,OAAOD,YAAYC,EAAE,IAAI,CAAC,MAAM;oBAClC;oBAGF,OAAOA;gBACT;gBAGA,IAAI,iBAAiBH,OACnB,OAAOA,MAAM,WAAW,IAAI;gBAG9B,MAAMK,cAAcH,YAAYF;gBAGhC,IAAI,iBAAiBK,aACnB,OAAOA,YAAY,WAAW,IAAI;gBAIpC,IAAIJ,AAAAA,SAAAA,CAAAA,oBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,kBAAkB,QAAQ,AAAD,MAAM,aACjC;oBAAA,IAAI,kCAAkCI,YAAY,IAAI,CAAC,KAAK,IAC1D,OAAO;gBACT;gBAGF,OAAO;YACT;YAEA,KAAK,MAAM,CAAC1B,KAAKqB,MAAM,IAAIpB,OAAO,OAAO,CAACkB,OACxC,IAAIE,SAAS,AAAiB,YAAjB,OAAOA,OAAoB;gBAEtC,MAAMkB,aACJ,AAAqC,cAArC,OAAQlB,MAAc,UAAU,IAC/BA,MAAc,UAAU;gBAC3B,MAAMmB,kBAAkBD,aAAa,GAAGvC,IAAI,CAAC,CAAC,GAAGA;gBAGjD,MAAMyB,WAAWL,YAAYC;gBAG7B,MAAMoB,cAAcH,eAAejB;gBAGnC,IAAIqB,YAAY,GAAGF,gBAAgB,EAAE,EAAEf,UAAU;gBACjD,IAAIgB,aACFC,aAAa,CAAC,IAAI,EAAED,aAAa;gBAGnCzB,WAAW,IAAI,CAAC0B;YAClB;YAIF,IAAI1B,WAAW,MAAM,GAAG,GAAG;gBACzBF,OAAO,IAAI,CAAC;gBACZE,WAAW,OAAO,CAAC,CAAC2B;oBAClB7B,OAAO,IAAI,CAAC,CAAC,IAAI,EAAE6B,MAAM;gBAC3B;YACF;QACF,OAAO;gBAGkBC;YAAvB,MAAMC,iBAAiB,QAAAD,CAAAA,gBAAAA,OAAO,IAAI,AAAD,IAAVA,KAAAA,IAAAA,cAAa,QAAQ;YAC5C,IAAInB,WAAW;YAEf,IAAIoB,AAAmB,gBAAnBA,gBAAgCpB,WAAW;iBAC1C,IAAIoB,AAAmB,gBAAnBA,gBAAgCpB,WAAW;iBAC/C,IAAIoB,AAAmB,iBAAnBA,gBAAiCpB,WAAW;YAGrD,MAAMgB,cAAc,iBAAiBxB,SAASA,OAAO,WAAW,GAAG;YAGnE,IAAI6B,mBAAmB,CAAC,SAAS,EAAErB,UAAU;YAC7C,IAAIgB,aACFK,oBAAoB,CAAC,IAAI,EAAEL,aAAa;YAE1CK,oBAAoB;YAEpBhC,OAAO,IAAI,CAACgC;QACd;IACF;IAEA,OAAO,CAAC,EAAE,EAAEnC,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;AAEA,MAAMkC,6BAA6B,CAAC,EAClCC,WAAW,EACXC,MAAM,EAIP;IACC,MAAMC,iBAAiBF,YAAY,GAAG,CAAC,CAACrC,SAAWA,OAAO,IAAI,EAAE,IAAI,CAAC;IACrE,MAAMwC,wBAAwBH,YAAY,GAAG,CAAC,CAACrC,SACtCD,qBAAqBC,QAAQH;IAEtC,MAAM4C,aAAaD,sBAAsB,IAAI,CAAC;IAE9C,OAAO,CAAC;;;;;yIAK+H,EAAED,eAAe;;kGAExD,EAAEG,AAAAA,IAAAA,qCAAAA,eAAAA,AAAAA,EAAgBJ,QAAQ;;;AAG5H,EAAEG,WAAW;;;;;;;EAOX,EAAE/C,aAAa;EACf,EAAEE,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;AA0BvB,CAAC;AACD;AAEA,MAAM+C,sBAAsB,CAAC,EAC3BN,WAAW,EAC0B;IACrC,MAAME,iBAAiBF,YAAY,GAAG,CAAC,CAACrC,SAAWA,OAAO,IAAI,EAAE,IAAI,CAAC;IACrE,MAAMwC,wBAAwBH,YAAY,GAAG,CAAC,CAACrC,SACtCD,qBAAqBC,QAAQF;IAEtC,MAAM2C,aAAaD,sBAAsB,IAAI,CAAC;IAE9C,OAAO,CAAC;;;;;;;;;;;;;;+IAcqI,EAAED,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;AAyBhK,EAAEE,WAAW;;AAEb,CAAC,CAAC,IAAI;AACN;AAEA,MAAMG,iBAAiB,CAAC;;;;;;;;;EAStB,EAAEjD,cAAc;EAChB,EAAEC,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgEvB,CAAC;AAEM,eAAeiD,2BAA2B,EAC/CR,WAAW,EACXC,MAAM,EAIP;IACC,IAAIA,QACF,OAAOF,2BAA2B;QAAEC;QAAaC;IAAO;IAG1D,OAAO,GAAGK,oBAAoB;QAAEN;IAAY,GAAG,IAAI,EAAEO,gBAAgB;AACvE;AAEO,MAAME,aAAuC;IAClD,MAAM;IACN,aAAa;QACX,MAAM;QACN,QAAQ;QACR,QAAQ;YACN,MAAM;YACN,QAAQ;YACR,YAAY;gBACV,SAAS;oBACP,MAAM;oBACN,OAAO;wBACL,MAAM;wBACN,QAAQ;wBACR,YAAY;4BACV,SAAS;gCACP,MAAM;gCACN,aACE;4BACJ;4BACA,MAAM;gCACJ,MAAM;gCACN,aAAa;4BACf;4BACA,OAAO;gCACL,OAAO;oCACL;wCAAE,MAAM;oCAAO;oCACf;wCACE,MAAM;wCACN,sBAAsB;oCACxB;iCACD;gCACD,aAAa;4BACf;4BACA,QAAQ;gCACN,MAAM;oCAAC;oCAAU;iCAAO;gCACxB,YAAY;oCACV,IAAI;wCAAE,MAAM;oCAAS;oCACrB,QAAQ;wCAAE,MAAM;oCAAS;gCAC3B;gCACA,UAAU;oCAAC;oCAAM;iCAAS;gCAC1B,sBAAsB;gCACtB,aAAa;4BACf;wBACF;wBACA,UAAU;4BAAC;4BAAW;4BAAQ;4BAAS;yBAAS;wBAChD,sBAAsB;oBACxB;oBACA,aAAa;gBACf;gBACA,oCAAoC;oBAClC,MAAM;oBACN,aACE;gBACJ;gBACA,KAAK;oBACH,MAAM;oBACN,aACE;gBACJ;gBACA,OAAO;oBACL,MAAM;wBAAC;wBAAU;qBAAO;oBACxB,aAAa;gBACf;YACF;YACA,UAAU;gBACR;gBACA;gBACA;gBACA;aACD;YACD,sBAAsB;QACxB;IACF;AACF"}
+{"version":3,"file":"ai-model/prompt/llm-planning.js","sources":["webpack://@midscene/core/webpack/runtime/define_property_getters","webpack://@midscene/core/webpack/runtime/has_own_property","webpack://@midscene/core/webpack/runtime/make_namespace_object","webpack://@midscene/core/./src/ai-model/prompt/llm-planning.ts"],"sourcesContent":["__webpack_require__.d = (exports, definition) => {\n\tfor(var key in definition) {\n        if(__webpack_require__.o(definition, key) && !__webpack_require__.o(exports, key)) {\n            Object.defineProperty(exports, key, { enumerable: true, get: definition[key] });\n        }\n    }\n};","__webpack_require__.o = (obj, prop) => (Object.prototype.hasOwnProperty.call(obj, prop))","// define __esModule on exports\n__webpack_require__.r = (exports) => {\n\tif(typeof Symbol !== 'undefined' && Symbol.toStringTag) {\n\t\tObject.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });\n\t}\n\tObject.defineProperty(exports, '__esModule', { value: true });\n};","import type { DeviceAction } from '@/types';\nimport type { TVlModeTypes } from '@midscene/shared/env';\nimport type { ResponseFormatJSONSchema } from 'openai/resources/index';\nimport type { z } from 'zod';\nimport { ifMidsceneLocatorField } from '../../common';\nimport { bboxDescription } from './common';\n\n// Note: put the log field first to trigger the CoT\n\nconst commonOutputFields = `\"error\"?: string, // Error messages about unexpected situations, if any. Only think it is an error when the situation is not foreseeable according to the instruction. Use the same language as the user's instruction.\n  \"more_actions_needed_by_instruction\": boolean, // Consider if there is still more action(s) to do after the action in \"Log\" is done, according to the instruction. If so, set this field to true. Otherwise, set it to false.`;\n\nconst vlLocateParam = (vlMode: TVlModeTypes | undefined) => {\n  if (vlMode) {\n    return `{bbox: [number, number, number, number], prompt: string } // ${bboxDescription(vlMode)}`;\n  }\n  return '{ prompt: string /* description of the target element */ }';\n};\n\nexport const descriptionForAction = (\n  action: DeviceAction<any>,\n  locatorSchemaTypeDescription: string,\n) => {\n  const tab = '  ';\n  const fields: string[] = [];\n\n  // Add the action type field\n  fields.push(`- type: \"${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 any;\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      // Helper function to get type name from zod schema\n      const getTypeName = (field: any): string => {\n        // Recursively unwrap optional, nullable, and other wrapper types to get the actual inner type\n        const unwrapField = (f: any): any => {\n          if (!f._def) return f;\n\n          const typeName = f._def.typeName;\n\n          // Handle wrapper types that have innerType\n          if (\n            typeName === 'ZodOptional' ||\n            typeName === 'ZodNullable' ||\n            typeName === 'ZodDefault'\n          ) {\n            return unwrapField(f._def.innerType);\n          }\n\n          // Handle ZodEffects (transformations, refinements, preprocessors)\n          if (typeName === 'ZodEffects') {\n            // For ZodEffects, unwrap the schema field which contains the underlying type\n            if (f._def.schema) {\n              return unwrapField(f._def.schema);\n            }\n          }\n\n          return f;\n        };\n\n        const actualField = unwrapField(field);\n        const fieldTypeName = actualField._def?.typeName;\n\n        if (fieldTypeName === 'ZodString') return 'string';\n        if (fieldTypeName === 'ZodNumber') return 'number';\n        if (fieldTypeName === 'ZodBoolean') return 'boolean';\n        if (fieldTypeName === 'ZodArray') return 'array';\n        if (fieldTypeName === 'ZodObject') {\n          // Check if this is a passthrough object (like MidsceneLocation)\n          if (ifMidsceneLocatorField(actualField)) {\n            return locatorSchemaTypeDescription;\n          }\n          return 'object';\n        }\n        if (fieldTypeName === 'ZodEnum') {\n          const values =\n            (actualField._def?.values as unknown[] | undefined)\n              ?.map((option: unknown) => String(`'${option}'`))\n              .join(', ') ?? 'enum';\n\n          return `enum(${values})`;\n        }\n        // Handle ZodUnion by taking the first option (for display purposes)\n        if (fieldTypeName === 'ZodUnion') {\n          const options = actualField._def?.options as any[] | undefined;\n          if (options && options.length > 0) {\n            // For unions, list all types\n            const types = options.map((opt: any) => getTypeName(opt));\n            return types.join(' | ');\n          }\n          return 'union';\n        }\n\n        console.warn(\n          'failed to parse Zod type. This may lead to wrong params from the LLM.\\n',\n          actualField._def,\n        );\n        return actualField.toString();\n      };\n\n      // Helper function to get description from zod schema\n      const getDescription = (field: z.ZodTypeAny): string | null => {\n        // Recursively unwrap optional, nullable, and other wrapper types to get the actual inner type\n        const unwrapField = (f: any): any => {\n          if (!f._def) return f;\n\n          const typeName = f._def.typeName;\n\n          // Handle wrapper types that have innerType\n          if (\n            typeName === 'ZodOptional' ||\n            typeName === 'ZodNullable' ||\n            typeName === 'ZodDefault'\n          ) {\n            return unwrapField(f._def.innerType);\n          }\n\n          // Handle ZodEffects (transformations, refinements, preprocessors)\n          if (typeName === 'ZodEffects') {\n            // For ZodEffects, unwrap the schema field which contains the underlying type\n            if (f._def.schema) {\n              return unwrapField(f._def.schema);\n            }\n          }\n\n          return f;\n        };\n\n        // Check for direct description on the original field (wrapper may have description)\n        if ('description' in field) {\n          return field.description || null;\n        }\n\n        const actualField = unwrapField(field);\n\n        // Check for description on the unwrapped field\n        if ('description' in actualField) {\n          return actualField.description || null;\n        }\n\n        // Check for MidsceneLocation fields and add description\n        if (actualField._def?.typeName === 'ZodObject') {\n          if ('midscene_location_field_flag' in actualField._def.shape()) {\n            return 'Location information for the target element';\n          }\n        }\n\n        return null;\n      };\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 any).isOptional === 'function' &&\n            (field as any).isOptional();\n          const keyWithOptional = isOptional ? `${key}?` : key;\n\n          // Get the type name\n          const typeName = getTypeName(field);\n\n          // Get description\n          const description = getDescription(field as z.ZodTypeAny);\n\n          // Build param line for this field\n          let paramLine = `${keyWithOptional}: ${typeName}`;\n          if (description) {\n            paramLine += ` // ${description}`;\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      // For simple primitive types, the param should be passed directly as the value\n      const schemaTypeName = schema._def?.typeName;\n      let typeName = 'unknown';\n\n      if (schemaTypeName === 'ZodString') typeName = 'string';\n      else if (schemaTypeName === 'ZodNumber') typeName = 'number';\n      else if (schemaTypeName === 'ZodBoolean') typeName = 'boolean';\n\n      // Get description if available\n      const description = 'description' in schema ? schema.description : null;\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  vlMode,\n  includeBbox,\n}: {\n  actionSpace: DeviceAction<any>[];\n  vlMode: TVlModeTypes | undefined;\n  includeBbox: boolean;\n}) {\n  // Validate parameters: if includeBbox is true, vlMode must be defined\n  if (includeBbox && !vlMode) {\n    throw new Error(\n      'vlMode cannot be undefined when includeBbox is true. A valid vlMode is required for bbox-based location.',\n    );\n  }\n\n  const actionDescriptionList = actionSpace.map((action) => {\n    return descriptionForAction(\n      action,\n      vlLocateParam(includeBbox ? vlMode : undefined),\n    );\n  });\n  const actionList = actionDescriptionList.join('\\n');\n\n  return `\nTarget: User will give you an instruction, some screenshots and previous logs indicating what have been done. Your task is to plan the next one action to accomplish the instruction.\n\nPlease tell what the next one action is (or null if no action should be done) to do the tasks the instruction requires. \n\nRestriction:\n- Don't give extra actions or plans beyond the instruction. ONLY plan for what the instruction requires. For example, don't try to submit the form if the instruction is only to fill something.\n- Give just the next ONE action you should do\n\nSupporting actions:\n${actionList}\n\nReturn in JSON format:\n{\n  \"log\": string, // Log your thoughts and what the next one action (ONLY ONE!) you can do according to the screenshot and the instruction. The log should contain the following information: \"The user wants to do ... . According to the instruction and the previous logs, next step is to .... Now i am going to compose an action '{ action-type }' to do this\". If no action should be done, log the reason. Use the same language as the user's instruction.\n  ${commonOutputFields}\n  \"action\": \n    {\n      // one of the supporting actions\n    } | null,\n  ,\n  \"sleep\"?: number, // The sleep time after the action, in milliseconds.\n}\n\nFor example, when the instruction is \"click 'Confirm' button, and click 'Yes' in popup\" and the previous log shows \"The 'Confirm' button has been clicked\", by viewing the screenshot and previous logs, you should consider: We have already clicked the 'Confirm' button, so next we should find and click 'Yes' in popup.\n\nthis and output the JSON:\n\n{\n  \"log\": \"The user wants to do click 'Confirm' button, and click 'Yes' in popup. According to the instruction and the previous logs, next step is to tap the 'Yes' button in the popup. Now i am going to compose an action 'Tap' to click 'Yes' in popup.\",\n  \"action\": {\n    \"type\": \"Tap\",\n    \"param\": {\n      \"locate\": {\n        ${vlMode ? `\"bbox\": [100, 100, 200, 200],` : ''}\n        \"prompt\": \"The 'Yes' button in popup\"\n      }\n    }\n  },\n  \"more_actions_needed_by_instruction\": false,\n}\n`;\n}\n\nexport const planSchema: ResponseFormatJSONSchema = {\n  type: 'json_schema',\n  json_schema: {\n    name: 'action_items',\n    strict: false,\n    schema: {\n      type: 'object',\n      strict: false,\n      properties: {\n        actions: {\n          type: 'array',\n          items: {\n            type: 'object',\n            strict: false,\n            properties: {\n              thought: {\n                type: 'string',\n                description:\n                  'Reasons for generating this task, and why this task is feasible on this page',\n              },\n              type: {\n                type: 'string',\n                description: 'Type of action',\n              },\n              param: {\n                anyOf: [\n                  { type: 'null' },\n                  {\n                    type: 'object',\n                    additionalProperties: true,\n                  },\n                ],\n                description: 'Parameter of the action',\n              },\n              locate: {\n                type: ['object', 'null'],\n                properties: {\n                  id: { type: 'string' },\n                  prompt: { type: 'string' },\n                },\n                required: ['id', 'prompt'],\n                additionalProperties: false,\n                description: 'Location information for the target element',\n              },\n            },\n            required: ['thought', 'type', 'param', 'locate'],\n            additionalProperties: false,\n          },\n          description: 'List of actions to be performed',\n        },\n        more_actions_needed_by_instruction: {\n          type: 'boolean',\n          description:\n            'If all the actions described in the instruction have been covered by this action and logs, set this field to false.',\n        },\n        log: {\n          type: 'string',\n          description:\n            'Log what these planned actions do. Do not include further actions that have not been planned.',\n        },\n        error: {\n          type: ['string', 'null'],\n          description: 'Error messages about unexpected situations',\n        },\n      },\n      required: [\n        'actions',\n        'more_actions_needed_by_instruction',\n        'log',\n        'error',\n      ],\n      additionalProperties: false,\n    },\n  },\n};\n"],"names":["__webpack_require__","definition","key","Object","obj","prop","Symbol","commonOutputFields","vlLocateParam","vlMode","bboxDescription","descriptionForAction","action","locatorSchemaTypeDescription","tab","fields","_schema__def","paramLines","schema","isZodObject","shape","getTypeName","field","_actualField__def","unwrapField","f","typeName","actualField","fieldTypeName","ifMidsceneLocatorField","_actualField__def_values","values","option","String","_actualField__def2","options","types","opt","console","getDescription","isOptional","keyWithOptional","description","paramLine","line","_schema__def1","schemaTypeName","paramDescription","systemPromptToTaskPlanning","actionSpace","includeBbox","Error","actionDescriptionList","undefined","actionList","planSchema"],"mappings":";;;IAAAA,oBAAoB,CAAC,GAAG,CAAC,UAASC;QACjC,IAAI,IAAIC,OAAOD,WACR,IAAGD,oBAAoB,CAAC,CAACC,YAAYC,QAAQ,CAACF,oBAAoB,CAAC,CAAC,UAASE,MACzEC,OAAO,cAAc,CAAC,UAASD,KAAK;YAAE,YAAY;YAAM,KAAKD,UAAU,CAACC,IAAI;QAAC;IAGzF;;;ICNAF,oBAAoB,CAAC,GAAG,CAACI,KAAKC,OAAUF,OAAO,SAAS,CAAC,cAAc,CAAC,IAAI,CAACC,KAAKC;;;ICClFL,oBAAoB,CAAC,GAAG,CAAC;QACxB,IAAG,AAAkB,eAAlB,OAAOM,UAA0BA,OAAO,WAAW,EACrDH,OAAO,cAAc,CAAC,UAASG,OAAO,WAAW,EAAE;YAAE,OAAO;QAAS;QAEtEH,OAAO,cAAc,CAAC,UAAS,cAAc;YAAE,OAAO;QAAK;IAC5D;;;;;;;;;;;ACGA,MAAMI,qBAAqB,CAAC;+NACmM,CAAC;AAEhO,MAAMC,gBAAgB,CAACC;IACrB,IAAIA,QACF,OAAO,CAAC,6DAA6D,EAAEC,AAAAA,IAAAA,qCAAAA,eAAAA,AAAAA,EAAgBD,SAAS;IAElG,OAAO;AACT;AAEO,MAAME,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;YAKFI;QAJpB,MAAMC,aAAuB,EAAE;QAG/B,MAAMC,SAASN,OAAO,WAAW;QACjC,MAAMO,cAAcH,AAAAA,SAAAA,CAAAA,eAAAA,OAAO,IAAI,AAAD,IAAVA,KAAAA,IAAAA,aAAa,QAAQ,AAAD,MAAM;QAE9C,IAAIG,eAAeD,OAAO,KAAK,EAAE;YAE/B,MAAME,QAAQF,OAAO,KAAK;YAG1B,MAAMG,cAAc,CAACC;oBA4BGC;gBA1BtB,MAAMC,cAAc,CAACC;oBACnB,IAAI,CAACA,EAAE,IAAI,EAAE,OAAOA;oBAEpB,MAAMC,WAAWD,EAAE,IAAI,CAAC,QAAQ;oBAGhC,IACEC,AAAa,kBAAbA,YACAA,AAAa,kBAAbA,YACAA,AAAa,iBAAbA,UAEA,OAAOF,YAAYC,EAAE,IAAI,CAAC,SAAS;oBAIrC,IAAIC,AAAa,iBAAbA,UAEF;wBAAA,IAAID,EAAE,IAAI,CAAC,MAAM,EACf,OAAOD,YAAYC,EAAE,IAAI,CAAC,MAAM;oBAClC;oBAGF,OAAOA;gBACT;gBAEA,MAAME,cAAcH,YAAYF;gBAChC,MAAMM,gBAAgB,QAAAL,CAAAA,oBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,kBAAkB,QAAQ;gBAEhD,IAAIK,AAAkB,gBAAlBA,eAA+B,OAAO;gBAC1C,IAAIA,AAAkB,gBAAlBA,eAA+B,OAAO;gBAC1C,IAAIA,AAAkB,iBAAlBA,eAAgC,OAAO;gBAC3C,IAAIA,AAAkB,eAAlBA,eAA8B,OAAO;gBACzC,IAAIA,AAAkB,gBAAlBA,eAA+B;oBAEjC,IAAIC,AAAAA,IAAAA,mCAAAA,sBAAAA,AAAAA,EAAuBF,cACzB,OAAOd;oBAET,OAAO;gBACT;gBACA,IAAIe,AAAkB,cAAlBA,eAA6B;wBAE5BE,0BAAAA;oBADH,MAAMC,SACJ,AAAC,SAAAD,CAAAA,qBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,QAAAA,CAAAA,2BAAAA,mBAAkB,MAAM,AAAD,IAAvBA,KAAAA,IAAAA,yBACG,GAAG,CAAC,CAACE,SAAoBC,OAAO,CAAC,CAAC,EAAED,OAAO,CAAC,CAAC,GAC9C,IAAI,CAAC,KAAI,KAAK;oBAEnB,OAAO,CAAC,KAAK,EAAED,OAAO,CAAC,CAAC;gBAC1B;gBAEA,IAAIH,AAAkB,eAAlBA,eAA8B;wBAChBM;oBAAhB,MAAMC,UAAU,QAAAD,CAAAA,qBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,mBAAkB,OAAO;oBACzC,IAAIC,WAAWA,QAAQ,MAAM,GAAG,GAAG;wBAEjC,MAAMC,QAAQD,QAAQ,GAAG,CAAC,CAACE,MAAahB,YAAYgB;wBACpD,OAAOD,MAAM,IAAI,CAAC;oBACpB;oBACA,OAAO;gBACT;gBAEAE,QAAQ,IAAI,CACV,2EACAX,YAAY,IAAI;gBAElB,OAAOA,YAAY,QAAQ;YAC7B;YAGA,MAAMY,iBAAiB,CAACjB;oBAwClBC;gBAtCJ,MAAMC,cAAc,CAACC;oBACnB,IAAI,CAACA,EAAE,IAAI,EAAE,OAAOA;oBAEpB,MAAMC,WAAWD,EAAE,IAAI,CAAC,QAAQ;oBAGhC,IACEC,AAAa,kBAAbA,YACAA,AAAa,kBAAbA,YACAA,AAAa,iBAAbA,UAEA,OAAOF,YAAYC,EAAE,IAAI,CAAC,SAAS;oBAIrC,IAAIC,AAAa,iBAAbA,UAEF;wBAAA,IAAID,EAAE,IAAI,CAAC,MAAM,EACf,OAAOD,YAAYC,EAAE,IAAI,CAAC,MAAM;oBAClC;oBAGF,OAAOA;gBACT;gBAGA,IAAI,iBAAiBH,OACnB,OAAOA,MAAM,WAAW,IAAI;gBAG9B,MAAMK,cAAcH,YAAYF;gBAGhC,IAAI,iBAAiBK,aACnB,OAAOA,YAAY,WAAW,IAAI;gBAIpC,IAAIJ,AAAAA,SAAAA,CAAAA,oBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,kBAAkB,QAAQ,AAAD,MAAM,aACjC;oBAAA,IAAI,kCAAkCI,YAAY,IAAI,CAAC,KAAK,IAC1D,OAAO;gBACT;gBAGF,OAAO;YACT;YAEA,KAAK,MAAM,CAACzB,KAAKoB,MAAM,IAAInB,OAAO,OAAO,CAACiB,OACxC,IAAIE,SAAS,AAAiB,YAAjB,OAAOA,OAAoB;gBAEtC,MAAMkB,aACJ,AAAqC,cAArC,OAAQlB,MAAc,UAAU,IAC/BA,MAAc,UAAU;gBAC3B,MAAMmB,kBAAkBD,aAAa,GAAGtC,IAAI,CAAC,CAAC,GAAGA;gBAGjD,MAAMwB,WAAWL,YAAYC;gBAG7B,MAAMoB,cAAcH,eAAejB;gBAGnC,IAAIqB,YAAY,GAAGF,gBAAgB,EAAE,EAAEf,UAAU;gBACjD,IAAIgB,aACFC,aAAa,CAAC,IAAI,EAAED,aAAa;gBAGnCzB,WAAW,IAAI,CAAC0B;YAClB;YAIF,IAAI1B,WAAW,MAAM,GAAG,GAAG;gBACzBF,OAAO,IAAI,CAAC;gBACZE,WAAW,OAAO,CAAC,CAAC2B;oBAClB7B,OAAO,IAAI,CAAC,CAAC,IAAI,EAAE6B,MAAM;gBAC3B;YACF;QACF,OAAO;gBAGkBC;YAAvB,MAAMC,iBAAiB,QAAAD,CAAAA,gBAAAA,OAAO,IAAI,AAAD,IAAVA,KAAAA,IAAAA,cAAa,QAAQ;YAC5C,IAAInB,WAAW;YAEf,IAAIoB,AAAmB,gBAAnBA,gBAAgCpB,WAAW;iBAC1C,IAAIoB,AAAmB,gBAAnBA,gBAAgCpB,WAAW;iBAC/C,IAAIoB,AAAmB,iBAAnBA,gBAAiCpB,WAAW;YAGrD,MAAMgB,cAAc,iBAAiBxB,SAASA,OAAO,WAAW,GAAG;YAGnE,IAAI6B,mBAAmB,CAAC,SAAS,EAAErB,UAAU;YAC7C,IAAIgB,aACFK,oBAAoB,CAAC,IAAI,EAAEL,aAAa;YAE1CK,oBAAoB;YAEpBhC,OAAO,IAAI,CAACgC;QACd;IACF;IAEA,OAAO,CAAC,EAAE,EAAEnC,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,eAAekC,2BAA2B,EAC/CC,WAAW,EACXxC,MAAM,EACNyC,WAAW,EAKZ;IAEC,IAAIA,eAAe,CAACzC,QAClB,MAAM,IAAI0C,MACR;IAIJ,MAAMC,wBAAwBH,YAAY,GAAG,CAAC,CAACrC,SACtCD,qBACLC,QACAJ,cAAc0C,cAAczC,SAAS4C;IAGzC,MAAMC,aAAaF,sBAAsB,IAAI,CAAC;IAE9C,OAAO,CAAC;;;;;;;;;;AAUV,EAAEE,WAAW;;;;;EAKX,EAAE/C,mBAAmB;;;;;;;;;;;;;;;;;;;QAmBf,EAAEE,SAAS,kCAAkC,GAAG;;;;;;;AAOxD,CAAC;AACD;AAEO,MAAM8C,aAAuC;IAClD,MAAM;IACN,aAAa;QACX,MAAM;QACN,QAAQ;QACR,QAAQ;YACN,MAAM;YACN,QAAQ;YACR,YAAY;gBACV,SAAS;oBACP,MAAM;oBACN,OAAO;wBACL,MAAM;wBACN,QAAQ;wBACR,YAAY;4BACV,SAAS;gCACP,MAAM;gCACN,aACE;4BACJ;4BACA,MAAM;gCACJ,MAAM;gCACN,aAAa;4BACf;4BACA,OAAO;gCACL,OAAO;oCACL;wCAAE,MAAM;oCAAO;oCACf;wCACE,MAAM;wCACN,sBAAsB;oCACxB;iCACD;gCACD,aAAa;4BACf;4BACA,QAAQ;gCACN,MAAM;oCAAC;oCAAU;iCAAO;gCACxB,YAAY;oCACV,IAAI;wCAAE,MAAM;oCAAS;oCACrB,QAAQ;wCAAE,MAAM;oCAAS;gCAC3B;gCACA,UAAU;oCAAC;oCAAM;iCAAS;gCAC1B,sBAAsB;gCACtB,aAAa;4BACf;wBACF;wBACA,UAAU;4BAAC;4BAAW;4BAAQ;4BAAS;yBAAS;wBAChD,sBAAsB;oBACxB;oBACA,aAAa;gBACf;gBACA,oCAAoC;oBAClC,MAAM;oBACN,aACE;gBACJ;gBACA,KAAK;oBACH,MAAM;oBACN,aACE;gBACJ;gBACA,OAAO;oBACL,MAAM;wBAAC;wBAAU;qBAAO;oBACxB,aAAa;gBACf;YACF;YACA,UAAU;gBACR;gBACA;gBACA;gBACA;aACD;YACD,sBAAsB;QACxB;IACF;AACF"}

package/dist/lib/ai-model/service-caller/index.js

@@ -50,7 +50,7 @@ const external_jsonrepair_namespaceObject = require("jsonrepair");
 const external_openai_namespaceObject = require("openai");
 var external_openai_default = /*#__PURE__*/ __webpack_require__.n(external_openai_namespaceObject);
 const external_socks_proxy_agent_namespaceObject = require("socks-proxy-agent");
-const external_common_js_namespaceObject = require("../common.js");
+const external_common_js_namespaceObject = require("../../common.js");
 const assertion_js_namespaceObject = require("../prompt/assertion.js");
 const llm_planning_js_namespaceObject = require("../prompt/llm-planning.js");
 async function createChatClient({ AIActionTypeValue, modelConfig }) {
@@ -118,9 +118,9 @@ async function callAI(messages, AIActionTypeValue, modelConfig, options) {
     let usage;
     let timeCost;
     const commonConfig = {
-        temperature: 'vlm-ui-tars' === vlMode ? 0.0 : 0.1,
+        temperature: 'vlm-ui-tars' === vlMode ? 0.0 : void 0,
         stream: !!isStreaming,
-        max_tokens: 'number' == typeof maxTokens ? maxTokens : Number.parseInt(maxTokens || '2048', 10),
+        max_tokens: 'number' == typeof maxTokens ? maxTokens : void 0,
         ...'qwen2.5-vl' === vlMode ? {
             vl_high_resolution_images: true
         } : {}