@midscene/core 1.0.1-beta-20251126084350.0 → 1.0.1-beta-20251126092253.0

package/dist/es/agent/tasks.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"agent/tasks.mjs","sources":["webpack://@midscene/core/./src/agent/tasks.ts"],"sourcesContent":["import { ConversationHistory, plan, uiTarsPlanning } from '@/ai-model';\nimport type { TMultimodalPrompt, TUserPrompt } from '@/common';\nimport type { AbstractInterface } from '@/device';\nimport type Service from '@/service';\nimport type { TaskRunner } from '@/task-runner';\nimport { TaskExecutionError } from '@/task-runner';\nimport type {\n  ExecutionTaskApply,\n  ExecutionTaskInsightQueryApply,\n  ExecutionTaskPlanningApply,\n  ExecutionTaskProgressOptions,\n  InterfaceType,\n  MidsceneYamlFlowItem,\n  PlanningAIResponse,\n  PlanningAction,\n  PlanningActionParamSleep,\n  PlanningActionParamWaitFor,\n  ServiceDump,\n  ServiceExtractOption,\n  ServiceExtractParam,\n  ThinkingStrategy,\n} from '@/types';\nimport { ServiceError } from '@/types';\nimport {\n  type IModelConfig,\n  MIDSCENE_REPLANNING_CYCLE_LIMIT,\n  globalConfigManager,\n} from '@midscene/shared/env';\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport { ExecutionSession } from './execution-session';\nimport { TaskBuilder } from './task-builder';\nimport type { TaskCache } from './task-cache';\nexport { locatePlanForLocate } from './task-builder';\nimport { descriptionOfTree } from '@midscene/shared/extractor';\nimport { taskTitleStr } from './ui-utils';\nimport { parsePrompt } from './utils';\n\ninterface ExecutionResult<OutputType = any> {\n  output: OutputType;\n  thought?: string;\n  runner: TaskRunner;\n}\n\ninterface TaskExecutorHooks {\n  onTaskUpdate?: (\n    runner: TaskRunner,\n    error?: TaskExecutionError,\n  ) => Promise<void> | void;\n}\n\nconst debug = getDebug('device-task-executor');\nconst defaultReplanningCycleLimit = 20;\nconst defaultVlmUiTarsReplanningCycleLimit = 40;\n\nexport { TaskExecutionError };\n\nexport class TaskExecutor {\n  interface: AbstractInterface;\n\n  service: Service;\n\n  taskCache?: TaskCache;\n\n  private readonly taskBuilder: TaskBuilder;\n\n  private conversationHistory: ConversationHistory;\n\n  onTaskStartCallback?: ExecutionTaskProgressOptions['onTaskStart'];\n\n  private readonly hooks?: TaskExecutorHooks;\n\n  replanningCycleLimit?: number;\n\n  // @deprecated use .interface instead\n  get page() {\n    return this.interface;\n  }\n\n  constructor(\n    interfaceInstance: AbstractInterface,\n    service: Service,\n    opts: {\n      taskCache?: TaskCache;\n      onTaskStart?: ExecutionTaskProgressOptions['onTaskStart'];\n      replanningCycleLimit?: number;\n      hooks?: TaskExecutorHooks;\n    },\n  ) {\n    this.interface = interfaceInstance;\n    this.service = service;\n    this.taskCache = opts.taskCache;\n    this.onTaskStartCallback = opts?.onTaskStart;\n    this.replanningCycleLimit = opts.replanningCycleLimit;\n    this.hooks = opts.hooks;\n    this.conversationHistory = new ConversationHistory();\n    this.taskBuilder = new TaskBuilder({\n      interfaceInstance,\n      service,\n      taskCache: opts.taskCache,\n    });\n  }\n\n  private createExecutionSession(\n    title: string,\n    options?: { tasks?: ExecutionTaskApply[] },\n  ) {\n    return new ExecutionSession(\n      title,\n      () => Promise.resolve(this.service.contextRetrieverFn()),\n      {\n        onTaskStart: this.onTaskStartCallback,\n        tasks: options?.tasks,\n        onTaskUpdate: this.hooks?.onTaskUpdate,\n      },\n    );\n  }\n\n  public async convertPlanToExecutable(\n    plans: PlanningAction[],\n    modelConfigForPlanning: IModelConfig,\n    modelConfigForDefaultIntent: IModelConfig,\n    options?: {\n      cacheable?: boolean;\n      subTask?: boolean;\n    },\n  ) {\n    return this.taskBuilder.build(\n      plans,\n      modelConfigForPlanning,\n      modelConfigForDefaultIntent,\n      options,\n    );\n  }\n\n  async loadYamlFlowAsPlanning(userInstruction: string, yamlString: string) {\n    const session = this.createExecutionSession(\n      taskTitleStr('Action', userInstruction),\n    );\n\n    const task: ExecutionTaskPlanningApply = {\n      type: 'Planning',\n      subType: 'LoadYaml',\n      param: {\n        userInstruction,\n      },\n      executor: async (param, executorContext) => {\n        const { uiContext } = executorContext;\n        assert(uiContext, 'uiContext is required for Planning task');\n        return {\n          output: {\n            actions: [],\n            more_actions_needed_by_instruction: false,\n            log: '',\n            yamlString,\n          },\n          cache: {\n            hit: true,\n          },\n          hitBy: {\n            from: 'Cache',\n            context: {\n              yamlString,\n            },\n          },\n        };\n      },\n    };\n    const runner = session.getRunner();\n    await session.appendAndRun(task);\n\n    return {\n      runner,\n    };\n  }\n\n  private createPlanningTask(\n    userInstruction: string,\n    actionContext: string | undefined,\n    modelConfigForPlanning: IModelConfig,\n    modelConfigForDefaultIntent: IModelConfig,\n    includeBbox: boolean,\n    thinkingStrategy: ThinkingStrategy,\n  ): ExecutionTaskPlanningApply {\n    const task: ExecutionTaskPlanningApply = {\n      type: 'Planning',\n      subType: 'Plan',\n      param: {\n        userInstruction,\n        aiActionContext: actionContext,\n      },\n      executor: async (param, executorContext) => {\n        const startTime = Date.now();\n        const { uiContext } = executorContext;\n        assert(uiContext, 'uiContext is required for Planning task');\n        const { vlMode } = modelConfigForPlanning;\n        const uiTarsModelVersion =\n          vlMode === 'vlm-ui-tars'\n            ? modelConfigForPlanning.uiTarsModelVersion\n            : undefined;\n\n        assert(\n          this.interface.actionSpace,\n          'actionSpace for device is not implemented',\n        );\n        const actionSpace = await this.interface.actionSpace();\n        debug(\n          'actionSpace for this interface is:',\n          actionSpace.map((action) => action.name).join(', '),\n        );\n        assert(Array.isArray(actionSpace), 'actionSpace must be an array');\n        if (actionSpace.length === 0) {\n          console.warn(\n            `ActionSpace for ${this.interface.interfaceType} is empty. This may lead to unexpected behavior.`,\n          );\n        }\n\n        const planResult = await (uiTarsModelVersion ? uiTarsPlanning : plan)(\n          param.userInstruction,\n          {\n            context: uiContext,\n            actionContext: param.aiActionContext,\n            interfaceType: this.interface.interfaceType as InterfaceType,\n            actionSpace,\n            modelConfig: modelConfigForPlanning,\n            conversationHistory: this.conversationHistory,\n            includeBbox,\n            thinkingStrategy,\n          },\n        );\n        debug('planResult', JSON.stringify(planResult, null, 2));\n\n        const {\n          actions,\n          log,\n          more_actions_needed_by_instruction,\n          error,\n          usage,\n          rawResponse,\n          sleep,\n        } = planResult;\n\n        executorContext.task.log = {\n          ...(executorContext.task.log || {}),\n          rawResponse,\n        };\n        executorContext.task.usage = usage;\n\n        const finalActions = actions || [];\n\n        if (sleep) {\n          const timeNow = Date.now();\n          const timeRemaining = sleep - (timeNow - startTime);\n          if (timeRemaining > 0) {\n            finalActions.push(this.sleepPlan(timeRemaining));\n          }\n        }\n\n        if (finalActions.length === 0) {\n          assert(\n            !more_actions_needed_by_instruction || sleep,\n            error ? `Failed to plan: ${error}` : 'No plan found',\n          );\n        }\n\n        return {\n          output: {\n            actions: finalActions,\n            more_actions_needed_by_instruction,\n            log,\n            yamlFlow: planResult.yamlFlow,\n          },\n          cache: {\n            hit: false,\n          },\n          uiContext,\n        };\n      },\n    };\n\n    return task;\n  }\n\n  async runPlans(\n    title: string,\n    plans: PlanningAction[],\n    modelConfigForPlanning: IModelConfig,\n    modelConfigForDefaultIntent: IModelConfig,\n  ): Promise<ExecutionResult> {\n    const session = this.createExecutionSession(title);\n    const { tasks } = await this.convertPlanToExecutable(\n      plans,\n      modelConfigForPlanning,\n      modelConfigForDefaultIntent,\n    );\n    const runner = session.getRunner();\n    const result = await session.appendAndRun(tasks);\n    const { output } = result ?? {};\n    return {\n      output,\n      runner,\n    };\n  }\n\n  private getReplanningCycleLimit(isVlmUiTars: boolean) {\n    return (\n      this.replanningCycleLimit ||\n      globalConfigManager.getEnvConfigInNumber(\n        MIDSCENE_REPLANNING_CYCLE_LIMIT,\n      ) ||\n      (isVlmUiTars\n        ? defaultVlmUiTarsReplanningCycleLimit\n        : defaultReplanningCycleLimit)\n    );\n  }\n\n  async action(\n    userPrompt: string,\n    modelConfigForPlanning: IModelConfig,\n    modelConfigForDefaultIntent: IModelConfig,\n    includeBboxInPlanning: boolean,\n    thinkingStrategy: ThinkingStrategy,\n    actionContext?: string,\n    cacheable?: boolean,\n  ): Promise<\n    ExecutionResult<\n      | {\n          yamlFlow?: MidsceneYamlFlowItem[]; // for cache use\n        }\n      | undefined\n    >\n  > {\n    this.conversationHistory.reset();\n\n    const session = this.createExecutionSession(\n      taskTitleStr('Action', userPrompt),\n    );\n    const runner = session.getRunner();\n\n    let replanCount = 0;\n    const yamlFlow: MidsceneYamlFlowItem[] = [];\n    const replanningCycleLimit = this.getReplanningCycleLimit(\n      modelConfigForPlanning.vlMode === 'vlm-ui-tars',\n    );\n\n    // Main planning loop - unified plan/replan logic\n    while (true) {\n      if (replanCount > replanningCycleLimit) {\n        const errorMsg = `Replanning ${replanningCycleLimit} times, which is more than the limit, please split the task into multiple steps`;\n        return session.appendErrorPlan(errorMsg);\n      }\n\n      // Create planning task (automatically includes execution history if available)\n      const planningTask = this.createPlanningTask(\n        userPrompt,\n        actionContext,\n        modelConfigForPlanning,\n        modelConfigForDefaultIntent,\n        includeBboxInPlanning,\n        thinkingStrategy,\n      );\n\n      const result = await session.appendAndRun(planningTask);\n      const planResult = result?.output as PlanningAIResponse | undefined;\n\n      // Execute planned actions\n      const plans = planResult?.actions || [];\n      yamlFlow.push(...(planResult?.yamlFlow || []));\n\n      let executables: Awaited<ReturnType<typeof this.convertPlanToExecutable>>;\n      try {\n        executables = await this.convertPlanToExecutable(\n          plans,\n          modelConfigForPlanning,\n          modelConfigForDefaultIntent,\n          {\n            cacheable,\n            subTask: true,\n          },\n        );\n      } catch (error) {\n        return session.appendErrorPlan(\n          `Error converting plans to executable tasks: ${error}, plans: ${JSON.stringify(\n            plans,\n          )}`,\n        );\n      }\n      await session.appendAndRun(executables.tasks);\n\n      // Check if task is complete\n      if (!planResult?.more_actions_needed_by_instruction) {\n        break;\n      }\n\n      // Increment replan count for next iteration\n      replanCount++;\n    }\n\n    const finalResult = {\n      output: {\n        yamlFlow,\n      },\n      runner,\n    };\n    return finalResult;\n  }\n\n  private createTypeQueryTask(\n    type: 'Query' | 'Boolean' | 'Number' | 'String' | 'Assert' | 'WaitFor',\n    demand: ServiceExtractParam,\n    modelConfig: IModelConfig,\n    opt?: ServiceExtractOption,\n    multimodalPrompt?: TMultimodalPrompt,\n  ) {\n    const queryTask: ExecutionTaskInsightQueryApply = {\n      type: 'Insight',\n      subType: type,\n      param: {\n        dataDemand: multimodalPrompt\n          ? ({\n              demand,\n              multimodalPrompt,\n            } as never)\n          : demand, // for user param presentation in report right sidebar\n      },\n      executor: async (param, taskContext) => {\n        const { task } = taskContext;\n        let queryDump: ServiceDump | undefined;\n        const applyDump = (dump: ServiceDump) => {\n          queryDump = dump;\n          task.log = {\n            dump,\n          };\n        };\n\n        // Get context for query operations\n        const uiContext = taskContext.uiContext;\n        assert(uiContext, 'uiContext is required for Query task');\n\n        const ifTypeRestricted = type !== 'Query';\n        let demandInput = demand;\n        let keyOfResult = 'result';\n        if (ifTypeRestricted && (type === 'Assert' || type === 'WaitFor')) {\n          keyOfResult = 'StatementIsTruthy';\n          const booleanPrompt =\n            type === 'Assert'\n              ? `Boolean, whether the following statement is true: ${demand}`\n              : `Boolean, the user wants to do some 'wait for' operation, please check whether the following statement is true: ${demand}`;\n          demandInput = {\n            [keyOfResult]: booleanPrompt,\n          };\n        } else if (ifTypeRestricted) {\n          demandInput = {\n            [keyOfResult]: `${type}, ${demand}`,\n          };\n        }\n\n        let extractResult;\n\n        let extraPageDescription = '';\n        if (opt?.domIncluded && this.interface.getElementsNodeTree) {\n          debug('appending tree info for page');\n          const tree = await this.interface.getElementsNodeTree();\n          extraPageDescription = await descriptionOfTree(\n            tree,\n            200,\n            false,\n            opt?.domIncluded === 'visible-only',\n          );\n        }\n\n        try {\n          extractResult = await this.service.extract<any>(\n            demandInput,\n            modelConfig,\n            opt,\n            extraPageDescription,\n            multimodalPrompt,\n          );\n        } catch (error) {\n          if (error instanceof ServiceError) {\n            applyDump(error.dump);\n          }\n          throw error;\n        }\n\n        const { data, usage, thought, dump } = extractResult;\n        applyDump(dump);\n\n        let outputResult = data;\n        if (ifTypeRestricted) {\n          // If AI returned a plain string instead of structured format, use it directly\n          if (typeof data === 'string') {\n            outputResult = data;\n          } else if (type === 'WaitFor') {\n            if (data === null || data === undefined) {\n              outputResult = false;\n            } else {\n              outputResult = (data as any)[keyOfResult];\n            }\n          } else if (data === null || data === undefined) {\n            outputResult = null;\n          } else {\n            assert(\n              data?.[keyOfResult] !== undefined,\n              'No result in query data',\n            );\n            outputResult = (data as any)[keyOfResult];\n          }\n        }\n\n        if (type === 'Assert' && !outputResult) {\n          task.usage = usage;\n          task.thought = thought;\n          throw new Error(`Assertion failed: ${thought}`);\n        }\n\n        return {\n          output: outputResult,\n          log: queryDump,\n          usage,\n          thought,\n        };\n      },\n    };\n\n    return queryTask;\n  }\n  async createTypeQueryExecution<T>(\n    type: 'Query' | 'Boolean' | 'Number' | 'String' | 'Assert',\n    demand: ServiceExtractParam,\n    modelConfig: IModelConfig,\n    opt?: ServiceExtractOption,\n    multimodalPrompt?: TMultimodalPrompt,\n  ): Promise<ExecutionResult<T>> {\n    const session = this.createExecutionSession(\n      taskTitleStr(\n        type,\n        typeof demand === 'string' ? demand : JSON.stringify(demand),\n      ),\n    );\n\n    const queryTask = await this.createTypeQueryTask(\n      type,\n      demand,\n      modelConfig,\n      opt,\n      multimodalPrompt,\n    );\n\n    const runner = session.getRunner();\n    const result = await session.appendAndRun(queryTask);\n\n    if (!result) {\n      throw new Error(\n        'result of taskExecutor.flush() is undefined in function createTypeQueryTask',\n      );\n    }\n\n    const { output, thought } = result;\n\n    return {\n      output,\n      thought,\n      runner,\n    };\n  }\n\n  private sleepPlan(timeMs: number): PlanningAction<PlanningActionParamSleep> {\n    return {\n      type: 'Sleep',\n      param: {\n        timeMs,\n      },\n    };\n  }\n\n  async taskForSleep(timeMs: number, _modelConfig: IModelConfig) {\n    return this.taskBuilder.createSleepTask({\n      timeMs,\n    });\n  }\n\n  async waitFor(\n    assertion: TUserPrompt,\n    opt: PlanningActionParamWaitFor,\n    modelConfig: IModelConfig,\n  ): Promise<ExecutionResult<void>> {\n    const { textPrompt, multimodalPrompt } = parsePrompt(assertion);\n\n    const description = `waitFor: ${textPrompt}`;\n    const session = this.createExecutionSession(\n      taskTitleStr('WaitFor', description),\n    );\n    const runner = session.getRunner();\n    const { timeoutMs, checkIntervalMs } = opt;\n\n    assert(assertion, 'No assertion for waitFor');\n    assert(timeoutMs, 'No timeoutMs for waitFor');\n    assert(checkIntervalMs, 'No checkIntervalMs for waitFor');\n\n    assert(\n      checkIntervalMs <= timeoutMs,\n      `wrong config for waitFor: checkIntervalMs must be less than timeoutMs, config: {checkIntervalMs: ${checkIntervalMs}, timeoutMs: ${timeoutMs}}`,\n    );\n\n    const overallStartTime = Date.now();\n    let lastCheckStart = overallStartTime;\n    let errorThought = '';\n    // Continue checking as long as the previous iteration began within the timeout window.\n    while (lastCheckStart - overallStartTime <= timeoutMs) {\n      const currentCheckStart = Date.now();\n      lastCheckStart = currentCheckStart;\n      const queryTask = await this.createTypeQueryTask(\n        'WaitFor',\n        textPrompt,\n        modelConfig,\n        undefined,\n        multimodalPrompt,\n      );\n\n      const result = (await session.appendAndRun(queryTask)) as\n        | {\n            output: boolean;\n            thought?: string;\n          }\n        | undefined;\n\n      if (result?.output) {\n        return {\n          output: undefined,\n          runner,\n        };\n      }\n\n      errorThought =\n        result?.thought ||\n        (!result && `No result from assertion: ${textPrompt}`) ||\n        `unknown error when waiting for assertion: ${textPrompt}`;\n      const now = Date.now();\n      if (now - currentCheckStart < checkIntervalMs) {\n        const timeRemaining = checkIntervalMs - (now - currentCheckStart);\n        const sleepTask = this.taskBuilder.createSleepTask({\n          timeMs: timeRemaining,\n        });\n        await session.append(sleepTask);\n      }\n    }\n\n    return session.appendErrorPlan(`waitFor timeout: ${errorThought}`);\n  }\n}\n"],"names":["debug","getDebug","defaultReplanningCycleLimit","defaultVlmUiTarsReplanningCycleLimit","TaskExecutor","title","options","_this_hooks","ExecutionSession","Promise","plans","modelConfigForPlanning","modelConfigForDefaultIntent","userInstruction","yamlString","session","taskTitleStr","task","param","executorContext","uiContext","assert","runner","actionContext","includeBbox","thinkingStrategy","startTime","Date","vlMode","uiTarsModelVersion","undefined","actionSpace","action","Array","console","planResult","uiTarsPlanning","plan","JSON","actions","log","more_actions_needed_by_instruction","error","usage","rawResponse","sleep","finalActions","timeNow","timeRemaining","tasks","result","output","isVlmUiTars","globalConfigManager","MIDSCENE_REPLANNING_CYCLE_LIMIT","userPrompt","includeBboxInPlanning","cacheable","replanCount","yamlFlow","replanningCycleLimit","errorMsg","planningTask","executables","finalResult","type","demand","modelConfig","opt","multimodalPrompt","queryTask","taskContext","queryDump","applyDump","dump","ifTypeRestricted","demandInput","keyOfResult","booleanPrompt","extractResult","extraPageDescription","tree","descriptionOfTree","ServiceError","data","thought","outputResult","Error","timeMs","_modelConfig","assertion","textPrompt","parsePrompt","description","timeoutMs","checkIntervalMs","overallStartTime","lastCheckStart","errorThought","currentCheckStart","now","sleepTask","interfaceInstance","service","opts","ConversationHistory","TaskBuilder"],"mappings":";;;;;;;;;;;;;;;;;;;;;AAmDA,MAAMA,QAAQC,SAAS;AACvB,MAAMC,8BAA8B;AACpC,MAAMC,uCAAuC;AAItC,MAAMC;IAkBX,IAAI,OAAO;QACT,OAAO,IAAI,CAAC,SAAS;IACvB;IA0BQ,uBACNC,KAAa,EACbC,OAA0C,EAC1C;YAOkBC;QANlB,OAAO,IAAIC,iBACTH,OACA,IAAMI,QAAQ,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,kBAAkB,KACrD;YACE,aAAa,IAAI,CAAC,mBAAmB;YACrC,OAAOH,QAAAA,UAAAA,KAAAA,IAAAA,QAAS,KAAK;YACrB,cAAc,QAAAC,CAAAA,cAAAA,IAAI,CAAC,KAAK,AAAD,IAATA,KAAAA,IAAAA,YAAY,YAAY;QACxC;IAEJ;IAEA,MAAa,wBACXG,KAAuB,EACvBC,sBAAoC,EACpCC,2BAAyC,EACzCN,OAGC,EACD;QACA,OAAO,IAAI,CAAC,WAAW,CAAC,KAAK,CAC3BI,OACAC,wBACAC,6BACAN;IAEJ;IAEA,MAAM,uBAAuBO,eAAuB,EAAEC,UAAkB,EAAE;QACxE,MAAMC,UAAU,IAAI,CAAC,sBAAsB,CACzCC,aAAa,UAAUH;QAGzB,MAAMI,OAAmC;YACvC,MAAM;YACN,SAAS;YACT,OAAO;gBACLJ;YACF;YACA,UAAU,OAAOK,OAAOC;gBACtB,MAAM,EAAEC,SAAS,EAAE,GAAGD;gBACtBE,OAAOD,WAAW;gBAClB,OAAO;oBACL,QAAQ;wBACN,SAAS,EAAE;wBACX,oCAAoC;wBACpC,KAAK;wBACLN;oBACF;oBACA,OAAO;wBACL,KAAK;oBACP;oBACA,OAAO;wBACL,MAAM;wBACN,SAAS;4BACPA;wBACF;oBACF;gBACF;YACF;QACF;QACA,MAAMQ,SAASP,QAAQ,SAAS;QAChC,MAAMA,QAAQ,YAAY,CAACE;QAE3B,OAAO;YACLK;QACF;IACF;IAEQ,mBACNT,eAAuB,EACvBU,aAAiC,EACjCZ,sBAAoC,EACpCC,2BAAyC,EACzCY,WAAoB,EACpBC,gBAAkC,EACN;QAC5B,MAAMR,OAAmC;YACvC,MAAM;YACN,SAAS;YACT,OAAO;gBACLJ;gBACA,iBAAiBU;YACnB;YACA,UAAU,OAAOL,OAAOC;gBACtB,MAAMO,YAAYC,KAAK,GAAG;gBAC1B,MAAM,EAAEP,SAAS,EAAE,GAAGD;gBACtBE,OAAOD,WAAW;gBAClB,MAAM,EAAEQ,MAAM,EAAE,GAAGjB;gBACnB,MAAMkB,qBACJD,AAAW,kBAAXA,SACIjB,uBAAuB,kBAAkB,GACzCmB;gBAENT,OACE,IAAI,CAAC,SAAS,CAAC,WAAW,EAC1B;gBAEF,MAAMU,cAAc,MAAM,IAAI,CAAC,SAAS,CAAC,WAAW;gBACpD/B,MACE,sCACA+B,YAAY,GAAG,CAAC,CAACC,SAAWA,OAAO,IAAI,EAAE,IAAI,CAAC;gBAEhDX,OAAOY,MAAM,OAAO,CAACF,cAAc;gBACnC,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EACpBG,QAAQ,IAAI,CACV,CAAC,gBAAgB,EAAE,IAAI,CAAC,SAAS,CAAC,aAAa,CAAC,gDAAgD,CAAC;gBAIrG,MAAMC,aAAa,MAAON,AAAAA,CAAAA,qBAAqBO,iBAAiBC,IAAG,EACjEnB,MAAM,eAAe,EACrB;oBACE,SAASE;oBACT,eAAeF,MAAM,eAAe;oBACpC,eAAe,IAAI,CAAC,SAAS,CAAC,aAAa;oBAC3Ca;oBACA,aAAapB;oBACb,qBAAqB,IAAI,CAAC,mBAAmB;oBAC7Ca;oBACAC;gBACF;gBAEFzB,MAAM,cAAcsC,KAAK,SAAS,CAACH,YAAY,MAAM;gBAErD,MAAM,EACJI,OAAO,EACPC,GAAG,EACHC,kCAAkC,EAClCC,KAAK,EACLC,KAAK,EACLC,WAAW,EACXC,KAAK,EACN,GAAGV;gBAEJhB,gBAAgB,IAAI,CAAC,GAAG,GAAG;oBACzB,GAAIA,gBAAgB,IAAI,CAAC,GAAG,IAAI,CAAC,CAAC;oBAClCyB;gBACF;gBACAzB,gBAAgB,IAAI,CAAC,KAAK,GAAGwB;gBAE7B,MAAMG,eAAeP,WAAW,EAAE;gBAElC,IAAIM,OAAO;oBACT,MAAME,UAAUpB,KAAK,GAAG;oBACxB,MAAMqB,gBAAgBH,QAASE,CAAAA,UAAUrB,SAAQ;oBACjD,IAAIsB,gBAAgB,GAClBF,aAAa,IAAI,CAAC,IAAI,CAAC,SAAS,CAACE;gBAErC;gBAEA,IAAIF,AAAwB,MAAxBA,aAAa,MAAM,EACrBzB,OACE,CAACoB,sCAAsCI,OACvCH,QAAQ,CAAC,gBAAgB,EAAEA,OAAO,GAAG;gBAIzC,OAAO;oBACL,QAAQ;wBACN,SAASI;wBACTL;wBACAD;wBACA,UAAUL,WAAW,QAAQ;oBAC/B;oBACA,OAAO;wBACL,KAAK;oBACP;oBACAf;gBACF;YACF;QACF;QAEA,OAAOH;IACT;IAEA,MAAM,SACJZ,KAAa,EACbK,KAAuB,EACvBC,sBAAoC,EACpCC,2BAAyC,EACf;QAC1B,MAAMG,UAAU,IAAI,CAAC,sBAAsB,CAACV;QAC5C,MAAM,EAAE4C,KAAK,EAAE,GAAG,MAAM,IAAI,CAAC,uBAAuB,CAClDvC,OACAC,wBACAC;QAEF,MAAMU,SAASP,QAAQ,SAAS;QAChC,MAAMmC,SAAS,MAAMnC,QAAQ,YAAY,CAACkC;QAC1C,MAAM,EAAEE,MAAM,EAAE,GAAGD,UAAU,CAAC;QAC9B,OAAO;YACLC;YACA7B;QACF;IACF;IAEQ,wBAAwB8B,WAAoB,EAAE;QACpD,OACE,IAAI,CAAC,oBAAoB,IACzBC,oBAAoB,oBAAoB,CACtCC,oCAEDF,CAAAA,cACGjD,uCACAD,2BAA0B;IAElC;IAEA,MAAM,OACJqD,UAAkB,EAClB5C,sBAAoC,EACpCC,2BAAyC,EACzC4C,qBAA8B,EAC9B/B,gBAAkC,EAClCF,aAAsB,EACtBkC,SAAmB,EAQnB;QACA,IAAI,CAAC,mBAAmB,CAAC,KAAK;QAE9B,MAAM1C,UAAU,IAAI,CAAC,sBAAsB,CACzCC,aAAa,UAAUuC;QAEzB,MAAMjC,SAASP,QAAQ,SAAS;QAEhC,IAAI2C,cAAc;QAClB,MAAMC,WAAmC,EAAE;QAC3C,MAAMC,uBAAuB,IAAI,CAAC,uBAAuB,CACvDjD,AAAkC,kBAAlCA,uBAAuB,MAAM;QAI/B,MAAO,KAAM;YACX,IAAI+C,cAAcE,sBAAsB;gBACtC,MAAMC,WAAW,CAAC,WAAW,EAAED,qBAAqB,+EAA+E,CAAC;gBACpI,OAAO7C,QAAQ,eAAe,CAAC8C;YACjC;YAGA,MAAMC,eAAe,IAAI,CAAC,kBAAkB,CAC1CP,YACAhC,eACAZ,wBACAC,6BACA4C,uBACA/B;YAGF,MAAMyB,SAAS,MAAMnC,QAAQ,YAAY,CAAC+C;YAC1C,MAAM3B,aAAae,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,MAAM;YAGjC,MAAMxC,QAAQyB,AAAAA,CAAAA,QAAAA,aAAAA,KAAAA,IAAAA,WAAY,OAAO,AAAD,KAAK,EAAE;YACvCwB,SAAS,IAAI,IAAKxB,AAAAA,CAAAA,QAAAA,aAAAA,KAAAA,IAAAA,WAAY,QAAQ,AAAD,KAAK,EAAE;YAE5C,IAAI4B;YACJ,IAAI;gBACFA,cAAc,MAAM,IAAI,CAAC,uBAAuB,CAC9CrD,OACAC,wBACAC,6BACA;oBACE6C;oBACA,SAAS;gBACX;YAEJ,EAAE,OAAOf,OAAO;gBACd,OAAO3B,QAAQ,eAAe,CAC5B,CAAC,4CAA4C,EAAE2B,MAAM,SAAS,EAAEJ,KAAK,SAAS,CAC5E5B,QACC;YAEP;YACA,MAAMK,QAAQ,YAAY,CAACgD,YAAY,KAAK;YAG5C,IAAI,CAAC5B,CAAAA,QAAAA,aAAAA,KAAAA,IAAAA,WAAY,kCAAkC,AAAD,GAChD;YAIFuB;QACF;QAEA,MAAMM,cAAc;YAClB,QAAQ;gBACNL;YACF;YACArC;QACF;QACA,OAAO0C;IACT;IAEQ,oBACNC,IAAsE,EACtEC,MAA2B,EAC3BC,WAAyB,EACzBC,GAA0B,EAC1BC,gBAAoC,EACpC;QACA,MAAMC,YAA4C;YAChD,MAAM;YACN,SAASL;YACT,OAAO;gBACL,YAAYI,mBACP;oBACCH;oBACAG;gBACF,IACAH;YACN;YACA,UAAU,OAAOhD,OAAOqD;gBACtB,MAAM,EAAEtD,IAAI,EAAE,GAAGsD;gBACjB,IAAIC;gBACJ,MAAMC,YAAY,CAACC;oBACjBF,YAAYE;oBACZzD,KAAK,GAAG,GAAG;wBACTyD;oBACF;gBACF;gBAGA,MAAMtD,YAAYmD,YAAY,SAAS;gBACvClD,OAAOD,WAAW;gBAElB,MAAMuD,mBAAmBV,AAAS,YAATA;gBACzB,IAAIW,cAAcV;gBAClB,IAAIW,cAAc;gBAClB,IAAIF,oBAAqBV,CAAAA,AAAS,aAATA,QAAqBA,AAAS,cAATA,IAAiB,GAAI;oBACjEY,cAAc;oBACd,MAAMC,gBACJb,AAAS,aAATA,OACI,CAAC,kDAAkD,EAAEC,QAAQ,GAC7D,CAAC,+GAA+G,EAAEA,QAAQ;oBAChIU,cAAc;wBACZ,CAACC,YAAY,EAAEC;oBACjB;gBACF,OAAO,IAAIH,kBACTC,cAAc;oBACZ,CAACC,YAAY,EAAE,GAAGZ,KAAK,EAAE,EAAEC,QAAQ;gBACrC;gBAGF,IAAIa;gBAEJ,IAAIC,uBAAuB;gBAC3B,IAAIZ,AAAAA,CAAAA,QAAAA,MAAAA,KAAAA,IAAAA,IAAK,WAAW,AAAD,KAAK,IAAI,CAAC,SAAS,CAAC,mBAAmB,EAAE;oBAC1DpE,MAAM;oBACN,MAAMiF,OAAO,MAAM,IAAI,CAAC,SAAS,CAAC,mBAAmB;oBACrDD,uBAAuB,MAAME,kBAC3BD,MACA,KACA,OACAb,AAAAA,CAAAA,QAAAA,MAAAA,KAAAA,IAAAA,IAAK,WAAW,AAAD,MAAM;gBAEzB;gBAEA,IAAI;oBACFW,gBAAgB,MAAM,IAAI,CAAC,OAAO,CAAC,OAAO,CACxCH,aACAT,aACAC,KACAY,sBACAX;gBAEJ,EAAE,OAAO3B,OAAO;oBACd,IAAIA,iBAAiByC,cACnBV,UAAU/B,MAAM,IAAI;oBAEtB,MAAMA;gBACR;gBAEA,MAAM,EAAE0C,IAAI,EAAEzC,KAAK,EAAE0C,OAAO,EAAEX,IAAI,EAAE,GAAGK;gBACvCN,UAAUC;gBAEV,IAAIY,eAAeF;gBACnB,IAAIT,kBAEF,IAAI,AAAgB,YAAhB,OAAOS,MACTE,eAAeF;qBACV,IAAInB,AAAS,cAATA,MAEPqB,eADEF,QAAAA,OACa,QAECA,IAAY,CAACP,YAAY;qBAEtC,IAAIO,QAAAA,MACTE,eAAe;qBACV;oBACLjE,OACE+D,AAAAA,CAAAA,QAAAA,OAAAA,KAAAA,IAAAA,IAAM,CAACP,YAAY,AAAD,MAAM/C,QACxB;oBAEFwD,eAAgBF,IAAY,CAACP,YAAY;gBAC3C;gBAGF,IAAIZ,AAAS,aAATA,QAAqB,CAACqB,cAAc;oBACtCrE,KAAK,KAAK,GAAG0B;oBACb1B,KAAK,OAAO,GAAGoE;oBACf,MAAM,IAAIE,MAAM,CAAC,kBAAkB,EAAEF,SAAS;gBAChD;gBAEA,OAAO;oBACL,QAAQC;oBACR,KAAKd;oBACL7B;oBACA0C;gBACF;YACF;QACF;QAEA,OAAOf;IACT;IACA,MAAM,yBACJL,IAA0D,EAC1DC,MAA2B,EAC3BC,WAAyB,EACzBC,GAA0B,EAC1BC,gBAAoC,EACP;QAC7B,MAAMtD,UAAU,IAAI,CAAC,sBAAsB,CACzCC,aACEiD,MACA,AAAkB,YAAlB,OAAOC,SAAsBA,SAAS5B,KAAK,SAAS,CAAC4B;QAIzD,MAAMI,YAAY,MAAM,IAAI,CAAC,mBAAmB,CAC9CL,MACAC,QACAC,aACAC,KACAC;QAGF,MAAM/C,SAASP,QAAQ,SAAS;QAChC,MAAMmC,SAAS,MAAMnC,QAAQ,YAAY,CAACuD;QAE1C,IAAI,CAACpB,QACH,MAAM,IAAIqC,MACR;QAIJ,MAAM,EAAEpC,MAAM,EAAEkC,OAAO,EAAE,GAAGnC;QAE5B,OAAO;YACLC;YACAkC;YACA/D;QACF;IACF;IAEQ,UAAUkE,MAAc,EAA4C;QAC1E,OAAO;YACL,MAAM;YACN,OAAO;gBACLA;YACF;QACF;IACF;IAEA,MAAM,aAAaA,MAAc,EAAEC,YAA0B,EAAE;QAC7D,OAAO,IAAI,CAAC,WAAW,CAAC,eAAe,CAAC;YACtCD;QACF;IACF;IAEA,MAAM,QACJE,SAAsB,EACtBtB,GAA+B,EAC/BD,WAAyB,EACO;QAChC,MAAM,EAAEwB,UAAU,EAAEtB,gBAAgB,EAAE,GAAGuB,YAAYF;QAErD,MAAMG,cAAc,CAAC,SAAS,EAAEF,YAAY;QAC5C,MAAM5E,UAAU,IAAI,CAAC,sBAAsB,CACzCC,aAAa,WAAW6E;QAE1B,MAAMvE,SAASP,QAAQ,SAAS;QAChC,MAAM,EAAE+E,SAAS,EAAEC,eAAe,EAAE,GAAG3B;QAEvC/C,OAAOqE,WAAW;QAClBrE,OAAOyE,WAAW;QAClBzE,OAAO0E,iBAAiB;QAExB1E,OACE0E,mBAAmBD,WACnB,CAAC,iGAAiG,EAAEC,gBAAgB,aAAa,EAAED,UAAU,CAAC,CAAC;QAGjJ,MAAME,mBAAmBrE,KAAK,GAAG;QACjC,IAAIsE,iBAAiBD;QACrB,IAAIE,eAAe;QAEnB,MAAOD,iBAAiBD,oBAAoBF,UAAW;YACrD,MAAMK,oBAAoBxE,KAAK,GAAG;YAClCsE,iBAAiBE;YACjB,MAAM7B,YAAY,MAAM,IAAI,CAAC,mBAAmB,CAC9C,WACAqB,YACAxB,aACArC,QACAuC;YAGF,MAAMnB,SAAU,MAAMnC,QAAQ,YAAY,CAACuD;YAO3C,IAAIpB,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,MAAM,EAChB,OAAO;gBACL,QAAQpB;gBACRR;YACF;YAGF4E,eACEhD,AAAAA,CAAAA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,OAAO,AAAD,KACb,CAACA,UAAU,CAAC,0BAA0B,EAAEyC,YAAY,IACrD,CAAC,0CAA0C,EAAEA,YAAY;YAC3D,MAAMS,MAAMzE,KAAK,GAAG;YACpB,IAAIyE,MAAMD,oBAAoBJ,iBAAiB;gBAC7C,MAAM/C,gBAAgB+C,kBAAmBK,CAAAA,MAAMD,iBAAgB;gBAC/D,MAAME,YAAY,IAAI,CAAC,WAAW,CAAC,eAAe,CAAC;oBACjD,QAAQrD;gBACV;gBACA,MAAMjC,QAAQ,MAAM,CAACsF;YACvB;QACF;QAEA,OAAOtF,QAAQ,eAAe,CAAC,CAAC,iBAAiB,EAAEmF,cAAc;IACnE;IA3jBA,YACEI,iBAAoC,EACpCC,OAAgB,EAChBC,IAKC,CACD;QA9BF;QAEA;QAEA;QAEA,uBAAiB,eAAjB;QAEA,uBAAQ,uBAAR;QAEA;QAEA,uBAAiB,SAAjB;QAEA;QAiBE,IAAI,CAAC,SAAS,GAAGF;QACjB,IAAI,CAAC,OAAO,GAAGC;QACf,IAAI,CAAC,SAAS,GAAGC,KAAK,SAAS;QAC/B,IAAI,CAAC,mBAAmB,GAAGA,QAAAA,OAAAA,KAAAA,IAAAA,KAAM,WAAW;QAC5C,IAAI,CAAC,oBAAoB,GAAGA,KAAK,oBAAoB;QACrD,IAAI,CAAC,KAAK,GAAGA,KAAK,KAAK;QACvB,IAAI,CAAC,mBAAmB,GAAG,IAAIC;QAC/B,IAAI,CAAC,WAAW,GAAG,IAAIC,YAAY;YACjCJ;YACAC;YACA,WAAWC,KAAK,SAAS;QAC3B;IACF;AAsiBF"}
+{"version":3,"file":"agent/tasks.mjs","sources":["webpack://@midscene/core/./src/agent/tasks.ts"],"sourcesContent":["import { ConversationHistory, plan, uiTarsPlanning } from '@/ai-model';\nimport type { TMultimodalPrompt, TUserPrompt } from '@/common';\nimport type { AbstractInterface } from '@/device';\nimport type Service from '@/service';\nimport type { TaskRunner } from '@/task-runner';\nimport { TaskExecutionError } from '@/task-runner';\nimport type {\n  ExecutionTaskApply,\n  ExecutionTaskInsightQueryApply,\n  ExecutionTaskPlanningApply,\n  ExecutionTaskProgressOptions,\n  InterfaceType,\n  MidsceneYamlFlowItem,\n  PlanningAIResponse,\n  PlanningAction,\n  PlanningActionParamSleep,\n  PlanningActionParamWaitFor,\n  ServiceDump,\n  ServiceExtractOption,\n  ServiceExtractParam,\n  ThinkingStrategy,\n} from '@/types';\nimport { ServiceError } from '@/types';\nimport {\n  type IModelConfig,\n  MIDSCENE_REPLANNING_CYCLE_LIMIT,\n  globalConfigManager,\n} from '@midscene/shared/env';\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport { ExecutionSession } from './execution-session';\nimport { TaskBuilder } from './task-builder';\nimport type { TaskCache } from './task-cache';\nexport { locatePlanForLocate } from './task-builder';\nimport { descriptionOfTree } from '@midscene/shared/extractor';\nimport { taskTitleStr } from './ui-utils';\nimport { parsePrompt } from './utils';\n\ninterface ExecutionResult<OutputType = any> {\n  output: OutputType;\n  thought?: string;\n  runner: TaskRunner;\n}\n\ninterface TaskExecutorHooks {\n  onTaskUpdate?: (\n    runner: TaskRunner,\n    error?: TaskExecutionError,\n  ) => Promise<void> | void;\n}\n\nconst debug = getDebug('device-task-executor');\nconst defaultReplanningCycleLimit = 20;\nconst defaultVlmUiTarsReplanningCycleLimit = 40;\nconst maxErrorCountAllowedInOnePlanningLoop = 5;\n\nexport { TaskExecutionError };\n\nexport class TaskExecutor {\n  interface: AbstractInterface;\n\n  service: Service;\n\n  taskCache?: TaskCache;\n\n  private readonly taskBuilder: TaskBuilder;\n\n  private conversationHistory: ConversationHistory;\n\n  onTaskStartCallback?: ExecutionTaskProgressOptions['onTaskStart'];\n\n  private readonly hooks?: TaskExecutorHooks;\n\n  replanningCycleLimit?: number;\n\n  // @deprecated use .interface instead\n  get page() {\n    return this.interface;\n  }\n\n  constructor(\n    interfaceInstance: AbstractInterface,\n    service: Service,\n    opts: {\n      taskCache?: TaskCache;\n      onTaskStart?: ExecutionTaskProgressOptions['onTaskStart'];\n      replanningCycleLimit?: number;\n      hooks?: TaskExecutorHooks;\n    },\n  ) {\n    this.interface = interfaceInstance;\n    this.service = service;\n    this.taskCache = opts.taskCache;\n    this.onTaskStartCallback = opts?.onTaskStart;\n    this.replanningCycleLimit = opts.replanningCycleLimit;\n    this.hooks = opts.hooks;\n    this.conversationHistory = new ConversationHistory();\n    this.taskBuilder = new TaskBuilder({\n      interfaceInstance,\n      service,\n      taskCache: opts.taskCache,\n    });\n  }\n\n  private createExecutionSession(\n    title: string,\n    options?: { tasks?: ExecutionTaskApply[] },\n  ) {\n    return new ExecutionSession(\n      title,\n      () => Promise.resolve(this.service.contextRetrieverFn()),\n      {\n        onTaskStart: this.onTaskStartCallback,\n        tasks: options?.tasks,\n        onTaskUpdate: this.hooks?.onTaskUpdate,\n      },\n    );\n  }\n\n  public async convertPlanToExecutable(\n    plans: PlanningAction[],\n    modelConfigForPlanning: IModelConfig,\n    modelConfigForDefaultIntent: IModelConfig,\n    options?: {\n      cacheable?: boolean;\n      subTask?: boolean;\n    },\n  ) {\n    return this.taskBuilder.build(\n      plans,\n      modelConfigForPlanning,\n      modelConfigForDefaultIntent,\n      options,\n    );\n  }\n\n  async loadYamlFlowAsPlanning(userInstruction: string, yamlString: string) {\n    const session = this.createExecutionSession(\n      taskTitleStr('Action', userInstruction),\n    );\n\n    const task: ExecutionTaskPlanningApply = {\n      type: 'Planning',\n      subType: 'LoadYaml',\n      param: {\n        userInstruction,\n      },\n      executor: async (param, executorContext) => {\n        const { uiContext } = executorContext;\n        assert(uiContext, 'uiContext is required for Planning task');\n        return {\n          output: {\n            actions: [],\n            more_actions_needed_by_instruction: false,\n            log: '',\n            yamlString,\n          },\n          cache: {\n            hit: true,\n          },\n          hitBy: {\n            from: 'Cache',\n            context: {\n              yamlString,\n            },\n          },\n        };\n      },\n    };\n    const runner = session.getRunner();\n    await session.appendAndRun(task);\n\n    return {\n      runner,\n    };\n  }\n\n  async runPlans(\n    title: string,\n    plans: PlanningAction[],\n    modelConfigForPlanning: IModelConfig,\n    modelConfigForDefaultIntent: IModelConfig,\n  ): Promise<ExecutionResult> {\n    const session = this.createExecutionSession(title);\n    const { tasks } = await this.convertPlanToExecutable(\n      plans,\n      modelConfigForPlanning,\n      modelConfigForDefaultIntent,\n    );\n    const runner = session.getRunner();\n    const result = await session.appendAndRun(tasks);\n    const { output } = result ?? {};\n    return {\n      output,\n      runner,\n    };\n  }\n\n  private getReplanningCycleLimit(isVlmUiTars: boolean) {\n    return (\n      this.replanningCycleLimit ||\n      globalConfigManager.getEnvConfigInNumber(\n        MIDSCENE_REPLANNING_CYCLE_LIMIT,\n      ) ||\n      (isVlmUiTars\n        ? defaultVlmUiTarsReplanningCycleLimit\n        : defaultReplanningCycleLimit)\n    );\n  }\n\n  async action(\n    userPrompt: string,\n    modelConfigForPlanning: IModelConfig,\n    modelConfigForDefaultIntent: IModelConfig,\n    includeBboxInPlanning: boolean,\n    thinkingStrategy: ThinkingStrategy,\n    backgroundKnowledge?: string,\n    cacheable?: boolean,\n  ): Promise<\n    ExecutionResult<\n      | {\n          yamlFlow?: MidsceneYamlFlowItem[]; // for cache use\n        }\n      | undefined\n    >\n  > {\n    this.conversationHistory.reset();\n\n    const session = this.createExecutionSession(\n      taskTitleStr('Action', userPrompt),\n    );\n    const runner = session.getRunner();\n\n    let replanCount = 0;\n    const yamlFlow: MidsceneYamlFlowItem[] = [];\n    const replanningCycleLimit = this.getReplanningCycleLimit(\n      modelConfigForPlanning.vlMode === 'vlm-ui-tars',\n    );\n\n    let errorCountInOnePlanningLoop = 0; // count the number of errors in one planning loop\n\n    // Main planning loop - unified plan/replan logic\n    while (true) {\n      const result = await session.appendAndRun(\n        {\n          type: 'Planning',\n          subType: 'Plan',\n          param: {\n            userInstruction: userPrompt,\n            aiActionContext: backgroundKnowledge,\n          },\n          executor: async (param, executorContext) => {\n            const startTime = Date.now();\n            const { uiContext } = executorContext;\n            assert(uiContext, 'uiContext is required for Planning task');\n            const { vlMode } = modelConfigForPlanning;\n            const uiTarsModelVersion =\n              vlMode === 'vlm-ui-tars'\n                ? modelConfigForPlanning.uiTarsModelVersion\n                : undefined;\n\n            assert(\n              this.interface.actionSpace,\n              'actionSpace for device is not implemented',\n            );\n            const actionSpace = await this.interface.actionSpace();\n            debug(\n              'actionSpace for this interface is:',\n              actionSpace.map((action) => action.name).join(', '),\n            );\n            assert(Array.isArray(actionSpace), 'actionSpace must be an array');\n            if (actionSpace.length === 0) {\n              console.warn(\n                `ActionSpace for ${this.interface.interfaceType} is empty. This may lead to unexpected behavior.`,\n              );\n            }\n\n            const planResult = await (uiTarsModelVersion\n              ? uiTarsPlanning\n              : plan)(param.userInstruction, {\n              context: uiContext,\n              actionContext: param.aiActionContext,\n              interfaceType: this.interface.interfaceType as InterfaceType,\n              actionSpace,\n              modelConfig: modelConfigForPlanning,\n              conversationHistory: this.conversationHistory,\n              includeBbox: includeBboxInPlanning,\n              thinkingStrategy,\n            });\n            debug('planResult', JSON.stringify(planResult, null, 2));\n\n            const {\n              actions,\n              log,\n              more_actions_needed_by_instruction,\n              error,\n              usage,\n              rawResponse,\n              sleep,\n            } = planResult;\n\n            executorContext.task.log = {\n              ...(executorContext.task.log || {}),\n              rawResponse,\n            };\n            executorContext.task.usage = usage;\n\n            const finalActions = actions || [];\n\n            if (sleep) {\n              const timeNow = Date.now();\n              const timeRemaining = sleep - (timeNow - startTime);\n              if (timeRemaining > 0) {\n                finalActions.push(this.sleepPlan(timeRemaining));\n              }\n            }\n\n            if (finalActions.length === 0) {\n              assert(\n                !more_actions_needed_by_instruction || sleep,\n                error ? `Failed to plan: ${error}` : 'No plan found',\n              );\n            }\n\n            return {\n              output: {\n                actions: finalActions,\n                more_actions_needed_by_instruction,\n                log,\n                yamlFlow: planResult.yamlFlow,\n              },\n              cache: {\n                hit: false,\n              },\n              uiContext,\n            };\n          },\n        },\n        {\n          allowWhenError: true,\n        },\n      );\n\n      const planResult = result?.output as PlanningAIResponse | undefined;\n\n      // Execute planned actions\n      const plans = planResult?.actions || [];\n      yamlFlow.push(...(planResult?.yamlFlow || []));\n\n      let executables: Awaited<ReturnType<typeof this.convertPlanToExecutable>>;\n      try {\n        executables = await this.convertPlanToExecutable(\n          plans,\n          modelConfigForPlanning,\n          modelConfigForDefaultIntent,\n          {\n            cacheable,\n            subTask: true,\n          },\n        );\n      } catch (error) {\n        return session.appendErrorPlan(\n          `Error converting plans to executable tasks: ${error}, plans: ${JSON.stringify(\n            plans,\n          )}`,\n        );\n      }\n      if (this.conversationHistory.pendingFeedbackMessage) {\n        console.warn(\n          'unconsumed pending feedback message detected, this may lead to unexpected planning result:',\n          this.conversationHistory.pendingFeedbackMessage,\n        );\n      }\n      try {\n        await session.appendAndRun(executables.tasks);\n      } catch (error: any) {\n        errorCountInOnePlanningLoop++;\n        this.conversationHistory.pendingFeedbackMessage = `Error executing running tasks: ${error?.message || String(error)}`;\n        debug(\n          'error when executing running tasks, but continue to run:',\n          error instanceof Error ? error.message : String(error),\n        );\n      }\n\n      if (errorCountInOnePlanningLoop > maxErrorCountAllowedInOnePlanningLoop) {\n        return session.appendErrorPlan('Too many errors in one planning loop');\n      }\n\n      // Check if task is complete\n      if (!planResult?.more_actions_needed_by_instruction) {\n        break;\n      }\n\n      // Increment replan count for next iteration\n      ++replanCount;\n\n      if (replanCount > replanningCycleLimit) {\n        const errorMsg = `Replanning ${replanningCycleLimit} times, which is more than the limit, please split the task into multiple steps`;\n        return session.appendErrorPlan(errorMsg);\n      }\n\n      if (!this.conversationHistory.pendingFeedbackMessage) {\n        this.conversationHistory.pendingFeedbackMessage =\n          'I have finished the action previously planned.';\n      }\n    }\n\n    const finalResult = {\n      output: {\n        yamlFlow,\n      },\n      runner,\n    };\n    return finalResult;\n  }\n\n  private createTypeQueryTask(\n    type: 'Query' | 'Boolean' | 'Number' | 'String' | 'Assert' | 'WaitFor',\n    demand: ServiceExtractParam,\n    modelConfig: IModelConfig,\n    opt?: ServiceExtractOption,\n    multimodalPrompt?: TMultimodalPrompt,\n  ) {\n    const queryTask: ExecutionTaskInsightQueryApply = {\n      type: 'Insight',\n      subType: type,\n      param: {\n        dataDemand: multimodalPrompt\n          ? ({\n              demand,\n              multimodalPrompt,\n            } as never)\n          : demand, // for user param presentation in report right sidebar\n      },\n      executor: async (param, taskContext) => {\n        const { task } = taskContext;\n        let queryDump: ServiceDump | undefined;\n        const applyDump = (dump: ServiceDump) => {\n          queryDump = dump;\n          task.log = {\n            dump,\n          };\n        };\n\n        // Get context for query operations\n        const uiContext = taskContext.uiContext;\n        assert(uiContext, 'uiContext is required for Query task');\n\n        const ifTypeRestricted = type !== 'Query';\n        let demandInput = demand;\n        let keyOfResult = 'result';\n        if (ifTypeRestricted && (type === 'Assert' || type === 'WaitFor')) {\n          keyOfResult = 'StatementIsTruthy';\n          const booleanPrompt =\n            type === 'Assert'\n              ? `Boolean, whether the following statement is true: ${demand}`\n              : `Boolean, the user wants to do some 'wait for' operation, please check whether the following statement is true: ${demand}`;\n          demandInput = {\n            [keyOfResult]: booleanPrompt,\n          };\n        } else if (ifTypeRestricted) {\n          demandInput = {\n            [keyOfResult]: `${type}, ${demand}`,\n          };\n        }\n\n        let extractResult;\n\n        let extraPageDescription = '';\n        if (opt?.domIncluded && this.interface.getElementsNodeTree) {\n          debug('appending tree info for page');\n          const tree = await this.interface.getElementsNodeTree();\n          extraPageDescription = await descriptionOfTree(\n            tree,\n            200,\n            false,\n            opt?.domIncluded === 'visible-only',\n          );\n        }\n\n        try {\n          extractResult = await this.service.extract<any>(\n            demandInput,\n            modelConfig,\n            opt,\n            extraPageDescription,\n            multimodalPrompt,\n          );\n        } catch (error) {\n          if (error instanceof ServiceError) {\n            applyDump(error.dump);\n          }\n          throw error;\n        }\n\n        const { data, usage, thought, dump } = extractResult;\n        applyDump(dump);\n\n        let outputResult = data;\n        if (ifTypeRestricted) {\n          // If AI returned a plain string instead of structured format, use it directly\n          if (typeof data === 'string') {\n            outputResult = data;\n          } else if (type === 'WaitFor') {\n            if (data === null || data === undefined) {\n              outputResult = false;\n            } else {\n              outputResult = (data as any)[keyOfResult];\n            }\n          } else if (data === null || data === undefined) {\n            outputResult = null;\n          } else {\n            assert(\n              data?.[keyOfResult] !== undefined,\n              'No result in query data',\n            );\n            outputResult = (data as any)[keyOfResult];\n          }\n        }\n\n        if (type === 'Assert' && !outputResult) {\n          task.usage = usage;\n          task.thought = thought;\n          throw new Error(`Assertion failed: ${thought}`);\n        }\n\n        return {\n          output: outputResult,\n          log: queryDump,\n          usage,\n          thought,\n        };\n      },\n    };\n\n    return queryTask;\n  }\n  async createTypeQueryExecution<T>(\n    type: 'Query' | 'Boolean' | 'Number' | 'String' | 'Assert',\n    demand: ServiceExtractParam,\n    modelConfig: IModelConfig,\n    opt?: ServiceExtractOption,\n    multimodalPrompt?: TMultimodalPrompt,\n  ): Promise<ExecutionResult<T>> {\n    const session = this.createExecutionSession(\n      taskTitleStr(\n        type,\n        typeof demand === 'string' ? demand : JSON.stringify(demand),\n      ),\n    );\n\n    const queryTask = await this.createTypeQueryTask(\n      type,\n      demand,\n      modelConfig,\n      opt,\n      multimodalPrompt,\n    );\n\n    const runner = session.getRunner();\n    const result = await session.appendAndRun(queryTask);\n\n    if (!result) {\n      throw new Error(\n        'result of taskExecutor.flush() is undefined in function createTypeQueryTask',\n      );\n    }\n\n    const { output, thought } = result;\n\n    return {\n      output,\n      thought,\n      runner,\n    };\n  }\n\n  private sleepPlan(timeMs: number): PlanningAction<PlanningActionParamSleep> {\n    return {\n      type: 'Sleep',\n      param: {\n        timeMs,\n      },\n    };\n  }\n\n  async taskForSleep(timeMs: number, _modelConfig: IModelConfig) {\n    return this.taskBuilder.createSleepTask({\n      timeMs,\n    });\n  }\n\n  async waitFor(\n    assertion: TUserPrompt,\n    opt: PlanningActionParamWaitFor,\n    modelConfig: IModelConfig,\n  ): Promise<ExecutionResult<void>> {\n    const { textPrompt, multimodalPrompt } = parsePrompt(assertion);\n\n    const description = `waitFor: ${textPrompt}`;\n    const session = this.createExecutionSession(\n      taskTitleStr('WaitFor', description),\n    );\n    const runner = session.getRunner();\n    const { timeoutMs, checkIntervalMs } = opt;\n\n    assert(assertion, 'No assertion for waitFor');\n    assert(timeoutMs, 'No timeoutMs for waitFor');\n    assert(checkIntervalMs, 'No checkIntervalMs for waitFor');\n\n    assert(\n      checkIntervalMs <= timeoutMs,\n      `wrong config for waitFor: checkIntervalMs must be less than timeoutMs, config: {checkIntervalMs: ${checkIntervalMs}, timeoutMs: ${timeoutMs}}`,\n    );\n\n    const overallStartTime = Date.now();\n    let lastCheckStart = overallStartTime;\n    let errorThought = '';\n    // Continue checking as long as the previous iteration began within the timeout window.\n    while (lastCheckStart - overallStartTime <= timeoutMs) {\n      const currentCheckStart = Date.now();\n      lastCheckStart = currentCheckStart;\n      const queryTask = await this.createTypeQueryTask(\n        'WaitFor',\n        textPrompt,\n        modelConfig,\n        undefined,\n        multimodalPrompt,\n      );\n\n      const result = (await session.appendAndRun(queryTask)) as\n        | {\n            output: boolean;\n            thought?: string;\n          }\n        | undefined;\n\n      if (result?.output) {\n        return {\n          output: undefined,\n          runner,\n        };\n      }\n\n      errorThought =\n        result?.thought ||\n        (!result && `No result from assertion: ${textPrompt}`) ||\n        `unknown error when waiting for assertion: ${textPrompt}`;\n      const now = Date.now();\n      if (now - currentCheckStart < checkIntervalMs) {\n        const timeRemaining = checkIntervalMs - (now - currentCheckStart);\n        const sleepTask = this.taskBuilder.createSleepTask({\n          timeMs: timeRemaining,\n        });\n        await session.append(sleepTask);\n      }\n    }\n\n    return session.appendErrorPlan(`waitFor timeout: ${errorThought}`);\n  }\n}\n"],"names":["debug","getDebug","defaultReplanningCycleLimit","defaultVlmUiTarsReplanningCycleLimit","maxErrorCountAllowedInOnePlanningLoop","TaskExecutor","title","options","_this_hooks","ExecutionSession","Promise","plans","modelConfigForPlanning","modelConfigForDefaultIntent","userInstruction","yamlString","session","taskTitleStr","task","param","executorContext","uiContext","assert","runner","tasks","result","output","isVlmUiTars","globalConfigManager","MIDSCENE_REPLANNING_CYCLE_LIMIT","userPrompt","includeBboxInPlanning","thinkingStrategy","backgroundKnowledge","cacheable","replanCount","yamlFlow","replanningCycleLimit","errorCountInOnePlanningLoop","startTime","Date","vlMode","uiTarsModelVersion","undefined","actionSpace","action","Array","console","planResult","uiTarsPlanning","plan","JSON","actions","log","more_actions_needed_by_instruction","error","usage","rawResponse","sleep","finalActions","timeNow","timeRemaining","executables","String","Error","errorMsg","finalResult","type","demand","modelConfig","opt","multimodalPrompt","queryTask","taskContext","queryDump","applyDump","dump","ifTypeRestricted","demandInput","keyOfResult","booleanPrompt","extractResult","extraPageDescription","tree","descriptionOfTree","ServiceError","data","thought","outputResult","timeMs","_modelConfig","assertion","textPrompt","parsePrompt","description","timeoutMs","checkIntervalMs","overallStartTime","lastCheckStart","errorThought","currentCheckStart","now","sleepTask","interfaceInstance","service","opts","ConversationHistory","TaskBuilder"],"mappings":";;;;;;;;;;;;;;;;;;;;;AAmDA,MAAMA,QAAQC,SAAS;AACvB,MAAMC,8BAA8B;AACpC,MAAMC,uCAAuC;AAC7C,MAAMC,wCAAwC;AAIvC,MAAMC;IAkBX,IAAI,OAAO;QACT,OAAO,IAAI,CAAC,SAAS;IACvB;IA0BQ,uBACNC,KAAa,EACbC,OAA0C,EAC1C;YAOkBC;QANlB,OAAO,IAAIC,iBACTH,OACA,IAAMI,QAAQ,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,kBAAkB,KACrD;YACE,aAAa,IAAI,CAAC,mBAAmB;YACrC,OAAOH,QAAAA,UAAAA,KAAAA,IAAAA,QAAS,KAAK;YACrB,cAAc,QAAAC,CAAAA,cAAAA,IAAI,CAAC,KAAK,AAAD,IAATA,KAAAA,IAAAA,YAAY,YAAY;QACxC;IAEJ;IAEA,MAAa,wBACXG,KAAuB,EACvBC,sBAAoC,EACpCC,2BAAyC,EACzCN,OAGC,EACD;QACA,OAAO,IAAI,CAAC,WAAW,CAAC,KAAK,CAC3BI,OACAC,wBACAC,6BACAN;IAEJ;IAEA,MAAM,uBAAuBO,eAAuB,EAAEC,UAAkB,EAAE;QACxE,MAAMC,UAAU,IAAI,CAAC,sBAAsB,CACzCC,aAAa,UAAUH;QAGzB,MAAMI,OAAmC;YACvC,MAAM;YACN,SAAS;YACT,OAAO;gBACLJ;YACF;YACA,UAAU,OAAOK,OAAOC;gBACtB,MAAM,EAAEC,SAAS,EAAE,GAAGD;gBACtBE,OAAOD,WAAW;gBAClB,OAAO;oBACL,QAAQ;wBACN,SAAS,EAAE;wBACX,oCAAoC;wBACpC,KAAK;wBACLN;oBACF;oBACA,OAAO;wBACL,KAAK;oBACP;oBACA,OAAO;wBACL,MAAM;wBACN,SAAS;4BACPA;wBACF;oBACF;gBACF;YACF;QACF;QACA,MAAMQ,SAASP,QAAQ,SAAS;QAChC,MAAMA,QAAQ,YAAY,CAACE;QAE3B,OAAO;YACLK;QACF;IACF;IAEA,MAAM,SACJjB,KAAa,EACbK,KAAuB,EACvBC,sBAAoC,EACpCC,2BAAyC,EACf;QAC1B,MAAMG,UAAU,IAAI,CAAC,sBAAsB,CAACV;QAC5C,MAAM,EAAEkB,KAAK,EAAE,GAAG,MAAM,IAAI,CAAC,uBAAuB,CAClDb,OACAC,wBACAC;QAEF,MAAMU,SAASP,QAAQ,SAAS;QAChC,MAAMS,SAAS,MAAMT,QAAQ,YAAY,CAACQ;QAC1C,MAAM,EAAEE,MAAM,EAAE,GAAGD,UAAU,CAAC;QAC9B,OAAO;YACLC;YACAH;QACF;IACF;IAEQ,wBAAwBI,WAAoB,EAAE;QACpD,OACE,IAAI,CAAC,oBAAoB,IACzBC,oBAAoB,oBAAoB,CACtCC,oCAEDF,CAAAA,cACGxB,uCACAD,2BAA0B;IAElC;IAEA,MAAM,OACJ4B,UAAkB,EAClBlB,sBAAoC,EACpCC,2BAAyC,EACzCkB,qBAA8B,EAC9BC,gBAAkC,EAClCC,mBAA4B,EAC5BC,SAAmB,EAQnB;QACA,IAAI,CAAC,mBAAmB,CAAC,KAAK;QAE9B,MAAMlB,UAAU,IAAI,CAAC,sBAAsB,CACzCC,aAAa,UAAUa;QAEzB,MAAMP,SAASP,QAAQ,SAAS;QAEhC,IAAImB,cAAc;QAClB,MAAMC,WAAmC,EAAE;QAC3C,MAAMC,uBAAuB,IAAI,CAAC,uBAAuB,CACvDzB,AAAkC,kBAAlCA,uBAAuB,MAAM;QAG/B,IAAI0B,8BAA8B;QAGlC,MAAO,KAAM;YACX,MAAMb,SAAS,MAAMT,QAAQ,YAAY,CACvC;gBACE,MAAM;gBACN,SAAS;gBACT,OAAO;oBACL,iBAAiBc;oBACjB,iBAAiBG;gBACnB;gBACA,UAAU,OAAOd,OAAOC;oBACtB,MAAMmB,YAAYC,KAAK,GAAG;oBAC1B,MAAM,EAAEnB,SAAS,EAAE,GAAGD;oBACtBE,OAAOD,WAAW;oBAClB,MAAM,EAAEoB,MAAM,EAAE,GAAG7B;oBACnB,MAAM8B,qBACJD,AAAW,kBAAXA,SACI7B,uBAAuB,kBAAkB,GACzC+B;oBAENrB,OACE,IAAI,CAAC,SAAS,CAAC,WAAW,EAC1B;oBAEF,MAAMsB,cAAc,MAAM,IAAI,CAAC,SAAS,CAAC,WAAW;oBACpD5C,MACE,sCACA4C,YAAY,GAAG,CAAC,CAACC,SAAWA,OAAO,IAAI,EAAE,IAAI,CAAC;oBAEhDvB,OAAOwB,MAAM,OAAO,CAACF,cAAc;oBACnC,IAAIA,AAAuB,MAAvBA,YAAY,MAAM,EACpBG,QAAQ,IAAI,CACV,CAAC,gBAAgB,EAAE,IAAI,CAAC,SAAS,CAAC,aAAa,CAAC,gDAAgD,CAAC;oBAIrG,MAAMC,aAAa,MAAON,AAAAA,CAAAA,qBACtBO,iBACAC,IAAG,EAAG/B,MAAM,eAAe,EAAE;wBAC/B,SAASE;wBACT,eAAeF,MAAM,eAAe;wBACpC,eAAe,IAAI,CAAC,SAAS,CAAC,aAAa;wBAC3CyB;wBACA,aAAahC;wBACb,qBAAqB,IAAI,CAAC,mBAAmB;wBAC7C,aAAamB;wBACbC;oBACF;oBACAhC,MAAM,cAAcmD,KAAK,SAAS,CAACH,YAAY,MAAM;oBAErD,MAAM,EACJI,OAAO,EACPC,GAAG,EACHC,kCAAkC,EAClCC,KAAK,EACLC,KAAK,EACLC,WAAW,EACXC,KAAK,EACN,GAAGV;oBAEJ5B,gBAAgB,IAAI,CAAC,GAAG,GAAG;wBACzB,GAAIA,gBAAgB,IAAI,CAAC,GAAG,IAAI,CAAC,CAAC;wBAClCqC;oBACF;oBACArC,gBAAgB,IAAI,CAAC,KAAK,GAAGoC;oBAE7B,MAAMG,eAAeP,WAAW,EAAE;oBAElC,IAAIM,OAAO;wBACT,MAAME,UAAUpB,KAAK,GAAG;wBACxB,MAAMqB,gBAAgBH,QAASE,CAAAA,UAAUrB,SAAQ;wBACjD,IAAIsB,gBAAgB,GAClBF,aAAa,IAAI,CAAC,IAAI,CAAC,SAAS,CAACE;oBAErC;oBAEA,IAAIF,AAAwB,MAAxBA,aAAa,MAAM,EACrBrC,OACE,CAACgC,sCAAsCI,OACvCH,QAAQ,CAAC,gBAAgB,EAAEA,OAAO,GAAG;oBAIzC,OAAO;wBACL,QAAQ;4BACN,SAASI;4BACTL;4BACAD;4BACA,UAAUL,WAAW,QAAQ;wBAC/B;wBACA,OAAO;4BACL,KAAK;wBACP;wBACA3B;oBACF;gBACF;YACF,GACA;gBACE,gBAAgB;YAClB;YAGF,MAAM2B,aAAavB,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,MAAM;YAGjC,MAAMd,QAAQqC,AAAAA,CAAAA,QAAAA,aAAAA,KAAAA,IAAAA,WAAY,OAAO,AAAD,KAAK,EAAE;YACvCZ,SAAS,IAAI,IAAKY,AAAAA,CAAAA,QAAAA,aAAAA,KAAAA,IAAAA,WAAY,QAAQ,AAAD,KAAK,EAAE;YAE5C,IAAIc;YACJ,IAAI;gBACFA,cAAc,MAAM,IAAI,CAAC,uBAAuB,CAC9CnD,OACAC,wBACAC,6BACA;oBACEqB;oBACA,SAAS;gBACX;YAEJ,EAAE,OAAOqB,OAAO;gBACd,OAAOvC,QAAQ,eAAe,CAC5B,CAAC,4CAA4C,EAAEuC,MAAM,SAAS,EAAEJ,KAAK,SAAS,CAC5ExC,QACC;YAEP;YACA,IAAI,IAAI,CAAC,mBAAmB,CAAC,sBAAsB,EACjDoC,QAAQ,IAAI,CACV,8FACA,IAAI,CAAC,mBAAmB,CAAC,sBAAsB;YAGnD,IAAI;gBACF,MAAM/B,QAAQ,YAAY,CAAC8C,YAAY,KAAK;YAC9C,EAAE,OAAOP,OAAY;gBACnBjB;gBACA,IAAI,CAAC,mBAAmB,CAAC,sBAAsB,GAAG,CAAC,+BAA+B,EAAEiB,AAAAA,CAAAA,QAAAA,QAAAA,KAAAA,IAAAA,MAAO,OAAO,AAAD,KAAKQ,OAAOR,QAAQ;gBACrHvD,MACE,4DACAuD,iBAAiBS,QAAQT,MAAM,OAAO,GAAGQ,OAAOR;YAEpD;YAEA,IAAIjB,8BAA8BlC,uCAChC,OAAOY,QAAQ,eAAe,CAAC;YAIjC,IAAI,CAACgC,CAAAA,QAAAA,aAAAA,KAAAA,IAAAA,WAAY,kCAAkC,AAAD,GAChD;YAIF,EAAEb;YAEF,IAAIA,cAAcE,sBAAsB;gBACtC,MAAM4B,WAAW,CAAC,WAAW,EAAE5B,qBAAqB,+EAA+E,CAAC;gBACpI,OAAOrB,QAAQ,eAAe,CAACiD;YACjC;YAEA,IAAI,CAAC,IAAI,CAAC,mBAAmB,CAAC,sBAAsB,EAClD,IAAI,CAAC,mBAAmB,CAAC,sBAAsB,GAC7C;QAEN;QAEA,MAAMC,cAAc;YAClB,QAAQ;gBACN9B;YACF;YACAb;QACF;QACA,OAAO2C;IACT;IAEQ,oBACNC,IAAsE,EACtEC,MAA2B,EAC3BC,WAAyB,EACzBC,GAA0B,EAC1BC,gBAAoC,EACpC;QACA,MAAMC,YAA4C;YAChD,MAAM;YACN,SAASL;YACT,OAAO;gBACL,YAAYI,mBACP;oBACCH;oBACAG;gBACF,IACAH;YACN;YACA,UAAU,OAAOjD,OAAOsD;gBACtB,MAAM,EAAEvD,IAAI,EAAE,GAAGuD;gBACjB,IAAIC;gBACJ,MAAMC,YAAY,CAACC;oBACjBF,YAAYE;oBACZ1D,KAAK,GAAG,GAAG;wBACT0D;oBACF;gBACF;gBAGA,MAAMvD,YAAYoD,YAAY,SAAS;gBACvCnD,OAAOD,WAAW;gBAElB,MAAMwD,mBAAmBV,AAAS,YAATA;gBACzB,IAAIW,cAAcV;gBAClB,IAAIW,cAAc;gBAClB,IAAIF,oBAAqBV,CAAAA,AAAS,aAATA,QAAqBA,AAAS,cAATA,IAAiB,GAAI;oBACjEY,cAAc;oBACd,MAAMC,gBACJb,AAAS,aAATA,OACI,CAAC,kDAAkD,EAAEC,QAAQ,GAC7D,CAAC,+GAA+G,EAAEA,QAAQ;oBAChIU,cAAc;wBACZ,CAACC,YAAY,EAAEC;oBACjB;gBACF,OAAO,IAAIH,kBACTC,cAAc;oBACZ,CAACC,YAAY,EAAE,GAAGZ,KAAK,EAAE,EAAEC,QAAQ;gBACrC;gBAGF,IAAIa;gBAEJ,IAAIC,uBAAuB;gBAC3B,IAAIZ,AAAAA,CAAAA,QAAAA,MAAAA,KAAAA,IAAAA,IAAK,WAAW,AAAD,KAAK,IAAI,CAAC,SAAS,CAAC,mBAAmB,EAAE;oBAC1DtE,MAAM;oBACN,MAAMmF,OAAO,MAAM,IAAI,CAAC,SAAS,CAAC,mBAAmB;oBACrDD,uBAAuB,MAAME,kBAC3BD,MACA,KACA,OACAb,AAAAA,CAAAA,QAAAA,MAAAA,KAAAA,IAAAA,IAAK,WAAW,AAAD,MAAM;gBAEzB;gBAEA,IAAI;oBACFW,gBAAgB,MAAM,IAAI,CAAC,OAAO,CAAC,OAAO,CACxCH,aACAT,aACAC,KACAY,sBACAX;gBAEJ,EAAE,OAAOhB,OAAO;oBACd,IAAIA,iBAAiB8B,cACnBV,UAAUpB,MAAM,IAAI;oBAEtB,MAAMA;gBACR;gBAEA,MAAM,EAAE+B,IAAI,EAAE9B,KAAK,EAAE+B,OAAO,EAAEX,IAAI,EAAE,GAAGK;gBACvCN,UAAUC;gBAEV,IAAIY,eAAeF;gBACnB,IAAIT,kBAEF,IAAI,AAAgB,YAAhB,OAAOS,MACTE,eAAeF;qBACV,IAAInB,AAAS,cAATA,MAEPqB,eADEF,QAAAA,OACa,QAECA,IAAY,CAACP,YAAY;qBAEtC,IAAIO,QAAAA,MACTE,eAAe;qBACV;oBACLlE,OACEgE,AAAAA,CAAAA,QAAAA,OAAAA,KAAAA,IAAAA,IAAM,CAACP,YAAY,AAAD,MAAMpC,QACxB;oBAEF6C,eAAgBF,IAAY,CAACP,YAAY;gBAC3C;gBAGF,IAAIZ,AAAS,aAATA,QAAqB,CAACqB,cAAc;oBACtCtE,KAAK,KAAK,GAAGsC;oBACbtC,KAAK,OAAO,GAAGqE;oBACf,MAAM,IAAIvB,MAAM,CAAC,kBAAkB,EAAEuB,SAAS;gBAChD;gBAEA,OAAO;oBACL,QAAQC;oBACR,KAAKd;oBACLlB;oBACA+B;gBACF;YACF;QACF;QAEA,OAAOf;IACT;IACA,MAAM,yBACJL,IAA0D,EAC1DC,MAA2B,EAC3BC,WAAyB,EACzBC,GAA0B,EAC1BC,gBAAoC,EACP;QAC7B,MAAMvD,UAAU,IAAI,CAAC,sBAAsB,CACzCC,aACEkD,MACA,AAAkB,YAAlB,OAAOC,SAAsBA,SAASjB,KAAK,SAAS,CAACiB;QAIzD,MAAMI,YAAY,MAAM,IAAI,CAAC,mBAAmB,CAC9CL,MACAC,QACAC,aACAC,KACAC;QAGF,MAAMhD,SAASP,QAAQ,SAAS;QAChC,MAAMS,SAAS,MAAMT,QAAQ,YAAY,CAACwD;QAE1C,IAAI,CAAC/C,QACH,MAAM,IAAIuC,MACR;QAIJ,MAAM,EAAEtC,MAAM,EAAE6D,OAAO,EAAE,GAAG9D;QAE5B,OAAO;YACLC;YACA6D;YACAhE;QACF;IACF;IAEQ,UAAUkE,MAAc,EAA4C;QAC1E,OAAO;YACL,MAAM;YACN,OAAO;gBACLA;YACF;QACF;IACF;IAEA,MAAM,aAAaA,MAAc,EAAEC,YAA0B,EAAE;QAC7D,OAAO,IAAI,CAAC,WAAW,CAAC,eAAe,CAAC;YACtCD;QACF;IACF;IAEA,MAAM,QACJE,SAAsB,EACtBrB,GAA+B,EAC/BD,WAAyB,EACO;QAChC,MAAM,EAAEuB,UAAU,EAAErB,gBAAgB,EAAE,GAAGsB,YAAYF;QAErD,MAAMG,cAAc,CAAC,SAAS,EAAEF,YAAY;QAC5C,MAAM5E,UAAU,IAAI,CAAC,sBAAsB,CACzCC,aAAa,WAAW6E;QAE1B,MAAMvE,SAASP,QAAQ,SAAS;QAChC,MAAM,EAAE+E,SAAS,EAAEC,eAAe,EAAE,GAAG1B;QAEvChD,OAAOqE,WAAW;QAClBrE,OAAOyE,WAAW;QAClBzE,OAAO0E,iBAAiB;QAExB1E,OACE0E,mBAAmBD,WACnB,CAAC,iGAAiG,EAAEC,gBAAgB,aAAa,EAAED,UAAU,CAAC,CAAC;QAGjJ,MAAME,mBAAmBzD,KAAK,GAAG;QACjC,IAAI0D,iBAAiBD;QACrB,IAAIE,eAAe;QAEnB,MAAOD,iBAAiBD,oBAAoBF,UAAW;YACrD,MAAMK,oBAAoB5D,KAAK,GAAG;YAClC0D,iBAAiBE;YACjB,MAAM5B,YAAY,MAAM,IAAI,CAAC,mBAAmB,CAC9C,WACAoB,YACAvB,aACA1B,QACA4B;YAGF,MAAM9C,SAAU,MAAMT,QAAQ,YAAY,CAACwD;YAO3C,IAAI/C,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,MAAM,EAChB,OAAO;gBACL,QAAQkB;gBACRpB;YACF;YAGF4E,eACE1E,AAAAA,CAAAA,QAAAA,SAAAA,KAAAA,IAAAA,OAAQ,OAAO,AAAD,KACb,CAACA,UAAU,CAAC,0BAA0B,EAAEmE,YAAY,IACrD,CAAC,0CAA0C,EAAEA,YAAY;YAC3D,MAAMS,MAAM7D,KAAK,GAAG;YACpB,IAAI6D,MAAMD,oBAAoBJ,iBAAiB;gBAC7C,MAAMnC,gBAAgBmC,kBAAmBK,CAAAA,MAAMD,iBAAgB;gBAC/D,MAAME,YAAY,IAAI,CAAC,WAAW,CAAC,eAAe,CAAC;oBACjD,QAAQzC;gBACV;gBACA,MAAM7C,QAAQ,MAAM,CAACsF;YACvB;QACF;QAEA,OAAOtF,QAAQ,eAAe,CAAC,CAAC,iBAAiB,EAAEmF,cAAc;IACnE;IAnkBA,YACEI,iBAAoC,EACpCC,OAAgB,EAChBC,IAKC,CACD;QA9BF;QAEA;QAEA;QAEA,uBAAiB,eAAjB;QAEA,uBAAQ,uBAAR;QAEA;QAEA,uBAAiB,SAAjB;QAEA;QAiBE,IAAI,CAAC,SAAS,GAAGF;QACjB,IAAI,CAAC,OAAO,GAAGC;QACf,IAAI,CAAC,SAAS,GAAGC,KAAK,SAAS;QAC/B,IAAI,CAAC,mBAAmB,GAAGA,QAAAA,OAAAA,KAAAA,IAAAA,KAAM,WAAW;QAC5C,IAAI,CAAC,oBAAoB,GAAGA,KAAK,oBAAoB;QACrD,IAAI,CAAC,KAAK,GAAGA,KAAK,KAAK;QACvB,IAAI,CAAC,mBAAmB,GAAG,IAAIC;QAC/B,IAAI,CAAC,WAAW,GAAG,IAAIC,YAAY;YACjCJ;YACAC;YACA,WAAWC,KAAK,SAAS;QAC3B;IACF;AA8iBF"}

package/dist/es/agent/utils.mjs

@@ -103,7 +103,7 @@ async function matchElementFromCache(context, cacheEntry, cachePrompt, cacheable
         return;
     }
 }
-const getMidsceneVersion = ()=>"1.0.1-beta-20251126084350.0";
+const getMidsceneVersion = ()=>"1.0.1-beta-20251126092253.0";
 const parsePrompt = (prompt)=>{
     if ('string' == typeof prompt) return {
         textPrompt: prompt,

package/dist/es/ai-model/conversation-history.mjs

@@ -12,6 +12,9 @@ var _computedKey;
 _computedKey = Symbol.iterator;
 let _computedKey1 = _computedKey;
 class ConversationHistory {
+    resetPendingFeedbackMessageIfExists() {
+        if (this.pendingFeedbackMessage) this.pendingFeedbackMessage = '';
+    }
     append(message) {
         this.messages.push(message);
     }
@@ -41,7 +44,9 @@ class ConversationHistory {
     constructor(options){
         var _options_initialMessages;
         _define_property(this, "messages", []);
+        _define_property(this, "pendingFeedbackMessage", void 0);
         if (null == options ? void 0 : null == (_options_initialMessages = options.initialMessages) ? void 0 : _options_initialMessages.length) this.seed(options.initialMessages);
+        this.pendingFeedbackMessage = '';
     }
 }
 export { ConversationHistory };

package/dist/es/ai-model/conversation-history.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"ai-model/conversation-history.mjs","sources":["webpack://@midscene/core/./src/ai-model/conversation-history.ts"],"sourcesContent":["import type { ChatCompletionMessageParam } from 'openai/resources/index';\n\nexport interface ConversationHistoryOptions {\n  initialMessages?: ChatCompletionMessageParam[];\n}\n\nexport class ConversationHistory {\n  private readonly messages: ChatCompletionMessageParam[] = [];\n\n  constructor(options?: ConversationHistoryOptions) {\n    if (options?.initialMessages?.length) {\n      this.seed(options.initialMessages);\n    }\n  }\n\n  append(message: ChatCompletionMessageParam) {\n    this.messages.push(message);\n  }\n\n  seed(messages: ChatCompletionMessageParam[]) {\n    this.reset();\n    messages.forEach((message) => {\n      this.append(message);\n    });\n  }\n\n  reset() {\n    this.messages.length = 0;\n  }\n\n  snapshot(): ChatCompletionMessageParam[] {\n    return [...this.messages];\n  }\n\n  get length(): number {\n    return this.messages.length;\n  }\n\n  [Symbol.iterator](): IterableIterator<ChatCompletionMessageParam> {\n    return this.messages[Symbol.iterator]();\n  }\n\n  toJSON(): ChatCompletionMessageParam[] {\n    return this.snapshot();\n  }\n}\n"],"names":["Symbol","ConversationHistory","message","messages","options","_options_initialMessages"],"mappings":";;;;;;;;;;;eAsCGA,OAAO,QAAQ;;AAhCX,MAAMC;IASX,OAAOC,OAAmC,EAAE;QAC1C,IAAI,CAAC,QAAQ,CAAC,IAAI,CAACA;IACrB;IAEA,KAAKC,QAAsC,EAAE;QAC3C,IAAI,CAAC,KAAK;QACVA,SAAS,OAAO,CAAC,CAACD;YAChB,IAAI,CAAC,MAAM,CAACA;QACd;IACF;IAEA,QAAQ;QACN,IAAI,CAAC,QAAQ,CAAC,MAAM,GAAG;IACzB;IAEA,WAAyC;QACvC,OAAO;eAAI,IAAI,CAAC,QAAQ;SAAC;IAC3B;IAEA,IAAI,SAAiB;QACnB,OAAO,IAAI,CAAC,QAAQ,CAAC,MAAM;IAC7B;IAEA,CAAC,cAAD,GAAkE;QAChE,OAAO,IAAI,CAAC,QAAQ,CAACF,OAAO,QAAQ,CAAC;IACvC;IAEA,SAAuC;QACrC,OAAO,IAAI,CAAC,QAAQ;IACtB;IAnCA,YAAYI,OAAoC,CAAE;YAC5CC;QAHN,uBAAiB,YAAyC,EAAE;QAG1D,IAAIA,QAAAA,UAAAA,KAAAA,IAAAA,QAAAA,CAAAA,2BAAAA,QAAS,eAAe,AAAD,IAAvBA,KAAAA,IAAAA,yBAA0B,MAAM,EAClC,IAAI,CAAC,IAAI,CAACD,QAAQ,eAAe;IAErC;AAgCF"}
+{"version":3,"file":"ai-model/conversation-history.mjs","sources":["webpack://@midscene/core/./src/ai-model/conversation-history.ts"],"sourcesContent":["import type { ChatCompletionMessageParam } from 'openai/resources/index';\n\nexport interface ConversationHistoryOptions {\n  initialMessages?: ChatCompletionMessageParam[];\n}\n\nexport class ConversationHistory {\n  private readonly messages: ChatCompletionMessageParam[] = [];\n\n  public pendingFeedbackMessage: string;\n\n  constructor(options?: ConversationHistoryOptions) {\n    if (options?.initialMessages?.length) {\n      this.seed(options.initialMessages);\n    }\n    this.pendingFeedbackMessage = '';\n  }\n\n  resetPendingFeedbackMessageIfExists() {\n    if (this.pendingFeedbackMessage) {\n      this.pendingFeedbackMessage = '';\n    }\n  }\n\n  append(message: ChatCompletionMessageParam) {\n    this.messages.push(message);\n  }\n\n  seed(messages: ChatCompletionMessageParam[]) {\n    this.reset();\n    messages.forEach((message) => {\n      this.append(message);\n    });\n  }\n\n  reset() {\n    this.messages.length = 0;\n  }\n\n  snapshot(): ChatCompletionMessageParam[] {\n    return [...this.messages];\n  }\n\n  get length(): number {\n    return this.messages.length;\n  }\n\n  [Symbol.iterator](): IterableIterator<ChatCompletionMessageParam> {\n    return this.messages[Symbol.iterator]();\n  }\n\n  toJSON(): ChatCompletionMessageParam[] {\n    return this.snapshot();\n  }\n}\n"],"names":["Symbol","ConversationHistory","message","messages","options","_options_initialMessages"],"mappings":";;;;;;;;;;;eA+CGA,OAAO,QAAQ;;AAzCX,MAAMC;IAYX,sCAAsC;QACpC,IAAI,IAAI,CAAC,sBAAsB,EAC7B,IAAI,CAAC,sBAAsB,GAAG;IAElC;IAEA,OAAOC,OAAmC,EAAE;QAC1C,IAAI,CAAC,QAAQ,CAAC,IAAI,CAACA;IACrB;IAEA,KAAKC,QAAsC,EAAE;QAC3C,IAAI,CAAC,KAAK;QACVA,SAAS,OAAO,CAAC,CAACD;YAChB,IAAI,CAAC,MAAM,CAACA;QACd;IACF;IAEA,QAAQ;QACN,IAAI,CAAC,QAAQ,CAAC,MAAM,GAAG;IACzB;IAEA,WAAyC;QACvC,OAAO;eAAI,IAAI,CAAC,QAAQ;SAAC;IAC3B;IAEA,IAAI,SAAiB;QACnB,OAAO,IAAI,CAAC,QAAQ,CAAC,MAAM;IAC7B;IAEA,CAAC,cAAD,GAAkE;QAChE,OAAO,IAAI,CAAC,QAAQ,CAACF,OAAO,QAAQ,CAAC;IACvC;IAEA,SAAuC;QACrC,OAAO,IAAI,CAAC,QAAQ;IACtB;IA1CA,YAAYI,OAAoC,CAAE;YAC5CC;QALN,uBAAiB,YAAyC,EAAE;QAE5D,uBAAO,0BAAP;QAGE,IAAIA,QAAAA,UAAAA,KAAAA,IAAAA,QAAAA,CAAAA,2BAAAA,QAAS,eAAe,AAAD,IAAvBA,KAAAA,IAAAA,yBAA0B,MAAM,EAClC,IAAI,CAAC,IAAI,CAACD,QAAQ,eAAe;QAEnC,IAAI,CAAC,sBAAsB,GAAG;IAChC;AAsCF"}

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

@@ -6,7 +6,6 @@ import { systemPromptToTaskPlanning } from "./prompt/llm-planning.mjs";
 import { callAIWithObjectResponse } from "./service-caller/index.mjs";
 const debug = getDebug('planning');
 async function plan(userInstruction, opts) {
-    var _opts_conversationHistory;
     const { context, modelConfig, conversationHistory } = opts;
     const { screenshotBase64, size } = context;
     const { vlMode } = modelConfig;
@@ -27,35 +26,42 @@ async function plan(userInstruction, opts) {
         imageHeight = paddedResult.height;
         imagePayload = paddedResult.imageBase64;
     }
-    const historyLog = (null == (_opts_conversationHistory = opts.conversationHistory) ? void 0 : _opts_conversationHistory.snapshot()) || [];
-    const knowledgeContext = opts.actionContext ? [
+    const instruction = [
         {
             role: 'user',
             content: [
                 {
                     type: 'text',
-                    text: `<high_priority_knowledge>${opts.actionContext}</high_priority_knowledge>`
+                    text: `<high_priority_knowledge>${opts.actionContext}</high_priority_knowledge>\n<user_instruction>${userInstruction}</user_instruction>`
                 }
             ]
         }
-    ] : [];
-    const instruction = [
-        {
+    ];
+    let latestFeedbackMessage;
+    if (conversationHistory.pendingFeedbackMessage) {
+        latestFeedbackMessage = {
             role: 'user',
             content: [
                 {
                     type: 'text',
-                    text: `<user_instruction>${userInstruction}</user_instruction>`
+                    text: `${conversationHistory.pendingFeedbackMessage}. The last screenshot is attached. Please going on according to the instruction.`
+                },
+                {
+                    type: 'image_url',
+                    image_url: {
+                        url: imagePayload,
+                        detail: 'high'
+                    }
                 }
             ]
-        }
-    ];
-    const latestImageMessage = {
+        };
+        conversationHistory.resetPendingFeedbackMessageIfExists();
+    } else latestFeedbackMessage = {
         role: 'user',
         content: [
             {
                 type: 'text',
-                text: 'I have finished the action previously planned, and the last screenshot is attached. Please going on according to the instruction.'
+                text: 'this is the latest screenshot'
             },
             {
                 type: 'image_url',
@@ -66,15 +72,15 @@ async function plan(userInstruction, opts) {
             }
         ]
     };
+    conversationHistory.append(latestFeedbackMessage);
+    const historyLog = conversationHistory.snapshot();
     const msgs = [
         {
             role: 'system',
             content: systemPrompt
         },
-        ...knowledgeContext,
         ...instruction,
-        ...historyLog,
-        latestImageMessage
+        ...historyLog
     ];
     const { content: planFromAI, contentString: rawResponse, usage } = await callAIWithObjectResponse(msgs, AIActionType.PLAN, modelConfig);
     const actions = planFromAI.action ? [
@@ -101,8 +107,7 @@ async function plan(userInstruction, opts) {
     });
     assert(!planFromAI.error, `Failed to plan actions: ${planFromAI.error}`);
     if (0 === actions.length && returnValue.more_actions_needed_by_instruction && !returnValue.sleep) console.warn('No actions planned for the prompt, but model said more actions are needed:', userInstruction);
-    null == conversationHistory || conversationHistory.append(latestImageMessage);
-    null == conversationHistory || conversationHistory.append({
+    conversationHistory.append({
         role: 'assistant',
         content: [
             {

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

@@ -1 +1 @@
-{"version":3,"file":"ai-model/llm-planning.mjs","sources":["webpack://@midscene/core/./src/ai-model/llm-planning.ts"],"sourcesContent":["import type {\n  DeviceAction,\n  InterfaceType,\n  PlanningAIResponse,\n  RawResponsePlanningAIResponse,\n  ThinkingStrategy,\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    thinkingStrategy: ThinkingStrategy;\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    thinkingStrategy: opts.thinkingStrategy,\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\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 latestImageMessage: ChatCompletionMessageParam = {\n    role: 'user',\n    content: [\n      {\n        type: 'text',\n        text: 'I have finished the action previously planned, and the last screenshot is attached. Please going on according to the instruction.',\n      },\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  const msgs: ChatCompletionMessageParam[] = [\n    { role: 'system', content: systemPrompt },\n    ...knowledgeContext,\n    ...instruction,\n    ...historyLog,\n    latestImageMessage,\n  ];\n\n  const {\n    content: planFromAI,\n    contentString: rawResponse,\n    usage,\n  } = await callAIWithObjectResponse<RawResponsePlanningAIResponse>(\n    msgs,\n    AIActionType.PLAN,\n    modelConfig,\n  );\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(latestImageMessage);\n\n  conversationHistory?.append({\n    role: 'assistant',\n    content: [\n      {\n        type: 'text',\n        text: rawResponse,\n      },\n    ],\n  });\n\n  return returnValue;\n}\n"],"names":["debug","getDebug","plan","userInstruction","opts","_opts_conversationHistory","context","modelConfig","conversationHistory","screenshotBase64","size","vlMode","systemPrompt","systemPromptToTaskPlanning","imagePayload","imageWidth","imageHeight","rightLimit","bottomLimit","paddedResult","paddingToMatchBlockByBase64","historyLog","knowledgeContext","instruction","latestImageMessage","msgs","planFromAI","rawResponse","usage","callAIWithObjectResponse","AIActionType","actions","returnValue","buildYamlFlowFromPlans","assert","action","type","actionInActionSpace","locateFields","findAllMidsceneLocatorField","field","locateResult","undefined","fillBboxParam","console"],"mappings":";;;;;;AAuBA,MAAMA,QAAQC,SAAS;AAEhB,eAAeC,KACpBC,eAAuB,EACvBC,IASC;QA4BkBC;IA1BnB,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,2BAA2B;QACpD,aAAaT,KAAK,WAAW;QAC7BO;QACA,aAAaP,KAAK,WAAW;QAC7B,kBAAkBA,KAAK,gBAAgB;IACzC;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,4BAA4BN;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;IAE7D,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,qBAAiD;QACrD,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAM;YACR;YACA;gBACE,MAAM;gBACN,WAAW;oBACT,KAAKV;oBACL,QAAQ;gBACV;YACF;SAED;IACH;IAEA,MAAMW,OAAqC;QACzC;YAAE,MAAM;YAAU,SAASb;QAAa;WACrCU;WACAC;WACAF;QACHG;KACD;IAED,MAAM,EACJ,SAASE,UAAU,EACnB,eAAeC,WAAW,EAC1BC,KAAK,EACN,GAAG,MAAMC,yBACRJ,MACAK,aAAa,IAAI,EACjBvB;IAGF,MAAMwB,UAAUL,WAAW,MAAM,GAAG;QAACA,WAAW,MAAM;KAAC,GAAG,EAAE;IAC5D,MAAMM,cAAkC;QACtC,GAAGN,UAAU;QACbK;QACAJ;QACAC;QACA,UAAUK,uBACRF,SACA3B,KAAK,WAAW,EAChBsB,WAAW,KAAK;IAEpB;IAEAQ,OAAOR,YAAY;IAEnBK,QAAQ,OAAO,CAAC,CAACI;QACf,MAAMC,OAAOD,OAAO,IAAI;QACxB,MAAME,sBAAsBjC,KAAK,WAAW,CAAC,IAAI,CAC/C,CAAC+B,SAAWA,OAAO,IAAI,KAAKC;QAG9BpC,MAAM,+BAA+BqC;QACrC,MAAMC,eAAeD,sBACjBE,4BAA4BF,oBAAoB,WAAW,IAC3D,EAAE;QAENrC,MAAM,gBAAgBsC;QAEtBA,aAAa,OAAO,CAAC,CAACE;YACpB,MAAMC,eAAeN,OAAO,KAAK,CAACK,MAAM;YACxC,IAAIC,gBAAgB9B,AAAW+B,WAAX/B,QAElBwB,OAAO,KAAK,CAACK,MAAM,GAAGG,cACpBF,cACA1B,YACAC,aACAC,YACAC,aACAP;QAGN;IACF;IAEAuB,OAAO,CAACR,WAAW,KAAK,EAAE,CAAC,wBAAwB,EAAEA,WAAW,KAAK,EAAE;IAEvE,IACEK,AAAmB,MAAnBA,QAAQ,MAAM,IACdC,YAAY,kCAAkC,IAC9C,CAACA,YAAY,KAAK,EAElBY,QAAQ,IAAI,CACV,8EACAzC;IAIJK,QAAAA,uBAAAA,oBAAqB,MAAM,CAACgB;IAE5BhB,QAAAA,uBAAAA,oBAAqB,MAAM,CAAC;QAC1B,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAMmB;YACR;SACD;IACH;IAEA,OAAOK;AACT"}
+{"version":3,"file":"ai-model/llm-planning.mjs","sources":["webpack://@midscene/core/./src/ai-model/llm-planning.ts"],"sourcesContent":["import type {\n  DeviceAction,\n  InterfaceType,\n  PlanningAIResponse,\n  RawResponsePlanningAIResponse,\n  ThinkingStrategy,\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    thinkingStrategy: ThinkingStrategy;\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    thinkingStrategy: opts.thinkingStrategy,\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 instruction: ChatCompletionMessageParam[] = [\n    {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `<high_priority_knowledge>${opts.actionContext}</high_priority_knowledge>\\n<user_instruction>${userInstruction}</user_instruction>`,\n        },\n      ],\n    },\n  ];\n\n  let latestFeedbackMessage: ChatCompletionMessageParam;\n\n  if (conversationHistory.pendingFeedbackMessage) {\n    latestFeedbackMessage = {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: `${conversationHistory.pendingFeedbackMessage}. The last screenshot is attached. Please going on according to the instruction.`,\n        },\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n      ],\n    };\n\n    conversationHistory.resetPendingFeedbackMessageIfExists();\n  } else {\n    latestFeedbackMessage = {\n      role: 'user',\n      content: [\n        {\n          type: 'text',\n          text: 'this is the latest screenshot',\n        },\n        {\n          type: 'image_url',\n          image_url: {\n            url: imagePayload,\n            detail: 'high',\n          },\n        },\n      ],\n    };\n  }\n  conversationHistory.append(latestFeedbackMessage);\n  const historyLog = conversationHistory.snapshot();\n\n  const msgs: ChatCompletionMessageParam[] = [\n    { role: 'system', content: systemPrompt },\n    ...instruction,\n    ...historyLog,\n  ];\n\n  const {\n    content: planFromAI,\n    contentString: rawResponse,\n    usage,\n  } = await callAIWithObjectResponse<RawResponsePlanningAIResponse>(\n    msgs,\n    AIActionType.PLAN,\n    modelConfig,\n  );\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\n  return returnValue;\n}\n"],"names":["debug","getDebug","plan","userInstruction","opts","context","modelConfig","conversationHistory","screenshotBase64","size","vlMode","systemPrompt","systemPromptToTaskPlanning","imagePayload","imageWidth","imageHeight","rightLimit","bottomLimit","paddedResult","paddingToMatchBlockByBase64","instruction","latestFeedbackMessage","historyLog","msgs","planFromAI","rawResponse","usage","callAIWithObjectResponse","AIActionType","actions","returnValue","buildYamlFlowFromPlans","assert","action","type","actionInActionSpace","locateFields","findAllMidsceneLocatorField","field","locateResult","undefined","fillBboxParam","console"],"mappings":";;;;;;AAuBA,MAAMA,QAAQC,SAAS;AAEhB,eAAeC,KACpBC,eAAuB,EACvBC,IASC;IAED,MAAM,EAAEC,OAAO,EAAEC,WAAW,EAAEC,mBAAmB,EAAE,GAAGH;IACtD,MAAM,EAAEI,gBAAgB,EAAEC,IAAI,EAAE,GAAGJ;IAEnC,MAAM,EAAEK,MAAM,EAAE,GAAGJ;IAEnB,MAAMK,eAAe,MAAMC,2BAA2B;QACpD,aAAaR,KAAK,WAAW;QAC7BM;QACA,aAAaN,KAAK,WAAW;QAC7B,kBAAkBA,KAAK,gBAAgB;IACzC;IAEA,IAAIS,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,4BAA4BN;QACvDC,aAAaI,aAAa,KAAK;QAC/BH,cAAcG,aAAa,MAAM;QACjCL,eAAeK,aAAa,WAAW;IACzC;IAEA,MAAME,cAA4C;QAChD;YACE,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,CAAC,yBAAyB,EAAEhB,KAAK,aAAa,CAAC,8CAA8C,EAAED,gBAAgB,mBAAmB,CAAC;gBAC3I;aACD;QACH;KACD;IAED,IAAIkB;IAEJ,IAAId,oBAAoB,sBAAsB,EAAE;QAC9Cc,wBAAwB;YACtB,MAAM;YACN,SAAS;gBACP;oBACE,MAAM;oBACN,MAAM,GAAGd,oBAAoB,sBAAsB,CAAC,gFAAgF,CAAC;gBACvI;gBACA;oBACE,MAAM;oBACN,WAAW;wBACT,KAAKM;wBACL,QAAQ;oBACV;gBACF;aACD;QACH;QAEAN,oBAAoB,mCAAmC;IACzD,OACEc,wBAAwB;QACtB,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAM;YACR;YACA;gBACE,MAAM;gBACN,WAAW;oBACT,KAAKR;oBACL,QAAQ;gBACV;YACF;SACD;IACH;IAEFN,oBAAoB,MAAM,CAACc;IAC3B,MAAMC,aAAaf,oBAAoB,QAAQ;IAE/C,MAAMgB,OAAqC;QACzC;YAAE,MAAM;YAAU,SAASZ;QAAa;WACrCS;WACAE;KACJ;IAED,MAAM,EACJ,SAASE,UAAU,EACnB,eAAeC,WAAW,EAC1BC,KAAK,EACN,GAAG,MAAMC,yBACRJ,MACAK,aAAa,IAAI,EACjBtB;IAGF,MAAMuB,UAAUL,WAAW,MAAM,GAAG;QAACA,WAAW,MAAM;KAAC,GAAG,EAAE;IAC5D,MAAMM,cAAkC;QACtC,GAAGN,UAAU;QACbK;QACAJ;QACAC;QACA,UAAUK,uBACRF,SACAzB,KAAK,WAAW,EAChBoB,WAAW,KAAK;IAEpB;IAEAQ,OAAOR,YAAY;IAEnBK,QAAQ,OAAO,CAAC,CAACI;QACf,MAAMC,OAAOD,OAAO,IAAI;QACxB,MAAME,sBAAsB/B,KAAK,WAAW,CAAC,IAAI,CAC/C,CAAC6B,SAAWA,OAAO,IAAI,KAAKC;QAG9BlC,MAAM,+BAA+BmC;QACrC,MAAMC,eAAeD,sBACjBE,4BAA4BF,oBAAoB,WAAW,IAC3D,EAAE;QAENnC,MAAM,gBAAgBoC;QAEtBA,aAAa,OAAO,CAAC,CAACE;YACpB,MAAMC,eAAeN,OAAO,KAAK,CAACK,MAAM;YACxC,IAAIC,gBAAgB7B,AAAW8B,WAAX9B,QAElBuB,OAAO,KAAK,CAACK,MAAM,GAAGG,cACpBF,cACAzB,YACAC,aACAC,YACAC,aACAP;QAGN;IACF;IAEAsB,OAAO,CAACR,WAAW,KAAK,EAAE,CAAC,wBAAwB,EAAEA,WAAW,KAAK,EAAE;IAEvE,IACEK,AAAmB,MAAnBA,QAAQ,MAAM,IACdC,YAAY,kCAAkC,IAC9C,CAACA,YAAY,KAAK,EAElBY,QAAQ,IAAI,CACV,8EACAvC;IAIJI,oBAAoB,MAAM,CAAC;QACzB,MAAM;QACN,SAAS;YACP;gBACE,MAAM;gBACN,MAAMkB;YACR;SACD;IACH;IAEA,OAAOK;AACT"}

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

@@ -121,6 +121,8 @@ Please tell what the next one action is (or null if no action should be done) to
 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.
 - Give just the next ONE action you should do
+- Make sure the previous actions are completed successfully before performing the next step
+- If there are some error messages reported by the previous actions, don't give up, try parse a new action to recover. If the error persists for more than 5 times, you should think this is an error and set the "error" field to the error message.
 - If the user mentions something to assert and the condition is not met, you should think this is an error and set the "error" field to the error message.
 
 Supporting actions:

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

@@ -1 +1 @@
-{"version":3,"file":"ai-model/prompt/llm-planning.mjs","sources":["webpack://@midscene/core/./src/ai-model/prompt/llm-planning.ts"],"sourcesContent":["import type { DeviceAction, ThinkingStrategy } 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  thinkingStrategy,\n}: {\n  actionSpace: DeviceAction<any>[];\n  vlMode: TVlModeTypes | undefined;\n  includeBbox: boolean;\n  thinkingStrategy: ThinkingStrategy;\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  // Conditionally include log field based on thinkingStrategy\n  const logFieldDefinition =\n    thinkingStrategy === 'off'\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  ';\n\n  const exampleLogField =\n    thinkingStrategy === 'off'\n      ? ''\n      : \"\\\"log\\\": \\\"The user wants to do click 'Confirm' button, and click 'Yes' in popup. The current progress is ..., we still need to ... .  Now i am going to compose an action '...' to click 'Yes' in popup.\\\",\\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- If the user mentions something to assert and the condition is not met, you should think this is an error and set the \"error\" field to the error message.\n\nSupporting actions:\n${actionList}\n\nReturn in JSON format:\n{\n  ${logFieldDefinition}${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  ${exampleLogField}\"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":["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","key","Object","isOptional","keyWithOptional","description","paramLine","line","_schema__def1","schemaTypeName","paramDescription","systemPromptToTaskPlanning","actionSpace","includeBbox","thinkingStrategy","Error","actionDescriptionList","undefined","actionList","logFieldDefinition","exampleLogField","planSchema"],"mappings":";;AASA,MAAMA,qBAAqB,CAAC;+NACmM,CAAC;AAEhO,MAAMC,gBAAgB,CAACC;IACrB,IAAIA,QACF,OAAO,CAAC,6DAA6D,EAAEC,gBAAgBD,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,uBAAuBF,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,CAACa,KAAKlB,MAAM,IAAImB,OAAO,OAAO,CAACrB,OACxC,IAAIE,SAAS,AAAiB,YAAjB,OAAOA,OAAoB;gBAEtC,MAAMoB,aACJ,AAAqC,cAArC,OAAQpB,MAAc,UAAU,IAC/BA,MAAc,UAAU;gBAC3B,MAAMqB,kBAAkBD,aAAa,GAAGF,IAAI,CAAC,CAAC,GAAGA;gBAGjD,MAAMd,WAAWL,YAAYC;gBAG7B,MAAMsB,cAAcL,eAAejB;gBAGnC,IAAIuB,YAAY,GAAGF,gBAAgB,EAAE,EAAEjB,UAAU;gBACjD,IAAIkB,aACFC,aAAa,CAAC,IAAI,EAAED,aAAa;gBAGnC3B,WAAW,IAAI,CAAC4B;YAClB;YAIF,IAAI5B,WAAW,MAAM,GAAG,GAAG;gBACzBF,OAAO,IAAI,CAAC;gBACZE,WAAW,OAAO,CAAC,CAAC6B;oBAClB/B,OAAO,IAAI,CAAC,CAAC,IAAI,EAAE+B,MAAM;gBAC3B;YACF;QACF,OAAO;gBAGkBC;YAAvB,MAAMC,iBAAiB,QAAAD,CAAAA,gBAAAA,OAAO,IAAI,AAAD,IAAVA,KAAAA,IAAAA,cAAa,QAAQ;YAC5C,IAAIrB,WAAW;YAEf,IAAIsB,AAAmB,gBAAnBA,gBAAgCtB,WAAW;iBAC1C,IAAIsB,AAAmB,gBAAnBA,gBAAgCtB,WAAW;iBAC/C,IAAIsB,AAAmB,iBAAnBA,gBAAiCtB,WAAW;YAGrD,MAAMkB,cAAc,iBAAiB1B,SAASA,OAAO,WAAW,GAAG;YAGnE,IAAI+B,mBAAmB,CAAC,SAAS,EAAEvB,UAAU;YAC7C,IAAIkB,aACFK,oBAAoB,CAAC,IAAI,EAAEL,aAAa;YAE1CK,oBAAoB;YAEpBlC,OAAO,IAAI,CAACkC;QACd;IACF;IAEA,OAAO,CAAC,EAAE,EAAErC,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,eAAeoC,2BAA2B,EAC/CC,WAAW,EACX1C,MAAM,EACN2C,WAAW,EACXC,gBAAgB,EAMjB;IAEC,IAAID,eAAe,CAAC3C,QAClB,MAAM,IAAI6C,MACR;IAIJ,MAAMC,wBAAwBJ,YAAY,GAAG,CAAC,CAACvC,SACtCD,qBACLC,QACAJ,cAAc4C,cAAc3C,SAAS+C;IAGzC,MAAMC,aAAaF,sBAAsB,IAAI,CAAC;IAG9C,MAAMG,qBACJL,AAAqB,UAArBA,mBACI,KACA;IAEN,MAAMM,kBACJN,AAAqB,UAArBA,mBACI,KACA;IAEN,OAAO,CAAC;;;;;;;;;;;AAWV,EAAEI,WAAW;;;;EAIX,EAAEC,qBAAqBnD,mBAAmB;;;;;;;;;;;;;;EAc1C,EAAEoD,gBAAgB;;;;QAIZ,EAAElD,SAAS,kCAAkC,GAAG;;;;;;;AAOxD,CAAC;AACD;AAEO,MAAMmD,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.mjs","sources":["webpack://@midscene/core/./src/ai-model/prompt/llm-planning.ts"],"sourcesContent":["import type { DeviceAction, ThinkingStrategy } 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  thinkingStrategy,\n}: {\n  actionSpace: DeviceAction<any>[];\n  vlMode: TVlModeTypes | undefined;\n  includeBbox: boolean;\n  thinkingStrategy: ThinkingStrategy;\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  // Conditionally include log field based on thinkingStrategy\n  const logFieldDefinition =\n    thinkingStrategy === 'off'\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  ';\n\n  const exampleLogField =\n    thinkingStrategy === 'off'\n      ? ''\n      : \"\\\"log\\\": \\\"The user wants to do click 'Confirm' button, and click 'Yes' in popup. The current progress is ..., we still need to ... .  Now i am going to compose an action '...' to click 'Yes' in popup.\\\",\\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- Make sure the previous actions are completed successfully before performing the next step\n- If there are some error messages reported by the previous actions, don't give up, try parse a new action to recover. If the error persists for more than 5 times, you should think this is an error and set the \"error\" field to the error message.\n- If the user mentions something to assert and the condition is not met, you should think this is an error and set the \"error\" field to the error message.\n\nSupporting actions:\n${actionList}\n\nReturn in JSON format:\n{\n  ${logFieldDefinition}${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  ${exampleLogField}\"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":["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","key","Object","isOptional","keyWithOptional","description","paramLine","line","_schema__def1","schemaTypeName","paramDescription","systemPromptToTaskPlanning","actionSpace","includeBbox","thinkingStrategy","Error","actionDescriptionList","undefined","actionList","logFieldDefinition","exampleLogField","planSchema"],"mappings":";;AASA,MAAMA,qBAAqB,CAAC;+NACmM,CAAC;AAEhO,MAAMC,gBAAgB,CAACC;IACrB,IAAIA,QACF,OAAO,CAAC,6DAA6D,EAAEC,gBAAgBD,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,uBAAuBF,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,CAACa,KAAKlB,MAAM,IAAImB,OAAO,OAAO,CAACrB,OACxC,IAAIE,SAAS,AAAiB,YAAjB,OAAOA,OAAoB;gBAEtC,MAAMoB,aACJ,AAAqC,cAArC,OAAQpB,MAAc,UAAU,IAC/BA,MAAc,UAAU;gBAC3B,MAAMqB,kBAAkBD,aAAa,GAAGF,IAAI,CAAC,CAAC,GAAGA;gBAGjD,MAAMd,WAAWL,YAAYC;gBAG7B,MAAMsB,cAAcL,eAAejB;gBAGnC,IAAIuB,YAAY,GAAGF,gBAAgB,EAAE,EAAEjB,UAAU;gBACjD,IAAIkB,aACFC,aAAa,CAAC,IAAI,EAAED,aAAa;gBAGnC3B,WAAW,IAAI,CAAC4B;YAClB;YAIF,IAAI5B,WAAW,MAAM,GAAG,GAAG;gBACzBF,OAAO,IAAI,CAAC;gBACZE,WAAW,OAAO,CAAC,CAAC6B;oBAClB/B,OAAO,IAAI,CAAC,CAAC,IAAI,EAAE+B,MAAM;gBAC3B;YACF;QACF,OAAO;gBAGkBC;YAAvB,MAAMC,iBAAiB,QAAAD,CAAAA,gBAAAA,OAAO,IAAI,AAAD,IAAVA,KAAAA,IAAAA,cAAa,QAAQ;YAC5C,IAAIrB,WAAW;YAEf,IAAIsB,AAAmB,gBAAnBA,gBAAgCtB,WAAW;iBAC1C,IAAIsB,AAAmB,gBAAnBA,gBAAgCtB,WAAW;iBAC/C,IAAIsB,AAAmB,iBAAnBA,gBAAiCtB,WAAW;YAGrD,MAAMkB,cAAc,iBAAiB1B,SAASA,OAAO,WAAW,GAAG;YAGnE,IAAI+B,mBAAmB,CAAC,SAAS,EAAEvB,UAAU;YAC7C,IAAIkB,aACFK,oBAAoB,CAAC,IAAI,EAAEL,aAAa;YAE1CK,oBAAoB;YAEpBlC,OAAO,IAAI,CAACkC;QACd;IACF;IAEA,OAAO,CAAC,EAAE,EAAErC,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,eAAeoC,2BAA2B,EAC/CC,WAAW,EACX1C,MAAM,EACN2C,WAAW,EACXC,gBAAgB,EAMjB;IAEC,IAAID,eAAe,CAAC3C,QAClB,MAAM,IAAI6C,MACR;IAIJ,MAAMC,wBAAwBJ,YAAY,GAAG,CAAC,CAACvC,SACtCD,qBACLC,QACAJ,cAAc4C,cAAc3C,SAAS+C;IAGzC,MAAMC,aAAaF,sBAAsB,IAAI,CAAC;IAG9C,MAAMG,qBACJL,AAAqB,UAArBA,mBACI,KACA;IAEN,MAAMM,kBACJN,AAAqB,UAArBA,mBACI,KACA;IAEN,OAAO,CAAC;;;;;;;;;;;;;AAaV,EAAEI,WAAW;;;;EAIX,EAAEC,qBAAqBnD,mBAAmB;;;;;;;;;;;;;;EAc1C,EAAEoD,gBAAgB;;;;QAIZ,EAAElD,SAAS,kCAAkC,GAAG;;;;;;;AAOxD,CAAC;AACD;AAEO,MAAMmD,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/es/task-runner.mjs

@@ -73,6 +73,12 @@ class TaskRunner {
             ...task
         };
     }
+    normalizeStatusFromError(options, errorMessage) {
+        var _this_latestErrorTask, _this_latestErrorTask1;
+        if ('error' !== this.status) return;
+        assert(null == options ? void 0 : options.allowWhenError, errorMessage || `task runner is in error state, cannot proceed\nerror=${null == (_this_latestErrorTask = this.latestErrorTask()) ? void 0 : _this_latestErrorTask.error}\n${null == (_this_latestErrorTask1 = this.latestErrorTask()) ? void 0 : _this_latestErrorTask1.errorStack}`);
+        this.status = this.tasks.length > 0 ? 'pending' : 'init';
+    }
     findPreviousNonSubTaskUIContext(currentIndex) {
         for(let i = currentIndex - 1; i >= 0; i--){
             const candidate = this.tasks[i];
@@ -81,23 +87,23 @@ class TaskRunner {
             }
         }
     }
-    async append(task) {
+    async append(task, options) {
         var _this_latestErrorTask, _this_latestErrorTask1;
-        assert('error' !== this.status, `task runner is in error state, cannot append task\nerror=${null == (_this_latestErrorTask = this.latestErrorTask()) ? void 0 : _this_latestErrorTask.error}\n${null == (_this_latestErrorTask1 = this.latestErrorTask()) ? void 0 : _this_latestErrorTask1.errorStack}`);
+        this.normalizeStatusFromError(options, `task runner is in error state, cannot append task\nerror=${null == (_this_latestErrorTask = this.latestErrorTask()) ? void 0 : _this_latestErrorTask.error}\n${null == (_this_latestErrorTask1 = this.latestErrorTask()) ? void 0 : _this_latestErrorTask1.errorStack}`);
         if (Array.isArray(task)) this.tasks.push(...task.map((item)=>this.markTaskAsPending(item)));
         else this.tasks.push(this.markTaskAsPending(task));
         if ('running' !== this.status) this.status = 'pending';
         await this.emitOnTaskUpdate();
     }
-    async appendAndFlush(task) {
-        await this.append(task);
-        return this.flush();
+    async appendAndFlush(task, options) {
+        await this.append(task, options);
+        return this.flush(options);
     }
-    async flush() {
+    async flush(options) {
         if ('init' === this.status && this.tasks.length > 0) console.warn('illegal state for task runner, status is init but tasks are not empty');
+        this.normalizeStatusFromError(options, 'task runner is in error state');
         assert('running' !== this.status, 'task runner is already running');
         assert('completed' !== this.status, 'task runner is already completed');
-        assert('error' !== this.status, 'task runner is in error state');
         const nextPendingIndex = this.tasks.findIndex((task)=>'pending' === task.status);
         if (nextPendingIndex < 0) return;
         this.status = 'running';