@midscene/core 0.30.5 → 1.0.1-beta-20251021060907.0

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

@@ -1,102 +1,58 @@
 import { AIResponseFormat } from "../../types.mjs";
-import { Anthropic } from "@anthropic-ai/sdk";
-import { DefaultAzureCredential, getBearerTokenProvider } from "@azure/identity";
-import { MIDSCENE_API_TYPE, MIDSCENE_LANGSMITH_DEBUG, OPENAI_MAX_TOKENS, globalConfigManager } from "@midscene/shared/env";
-import { parseBase64 } from "@midscene/shared/img";
+import { OPENAI_MAX_TOKENS, globalConfigManager } from "@midscene/shared/env";
 import { getDebug } from "@midscene/shared/logger";
-import { assert, ifInBrowser } from "@midscene/shared/utils";
+import { assert } from "@midscene/shared/utils";
 import { HttpsProxyAgent } from "https-proxy-agent";
 import { jsonrepair } from "jsonrepair";
-import openai_0, { AzureOpenAI } from "openai";
+import openai_0 from "openai";
 import { SocksProxyAgent } from "socks-proxy-agent";
 import { AIActionType } from "../common.mjs";
 import { assertSchema } from "../prompt/assertion.mjs";
-import { locatorSchema } from "../prompt/llm-locator.mjs";
 import { planSchema } from "../prompt/llm-planning.mjs";
 async function createChatClient({ AIActionTypeValue, modelConfig }) {
-    const { socksProxy, httpProxy, modelName, openaiBaseURL, openaiApiKey, openaiExtraConfig, openaiUseAzureDeprecated, useAzureOpenai, azureOpenaiScope, azureOpenaiKey, azureOpenaiEndpoint, azureOpenaiApiVersion, azureOpenaiDeployment, azureExtraConfig, useAnthropicSdk, anthropicApiKey, modelDescription, uiTarsModelVersion: uiTarsVersion, vlMode } = modelConfig;
+    const { socksProxy, httpProxy, modelName, openaiBaseURL, openaiApiKey, openaiExtraConfig, modelDescription, uiTarsModelVersion: uiTarsVersion, vlMode, createOpenAIClient } = modelConfig;
     let openai;
-    let proxyAgent;
-    const debugProxy = getDebug('ai:call:proxy');
-    if (httpProxy) {
-        debugProxy('using http proxy', httpProxy);
-        proxyAgent = new HttpsProxyAgent(httpProxy);
-    } else if (socksProxy) {
-        debugProxy('using socks proxy', socksProxy);
-        proxyAgent = new SocksProxyAgent(socksProxy);
-    }
-    if (openaiUseAzureDeprecated) openai = new AzureOpenAI({
-        baseURL: openaiBaseURL,
-        apiKey: openaiApiKey,
-        httpAgent: proxyAgent,
-        ...openaiExtraConfig,
-        dangerouslyAllowBrowser: true
+    if (createOpenAIClient) openai = createOpenAIClient({
+        modelName,
+        openaiApiKey,
+        openaiBaseURL,
+        socksProxy,
+        httpProxy,
+        openaiExtraConfig,
+        vlMode,
+        intent: modelConfig.intent,
+        modelDescription
     });
-    else if (useAzureOpenai) {
-        let tokenProvider;
-        if (azureOpenaiScope) {
-            assert(!ifInBrowser, 'Azure OpenAI is not supported in browser with Midscene.');
-            const credential = new DefaultAzureCredential();
-            tokenProvider = getBearerTokenProvider(credential, azureOpenaiScope);
-            openai = new AzureOpenAI({
-                azureADTokenProvider: tokenProvider,
-                endpoint: azureOpenaiEndpoint,
-                apiVersion: azureOpenaiApiVersion,
-                deployment: azureOpenaiDeployment,
-                ...openaiExtraConfig,
-                ...azureExtraConfig
-            });
-        } else openai = new AzureOpenAI({
-            apiKey: azureOpenaiKey,
-            endpoint: azureOpenaiEndpoint,
-            apiVersion: azureOpenaiApiVersion,
-            deployment: azureOpenaiDeployment,
-            dangerouslyAllowBrowser: true,
+    else {
+        let proxyAgent;
+        const debugProxy = getDebug('ai:call:proxy');
+        if (httpProxy) {
+            debugProxy('using http proxy', httpProxy);
+            proxyAgent = new HttpsProxyAgent(httpProxy);
+        } else if (socksProxy) {
+            debugProxy('using socks proxy', socksProxy);
+            proxyAgent = new SocksProxyAgent(socksProxy);
+        }
+        openai = new openai_0({
+            baseURL: openaiBaseURL,
+            apiKey: openaiApiKey,
+            ...proxyAgent ? {
+                httpAgent: proxyAgent
+            } : {},
             ...openaiExtraConfig,
-            ...azureExtraConfig
+            dangerouslyAllowBrowser: true
         });
-    } else if (!useAnthropicSdk) openai = new openai_0({
-        baseURL: openaiBaseURL,
-        apiKey: openaiApiKey,
-        httpAgent: proxyAgent,
-        ...openaiExtraConfig,
-        defaultHeaders: {
-            ...(null == openaiExtraConfig ? void 0 : openaiExtraConfig.defaultHeaders) || {},
-            [MIDSCENE_API_TYPE]: AIActionTypeValue.toString()
-        },
-        dangerouslyAllowBrowser: true
-    });
-    if (openai && globalConfigManager.getEnvConfigInBoolean(MIDSCENE_LANGSMITH_DEBUG)) {
-        if (ifInBrowser) throw new Error('langsmith is not supported in browser');
-        console.log('DEBUGGING MODE: langsmith wrapper enabled');
-        const { wrapOpenAI } = await import("langsmith/wrappers");
-        openai = wrapOpenAI(openai);
     }
-    if (void 0 !== openai) return {
+    return {
         completion: openai.chat.completions,
-        style: 'openai',
-        modelName,
-        modelDescription,
-        uiTarsVersion,
-        vlMode
-    };
-    if (useAnthropicSdk) openai = new Anthropic({
-        apiKey: anthropicApiKey,
-        httpAgent: proxyAgent,
-        dangerouslyAllowBrowser: true
-    });
-    if (void 0 !== openai && openai.messages) return {
-        completion: openai.messages,
-        style: 'anthropic',
         modelName,
         modelDescription,
         uiTarsVersion,
         vlMode
     };
-    throw new Error('Openai SDK or Anthropic SDK is not initialized');
 }
 async function callAI(messages, AIActionTypeValue, modelConfig, options) {
-    const { completion, style, modelName, modelDescription, uiTarsVersion, vlMode } = await createChatClient({
+    const { completion, modelName, modelDescription, uiTarsVersion, vlMode } = await createChatClient({
         AIActionTypeValue,
         modelConfig
     });
@@ -115,168 +71,85 @@ async function callAI(messages, AIActionTypeValue, modelConfig, options) {
         temperature: 'vlm-ui-tars' === vlMode ? 0.0 : 0.1,
         stream: !!isStreaming,
         max_tokens: 'number' == typeof maxTokens ? maxTokens : Number.parseInt(maxTokens || '2048', 10),
-        ...'qwen-vl' === vlMode || 'qwen3-vl' === vlMode ? {
+        ...'qwen-vl' === vlMode ? {
             vl_high_resolution_images: true
         } : {}
     };
     try {
-        if ('openai' === style) {
-            debugCall(`sending ${isStreaming ? 'streaming ' : ''}request to ${modelName}`);
-            if (isStreaming) {
-                const stream = await completion.create({
-                    model: modelName,
-                    messages,
-                    response_format: responseFormat,
-                    ...commonConfig
-                }, {
-                    stream: true
-                });
-                for await (const chunk of stream){
-                    var _chunk_choices__delta, _chunk_choices_, _chunk_choices, _chunk_choices__delta1, _chunk_choices_1, _chunk_choices1, _chunk_choices_2, _chunk_choices2;
-                    const content = (null == (_chunk_choices = chunk.choices) ? void 0 : null == (_chunk_choices_ = _chunk_choices[0]) ? void 0 : null == (_chunk_choices__delta = _chunk_choices_.delta) ? void 0 : _chunk_choices__delta.content) || '';
-                    const reasoning_content = (null == (_chunk_choices1 = chunk.choices) ? void 0 : null == (_chunk_choices_1 = _chunk_choices1[0]) ? void 0 : null == (_chunk_choices__delta1 = _chunk_choices_1.delta) ? void 0 : _chunk_choices__delta1.reasoning_content) || '';
-                    if (chunk.usage) usage = chunk.usage;
-                    if (content || reasoning_content) {
-                        accumulated += content;
-                        const chunkData = {
-                            content,
-                            reasoning_content,
-                            accumulated,
-                            isComplete: false,
-                            usage: void 0
-                        };
-                        options.onChunk(chunkData);
-                    }
-                    if (null == (_chunk_choices2 = chunk.choices) ? void 0 : null == (_chunk_choices_2 = _chunk_choices2[0]) ? void 0 : _chunk_choices_2.finish_reason) {
-                        timeCost = Date.now() - startTime;
-                        if (!usage) {
-                            const estimatedTokens = Math.max(1, Math.floor(accumulated.length / 4));
-                            usage = {
-                                prompt_tokens: estimatedTokens,
-                                completion_tokens: estimatedTokens,
-                                total_tokens: 2 * estimatedTokens
-                            };
-                        }
-                        const finalChunk = {
-                            content: '',
-                            accumulated,
-                            reasoning_content: '',
-                            isComplete: true,
-                            usage: {
-                                prompt_tokens: usage.prompt_tokens ?? 0,
-                                completion_tokens: usage.completion_tokens ?? 0,
-                                total_tokens: usage.total_tokens ?? 0,
-                                time_cost: timeCost ?? 0,
-                                model_name: modelName,
-                                model_description: modelDescription,
-                                intent: modelConfig.intent
-                            }
-                        };
-                        options.onChunk(finalChunk);
-                        break;
-                    }
-                }
-                content = accumulated;
-                debugProfileStats(`streaming model, ${modelName}, mode, ${vlMode || 'default'}, cost-ms, ${timeCost}`);
-            } else {
-                var _result_usage, _result_usage1, _result_usage2;
-                const result = await completion.create({
-                    model: modelName,
-                    messages,
-                    response_format: responseFormat,
-                    ...commonConfig
-                });
-                timeCost = Date.now() - startTime;
-                debugProfileStats(`model, ${modelName}, mode, ${vlMode || 'default'}, ui-tars-version, ${uiTarsVersion}, prompt-tokens, ${(null == (_result_usage = result.usage) ? void 0 : _result_usage.prompt_tokens) || ''}, completion-tokens, ${(null == (_result_usage1 = result.usage) ? void 0 : _result_usage1.completion_tokens) || ''}, total-tokens, ${(null == (_result_usage2 = result.usage) ? void 0 : _result_usage2.total_tokens) || ''}, cost-ms, ${timeCost}, requestId, ${result._request_id || ''}`);
-                debugProfileDetail(`model usage detail: ${JSON.stringify(result.usage)}`);
-                assert(result.choices, `invalid response from LLM service: ${JSON.stringify(result)}`);
-                content = result.choices[0].message.content;
-                usage = result.usage;
-            }
-            debugCall(`response: ${content}`);
-            assert(content, 'empty content');
-        } else if ('anthropic' === style) {
-            const convertImageContent = (content)=>{
-                if ('image_url' === content.type) {
-                    const imgBase64 = content.image_url.url;
-                    assert(imgBase64, 'image_url is required');
-                    const { mimeType, body } = parseBase64(content.image_url.url);
-                    return {
-                        source: {
-                            type: 'base64',
-                            media_type: mimeType,
-                            data: body
-                        },
-                        type: 'image'
+        debugCall(`sending ${isStreaming ? 'streaming ' : ''}request to ${modelName}`);
+        if (isStreaming) {
+            const stream = await completion.create({
+                model: modelName,
+                messages,
+                response_format: responseFormat,
+                ...commonConfig
+            }, {
+                stream: true
+            });
+            for await (const chunk of stream){
+                var _chunk_choices__delta, _chunk_choices_, _chunk_choices, _chunk_choices__delta1, _chunk_choices_1, _chunk_choices1, _chunk_choices_2, _chunk_choices2;
+                const content = (null == (_chunk_choices = chunk.choices) ? void 0 : null == (_chunk_choices_ = _chunk_choices[0]) ? void 0 : null == (_chunk_choices__delta = _chunk_choices_.delta) ? void 0 : _chunk_choices__delta.content) || '';
+                const reasoning_content = (null == (_chunk_choices1 = chunk.choices) ? void 0 : null == (_chunk_choices_1 = _chunk_choices1[0]) ? void 0 : null == (_chunk_choices__delta1 = _chunk_choices_1.delta) ? void 0 : _chunk_choices__delta1.reasoning_content) || '';
+                if (chunk.usage) usage = chunk.usage;
+                if (content || reasoning_content) {
+                    accumulated += content;
+                    const chunkData = {
+                        content,
+                        reasoning_content,
+                        accumulated,
+                        isComplete: false,
+                        usage: void 0
                     };
+                    options.onChunk(chunkData);
                 }
-                return content;
-            };
-            if (isStreaming) {
-                const stream = await completion.create({
-                    model: modelName,
-                    system: 'You are a versatile professional in software UI automation',
-                    messages: messages.map((m)=>({
-                            role: 'user',
-                            content: Array.isArray(m.content) ? m.content.map(convertImageContent) : m.content
-                        })),
-                    response_format: responseFormat,
-                    ...commonConfig
-                });
-                for await (const chunk of stream){
-                    var _chunk_delta;
-                    const content = (null == (_chunk_delta = chunk.delta) ? void 0 : _chunk_delta.text) || '';
-                    if (content) {
-                        accumulated += content;
-                        const chunkData = {
-                            content,
-                            accumulated,
-                            reasoning_content: '',
-                            isComplete: false,
-                            usage: void 0
+                if (null == (_chunk_choices2 = chunk.choices) ? void 0 : null == (_chunk_choices_2 = _chunk_choices2[0]) ? void 0 : _chunk_choices_2.finish_reason) {
+                    timeCost = Date.now() - startTime;
+                    if (!usage) {
+                        const estimatedTokens = Math.max(1, Math.floor(accumulated.length / 4));
+                        usage = {
+                            prompt_tokens: estimatedTokens,
+                            completion_tokens: estimatedTokens,
+                            total_tokens: 2 * estimatedTokens
                         };
-                        options.onChunk(chunkData);
-                    }
-                    if ('message_stop' === chunk.type) {
-                        timeCost = Date.now() - startTime;
-                        const anthropicUsage = chunk.usage;
-                        const finalChunk = {
-                            content: '',
-                            accumulated,
-                            reasoning_content: '',
-                            isComplete: true,
-                            usage: anthropicUsage ? {
-                                prompt_tokens: anthropicUsage.input_tokens ?? 0,
-                                completion_tokens: anthropicUsage.output_tokens ?? 0,
-                                total_tokens: (anthropicUsage.input_tokens ?? 0) + (anthropicUsage.output_tokens ?? 0),
-                                time_cost: timeCost ?? 0,
-                                model_name: modelName,
-                                model_description: modelDescription,
-                                intent: modelConfig.intent
-                            } : void 0
-                        };
-                        options.onChunk(finalChunk);
-                        break;
                     }
+                    const finalChunk = {
+                        content: '',
+                        accumulated,
+                        reasoning_content: '',
+                        isComplete: true,
+                        usage: {
+                            prompt_tokens: usage.prompt_tokens ?? 0,
+                            completion_tokens: usage.completion_tokens ?? 0,
+                            total_tokens: usage.total_tokens ?? 0,
+                            time_cost: timeCost ?? 0,
+                            model_name: modelName,
+                            model_description: modelDescription,
+                            intent: modelConfig.intent
+                        }
+                    };
+                    options.onChunk(finalChunk);
+                    break;
                 }
-                content = accumulated;
-            } else {
-                const result = await completion.create({
-                    model: modelName,
-                    system: 'You are a versatile professional in software UI automation',
-                    messages: messages.map((m)=>({
-                            role: 'user',
-                            content: Array.isArray(m.content) ? m.content.map(convertImageContent) : m.content
-                        })),
-                    response_format: responseFormat,
-                    ...commonConfig
-                });
-                timeCost = Date.now() - startTime;
-                content = result.content[0].text;
-                usage = result.usage;
             }
-            assert(content, 'empty content');
+            content = accumulated;
+            debugProfileStats(`streaming model, ${modelName}, mode, ${vlMode || 'default'}, cost-ms, ${timeCost}`);
+        } else {
+            var _result_usage, _result_usage1, _result_usage2;
+            const result = await completion.create({
+                model: modelName,
+                messages,
+                response_format: responseFormat,
+                ...commonConfig
+            });
+            timeCost = Date.now() - startTime;
+            debugProfileStats(`model, ${modelName}, mode, ${vlMode || 'default'}, ui-tars-version, ${uiTarsVersion}, prompt-tokens, ${(null == (_result_usage = result.usage) ? void 0 : _result_usage.prompt_tokens) || ''}, completion-tokens, ${(null == (_result_usage1 = result.usage) ? void 0 : _result_usage1.completion_tokens) || ''}, total-tokens, ${(null == (_result_usage2 = result.usage) ? void 0 : _result_usage2.total_tokens) || ''}, cost-ms, ${timeCost}, requestId, ${result._request_id || ''}`);
+            debugProfileDetail(`model usage detail: ${JSON.stringify(result.usage)}`);
+            assert(result.choices, `invalid response from LLM service: ${JSON.stringify(result)}`);
+            content = result.choices[0].message.content;
+            usage = result.usage;
         }
+        debugCall(`response: ${content}`);
+        assert(content, 'empty content');
         if (isStreaming && !usage) {
             const estimatedTokens = Math.max(1, Math.floor((content || '').length / 4));
             usage = {
@@ -312,9 +185,6 @@ const getResponseFormat = (modelName, AIActionTypeValue)=>{
         case AIActionType.ASSERT:
             responseFormat = assertSchema;
             break;
-        case AIActionType.INSPECT_ELEMENT:
-            responseFormat = locatorSchema;
-            break;
         case AIActionType.PLAN:
             responseFormat = planSchema;
             break;

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

@@ -1 +1 @@
-{"version":3,"file":"ai-model/service-caller/index.mjs","sources":["webpack://@midscene/core/./src/ai-model/service-caller/index.ts"],"sourcesContent":["import { AIResponseFormat, type AIUsageInfo } from '@/types';\nimport type { CodeGenerationChunk, StreamingCallback } from '@/types';\nimport { Anthropic } from '@anthropic-ai/sdk';\nimport {\n  DefaultAzureCredential,\n  getBearerTokenProvider,\n} from '@azure/identity';\nimport {\n  type IModelConfig,\n  MIDSCENE_API_TYPE,\n  MIDSCENE_LANGSMITH_DEBUG,\n  OPENAI_MAX_TOKENS,\n  type TVlModeTypes,\n  type UITarsModelVersion,\n  globalConfigManager,\n} from '@midscene/shared/env';\n\nimport { parseBase64 } from '@midscene/shared/img';\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport { ifInBrowser } from '@midscene/shared/utils';\nimport { HttpsProxyAgent } from 'https-proxy-agent';\nimport { jsonrepair } from 'jsonrepair';\nimport OpenAI, { AzureOpenAI } from 'openai';\nimport type { ChatCompletionMessageParam } from 'openai/resources/index';\nimport type { Stream } from 'openai/streaming';\nimport { SocksProxyAgent } from 'socks-proxy-agent';\nimport { AIActionType, type AIArgs } from '../common';\nimport { assertSchema } from '../prompt/assertion';\nimport { locatorSchema } from '../prompt/llm-locator';\nimport { planSchema } from '../prompt/llm-planning';\n\nasync function createChatClient({\n  AIActionTypeValue,\n  modelConfig,\n}: {\n  AIActionTypeValue: AIActionType;\n  modelConfig: IModelConfig;\n}): Promise<{\n  completion: OpenAI.Chat.Completions;\n  style: 'openai' | 'anthropic';\n  modelName: string;\n  modelDescription: string;\n  uiTarsVersion?: UITarsModelVersion;\n  vlMode: TVlModeTypes | undefined;\n}> {\n  const {\n    socksProxy,\n    httpProxy,\n    modelName,\n    openaiBaseURL,\n    openaiApiKey,\n    openaiExtraConfig,\n    openaiUseAzureDeprecated,\n    useAzureOpenai,\n    azureOpenaiScope,\n    azureOpenaiKey,\n    azureOpenaiEndpoint,\n    azureOpenaiApiVersion,\n    azureOpenaiDeployment,\n    azureExtraConfig,\n    useAnthropicSdk,\n    anthropicApiKey,\n    modelDescription,\n    uiTarsModelVersion: uiTarsVersion,\n    vlMode,\n  } = modelConfig;\n\n  let openai: OpenAI | AzureOpenAI | undefined;\n\n  let proxyAgent = undefined;\n  const debugProxy = getDebug('ai:call:proxy');\n  if (httpProxy) {\n    debugProxy('using http proxy', httpProxy);\n    proxyAgent = new HttpsProxyAgent(httpProxy);\n  } else if (socksProxy) {\n    debugProxy('using socks proxy', socksProxy);\n    proxyAgent = new SocksProxyAgent(socksProxy);\n  }\n\n  if (openaiUseAzureDeprecated) {\n    // this is deprecated\n    openai = new AzureOpenAI({\n      baseURL: openaiBaseURL,\n      apiKey: openaiApiKey,\n      httpAgent: proxyAgent,\n      ...openaiExtraConfig,\n      dangerouslyAllowBrowser: true,\n    }) as OpenAI;\n  } else if (useAzureOpenai) {\n    // https://learn.microsoft.com/en-us/azure/ai-services/openai/chatgpt-quickstart?tabs=bash%2Cjavascript-key%2Ctypescript-keyless%2Cpython&pivots=programming-language-javascript#rest-api\n    // keyless authentication\n    let tokenProvider: any = undefined;\n    if (azureOpenaiScope) {\n      assert(\n        !ifInBrowser,\n        'Azure OpenAI is not supported in browser with Midscene.',\n      );\n      const credential = new DefaultAzureCredential();\n\n      tokenProvider = getBearerTokenProvider(credential, azureOpenaiScope);\n\n      openai = new AzureOpenAI({\n        azureADTokenProvider: tokenProvider,\n        endpoint: azureOpenaiEndpoint,\n        apiVersion: azureOpenaiApiVersion,\n        deployment: azureOpenaiDeployment,\n        ...openaiExtraConfig,\n        ...azureExtraConfig,\n      });\n    } else {\n      // endpoint, apiKey, apiVersion, deployment\n      openai = new AzureOpenAI({\n        apiKey: azureOpenaiKey,\n        endpoint: azureOpenaiEndpoint,\n        apiVersion: azureOpenaiApiVersion,\n        deployment: azureOpenaiDeployment,\n        dangerouslyAllowBrowser: true,\n        ...openaiExtraConfig,\n        ...azureExtraConfig,\n      });\n    }\n  } else if (!useAnthropicSdk) {\n    openai = new OpenAI({\n      baseURL: openaiBaseURL,\n      apiKey: openaiApiKey,\n      httpAgent: proxyAgent,\n      ...openaiExtraConfig,\n      defaultHeaders: {\n        ...(openaiExtraConfig?.defaultHeaders || {}),\n        [MIDSCENE_API_TYPE]: AIActionTypeValue.toString(),\n      },\n      dangerouslyAllowBrowser: true,\n    });\n  }\n\n  if (\n    openai &&\n    globalConfigManager.getEnvConfigInBoolean(MIDSCENE_LANGSMITH_DEBUG)\n  ) {\n    if (ifInBrowser) {\n      throw new Error('langsmith is not supported in browser');\n    }\n    console.log('DEBUGGING MODE: langsmith wrapper enabled');\n    const { wrapOpenAI } = await import('langsmith/wrappers');\n    openai = wrapOpenAI(openai);\n  }\n\n  if (typeof openai !== 'undefined') {\n    return {\n      completion: openai.chat.completions,\n      style: 'openai',\n      modelName,\n      modelDescription,\n      uiTarsVersion,\n      vlMode,\n    };\n  }\n\n  // Anthropic\n  if (useAnthropicSdk) {\n    openai = new Anthropic({\n      apiKey: anthropicApiKey,\n      httpAgent: proxyAgent,\n      dangerouslyAllowBrowser: true,\n    }) as any;\n  }\n\n  if (typeof openai !== 'undefined' && (openai as any).messages) {\n    return {\n      completion: (openai as any).messages,\n      style: 'anthropic',\n      modelName,\n      modelDescription,\n      uiTarsVersion,\n      vlMode,\n    };\n  }\n\n  throw new Error('Openai SDK or Anthropic SDK is not initialized');\n}\n\nexport async function callAI(\n  messages: ChatCompletionMessageParam[],\n  AIActionTypeValue: AIActionType,\n  modelConfig: IModelConfig,\n  options?: {\n    stream?: boolean;\n    onChunk?: StreamingCallback;\n  },\n): Promise<{ content: string; usage?: AIUsageInfo; isStreamed: boolean }> {\n  const {\n    completion,\n    style,\n    modelName,\n    modelDescription,\n    uiTarsVersion,\n    vlMode,\n  } = await createChatClient({\n    AIActionTypeValue,\n    modelConfig,\n  });\n\n  const responseFormat = getResponseFormat(modelName, AIActionTypeValue);\n\n  const maxTokens = globalConfigManager.getEnvConfigValue(OPENAI_MAX_TOKENS);\n  const debugCall = getDebug('ai:call');\n  const debugProfileStats = getDebug('ai:profile:stats');\n  const debugProfileDetail = getDebug('ai:profile:detail');\n\n  const startTime = Date.now();\n\n  const isStreaming = options?.stream && options?.onChunk;\n  let content: string | undefined;\n  let accumulated = '';\n  let usage: OpenAI.CompletionUsage | undefined;\n  let timeCost: number | undefined;\n\n  const commonConfig = {\n    temperature: vlMode === 'vlm-ui-tars' ? 0.0 : 0.1,\n    stream: !!isStreaming,\n    max_tokens:\n      typeof maxTokens === 'number'\n        ? maxTokens\n        : Number.parseInt(maxTokens || '2048', 10),\n    ...(vlMode === 'qwen-vl' || vlMode === 'qwen3-vl' // qwen specific config\n      ? {\n          vl_high_resolution_images: true,\n        }\n      : {}),\n  };\n\n  try {\n    if (style === 'openai') {\n      debugCall(\n        `sending ${isStreaming ? 'streaming ' : ''}request to ${modelName}`,\n      );\n\n      if (isStreaming) {\n        const stream = (await completion.create(\n          {\n            model: modelName,\n            messages,\n            response_format: responseFormat,\n            ...commonConfig,\n          },\n          {\n            stream: true,\n          },\n        )) as Stream<OpenAI.Chat.Completions.ChatCompletionChunk> & {\n          _request_id?: string | null;\n        };\n\n        for await (const chunk of stream) {\n          const content = chunk.choices?.[0]?.delta?.content || '';\n          const reasoning_content =\n            (chunk.choices?.[0]?.delta as any)?.reasoning_content || '';\n\n          // Check for usage info in any chunk (OpenAI provides usage in separate chunks)\n          if (chunk.usage) {\n            usage = chunk.usage;\n          }\n\n          if (content || reasoning_content) {\n            accumulated += content;\n            const chunkData: CodeGenerationChunk = {\n              content,\n              reasoning_content,\n              accumulated,\n              isComplete: false,\n              usage: undefined,\n            };\n            options.onChunk!(chunkData);\n          }\n\n          // Check if stream is complete\n          if (chunk.choices?.[0]?.finish_reason) {\n            timeCost = Date.now() - startTime;\n\n            // If usage is not available from the stream, provide a basic usage info\n            if (!usage) {\n              // Estimate token counts based on content length (rough approximation)\n              const estimatedTokens = Math.max(\n                1,\n                Math.floor(accumulated.length / 4),\n              );\n              usage = {\n                prompt_tokens: estimatedTokens,\n                completion_tokens: estimatedTokens,\n                total_tokens: estimatedTokens * 2,\n              };\n            }\n\n            // Send final chunk\n            const finalChunk: CodeGenerationChunk = {\n              content: '',\n              accumulated,\n              reasoning_content: '',\n              isComplete: true,\n              usage: {\n                prompt_tokens: usage.prompt_tokens ?? 0,\n                completion_tokens: usage.completion_tokens ?? 0,\n                total_tokens: usage.total_tokens ?? 0,\n                time_cost: timeCost ?? 0,\n                model_name: modelName,\n                model_description: modelDescription,\n                intent: modelConfig.intent,\n              },\n            };\n            options.onChunk!(finalChunk);\n            break;\n          }\n        }\n        content = accumulated;\n        debugProfileStats(\n          `streaming model, ${modelName}, mode, ${vlMode || 'default'}, cost-ms, ${timeCost}`,\n        );\n      } else {\n        const result = await completion.create({\n          model: modelName,\n          messages,\n          response_format: responseFormat,\n          ...commonConfig,\n        } as any);\n        timeCost = Date.now() - startTime;\n\n        debugProfileStats(\n          `model, ${modelName}, mode, ${vlMode || 'default'}, ui-tars-version, ${uiTarsVersion}, prompt-tokens, ${result.usage?.prompt_tokens || ''}, completion-tokens, ${result.usage?.completion_tokens || ''}, total-tokens, ${result.usage?.total_tokens || ''}, cost-ms, ${timeCost}, requestId, ${result._request_id || ''}`,\n        );\n\n        debugProfileDetail(\n          `model usage detail: ${JSON.stringify(result.usage)}`,\n        );\n\n        assert(\n          result.choices,\n          `invalid response from LLM service: ${JSON.stringify(result)}`,\n        );\n        content = result.choices[0].message.content!;\n        usage = result.usage;\n      }\n\n      debugCall(`response: ${content}`);\n      assert(content, 'empty content');\n    } else if (style === 'anthropic') {\n      const convertImageContent = (content: any) => {\n        if (content.type === 'image_url') {\n          const imgBase64 = content.image_url.url;\n          assert(imgBase64, 'image_url is required');\n          const { mimeType, body } = parseBase64(content.image_url.url);\n          return {\n            source: {\n              type: 'base64',\n              media_type: mimeType,\n              data: body,\n            },\n            type: 'image',\n          };\n        }\n        return content;\n      };\n\n      if (isStreaming) {\n        const stream = (await completion.create({\n          model: modelName,\n          system: 'You are a versatile professional in software UI automation',\n          messages: messages.map((m) => ({\n            role: 'user',\n            content: Array.isArray(m.content)\n              ? (m.content as any).map(convertImageContent)\n              : m.content,\n          })),\n          response_format: responseFormat,\n          ...commonConfig,\n        } as any)) as any;\n\n        for await (const chunk of stream) {\n          const content = chunk.delta?.text || '';\n          if (content) {\n            accumulated += content;\n            const chunkData: CodeGenerationChunk = {\n              content,\n              accumulated,\n              reasoning_content: '',\n              isComplete: false,\n              usage: undefined,\n            };\n            options.onChunk!(chunkData);\n          }\n\n          // Check if stream is complete\n          if (chunk.type === 'message_stop') {\n            timeCost = Date.now() - startTime;\n            const anthropicUsage = chunk.usage;\n\n            // Send final chunk\n            const finalChunk: CodeGenerationChunk = {\n              content: '',\n              accumulated,\n              reasoning_content: '',\n              isComplete: true,\n              usage: anthropicUsage\n                ? {\n                    prompt_tokens: anthropicUsage.input_tokens ?? 0,\n                    completion_tokens: anthropicUsage.output_tokens ?? 0,\n                    total_tokens:\n                      (anthropicUsage.input_tokens ?? 0) +\n                      (anthropicUsage.output_tokens ?? 0),\n                    time_cost: timeCost ?? 0,\n                    model_name: modelName,\n                    model_description: modelDescription,\n                    intent: modelConfig.intent,\n                  }\n                : undefined,\n            };\n            options.onChunk!(finalChunk);\n            break;\n          }\n        }\n        content = accumulated;\n      } else {\n        const result = await completion.create({\n          model: modelName,\n          system: 'You are a versatile professional in software UI automation',\n          messages: messages.map((m) => ({\n            role: 'user',\n            content: Array.isArray(m.content)\n              ? (m.content as any).map(convertImageContent)\n              : m.content,\n          })),\n          response_format: responseFormat,\n          ...commonConfig,\n        } as any);\n        timeCost = Date.now() - startTime;\n        content = (result as any).content[0].text as string;\n        usage = result.usage;\n      }\n\n      assert(content, 'empty content');\n    }\n    // Ensure we always have usage info for streaming responses\n    if (isStreaming && !usage) {\n      // Estimate token counts based on content length (rough approximation)\n      const estimatedTokens = Math.max(\n        1,\n        Math.floor((content || '').length / 4),\n      );\n      usage = {\n        prompt_tokens: estimatedTokens,\n        completion_tokens: estimatedTokens,\n        total_tokens: estimatedTokens * 2,\n      };\n    }\n\n    return {\n      content: content || '',\n      usage: usage\n        ? {\n            prompt_tokens: usage.prompt_tokens ?? 0,\n            completion_tokens: usage.completion_tokens ?? 0,\n            total_tokens: usage.total_tokens ?? 0,\n            time_cost: timeCost ?? 0,\n            model_name: modelName,\n            model_description: modelDescription,\n            intent: modelConfig.intent,\n          }\n        : undefined,\n      isStreamed: !!isStreaming,\n    };\n  } catch (e: any) {\n    console.error(' call AI error', e);\n    const newError = new Error(\n      `failed to call ${isStreaming ? 'streaming ' : ''}AI model service: ${e.message}. Trouble shooting: https://midscenejs.com/model-provider.html`,\n      {\n        cause: e,\n      },\n    );\n    throw newError;\n  }\n}\n\nexport const getResponseFormat = (\n  modelName: string,\n  AIActionTypeValue: AIActionType,\n):\n  | OpenAI.ChatCompletionCreateParams['response_format']\n  | OpenAI.ResponseFormatJSONObject => {\n  let responseFormat:\n    | OpenAI.ChatCompletionCreateParams['response_format']\n    | OpenAI.ResponseFormatJSONObject\n    | undefined;\n\n  if (modelName.includes('gpt-4')) {\n    switch (AIActionTypeValue) {\n      case AIActionType.ASSERT:\n        responseFormat = assertSchema;\n        break;\n      case AIActionType.INSPECT_ELEMENT:\n        responseFormat = locatorSchema;\n        break;\n      case AIActionType.PLAN:\n        responseFormat = planSchema;\n        break;\n      case AIActionType.EXTRACT_DATA:\n      case AIActionType.DESCRIBE_ELEMENT:\n        responseFormat = { type: AIResponseFormat.JSON };\n        break;\n      case AIActionType.TEXT:\n        // No response format for plain text - return as-is\n        responseFormat = undefined;\n        break;\n    }\n  }\n\n  // gpt-4o-2024-05-13 only supports json_object response format\n  // Skip for plain text to allow string output\n  if (\n    modelName === 'gpt-4o-2024-05-13' &&\n    AIActionTypeValue !== AIActionType.TEXT\n  ) {\n    responseFormat = { type: AIResponseFormat.JSON };\n  }\n\n  return responseFormat;\n};\n\nexport async function callAIWithObjectResponse<T>(\n  messages: ChatCompletionMessageParam[],\n  AIActionTypeValue: AIActionType,\n  modelConfig: IModelConfig,\n): Promise<{ content: T; usage?: AIUsageInfo }> {\n  const response = await callAI(messages, AIActionTypeValue, modelConfig);\n  assert(response, 'empty response');\n  const vlMode = modelConfig.vlMode;\n  const jsonContent = safeParseJson(response.content, vlMode);\n  return { content: jsonContent, usage: response.usage };\n}\n\nexport async function callAIWithStringResponse(\n  msgs: AIArgs,\n  AIActionTypeValue: AIActionType,\n  modelConfig: IModelConfig,\n): Promise<{ content: string; usage?: AIUsageInfo }> {\n  const { content, usage } = await callAI(msgs, AIActionTypeValue, modelConfig);\n  return { content, usage };\n}\n\nexport function extractJSONFromCodeBlock(response: string) {\n  try {\n    // First, try to match a JSON object directly in the response\n    const jsonMatch = response.match(/^\\s*(\\{[\\s\\S]*\\})\\s*$/);\n    if (jsonMatch) {\n      return jsonMatch[1];\n    }\n\n    // If no direct JSON object is found, try to extract JSON from a code block\n    const codeBlockMatch = response.match(\n      /```(?:json)?\\s*(\\{[\\s\\S]*?\\})\\s*```/,\n    );\n    if (codeBlockMatch) {\n      return codeBlockMatch[1];\n    }\n\n    // If no code block is found, try to find a JSON-like structure in the text\n    const jsonLikeMatch = response.match(/\\{[\\s\\S]*\\}/);\n    if (jsonLikeMatch) {\n      return jsonLikeMatch[0];\n    }\n  } catch {}\n  // If no JSON-like structure is found, return the original response\n  return response;\n}\n\nexport function preprocessDoubaoBboxJson(input: string) {\n  if (input.includes('bbox')) {\n    // when its values like 940 445 969 490, replace all /\\d+\\s+\\d+/g with /$1,$2/g\n    while (/\\d+\\s+\\d+/.test(input)) {\n      input = input.replace(/(\\d+)\\s+(\\d+)/g, '$1,$2');\n    }\n  }\n  return input;\n}\n\nexport function safeParseJson(input: string, vlMode: TVlModeTypes | undefined) {\n  const cleanJsonString = extractJSONFromCodeBlock(input);\n  // match the point\n  if (cleanJsonString?.match(/\\((\\d+),(\\d+)\\)/)) {\n    return cleanJsonString\n      .match(/\\((\\d+),(\\d+)\\)/)\n      ?.slice(1)\n      .map(Number);\n  }\n  try {\n    return JSON.parse(cleanJsonString);\n  } catch {}\n  try {\n    return JSON.parse(jsonrepair(cleanJsonString));\n  } catch (e) {}\n\n  if (vlMode === 'doubao-vision' || vlMode === 'vlm-ui-tars') {\n    const jsonString = preprocessDoubaoBboxJson(cleanJsonString);\n    return JSON.parse(jsonrepair(jsonString));\n  }\n  throw Error(`failed to parse json response: ${input}`);\n}\n"],"names":["createChatClient","AIActionTypeValue","modelConfig","socksProxy","httpProxy","modelName","openaiBaseURL","openaiApiKey","openaiExtraConfig","openaiUseAzureDeprecated","useAzureOpenai","azureOpenaiScope","azureOpenaiKey","azureOpenaiEndpoint","azureOpenaiApiVersion","azureOpenaiDeployment","azureExtraConfig","useAnthropicSdk","anthropicApiKey","modelDescription","uiTarsVersion","vlMode","openai","proxyAgent","debugProxy","getDebug","HttpsProxyAgent","SocksProxyAgent","AzureOpenAI","tokenProvider","assert","ifInBrowser","credential","DefaultAzureCredential","getBearerTokenProvider","OpenAI","MIDSCENE_API_TYPE","globalConfigManager","MIDSCENE_LANGSMITH_DEBUG","Error","console","wrapOpenAI","Anthropic","callAI","messages","options","completion","style","responseFormat","getResponseFormat","maxTokens","OPENAI_MAX_TOKENS","debugCall","debugProfileStats","debugProfileDetail","startTime","Date","isStreaming","content","accumulated","usage","timeCost","commonConfig","Number","stream","chunk","_chunk_choices__delta","_chunk_choices__delta1","_chunk_choices_2","reasoning_content","chunkData","undefined","estimatedTokens","Math","finalChunk","_result_usage","_result_usage1","_result_usage2","result","JSON","convertImageContent","imgBase64","mimeType","body","parseBase64","m","Array","_chunk_delta","anthropicUsage","e","newError","AIActionType","assertSchema","locatorSchema","planSchema","AIResponseFormat","callAIWithObjectResponse","response","jsonContent","safeParseJson","callAIWithStringResponse","msgs","extractJSONFromCodeBlock","jsonMatch","codeBlockMatch","jsonLikeMatch","preprocessDoubaoBboxJson","input","cleanJsonString","_cleanJsonString_match","jsonrepair","jsonString"],"mappings":";;;;;;;;;;;;;;;AAgCA,eAAeA,iBAAiB,EAC9BC,iBAAiB,EACjBC,WAAW,EAIZ;IAQC,MAAM,EACJC,UAAU,EACVC,SAAS,EACTC,SAAS,EACTC,aAAa,EACbC,YAAY,EACZC,iBAAiB,EACjBC,wBAAwB,EACxBC,cAAc,EACdC,gBAAgB,EAChBC,cAAc,EACdC,mBAAmB,EACnBC,qBAAqB,EACrBC,qBAAqB,EACrBC,gBAAgB,EAChBC,eAAe,EACfC,eAAe,EACfC,gBAAgB,EAChB,oBAAoBC,aAAa,EACjCC,MAAM,EACP,GAAGnB;IAEJ,IAAIoB;IAEJ,IAAIC;IACJ,MAAMC,aAAaC,SAAS;IAC5B,IAAIrB,WAAW;QACboB,WAAW,oBAAoBpB;QAC/BmB,aAAa,IAAIG,gBAAgBtB;IACnC,OAAO,IAAID,YAAY;QACrBqB,WAAW,qBAAqBrB;QAChCoB,aAAa,IAAII,gBAAgBxB;IACnC;IAEA,IAAIM,0BAEFa,SAAS,IAAIM,YAAY;QACvB,SAAStB;QACT,QAAQC;QACR,WAAWgB;QACX,GAAGf,iBAAiB;QACpB,yBAAyB;IAC3B;SACK,IAAIE,gBAAgB;QAGzB,IAAImB;QACJ,IAAIlB,kBAAkB;YACpBmB,OACE,CAACC,aACD;YAEF,MAAMC,aAAa,IAAIC;YAEvBJ,gBAAgBK,uBAAuBF,YAAYrB;YAEnDW,SAAS,IAAIM,YAAY;gBACvB,sBAAsBC;gBACtB,UAAUhB;gBACV,YAAYC;gBACZ,YAAYC;gBACZ,GAAGP,iBAAiB;gBACpB,GAAGQ,gBAAgB;YACrB;QACF,OAEEM,SAAS,IAAIM,YAAY;YACvB,QAAQhB;YACR,UAAUC;YACV,YAAYC;YACZ,YAAYC;YACZ,yBAAyB;YACzB,GAAGP,iBAAiB;YACpB,GAAGQ,gBAAgB;QACrB;IAEJ,OAAO,IAAI,CAACC,iBACVK,SAAS,IAAIa,SAAO;QAClB,SAAS7B;QACT,QAAQC;QACR,WAAWgB;QACX,GAAGf,iBAAiB;QACpB,gBAAgB;YACd,GAAIA,AAAAA,CAAAA,QAAAA,oBAAAA,KAAAA,IAAAA,kBAAmB,cAAc,AAAD,KAAK,CAAC,CAAC;YAC3C,CAAC4B,kBAAkB,EAAEnC,kBAAkB,QAAQ;QACjD;QACA,yBAAyB;IAC3B;IAGF,IACEqB,UACAe,oBAAoB,qBAAqB,CAACC,2BAC1C;QACA,IAAIP,aACF,MAAM,IAAIQ,MAAM;QAElBC,QAAQ,GAAG,CAAC;QACZ,MAAM,EAAEC,UAAU,EAAE,GAAG,MAAM,MAAM,CAAC;QACpCnB,SAASmB,WAAWnB;IACtB;IAEA,IAAI,AAAkB,WAAXA,QACT,OAAO;QACL,YAAYA,OAAO,IAAI,CAAC,WAAW;QACnC,OAAO;QACPjB;QACAc;QACAC;QACAC;IACF;IAIF,IAAIJ,iBACFK,SAAS,IAAIoB,UAAU;QACrB,QAAQxB;QACR,WAAWK;QACX,yBAAyB;IAC3B;IAGF,IAAI,AAAkB,WAAXD,UAA2BA,OAAe,QAAQ,EAC3D,OAAO;QACL,YAAaA,OAAe,QAAQ;QACpC,OAAO;QACPjB;QACAc;QACAC;QACAC;IACF;IAGF,MAAM,IAAIkB,MAAM;AAClB;AAEO,eAAeI,OACpBC,QAAsC,EACtC3C,iBAA+B,EAC/BC,WAAyB,EACzB2C,OAGC;IAED,MAAM,EACJC,UAAU,EACVC,KAAK,EACL1C,SAAS,EACTc,gBAAgB,EAChBC,aAAa,EACbC,MAAM,EACP,GAAG,MAAMrB,iBAAiB;QACzBC;QACAC;IACF;IAEA,MAAM8C,iBAAiBC,kBAAkB5C,WAAWJ;IAEpD,MAAMiD,YAAYb,oBAAoB,iBAAiB,CAACc;IACxD,MAAMC,YAAY3B,SAAS;IAC3B,MAAM4B,oBAAoB5B,SAAS;IACnC,MAAM6B,qBAAqB7B,SAAS;IAEpC,MAAM8B,YAAYC,KAAK,GAAG;IAE1B,MAAMC,cAAcZ,AAAAA,CAAAA,QAAAA,UAAAA,KAAAA,IAAAA,QAAS,MAAM,AAAD,KAAKA,CAAAA,QAAAA,UAAAA,KAAAA,IAAAA,QAAS,OAAO,AAAD;IACtD,IAAIa;IACJ,IAAIC,cAAc;IAClB,IAAIC;IACJ,IAAIC;IAEJ,MAAMC,eAAe;QACnB,aAAazC,AAAW,kBAAXA,SAA2B,MAAM;QAC9C,QAAQ,CAAC,CAACoC;QACV,YACE,AAAqB,YAArB,OAAOP,YACHA,YACAa,OAAO,QAAQ,CAACb,aAAa,QAAQ;QAC3C,GAAI7B,AAAW,cAAXA,UAAwBA,AAAW,eAAXA,SACxB;YACE,2BAA2B;QAC7B,IACA,CAAC,CAAC;IACR;IAEA,IAAI;QACF,IAAI0B,AAAU,aAAVA,OAAoB;YACtBK,UACE,CAAC,QAAQ,EAAEK,cAAc,eAAe,GAAG,WAAW,EAAEpD,WAAW;YAGrE,IAAIoD,aAAa;gBACf,MAAMO,SAAU,MAAMlB,WAAW,MAAM,CACrC;oBACE,OAAOzC;oBACPuC;oBACA,iBAAiBI;oBACjB,GAAGc,YAAY;gBACjB,GACA;oBACE,QAAQ;gBACV;gBAKF,WAAW,MAAMG,SAASD,OAAQ;wBAChBE,uBAAAA,iBAAAA,gBAEbC,wBAAAA,kBAAAA,iBAoBCC,kBAAAA;oBAtBJ,MAAMV,UAAUQ,AAAAA,SAAAA,CAAAA,iBAAAA,MAAM,OAAO,AAAD,IAAZA,KAAAA,IAAAA,QAAAA,CAAAA,kBAAAA,cAAe,CAAC,EAAE,AAAD,IAAjBA,KAAAA,IAAAA,QAAAA,CAAAA,wBAAAA,gBAAoB,KAAK,AAAD,IAAxBA,KAAAA,IAAAA,sBAA2B,OAAO,AAAD,KAAK;oBACtD,MAAMG,oBACJ,AAAC,SAAAF,CAAAA,kBAAAA,MAAM,OAAO,AAAD,IAAZA,KAAAA,IAAAA,QAAAA,CAAAA,mBAAAA,eAAe,CAAC,EAAE,AAAD,IAAjBA,KAAAA,IAAAA,QAAAA,CAAAA,yBAAAA,iBAAoB,KAAK,AAAD,IAAxBA,KAAAA,IAAAA,uBAAmC,iBAAiB,AAAD,KAAK;oBAG3D,IAAIF,MAAM,KAAK,EACbL,QAAQK,MAAM,KAAK;oBAGrB,IAAIP,WAAWW,mBAAmB;wBAChCV,eAAeD;wBACf,MAAMY,YAAiC;4BACrCZ;4BACAW;4BACAV;4BACA,YAAY;4BACZ,OAAOY;wBACT;wBACA1B,QAAQ,OAAO,CAAEyB;oBACnB;oBAGA,IAAI,QAAAF,CAAAA,kBAAAA,MAAM,OAAO,AAAD,IAAZA,KAAAA,IAAAA,QAAAA,CAAAA,mBAAAA,eAAe,CAAC,EAAE,AAAD,IAAjBA,KAAAA,IAAAA,iBAAoB,aAAa,EAAE;wBACrCP,WAAWL,KAAK,GAAG,KAAKD;wBAGxB,IAAI,CAACK,OAAO;4BAEV,MAAMY,kBAAkBC,KAAK,GAAG,CAC9B,GACAA,KAAK,KAAK,CAACd,YAAY,MAAM,GAAG;4BAElCC,QAAQ;gCACN,eAAeY;gCACf,mBAAmBA;gCACnB,cAAcA,AAAkB,IAAlBA;4BAChB;wBACF;wBAGA,MAAME,aAAkC;4BACtC,SAAS;4BACTf;4BACA,mBAAmB;4BACnB,YAAY;4BACZ,OAAO;gCACL,eAAeC,MAAM,aAAa,IAAI;gCACtC,mBAAmBA,MAAM,iBAAiB,IAAI;gCAC9C,cAAcA,MAAM,YAAY,IAAI;gCACpC,WAAWC,YAAY;gCACvB,YAAYxD;gCACZ,mBAAmBc;gCACnB,QAAQjB,YAAY,MAAM;4BAC5B;wBACF;wBACA2C,QAAQ,OAAO,CAAE6B;wBACjB;oBACF;gBACF;gBACAhB,UAAUC;gBACVN,kBACE,CAAC,iBAAiB,EAAEhD,UAAU,QAAQ,EAAEgB,UAAU,UAAU,WAAW,EAAEwC,UAAU;YAEvF,OAAO;oBAUqGc,eAAyDC,gBAAwDC;gBAT3N,MAAMC,SAAS,MAAMhC,WAAW,MAAM,CAAC;oBACrC,OAAOzC;oBACPuC;oBACA,iBAAiBI;oBACjB,GAAGc,YAAY;gBACjB;gBACAD,WAAWL,KAAK,GAAG,KAAKD;gBAExBF,kBACE,CAAC,OAAO,EAAEhD,UAAU,QAAQ,EAAEgB,UAAU,UAAU,mBAAmB,EAAED,cAAc,iBAAiB,EAAEuD,AAAAA,SAAAA,CAAAA,gBAAAA,OAAO,KAAK,AAAD,IAAXA,KAAAA,IAAAA,cAAc,aAAa,AAAD,KAAK,GAAG,qBAAqB,EAAEC,AAAAA,SAAAA,CAAAA,iBAAAA,OAAO,KAAK,AAAD,IAAXA,KAAAA,IAAAA,eAAc,iBAAiB,AAAD,KAAK,GAAG,gBAAgB,EAAEC,AAAAA,SAAAA,CAAAA,iBAAAA,OAAO,KAAK,AAAD,IAAXA,KAAAA,IAAAA,eAAc,YAAY,AAAD,KAAK,GAAG,WAAW,EAAEhB,SAAS,aAAa,EAAEiB,OAAO,WAAW,IAAI,IAAI;gBAG3TxB,mBACE,CAAC,oBAAoB,EAAEyB,KAAK,SAAS,CAACD,OAAO,KAAK,GAAG;gBAGvDhD,OACEgD,OAAO,OAAO,EACd,CAAC,mCAAmC,EAAEC,KAAK,SAAS,CAACD,SAAS;gBAEhEpB,UAAUoB,OAAO,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,OAAO;gBAC3ClB,QAAQkB,OAAO,KAAK;YACtB;YAEA1B,UAAU,CAAC,UAAU,EAAEM,SAAS;YAChC5B,OAAO4B,SAAS;QAClB,OAAO,IAAIX,AAAU,gBAAVA,OAAuB;YAChC,MAAMiC,sBAAsB,CAACtB;gBAC3B,IAAIA,AAAiB,gBAAjBA,QAAQ,IAAI,EAAkB;oBAChC,MAAMuB,YAAYvB,QAAQ,SAAS,CAAC,GAAG;oBACvC5B,OAAOmD,WAAW;oBAClB,MAAM,EAAEC,QAAQ,EAAEC,IAAI,EAAE,GAAGC,YAAY1B,QAAQ,SAAS,CAAC,GAAG;oBAC5D,OAAO;wBACL,QAAQ;4BACN,MAAM;4BACN,YAAYwB;4BACZ,MAAMC;wBACR;wBACA,MAAM;oBACR;gBACF;gBACA,OAAOzB;YACT;YAEA,IAAID,aAAa;gBACf,MAAMO,SAAU,MAAMlB,WAAW,MAAM,CAAC;oBACtC,OAAOzC;oBACP,QAAQ;oBACR,UAAUuC,SAAS,GAAG,CAAC,CAACyC,IAAO;4BAC7B,MAAM;4BACN,SAASC,MAAM,OAAO,CAACD,EAAE,OAAO,IAC3BA,EAAE,OAAO,CAAS,GAAG,CAACL,uBACvBK,EAAE,OAAO;wBACf;oBACA,iBAAiBrC;oBACjB,GAAGc,YAAY;gBACjB;gBAEA,WAAW,MAAMG,SAASD,OAAQ;wBAChBuB;oBAAhB,MAAM7B,UAAU6B,AAAAA,SAAAA,CAAAA,eAAAA,MAAM,KAAK,AAAD,IAAVA,KAAAA,IAAAA,aAAa,IAAI,AAAD,KAAK;oBACrC,IAAI7B,SAAS;wBACXC,eAAeD;wBACf,MAAMY,YAAiC;4BACrCZ;4BACAC;4BACA,mBAAmB;4BACnB,YAAY;4BACZ,OAAOY;wBACT;wBACA1B,QAAQ,OAAO,CAAEyB;oBACnB;oBAGA,IAAIL,AAAe,mBAAfA,MAAM,IAAI,EAAqB;wBACjCJ,WAAWL,KAAK,GAAG,KAAKD;wBACxB,MAAMiC,iBAAiBvB,MAAM,KAAK;wBAGlC,MAAMS,aAAkC;4BACtC,SAAS;4BACTf;4BACA,mBAAmB;4BACnB,YAAY;4BACZ,OAAO6B,iBACH;gCACE,eAAeA,eAAe,YAAY,IAAI;gCAC9C,mBAAmBA,eAAe,aAAa,IAAI;gCACnD,cACGA,AAAAA,CAAAA,eAAe,YAAY,IAAI,KAC/BA,CAAAA,eAAe,aAAa,IAAI;gCACnC,WAAW3B,YAAY;gCACvB,YAAYxD;gCACZ,mBAAmBc;gCACnB,QAAQjB,YAAY,MAAM;4BAC5B,IACAqE;wBACN;wBACA1B,QAAQ,OAAO,CAAE6B;wBACjB;oBACF;gBACF;gBACAhB,UAAUC;YACZ,OAAO;gBACL,MAAMmB,SAAS,MAAMhC,WAAW,MAAM,CAAC;oBACrC,OAAOzC;oBACP,QAAQ;oBACR,UAAUuC,SAAS,GAAG,CAAC,CAACyC,IAAO;4BAC7B,MAAM;4BACN,SAASC,MAAM,OAAO,CAACD,EAAE,OAAO,IAC3BA,EAAE,OAAO,CAAS,GAAG,CAACL,uBACvBK,EAAE,OAAO;wBACf;oBACA,iBAAiBrC;oBACjB,GAAGc,YAAY;gBACjB;gBACAD,WAAWL,KAAK,GAAG,KAAKD;gBACxBG,UAAWoB,OAAe,OAAO,CAAC,EAAE,CAAC,IAAI;gBACzClB,QAAQkB,OAAO,KAAK;YACtB;YAEAhD,OAAO4B,SAAS;QAClB;QAEA,IAAID,eAAe,CAACG,OAAO;YAEzB,MAAMY,kBAAkBC,KAAK,GAAG,CAC9B,GACAA,KAAK,KAAK,CAAEf,AAAAA,CAAAA,WAAW,EAAC,EAAG,MAAM,GAAG;YAEtCE,QAAQ;gBACN,eAAeY;gBACf,mBAAmBA;gBACnB,cAAcA,AAAkB,IAAlBA;YAChB;QACF;QAEA,OAAO;YACL,SAASd,WAAW;YACpB,OAAOE,QACH;gBACE,eAAeA,MAAM,aAAa,IAAI;gBACtC,mBAAmBA,MAAM,iBAAiB,IAAI;gBAC9C,cAAcA,MAAM,YAAY,IAAI;gBACpC,WAAWC,YAAY;gBACvB,YAAYxD;gBACZ,mBAAmBc;gBACnB,QAAQjB,YAAY,MAAM;YAC5B,IACAqE;YACJ,YAAY,CAAC,CAACd;QAChB;IACF,EAAE,OAAOgC,GAAQ;QACfjD,QAAQ,KAAK,CAAC,kBAAkBiD;QAChC,MAAMC,WAAW,IAAInD,MACnB,CAAC,eAAe,EAAEkB,cAAc,eAAe,GAAG,kBAAkB,EAAEgC,EAAE,OAAO,CAAC,8DAA8D,CAAC,EAC/I;YACE,OAAOA;QACT;QAEF,MAAMC;IACR;AACF;AAEO,MAAMzC,oBAAoB,CAC/B5C,WACAJ;IAIA,IAAI+C;IAKJ,IAAI3C,UAAU,QAAQ,CAAC,UACrB,OAAQJ;QACN,KAAK0F,aAAa,MAAM;YACtB3C,iBAAiB4C;YACjB;QACF,KAAKD,aAAa,eAAe;YAC/B3C,iBAAiB6C;YACjB;QACF,KAAKF,aAAa,IAAI;YACpB3C,iBAAiB8C;YACjB;QACF,KAAKH,aAAa,YAAY;QAC9B,KAAKA,aAAa,gBAAgB;YAChC3C,iBAAiB;gBAAE,MAAM+C,iBAAiB,IAAI;YAAC;YAC/C;QACF,KAAKJ,aAAa,IAAI;YAEpB3C,iBAAiBuB;YACjB;IACJ;IAKF,IACElE,AAAc,wBAAdA,aACAJ,sBAAsB0F,aAAa,IAAI,EAEvC3C,iBAAiB;QAAE,MAAM+C,iBAAiB,IAAI;IAAC;IAGjD,OAAO/C;AACT;AAEO,eAAegD,yBACpBpD,QAAsC,EACtC3C,iBAA+B,EAC/BC,WAAyB;IAEzB,MAAM+F,WAAW,MAAMtD,OAAOC,UAAU3C,mBAAmBC;IAC3D4B,OAAOmE,UAAU;IACjB,MAAM5E,SAASnB,YAAY,MAAM;IACjC,MAAMgG,cAAcC,cAAcF,SAAS,OAAO,EAAE5E;IACpD,OAAO;QAAE,SAAS6E;QAAa,OAAOD,SAAS,KAAK;IAAC;AACvD;AAEO,eAAeG,yBACpBC,IAAY,EACZpG,iBAA+B,EAC/BC,WAAyB;IAEzB,MAAM,EAAEwD,OAAO,EAAEE,KAAK,EAAE,GAAG,MAAMjB,OAAO0D,MAAMpG,mBAAmBC;IACjE,OAAO;QAAEwD;QAASE;IAAM;AAC1B;AAEO,SAAS0C,yBAAyBL,QAAgB;IACvD,IAAI;QAEF,MAAMM,YAAYN,SAAS,KAAK,CAAC;QACjC,IAAIM,WACF,OAAOA,SAAS,CAAC,EAAE;QAIrB,MAAMC,iBAAiBP,SAAS,KAAK,CACnC;QAEF,IAAIO,gBACF,OAAOA,cAAc,CAAC,EAAE;QAI1B,MAAMC,gBAAgBR,SAAS,KAAK,CAAC;QACrC,IAAIQ,eACF,OAAOA,aAAa,CAAC,EAAE;IAE3B,EAAE,OAAM,CAAC;IAET,OAAOR;AACT;AAEO,SAASS,yBAAyBC,KAAa;IACpD,IAAIA,MAAM,QAAQ,CAAC,SAEjB,MAAO,YAAY,IAAI,CAACA,OACtBA,QAAQA,MAAM,OAAO,CAAC,kBAAkB;IAG5C,OAAOA;AACT;AAEO,SAASR,cAAcQ,KAAa,EAAEtF,MAAgC;IAC3E,MAAMuF,kBAAkBN,yBAAyBK;IAEjD,IAAIC,QAAAA,kBAAAA,KAAAA,IAAAA,gBAAiB,KAAK,CAAC,oBAAoB;YACtCC;QAAP,OAAO,QAAAA,CAAAA,yBAAAA,gBACJ,KAAK,CAAC,kBAAiB,IADnBA,KAAAA,IAAAA,uBAEH,KAAK,CAAC,GACP,GAAG,CAAC9C;IACT;IACA,IAAI;QACF,OAAOgB,KAAK,KAAK,CAAC6B;IACpB,EAAE,OAAM,CAAC;IACT,IAAI;QACF,OAAO7B,KAAK,KAAK,CAAC+B,WAAWF;IAC/B,EAAE,OAAOnB,GAAG,CAAC;IAEb,IAAIpE,AAAW,oBAAXA,UAA8BA,AAAW,kBAAXA,QAA0B;QAC1D,MAAM0F,aAAaL,yBAAyBE;QAC5C,OAAO7B,KAAK,KAAK,CAAC+B,WAAWC;IAC/B;IACA,MAAMxE,MAAM,CAAC,+BAA+B,EAAEoE,OAAO;AACvD"}
+{"version":3,"file":"ai-model/service-caller/index.mjs","sources":["webpack://@midscene/core/./src/ai-model/service-caller/index.ts"],"sourcesContent":["import { AIResponseFormat, type AIUsageInfo } from '@/types';\nimport type { CodeGenerationChunk, StreamingCallback } from '@/types';\nimport {\n  type IModelConfig,\n  OPENAI_MAX_TOKENS,\n  type TVlModeTypes,\n  type UITarsModelVersion,\n  globalConfigManager,\n} from '@midscene/shared/env';\n\nimport { getDebug } from '@midscene/shared/logger';\nimport { assert } from '@midscene/shared/utils';\nimport { HttpsProxyAgent } from 'https-proxy-agent';\nimport { jsonrepair } from 'jsonrepair';\nimport OpenAI from 'openai';\nimport type { ChatCompletionMessageParam } from 'openai/resources/index';\nimport type { Stream } from 'openai/streaming';\nimport { SocksProxyAgent } from 'socks-proxy-agent';\nimport { AIActionType, type AIArgs } from '../common';\nimport { assertSchema } from '../prompt/assertion';\nimport { planSchema } from '../prompt/llm-planning';\n\nasync function createChatClient({\n  AIActionTypeValue,\n  modelConfig,\n}: {\n  AIActionTypeValue: AIActionType;\n  modelConfig: IModelConfig;\n}): Promise<{\n  completion: OpenAI.Chat.Completions;\n  modelName: string;\n  modelDescription: string;\n  uiTarsVersion?: UITarsModelVersion;\n  vlMode: TVlModeTypes | undefined;\n}> {\n  const {\n    socksProxy,\n    httpProxy,\n    modelName,\n    openaiBaseURL,\n    openaiApiKey,\n    openaiExtraConfig,\n    modelDescription,\n    uiTarsModelVersion: uiTarsVersion,\n    vlMode,\n    createOpenAIClient,\n  } = modelConfig;\n\n  let openai: OpenAI;\n\n  // Use custom client factory if provided\n  if (createOpenAIClient) {\n    openai = createOpenAIClient({\n      modelName,\n      openaiApiKey,\n      openaiBaseURL,\n      socksProxy,\n      httpProxy,\n      openaiExtraConfig,\n      vlMode,\n      intent: modelConfig.intent,\n      modelDescription,\n    });\n  } else {\n    // Default OpenAI client creation\n    let proxyAgent = undefined;\n    const debugProxy = getDebug('ai:call:proxy');\n    if (httpProxy) {\n      debugProxy('using http proxy', httpProxy);\n      proxyAgent = new HttpsProxyAgent(httpProxy);\n    } else if (socksProxy) {\n      debugProxy('using socks proxy', socksProxy);\n      proxyAgent = new SocksProxyAgent(socksProxy);\n    }\n\n    openai = new OpenAI({\n      baseURL: openaiBaseURL,\n      apiKey: openaiApiKey,\n      ...(proxyAgent ? { httpAgent: proxyAgent as any } : {}),\n      ...openaiExtraConfig,\n      dangerouslyAllowBrowser: true,\n    });\n  }\n\n  return {\n    completion: openai.chat.completions,\n    modelName,\n    modelDescription,\n    uiTarsVersion,\n    vlMode,\n  };\n}\n\nexport async function callAI(\n  messages: ChatCompletionMessageParam[],\n  AIActionTypeValue: AIActionType,\n  modelConfig: IModelConfig,\n  options?: {\n    stream?: boolean;\n    onChunk?: StreamingCallback;\n  },\n): Promise<{ content: string; usage?: AIUsageInfo; isStreamed: boolean }> {\n  const { completion, modelName, modelDescription, uiTarsVersion, vlMode } =\n    await createChatClient({\n      AIActionTypeValue,\n      modelConfig,\n    });\n\n  const responseFormat = getResponseFormat(modelName, AIActionTypeValue);\n\n  const maxTokens = globalConfigManager.getEnvConfigValue(OPENAI_MAX_TOKENS);\n  const debugCall = getDebug('ai:call');\n  const debugProfileStats = getDebug('ai:profile:stats');\n  const debugProfileDetail = getDebug('ai:profile:detail');\n\n  const startTime = Date.now();\n\n  const isStreaming = options?.stream && options?.onChunk;\n  let content: string | undefined;\n  let accumulated = '';\n  let usage: OpenAI.CompletionUsage | undefined;\n  let timeCost: number | undefined;\n\n  const commonConfig = {\n    temperature: vlMode === 'vlm-ui-tars' ? 0.0 : 0.1,\n    stream: !!isStreaming,\n    max_tokens:\n      typeof maxTokens === 'number'\n        ? maxTokens\n        : Number.parseInt(maxTokens || '2048', 10),\n    ...(vlMode === 'qwen-vl' // qwen vl v2 specific config\n      ? {\n          vl_high_resolution_images: true,\n        }\n      : {}),\n  };\n\n  try {\n    debugCall(\n      `sending ${isStreaming ? 'streaming ' : ''}request to ${modelName}`,\n    );\n\n    if (isStreaming) {\n      const stream = (await completion.create(\n        {\n          model: modelName,\n          messages,\n          response_format: responseFormat,\n          ...commonConfig,\n        },\n        {\n          stream: true,\n        },\n      )) as Stream<OpenAI.Chat.Completions.ChatCompletionChunk> & {\n        _request_id?: string | null;\n      };\n\n      for await (const chunk of stream) {\n        const content = chunk.choices?.[0]?.delta?.content || '';\n        const reasoning_content =\n          (chunk.choices?.[0]?.delta as any)?.reasoning_content || '';\n\n        // Check for usage info in any chunk (OpenAI provides usage in separate chunks)\n        if (chunk.usage) {\n          usage = chunk.usage;\n        }\n\n        if (content || reasoning_content) {\n          accumulated += content;\n          const chunkData: CodeGenerationChunk = {\n            content,\n            reasoning_content,\n            accumulated,\n            isComplete: false,\n            usage: undefined,\n          };\n          options.onChunk!(chunkData);\n        }\n\n        // Check if stream is complete\n        if (chunk.choices?.[0]?.finish_reason) {\n          timeCost = Date.now() - startTime;\n\n          // If usage is not available from the stream, provide a basic usage info\n          if (!usage) {\n            // Estimate token counts based on content length (rough approximation)\n            const estimatedTokens = Math.max(\n              1,\n              Math.floor(accumulated.length / 4),\n            );\n            usage = {\n              prompt_tokens: estimatedTokens,\n              completion_tokens: estimatedTokens,\n              total_tokens: estimatedTokens * 2,\n            };\n          }\n\n          // Send final chunk\n          const finalChunk: CodeGenerationChunk = {\n            content: '',\n            accumulated,\n            reasoning_content: '',\n            isComplete: true,\n            usage: {\n              prompt_tokens: usage.prompt_tokens ?? 0,\n              completion_tokens: usage.completion_tokens ?? 0,\n              total_tokens: usage.total_tokens ?? 0,\n              time_cost: timeCost ?? 0,\n              model_name: modelName,\n              model_description: modelDescription,\n              intent: modelConfig.intent,\n            },\n          };\n          options.onChunk!(finalChunk);\n          break;\n        }\n      }\n      content = accumulated;\n      debugProfileStats(\n        `streaming model, ${modelName}, mode, ${vlMode || 'default'}, cost-ms, ${timeCost}`,\n      );\n    } else {\n      const result = await completion.create({\n        model: modelName,\n        messages,\n        response_format: responseFormat,\n        ...commonConfig,\n      } as any);\n      timeCost = Date.now() - startTime;\n\n      debugProfileStats(\n        `model, ${modelName}, mode, ${vlMode || 'default'}, ui-tars-version, ${uiTarsVersion}, prompt-tokens, ${result.usage?.prompt_tokens || ''}, completion-tokens, ${result.usage?.completion_tokens || ''}, total-tokens, ${result.usage?.total_tokens || ''}, cost-ms, ${timeCost}, requestId, ${result._request_id || ''}`,\n      );\n\n      debugProfileDetail(`model usage detail: ${JSON.stringify(result.usage)}`);\n\n      assert(\n        result.choices,\n        `invalid response from LLM service: ${JSON.stringify(result)}`,\n      );\n      content = result.choices[0].message.content!;\n      usage = result.usage;\n    }\n\n    debugCall(`response: ${content}`);\n    assert(content, 'empty content');\n\n    // Ensure we always have usage info for streaming responses\n    if (isStreaming && !usage) {\n      // Estimate token counts based on content length (rough approximation)\n      const estimatedTokens = Math.max(\n        1,\n        Math.floor((content || '').length / 4),\n      );\n      usage = {\n        prompt_tokens: estimatedTokens,\n        completion_tokens: estimatedTokens,\n        total_tokens: estimatedTokens * 2,\n      };\n    }\n\n    return {\n      content: content || '',\n      usage: usage\n        ? {\n            prompt_tokens: usage.prompt_tokens ?? 0,\n            completion_tokens: usage.completion_tokens ?? 0,\n            total_tokens: usage.total_tokens ?? 0,\n            time_cost: timeCost ?? 0,\n            model_name: modelName,\n            model_description: modelDescription,\n            intent: modelConfig.intent,\n          }\n        : undefined,\n      isStreamed: !!isStreaming,\n    };\n  } catch (e: any) {\n    console.error(' call AI error', e);\n    const newError = new Error(\n      `failed to call ${isStreaming ? 'streaming ' : ''}AI model service: ${e.message}. Trouble shooting: https://midscenejs.com/model-provider.html`,\n      {\n        cause: e,\n      },\n    );\n    throw newError;\n  }\n}\n\nexport const getResponseFormat = (\n  modelName: string,\n  AIActionTypeValue: AIActionType,\n):\n  | OpenAI.ChatCompletionCreateParams['response_format']\n  | OpenAI.ResponseFormatJSONObject => {\n  let responseFormat:\n    | OpenAI.ChatCompletionCreateParams['response_format']\n    | OpenAI.ResponseFormatJSONObject\n    | undefined;\n\n  if (modelName.includes('gpt-4')) {\n    switch (AIActionTypeValue) {\n      case AIActionType.ASSERT:\n        responseFormat = assertSchema;\n        break;\n      case AIActionType.PLAN:\n        responseFormat = planSchema;\n        break;\n      case AIActionType.EXTRACT_DATA:\n      case AIActionType.DESCRIBE_ELEMENT:\n        responseFormat = { type: AIResponseFormat.JSON };\n        break;\n      case AIActionType.TEXT:\n        // No response format for plain text - return as-is\n        responseFormat = undefined;\n        break;\n    }\n  }\n\n  // gpt-4o-2024-05-13 only supports json_object response format\n  // Skip for plain text to allow string output\n  if (\n    modelName === 'gpt-4o-2024-05-13' &&\n    AIActionTypeValue !== AIActionType.TEXT\n  ) {\n    responseFormat = { type: AIResponseFormat.JSON };\n  }\n\n  return responseFormat;\n};\n\nexport async function callAIWithObjectResponse<T>(\n  messages: ChatCompletionMessageParam[],\n  AIActionTypeValue: AIActionType,\n  modelConfig: IModelConfig,\n): Promise<{ content: T; usage?: AIUsageInfo }> {\n  const response = await callAI(messages, AIActionTypeValue, modelConfig);\n  assert(response, 'empty response');\n  const vlMode = modelConfig.vlMode;\n  const jsonContent = safeParseJson(response.content, vlMode);\n  return { content: jsonContent, usage: response.usage };\n}\n\nexport async function callAIWithStringResponse(\n  msgs: AIArgs,\n  AIActionTypeValue: AIActionType,\n  modelConfig: IModelConfig,\n): Promise<{ content: string; usage?: AIUsageInfo }> {\n  const { content, usage } = await callAI(msgs, AIActionTypeValue, modelConfig);\n  return { content, usage };\n}\n\nexport function extractJSONFromCodeBlock(response: string) {\n  try {\n    // First, try to match a JSON object directly in the response\n    const jsonMatch = response.match(/^\\s*(\\{[\\s\\S]*\\})\\s*$/);\n    if (jsonMatch) {\n      return jsonMatch[1];\n    }\n\n    // If no direct JSON object is found, try to extract JSON from a code block\n    const codeBlockMatch = response.match(\n      /```(?:json)?\\s*(\\{[\\s\\S]*?\\})\\s*```/,\n    );\n    if (codeBlockMatch) {\n      return codeBlockMatch[1];\n    }\n\n    // If no code block is found, try to find a JSON-like structure in the text\n    const jsonLikeMatch = response.match(/\\{[\\s\\S]*\\}/);\n    if (jsonLikeMatch) {\n      return jsonLikeMatch[0];\n    }\n  } catch {}\n  // If no JSON-like structure is found, return the original response\n  return response;\n}\n\nexport function preprocessDoubaoBboxJson(input: string) {\n  if (input.includes('bbox')) {\n    // when its values like 940 445 969 490, replace all /\\d+\\s+\\d+/g with /$1,$2/g\n    while (/\\d+\\s+\\d+/.test(input)) {\n      input = input.replace(/(\\d+)\\s+(\\d+)/g, '$1,$2');\n    }\n  }\n  return input;\n}\n\nexport function safeParseJson(input: string, vlMode: TVlModeTypes | undefined) {\n  const cleanJsonString = extractJSONFromCodeBlock(input);\n  // match the point\n  if (cleanJsonString?.match(/\\((\\d+),(\\d+)\\)/)) {\n    return cleanJsonString\n      .match(/\\((\\d+),(\\d+)\\)/)\n      ?.slice(1)\n      .map(Number);\n  }\n  try {\n    return JSON.parse(cleanJsonString);\n  } catch {}\n  try {\n    return JSON.parse(jsonrepair(cleanJsonString));\n  } catch (e) {}\n\n  if (vlMode === 'doubao-vision' || vlMode === 'vlm-ui-tars') {\n    const jsonString = preprocessDoubaoBboxJson(cleanJsonString);\n    return JSON.parse(jsonrepair(jsonString));\n  }\n  throw Error(`failed to parse json response: ${input}`);\n}\n"],"names":["createChatClient","AIActionTypeValue","modelConfig","socksProxy","httpProxy","modelName","openaiBaseURL","openaiApiKey","openaiExtraConfig","modelDescription","uiTarsVersion","vlMode","createOpenAIClient","openai","proxyAgent","debugProxy","getDebug","HttpsProxyAgent","SocksProxyAgent","OpenAI","callAI","messages","options","completion","responseFormat","getResponseFormat","maxTokens","globalConfigManager","OPENAI_MAX_TOKENS","debugCall","debugProfileStats","debugProfileDetail","startTime","Date","isStreaming","content","accumulated","usage","timeCost","commonConfig","Number","stream","chunk","_chunk_choices__delta","_chunk_choices__delta1","_chunk_choices_2","reasoning_content","chunkData","undefined","estimatedTokens","Math","finalChunk","_result_usage","_result_usage1","_result_usage2","result","JSON","assert","e","console","newError","Error","AIActionType","assertSchema","planSchema","AIResponseFormat","callAIWithObjectResponse","response","jsonContent","safeParseJson","callAIWithStringResponse","msgs","extractJSONFromCodeBlock","jsonMatch","codeBlockMatch","jsonLikeMatch","preprocessDoubaoBboxJson","input","cleanJsonString","_cleanJsonString_match","jsonrepair","jsonString"],"mappings":";;;;;;;;;;;AAsBA,eAAeA,iBAAiB,EAC9BC,iBAAiB,EACjBC,WAAW,EAIZ;IAOC,MAAM,EACJC,UAAU,EACVC,SAAS,EACTC,SAAS,EACTC,aAAa,EACbC,YAAY,EACZC,iBAAiB,EACjBC,gBAAgB,EAChB,oBAAoBC,aAAa,EACjCC,MAAM,EACNC,kBAAkB,EACnB,GAAGV;IAEJ,IAAIW;IAGJ,IAAID,oBACFC,SAASD,mBAAmB;QAC1BP;QACAE;QACAD;QACAH;QACAC;QACAI;QACAG;QACA,QAAQT,YAAY,MAAM;QAC1BO;IACF;SACK;QAEL,IAAIK;QACJ,MAAMC,aAAaC,SAAS;QAC5B,IAAIZ,WAAW;YACbW,WAAW,oBAAoBX;YAC/BU,aAAa,IAAIG,gBAAgBb;QACnC,OAAO,IAAID,YAAY;YACrBY,WAAW,qBAAqBZ;YAChCW,aAAa,IAAII,gBAAgBf;QACnC;QAEAU,SAAS,IAAIM,SAAO;YAClB,SAASb;YACT,QAAQC;YACR,GAAIO,aAAa;gBAAE,WAAWA;YAAkB,IAAI,CAAC,CAAC;YACtD,GAAGN,iBAAiB;YACpB,yBAAyB;QAC3B;IACF;IAEA,OAAO;QACL,YAAYK,OAAO,IAAI,CAAC,WAAW;QACnCR;QACAI;QACAC;QACAC;IACF;AACF;AAEO,eAAeS,OACpBC,QAAsC,EACtCpB,iBAA+B,EAC/BC,WAAyB,EACzBoB,OAGC;IAED,MAAM,EAAEC,UAAU,EAAElB,SAAS,EAAEI,gBAAgB,EAAEC,aAAa,EAAEC,MAAM,EAAE,GACtE,MAAMX,iBAAiB;QACrBC;QACAC;IACF;IAEF,MAAMsB,iBAAiBC,kBAAkBpB,WAAWJ;IAEpD,MAAMyB,YAAYC,oBAAoB,iBAAiB,CAACC;IACxD,MAAMC,YAAYb,SAAS;IAC3B,MAAMc,oBAAoBd,SAAS;IACnC,MAAMe,qBAAqBf,SAAS;IAEpC,MAAMgB,YAAYC,KAAK,GAAG;IAE1B,MAAMC,cAAcZ,AAAAA,CAAAA,QAAAA,UAAAA,KAAAA,IAAAA,QAAS,MAAM,AAAD,KAAKA,CAAAA,QAAAA,UAAAA,KAAAA,IAAAA,QAAS,OAAO,AAAD;IACtD,IAAIa;IACJ,IAAIC,cAAc;IAClB,IAAIC;IACJ,IAAIC;IAEJ,MAAMC,eAAe;QACnB,aAAa5B,AAAW,kBAAXA,SAA2B,MAAM;QAC9C,QAAQ,CAAC,CAACuB;QACV,YACE,AAAqB,YAArB,OAAOR,YACHA,YACAc,OAAO,QAAQ,CAACd,aAAa,QAAQ;QAC3C,GAAIf,AAAW,cAAXA,SACA;YACE,2BAA2B;QAC7B,IACA,CAAC,CAAC;IACR;IAEA,IAAI;QACFkB,UACE,CAAC,QAAQ,EAAEK,cAAc,eAAe,GAAG,WAAW,EAAE7B,WAAW;QAGrE,IAAI6B,aAAa;YACf,MAAMO,SAAU,MAAMlB,WAAW,MAAM,CACrC;gBACE,OAAOlB;gBACPgB;gBACA,iBAAiBG;gBACjB,GAAGe,YAAY;YACjB,GACA;gBACE,QAAQ;YACV;YAKF,WAAW,MAAMG,SAASD,OAAQ;oBAChBE,uBAAAA,iBAAAA,gBAEbC,wBAAAA,kBAAAA,iBAoBCC,kBAAAA;gBAtBJ,MAAMV,UAAUQ,AAAAA,SAAAA,CAAAA,iBAAAA,MAAM,OAAO,AAAD,IAAZA,KAAAA,IAAAA,QAAAA,CAAAA,kBAAAA,cAAe,CAAC,EAAE,AAAD,IAAjBA,KAAAA,IAAAA,QAAAA,CAAAA,wBAAAA,gBAAoB,KAAK,AAAD,IAAxBA,KAAAA,IAAAA,sBAA2B,OAAO,AAAD,KAAK;gBACtD,MAAMG,oBACJ,AAAC,SAAAF,CAAAA,kBAAAA,MAAM,OAAO,AAAD,IAAZA,KAAAA,IAAAA,QAAAA,CAAAA,mBAAAA,eAAe,CAAC,EAAE,AAAD,IAAjBA,KAAAA,IAAAA,QAAAA,CAAAA,yBAAAA,iBAAoB,KAAK,AAAD,IAAxBA,KAAAA,IAAAA,uBAAmC,iBAAiB,AAAD,KAAK;gBAG3D,IAAIF,MAAM,KAAK,EACbL,QAAQK,MAAM,KAAK;gBAGrB,IAAIP,WAAWW,mBAAmB;oBAChCV,eAAeD;oBACf,MAAMY,YAAiC;wBACrCZ;wBACAW;wBACAV;wBACA,YAAY;wBACZ,OAAOY;oBACT;oBACA1B,QAAQ,OAAO,CAAEyB;gBACnB;gBAGA,IAAI,QAAAF,CAAAA,kBAAAA,MAAM,OAAO,AAAD,IAAZA,KAAAA,IAAAA,QAAAA,CAAAA,mBAAAA,eAAe,CAAC,EAAE,AAAD,IAAjBA,KAAAA,IAAAA,iBAAoB,aAAa,EAAE;oBACrCP,WAAWL,KAAK,GAAG,KAAKD;oBAGxB,IAAI,CAACK,OAAO;wBAEV,MAAMY,kBAAkBC,KAAK,GAAG,CAC9B,GACAA,KAAK,KAAK,CAACd,YAAY,MAAM,GAAG;wBAElCC,QAAQ;4BACN,eAAeY;4BACf,mBAAmBA;4BACnB,cAAcA,AAAkB,IAAlBA;wBAChB;oBACF;oBAGA,MAAME,aAAkC;wBACtC,SAAS;wBACTf;wBACA,mBAAmB;wBACnB,YAAY;wBACZ,OAAO;4BACL,eAAeC,MAAM,aAAa,IAAI;4BACtC,mBAAmBA,MAAM,iBAAiB,IAAI;4BAC9C,cAAcA,MAAM,YAAY,IAAI;4BACpC,WAAWC,YAAY;4BACvB,YAAYjC;4BACZ,mBAAmBI;4BACnB,QAAQP,YAAY,MAAM;wBAC5B;oBACF;oBACAoB,QAAQ,OAAO,CAAE6B;oBACjB;gBACF;YACF;YACAhB,UAAUC;YACVN,kBACE,CAAC,iBAAiB,EAAEzB,UAAU,QAAQ,EAAEM,UAAU,UAAU,WAAW,EAAE2B,UAAU;QAEvF,OAAO;gBAUqGc,eAAyDC,gBAAwDC;YAT3N,MAAMC,SAAS,MAAMhC,WAAW,MAAM,CAAC;gBACrC,OAAOlB;gBACPgB;gBACA,iBAAiBG;gBACjB,GAAGe,YAAY;YACjB;YACAD,WAAWL,KAAK,GAAG,KAAKD;YAExBF,kBACE,CAAC,OAAO,EAAEzB,UAAU,QAAQ,EAAEM,UAAU,UAAU,mBAAmB,EAAED,cAAc,iBAAiB,EAAE0C,AAAAA,SAAAA,CAAAA,gBAAAA,OAAO,KAAK,AAAD,IAAXA,KAAAA,IAAAA,cAAc,aAAa,AAAD,KAAK,GAAG,qBAAqB,EAAEC,AAAAA,SAAAA,CAAAA,iBAAAA,OAAO,KAAK,AAAD,IAAXA,KAAAA,IAAAA,eAAc,iBAAiB,AAAD,KAAK,GAAG,gBAAgB,EAAEC,AAAAA,SAAAA,CAAAA,iBAAAA,OAAO,KAAK,AAAD,IAAXA,KAAAA,IAAAA,eAAc,YAAY,AAAD,KAAK,GAAG,WAAW,EAAEhB,SAAS,aAAa,EAAEiB,OAAO,WAAW,IAAI,IAAI;YAG3TxB,mBAAmB,CAAC,oBAAoB,EAAEyB,KAAK,SAAS,CAACD,OAAO,KAAK,GAAG;YAExEE,OACEF,OAAO,OAAO,EACd,CAAC,mCAAmC,EAAEC,KAAK,SAAS,CAACD,SAAS;YAEhEpB,UAAUoB,OAAO,OAAO,CAAC,EAAE,CAAC,OAAO,CAAC,OAAO;YAC3ClB,QAAQkB,OAAO,KAAK;QACtB;QAEA1B,UAAU,CAAC,UAAU,EAAEM,SAAS;QAChCsB,OAAOtB,SAAS;QAGhB,IAAID,eAAe,CAACG,OAAO;YAEzB,MAAMY,kBAAkBC,KAAK,GAAG,CAC9B,GACAA,KAAK,KAAK,CAAEf,AAAAA,CAAAA,WAAW,EAAC,EAAG,MAAM,GAAG;YAEtCE,QAAQ;gBACN,eAAeY;gBACf,mBAAmBA;gBACnB,cAAcA,AAAkB,IAAlBA;YAChB;QACF;QAEA,OAAO;YACL,SAASd,WAAW;YACpB,OAAOE,QACH;gBACE,eAAeA,MAAM,aAAa,IAAI;gBACtC,mBAAmBA,MAAM,iBAAiB,IAAI;gBAC9C,cAAcA,MAAM,YAAY,IAAI;gBACpC,WAAWC,YAAY;gBACvB,YAAYjC;gBACZ,mBAAmBI;gBACnB,QAAQP,YAAY,MAAM;YAC5B,IACA8C;YACJ,YAAY,CAAC,CAACd;QAChB;IACF,EAAE,OAAOwB,GAAQ;QACfC,QAAQ,KAAK,CAAC,kBAAkBD;QAChC,MAAME,WAAW,IAAIC,MACnB,CAAC,eAAe,EAAE3B,cAAc,eAAe,GAAG,kBAAkB,EAAEwB,EAAE,OAAO,CAAC,8DAA8D,CAAC,EAC/I;YACE,OAAOA;QACT;QAEF,MAAME;IACR;AACF;AAEO,MAAMnC,oBAAoB,CAC/BpB,WACAJ;IAIA,IAAIuB;IAKJ,IAAInB,UAAU,QAAQ,CAAC,UACrB,OAAQJ;QACN,KAAK6D,aAAa,MAAM;YACtBtC,iBAAiBuC;YACjB;QACF,KAAKD,aAAa,IAAI;YACpBtC,iBAAiBwC;YACjB;QACF,KAAKF,aAAa,YAAY;QAC9B,KAAKA,aAAa,gBAAgB;YAChCtC,iBAAiB;gBAAE,MAAMyC,iBAAiB,IAAI;YAAC;YAC/C;QACF,KAAKH,aAAa,IAAI;YAEpBtC,iBAAiBwB;YACjB;IACJ;IAKF,IACE3C,AAAc,wBAAdA,aACAJ,sBAAsB6D,aAAa,IAAI,EAEvCtC,iBAAiB;QAAE,MAAMyC,iBAAiB,IAAI;IAAC;IAGjD,OAAOzC;AACT;AAEO,eAAe0C,yBACpB7C,QAAsC,EACtCpB,iBAA+B,EAC/BC,WAAyB;IAEzB,MAAMiE,WAAW,MAAM/C,OAAOC,UAAUpB,mBAAmBC;IAC3DuD,OAAOU,UAAU;IACjB,MAAMxD,SAAST,YAAY,MAAM;IACjC,MAAMkE,cAAcC,cAAcF,SAAS,OAAO,EAAExD;IACpD,OAAO;QAAE,SAASyD;QAAa,OAAOD,SAAS,KAAK;IAAC;AACvD;AAEO,eAAeG,yBACpBC,IAAY,EACZtE,iBAA+B,EAC/BC,WAAyB;IAEzB,MAAM,EAAEiC,OAAO,EAAEE,KAAK,EAAE,GAAG,MAAMjB,OAAOmD,MAAMtE,mBAAmBC;IACjE,OAAO;QAAEiC;QAASE;IAAM;AAC1B;AAEO,SAASmC,yBAAyBL,QAAgB;IACvD,IAAI;QAEF,MAAMM,YAAYN,SAAS,KAAK,CAAC;QACjC,IAAIM,WACF,OAAOA,SAAS,CAAC,EAAE;QAIrB,MAAMC,iBAAiBP,SAAS,KAAK,CACnC;QAEF,IAAIO,gBACF,OAAOA,cAAc,CAAC,EAAE;QAI1B,MAAMC,gBAAgBR,SAAS,KAAK,CAAC;QACrC,IAAIQ,eACF,OAAOA,aAAa,CAAC,EAAE;IAE3B,EAAE,OAAM,CAAC;IAET,OAAOR;AACT;AAEO,SAASS,yBAAyBC,KAAa;IACpD,IAAIA,MAAM,QAAQ,CAAC,SAEjB,MAAO,YAAY,IAAI,CAACA,OACtBA,QAAQA,MAAM,OAAO,CAAC,kBAAkB;IAG5C,OAAOA;AACT;AAEO,SAASR,cAAcQ,KAAa,EAAElE,MAAgC;IAC3E,MAAMmE,kBAAkBN,yBAAyBK;IAEjD,IAAIC,QAAAA,kBAAAA,KAAAA,IAAAA,gBAAiB,KAAK,CAAC,oBAAoB;YACtCC;QAAP,OAAO,QAAAA,CAAAA,yBAAAA,gBACJ,KAAK,CAAC,kBAAiB,IADnBA,KAAAA,IAAAA,uBAEH,KAAK,CAAC,GACP,GAAG,CAACvC;IACT;IACA,IAAI;QACF,OAAOgB,KAAK,KAAK,CAACsB;IACpB,EAAE,OAAM,CAAC;IACT,IAAI;QACF,OAAOtB,KAAK,KAAK,CAACwB,WAAWF;IAC/B,EAAE,OAAOpB,GAAG,CAAC;IAEb,IAAI/C,AAAW,oBAAXA,UAA8BA,AAAW,kBAAXA,QAA0B;QAC1D,MAAMsE,aAAaL,yBAAyBE;QAC5C,OAAOtB,KAAK,KAAK,CAACwB,WAAWC;IAC/B;IACA,MAAMpB,MAAM,CAAC,+BAA+B,EAAEgB,OAAO;AACvD"}

package/dist/es/index.mjs

@@ -1,11 +1,12 @@
 import { z } from "zod";
-import { Executor } from "./ai-model/action-executor.mjs";
+import { TaskRunner } from "./task-runner.mjs";
 import insight from "./insight/index.mjs";
 import { getVersion } from "./utils.mjs";
 import { AiLocateElement, PointSchema, RectSchema, SizeSchema, TMultimodalPromptSchema, TUserPromptSchema, describeUserPage, getMidsceneLocationSchema, plan } from "./ai-model/index.mjs";
 import { MIDSCENE_MODEL_NAME } from "@midscene/shared/env";
+import { InsightError } from "./types.mjs";
 import { Agent, createAgent } from "./agent/index.mjs";
 const src = insight;
-export { Agent, AiLocateElement, Executor, insight as Insight, MIDSCENE_MODEL_NAME, PointSchema, RectSchema, SizeSchema, TMultimodalPromptSchema, TUserPromptSchema, createAgent, src as default, describeUserPage, getMidsceneLocationSchema, getVersion, plan, z };
+export { Agent, AiLocateElement, insight as Insight, InsightError, MIDSCENE_MODEL_NAME, PointSchema, RectSchema, SizeSchema, TMultimodalPromptSchema, TUserPromptSchema, TaskRunner, createAgent, src as default, describeUserPage, getMidsceneLocationSchema, getVersion, plan, z };
 
 //# sourceMappingURL=index.mjs.map

package/dist/es/index.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","sources":["webpack://@midscene/core/./src/index.ts"],"sourcesContent":["import { z } from 'zod';\nimport { Executor } from './ai-model/action-executor';\nimport Insight from './insight/index';\nimport { getVersion } from './utils';\n\nexport {\n  plan,\n  describeUserPage,\n  AiLocateElement,\n  getMidsceneLocationSchema,\n  type MidsceneLocationResultType,\n  PointSchema,\n  SizeSchema,\n  RectSchema,\n  TMultimodalPromptSchema,\n  TUserPromptSchema,\n  type TMultimodalPrompt,\n  type TUserPrompt,\n} from './ai-model/index';\n\nexport { MIDSCENE_MODEL_NAME } from '@midscene/shared/env';\n\nexport type * from './types';\n\nexport { z };\n\nexport default Insight;\nexport { Executor, Insight, getVersion };\n\nexport type {\n  MidsceneYamlScript,\n  MidsceneYamlTask,\n  MidsceneYamlFlowItem,\n  MidsceneYamlConfigResult,\n  MidsceneYamlConfig,\n  MidsceneYamlScriptWebEnv,\n  MidsceneYamlScriptAndroidEnv,\n  MidsceneYamlScriptIOSEnv,\n  MidsceneYamlScriptEnv,\n  LocateOption,\n  DetailedLocateParam,\n} from './yaml';\n\nexport { Agent, type AgentOpt, createAgent } from './agent';\n"],"names":["Insight"],"mappings":";;;;;;;AA0BA,YAAeA"}
+{"version":3,"file":"index.mjs","sources":["webpack://@midscene/core/./src/index.ts"],"sourcesContent":["import { z } from 'zod';\nimport { TaskRunner } from './task-runner';\nimport Insight from './insight/index';\nimport { getVersion } from './utils';\n\nexport {\n  plan,\n  describeUserPage,\n  AiLocateElement,\n  getMidsceneLocationSchema,\n  type MidsceneLocationResultType,\n  PointSchema,\n  SizeSchema,\n  RectSchema,\n  TMultimodalPromptSchema,\n  TUserPromptSchema,\n  type TMultimodalPrompt,\n  type TUserPrompt,\n} from './ai-model/index';\n\nexport {\n  MIDSCENE_MODEL_NAME,\n  type CreateOpenAIClientFn,\n} from '@midscene/shared/env';\n\nexport type * from './types';\nexport { InsightError } from './types';\n\nexport { z };\n\nexport default Insight;\nexport { TaskRunner, Insight, getVersion };\n\nexport type {\n  MidsceneYamlScript,\n  MidsceneYamlTask,\n  MidsceneYamlFlowItem,\n  MidsceneYamlConfigResult,\n  MidsceneYamlConfig,\n  MidsceneYamlScriptWebEnv,\n  MidsceneYamlScriptAndroidEnv,\n  MidsceneYamlScriptIOSEnv,\n  MidsceneYamlScriptEnv,\n  LocateOption,\n  DetailedLocateParam,\n} from './yaml';\n\nexport { Agent, type AgentOpt, createAgent } from './agent';\n"],"names":["Insight"],"mappings":";;;;;;;;AA8BA,YAAeA"}