​Chat BI 界面​业务用户可通过自然语言查询数据并以图表形式可视化结果​Headless BI 界面​分析工程师可构建语义数据模型指标/维度/标签的定义及其含义和关系​可扩展架构​通过 Java SPI 机制支持自定义实现​语义增强​将数据语义业务术语、列值等注入 Prompt减少 LLM 幻觉​查询简化​将复杂 SQL 语法JOIN、公式等从 LLM 卸载到语义层架构组件组件 功能Knowledge Base 从语义模型提取 Schema 信息构建字典和索引Schema Mapper 识别用户查询中的 Schema 元素引用Semantic Parser 理解用户查询生成语义查询语句规则 LLMSemantic Corrector 检查语义查询语句的有效性并修正Semantic Translator 将语义查询语句转换为可执行的 SQLChat Plugin 通过第三方工具扩展功能Chat Memory 封装历史查询轨迹用于 few-shot promptingGitHub: https://github.com/tencentmusic/supersonic (⭐ 4.9k)Apache DolphinSchedulerApache DolphinScheduler 是 Apache 基金会的​现代数据编排平台​专注于处理数据管道中的复杂任务依赖。核心特性​易于部署​支持 Standalone、Cluster、Docker、Kubernetes 四种部署模式​易于使用​通过 Web UI、Python SDK 或 Open API 创建和管理工作流​高可靠​去中心化、多 Master 多 Worker 架构原生支持水平扩展​高性能​性能是其他编排平台的数倍每天可处理数千万任务​云原生​支持跨多云和数据中心编排工作流​版本控制​提供工作流和任务实例的版本控制​灵活状态控制​支持随时暂停、停止和恢复工作流多租户支持GitHub: https://github.com/apache/dolphinscheduler (⭐ 14.3k)dsctl (DolphinScheduler CLI)dsctl 是一个​命令行工具​用于管理 DolphinScheduler 资源。主要功能项目管理project工作流管理workflow任务管理task数据源管理datasource环境管理environment用户管理user监控和审计monitor/audit示例命令dsctl project list # 列出所有项目dsctl workflow run daily-etl # 运行工作流dsctl workflow-instance watch 123 # 监控工作流实例dsctl task-instance log 456 --raw # 获取任务日志集成背景为什么需要集成场景 传统方式 集成后运行工作流 登录 DS Web UI → 找到工作流 → 点击运行 对话输入运行 daily-etl 工作流监控实例 登录 DS Web UI → 查看实例列表 → 找到目标 对话输入监控实例 123查看日志 登录 DS Web UI → 找到任务实例 → 查看日志 对话输入查看任务 456 的日志批量操作 重复多次 UI 操作 对话输入批量指令集成价值​提升效率​减少 UI 操作步骤一句话完成复杂操作​降低门槛​非技术人员可通过自然语言操作工作流​统一入口​在 SuperSonic 对话界面中统一管理数据查询和工作流操作架构设计用户自然语言输入↓WorkflowParserChatQueryParser SPI├── accept(): 检查 Agent 是否配置了 DSCTL 工具└── parse(): 调用 LLM 将自然语言映射为 dsctl 子命令结果写入 SemanticParseInfo.properties↓WorkflowExecutorChatQueryExecutor SPI├── accept(): 检查 queryMode “WORKFLOW_CTL”└── execute(): 调用 ProcessBuilder 执行 dsctl 命令注入 DS_API_URL / DS_TOKEN 环境变量返回 QueryResult.textResultMarkdown 格式↓前端 ChatItem修改白名单支持 WORKFLOW_CTL 模式渲染数据流用户输入自然语言 → SuperSonic Chat APIWorkflowParser 接收请求 → 调用 LLM 生成 dsctl 命令WorkflowExecutor 执行命令 → 调用 dsctl 进程dsctl 调用 DolphinScheduler API → 返回结果结果格式化为 Markdown → 前端渲染后端实现文件 1AgentToolType.java修改路径chat/server/src/main/java/com/tencent/supersonic/chat/server/agent/AgentToolType.javapackage com.tencent.supersonic.chat.server.agent;import java.util.HashMap;import java.util.Map;publicenumAgentToolType {DATASET(“Text2SQL数据集”),PLUGIN(“第三方插件”),WORK_FLOW_CTL(“工作流命令行”);privatefinal String title; AgentToolType(String title) { this.title title; } publicstatic MapAgentToolType, String getToolTypes() { MapAgentToolType, String map newHashMap(); map.put(DATASET, DATASET.title); map.put(PLUGIN, PLUGIN.title); map.put(DSCTL, DSCTL.title); // 新增 return map; }}注册后前端 getToolTypes() API 会自动返回新类型Agent 管理页面的工具类型下拉框中会出现 “DolphinScheduler CLI”。文件 2WorkflowTool.java新建路径chat/server/src/main/java/com/tencent/supersonic/chat/server/agent/WorkflowTool.javapackage com.tencent.supersonic.chat.server.agent;import java.util.List;import lombok.AllArgsConstructor;import lombok.Data;import lombok.NoArgsConstructor;DataNoArgsConstructorAllArgsConstructorpublicclassWorkflowToolextendsAgentTool {/** dsctl 可执行文件路径默认 “dsctl” */privateStringdsctlPath“dsctl”;/** 默认使用的 DolphinScheduler 项目名 */private String defaultProject;private String dsApiUrl;private String dsToken;/** 示例问题用于 embedding 召回 */private List exampleQuestions;}文件 3WorkflowParser.java新建路径chat/server/src/main/java/com/tencent/supersonic/chat/server/parser/WorkflowParser.javapackage com.tencent.supersonic.chat.server.parser;import com.alibaba.fastjson2.JSON;import com.alibaba.fastjson2.JSONArray;import com.alibaba.fastjson2.JSONObject;import com.tencent.supersonic.chat.server.agent.AgentToolType;import com.tencent.supersonic.chat.server.agent.WorkflowTool;import com.tencent.supersonic.chat.server.pojo.ParseContext;import com.tencent.supersonic.common.pojo.ChatApp;import com.tencent.supersonic.common.pojo.enums.AppModule;import com.tencent.supersonic.common.util.ChatAppManager;import com.tencent.supersonic.headless.api.pojo.SemanticParseInfo;import com.tencent.supersonic.headless.api.pojo.response.ParseResp;import dev.langchain4j.model.chat.ChatModel;import dev.langchain4j.model.input.Prompt;import dev.langchain4j.model.input.PromptTemplate;import dev.langchain4j.model.output.structured.Description;import dev.langchain4j.provider.ModelProvider;import dev.langchain4j.service.AiServices;import java.io.BufferedReader;import java.io.InputStreamReader;import java.util.HashMap;import java.util.List;import java.util.Map;import java.util.Objects;import java.util.concurrent.TimeUnit;import java.util.stream.Collectors;import lombok.Data;import lombok.extern.slf4j.Slf4j;import org.apache.commons.collections4.CollectionUtils;/**Agent 管理页面 └── 添加工具: typeDSCTL, dsctlPath/usr/bin/dsctl, defaultProjectetl-prod ↓ 存入Agent.toolConfig (JSON) WorkflowParser.accept() └── agent.getTools(AgentToolType.DSCTL) 非空 → 激活WorkflowParser.parse() └── 读取 DsctlTool 配置 → LLM 解析命令 → 写入 parseInfo.propertiesWorkflowExecutor.execute() └── 读取 dsctl_path dsctl_cmd → Runtime.exec() → 返回结果*/Slf4jpublicclassWorkflowParserimplementsChatQueryParser {publicstaticfinalStringQUERY_MODE“WORKFLOW_CTL”;publicstaticfinalStringAPP_KEY“WORKFLOW_PARSER”;/** 缓存 dsctl schema 结果避免重复执行 */privatestaticvolatile String cachedSchema;privatestaticvolatilelong schemaCacheTime;privatestaticfinallongSCHEMA_CACHE_TTL_MS5 * 60 * 1000; // 5 minutesprivatestaticfinalStringINSTRUCTION“”# Role: You are a DolphinScheduler CLI expert using dsctl.# Task: Convert users natural language into dsctl commands. # dsctl Usage: {{dsctl_help}} # Available Commands: {{dsctl_commands}} # Command Generation Rules: 1. Match user intent to the appropriate command 2. For resource selectors, use discovery commands listed below 3. Include required arguments and options 4. Output ONLY the dsctl command # Intent Patterns: - list/view/show resource → dsctl resource list - get/detail resource → dsctl resource get selector - create/new resource → dsctl resource create --name name - delete/remove resource → dsctl resource delete selector --force - run/execute workflow → dsctl workflow run workflow - stop instance → dsctl workflow-instance stop id - watch/monitor instance → dsctl workflow-instance watch id - log task → dsctl task-instance log id --raw # Resource Selectors: - project: name or code (discover: dsctl project list) - workflow: name or code (discover: dsctl workflow list) - task: name or code (discover: dsctl task list) - workflow-instance: numeric id (discover: dsctl workflow-instance list) - task-instance: numeric id (discover: dsctl task-instance list) # Output Rules: 1. Output ONLY the dsctl command 2. If intent is unclear, output: unknown 3. Do NOT include explanations # Question: {{question}} # Command: ;publicWorkflowParser() {ChatAppManager.register(APP_KEY,ChatApp.builder().prompt(INSTRUCTION).name(“工作流命令行”).appModule(AppModule.CHAT).description(“将自然语言转换为工作流命令”).enable(true).build());}/** 结构化输出类Description 告诉 AiServices 每个字段的含义 */DatastaticclassDsctlCommand {Description(“the dsctl subcommand without ‘dsctl’ prefix, e.g. ‘workflow run daily-etl’”)private String command;Description(brief explanation in Chinese of what this command does) private String thought;}/** 提取接口AiServices 自动生成实现 */interfaceDsctlCommandExtractor {DsctlCommand extractCommand(String text);}Overridepublicbooleanaccept(ParseContext parseContext) {// 只有 Agent 配置了 Workflow Cli 工具才激活List tools parseContext.getAgent().getTools(AgentToolType.WORK_FLOW_CTL);return !CollectionUtils.isEmpty(tools);}Overridepublicvoidparse(ParseContext parseContext) {// 读取 Agent 上配置的 workflowTool取第一个List toolJsonList parseContext.getAgent().getTools(AgentToolType.WORK_FLOW_CTL);WorkflowToolworkflowTool JSONObject.parseObject(toolJsonList.getFirst(), WorkflowTool.class);// 读取 ChatApp 配置LLM 模型绑定 prompt 模板 ChatAppchatApp parseContext.getAgent().getChatAppConfig().get(APP_KEY); if (Objects.isNull(chatApp) || !chatApp.isEnable()) { log.warn(WorkflowParser ChatApp [{}] is not enabled, skip., APP_KEY); return; } // 构建 prompt替换占位符 MapString, Object variables newHashMap(); // 1. 执行 dsctl --help 获取使用说明 StringdsctlHelp executeDsctlCommand(workflowTool.getDsctlPath(), workflowTool, --help); // 2. 执行 dsctl schema 获取命令列表并格式化带缓存 StringdsctlSchema getCachedOrExecuteSchema(workflowTool); StringformattedCommands formatCommandsFromSchema(dsctlSchema); variables.put(dsctl_help, dsctlHelp); variables.put(dsctl_commands, formattedCommands); variables.put(question, parseContext.getRequest().getQueryText()); Promptprompt PromptTemplate.from(chatApp.getPrompt()).apply(variables); // 调用 LLM使用 AiServices 做结构化输出 ChatModelmodel ModelProvider.getChatModel(chatApp.getChatModelConfig()); DsctlCommandExtractorextractor AiServices.create(DsctlCommandExtractor.class, model); DsctlCommandcmd extractor.extractCommand(prompt.toUserMessage().singleText()); log.info( WorkflowParser input[{}] - command[{}] thought[{}], parseContext.getRequest().getQueryText(), cmd null ? null : cmd.getCommand(), cmd null ? null : cmd.getThought()); if (cmd null || unknown.equalsIgnoreCase(cmd.getCommand())) { // 无法识别不设置 parseInfo交给后续 parser 处理 return; } // 将解析结果写入 parseInfo供 DsctlExecutor 使用 SemanticParseInfoparseInfonewSemanticParseInfo(); parseInfo.setQueryMode(QUERY_MODE); parseInfo.setId(1); MapString, Object props newHashMap(); // e.g. workflow run daily-etl props.put(workflow_ctl_cmd, cmd.getCommand()); // LLM 的解释 props.put(workflow_ctl_thought, cmd.getThought()); // 可执行文件路径 props.put(workflow_ctl_path, workflowTool.getDsctlPath()); if (workflowTool.getDefaultProject() ! null) { props.put(default_project, workflowTool.getDefaultProject()); } if (workflowTool.getDsApiUrl() ! null) { props.put(ds_api_url, workflowTool.getDsApiUrl()); } if (workflowTool.getDsToken() ! null) { props.put(ds_token, workflowTool.getDsToken()); } parseInfo.setProperties(props); parseContext.getResponse().getSelectedParses().add(parseInfo); parseContext.getResponse().setState(ParseResp.ParseState.COMPLETED);}/**执行 dsctl 命令并返回输出param dsctlPath dsctl 可执行文件路径param workflowTool 工作流工具配置包含环境变量param args 命令参数return 命令输出*/private String executeDsctlCommand(String dsctlPath, WorkflowTool workflowTool, String… args) {try {String[] cmdArray newString[args.length 1];cmdArray[0] dsctlPath;System.arraycopy(args, 0, cmdArray, 1, args.length);ProcessBuilderpbnewProcessBuilder(cmdArray);pb.redirectErrorStream(true);// 添加环境变量MapString, String env pb.environment();if (workflowTool.getDsApiUrl() ! null) {env.put(“DS_API_URL”, workflowTool.getDsApiUrl());}if (workflowTool.getDsToken() ! null) {env.put(“DS_API_TOKEN”, workflowTool.getDsToken());}Processprocess pb.start();String output;try (BufferedReaderreadernewBufferedReader(newInputStreamReader(process.getInputStream()))) {output reader.lines().collect(Collectors.joining(“\n”));}booleanfinished process.waitFor(30, TimeUnit.SECONDS);if (!finished) {process.destroyForcibly();log.warn(“dsctl command timed out: {}”, String.join( “, cmdArray));return”;}if (process.exitValue() ! 0) {log.warn(“dsctl command failed with exit code {}: {}”, process.exitValue(), output);}return output;} catch (Exception e) {log.error(“Failed to execute dsctl command: {}”, e.getMessage(), e);return;}}/**获取缓存的 schema 或执行新命令param workflowTool 工作流工具配置return dsctl schema JSON*/private String getCachedOrExecuteSchema(WorkflowTool workflowTool) {longnow System.currentTimeMillis();if (cachedSchema ! null (now - schemaCacheTime) SCHEMA_CACHE_TTL_MS) {log.debug(“Using cached dsctl schema”);return cachedSchema;}synchronized (WorkflowParser.class) { // Double-check locking if (cachedSchema ! null (now - schemaCacheTime) SCHEMA_CACHE_TTL_MS) { return cachedSchema; } Stringschema executeDsctlCommand(workflowTool.getDsctlPath(), workflowTool, schema); if (schema ! null !schema.isBlank()) { cachedSchema schema; schemaCacheTime now; } return schema; }}/**从 dsctl schema JSON 中格式化命令列表param schemaJson dsctl schema 的 JSON 输出return 格式化后的命令列表*/private String formatCommandsFromSchema(String schemaJson) {if (schemaJson null || schemaJson.isBlank()) {return getDefaultCommands();}try { JSONObjectschema JSON.parseObject(schemaJson); JSONObjectdata schema.getJSONObject(data); if (data null) { return getDefaultCommands(); } JSONArraycommands data.getJSONArray(commands); if (commands null) { return getDefaultCommands(); } StringBuildersbnewStringBuilder(); for (inti0; i commands.size(); i) { JSONObjectgroup commands.getJSONObject(i); if (!group.equals(group.getString(kind))) { continue; } StringgroupName group.getString(name); StringgroupSummary group.getString(summary); sb.append(\n## ).append(groupName).append(\n); sb.append(# ).append(groupSummary).append(\n); JSONArraycmdList group.getJSONArray(commands); if (cmdList ! null) { for (intj0; j cmdList.size(); j) { JSONObjectcmd cmdList.getJSONObject(j); StringcmdName cmd.getString(name); Stringsummary cmd.getString(summary); // Build command syntax StringBuildersyntaxnewStringBuilder(); syntax.append(dsctl ).append(groupName).append( ).append(cmdName); // Add required arguments JSONArrayargs cmd.getJSONArray(arguments); if (args ! null) { for (intk0; k args.size(); k) { JSONObjectarg args.getJSONObject(k); if (arg.getBooleanValue(required)) { syntax.append( ).append(arg.getString(name)).append(); } } } sb.append(- ).append(syntax).append(: ).append(summary).append(\n); // Add discovery command if available if (args ! null) { for (intk0; k args.size(); k) { JSONObjectarg args.getJSONObject(k); Stringdiscovery arg.getString(discovery_command); if (discovery ! null !discovery.isEmpty()) { sb.append( Discovery: ).append(discovery).append(\n); break; } } } } } } return sb.toString(); } catch (Exception e) { log.error(Failed to parse dsctl schema: {}, e.getMessage(), e); return getDefaultCommands(); }}/**返回默认的命令列表当 schema 解析失败时使用return 默认命令列表*/private String getDefaultCommands() {return“”dsctl project list: List projectsdsctl project get : Get one projectdsctl project create --name : Create a projectdsctl project update : Update a projectdsctl project delete : Delete a projectdsctl workflow list: List workflowsdsctl workflow get : Get one workflowdsctl workflow run : Run a workflowdsctl workflow delete : Delete a workflowdsctl workflow-instance list: List instancesdsctl workflow-instance get : Get one instancedsctl workflow-instance watch : Watch instancedsctl workflow-instance stop : Stop instancedsctl workflow-instance digest : Get instance summarydsctl task list --workflow : List tasksdsctl task get : Get one taskdsctl task-instance list --workflow-instance : List task instancesdsctl task-instance log : Get task logdsctl cluster list: List clustersdsctl datasource list: List datasourcesdsctl environment list: List environmentsdsctl user list: List usersdsctl tenant list: List tenantsdsctl schedule list: List schedulesdsctl doctor: Run diagnostics“”;}}关键设计说明ChatAppManager.register() 在构造函数中注册 prompt 模板使其在 Agent 管理页面中可配置、可绑定不同 LLMAiServices.create() 自动在 prompt 中追加 JSON schema 约束将 LLM 输出反序列化为 DsctlCommand 对象accept() 通过检查 Agent 是否配置了 DSCTL 类型工具来决定是否激活实现按 Agent 粒度控制文件 4WorkflowExecutor.java新建路径chat/server/src/main/java/com/tencent/supersonic/chat/server/executor/WorkflowExecutor.javapackage com.tencent.supersonic.chat.server.executor;import com.tencent.supersonic.chat.api.pojo.response.QueryResult;import com.tencent.supersonic.chat.server.parser.WorkflowParser;import com.tencent.supersonic.chat.server.pojo.ExecuteContext;import com.tencent.supersonic.headless.api.pojo.response.QueryState;import java.io.BufferedReader;import java.io.InputStreamReader;import java.nio.charset.StandardCharsets;import java.util.Map;import java.util.concurrent.TimeUnit;import lombok.extern.slf4j.Slf4j;Slf4jpublicclassWorkflowExecutorimplementsChatQueryExecutor {Overridepublicbooleanaccept(ExecuteContext executeContext) {return WorkflowParser.QUERY_MODE.equals(executeContext.getParseInfo().getQueryMode());}Overridepublic QueryResult execute(ExecuteContext executeContext) {MapString, Object props executeContext.getParseInfo().getProperties();StringdsctlPath (String) props.getOrDefault(“workflow_ctl_path”, “dsctl”);StringsubCmd (String) props.get(“workflow_ctl_cmd”);Stringthought (String) props.getOrDefault(“workflow_ctl_thought”, “”);QueryResultresultnewQueryResult(); result.setQueryId(executeContext.getRequest().getQueryId()); result.setQueryMode(WorkflowParser.QUERY_MODE); result.setQueryMode(PLAIN_TEXT); if (subCmd null || subCmd.isBlank()) { result.setQueryState(QueryState.INVALID); result.setErrorMsg(无法识别 dsctl 命令); return result; } try { Stringoutput runDsctl(dsctlPath, subCmd, props); result.setQueryState(QueryState.SUCCESS); // textResult 会显示在对话界面中 result.setTextResult(buildTextResult(thought, subCmd, output)); } catch (Exception e) { log.error(WorkflowExecutor failed, cmd[dsctl {}], subCmd, e); result.setQueryState(QueryState.SEARCH_EXCEPTION); result.setErrorMsg(dsctl 执行失败: e.getMessage()); } return result;}/** 执行 dsctl 子命令返回标准输出内容。 超时时间 60 秒watch 类命令可能需要更长可按需调整。 */private String runDsctl(String dsctlPath, String subCmd, MapString, Object props)throws Exception {StringfullCmd dsctlPath subCmd;log.info(“DsctlExecutor executing: {}”, fullCmd);ProcessBuilderpbnewProcessBuilder(fullCmd.split(\\s)); pb.redirectErrorStream(true); // 显式注入环境变量到子进程 MapString, String env pb.environment(); StringdsApiUrl (String) props.get(ds_api_url); StringdsToken (String) props.get(ds_token); if (dsApiUrl ! null !dsApiUrl.isBlank()) { env.put(DS_API_URL, dsApiUrl); } if (dsToken ! null !dsToken.isBlank()) { env.put(DS_API_TOKEN, dsToken); } Processprocess pb.start(); booleanfinished process.waitFor(60, TimeUnit.SECONDS); if (!finished) { process.destroyForcibly(); thrownewRuntimeException(dsctl 命令超时60s: fullCmd); } try (BufferedReaderreader newBufferedReader( newInputStreamReader(process.getInputStream(), StandardCharsets.UTF_8))) { StringBuildersbnewStringBuilder(); String line; while ((line reader.readLine()) ! null) { sb.append(line).append(\n); } return sb.toString().trim(); }}private String buildTextResult(String thought, String subCmd, String output) {StringBuildersbnewStringBuilder();if (thought ! null !thought.isBlank()) {sb.append(““).append(thought).append(”\n\n”);}sb.append(“\n$ dsctl ).append(subCmd).append(\n); sb.append(output).append(\n”);return sb.toString();}}关键设计说明ProcessBuilder.environment() 显式注入 DS_API_URL / DS_API_TOKEN解决 JVM 子进程不继承 shell 环境变量的问题redirectErrorStream(true) 将 stderr 合并到 stdout方便统一处理错误输出buildTextResult() 将输出格式化为 Markdown供前端渲染文件 5spring.factories修改路径launchers/standalone/src/main/resources/META-INF/spring.factories在现有内容基础上追加DsctlParser 放在 PlainTextParser 前DsctlExecutor 放在 SqlExecutor 前com.tencent.supersonic.chat.server.parser.ChatQueryParsercom.tencent.supersonic.chat.server.parser.NL2PluginParser,com.tencent.supersonic.chat.server.parser.NL2SQLParser,com.tencent.supersonic.chat.server.parser.WorkflowParser,com.tencent.supersonic.chat.server.parser.PlainTextParsercom.tencent.supersonic.chat.server.executor.ChatQueryExecutorcom.tencent.supersonic.chat.server.executor.PluginExecutor,com.tencent.supersonic.chat.server.executor.WorkflowExecutor,com.tencent.supersonic.chat.server.executor.SqlExecutor,com.tencent.supersonic.chat.server.executor.PlainTextExecutor​注意​WorkflowExecutor 必须排在 SqlExecutor 前面因为 SqlExecutor.accept() 永远返回 true会拦截所有请求。前端实现文件 6type.ts修改路径webapp/packages/supersonic-fe/src/pages/Agent/type.tsexport enumAgentToolTypeEnum {NL2SQL_RULE ‘NL2SQL_RULE’,NL2SQL_LLM ‘NL2SQL_LLM’,PLUGIN ‘PLUGIN’,DATASET ‘DATASET’,WORK_FLOW_CTL ‘WORK_FLOW_CTL’, // 新增}exporttypeAgentToolType {id?: string;type: AgentToolTypeEnum;name: string;queryModes?: QueryModeEnum[];plugins?: number[];metricOptions?: MetricOptionType[];exampleQuestions?: string[];modelIds?: number[];// 新增 WORK_FLOW_CTL相关字段dsctlPath?: string;dsApiUrl?: string;dsToken?: string;defaultProject?: string;};文件 7ToolModal.tsx修改路径webapp/packages/supersonic-fe/src/pages/Agent/ToolModal.tsx在现有 PLUGIN 的 FormItem 之后追加 WORK_FLOW_CTL 的表单字段{toolType AgentToolTypeEnum.WORK_FLOW_CTL (FormItemname“dsApiUrl”label“DS API 地址”rules{[{ required: true, message: ‘请输入 DolphinScheduler API 地址’ }]}Input.Password placeholder“DolphinScheduler API Token” allowClear //)}工具类型下拉框的选项由后端 getToolTypes() API 动态加载后端枚举添加 WORK_FLOW_CTL 后前端下拉框自动出现 “DolphinScheduler CLI”无需额外修改选项列表。文件 8index.tsx修改路径webapp/packages/chat-sdk/src/components/ChatItem/index.tsx在 updateData 函数的白名单中添加 WORKFLOW_CTL并兼容 code: 1 的响应格式const updateData (res: Result) {lettip: string ‘’;letdata: MsgDataType | undefined undefined;const { queryColumns, queryResults, queryState, queryMode, response, chatContext, errorMsg } res.data || {};setExecuteErrorMsg(errorMsg);if (res.code 400 || res.code 401 || res.code 412) {tip res.msg;} elseif (res.code ! 200 res.code ! 1) { // 兼容 code:1tip SEARCH_EXCEPTION_TIP;} elseif (queryState ! ‘SUCCESS’) {tip response typeof response ‘string’ ? response : SEARCH_EXCEPTION_TIP;} elseif ((queryColumns queryColumns.length 0 queryResults) ||queryMode ‘WEB_PAGE’ ||queryMode ‘WEB_SERVICE’ ||queryMode ‘PLAIN_TEXT’ ||queryMode ‘WORKFLOW_CTL’ // 新增) {data res.data;tip ‘’;}// … 其余逻辑不变文件 9ExecuteItem.tsx修改路径webapp/packages/chat-sdk/src/components/ChatItem/ExecuteItem.tsx​修改 1​标题前缀第 50 行附近const titlePrefix queryMode ‘PLAIN_TEXT’ || queryMode ‘WEB_SERVICE’ || queryMode ‘WORKFLOW_CTL’? ‘问答’: ‘数据’;​修改 2​渲染逻辑第 163 行附近添加 WORKFLOW_CTL 分支使用 ReactMarkdown 渲染 Markdown 格式的命令输出{renderCustomExecuteNode executeItemNode ? (executeItemNode) : data?.queryMode ‘PLAIN_TEXT’ || data?.queryMode ‘WEB_SERVICE’ ? (data?.textResult) : data?.queryMode ‘WORKFLOW_CTL’ ? ({data?.textResult ?? ‘’}) : data?.queryMode ‘WEB_PAGE’ ? () : ()}ReactMarkdown 已在该文件中 import无需额外引入。使用配置前置条件​安装 dsctl​pip install -e .​配置环境变量​export DS_API_URLhttp://your-dolphinscheduler-host:12345/dolphinschedulerexport DS_API_TOKENyour-api-token​验证安装​dsctl doctorAgent 配置步骤进入 Agent 管理页面新建或编辑 Agent在工具Tab 中点击新增工具工具类型选择工作流命令行填写配置项字段 说明 示例DS API 地址 DolphinScheduler API 地址 http://ds-host:12345/dolphinschedulerDS Token API 认证 Token your_token_heredsctl 路径 dsctl 可执行文件路径 /usr/local/bin/dsctl默认项目 默认使用的项目名 etl-prod在对话聊天中进行自然语言对话支持的自然语言指令用户输入示例 对应 dsctl 命令“健康检查” dsctl doctor“列出所有项目” dsctl project list“切换到 etl-prod 项目” dsctl use project etl-prod“列出所有工作流” dsctl workflow list“运行 daily-etl 工作流” dsctl workflow run daily-etl“监控工作流实例 123” dsctl workflow-instance watch 123“查看工作流实例 123 的摘要” dsctl workflow-instance digest 123“列出工作流实例 123 的任务” dsctl task-instance list --workflow-instance 123“查看任务 456 的原始日志” dsctl task-instance log 456 --raw扩展要点总结扩展点 接口 作用意图识别 ChatQueryParser 将自然语言映射为 dsctl 命令命令执行 ChatQueryExecutor 调用 dsctl 进程返回结果工具类型 AgentToolType 在 Agent 管理页面中配置连接参数工具配置 AgentTool子类 携带 dsctl 路径、API 地址等配置LLM 调用 AiServicesChatAppManager 结构化提取命令prompt 可在 UI 中修改