feat(workflow): natural-language → workflow draft generator + agent tool

mateclaw-server/src/main/java/vip/mate/workflow/api/WorkflowController.java

@@ -41,6 +41,13 @@ public class WorkflowController {
     private final WorkflowRunPauseMapper pauseMapper;
     private final WorkflowCompiler compiler;
     private final WorkflowAclPort aclPort;
+    /** Optional — only present when the LLM module is wired (production).
+     *  Tests that don't boot the chat-model factory get a null and the
+     *  /draft/generate endpoint returns 503 instead of crashing. */
+    @org.springframework.beans.factory.annotation.Autowired(required = false)
+    private vip.mate.workflow.draftgen.WorkflowDraftGenerator draftGenerator;
+    @org.springframework.beans.factory.annotation.Autowired(required = false)
+    private vip.mate.workflow.draftgen.WorkflowDraftTemplateLibrary draftTemplates;
 
     @Operation(summary = "List workflows in the workspace")
     @GetMapping
@@ -229,6 +236,35 @@ public class WorkflowController {
         return R.ok(new RunDetail(run, steps, activePause));
     }
 
+    @Operation(summary = "Generate a workflow draft from a natural-language description.")
+    @PostMapping("/draft/generate")
+    public ResponseEntity<?> generateDraft(@RequestBody DraftGenerateRequest body,
+                                           @RequestHeader("X-Workspace-Id") long workspaceId) {
+        if (draftGenerator == null) {
+            return ResponseEntity.status(HttpStatus.SERVICE_UNAVAILABLE)
+                    .body(R.fail("workflow draft generator is not configured on this deployment"));
+        }
+        if (body == null || body.description() == null || body.description().isBlank()) {
+            return ResponseEntity.badRequest().body(R.fail("description is required"));
+        }
+        try {
+            return ResponseEntity.ok(R.ok(draftGenerator.generate(body.description(), workspaceId)));
+        } catch (IllegalArgumentException e) {
+            return ResponseEntity.badRequest().body(R.fail(e.getMessage()));
+        } catch (IllegalStateException e) {
+            return ResponseEntity.status(HttpStatus.BAD_GATEWAY).body(R.fail(e.getMessage()));
+        }
+    }
+
+    @Operation(summary = "List the canonical workflow templates the generator can apply directly.")
+    @GetMapping("/draft/templates")
+    public R<List<vip.mate.workflow.draftgen.WorkflowDraftTemplate>> listDraftTemplates() {
+        if (draftTemplates == null) return R.ok(List.of());
+        return R.ok(draftTemplates.all());
+    }
+
+    public record DraftGenerateRequest(String description) {}
+
     /** Narrow patch shape for {@link #update}; keeps the metadata path
      *  from accepting fields that would clobber the draft. */
     public record WorkflowMetadataRequest(String name, String description, Boolean enabled) {}

mateclaw-server/src/main/java/vip/mate/workflow/draftgen/GeneratedWorkflowDraft.java

@@ -0,0 +1,38 @@
+package vip.mate.workflow.draftgen;
+
+import java.util.List;
+import java.util.Map;
+
+/**
+ * Result of a natural-language → workflow draft generation. Crosses the
+ * REST boundary as JSON; the controller returns this verbatim.
+ *
+ * <p>{@code draftJson} is the {@code {"steps":[...]}} shape the
+ * runtime expects — same string the UI's JSON tab edits, same one
+ * {@link vip.mate.workflow.compiler.WorkflowCompiler} consumes. The
+ * generator pre-runs the compiler against it and reports compile
+ * failures via {@code compileErrors} without auto-publishing — v0
+ * always lets the operator review before pushing the row to a
+ * revision.
+ *
+ * <p>{@code triggerDrafts} is a list of suggested triggers the user
+ * can choose to create alongside the workflow; they're NOT created
+ * automatically and arrive with {@code enabled=false} per the
+ * generator system prompt's contract.
+ *
+ * <p>{@code warnings} / {@code missingFields} surface anywhere the
+ * model had to hedge — unfilled {@code TODO_*} placeholders, ambiguous
+ * approval policy, missing channel target. The UI displays these
+ * inline so the operator can finish the draft.
+ */
+public record GeneratedWorkflowDraft(
+        String name,
+        String description,
+        String draftJson,
+        List<Map<String, Object>> triggerDrafts,
+        List<String> warnings,
+        List<String> missingFields,
+        Double confidence,
+        boolean compileOk,
+        List<vip.mate.workflow.compiler.CompileError> compileErrors
+) {}

mateclaw-server/src/main/java/vip/mate/workflow/draftgen/WorkflowAuthoringTool.java

@@ -0,0 +1,112 @@
+package vip.mate.workflow.draftgen;
+
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.ai.chat.model.ToolContext;
+import org.springframework.ai.tool.annotation.Tool;
+import org.springframework.ai.tool.annotation.ToolParam;
+import org.springframework.lang.Nullable;
+import org.springframework.stereotype.Component;
+import vip.mate.agent.context.ChatOrigin;
+import vip.mate.workflow.model.WorkflowEntity;
+import vip.mate.workflow.service.WorkflowService;
+
+/**
+ * Agent-callable workflow drafting tool.
+ *
+ * <p>Lets a user say in chat «帮我把每周一汇总销售这件事做成 workflow»
+ * and have the agent compose a draft + persist it as a fresh
+ * {@link WorkflowEntity} row + return a short natural-language summary
+ * the user can act on. The created workflow stays as a draft (no
+ * publish, no triggers wired) — same v0 safety contract as the
+ * controller endpoint.
+ *
+ * <p>Workspace is taken from {@link ChatOrigin} on the active
+ * {@link ToolContext}, so the tool can never write into a foreign
+ * workspace even if the agent prompt tried to forge one.
+ */
+@Slf4j
+@Component
+public class WorkflowAuthoringTool {
+
+    private final WorkflowDraftGenerator generator;
+    private final WorkflowService workflowService;
+
+    public WorkflowAuthoringTool(WorkflowDraftGenerator generator,
+                                 WorkflowService workflowService) {
+        this.generator = generator;
+        this.workflowService = workflowService;
+    }
+
+    @Tool(description = "把用户描述的业务流程转换成一个 MateClaw workflow 草稿并保存到当前 workspace。"
+            + "适用场景：用户说「把 X 这件事做成 workflow / 自动化 / 流程」、「每周一让 X 员工 ...」、"
+            + "「客户消息进来时让 X 应对」。工具会输出 workflowId + 简短摘要，前端会自动在 workflow 编辑器里打开。"
+            + "不会自动发布，不会自动启用 trigger — 用户需要在编辑器里 review 后再 publish。")
+    public String workflow_draft_generate(
+            @ToolParam(description = "用户对业务流程的自然语言描述，越具体越好；可以包含触发条件、参与员工、是否要审批、要发到哪个渠道。")
+            String description,
+            // ChatOrigin-scoped workspace lookup; never trust the LLM to pass workspaceId.
+            @Nullable ToolContext ctx) {
+
+        Long workspaceId = ctx == null ? null : ChatOrigin.from(ctx).workspaceId();
+        if (workspaceId == null || workspaceId <= 0) {
+            return "无法确定当前 workspace，工具放弃执行。请在 workspace 上下文里调用我。";
+        }
+
+        GeneratedWorkflowDraft draft;
+        try {
+            draft = generator.generate(description, workspaceId);
+        } catch (Exception e) {
+            log.warn("[workflow_draft_generate] generation failed for ws={}: {}",
+                    workspaceId, e.getMessage());
+            return "生成失败：" + e.getMessage();
+        }
+
+        // Persist as a draft. No publish, no triggers — that's a separate
+        // user action via the editor / approve flow. We name it from the
+        // generator output so the editor surfaces something useful in
+        // the list immediately.
+        WorkflowEntity wf = new WorkflowEntity();
+        wf.setName(draft.name());
+        wf.setDescription(draft.description());
+        wf.setEnabled(true);
+        wf.setWorkspaceId(workspaceId);
+        WorkflowEntity created;
+        try {
+            created = workflowService.create(wf);
+            workflowService.saveDraft(created.getId(), workspaceId, draft.draftJson(), null);
+        } catch (Exception e) {
+            log.warn("[workflow_draft_generate] persist failed: {}", e.getMessage());
+            return "草稿生成成功但保存失败：" + e.getMessage();
+        }
+
+        StringBuilder out = new StringBuilder();
+        out.append("已生成 workflow 草稿 ").append(draft.name())
+           .append("（id=").append(created.getId()).append("）。\n");
+        if (draft.compileOk()) {
+            out.append("✓ 编译预校验通过。\n");
+        } else {
+            out.append("⚠ 编译预校验未通过 (").append(draft.compileErrors().size()).append(" 处)，需在编辑器里修正。\n");
+        }
+        if (draft.missingFields() != null && !draft.missingFields().isEmpty()) {
+            out.append("缺失字段：");
+            for (int i = 0; i < draft.missingFields().size(); i++) {
+                if (i > 0) out.append("；");
+                out.append(draft.missingFields().get(i));
+            }
+            out.append("\n");
+        }
+        if (draft.warnings() != null && !draft.warnings().isEmpty()) {
+            out.append("警告：");
+            for (int i = 0; i < draft.warnings().size(); i++) {
+                if (i > 0) out.append("；");
+                out.append(draft.warnings().get(i));
+            }
+            out.append("\n");
+        }
+        if (draft.triggerDrafts() != null && !draft.triggerDrafts().isEmpty()) {
+            out.append("建议触发器：").append(draft.triggerDrafts().size()).append(" 个 (默认未启用，需在编辑器里确认后创建)。\n");
+        }
+        out.append("请到 workflow 编辑器查看并继续完善。");
+        return out.toString();
+    }
+}

mateclaw-server/src/main/java/vip/mate/workflow/draftgen/WorkflowDraftGenerator.java

@@ -0,0 +1,355 @@
+package vip.mate.workflow.draftgen;
+
+import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
+import com.fasterxml.jackson.core.type.TypeReference;
+import com.fasterxml.jackson.databind.JsonNode;
+import com.fasterxml.jackson.databind.ObjectMapper;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.ai.chat.client.ChatClient;
+import org.springframework.ai.chat.model.ChatModel;
+import org.springframework.retry.support.RetryTemplate;
+import org.springframework.stereotype.Service;
+import vip.mate.agent.model.AgentEntity;
+import vip.mate.agent.repository.AgentMapper;
+import vip.mate.channel.model.ChannelEntity;
+import vip.mate.channel.repository.ChannelMapper;
+import vip.mate.llm.chatmodel.ProviderChatModelFactory;
+import vip.mate.llm.model.ModelConfigEntity;
+import vip.mate.llm.service.ModelConfigService;
+import vip.mate.workflow.compiler.PublishContext;
+import vip.mate.workflow.compiler.WorkflowAclPort;
+import vip.mate.workflow.compiler.WorkflowCompiler;
+
+import java.util.ArrayList;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+
+/**
+ * Natural-language → workflow draft generator.
+ *
+ * <p>Composes a system prompt + workspace-scoped context (available
+ * digital employees + channels) + the user description, dispatches to
+ * the workspace's default chat model, parses the JSON response, and
+ * runs {@link WorkflowCompiler} against it without persisting. The
+ * compile pass is "preview-only" — auto-publish is explicitly
+ * forbidden in the system prompt and we don't insert any rows here.
+ *
+ * <p>The generator is also the shared core called by the
+ * {@code workflow_draft_generate} agent tool, so a chat user can ask
+ * an agent "把每周一汇总销售这件事做成 workflow" and the agent gets back
+ * the same draft shape.
+ *
+ * <p>Failures are surfaced rather than swallowed: if the model returns
+ * non-JSON or the JSON doesn't carry a {@code steps} array, the
+ * generator throws so the controller / tool returns a clear error
+ * instead of a silently-broken draft.
+ */
+@Slf4j
+@Service
+public class WorkflowDraftGenerator {
+
+    /** System prompt — the contract the LLM must honor. Embedded as a
+     *  text block so the file is the canonical version (no resource
+     *  loading, no separate prompt-management infra in v0). */
+    static final String SYSTEM_PROMPT = """
+            你是 MateClaw 的工作流草稿生成器。你的任务是把用户用自然语言描述的业务流程，转换成 MateClaw RFC-29 v0 workflow JSON 草稿。
+
+            你只输出 JSON，不输出 Markdown，不输出解释，不输出代码块。
+
+            # 输出形态
+
+            必须输出一个 JSON object，结构如下：
+
+            {
+              "schemaVersion": "1.0",
+              "name": "...",
+              "description": "...",
+              "metadata": {
+                "generatedFrom": "natural_language",
+                "confidence": 0.0,
+                "warnings": [],
+                "missingFields": []
+              },
+              "triggerDrafts": [],
+              "steps": []
+            }
+
+            # v0 支持的 7 种 mode
+
+            sequential — 一个员工执行；必须 agentId/agentName + promptTemplate。outputContentType 只能 text 或 json。
+            fan_out — 至少 2 个连续 fan_out，后接 collect；每个分支必须 agentId/agentName + promptTemplate。
+            collect — 不带 agentId、agentName、promptTemplate；只能跟在 fan_out group 后。
+            conditional — mode.expression 必填（Pebble 子集，如 {{ outputs.x.approved == true }}）；agentId/agentName + promptTemplate 必填。
+            await_approval — approvalKind + approverChannels[] + approvalMessage 必填；可选 timeoutSecs；不要 agentId / agentName / promptTemplate。
+            dispatch_channel — channels[] + targets{} + content 必填；不要 agentId / agentName / promptTemplate。
+            write_memory — employeeId + file + mergeStrategy(append/prepend/replace_section/upsert_kv/overwrite) + content 必填；不要 agentId / agentName / promptTemplate。
+
+            # 不支持
+
+            不要生成 loop / invoke_skill / subflow。不要生成 agent_lifecycle / content_match 触发器。
+            遇到循环、重复直到成功、调用技能、复杂嵌套，用最接近的线性步骤，并在 metadata.warnings 写明需人工确认。
+
+            # 触发器（triggerDrafts）
+
+            只允许 patternType: cron / channel_message / workflow_completion / webhook。
+            triggerDrafts 默认 enabled=false，绝不自动启用。
+
+            # 命名
+
+            workflow.name 与 step.name 用英文 kebab-case (collect-sales-data / ask-finance-approval)。description 用用户母语。
+
+            # 占位字段
+
+            找不到匹配的真实 ID/渠道/员工时使用占位：
+            - agentName: "TODO_*_AGENT"
+            - employeeId: "TODO_EMPLOYEE_ID"
+            - channels[*]: "TODO_SELECT_CHANNEL"
+            - targets["TODO_SELECT_CHANNEL"]: "TODO_TARGET_ID"
+            - sourceWorkflowId: "TODO_WORKFLOW_ID"
+            每个 TODO 都要在 metadata.missingFields 中解释。
+            绝不能编造不存在的 agentId / channelType / 群 ID。
+
+            # 默认值
+
+            approvalKind: manager / finance / manual / legal / oncall 之一。
+            approverChannels: 默认 ["web"]，除非用户明确说企业 IM 渠道。
+            mergeStrategy: 默认 "append"。
+            schemaVersion: 始终 "1.0"。
+
+            # 质量
+
+            只使用 v0 字段；无注释；无 trailing comma；无 Markdown；不自动启用 trigger；不自动发布。
+            """;
+
+    private final ProviderChatModelFactory chatModelFactory;
+    private final ModelConfigService modelConfigService;
+    private final RetryTemplate retryTemplate;
+    private final AgentMapper agentMapper;
+    private final ChannelMapper channelMapper;
+    private final ObjectMapper objectMapper;
+    private final WorkflowCompiler compiler;
+    private final WorkflowAclPort aclPort;
+    private final WorkflowDraftTemplateLibrary templateLibrary;
+
+    public WorkflowDraftGenerator(ProviderChatModelFactory chatModelFactory,
+                                  ModelConfigService modelConfigService,
+                                  RetryTemplate retryTemplate,
+                                  AgentMapper agentMapper,
+                                  ChannelMapper channelMapper,
+                                  ObjectMapper objectMapper,
+                                  WorkflowCompiler compiler,
+                                  WorkflowAclPort aclPort,
+                                  WorkflowDraftTemplateLibrary templateLibrary) {
+        this.chatModelFactory = chatModelFactory;
+        this.modelConfigService = modelConfigService;
+        this.retryTemplate = retryTemplate;
+        this.agentMapper = agentMapper;
+        this.channelMapper = channelMapper;
+        this.objectMapper = objectMapper;
+        this.compiler = compiler;
+        this.aclPort = aclPort;
+        this.templateLibrary = templateLibrary;
+    }
+
+    public GeneratedWorkflowDraft generate(String description, long workspaceId) {
+        if (description == null || description.isBlank()) {
+            throw new IllegalArgumentException("description must not be empty");
+        }
+
+        // --- 1. workspace context ---------------------------------------
+        String contextPrompt = buildContextPrompt(workspaceId);
+
+        // --- 2. resolve runtime model ----------------------------------
+        ModelConfigEntity model = modelConfigService.getDefaultModel();
+        if (model == null) {
+            throw new IllegalStateException(
+                    "No default chat model configured; cannot generate workflow draft");
+        }
+        ChatModel chatModel = chatModelFactory.buildFor(model, retryTemplate);
+        ChatClient client = ChatClient.create(chatModel);
+
+        // --- 3. call the model -----------------------------------------
+        String raw;
+        try {
+            raw = client.prompt()
+                    .system(SYSTEM_PROMPT + "\n\n" + contextPrompt)
+                    .user(description)
+                    .call()
+                    .content();
+        } catch (Exception e) {
+            throw new IllegalStateException(
+                    "Workflow draft generator chat call failed: " + e.getMessage(), e);
+        }
+        if (raw == null || raw.isBlank()) {
+            throw new IllegalStateException("Workflow draft generator returned empty content");
+        }
+
+        // --- 4. parse + validate shape ---------------------------------
+        JsonNode root = parseStrict(raw);
+        if (!root.has("steps") || !root.get("steps").isArray()) {
+            throw new IllegalStateException(
+                    "Generated draft has no steps[] array; raw output: " + truncate(raw));
+        }
+
+        // --- 5. extract fields -----------------------------------------
+        String name = root.path("name").asText("");
+        String userDescription = root.path("description").asText("");
+        Double confidence = root.path("metadata").path("confidence").isNumber()
+                ? root.path("metadata").path("confidence").asDouble() : null;
+
+        List<String> warnings = readStringArray(root, "metadata", "warnings");
+        List<String> missingFields = readStringArray(root, "metadata", "missingFields");
+
+        // The runtime only consumes the steps part of the draft — strip
+        // everything else into a clean {steps:[...]} shape.
+        Map<String, Object> draftRoot = new LinkedHashMap<>();
+        draftRoot.put("steps", objectMapper.convertValue(root.get("steps"),
+                new TypeReference<List<Map<String, Object>>>() {}));
+        String draftJson;
+        try {
+            draftJson = objectMapper.writerWithDefaultPrettyPrinter().writeValueAsString(draftRoot);
+        } catch (Exception e) {
+            throw new IllegalStateException("Failed to re-serialize generated steps: " + e.getMessage(), e);
+        }
+
+        // --- 6. trigger drafts -----------------------------------------
+        List<Map<String, Object>> triggerDrafts = new ArrayList<>();
+        if (root.has("triggerDrafts") && root.get("triggerDrafts").isArray()) {
+            triggerDrafts = objectMapper.convertValue(root.get("triggerDrafts"),
+                    new TypeReference<List<Map<String, Object>>>() {});
+            for (Map<String, Object> td : triggerDrafts) {
+                // Belt-and-suspenders: never trust the LLM to honor enabled=false.
+                td.put("enabled", false);
+            }
+        }
+
+        // --- 7. compile preview ---------------------------------------
+        boolean compileOk;
+        List<vip.mate.workflow.compiler.CompileError> compileErrors;
+        try {
+            // PublishContext is (workspaceId, publisherId).
+            WorkflowCompiler.Result result = compiler.compile(draftJson,
+                    new PublishContext(workspaceId, 0L), aclPort);
+            compileOk = result.ok();
+            compileErrors = compileOk ? List.of() : result.errors();
+        } catch (Exception e) {
+            // Compile preview failures are not fatal — the operator can
+            // still edit the draft. We surface them as warnings.
+            log.warn("[WorkflowDraftGenerator] preview compile failed: {}", e.getMessage());
+            compileOk = false;
+            compileErrors = List.of();
+            warnings = appendWarning(warnings, "preview compile threw: " + e.getMessage());
+        }
+
+        return new GeneratedWorkflowDraft(
+                name == null || name.isBlank() ? "untitled-workflow" : name,
+                userDescription,
+                draftJson,
+                triggerDrafts,
+                warnings,
+                missingFields,
+                confidence,
+                compileOk,
+                compileErrors);
+    }
+
+    /** Compose the workspace-scoped context prompt: agent + channel
+     *  inventory the model can pick from. Agents are filtered to enabled
+     *  rows; channels likewise. The model is told to prefer real ids
+     *  over TODOs but never to fabricate. */
+    private String buildContextPrompt(long workspaceId) {
+        List<AgentEntity> agents = agentMapper.selectList(new LambdaQueryWrapper<AgentEntity>()
+                .eq(AgentEntity::getWorkspaceId, workspaceId)
+                .eq(AgentEntity::getEnabled, true));
+        List<ChannelEntity> channels = channelMapper.selectList(new LambdaQueryWrapper<ChannelEntity>()
+                .eq(ChannelEntity::getWorkspaceId, workspaceId)
+                .eq(ChannelEntity::getEnabled, true));
+
+        StringBuilder sb = new StringBuilder();
+        sb.append("# 当前 workspace 可用数字员工\n[");
+        boolean first = true;
+        for (AgentEntity a : agents) {
+            if (!first) sb.append(",");
+            first = false;
+            sb.append("{\"agentId\":").append(a.getId())
+              .append(",\"name\":\"").append(escape(a.getName()))
+              .append("\",\"description\":\"")
+              .append(escape(a.getDescription() == null ? "" : a.getDescription()))
+              .append("\"}");
+        }
+        sb.append("]\n\n# 当前 workspace 可用渠道\n[");
+        first = true;
+        for (ChannelEntity c : channels) {
+            if (!first) sb.append(",");
+            first = false;
+            sb.append("{\"channelType\":\"").append(escape(c.getChannelType()))
+              .append("\",\"name\":\"").append(escape(c.getName()))
+              .append("\"}");
+        }
+        sb.append("]\n\n优先使用这些真实 agentId 和 channelType。不存在的 ID 必须用 TODO_* 占位，不要编造。\n");
+
+        // Few-shot exemplars from the template library — the LLM stays
+        // closer to canonical shapes when it has 2-3 concrete examples
+        // in the system prompt.
+        sb.append("\n# 模板示例（参考，不必照抄）\n");
+        for (WorkflowDraftTemplate t : templateLibrary.all()) {
+            sb.append("## ").append(t.id()).append(" — ").append(t.label()).append("\n");
+            sb.append(t.description()).append("\n");
+            sb.append("draft: ").append(t.draftJson()).append("\n");
+            if (t.triggerDraftsJson() != null && !"[]".equals(t.triggerDraftsJson())) {
+                sb.append("triggerDrafts: ").append(t.triggerDraftsJson()).append("\n");
+            }
+        }
+        return sb.toString();
+    }
+
+    private JsonNode parseStrict(String raw) {
+        // Some models still wrap the JSON in a ```json fence even when
+        // the prompt says "no Markdown". Strip the fences before parsing
+        // so we don't reject otherwise-valid output.
+        String cleaned = raw.trim();
+        if (cleaned.startsWith("```")) {
+            int firstNl = cleaned.indexOf('\n');
+            if (firstNl > 0) cleaned = cleaned.substring(firstNl + 1);
+            int closeFence = cleaned.lastIndexOf("```");
+            if (closeFence > 0) cleaned = cleaned.substring(0, closeFence);
+            cleaned = cleaned.trim();
+        }
+        try {
+            return objectMapper.readTree(cleaned);
+        } catch (Exception e) {
+            throw new IllegalStateException(
+                    "Workflow draft generator returned non-JSON: " + e.getMessage()
+                            + " — raw: " + truncate(raw), e);
+        }
+    }
+
+    private List<String> readStringArray(JsonNode root, String... path) {
+        JsonNode node = root;
+        for (String p : path) node = node.path(p);
+        if (!node.isArray()) return List.of();
+        List<String> out = new ArrayList<>(node.size());
+        for (JsonNode item : node) {
+            if (item.isTextual()) out.add(item.asText());
+        }
+        return out;
+    }
+
+    private static List<String> appendWarning(List<String> existing, String msg) {
+        List<String> next = new ArrayList<>(existing == null ? List.of() : existing);
+        next.add(msg);
+        return next;
+    }
+
+    private static String escape(String s) {
+        if (s == null) return "";
+        return s.replace("\\", "\\\\").replace("\"", "\\\"")
+                .replace("\n", " ").replace("\r", " ");
+    }
+
+    private static String truncate(String s) {
+        if (s == null) return "";
+        return s.length() <= 400 ? s : s.substring(0, 400) + "…";
+    }
+}

mateclaw-server/src/main/java/vip/mate/workflow/draftgen/WorkflowDraftTemplate.java

@@ -0,0 +1,43 @@
+package vip.mate.workflow.draftgen;
+
+import java.util.List;
+
+/**
+ * One named exemplar in the workflow template library.
+ *
+ * <p>Templates serve two purposes:
+ * <ol>
+ *   <li>As few-shot examples inside the system prompt — the LLM sees
+ *       "here are five canonical shapes; pick the closest and adapt the
+ *       fields" instead of inventing structure from scratch. RFC v0
+ *       authors should never see anything more exotic than these
+ *       shapes.</li>
+ *   <li>As "apply template" entries the UI or the
+ *       workflow_draft_generate tool can drop in directly when the
+ *       user's description matches a canonical pattern (saves a
+ *       generation roundtrip and stays cheaper / faster).</li>
+ * </ol>
+ *
+ * <p>{@code matchHints} is a small bag of natural-language phrases the
+ * tool can use to short-circuit to a template before calling the LLM —
+ * if the user says "周一汇总" or "weekly summary" we already know which
+ * shape they mean.
+ */
+public record WorkflowDraftTemplate(
+        /** Stable kebab-case id; surfaces in the API response. */
+        String id,
+        /** Short bilingual label; the UI's "apply template" picker shows this. */
+        String label,
+        /** One-sentence description in user-facing prose. */
+        String description,
+        /** Natural-language phrases that should bias toward this template. */
+        List<String> matchHints,
+        /** Workflow draft JSON; placeholders like TODO_AGENT_ID stay
+         *  in the body until the UI / tool fills them. */
+        String draftJson,
+        /** Trigger drafts attached to this template, if any. Stored as
+         *  serialised JSON arrays so the prompt doesn't have to know
+         *  about Java types. */
+        String triggerDraftsJson
+) {
+}

mateclaw-server/src/main/java/vip/mate/workflow/draftgen/WorkflowDraftTemplateLibrary.java

@@ -0,0 +1,210 @@
+package vip.mate.workflow.draftgen;
+
+import org.springframework.stereotype.Component;
+
+import java.util.List;
+
+/**
+ * Small in-process library of canonical workflow shapes. Used as
+ * few-shot exemplars in the system prompt AND as "apply template"
+ * entries operators / agents can drop in directly. Kept as code
+ * constants rather than a DB table so the templates version with the
+ * runtime that interprets them — a template that references modes the
+ * runtime doesn't support yet should never ship.
+ *
+ * <p>Templates are intentionally minimal: 5-7 shapes that cover what
+ * the v0 reviewer flagged as the actual customer use cases (weekly
+ * summary, approval-and-notify, customer-message routing, chained
+ * workflow, daily memory write). New shapes only get added when the
+ * customer evidence is in.
+ */
+@Component
+public class WorkflowDraftTemplateLibrary {
+
+    private final List<WorkflowDraftTemplate> templates = List.of(
+            weeklySummary(),
+            approvalAndNotify(),
+            customerMessageRouting(),
+            chainedWorkflow(),
+            dailyMemoryWrite(),
+            parallelAnalysis(),
+            channelAlertOnFailure()
+    );
+
+    public List<WorkflowDraftTemplate> all() {
+        return templates;
+    }
+
+    /** Look up a template by id; returns null when no match. */
+    public WorkflowDraftTemplate byId(String id) {
+        if (id == null) return null;
+        return templates.stream()
+                .filter(t -> id.equals(t.id()))
+                .findFirst().orElse(null);
+    }
+
+    // ===== template definitions =====
+
+    private static WorkflowDraftTemplate weeklySummary() {
+        return new WorkflowDraftTemplate(
+                "weekly-summary",
+                "周报汇总 / Weekly summary",
+                "每周固定时间让数字员工汇总数据，再发到群里。常见于销售周报、运营日报。",
+                List.of("每周", "周报", "周一", "weekly", "summary", "汇总"),
+                """
+                {"steps":[
+                  {"name":"collect-data","agentName":"TODO_DATA_AGENT","mode":{"type":"sequential"},
+                   "promptTemplate":"汇总本周的{{ inputs.topic }}并输出 JSON","outputVar":"summary","outputContentType":"json"},
+                  {"name":"notify-group",
+                   "mode":{"type":"dispatch_channel","channels":["TODO_SELECT_CHANNEL"],
+                           "targets":{"TODO_SELECT_CHANNEL":"TODO_TARGET_ID"},
+                           "content":"本周汇总：{{ outputs.summary }}"}}
+                ]}""",
+                """
+                [{"name":"weekly-summary-cron","patternType":"cron","enabled":false,
+                  "patternJson":{"cron":"0 0 9 ? * MON","timezone":"Asia/Shanghai"},
+                  "targetType":"workflow",
+                  "payloadTemplate":"{\\"topic\\":\\"销售\\"}"}]"""
+        );
+    }
+
+    private static WorkflowDraftTemplate approvalAndNotify() {
+        return new WorkflowDraftTemplate(
+                "approval-and-notify",
+                "审批后通知 / Approval then notify",
+                "数字员工出方案 → 老板审批 → 通过后发到群里。常见于费用申请、采购、合同。",
+                List.of("审批", "确认", "老板", "approval", "approve", "确认通过"),
+                """
+                {"steps":[
+                  {"name":"draft-proposal","agentName":"TODO_DRAFTER","mode":{"type":"sequential"},
+                   "promptTemplate":"为 {{ inputs.topic }} 起草一个方案","outputVar":"proposal","outputContentType":"text"},
+                  {"name":"manager-approve",
+                   "mode":{"type":"await_approval","approvalKind":"manager",
+                           "approverChannels":["web"],
+                           "approvalMessage":"请审批方案：{{ outputs.proposal }}",
+                           "timeoutSecs":86400}},
+                  {"name":"notify-group",
+                   "mode":{"type":"dispatch_channel","channels":["TODO_SELECT_CHANNEL"],
+                           "targets":{"TODO_SELECT_CHANNEL":"TODO_TARGET_ID"},
+                           "content":"方案已通过：{{ outputs.proposal }}"}}
+                ]}""",
+                "[]"
+        );
+    }
+
+    private static WorkflowDraftTemplate customerMessageRouting() {
+        return new WorkflowDraftTemplate(
+                "customer-message-routing",
+                "客户消息路由 / Customer message routing",
+                "渠道里出现关键词时，让客服员工应对，并把结果记到员工记忆。",
+                List.of("客户", "客服", "关键词", "customer", "support", "回复"),
+                """
+                {"steps":[
+                  {"name":"answer-customer","agentName":"TODO_SUPPORT_AGENT","mode":{"type":"sequential"},
+                   "promptTemplate":"客户说：{{ inputs.content }}。请用礼貌的语气回复。",
+                   "outputVar":"reply","outputContentType":"text"},
+                  {"name":"send-reply",
+                   "mode":{"type":"dispatch_channel","channels":["TODO_SELECT_CHANNEL"],
+                           "targets":{"TODO_SELECT_CHANNEL":"{{ inputs.sender }}"},
+                           "content":"{{ outputs.reply }}"}},
+                  {"name":"remember-issue",
+                   "mode":{"type":"write_memory","employeeId":"TODO_SUPPORT_AGENT",
+                           "file":"customer-issues.md","mergeStrategy":"append",
+                           "content":"### {{ inputs.sender }}\\n{{ inputs.content }}\\n回复：{{ outputs.reply }}\\n"}}
+                ]}""",
+                """
+                [{"name":"customer-keyword","patternType":"channel_message","enabled":false,
+                  "patternJson":{"channelType":"TODO_SELECT_CHANNEL","contentContains":"发票"},
+                  "targetType":"workflow",
+                  "payloadTemplate":"{\\"content\\":\\"{{ event.content }}\\",\\"sender\\":\\"{{ event.senderId }}\\"}"}]"""
+        );
+    }
+
+    private static WorkflowDraftTemplate chainedWorkflow() {
+        return new WorkflowDraftTemplate(
+                "chained-workflow",
+                "上游完成后接力 / Chained on upstream completion",
+                "上游 workflow 跑完后自动接一段处理：常见于 ETL 接出报表、运营接审计。",
+                List.of("接力", "上游", "完成后", "chained", "after"),
+                """
+                {"steps":[
+                  {"name":"post-process","agentName":"TODO_AGENT","mode":{"type":"sequential"},
+                   "promptTemplate":"上游 run {{ inputs.sourceWorkflowId }} 已完成（state={{ inputs.state }}），请处理后续。",
+                   "outputVar":"summary","outputContentType":"text"}
+                ]}""",
+                """
+                [{"name":"after-upstream","patternType":"workflow_completion","enabled":false,
+                  "patternJson":{"sourceWorkflowId":"TODO_WORKFLOW_ID","stateFilter":"succeeded"},
+                  "targetType":"workflow",
+                  "payloadTemplate":"{\\"sourceWorkflowId\\":\\"{{ event.sourceWorkflowId }}\\",\\"state\\":\\"{{ event.state }}\\"}"}]"""
+        );
+    }
+
+    private static WorkflowDraftTemplate dailyMemoryWrite() {
+        return new WorkflowDraftTemplate(
+                "daily-memory-write",
+                "每日记入员工记忆 / Daily memory append",
+                "每天定时让员工写一段记忆，作为后续对话的上下文。",
+                List.of("每天", "daily", "记忆", "写入", "memory"),
+                """
+                {"steps":[
+                  {"name":"summarize-day","agentName":"TODO_AGENT","mode":{"type":"sequential"},
+                   "promptTemplate":"用一段话总结今天的{{ inputs.topic }}。",
+                   "outputVar":"summary","outputContentType":"text"},
+                  {"name":"persist-memory",
+                   "mode":{"type":"write_memory","employeeId":"TODO_EMPLOYEE_ID",
+                           "file":"daily-log.md","mergeStrategy":"append",
+                           "content":"### {{ inputs.date }}\\n{{ outputs.summary }}\\n"}}
+                ]}""",
+                """
+                [{"name":"daily-memory-cron","patternType":"cron","enabled":false,
+                  "patternJson":{"cron":"0 0 22 * * ?","timezone":"Asia/Shanghai"},
+                  "targetType":"workflow",
+                  "payloadTemplate":"{\\"topic\\":\\"工作\\",\\"date\\":\\"{{ trigger.firedAt }}\\"}"}]"""
+        );
+    }
+
+    private static WorkflowDraftTemplate parallelAnalysis() {
+        return new WorkflowDraftTemplate(
+                "parallel-analysis",
+                "并行多角度分析 / Parallel multi-angle analysis",
+                "三个不同员工同时从不同角度分析同一份输入，最后由 collect 汇合。",
+                List.of("分别", "并行", "多角度", "parallel", "fan_out"),
+                """
+                {"steps":[
+                  {"name":"angle-finance","agentName":"TODO_FINANCE_AGENT","mode":{"type":"fan_out"},
+                   "promptTemplate":"从财务角度分析：{{ inputs.topic }}",
+                   "outputVar":"finance","outputContentType":"text"},
+                  {"name":"angle-operations","agentName":"TODO_OPS_AGENT","mode":{"type":"fan_out"},
+                   "promptTemplate":"从运营角度分析：{{ inputs.topic }}",
+                   "outputVar":"ops","outputContentType":"text"},
+                  {"name":"angle-customer","agentName":"TODO_CUSTOMER_AGENT","mode":{"type":"fan_out"},
+                   "promptTemplate":"从客户角度分析：{{ inputs.topic }}",
+                   "outputVar":"customer","outputContentType":"text"},
+                  {"name":"merge-views","mode":{"type":"collect"}}
+                ]}""",
+                "[]"
+        );
+    }
+
+    private static WorkflowDraftTemplate channelAlertOnFailure() {
+        return new WorkflowDraftTemplate(
+                "channel-alert-on-failure",
+                "上游失败时报警 / Alert on upstream failure",
+                "上游 workflow 跑失败时立即推送到值班渠道，常见于关键 ETL / 自动化作业的兜底。",
+                List.of("失败", "报警", "alert", "failure", "失败时"),
+                """
+                {"steps":[
+                  {"name":"alert-oncall",
+                   "mode":{"type":"dispatch_channel","channels":["TODO_SELECT_CHANNEL"],
+                           "targets":{"TODO_SELECT_CHANNEL":"TODO_ONCALL_TARGET"},
+                           "content":"⚠ 上游 workflow run {{ inputs.runId }} 失败：{{ inputs.errorMessage }}"}}
+                ]}""",
+                """
+                [{"name":"upstream-failure","patternType":"workflow_completion","enabled":false,
+                  "patternJson":{"sourceWorkflowId":"TODO_WORKFLOW_ID","stateFilter":"failed"},
+                  "targetType":"workflow",
+                  "payloadTemplate":"{\\"runId\\":\\"{{ event.runId }}\\",\\"errorMessage\\":\\"{{ event.errorMessage }}\\"}"}]"""
+        );
+    }
+}

mateclaw-ui/src/api/index.ts

@@ -891,6 +891,35 @@ export const workflowApi = {
       outcome,
       payload,
     }),
+  /** Generate a workflow draft from natural language. Returns the parsed
+   *  shape; the caller is responsible for creating a workflow row +
+   *  saving the draft if the user accepts it. */
+  generateDraft: (description: string) =>
+    http.post<GeneratedDraft>('/workflows/draft/generate', { description }),
+  /** Canonical workflow templates the generator can apply directly. */
+  listDraftTemplates: () =>
+    http.get<WorkflowDraftTemplate[]>('/workflows/draft/templates'),
+}
+
+export interface GeneratedDraft {
+  name: string
+  description: string
+  draftJson: string
+  triggerDrafts: Array<Record<string, unknown>>
+  warnings: string[]
+  missingFields: string[]
+  confidence?: number | null
+  compileOk: boolean
+  compileErrors: WorkflowCompileError[]
+}
+
+export interface WorkflowDraftTemplate {
+  id: string
+  label: string
+  description: string
+  matchHints: string[]
+  draftJson: string
+  triggerDraftsJson: string
 }
 
 // ==================== Trigger ====================

mateclaw-ui/src/components/workflow/GenerateWorkflowDialog.vue

@@ -0,0 +1,317 @@
+<template>
+  <Teleport to="body">
+    <div v-if="visible" class="modal-overlay" @click.self="close">
+      <div class="modal" role="dialog" aria-modal="true">
+        <div class="modal-header">
+          <h3>{{ t('workflows.generate.title') }}</h3>
+          <button class="modal-close" @click="close" aria-label="close">×</button>
+        </div>
+        <div class="modal-body">
+          <p class="hint">{{ t('workflows.generate.hint') }}</p>
+          <textarea
+            ref="descRef"
+            v-model="description"
+            class="form-input form-textarea"
+            rows="5"
+            spellcheck="false"
+            :placeholder="t('workflows.generate.placeholder')"
+          />
+
+          <div v-if="result" class="generate-result">
+            <div class="result-summary">
+              <strong>{{ result.name }}</strong>
+              <span class="confidence" v-if="result.confidence != null">
+                · confidence {{ Math.round((result.confidence ?? 0) * 100) }}%
+              </span>
+              <span v-if="result.compileOk" class="status-pill ok">
+                ✓ {{ t('workflows.generate.compileOk') }}
+              </span>
+              <span v-else class="status-pill err">
+                ⚠ {{ t('workflows.generate.compileFail', { count: result.compileErrors.length }) }}
+              </span>
+            </div>
+            <p v-if="result.description" class="result-desc">{{ result.description }}</p>
+            <ul v-if="result.missingFields.length" class="result-list">
+              <li v-for="(m, i) in result.missingFields" :key="`miss-${i}`">
+                <span class="result-tag">{{ t('workflows.generate.missing') }}</span> {{ m }}
+              </li>
+            </ul>
+            <ul v-if="result.warnings.length" class="result-list">
+              <li v-for="(w, i) in result.warnings" :key="`warn-${i}`">
+                <span class="result-tag warn">{{ t('workflows.generate.warning') }}</span> {{ w }}
+              </li>
+            </ul>
+            <details class="result-raw">
+              <summary>{{ t('workflows.generate.previewDraft') }}</summary>
+              <pre>{{ result.draftJson }}</pre>
+            </details>
+          </div>
+        </div>
+        <div class="modal-footer">
+          <button class="btn-secondary" @click="close">{{ t('common.cancel') }}</button>
+          <button
+            v-if="!result"
+            class="btn-primary"
+            :disabled="loading || !description.trim()"
+            @click="onGenerate"
+          >
+            {{ loading ? t('workflows.generate.running') : t('workflows.generate.submit') }}
+          </button>
+          <template v-else>
+            <button class="btn-secondary" @click="onRetry">
+              {{ t('workflows.generate.retry') }}
+            </button>
+            <button class="btn-primary" :disabled="loading" @click="onAccept">
+              {{ t('workflows.generate.accept') }}
+            </button>
+          </template>
+        </div>
+      </div>
+    </div>
+  </Teleport>
+</template>
+
+<script setup lang="ts">
+import { nextTick, ref, watch } from 'vue'
+import { useI18n } from 'vue-i18n'
+import { ElMessage } from 'element-plus'
+import { workflowApi, type GeneratedDraft } from '@/api'
+
+interface Props {
+  modelValue: boolean
+}
+const props = defineProps<Props>()
+const emit = defineEmits<{
+  (e: 'update:modelValue', v: boolean): void
+  /** User accepted the draft — parent should create the workflow + save the draft. */
+  (e: 'accept', payload: GeneratedDraft): void
+}>()
+
+const { t } = useI18n()
+const visible = ref(props.modelValue)
+const description = ref('')
+const loading = ref(false)
+const result = ref<GeneratedDraft | null>(null)
+const descRef = ref<HTMLTextAreaElement | null>(null)
+
+watch(
+  () => props.modelValue,
+  async (open) => {
+    visible.value = open
+    if (open) {
+      description.value = ''
+      result.value = null
+      loading.value = false
+      await nextTick()
+      descRef.value?.focus()
+      document.addEventListener('keydown', onKey)
+    } else {
+      document.removeEventListener('keydown', onKey)
+    }
+  }
+)
+watch(visible, (v) => emit('update:modelValue', v))
+
+function onKey(e: KeyboardEvent) {
+  if (e.key === 'Escape' && visible.value) close()
+}
+function close() {
+  visible.value = false
+}
+
+async function onGenerate() {
+  const desc = description.value.trim()
+  if (!desc) return
+  loading.value = true
+  try {
+    const res = await workflowApi.generateDraft(desc)
+    result.value = res.data as unknown as GeneratedDraft
+  } catch (e) {
+    ElMessage.error(t('workflows.generate.failed', { msg: (e as Error).message }))
+  } finally {
+    loading.value = false
+  }
+}
+
+function onRetry() {
+  result.value = null
+}
+
+function onAccept() {
+  if (!result.value) return
+  emit('accept', result.value)
+  close()
+}
+</script>
+
+<style scoped>
+.modal-overlay {
+  position: fixed;
+  inset: 0;
+  background: rgba(15, 10, 8, 0.45);
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  padding: 20px;
+  z-index: 2100;
+}
+.modal {
+  width: 640px;
+  max-width: 100%;
+  background: var(--mc-bg-elevated);
+  border: 1px solid var(--mc-border);
+  border-radius: 14px;
+  overflow: hidden;
+  max-height: 80vh;
+  display: flex;
+  flex-direction: column;
+}
+.modal-header {
+  display: flex;
+  justify-content: space-between;
+  align-items: center;
+  padding: 14px 18px;
+  border-bottom: 1px solid var(--mc-border-light);
+}
+.modal-header h3 {
+  margin: 0;
+  font-size: 16px;
+  font-weight: 600;
+}
+.modal-close {
+  width: 26px;
+  height: 26px;
+  border: none;
+  background: none;
+  color: var(--mc-text-tertiary);
+  font-size: 22px;
+  cursor: pointer;
+  border-radius: 6px;
+}
+.modal-close:hover {
+  background: var(--mc-bg-muted);
+  color: var(--mc-text-primary);
+}
+.modal-body {
+  padding: 18px;
+  overflow-y: auto;
+  display: flex;
+  flex-direction: column;
+  gap: 12px;
+}
+.hint {
+  margin: 0;
+  font-size: 12.5px;
+  color: var(--mc-text-secondary);
+  line-height: 1.5;
+}
+.form-input,
+.form-textarea {
+  width: 100%;
+  padding: 9px 12px;
+  border: 1px solid var(--mc-border);
+  border-radius: 8px;
+  background: var(--mc-bg-sunken);
+  color: var(--mc-text-primary);
+  font-size: 13.5px;
+  box-sizing: border-box;
+  font-family: inherit;
+}
+.form-textarea { resize: vertical; }
+.form-input:focus,
+.form-textarea:focus {
+  outline: none;
+  border-color: var(--mc-primary);
+}
+.generate-result {
+  margin-top: 6px;
+  padding: 12px;
+  border: 1px solid var(--mc-border-light);
+  border-radius: 8px;
+  background: var(--mc-bg-sunken);
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+.result-summary {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  flex-wrap: wrap;
+  font-size: 13px;
+}
+.confidence { font-size: 11.5px; opacity: 0.7; }
+.status-pill {
+  font-size: 11px;
+  padding: 2px 8px;
+  border-radius: 999px;
+  font-weight: 500;
+}
+.status-pill.ok { background: rgba(46, 204, 113, 0.18); color: #1e8449; }
+.status-pill.err { background: rgba(231, 76, 60, 0.14); color: var(--mc-danger); }
+.result-desc { margin: 0; font-size: 12.5px; color: var(--mc-text-secondary); }
+.result-list { margin: 4px 0 0; padding: 0 0 0 4px; list-style: none; font-size: 12px; }
+.result-list li { padding: 3px 0; line-height: 1.5; }
+.result-tag {
+  display: inline-block;
+  padding: 1px 6px;
+  margin-right: 5px;
+  border-radius: 3px;
+  background: var(--mc-bg-muted);
+  font-size: 10px;
+  color: var(--mc-text-tertiary);
+  text-transform: uppercase;
+  letter-spacing: 0.04em;
+}
+.result-tag.warn { background: rgba(245, 158, 11, 0.16); color: #b45309; }
+.result-raw summary {
+  font-size: 11px;
+  color: var(--mc-text-tertiary);
+  cursor: pointer;
+}
+.result-raw pre {
+  margin: 6px 0 0;
+  font-family: 'JetBrains Mono', Consolas, monospace;
+  font-size: 11px;
+  background: var(--mc-bg-elevated);
+  border: 1px solid var(--mc-border-light);
+  padding: 8px;
+  border-radius: 6px;
+  white-space: pre-wrap;
+  word-break: break-word;
+  max-height: 200px;
+  overflow: auto;
+}
+.modal-footer {
+  display: flex;
+  justify-content: flex-end;
+  gap: 8px;
+  padding: 12px 18px 16px;
+  border-top: 1px solid var(--mc-border-light);
+}
+.btn-primary,
+.btn-secondary {
+  padding: 8px 16px;
+  border-radius: 8px;
+  font-size: 13px;
+  font-weight: 500;
+  cursor: pointer;
+  border: 1px solid transparent;
+}
+.btn-primary {
+  background: var(--mc-primary);
+  color: var(--mc-text-inverse, #ffffff);
+  border-color: var(--mc-primary);
+}
+.btn-primary:disabled { opacity: 0.55; cursor: not-allowed; }
+.btn-primary:hover:not(:disabled) { background: var(--mc-primary-hover, var(--mc-primary)); }
+.btn-secondary {
+  background: transparent;
+  color: var(--mc-text-secondary);
+  border-color: var(--mc-border);
+}
+.btn-secondary:hover {
+  background: var(--mc-bg-muted);
+  color: var(--mc-text-primary);
+}
+</style>

mateclaw-ui/src/i18n/locales/en-US.ts

@@ -2021,6 +2021,22 @@ export default {
       deleteContent: 'Delete workflow "{name}"? This is reversible via the audit log.',
     },
     selectHint: 'Select a workflow on the left, or click "New Workflow" to start a fresh draft.',
+    generate: {
+      entryButton: 'Generate from description',
+      title: 'Generate workflow draft from natural language',
+      hint: 'Describe your business process in a sentence — the AI generates a workflow JSON draft. The draft is saved unpublished and no triggers are auto-enabled; you review and publish afterwards.',
+      placeholder: 'e.g. Every Monday at 9 AM, the sales agent summarises last week\'s data; once the manager approves, post to the Feishu ops group',
+      submit: 'Generate',
+      running: 'Generating…',
+      retry: 'Try again',
+      accept: 'Accept & open editor',
+      compileOk: 'Compile preview OK',
+      compileFail: 'Compile preview failed ({count})',
+      missing: 'missing',
+      warning: 'warning',
+      previewDraft: 'Preview generated JSON',
+      failed: 'Generate failed: {msg}',
+    },
     paused: {
       header: 'Paused runs ({count})',
       empty: 'No paused runs.',

mateclaw-ui/src/i18n/locales/zh-CN.ts

@@ -2033,6 +2033,22 @@ export default {
       deleteContent: '将删除工作流「{name}」？此操作可通过审计日志撤销。',
     },
     selectHint: '在左侧选择一个工作流，或点击「新建工作流」开始。',
+    generate: {
+      entryButton: '从描述生成',
+      title: '从自然语言生成工作流草稿',
+      hint: '用一句话描述你的业务流程，AI 会生成 workflow JSON 草稿。生成后会保留为草稿，需要你确认后再发布；不会自动启用任何触发器。',
+      placeholder: '例如：每周一早上 9 点让销售员工汇总上周数据，老板确认后发到飞书运营群',
+      submit: '生成草稿',
+      running: '生成中...',
+      retry: '重试',
+      accept: '采用并打开编辑器',
+      compileOk: '编译预校验通过',
+      compileFail: '编译预校验未通过 ({count})',
+      missing: '缺失',
+      warning: '警告',
+      previewDraft: '预览生成的 JSON',
+      failed: '生成失败：{msg}',
+    },
     paused: {
       header: '待恢复运行 ({count})',
       empty: '当前没有暂停中的运行。',

mateclaw-ui/src/views/Workflows.vue

@@ -8,7 +8,10 @@
             <h1 class="mc-page-title">{{ t('workflows.title') }}</h1>
             <p class="mc-page-desc">{{ t('workflows.desc') }}</p>
           </div>
-          <button class="btn-primary" @click="openCreate">{{ t('workflows.newWorkflow') }}</button>
+          <div class="header-actions">
+            <button class="btn-ghost" @click="openGenerate">{{ t('workflows.generate.entryButton') }}</button>
+            <button class="btn-primary" @click="openCreate">{{ t('workflows.newWorkflow') }}</button>
+          </div>
         </div>
 
         <div class="workflows-grid">
@@ -205,6 +208,7 @@
 
     <CreateWorkflowDialog v-model="createDialogOpen" :loading="busy" @submit="onCreateSubmit" />
     <PublishDialog v-model="publishDialogOpen" :loading="busy" @submit="onPublishSubmit" />
+    <GenerateWorkflowDialog v-model="generateDialogOpen" @accept="onGenerateAccept" />
   </div>
 </template>
 
@@ -224,6 +228,7 @@ import {
   type WorkflowCompileFailure,
   type PausedRunSummary,
   type ResumeOutcome,
+  type GeneratedDraft,
 } from '@/api'
 import type { Channel } from '@/types'
 import { useWorkspaceStore } from '@/stores/useWorkspaceStore'
@@ -232,6 +237,7 @@ import StepPropertyPanel from '@/components/workflow/StepPropertyPanel.vue'
 import WorkflowJsonEditor from '@/components/workflow/WorkflowJsonEditor.vue'
 import CreateWorkflowDialog from '@/components/workflow/CreateWorkflowDialog.vue'
 import PublishDialog from '@/components/workflow/PublishDialog.vue'
+import GenerateWorkflowDialog from '@/components/workflow/GenerateWorkflowDialog.vue'
 import type { StepNodeData, RawStep } from '@/composables/useWorkflowGraph'
 import {
   readStepAtIndex,
@@ -319,6 +325,43 @@ function onStepDelete(payload: { index: number }) {
 
 const createDialogOpen = ref(false)
 const publishDialogOpen = ref(false)
+const generateDialogOpen = ref(false)
+
+function openGenerate() {
+  if (!workspaceId.value) return
+  generateDialogOpen.value = true
+}
+
+async function onGenerateAccept(draft: GeneratedDraft) {
+  if (!workspaceId.value) return
+  busy.value = true
+  try {
+    // Create a workflow row + save the generated draft, then jump
+    // straight into the canvas so the operator can finish the TODOs.
+    const res = await workflowApi.create({
+      workspaceId: workspaceId.value,
+      name: draft.name,
+      description: draft.description || undefined,
+      enabled: true,
+    })
+    const created = res.data as unknown as WorkflowSummary
+    if (created?.id) {
+      await workflowApi.saveDraft(created.id, draft.draftJson)
+      await reload()
+      await select(created.id)
+      // If the generator's preview compile failed, surface the errors
+      // inline so the operator sees them immediately on first load.
+      if (!draft.compileOk && draft.compileErrors.length) {
+        compileErrors.value = draft.compileErrors
+      }
+      ElMessage.success(t('workflows.generate.compileOk'))
+    }
+  } catch (e) {
+    ElMessage.error(t('workflows.generate.failed', { msg: (e as Error).message }))
+  } finally {
+    busy.value = false
+  }
+}
 
 // Live JSON-syntax check on the textarea so the operator sees parse errors
 // immediately, instead of waiting for compile to round-trip.