feat: implement streaming chat, skill routing, and SAFe PI planning tools

- Add /api/chat/stream endpoint with Server-Sent Events (SSE) for real-time message streaming
  * Implement StreamEvent types (thought, tool_call, tool_result, final, error)
  * Add StreamEventCallback mechanism for event propagation
  * Create StreamChatHandler in webui/bot with proper HTTP headers and flushing

- Implement LLM-based skill router for intelligent capability selection
  * Add optional routerLLM client for semantic routing
  * Implement routeSkillsWithLLM() to match user intent to available skills
  * Add matchSkillsByName() for fuzzy skill matching
  * Update buildUnifiedSystemPrompt() to use routed skills

- Add streaming support to ReAct pipeline
  * Implement runUnifiedReActStream() for streaming thought/action/observation
  * Emit StreamEvent at each ReAct step
  * Support callback error handling in streaming mode

- Integrate three new DevOps tools
  * tools/filedoc: Extract document content from file_id via OpenAI
  * tools/giteaticket: Create Gitea issues from PI plan items with SAFe metadata
  * tools/piplan: Publish PI planning blueprints with dependency tracking

- Add SAFe PI Planning skill
  * Implement PM/SA/RTE (iron triangle) workflow
  * Support for Feature, Enabler, and Dependency definition
  * Automatic task decomposition and Gitea integration

- Create frontend integration documentation
  * Complete SSE protocol specification
  * TypeScript fetch + ReadableStream example
  * LLM-ready refactoring template for other projects

- Simplify file handling
  * Remove legacy file context structures and dual-mode processing
  * Consolidate file operations into UploadAndCacheFiles()
  * Remove FilePromptMode configuration and related complexity

- Update configuration
  * Add Router model support (LLM_ROUTER_MODEL)
  * Add Gitea configuration (BaseURL, Token, Owner, Repo)
  * WebSearch and additional tool infrastructure

Tests: All 22 test packages passing, 8/8 webui tests including 3 new stream tests

cmd/bot/main.go

@@ -126,9 +126,19 @@ func main() {
 	// 实例化 LLM 客户端
 	llmClient := llm.NewOpenAICompatibleClient(cfg.LLM, appLogger.WithComponent("llm"))
 
+	// 实例化路由 LLM 客户端（如果配置了独立的路由模型）
+	var routerLLMClient llm.Client
+	if cfg.LLM.RouterModel != "" {
+		routerCfg := cfg.LLM
+		routerCfg.Model = cfg.LLM.RouterModel
+		routerLLMClient = llm.NewOpenAICompatibleClient(routerCfg, appLogger.WithComponent("llm.router"))
+		appLogger.Infof("skill router enabled, model=%s", cfg.LLM.RouterModel)
+	}
+
 	// 创建编排器，整合 LLM、记忆系统、知识技能库与各种工具
 	engine := agent.NewOrchestrator(
 		llmClient,
+		routerLLMClient,
 		store,
 		toolRegistry,
 		soul,
@@ -207,11 +217,18 @@ func runMessageChannel(ctx context.Context, cfg config.Config, engine *agent.Orc
 		return wb.Run(
 			ctx,
 			func(ctx context.Context, msg webui.IncomingMessage) (string, error) {
-				if len(msg.FileIDs) > 0 {
-					return engine.HandleMessageWithFileIDs(ctx, msg.ChatID, msg.UserID, msg.Text, msg.FileIDs)
-				}
 				return engine.HandleMessage(ctx, msg.ChatID, msg.UserID, msg.Text)
 			},
+			func(ctx context.Context, msg webui.IncomingMessage, callback webui.StreamEventCallback) (string, error) {
+				return engine.HandleMessageStream(ctx, msg.ChatID, msg.UserID, msg.Text, func(event agent.StreamEvent) error {
+					return callback(webui.StreamEvent{
+						Type:     webui.StreamEventType(event.Type),
+						Content:  event.Content,
+						Step:     event.Step,
+						ToolName: event.ToolName,
+					})
+				})
+			},
 			func(ctx context.Context, chatID, userID string, files []llm.InputFile) ([]string, error) {
 				return engine.UploadAndCacheFiles(ctx, chatID, userID, files)
 			},

configs/env.sample

@@ -17,6 +17,8 @@ TELEGRAM_POLL_TIMEOUT_SECONDS=30
 FEISHU_APP_ID=
 FEISHU_APP_SECRET=
 FEISHU_VERIFY_TOKEN=
+FEISHU_LISTEN_ADDR=:8080
+FEISHU_EVENT_PATH=/feishu/events
 WEBUI_LISTEN_ADDR=:8090
 WEBUI_MAX_UPLOAD_MB=20
 
@@ -24,7 +26,15 @@ LLM_BASE_URL=https://api.openai.com/v1
 LLM_API_KEY=
 LLM_MODEL=gpt-4o-mini
 LLM_FILE_MODEL=gpt-4o-mini
-LLM_FILE_PROMPT_MODE=user_content_file_parts
+LLM_ROUTER_MODEL=
+
+WEB_SEARCH_ENGINE=duckduckgo
+WEB_SEARCH_API_KEY=
+
+GITEA_BASE_URL=
+GITEA_TOKEN=
+GITEA_OWNER=
+GITEA_REPO=
 
 SQLITE_PATH=./data/laodingbot.db
 ALLOWED_DIRS=./workspace,./data,./skills

doc/WebUI_Stream_API_前端对接说明.md

@@ -0,0 +1,187 @@
+# WebUI `/api/chat/stream` 前端对接说明
+
+本文档用于指导前端项目接入 LaodingBot 的流式聊天接口，并可直接作为提示词输入给 LLM，批量改造其他前端代码。
+
+## 1. 接口总览
+
+- 方法: `POST`
+- 路径: `/api/chat/stream`
+- 请求头: `Content-Type: application/json`
+- 响应类型: `text/event-stream`
+- 协议: SSE (Server-Sent Events)
+
+说明: 该接口为单次请求、多次推送。后端会持续推送 `data: <json>\n\n` 格式的事件。
+
+## 2. 请求体
+
+```json
+{
+  "text": "请帮我分析当前目录",
+  "session_id": "sess_abc",
+  "user_id": "user_001"
+}
+```
+
+字段说明:
+
+- `text` (string, required): 用户输入文本，去除空白后不能为空。
+- `session_id` (string, optional): 会话 ID，不传时后端自动生成。
+- `user_id` (string, optional): 用户 ID，不传时后端自动生成。
+
+兼容字段:
+
+- `sessionId` 等价于 `session_id`
+- `userId` 等价于 `user_id`
+
+## 3. SSE 事件格式
+
+每条 SSE 消息只包含 `data` 字段，内容是 JSON:
+
+```text
+data: {"type":"thought","content":"我先判断是否需要调用工具","step":1}
+
+data: {"type":"tool_call","content":"{\"input\":\"pwd\"}","step":1,"tool_name":"shell"}
+
+data: {"type":"tool_result","content":"C:/Project/MyProject","step":1,"tool_name":"shell"}
+
+data: {"type":"final","content":"当前目录是 C:/Project/MyProject","step":2}
+
+```
+
+事件字段:
+
+- `type` (string): 事件类型
+- `content` (string): 事件文本内容
+- `step` (number, optional): ReAct 步骤编号
+- `tool_name` (string, optional): 工具名
+
+事件类型:
+
+- `thought`: 模型思考过程
+- `tool_call`: 工具调用请求
+- `tool_result`: 工具返回结果
+- `final`: 最终回答
+- `error`: 错误信息
+
+## 4. 连接生命周期
+
+- 正常结束: 收到 `type=final` 后结束渲染，连接可由浏览器自然关闭。
+- 异常结束: 收到 `type=error`，前端应显示错误并结束当前轮次。
+- 网络中断: 前端应允许用户重试，并保留已收到的事件记录。
+
+## 5. 前端渲染建议
+
+推荐将一次请求的事件按 `step` 分组后渲染，典型展示区块:
+
+- 思考区: 逐条显示 `thought`
+- 工具区: 成对显示 `tool_call` 与 `tool_result`
+- 答案区: 显示最后一个 `final`
+- 错误区: 显示 `error`
+
+建议状态机:
+
+- `idle`: 初始状态
+- `streaming`: 请求中且持续接收事件
+- `done`: 收到 `final`
+- `failed`: 收到 `error` 或请求异常
+
+## 6. TypeScript 对接示例 (fetch + ReadableStream)
+
+```ts
+type StreamEventType = 'thought' | 'tool_call' | 'tool_result' | 'final' | 'error';
+
+interface StreamEvent {
+  type: StreamEventType;
+  content: string;
+  step?: number;
+  tool_name?: string;
+}
+
+export async function streamChat(
+  payload: { text: string; session_id?: string; user_id?: string },
+  onEvent: (event: StreamEvent) => void,
+  signal?: AbortSignal,
+): Promise<void> {
+  const resp = await fetch('/api/chat/stream', {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify(payload),
+    signal,
+  });
+
+  if (!resp.ok) {
+    throw new Error(`HTTP ${resp.status}`);
+  }
+  if (!resp.body) {
+    throw new Error('ReadableStream is not available');
+  }
+
+  const reader = resp.body.getReader();
+  const decoder = new TextDecoder('utf-8');
+  let buffer = '';
+
+  while (true) {
+    const { value, done } = await reader.read();
+    if (done) break;
+
+    buffer += decoder.decode(value, { stream: true });
+
+    // SSE message delimiter: blank line
+    let idx = buffer.indexOf('\n\n');
+    while (idx >= 0) {
+      const chunk = buffer.slice(0, idx).trim();
+      buffer = buffer.slice(idx + 2);
+
+      for (const line of chunk.split('\n')) {
+        const text = line.trim();
+        if (!text.startsWith('data:')) continue;
+        const raw = text.slice(5).trim();
+        if (!raw) continue;
+
+        const event = JSON.parse(raw) as StreamEvent;
+        onEvent(event);
+      }
+
+      idx = buffer.indexOf('\n\n');
+    }
+  }
+}
+```
+
+## 7. 给 LLM 的改造任务提示词模板
+
+将下面模板发给 LLM，可用于自动改造其他前端项目:
+
+```text
+你要改造一个前端项目的聊天页面，把非流式接口 `/api/chat` 改为流式接口 `/api/chat/stream`。
+
+后端协议约束:
+1) 请求方法 POST，Content-Type=application/json
+2) 请求体: { text, session_id?, user_id? }
+3) 响应是 SSE 文本流，事件格式为 `data: <json>\n\n`
+4) JSON 结构:
+   - type: thought | tool_call | tool_result | final | error
+   - content: string
+   - step?: number
+   - tool_name?: string
+5) 收到 final 视为本轮完成；收到 error 视为失败
+
+你的改造要求:
+1) 保留现有 UI 风格和组件结构，不做无关重构
+2) 新增流式读取逻辑，支持中途取消 (AbortController)
+3) 将事件按 step 渲染到消息区
+4) 兼容旧会话字段命名 (session_id / sessionId, user_id / userId)
+5) 增加错误态与重试按钮
+6) 不破坏原有上传、历史消息和输入框行为
+7) 输出改动文件列表 + 每个文件的关键变更说明
+
+请直接给出可运行代码补丁。
+```
+
+## 8. 调试清单
+
+- 检查响应头是否为 `text/event-stream`
+- 检查每条事件是否以 `data:` 开头并以空行结尾
+- 确认 `final` 和 `error` 都能正确结束当前轮次
+- 验证弱网下不会丢失已收到的事件
+- 验证用户快速连续提问时，旧流可被取消

internal/agent/orchestrator.go

@@ -18,9 +18,32 @@ import (
 	"laodingbot/internal/tools"
 )
 
+// StreamEventType 定义流式输出事件类型
+type StreamEventType string
+
+const (
+	StreamEventTypeThought    StreamEventType = "thought"     // LLM 思考过程
+	StreamEventTypeToolCall   StreamEventType = "tool_call"   // 工具调用请求
+	StreamEventTypeToolResult StreamEventType = "tool_result" // 工具执行结果
+	StreamEventTypeFinal      StreamEventType = "final"       // 最终答案
+	StreamEventTypeError      StreamEventType = "error"       // 错误信息
+)
+
+// StreamEvent 代表流式输出中的一个事件
+type StreamEvent struct {
+	Type     StreamEventType `json:"type"`
+	Content  string          `json:"content"`
+	Step     int             `json:"step,omitempty"`
+	ToolName string          `json:"tool_name,omitempty"`
+}
+
+// StreamEventCallback 是流式事件回调函数类型，用于推送事件到客户端
+type StreamEventCallback func(event StreamEvent) error
+
 // Orchestrator 负责协调和组合业务逻辑，包含 LLM 计算、上下文管理、技能匹配计算和工具调用。
 type Orchestrator struct {
 	llm                  llm.Client
+	routerLLM            llm.Client // 可选：轻量路由模型，用于技能意图路由；为 nil 则仅用关键词匹配
 	store                *memory.SQLiteStore
 	tools                *tools.Registry
 	soul                 string
@@ -44,16 +67,10 @@ type pendingFileRef struct {
 	MimeType string
 }
 
-type filePromptContext struct {
-	Summary     string
-	FatalReason string
-	FileIDs     []string
-	Uploaded    []pendingFileRef
-}
-
 // NewOrchestrator 创建一个新的编排器对象，初始化关键路径和超时控制等。
 func NewOrchestrator(
 	llmClient llm.Client,
+	routerLLM llm.Client,
 	store *memory.SQLiteStore,
 	registry *tools.Registry,
 	soul string,
@@ -81,6 +98,7 @@ func NewOrchestrator(
 	}
 	return &Orchestrator{
 		llm:                  llmClient,
+		routerLLM:            routerLLM,
 		store:                store,
 		tools:                registry,
 		soul:                 soul,
@@ -103,52 +121,88 @@ func NewOrchestrator(
 // - 是否需要调用工具（action + action_input）
 // 循环持续进行，直到 LLM 返回 is_final_answer=true。
 func (o *Orchestrator) HandleMessage(ctx context.Context, chatID, userID, text string) (string, error) {
-	return o.handleMessageInternal(ctx, chatID, userID, text, nil, false)
+	return o.handleMessageInternal(ctx, chatID, userID, text)
 }
 
+// HandleMessageWithFiles 接收用户消息和文件，上传文件获取 file_id 并缓存，然后进入普通消息处理流程。
 func (o *Orchestrator) HandleMessageWithFiles(ctx context.Context, chatID, userID, text string, files []llm.InputFile) (string, error) {
-	return o.handleMessageInternal(ctx, chatID, userID, text, files, false)
+	if len(files) > 0 {
+		ids, err := o.UploadAndCacheFiles(ctx, chatID, userID, files)
+		if err != nil && o.log != nil {
+			o.log.Warnf("upload files failed chat_id=%s err=%v", chatID, err)
+		}
+		_ = ids
+	}
+	if strings.TrimSpace(text) == "" {
+		return "文件已接收。请继续发送你的问题。", nil
+	}
+	return o.handleMessageInternal(ctx, chatID, userID, text)
 }
 
-// HandleMessageWithFileIDs 接收用户文本与外部 file_id 列表，复用统一 ReAct 链路。
+// HandleMessageStream 接收用户消息并通过流式方式返回回复。
-// 该方法会先把 file_id 注入当前会话上下文，然后调用常规 HandleMessage 流程。
+// 通过 callback 推送实时事件，包括思考过程、工具调用、工具结果和最终答案。
-func (o *Orchestrator) HandleMessageWithFileIDs(ctx context.Context, chatID, userID, text string, fileIDs []string) (string, error) {
+func (o *Orchestrator) HandleMessageStream(ctx context.Context, chatID, userID, text string, callback StreamEventCallback) (string, error) {
-	ids := nonEmptyIDs(fileIDs)
+	if callback == nil {
-	if len(ids) > 0 {
+		return "", fmt.Errorf("stream callback is required")
-		refs := make([]pendingFileRef, 0, len(ids))
-		for _, id := range ids {
-			refs = append(refs, pendingFileRef{ID: id})
 	}
-		o.appendPendingFiles(chatID, userID, refs)
+	return o.handleMessageStreamInternal(ctx, chatID, userID, text, callback)
+}
+
+// HandleMessageStreamWithFiles 接收用户消息和文件，上传文件后进入流式处理流程。
+func (o *Orchestrator) HandleMessageStreamWithFiles(ctx context.Context, chatID, userID, text string, files []llm.InputFile, callback StreamEventCallback) (string, error) {
+	if callback == nil {
+		return "", fmt.Errorf("stream callback is required")
 	}
-	return o.handleMessageInternal(ctx, chatID, userID, text, nil, true)
+	if len(files) > 0 {
+		ids, err := o.UploadAndCacheFiles(ctx, chatID, userID, files)
+		if err != nil && o.log != nil {
+			o.log.Warnf("upload files failed chat_id=%s err=%v", chatID, err)
+		}
+		_ = ids
+	}
+	if strings.TrimSpace(text) == "" {
+		return "文件已接收。请继续发送你的问题。", nil
+	}
+	return o.handleMessageStreamInternal(ctx, chatID, userID, text, callback)
 }
 
 // UploadAndCacheFiles 上传文件到 LLM 并缓存 file_id，供后续同会话文本问答复用。
-// 该方法不会写入 messages 表，仅更新内存中的 pending file 上下文。
 func (o *Orchestrator) UploadAndCacheFiles(ctx context.Context, chatID, userID string, files []llm.InputFile) ([]string, error) {
 	if len(files) == 0 {
 		return nil, fmt.Errorf("no files provided")
 	}
-	uploadCtx := o.prepareFilePromptContext(ctx, files, nil)
+	uploader, ok := o.llm.(llm.FileUploader)
-	if strings.TrimSpace(uploadCtx.FatalReason) != "" {
+	if !ok {
-		return nil, fmt.Errorf(uploadCtx.FatalReason)
+		return nil, fmt.Errorf("当前 LLM 客户端不支持文件上传接口")
 	}
-	ids := nonEmptyIDs(uploadCtx.FileIDs)
+	var ids []string
-	if len(ids) == 0 {
+	var refs []pendingFileRef
-		return nil, fmt.Errorf("file upload completed but no valid file_id returned")
+	for i, f := range files {
+		if strings.TrimSpace(f.FileName) == "" || len(f.Content) == 0 {
+			return nil, fmt.Errorf("file[%d] 缺少文件名或内容", i+1)
 		}
-	o.appendPendingFiles(chatID, userID, uploadCtx.toPendingRefs())
+		fileID, err := uploader.UploadFile(ctx, f, "file-extract")
+		if err != nil {
+			return nil, fmt.Errorf("file[%d] name=%s 上传失败: %w", i+1, f.FileName, err)
+		}
+		ids = append(ids, fileID)
+		refs = append(refs, pendingFileRef{
+			ID:       fileID,
+			Name:     strings.TrimSpace(f.FileName),
+			MimeType: defaultIfEmpty(strings.TrimSpace(f.MimeType), "application/octet-stream"),
+		})
+	}
+	o.appendPendingFiles(chatID, userID, refs)
 	return ids, nil
 }
 
-func (o *Orchestrator) handleMessageInternal(ctx context.Context, chatID, userID, text string, files []llm.InputFile, appendFileIDText bool) (string, error) {
+func (o *Orchestrator) handleMessageInternal(ctx context.Context, chatID, userID, text string) (string, error) {
 	// 为链路追踪设置唯一的 TraceID
 	traceID := logger.NewTraceID()
 	ctx = logger.WithTraceID(ctx, traceID)
 	traceLogPrefix := "trace_id=" + traceID
 	if o.log != nil {
-		o.log.Infof("%s handle message chat_id=%s user_id=%s text_len=%d files=%d", traceLogPrefix, chatID, userID, len(text), len(files))
+		o.log.Infof("%s handle message chat_id=%s user_id=%s text_len=%d", traceLogPrefix, chatID, userID, len(text))
 		o.log.Debugf("%s handle message text=%q", traceLogPrefix, text)
 	}
 
@@ -169,38 +223,6 @@ func (o *Orchestrator) handleMessageInternal(ctx context.Context, chatID, userID
 		return report, nil
 	}
 
-	trimmedText := strings.TrimSpace(text)
-	isFileOnly := len(files) > 0 && trimmedText == ""
-
-	if isFileOnly {
-		if err := o.store.SaveMessage(chatID, userID, "user", "[FILE_UPLOAD]"); err != nil {
-			if o.log != nil {
-				o.log.Errorf("%s save file-only user marker failed chat_id=%s err=%v", traceLogPrefix, chatID, err)
-			}
-			return "", err
-		}
-		uploadCtx := o.prepareFilePromptContext(ctx, files, nil)
-		if strings.TrimSpace(uploadCtx.FatalReason) != "" {
-			finalText := "文件上传失败，无法建立文档上下文。" + "\n" + uploadCtx.FatalReason
-			if err := o.store.SaveMessage(chatID, userID, "assistant", finalText); err != nil && o.log != nil {
-				o.log.Warnf("%s save upload failure message failed chat_id=%s err=%v", traceLogPrefix, chatID, err)
-			}
-			return finalText, nil
-		}
-		o.appendPendingFiles(chatID, userID, uploadCtx.toPendingRefs())
-		finalText := o.buildFileUploadAck(uploadCtx)
-		if err := o.store.SaveMessage(chatID, userID, "assistant", finalText); err != nil {
-			if o.log != nil {
-				o.log.Errorf("%s save file upload ack failed chat_id=%s err=%v", traceLogPrefix, chatID, err)
-			}
-			return "", err
-		}
-		if o.log != nil {
-			o.log.Infof("%s file-only message handled chat_id=%s cached_files=%d", traceLogPrefix, chatID, len(uploadCtx.FileIDs))
-		}
-		return finalText, nil
-	}
-
 	// 保存用户消息到 SQLite 中
 	if err := o.store.SaveMessage(chatID, userID, "user", text); err != nil {
 		if o.log != nil {
@@ -223,28 +245,13 @@ func (o *Orchestrator) handleMessageInternal(ctx context.Context, chatID, userID
 	}
 
 	// 进入统一 ReAct 循环
-	pendingRefs := o.getPendingFiles(chatID, userID)
+	response, err := o.runUnifiedReAct(ctx, chatID, userID, compressed, text)
-	fileCtx := o.prepareFilePromptContext(ctx, files, pendingRefs)
-	if strings.TrimSpace(fileCtx.FatalReason) != "" {
-		finalText := "文件上传失败，无法继续进行文档解析。" + "\n" + fileCtx.FatalReason
-		if err := o.store.SaveMessage(chatID, userID, "assistant", finalText); err != nil && o.log != nil {
-			o.log.Warnf("%s save assistant failure message failed chat_id=%s err=%v", traceLogPrefix, chatID, err)
-		}
-		if o.log != nil {
-			o.log.Warnf("%s stop before react due to file upload failure reason=%s", traceLogPrefix, fileCtx.FatalReason)
-		}
-		return finalText, nil
-	}
-	response, err := o.runUnifiedReAct(ctx, chatID, userID, compressed, text, fileCtx, appendFileIDText)
 	if err != nil {
 		if o.log != nil {
 			o.log.Errorf("%s message generation failed chat_id=%s err=%v", traceLogPrefix, chatID, err)
 		}
 		return "", err
 	}
-	if len(pendingRefs) > 0 {
-		o.clearPendingFiles(chatID, userID)
-	}
 
 	// 最终将机器人的回复也加入记忆缓存
 	if err := o.store.SaveMessage(chatID, userID, "assistant", response); err != nil {
@@ -260,11 +267,89 @@ func (o *Orchestrator) handleMessageInternal(ctx context.Context, chatID, userID
 	return response, nil
 }
 
+// handleMessageStreamInternal 处理流式消息的内部逻辑，类似于handleMessageInternal但支持流式回调
+func (o *Orchestrator) handleMessageStreamInternal(ctx context.Context, chatID, userID, text string, callback StreamEventCallback) (string, error) {
+	// 为链路追踪设置唯一的 TraceID
+	traceID := logger.NewTraceID()
+	ctx = logger.WithTraceID(ctx, traceID)
+	traceLogPrefix := "trace_id=" + traceID
+	if o.log != nil {
+		o.log.Infof("%s handle message stream chat_id=%s user_id=%s text_len=%d", traceLogPrefix, chatID, userID, len(text))
+		o.log.Debugf("%s handle message stream text=%q", traceLogPrefix, text)
+	}
+
+	// 处理特殊的重载指令
+	if strings.EqualFold(strings.TrimSpace(text), "/reload_skills") {
+		if err := o.ReloadSkills(); err != nil {
+			return "技能热加载失败: " + err.Error(), nil
+		}
+		return "技能已热加载完成。", nil
+	}
+
+	// 如果用户请求能力缺口报告，则生成报告格式化输出
+	if strings.EqualFold(strings.TrimSpace(text), "/capability_gaps") {
+		report, err := o.BuildCapabilityGapReport(10)
+		if err != nil {
+			return "缺口报告生成失败: " + err.Error(), nil
+		}
+		return report, nil
+	}
+
+	// 保存用户消息到 SQLite 中
+	if err := o.store.SaveMessage(chatID, userID, "user", text); err != nil {
+		if o.log != nil {
+			o.log.Errorf("%s save user message failed chat_id=%s err=%v", traceLogPrefix, chatID, err)
+		}
+		return "", err
+	}
+
+	// 读取最近的会话记忆并压缩成 Prompt 上下文
+	recent, err := o.store.LoadRecent(chatID, 16)
+	if err != nil {
+		if o.log != nil {
+			o.log.Errorf("%s load recent failed chat_id=%s err=%v", traceLogPrefix, chatID, err)
+		}
+		return "", err
+	}
+	compressed := memory.CompressForPrompt(recent, 6000)
+	if o.log != nil {
+		o.log.Debugf("%s stream prompt context prepared chat_id=%s recent_count=%d compressed_len=%d", traceLogPrefix, chatID, len(recent), len(compressed))
+	}
+
+	// 进入流式统一 ReAct 循环
+	response, err := o.runUnifiedReActStream(ctx, chatID, userID, compressed, text, callback)
+	if err != nil {
+		if o.log != nil {
+			o.log.Errorf("%s stream message generation failed chat_id=%s err=%v", traceLogPrefix, chatID, err)
+		}
+		return "", err
+	}
+
+	// 最终将机器人的回复也加入记忆缓存
+	if err := o.store.SaveMessage(chatID, userID, "assistant", response); err != nil {
+		if o.log != nil {
+			o.log.Errorf("%s save assistant response failed chat_id=%s err=%v", traceLogPrefix, chatID, err)
+		}
+		return "", err
+	}
+
+	if o.log != nil {
+		o.log.Infof("%s stream message handled chat_id=%s response_len=%d", traceLogPrefix, chatID, len(response))
+	}
+	return response, nil
+}
+
 // buildUnifiedSystemPrompt 构建统一 ReAct 循环的 system prompt。
 // 工具定义通过 API 的 tools 字段传递；此处只需包含人格、技能、运行环境和思考指引。
-func (o *Orchestrator) buildUnifiedSystemPrompt(userInput string) string {
+// routedSkills 为 LLM 路由预选的技能列表；如果为 nil，则回退到关键词匹配。
+func (o *Orchestrator) buildUnifiedSystemPrompt(userInput string, routedSkills []knowledge.Skill) string {
 	skillMetaDoc := o.formatSkillSummariesForPrompt()
-	relevantSkillsDoc := o.formatSelectedSkillsForPrompt(userInput, nil)
+	var relevantSkillsDoc string
+	if routedSkills != nil {
+		relevantSkillsDoc = o.formatSelectedSkillsForPrompt(userInput, routedSkills)
+	} else {
+		relevantSkillsDoc = o.formatSelectedSkillsForPrompt(userInput, nil)
+	}
 	runtimeDoc := formatRuntimeContextForPrompt()
 
 	return strings.Join([]string{
@@ -292,20 +377,151 @@ func (o *Orchestrator) buildUnifiedSystemPrompt(userInput string) string {
 		"",
 		"===== 本轮相关技能（按用户问题筛选） =====",
 		relevantSkillsDoc,
+		"",
+		"===== 关键约束 =====",
 	}, "\n")
 }
 
+// routeSkillsWithLLM 使用轻量 LLM 模型对用户输入进行语义路由，判断是否需要加载技能以及选择哪些技能。
+// 返回匹配到的技能列表（可能为空切片表示不需要技能，nil 表示调用失败应回退）。
+func (o *Orchestrator) routeSkillsWithLLM(ctx context.Context, userInput string) ([]knowledge.Skill, error) {
+	traceLogPrefix := "trace_id=" + logger.TraceIDFromContext(ctx)
+
+	summaries := o.getSkillSummariesSnapshot()
+	if len(summaries) == 0 {
+		if o.log != nil {
+			o.log.Debugf("%s skill router: no skills available, skip", traceLogPrefix)
+		}
+		return []knowledge.Skill{}, nil
+	}
+
+	// 构建技能池描述
+	skillPool := strings.Builder{}
+	for _, s := range summaries {
+		name := strings.TrimSpace(s.Name)
+		desc := strings.TrimSpace(s.Description)
+		if name == "" {
+			continue
+		}
+		skillPool.WriteString("- ")
+		skillPool.WriteString(name)
+		if desc != "" {
+			skillPool.WriteString(": ")
+			skillPool.WriteString(desc)
+		}
+		skillPool.WriteString("\n")
+	}
+
+	routerSystemPrompt := strings.Join([]string{
+		"你是一个意图路由器。根据用户输入，从技能池中挑选最合适的技能。",
+		"",
+		"规则：",
+		"1. 如果用户的问题可以直接回答（闲聊、简单问答）或只需简单工具调用，设置 need_skills=false，selected_skills 为空数组。",
+		"2. 如果用户的问题涉及专业流程、复杂任务或与某个技能高度相关，设置 need_skills=true 并选择最相关的技能名称。",
+		"3. 最多选择 3 个技能。",
+		"4. 仅返回 JSON，不要附加任何其他文字。",
+		"",
+		"可用技能池：",
+		strings.TrimSpace(skillPool.String()),
+		"",
+		"输出格式（严格 JSON）：",
+		`{"need_skills": true, "selected_skills": ["技能名称1"], "reason": "简要说明"}`,
+	}, "\n")
+
+	routerUserPrompt := "用户输入：" + userInput
+
+	if o.log != nil {
+		o.log.Debugf("%s skill router request: skills_count=%d input_len=%d", traceLogPrefix, len(summaries), len(userInput))
+	}
+
+	raw, err := o.routerLLM.Generate(ctx, routerSystemPrompt, routerUserPrompt)
+	if err != nil {
+		return nil, fmt.Errorf("router llm call failed: %w", err)
+	}
+
+	if o.log != nil {
+		o.log.Debugf("%s skill router response: %s", traceLogPrefix, truncateForLog(raw, 500))
+	}
+
+	decision, err := parseCapabilityRoute(raw)
+	if err != nil {
+		return nil, fmt.Errorf("router response parse failed: %w", err)
+	}
+
+	if !decision.NeedSkills || len(decision.SelectedSkills) == 0 {
+		if o.log != nil {
+			o.log.Infof("%s skill router: no skills needed, reason=%s", traceLogPrefix, decision.Reason)
+		}
+		return []knowledge.Skill{}, nil
+	}
+
+	// 根据路由结果匹配完整技能内容
+	allSkills := o.getSkillsSnapshot()
+	selected := matchSkillsByName(allSkills, decision.SelectedSkills)
+
+	if o.log != nil {
+		o.log.Infof("%s skill router: need_skills=true requested=%v matched=%d reason=%s",
+			traceLogPrefix, decision.SelectedSkills, len(selected), decision.Reason)
+	}
+
+	return selected, nil
+}
+
+// matchSkillsByName 根据名称列表从全量技能中模糊匹配。
+func matchSkillsByName(allSkills []knowledge.Skill, names []string) []knowledge.Skill {
+	if len(names) == 0 {
+		return nil
+	}
+	matched := make([]knowledge.Skill, 0, len(names))
+	for _, wantName := range names {
+		want := strings.ToLower(strings.TrimSpace(wantName))
+		if want == "" {
+			continue
+		}
+		for _, sk := range allSkills {
+			skName := strings.ToLower(strings.TrimSpace(sk.Name))
+			if skName == want || strings.Contains(skName, want) || strings.Contains(want, skName) {
+				matched = append(matched, sk)
+				break
+			}
+		}
+	}
+	return matched
+}
+
 // runUnifiedReAct 执行统一的 ReAct 循环，使用原生 function calling API。
 // messages 数组随交互动态增长：system → history → user → assistant(tool_calls) → tool → ...
 // 循环持续到 LLM 返回无 tool_calls 的纯文本回复（即最终回答）或达到安全上限。
-func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, compressedContext, userInput string, fileCtx filePromptContext, appendFileIDText bool) (string, error) {
+func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, compressedContext, userInput string) (string, error) {
 	traceID := logger.TraceIDFromContext(ctx)
 	traceLogPrefix := "trace_id=" + traceID
 
-	systemPrompt := o.buildUnifiedSystemPrompt(userInput)
+	// ===== LLM 意图路由：使用轻量模型判断是否需要加载技能 =====
+	var routedSkills []knowledge.Skill
+	if o.routerLLM != nil {
+		routed, routeErr := o.routeSkillsWithLLM(ctx, userInput)
+		if routeErr != nil {
+			if o.log != nil {
+				o.log.Warnf("%s skill router failed, fallback to keyword matching err=%v", traceLogPrefix, routeErr)
+			}
+			// 路由失败时 routedSkills 保持 nil，buildUnifiedSystemPrompt 回退到关键词匹配
+		} else {
+			routedSkills = routed
+			if o.log != nil {
+				names := make([]string, 0, len(routedSkills))
+				for _, sk := range routedSkills {
+					names = append(names, sk.Name)
+				}
+				o.log.Infof("%s skill router selected %d skills: %v", traceLogPrefix, len(routedSkills), names)
+			}
+		}
+	}
+
+	systemPrompt := o.buildUnifiedSystemPrompt(userInput, routedSkills)
 
 	if o.log != nil {
 		o.log.Infof("%s unified react start", traceLogPrefix)
+		o.log.Debugf("%s system_prompt_len=%d", traceLogPrefix, len(systemPrompt))
 	}
 
 	// 检查 LLM 客户端是否支持原生 tool_calls
@@ -314,7 +530,7 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
 		if o.log != nil {
 			o.log.Warnf("%s llm client does not support ToolCallChatClient, falling back to legacy ReAct", traceLogPrefix)
 		}
-		return o.runLegacyReAct(ctx, chatID, userID, compressedContext, userInput, fileCtx, appendFileIDText)
+		return o.runLegacyReAct(ctx, chatID, userID, compressedContext, userInput)
 	}
 
 	// 构建初始 messages 数组
@@ -322,13 +538,20 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
 	messages = append(messages, llm.PromptMessage{Role: "system", Content: systemPrompt})
 
 	// 加入历史会话上下文
-	//messages = append(messages, parseCompressedHistoryMessages(compressedContext)...)
+	messages = append(messages, parseCompressedHistoryMessages(compressedContext)...)
 
 	// 加入当前用户消息
 	messages = append(messages, llm.PromptMessage{Role: "user", Content: userInput})
 
 	// 构建工具定义列表（通过 API tools 字段传递）
 	toolDefs := o.buildToolDefinitions()
+	if o.log != nil {
+		toolNames := make([]string, 0, len(toolDefs))
+		for _, td := range toolDefs {
+			toolNames = append(toolNames, td.Function.Name)
+		}
+		o.log.Debugf("%s tool_defs_count=%d names=%v", traceLogPrefix, len(toolDefs), toolNames)
+	}
 
 	const maxSteps = 20
 	for step := 1; step <= maxSteps; step++ {
@@ -337,7 +560,7 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
 		}
 
 		// 调用 LLM（传入完整 messages + tools 定义）
-		completion, err := toolCallClient.GenerateWithTools(ctx, messages, toolDefs, fileCtx.FileIDs, appendFileIDText)
+		completion, err := toolCallClient.GenerateWithTools(ctx, messages, toolDefs)
 		if err != nil {
 			return "", err
 		}
@@ -391,7 +614,8 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
 			}
 
 			if o.log != nil {
-				o.log.Infof("%s react step=%d tool_call tool=%s input=%q", traceLogPrefix, step, toolName, toolInput)
+				o.log.Infof("%s react step=%d tool_call tool=%s input_len=%d", traceLogPrefix, step, toolName, len(toolInput))
+				o.log.Debugf("%s react step=%d tool=%s input=%q", traceLogPrefix, step, toolName, toolInput)
 			}
 
 			toolOut, toolErr := tool.Call(ctx, toolInput)
@@ -410,6 +634,7 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
 
 			if o.log != nil {
 				o.log.Infof("%s react step=%d tool=%s observation_len=%d", traceLogPrefix, step, toolName, len(obs))
+				o.log.Debugf("%s react step=%d tool=%s observation=%q", traceLogPrefix, step, toolName, truncateForLog(obs, 500))
 			}
 
 			messages = append(messages, llm.PromptMessage{
@@ -426,8 +651,217 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
 	return "我尝试了多轮推理与工具调用，但仍未得到稳定结论。请给我更具体的约束或允许我继续尝试。", nil
 }
 
+// runUnifiedReActStream 执行统一的 ReAct 循环并通过回调推送流式事件。
+func (o *Orchestrator) runUnifiedReActStream(ctx context.Context, chatID, userID, compressedContext, userInput string, callback StreamEventCallback) (string, error) {
+	traceID := logger.TraceIDFromContext(ctx)
+	traceLogPrefix := "trace_id=" + traceID
+
+	// ===== LLM 意图路由：使用轻量模型判断是否需要加载技能 =====
+	var routedSkills []knowledge.Skill
+	if o.routerLLM != nil {
+		routed, routeErr := o.routeSkillsWithLLM(ctx, userInput)
+		if routeErr != nil {
+			if o.log != nil {
+				o.log.Warnf("%s skill router failed, fallback to keyword matching err=%v", traceLogPrefix, routeErr)
+			}
+		} else {
+			routedSkills = routed
+			if o.log != nil {
+				names := make([]string, 0, len(routedSkills))
+				for _, sk := range routedSkills {
+					names = append(names, sk.Name)
+				}
+				o.log.Infof("%s skill router selected %d skills: %v", traceLogPrefix, len(routedSkills), names)
+			}
+		}
+	}
+
+	systemPrompt := o.buildUnifiedSystemPrompt(userInput, routedSkills)
+
+	if o.log != nil {
+		o.log.Infof("%s unified react stream start", traceLogPrefix)
+		o.log.Debugf("%s system_prompt_len=%d", traceLogPrefix, len(systemPrompt))
+	}
+
+	// 检查 LLM 客户端是否支持原生 tool_calls
+	toolCallClient, supportsToolCalls := o.llm.(llm.ToolCallChatClient)
+	if !supportsToolCalls {
+		if o.log != nil {
+			o.log.Warnf("%s llm client does not support ToolCallChatClient, stream mode not available", traceLogPrefix)
+		}
+		return "", fmt.Errorf("stream mode requires ToolCallChatClient support")
+	}
+
+	// 构建初始 messages 数组
+	messages := make([]llm.PromptMessage, 0, 32)
+	messages = append(messages, llm.PromptMessage{Role: "system", Content: systemPrompt})
+	messages = append(messages, parseCompressedHistoryMessages(compressedContext)...)
+	messages = append(messages, llm.PromptMessage{Role: "user", Content: userInput})
+
+	// 构建工具定义列表
+	toolDefs := o.buildToolDefinitions()
+	if o.log != nil {
+		toolNames := make([]string, 0, len(toolDefs))
+		for _, td := range toolDefs {
+			toolNames = append(toolNames, td.Function.Name)
+		}
+		o.log.Debugf("%s tool_defs_count=%d names=%v", traceLogPrefix, len(toolDefs), toolNames)
+	}
+
+	const maxSteps = 20
+	for step := 1; step <= maxSteps; step++ {
+		if o.log != nil {
+			o.log.Infof("%s react stream step=%d start messages_count=%d", traceLogPrefix, step, len(messages))
+		}
+
+		// 调用 LLM
+		completion, err := toolCallClient.GenerateWithTools(ctx, messages, toolDefs)
+		if err != nil {
+			return "", err
+		}
+
+		if o.log != nil {
+			o.log.Infof("%s react stream step=%d content_len=%d tool_calls=%d",
+				traceLogPrefix, step, len(completion.Content), len(completion.ToolCalls))
+			if completion.Content != "" {
+				o.log.Debugf("%s react stream step=%d thought=%q", traceLogPrefix, step, completion.Content)
+			}
+		}
+
+		// 推送思考过程事件
+		if completion.Content != "" {
+			if err := callback(StreamEvent{
+				Type:    StreamEventTypeThought,
+				Content: completion.Content,
+				Step:    step,
+			}); err != nil {
+				return "", fmt.Errorf("callback error: %w", err)
+			}
+		}
+
+		// ========== 无 tool_calls → 最终回答 ==========
+		if len(completion.ToolCalls) == 0 {
+			finalText := strings.TrimSpace(completion.Content)
+			if finalText == "" {
+				finalText = "已完成处理。"
+			}
+			if o.log != nil {
+				o.log.Infof("%s react stream final at step=%d answer_len=%d", traceLogPrefix, step, len(finalText))
+			}
+			// 推送最终答案事件
+			if err := callback(StreamEvent{
+				Type:    StreamEventTypeFinal,
+				Content: finalText,
+				Step:    step,
+			}); err != nil {
+				return "", fmt.Errorf("callback error: %w", err)
+			}
+			return finalText, nil
+		}
+
+		// ========== 有 tool_calls → 执行工具 ==========
+		assistantMsg := llm.PromptMessage{
+			Role:      "assistant",
+			Content:   completion.Content,
+			ToolCalls: completion.ToolCalls,
+		}
+		messages = append(messages, assistantMsg)
+
+		// 逐个执行工具调用
+		for _, tc := range completion.ToolCalls {
+			toolName := strings.ToLower(strings.TrimSpace(tc.Function.Name))
+			toolInput := extractToolInput(tc.Function.Arguments)
+
+			// 推送工具调用事件
+			if err := callback(StreamEvent{
+				Type:     StreamEventTypeToolCall,
+				Content:  toolInput,
+				Step:     step,
+				ToolName: toolName,
+			}); err != nil {
+				return "", fmt.Errorf("callback error: %w", err)
+			}
+
+			tool, ok := o.tools.Get(toolName)
+			if !ok {
+				if o.log != nil {
+					o.log.Warnf("%s react stream step=%d tool_not_found=%s", traceLogPrefix, step, toolName)
+				}
+				// 推送错误事件
+				errMsg := "工具不存在：" + toolName
+				if err := callback(StreamEvent{
+					Type:     StreamEventTypeError,
+					Content:  errMsg,
+					Step:     step,
+					ToolName: toolName,
+				}); err != nil {
+					return "", fmt.Errorf("callback error: %w", err)
+				}
+				messages = append(messages, llm.PromptMessage{
+					Role:       "tool",
+					ToolCallID: tc.ID,
+					Name:       tc.Function.Name,
+					Content:    formatToolErrorObservation("TOOL_NOT_FOUND", toolName, "该工具不存在，请检查工具名称后重试"),
+				})
+				o.emitCapabilityGap(chatID, userID, userInput, "tool_not_found:"+toolName)
+				continue
+			}
+
+			if o.log != nil {
+				o.log.Infof("%s react stream step=%d tool_call tool=%s input_len=%d", traceLogPrefix, step, toolName, len(toolInput))
+				o.log.Debugf("%s react stream step=%d tool=%s input=%q", traceLogPrefix, step, toolName, toolInput)
+			}
+
+			toolOut, toolErr := tool.Call(ctx, toolInput)
+			obs := strings.TrimSpace(toolOut)
+			if obs == "" {
+				obs = "(empty output)"
+			}
+			if toolErr != nil {
+				obs = formatToolErrorObservation("TOOL_EXEC_ERROR", toolName, toolErr.Error()) + "\nOUTPUT:\n" + obs
+				o.emitCapabilityGap(chatID, userID, userInput, "tool_call_failed:"+toolName)
+			}
+			// 限制观察值长度防止超出 LLM 上下文窗口
+			if len(obs) > 4000 {
+				obs = obs[:4000] + "\n...(truncated)"
+			}
+
+			if o.log != nil {
+				o.log.Infof("%s react stream step=%d tool=%s observation_len=%d", traceLogPrefix, step, toolName, len(obs))
+				o.log.Debugf("%s react stream step=%d tool=%s observation=%q", traceLogPrefix, step, toolName, truncateForLog(obs, 500))
+			}
+
+			// 推送工具结果事件
+			if err := callback(StreamEvent{
+				Type:     StreamEventTypeToolResult,
+				Content:  obs,
+				Step:     step,
+				ToolName: toolName,
+			}); err != nil {
+				return "", fmt.Errorf("callback error: %w", err)
+			}
+
+			messages = append(messages, llm.PromptMessage{
+				Role:       "tool",
+				ToolCallID: tc.ID,
+				Name:       tc.Function.Name,
+				Content:    obs,
+			})
+		}
+	}
+
+	// 达到安全上限
+	o.emitCapabilityGap(chatID, userID, userInput, "react_step_exhausted")
+	errMsg := "我尝试了多轮推理与工具调用，但仍未得到稳定结论。请给我更具体的约束或允许我继续尝试。"
+	_ = callback(StreamEvent{
+		Type:    StreamEventTypeError,
+		Content: errMsg,
+	})
+	return errMsg, nil
+}
+
 // runLegacyReAct 是旧版基于 JSON 决策解析的 ReAct 循环，作为不支持 tool_calls 的 LLM 的降级方案。
-func (o *Orchestrator) runLegacyReAct(ctx context.Context, chatID, userID, compressedContext, userInput string, fileCtx filePromptContext, appendFileIDText bool) (string, error) {
+func (o *Orchestrator) runLegacyReAct(ctx context.Context, chatID, userID, compressedContext, userInput string) (string, error) {
 	traceID := logger.TraceIDFromContext(ctx)
 	traceLogPrefix := "trace_id=" + traceID
 
@@ -441,8 +875,8 @@ func (o *Orchestrator) runLegacyReAct(ctx context.Context, chatID, userID, compr
 			o.log.Infof("%s legacy react step=%d start", traceLogPrefix, step)
 		}
 
-		messages := buildReActMessages(systemPrompt, compressedContext, userInput, fileCtx.Summary, scratchpad)
+		messages := buildReActMessages(systemPrompt, compressedContext, userInput, scratchpad)
-		raw, err := o.generateWithOptionalFilesMessages(ctx, messages, fileCtx.FileIDs, appendFileIDText)
+		raw, err := o.generateMessages(ctx, messages)
 		if err != nil {
 			return "", err
 		}
@@ -584,105 +1018,19 @@ func extractToolInput(arguments string) string {
 	return arguments
 }
 
-func (o *Orchestrator) prepareFilePromptContext(ctx context.Context, files []llm.InputFile, pending []pendingFileRef) filePromptContext {
+func (o *Orchestrator) generateMessages(ctx context.Context, messages []llm.PromptMessage) (string, error) {
-	ctxOut := filePromptContext{}
-	if len(pending) > 0 {
-		for _, p := range pending {
-			id := strings.TrimSpace(p.ID)
-			if id == "" {
-				continue
-			}
-			ctxOut.FileIDs = append(ctxOut.FileIDs, id)
-		}
-	}
-	if len(files) == 0 {
-		ctxOut.Summary = buildFileSummary(pending, nil)
-		return ctxOut
-	}
-	uploader, ok := o.llm.(llm.FileUploader)
-	if !ok {
-		return filePromptContext{FatalReason: "检测到文件输入，但当前 LLM 客户端不支持文件上传接口。"}
-	}
-
-	uploaded := make([]pendingFileRef, 0, len(files))
-	for i, f := range files {
-		if strings.TrimSpace(f.FileName) == "" || len(f.Content) == 0 {
-			return filePromptContext{FatalReason: fmt.Sprintf("file[%d] 缺少文件名或内容，无法上传。", i+1)}
-		}
-		fileID, err := uploader.UploadFile(ctx, f, "file-extract")
-		if err != nil {
-			return filePromptContext{FatalReason: fmt.Sprintf("file[%d] name=%s 上传失败: %v", i+1, f.FileName, err)}
-		}
-		ctxOut.FileIDs = append(ctxOut.FileIDs, fileID)
-		uploaded = append(uploaded, pendingFileRef{
-			ID:       fileID,
-			Name:     strings.TrimSpace(f.FileName),
-			MimeType: defaultIfEmpty(strings.TrimSpace(f.MimeType), "application/octet-stream"),
-		})
-	}
-	ctxOut.Uploaded = uploaded
-	ctxOut.Summary = buildFileSummary(pending, uploaded)
-	return ctxOut
-}
-
-func buildFileSummary(pending, uploaded []pendingFileRef) string {
-	if len(pending) == 0 && len(uploaded) == 0 {
-		return ""
-	}
-	lines := make([]string, 0, len(pending)+len(uploaded)+2)
-	lines = append(lines, "以下文件 file_id 可用于本轮问答：")
-	idx := 1
-	for _, p := range pending {
-		id := strings.TrimSpace(p.ID)
-		if id == "" {
-			continue
-		}
-		lines = append(lines, fmt.Sprintf("- cached_file[%d] name=%s mime=%s file_id=%s", idx, defaultIfEmpty(strings.TrimSpace(p.Name), "(unknown)"), defaultIfEmpty(strings.TrimSpace(p.MimeType), "application/octet-stream"), id))
-		idx++
-	}
-	for _, p := range uploaded {
-		id := strings.TrimSpace(p.ID)
-		if id == "" {
-			continue
-		}
-		lines = append(lines, fmt.Sprintf("- uploaded_file[%d] name=%s mime=%s file_id=%s", idx, defaultIfEmpty(strings.TrimSpace(p.Name), "(unknown)"), defaultIfEmpty(strings.TrimSpace(p.MimeType), "application/octet-stream"), id))
-		idx++
-	}
-	if len(lines) == 1 {
-		return ""
-	}
-	return strings.Join(lines, "\n")
-}
-
-func (o *Orchestrator) generateWithOptionalFilesMessages(ctx context.Context, messages []llm.PromptMessage, fileIDs []string, appendFileIDText bool) (string, error) {
-	ids := nonEmptyIDs(fileIDs)
-	if len(ids) == 0 {
 	if client, ok := o.llm.(llm.MessageChatClient); ok {
 		return client.GenerateMessages(ctx, messages)
 	}
 	systemPrompt, userPrompt := fallbackPromptsFromMessages(messages)
 	return o.llm.Generate(ctx, systemPrompt, userPrompt)
-	}
-	if client, ok := o.llm.(llm.FileMessageChatClient); ok {
-		return client.GenerateMessagesWithFiles(ctx, messages, ids, appendFileIDText)
-	}
-	client, ok := o.llm.(llm.FileChatClient)
-	if !ok {
-		systemPrompt, userPrompt := fallbackPromptsFromMessages(messages)
-		return o.llm.Generate(ctx, systemPrompt, userPrompt)
-	}
-	systemPrompt, userPrompt := fallbackPromptsFromMessages(messages)
-	return client.GenerateWithFiles(ctx, systemPrompt, userPrompt, ids, appendFileIDText)
 }
 
-func buildReActMessages(systemPrompt, compressedContext, userInput, fileSummary, scratchpad string) []llm.PromptMessage {
+func buildReActMessages(systemPrompt, compressedContext, userInput, scratchpad string) []llm.PromptMessage {
 	msgs := make([]llm.PromptMessage, 0, 16)
 	msgs = append(msgs, llm.PromptMessage{Role: "system", Content: systemPrompt})
 	msgs = append(msgs, parseCompressedHistoryMessages(compressedContext)...)
 
-	if strings.TrimSpace(fileSummary) != "" {
-		msgs = append(msgs, llm.PromptMessage{Role: "assistant", Content: "文件上下文摘要:\n" + strings.TrimSpace(fileSummary)})
-	}
 	if strings.TrimSpace(scratchpad) != "" {
 		msgs = append(msgs, llm.PromptMessage{Role: "assistant", Content: "推理记录:\n" + strings.TrimSpace(scratchpad)})
 	}
@@ -735,20 +1083,6 @@ func fallbackPromptsFromMessages(messages []llm.PromptMessage) (string, string)
 	return strings.Join(sysParts, "\n\n"), strings.Join(userParts, "\n")
 }
 
-func (o *Orchestrator) buildFileUploadAck(ctx filePromptContext) string {
-	if len(ctx.FileIDs) == 0 {
-		return "文件已接收，但未拿到有效 file_id。请重新上传一次。"
-	}
-	lines := []string{
-		fmt.Sprintf("文件上传完成，已缓存 %d 个 file_id。", len(ctx.FileIDs)),
-		"请继续发送你的问题，我会结合这些文件内容和历史对话一起回答。",
-	}
-	if strings.TrimSpace(ctx.Summary) != "" {
-		lines = append(lines, "", ctx.Summary)
-	}
-	return strings.Join(lines, "\n")
-}
-
 func nonEmptyIDs(ids []string) []string {
 	if len(ids) == 0 {
 		return nil
@@ -769,20 +1103,6 @@ func nonEmptyIDs(ids []string) []string {
 	return out
 }
 
-func (c filePromptContext) toPendingRefs() []pendingFileRef {
-	if len(c.Uploaded) > 0 {
-		copied := make([]pendingFileRef, len(c.Uploaded))
-		copy(copied, c.Uploaded)
-		return sanitizePendingRefs(copied)
-	}
-	ids := nonEmptyIDs(c.FileIDs)
-	out := make([]pendingFileRef, 0, len(ids))
-	for _, id := range ids {
-		out = append(out, pendingFileRef{ID: id})
-	}
-	return out
-}
-
 func (o *Orchestrator) appendPendingFiles(chatID, userID string, refs []pendingFileRef) {
 	refs = sanitizePendingRefs(refs)
 	if len(refs) == 0 {
@@ -890,6 +1210,9 @@ func (o *Orchestrator) selectRelevantSkills(userInput string, maxCount int) []kn
 			continue
 		}
 		ranked = append(ranked, item{skill: sk, score: score})
+		if o.log != nil {
+			o.log.Debugf("selectRelevantSkills skill=%q score=%d", sk.Name, score)
+		}
 	}
 
 	if len(ranked) == 0 {
@@ -910,6 +1233,13 @@ func (o *Orchestrator) selectRelevantSkills(userInput string, maxCount int) []kn
 	for _, r := range ranked {
 		out = append(out, r.skill)
 	}
+	if o.log != nil {
+		selectedNames := make([]string, 0, len(out))
+		for _, sk := range out {
+			selectedNames = append(selectedNames, sk.Name)
+		}
+		o.log.Debugf("selectRelevantSkills query=%q matched=%d selected=%v", query, len(ranked), selectedNames)
+	}
 	return out
 }
 
@@ -1181,3 +1511,10 @@ func (o *Orchestrator) formatToolDoc() string {
 	}
 	return strings.TrimSpace(b.String())
 }
+
+func truncateForLog(s string, maxLen int) string {
+	if len(s) <= maxLen {
+		return s
+	}
+	return s[:maxLen] + "...(truncated)"
+}

internal/agent/orchestrator_skill_selection_test.go

@@ -46,3 +46,50 @@ func TestFormatRuntimeContextForPromptIncludesGOOS(t *testing.T) {
 		t.Fatalf("expected runtime context contains GOOS=%s, got: %s", runtime.GOOS, doc)
 	}
 }
+
+func TestMatchSkillsByNameExact(t *testing.T) {
+	all := []knowledge.Skill{
+		{Name: "SAFe PI Planning", Content: "PI规划技能"},
+		{Name: "文件系统查询专家", Content: "文件查询"},
+		{Name: "代码生成", Content: "代码生成技能"},
+	}
+	matched := matchSkillsByName(all, []string{"SAFe PI Planning"})
+	if len(matched) != 1 {
+		t.Fatalf("expected 1 match, got %d", len(matched))
+	}
+	if matched[0].Name != "SAFe PI Planning" {
+		t.Fatalf("expected SAFe PI Planning, got %s", matched[0].Name)
+	}
+}
+
+func TestMatchSkillsByNameFuzzy(t *testing.T) {
+	all := []knowledge.Skill{
+		{Name: "SAFe PI Planning", Content: "PI规划技能"},
+		{Name: "文件系统查询专家", Content: "文件查询"},
+	}
+	matched := matchSkillsByName(all, []string{"pi planning", "文件"})
+	if len(matched) != 2 {
+		t.Fatalf("expected 2 matches, got %d", len(matched))
+	}
+}
+
+func TestMatchSkillsByNameNoMatch(t *testing.T) {
+	all := []knowledge.Skill{
+		{Name: "文件系统查询专家", Content: "文件查询"},
+	}
+	matched := matchSkillsByName(all, []string{"不存在的技能"})
+	if len(matched) != 0 {
+		t.Fatalf("expected 0 matches, got %d", len(matched))
+	}
+}
+
+func TestMatchSkillsByNameEmpty(t *testing.T) {
+	matched := matchSkillsByName(nil, []string{"any"})
+	if len(matched) != 0 {
+		t.Fatalf("expected 0 matches, got %d", len(matched))
+	}
+	matched = matchSkillsByName([]knowledge.Skill{{Name: "test"}}, nil)
+	if len(matched) != 0 {
+		t.Fatalf("expected 0 matches, got %d", len(matched))
+	}
+}

internal/config/config.go

@@ -29,6 +29,7 @@ type Config struct {
 	LLM       LLMConfig
 	Security  SecurityConfig
 	WebSearch WebSearchConfig
+	Gitea     GiteaConfig
 
 	SQLitePath string
 }
@@ -56,7 +57,7 @@ type LLMConfig struct {
 	APIKey      string
 	Model       string
 	FileModel   string
-	FilePromptMode string
+	RouterModel string // 轻量路由模型，用于技能意图路由；为空则仅用关键词匹配
 }
 
 type SecurityConfig struct {
@@ -70,6 +71,13 @@ type WebSearchConfig struct {
 	APIKey string
 }
 
+type GiteaConfig struct {
+	BaseURL string // Gitea 实例地址
+	Token   string // Personal Access Token
+	Owner   string // 仓库所有者
+	Repo    string // 仓库名称
+}
+
 func Load() (Config, error) {
 	agentWorkspaceDir := resolveAgentWorkspaceDir()
 	if err := preloadEnvFiles(); err != nil {
@@ -110,7 +118,7 @@ func Load() (Config, error) {
 			APIKey:      strings.TrimSpace(os.Getenv("LLM_API_KEY")),
 			Model:       defaultIfEmpty(os.Getenv("LLM_MODEL"), "gpt-4o-mini"),
 			FileModel:   defaultIfEmpty(os.Getenv("LLM_FILE_MODEL"), defaultIfEmpty(os.Getenv("LLM_MODEL"), "gpt-4o-mini")),
-			FilePromptMode: normalizeFilePromptMode(defaultIfEmpty(os.Getenv("LLM_FILE_PROMPT_MODE"), "user_content_file_parts")),
+			RouterModel: strings.TrimSpace(os.Getenv("LLM_ROUTER_MODEL")),
 		},
 		SQLitePath: defaultIfEmpty(os.Getenv("SQLITE_PATH"), filepath.Join(defaultDataDir, "laodingbot.db")),
 		WebSearch: WebSearchConfig{
@@ -122,6 +130,12 @@ func Load() (Config, error) {
 			AllowedCommands: splitCSV(defaultIfEmpty(os.Getenv("ALLOWED_COMMANDS"), "pwd,ls,cat,echo,grep,find,head,tail,go")),
 			WorkDir:         defaultIfEmpty(os.Getenv("WORK_DIR"), defaultWorkSubdir),
 		},
+		Gitea: GiteaConfig{
+			BaseURL: strings.TrimRight(strings.TrimSpace(os.Getenv("GITEA_BASE_URL")), "/"),
+			Token:   strings.TrimSpace(os.Getenv("GITEA_TOKEN")),
+			Owner:   strings.TrimSpace(os.Getenv("GITEA_OWNER")),
+			Repo:    strings.TrimSpace(os.Getenv("GITEA_REPO")),
+		},
 	}
 
 	cfg.MessageChannel = strings.ToLower(strings.TrimSpace(cfg.MessageChannel))
@@ -178,9 +192,6 @@ func Load() (Config, error) {
 	if cfg.LLM.APIKey == "" {
 		return Config{}, fmt.Errorf("LLM_API_KEY is required")
 	}
-	if cfg.LLM.FilePromptMode != "user_content_file_parts" && cfg.LLM.FilePromptMode != "system_fileid_uri" {
-		return Config{}, fmt.Errorf("LLM_FILE_PROMPT_MODE must be one of: user_content_file_parts, system_fileid_uri")
-	}
 
 	cfg.SoulPath = resolvePathInWorkspace(cfg.SoulPath, agentWorkspaceDir)
 	cfg.SkillsDir = resolvePathInWorkspace(cfg.SkillsDir, agentWorkspaceDir)
@@ -417,14 +428,3 @@ func splitCSV(raw string) []string {
 	}
 	return out
 }
-
-func normalizeFilePromptMode(v string) string {
-	v = strings.ToLower(strings.TrimSpace(v))
-	if v == "" {
-		return "user_content_file_parts"
-	}
-	if v == "system_fileid" || v == "system_fileid_url" || v == "system_fileid_uri" {
-		return "system_fileid_uri"
-	}
-	return v
-}

internal/llm/client.go

@@ -33,21 +33,13 @@ type MessageChatClient interface {
 	GenerateMessages(ctx context.Context, messages []PromptMessage) (string, error)
 }
 
-type FileChatClient interface {
-	GenerateWithFiles(ctx context.Context, systemPrompt, userPrompt string, fileIDs []string, appendFileIDText bool) (string, error)
-}
-
-type FileMessageChatClient interface {
-	GenerateMessagesWithFiles(ctx context.Context, messages []PromptMessage, fileIDs []string, appendFileIDText bool) (string, error)
-}
-
 type FileUploader interface {
 	UploadFile(ctx context.Context, file InputFile, purpose string) (string, error)
 }
 
 // ToolCallChatClient 支持原生 function calling 的 LLM 客户端接口。
 type ToolCallChatClient interface {
-	GenerateWithTools(ctx context.Context, messages []PromptMessage, tools []ToolDefinition, fileIDs []string, appendFileIDText bool) (*ChatCompletion, error)
+	GenerateWithTools(ctx context.Context, messages []PromptMessage, tools []ToolDefinition) (*ChatCompletion, error)
 }
 
 // ToolDefinition 描述一个可供 LLM 调用的工具函数定义。
@@ -91,8 +83,6 @@ type InputFile struct {
 type OpenAICompatibleClient struct {
 	client openai.Client
 	model  string
-	fileModel      string
-	filePromptMode string
 	log    *logger.Logger
 }
 
@@ -107,8 +97,6 @@ func NewOpenAICompatibleClient(cfg config.LLMConfig, log *logger.Logger) *OpenAI
 	return &OpenAICompatibleClient{
 		client: openai.NewClient(opts...),
 		model:  cfg.Model,
-		fileModel:      cfg.FileModel,
-		filePromptMode: cfg.FilePromptMode,
 		log:    log,
 	}
 }
@@ -118,38 +106,22 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, systemPrompt, use
 		{Role: "system", Content: systemPrompt},
 		{Role: "user", Content: userPrompt},
 	}
-	return c.generateWithMessagesInternal(ctx, messages, nil, false)
+	return c.generateWithMessagesInternal(ctx, messages)
-}
-
-func (c *OpenAICompatibleClient) GenerateWithFiles(ctx context.Context, systemPrompt, userPrompt string, fileIDs []string, appendFileIDText bool) (string, error) {
-	messages := []PromptMessage{
-		{Role: "system", Content: systemPrompt},
-		{Role: "user", Content: userPrompt},
-	}
-	return c.generateWithMessagesInternal(ctx, messages, fileIDs, appendFileIDText)
 }
 
 func (c *OpenAICompatibleClient) GenerateMessages(ctx context.Context, messages []PromptMessage) (string, error) {
-	return c.generateWithMessagesInternal(ctx, messages, nil, false)
+	return c.generateWithMessagesInternal(ctx, messages)
-}
-
-func (c *OpenAICompatibleClient) GenerateMessagesWithFiles(ctx context.Context, messages []PromptMessage, fileIDs []string, appendFileIDText bool) (string, error) {
-	return c.generateWithMessagesInternal(ctx, messages, fileIDs, appendFileIDText)
 }
 
 // GenerateWithTools 使用原生 function calling 发送请求，返回结构化的 ChatCompletion。
-func (c *OpenAICompatibleClient) GenerateWithTools(ctx context.Context, messages []PromptMessage, tools []ToolDefinition, fileIDs []string, appendFileIDText bool) (*ChatCompletion, error) {
+func (c *OpenAICompatibleClient) GenerateWithTools(ctx context.Context, messages []PromptMessage, tools []ToolDefinition) (*ChatCompletion, error) {
 	model := c.model
-	ids := nonEmptyIDs(fileIDs)
-	if len(ids) > 0 && strings.TrimSpace(c.fileModel) != "" {
-		model = c.fileModel
-	}
 
-	sdkMessages := buildSDKMessages(messages, ids, c.normalizedFilePromptMode(), appendFileIDText)
+	sdkMessages := buildSDKMessages(messages)
 	sdkTools := toSDKTools(tools)
 
 	if c.log != nil {
-		c.log.Debugf("llm tool-call request start model=%s messages=%d tools=%d files=%d", model, len(sdkMessages), len(sdkTools), len(ids))
+		c.log.Debugf("llm tool-call request start model=%s messages=%d tools=%d", model, len(sdkMessages), len(sdkTools))
 	}
 
 	params := openai.ChatCompletionNewParams{
@@ -188,12 +160,8 @@ func (c *OpenAICompatibleClient) GenerateWithTools(ctx context.Context, messages
 	}, nil
 }
 
-func (c *OpenAICompatibleClient) generateWithMessagesInternal(ctx context.Context, messages []PromptMessage, fileIDs []string, appendFileIDText bool) (string, error) {
+func (c *OpenAICompatibleClient) generateWithMessagesInternal(ctx context.Context, messages []PromptMessage) (string, error) {
 	model := c.model
-	ids := nonEmptyIDs(fileIDs)
-	if len(ids) > 0 && strings.TrimSpace(c.fileModel) != "" {
-		model = c.fileModel
-	}
 
 	baseMessages := normalizePromptMessages(messages)
 	if len(baseMessages) == 0 {
@@ -202,10 +170,10 @@ func (c *OpenAICompatibleClient) generateWithMessagesInternal(ctx context.Contex
 
 	systemLen, userLen := promptMessageLengths(baseMessages)
 	if c.log != nil {
-		c.log.Debugf("llm request start model=%s system_len=%d user_len=%d file_count=%d file_prompt_mode=%s", model, systemLen, userLen, len(ids), c.normalizedFilePromptMode())
+		c.log.Debugf("llm request start model=%s system_len=%d user_len=%d", model, systemLen, userLen)
 	}
 
-	sdkMessages := buildSDKMessages(baseMessages, ids, c.normalizedFilePromptMode(), appendFileIDText)
+	sdkMessages := buildSDKMessages(baseMessages)
 
 	params := openai.ChatCompletionNewParams{
 		Model:    shared.ChatModel(model),
@@ -234,10 +202,9 @@ func (c *OpenAICompatibleClient) generateWithMessagesInternal(ctx context.Contex
 	return content, nil
 }
 
-// buildSDKMessages 将 PromptMessage 列表转换为 openai SDK 的消息格式，并注入 file_id（如需要）。
+// buildSDKMessages 将 PromptMessage 列表转换为 openai SDK 的消息格式。
-func buildSDKMessages(base []PromptMessage, fileIDs []string, mode string, appendFileIDText bool) []openai.ChatCompletionMessageParamUnion {
+func buildSDKMessages(base []PromptMessage) []openai.ChatCompletionMessageParamUnion {
-	mode = strings.ToLower(strings.TrimSpace(mode))
+	out := make([]openai.ChatCompletionMessageParamUnion, 0, len(base))
-	out := make([]openai.ChatCompletionMessageParamUnion, 0, len(base)+2)
 
 	for _, m := range base {
 		role := normalizeRole(m.Role)
@@ -247,34 +214,6 @@ func buildSDKMessages(base []PromptMessage, fileIDs []string, mode string, appen
 		out = append(out, toSDKMessage(m, role))
 	}
 
-	if len(fileIDs) == 0 {
-		return out
-	}
-
-	if appendFileIDText {
-		// WebUI 场景：将首个 fileID 作为 text part 追加到最后一个 user 消息。
-		firstFileID := strings.TrimSpace(fileIDs[0])
-		if firstFileID == "" {
-			return out
-		}
-		for i := len(out) - 1; i >= 0; i-- {
-			if r := out[i].GetRole(); r != nil && *r == "user" {
-				out[i] = buildUserMessageWithFileIDText(out[i], firstFileID)
-				return out
-			}
-		}
-		out = append(out, buildUserMessageWithFileIDText(openai.UserMessage(""), firstFileID))
-		return out
-	}
-
-	// 非 WebUI 场景：保持原有 file content part 方式。
-	for i := len(out) - 1; i >= 0; i-- {
-		if r := out[i].GetRole(); r != nil && *r == "user" {
-			out[i] = buildUserMessageWithFiles(out[i], fileIDs)
-			return out
-		}
-	}
-	out = append(out, buildUserMessageWithFiles(openai.UserMessage(""), fileIDs))
 	return out
 }
 
@@ -309,53 +248,6 @@ func toSDKMessage(m PromptMessage, role string) openai.ChatCompletionMessagePara
 	}
 }
 
-// buildUserMessageWithFileIDText 为 user 消息追加一个 text part，内容为 fileID。
-func buildUserMessageWithFileIDText(msg openai.ChatCompletionMessageParamUnion, fileID string) openai.ChatCompletionMessageParamUnion {
-	// 提取已有的文本内容
-	text := ""
-	if s, ok := msg.GetContent().AsAny().(*string); ok && s != nil {
-		text = *s
-	}
-	fileID = strings.TrimSpace(fileID)
-	if fileID == "" {
-		return msg
-	}
-
-	parts := make([]openai.ChatCompletionContentPartUnionParam, 0, 2)
-	if strings.TrimSpace(text) != "" {
-		parts = append(parts, openai.TextContentPart(text))
-	}
-	parts = append(parts, openai.TextContentPart(fileID))
-	if len(parts) == 0 {
-		return msg
-	}
-	return openai.UserMessage(parts)
-}
-
-// buildUserMessageWithFiles 为 user 消息追加 file content parts。
-func buildUserMessageWithFiles(msg openai.ChatCompletionMessageParamUnion, fileIDs []string) openai.ChatCompletionMessageParamUnion {
-	text := ""
-	if s, ok := msg.GetContent().AsAny().(*string); ok && s != nil {
-		text = *s
-	}
-
-	parts := make([]openai.ChatCompletionContentPartUnionParam, 0, len(fileIDs)+1)
-	if strings.TrimSpace(text) != "" {
-		parts = append(parts, openai.TextContentPart(text))
-	}
-	for _, id := range fileIDs {
-		id = strings.TrimSpace(id)
-		if id == "" {
-			continue
-		}
-		parts = append(parts, openai.FileContentPart(openai.ChatCompletionContentPartFileFileParam{FileID: param.NewOpt(id)}))
-	}
-	if len(parts) == 0 {
-		return msg
-	}
-	return openai.UserMessage(parts)
-}
-
 // toSDKTools 将内部 ToolDefinition 列表转换为 openai SDK 的 ChatCompletionToolParam 列表。
 func toSDKTools(tools []ToolDefinition) []openai.ChatCompletionToolParam {
 	if len(tools) == 0 {
@@ -397,6 +289,46 @@ func fromSDKToolCalls(sdkCalls []openai.ChatCompletionMessageToolCall) []ToolCal
 	return out
 }
 
+func normalizePromptMessages(messages []PromptMessage) []PromptMessage {
+	out := make([]PromptMessage, 0, len(messages))
+	for _, m := range messages {
+		role := normalizeRole(m.Role)
+		if role == "" {
+			continue
+		}
+		out = append(out, PromptMessage{
+			Role:       role,
+			Content:    m.Content,
+			ToolCalls:  m.ToolCalls,
+			ToolCallID: m.ToolCallID,
+			Name:       m.Name,
+		})
+	}
+	return out
+}
+
+func normalizeRole(role string) string {
+	r := strings.ToLower(strings.TrimSpace(role))
+	if r != "system" && r != "user" && r != "assistant" && r != "tool" {
+		return ""
+	}
+	return r
+}
+
+func promptMessageLengths(messages []PromptMessage) (int, int) {
+	systemLen := 0
+	userLen := 0
+	for _, m := range messages {
+		switch normalizeRole(m.Role) {
+		case "system":
+			systemLen += len(m.Content)
+		case "user":
+			userLen += len(m.Content)
+		}
+	}
+	return systemLen, userLen
+}
+
 func (c *OpenAICompatibleClient) UploadFile(ctx context.Context, file InputFile, purpose string) (string, error) {
 	if strings.TrimSpace(file.FileName) == "" {
 		return "", fmt.Errorf("empty file name")
@@ -460,71 +392,3 @@ func appendIfMissing(items []string, value string) []string {
 	}
 	return append(items, value)
 }
-
-func nonEmptyIDs(ids []string) []string {
-	if len(ids) == 0 {
-		return nil
-	}
-	out := make([]string, 0, len(ids))
-	seen := map[string]struct{}{}
-	for _, id := range ids {
-		id = strings.TrimSpace(id)
-		if id == "" {
-			continue
-		}
-		if _, ok := seen[id]; ok {
-			continue
-		}
-		seen[id] = struct{}{}
-		out = append(out, id)
-	}
-	return out
-}
-
-func normalizePromptMessages(messages []PromptMessage) []PromptMessage {
-	out := make([]PromptMessage, 0, len(messages))
-	for _, m := range messages {
-		role := normalizeRole(m.Role)
-		if role == "" {
-			continue
-		}
-		out = append(out, PromptMessage{
-			Role:       role,
-			Content:    m.Content,
-			ToolCalls:  m.ToolCalls,
-			ToolCallID: m.ToolCallID,
-			Name:       m.Name,
-		})
-	}
-	return out
-}
-
-func normalizeRole(role string) string {
-	r := strings.ToLower(strings.TrimSpace(role))
-	if r != "system" && r != "user" && r != "assistant" && r != "tool" {
-		return ""
-	}
-	return r
-}
-
-func promptMessageLengths(messages []PromptMessage) (int, int) {
-	systemLen := 0
-	userLen := 0
-	for _, m := range messages {
-		switch normalizeRole(m.Role) {
-		case "system":
-			systemLen += len(m.Content)
-		case "user":
-			userLen += len(m.Content)
-		}
-	}
-	return systemLen, userLen
-}
-
-func (c *OpenAICompatibleClient) normalizedFilePromptMode() string {
-	mode := strings.ToLower(strings.TrimSpace(c.filePromptMode))
-	if mode == "system_fileid" || mode == "system_fileid_url" || mode == "system_fileid_uri" {
-		return "system_fileid_uri"
-	}
-	return "user_content_file_parts"
-}

internal/toolhost/runtime.go

@@ -8,8 +8,11 @@ import (
 	"laodingbot/internal/config"
 	"laodingbot/internal/logger"
 	"laodingbot/internal/tools"
+	"laodingbot/tools/filedoc"
 	"laodingbot/tools/fileoperation"
 	"laodingbot/tools/git"
+	"laodingbot/tools/giteaticket"
+	"laodingbot/tools/piplan"
 	"laodingbot/tools/shell"
 	"laodingbot/tools/websearch"
 )
@@ -20,6 +23,9 @@ func RunChild(ctx context.Context, cfg config.Config, log *logger.Logger) error
 	var gitLog *logger.Logger
 	var shellLog *logger.Logger
 	var searchLog *logger.Logger
+	var fileDocLog *logger.Logger
+	var piPlanLog *logger.Logger
+	var giteaTicketLog *logger.Logger
 	var serverLog *logger.Logger
 	if log != nil {
 		log.Infof("toolhost child starting")
@@ -28,6 +34,9 @@ func RunChild(ctx context.Context, cfg config.Config, log *logger.Logger) error
 		gitLog = log.WithComponent("toolhost.git")
 		shellLog = log.WithComponent("toolhost.shell")
 		searchLog = log.WithComponent("toolhost.websearch")
+		fileDocLog = log.WithComponent("toolhost.filedoc")
+		piPlanLog = log.WithComponent("toolhost.piplan")
+		giteaTicketLog = log.WithComponent("toolhost.giteaticket")
 		serverLog = log.WithComponent("toolhost.server")
 	}
 	registry := tools.NewRegistry(registryLog)
@@ -53,6 +62,27 @@ func RunChild(ctx context.Context, cfg config.Config, log *logger.Logger) error
 		cfg.ToolOutputMaxChars,
 		searchLog,
 	))
+	registry.Register(filedoc.New(
+		filedoc.Config{
+			APIKey:  cfg.LLM.APIKey,
+			BaseURL: cfg.LLM.BaseURL,
+			Model:   cfg.LLM.FileModel,
+			Timeout: time.Duration(cfg.ToolCallTimeoutSec) * time.Second,
+		},
+		cfg.ToolOutputMaxChars,
+		fileDocLog,
+	))
+	registry.Register(piplan.New(cfg.ToolOutputMaxChars, piPlanLog))
+	registry.Register(giteaticket.New(
+		giteaticket.Config{
+			BaseURL: cfg.Gitea.BaseURL,
+			Token:   cfg.Gitea.Token,
+			Owner:   cfg.Gitea.Owner,
+			Repo:    cfg.Gitea.Repo,
+			Timeout: time.Duration(cfg.ToolCallTimeoutSec) * time.Second,
+		},
+		giteaTicketLog,
+	))
 
 	server := NewServer(registry, serverLog)
 	if err := server.Serve(ctx, stdin(), stdout()); err != nil && ctx.Err() == nil {

internal/transport/webui/bot.go

@@ -21,10 +21,30 @@ type IncomingMessage struct {
 	ChatID string
 	UserID string
 	Text   string
-	FileIDs []string
+}
+
+// StreamEventType 定义流式输出的事件类型
+type StreamEventType string
+
+const (
+	StreamEventTypeThought    StreamEventType = "thought"     // LLM 思考过程
+	StreamEventTypeToolCall   StreamEventType = "tool_call"   // 工具调用请求
+	StreamEventTypeToolResult StreamEventType = "tool_result" // 工具执行结果
+	StreamEventTypeFinal      StreamEventType = "final"       // 最终答案
+	StreamEventTypeError      StreamEventType = "error"       // 错误信息
+)
+
+// StreamEvent 代表流式输出中的一个事件
+type StreamEvent struct {
+	Type     StreamEventType `json:"type"`
+	Content  string          `json:"content"`
+	Step     int             `json:"step,omitempty"`
+	ToolName string          `json:"tool_name,omitempty"`
 }
 
 type ChatHandler func(context.Context, IncomingMessage) (string, error)
+type StreamChatHandler func(context.Context, IncomingMessage, StreamEventCallback) (string, error)
+type StreamEventCallback func(event StreamEvent) error
 type UploadHandler func(context.Context, string, string, []llm.InputFile) ([]string, error)
 
 type Bot struct {
@@ -33,6 +53,7 @@ type Bot struct {
 	log            *logger.Logger
 
 	chatHandler       ChatHandler
+	streamChatHandler StreamChatHandler
 	uploadHandler     UploadHandler
 	counter           uint64
 }
@@ -41,7 +62,6 @@ type chatRequest struct {
 	Text      string `json:"text"`
 	SessionID string `json:"session_id"`
 	UserID    string `json:"user_id"`
-	FileIDs   []string `json:"file_ids"`
 }
 
 func (r *chatRequest) UnmarshalJSON(data []byte) error {
@@ -51,10 +71,6 @@ func (r *chatRequest) UnmarshalJSON(data []byte) error {
 		SessionIDCamel string `json:"sessionId"`
 		UserID         string `json:"user_id"`
 		UserIDCamel    string `json:"userId"`
-		FileIDs        json.RawMessage `json:"file_ids"`
-		FileIDsCamel   json.RawMessage `json:"fileIds"`
-		FileIDsFlat    json.RawMessage `json:"fileids"`
-		FileID         json.RawMessage `json:"file_id"`
 	}
 
 	var raw rawChatRequest
@@ -65,13 +81,6 @@ func (r *chatRequest) UnmarshalJSON(data []byte) error {
 	r.Text = raw.Text
 	r.SessionID = firstNonEmpty(raw.SessionID, raw.SessionIDCamel)
 	r.UserID = firstNonEmpty(raw.UserID, raw.UserIDCamel)
-
-	rawIDs := firstNonEmptyRaw(raw.FileIDs, raw.FileIDsCamel, raw.FileIDsFlat, raw.FileID)
-	ids, err := decodeStringList(rawIDs)
-	if err != nil {
-		return err
-	}
-	r.FileIDs = ids
 	return nil
 }
 
@@ -109,7 +118,7 @@ func NewBot(cfg config.WebUIConfig, log *logger.Logger) (*Bot, error) {
 	}, nil
 }
 
-func (b *Bot) Run(ctx context.Context, chatHandler ChatHandler, uploadHandler UploadHandler) error {
+func (b *Bot) Run(ctx context.Context, chatHandler ChatHandler, streamChatHandler StreamChatHandler, uploadHandler UploadHandler) error {
 	if chatHandler == nil {
 		return fmt.Errorf("nil webui chat handler")
 	}
@@ -117,10 +126,12 @@ func (b *Bot) Run(ctx context.Context, chatHandler ChatHandler, uploadHandler Up
 		return fmt.Errorf("nil webui upload handler")
 	}
 	b.chatHandler = chatHandler
+	b.streamChatHandler = streamChatHandler
 	b.uploadHandler = uploadHandler
 
 	mux := http.NewServeMux()
 	mux.HandleFunc("/api/chat", b.handleChat)
+	mux.HandleFunc("/api/chat/stream", b.handleChatStream)
 	mux.HandleFunc("/api/upload", b.handleUpload)
 
 	srv := &http.Server{
@@ -194,7 +205,6 @@ func (b *Bot) handleChat(w http.ResponseWriter, r *http.Request) {
 		ChatID: sessionID,
 		UserID: userID,
 		Text:   req.Text,
-		FileIDs: req.FileIDs,
 	})
 	if err != nil {
 		if b.log != nil {
@@ -210,37 +220,8 @@ func (b *Bot) handleChat(w http.ResponseWriter, r *http.Request) {
 	})
 }
 
-func decodeStringList(raw json.RawMessage) ([]string, error) {
-	if len(raw) == 0 {
-		return nil, nil
-	}
-
-	var list []string
-	if err := json.Unmarshal(raw, &list); err == nil {
-		return nonEmptyIDs(list), nil
-	}
-
-	var single string
-	if err := json.Unmarshal(raw, &single); err == nil {
-		if strings.TrimSpace(single) == "" {
-			return nil, nil
-		}
-		return nonEmptyIDs(strings.Split(single, ",")), nil
-	}
-
-	return nil, fmt.Errorf("invalid file ids format")
-}
-
-func firstNonEmptyRaw(vals ...json.RawMessage) json.RawMessage {
-	for _, v := range vals {
-		if len(v) > 0 {
-			return v
-		}
-	}
-	return nil
-}
-
 func firstNonEmpty(vals ...string) string {
+
 	for _, v := range vals {
 		if strings.TrimSpace(v) != "" {
 			return v
@@ -249,24 +230,82 @@ func firstNonEmpty(vals ...string) string {
 	return ""
 }
 
-func nonEmptyIDs(ids []string) []string {
+func (b *Bot) handleChatStream(w http.ResponseWriter, r *http.Request) {
-	if len(ids) == 0 {
+	if r.Method != http.MethodPost {
+		writeJSON(w, http.StatusMethodNotAllowed, errorResponse{Error: "method not allowed"})
+		return
+	}
+	if !strings.Contains(strings.ToLower(strings.TrimSpace(r.Header.Get("Content-Type"))), "application/json") {
+		writeJSON(w, http.StatusBadRequest, errorResponse{Error: "content-type must be application/json"})
+		return
+	}
+	if b.streamChatHandler == nil {
+		writeJSON(w, http.StatusInternalServerError, errorResponse{Error: "stream chat handler not ready"})
+		return
+	}
+
+	var req chatRequest
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		writeJSON(w, http.StatusBadRequest, errorResponse{Error: "invalid json body"})
+		return
+	}
+	req.Text = strings.TrimSpace(req.Text)
+	if req.Text == "" {
+		writeJSON(w, http.StatusBadRequest, errorResponse{Error: "text is required"})
+		return
+	}
+	sessionID := b.resolveID(req.SessionID, "sess")
+	userID := b.resolveID(req.UserID, "user")
+
+	// 设置 SSE 响应头
+	w.Header().Set("Content-Type", "text/event-stream")
+	w.Header().Set("Cache-Control", "no-cache")
+	w.Header().Set("Connection", "keep-alive")
+	w.Header().Set("Access-Control-Allow-Origin", "*")
+	w.Header().Set("Access-Control-Allow-Headers", "Content-Type")
+	w.WriteHeader(http.StatusOK)
+
+	flusher, ok := w.(http.Flusher)
+	if !ok {
+		http.Error(w, "streaming not supported", http.StatusInternalServerError)
+		return
+	}
+
+	// 创建回调函数来推送 SSE 事件
+	callback := func(event StreamEvent) error {
+		data, err := json.Marshal(event)
+		if err != nil {
+			return err
+		}
+		fmt.Fprintf(w, "data: %s\n\n", string(data))
+		flusher.Flush()
 		return nil
 	}
-	out := make([]string, 0, len(ids))
+
-	seen := map[string]struct{}{}
+	// 调用流式处理器
-	for _, id := range ids {
+	reply, err := b.streamChatHandler(r.Context(), IncomingMessage{
-		id = strings.TrimSpace(id)
+		ChatID: sessionID,
-		if id == "" {
+		UserID: userID,
-			continue
+		Text:   req.Text,
+	}, callback)
+	if err != nil {
+		if b.log != nil {
+			b.log.Errorf("webui stream chat handler failed session_id=%s user_id=%s err=%v", sessionID, userID, err)
 		}
-		if _, ok := seen[id]; ok {
+		// 推送错误事件
-			continue
+		errEvent := StreamEvent{
+			Type:    StreamEventTypeError,
+			Content: "stream error: " + err.Error(),
 		}
-		seen[id] = struct{}{}
+		data, _ := json.Marshal(errEvent)
-		out = append(out, id)
+		fmt.Fprintf(w, "data: %s\n\n", string(data))
+		flusher.Flush()
+		return
+	}
+
+	if b.log != nil {
+		b.log.Infof("webui stream chat completed session_id=%s user_id=%s reply_len=%d", sessionID, userID, len(reply))
 	}
-	return out
 }
 
 func (b *Bot) handleUpload(w http.ResponseWriter, r *http.Request) {

internal/transport/webui/bot_test.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"mime/multipart"
 	"net/http"
 	"net/http/httptest"
@@ -51,66 +52,6 @@ func TestHandleChatSuccess(t *testing.T) {
 	}
 }
 
-func TestHandleChatWithFileIDs(t *testing.T) {
-	b := newTestBot(t, 1024*1024)
-	b.chatHandler = func(_ context.Context, msg IncomingMessage) (string, error) {
-		if msg.ChatID != "s1" || msg.UserID != "u1" || msg.Text != "hello" {
-			t.Fatalf("unexpected message: %+v", msg)
-		}
-		if len(msg.FileIDs) != 2 || msg.FileIDs[0] != "file_a" || msg.FileIDs[1] != "file_b" {
-			t.Fatalf("unexpected file ids: %+v", msg.FileIDs)
-		}
-		return "ok", nil
-	}
-
-	body := strings.NewReader(`{"text":"hello","session_id":"s1","user_id":"u1","file_ids":["file_a","file_b"]}`)
-	req := httptest.NewRequest(http.MethodPost, "/api/chat", body)
-	req.Header.Set("Content-Type", "application/json")
-	w := httptest.NewRecorder()
-
-	b.handleChat(w, req)
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
-	}
-}
-
-func TestHandleChatWithFileIDsAliases(t *testing.T) {
-	tests := []struct {
-		name string
-		body string
-	}{
-		{name: "camel array", body: `{"text":"hello","sessionId":"s1","userId":"u1","fileIds":["file_a","file_b"]}`},
-		{name: "flat array", body: `{"text":"hello","session_id":"s1","user_id":"u1","fileids":["file_a","file_b"]}`},
-		{name: "single key", body: `{"text":"hello","session_id":"s1","user_id":"u1","file_id":"file_a"}`},
-		{name: "csv string", body: `{"text":"hello","session_id":"s1","user_id":"u1","file_ids":"file_a, file_b"}`},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			b := newTestBot(t, 1024*1024)
-			b.chatHandler = func(_ context.Context, msg IncomingMessage) (string, error) {
-				if msg.ChatID != "s1" || msg.UserID != "u1" || msg.Text != "hello" {
-					t.Fatalf("unexpected message: %+v", msg)
-				}
-				if len(msg.FileIDs) == 0 {
-					t.Fatalf("expected file ids from alias payload, got empty")
-				}
-				return "ok", nil
-			}
-
-			body := strings.NewReader(tt.body)
-			req := httptest.NewRequest(http.MethodPost, "/api/chat", body)
-			req.Header.Set("Content-Type", "application/json")
-			w := httptest.NewRecorder()
-
-			b.handleChat(w, req)
-			if w.Code != http.StatusOK {
-				t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
-			}
-		})
-	}
-}
-
 func TestHandleChatMissingText(t *testing.T) {
 	b := newTestBot(t, 1024*1024)
 	b.chatHandler = func(_ context.Context, _ IncomingMessage) (string, error) { return "", nil }
@@ -215,3 +156,149 @@ func TestHandleUploadMissingFile(t *testing.T) {
 		t.Fatalf("expected 400, got %d", w.Code)
 	}
 }
+
+func TestHandleChatStreamSuccess(t *testing.T) {
+	b := newTestBot(t, 1024*1024)
+	b.streamChatHandler = func(_ context.Context, msg IncomingMessage, cb StreamEventCallback) (string, error) {
+		if msg.ChatID != "s1" || msg.UserID != "u1" || msg.Text != "hello" {
+			t.Fatalf("unexpected message: %+v", msg)
+		}
+		if err := cb(StreamEvent{Type: StreamEventTypeThought, Content: "thinking", Step: 1}); err != nil {
+			return "", err
+		}
+		if err := cb(StreamEvent{Type: StreamEventTypeToolCall, Content: "{\"input\":\"pwd\"}", Step: 1, ToolName: "shell"}); err != nil {
+			return "", err
+		}
+		if err := cb(StreamEvent{Type: StreamEventTypeToolResult, Content: "C:/Project", Step: 1, ToolName: "shell"}); err != nil {
+			return "", err
+		}
+		if err := cb(StreamEvent{Type: StreamEventTypeFinal, Content: "done", Step: 2}); err != nil {
+			return "", err
+		}
+		return "done", nil
+	}
+
+	body := strings.NewReader(`{"text":"hello","session_id":"s1","user_id":"u1"}`)
+	req := httptest.NewRequest(http.MethodPost, "/api/chat/stream", body)
+	req.Header.Set("Content-Type", "application/json")
+	w := httptest.NewRecorder()
+
+	b.handleChatStream(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
+	}
+	if got := w.Header().Get("Content-Type"); got != "text/event-stream" {
+		t.Fatalf("expected text/event-stream, got %q", got)
+	}
+
+	var events []StreamEvent
+	chunks := strings.Split(strings.TrimSpace(w.Body.String()), "\n\n")
+	for _, chunk := range chunks {
+		line := strings.TrimSpace(chunk)
+		if line == "" {
+			continue
+		}
+		if !strings.HasPrefix(line, "data: ") {
+			t.Fatalf("invalid sse line: %q", line)
+		}
+		payload := strings.TrimPrefix(line, "data: ")
+		var ev StreamEvent
+		if err := json.Unmarshal([]byte(payload), &ev); err != nil {
+			t.Fatalf("unmarshal stream event failed: %v payload=%s", err, payload)
+		}
+		events = append(events, ev)
+	}
+
+	if len(events) != 4 {
+		t.Fatalf("expected 4 events, got %d", len(events))
+	}
+	if events[0].Type != StreamEventTypeThought {
+		t.Fatalf("event[0] type mismatch: %s", events[0].Type)
+	}
+	if events[1].Type != StreamEventTypeToolCall || events[1].ToolName != "shell" {
+		t.Fatalf("event[1] mismatch: %+v", events[1])
+	}
+	if events[2].Type != StreamEventTypeToolResult || events[2].ToolName != "shell" {
+		t.Fatalf("event[2] mismatch: %+v", events[2])
+	}
+	if events[3].Type != StreamEventTypeFinal || events[3].Content != "done" {
+		t.Fatalf("event[3] mismatch: %+v", events[3])
+	}
+}
+
+func TestHandleChatStreamHandlerError(t *testing.T) {
+	b := newTestBot(t, 1024*1024)
+	b.streamChatHandler = func(_ context.Context, _ IncomingMessage, _ StreamEventCallback) (string, error) {
+		return "", errors.New("boom")
+	}
+
+	body := strings.NewReader(`{"text":"hello"}`)
+	req := httptest.NewRequest(http.MethodPost, "/api/chat/stream", body)
+	req.Header.Set("Content-Type", "application/json")
+	w := httptest.NewRecorder()
+
+	b.handleChatStream(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d", w.Code)
+	}
+	respBody := w.Body.String()
+	if !strings.Contains(respBody, `"type":"error"`) {
+		t.Fatalf("expected error event in stream, body=%q", respBody)
+	}
+	if !strings.Contains(respBody, "stream error: boom") {
+		t.Fatalf("expected error detail in stream, body=%q", respBody)
+	}
+}
+
+func TestHandleChatStreamValidation(t *testing.T) {
+	t.Run("method not allowed", func(t *testing.T) {
+		b := newTestBot(t, 1024*1024)
+		req := httptest.NewRequest(http.MethodGet, "/api/chat/stream", nil)
+		w := httptest.NewRecorder()
+
+		b.handleChatStream(w, req)
+		if w.Code != http.StatusMethodNotAllowed {
+			t.Fatalf("expected 405, got %d", w.Code)
+		}
+	})
+
+	t.Run("content type must be json", func(t *testing.T) {
+		b := newTestBot(t, 1024*1024)
+		req := httptest.NewRequest(http.MethodPost, "/api/chat/stream", strings.NewReader(`{"text":"hello"}`))
+		w := httptest.NewRecorder()
+
+		b.handleChatStream(w, req)
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected 400, got %d", w.Code)
+		}
+	})
+
+	t.Run("handler not ready", func(t *testing.T) {
+		b := newTestBot(t, 1024*1024)
+		req := httptest.NewRequest(http.MethodPost, "/api/chat/stream", strings.NewReader(`{"text":"hello"}`))
+		req.Header.Set("Content-Type", "application/json")
+		w := httptest.NewRecorder()
+
+		b.handleChatStream(w, req)
+		if w.Code != http.StatusInternalServerError {
+			t.Fatalf("expected 500, got %d", w.Code)
+		}
+	})
+
+	t.Run("text required", func(t *testing.T) {
+		b := newTestBot(t, 1024*1024)
+		b.streamChatHandler = func(_ context.Context, _ IncomingMessage, _ StreamEventCallback) (string, error) {
+			return "", nil
+		}
+		req := httptest.NewRequest(http.MethodPost, "/api/chat/stream", strings.NewReader(`{"text":"   "}`))
+		req.Header.Set("Content-Type", "application/json")
+		w := httptest.NewRecorder()
+
+		b.handleChatStream(w, req)
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected 400, got %d", w.Code)
+		}
+	})
+}

skills/safe_pi_planning/skill.md

@@ -0,0 +1,247 @@
+---
+name: SAFe 铁三角 PI 规划
+description: 扮演 SAFe 铁三角（PM、架构师、RTE），将宏观 Epic 拆解为 PI 规划，输出标准化架构蓝图，并在 Gitea 创建可执行工单。支持从 PDF 等文档提取需求。
+---
+
+# Skill: SAFe 铁三角 PI 规划
+
+## 1. 触发条件
+
+当用户的意图匹配以下任意一种时，**必须启用本技能**：
+
+- 提交了一份**宏观业务需求**（Epic），要求进行 PI 级别拆解
+- 要求进行**下一个 PI (Program Increment) 的规划**
+- 提到 SAFe、PI Planning、铁三角、Feature 拆解、架构跑道等关键词
+- 上传了**需求文档**（PDF、Word、Markdown）并要求分析和拆解
+- 要求将规划结果**同步到 Gitea / 创建工单**
+
+**不适用场景**：单纯的代码编写、Bug 修复、文件查询等操作性任务。
+
+## 2. 可用工具
+
+| 工具名 | 用途 | 阶段 |
+|--------|------|------|
+| `extract_file_document` | 提取用户上传的 PDF/文档内容 | 输入准备 |
+| `publish_pi_plan` | 将推演结果渲染为标准化架构蓝图 | 规划输出 |
+| `create_gitea_ticket` | 在 Gitea 中创建 User Story 工单 | 任务下发 |
+| `web_search` | 搜索技术方案、行业标准等参考信息 | 辅助决策 |
+
+## 3. 执行流程
+
+执行本技能时，严格遵循以下分阶段流程。每个阶段对应 ReAct 循环中的一个或多个 Thought → Action → Observation 步骤。
+
+### 阶段 0：输入准备 — 文档提取（如有附件）
+
+如果用户提交了文件（file_id），**必须先提取文档内容**，再进入正式推演。
+
+**Action**: 调用 `extract_file_document`
+```
+输入: 用户提供的 file_id
+```
+**Observation**: 获取文档全文结构化摘要，包含标题、核心观点、关键数据。
+
+将提取到的文档内容作为后续推演的输入素材。如果用户同时提交了多个文件，逐个提取后合并为统一的需求上下文。
+
+---
+
+### 阶段 1：铁三角推演（核心 Thought 过程）
+
+**这是本技能的核心。** 你必须在 Thought 中同时扮演三个角色，按以下顺序进行严格的自我推演和博弈。不可跳过任何视角。
+
+#### 视角 1: 产品管理 (PM) — 决定"做什么"和"为什么做"
+
+在 Thought 中以 `[PM]` 标记此视角的思考：
+
+1. **需求拆解**: 将宏观 Epic 拆解为 2-4 个可在 8-12 周内交付的业务特性 (Feature)
+2. **价值假设**: 为每个 Feature 写出清晰的 Benefit Hypothesis — 如果做了这个功能，会带来什么可量化的业务收益
+3. **验收标准**: 从业务视角列出每个 Feature 的验收条件 (Acceptance Criteria)，不涉及技术实现细节
+4. **优先级排序**: 根据业务价值和紧迫性，给出 Feature 的建议交付顺序
+
+**PM 视角的自检问题**：
+- 每个 Feature 是否独立可交付，还是必须和其他 Feature 一起才有意义？
+- 验收标准是否可以被测试团队直接转化为测试用例？
+- 是否遗漏了用户（终端使用者）会关心的场景？
+
+#### 视角 2: 系统架构师 (SA) — 决定"技术怎么接"和"底座跑道"
+
+在 Thought 中以 `[SA]` 标记此视角的思考：
+
+1. **架构跑道 (Architectural Runway)**: PM 提出的每个 Feature，当前系统能直接支撑吗？如果不能，需要提前铺设哪些底层基础设施？将这些转化为 Enabler
+2. **瓶颈识别**: 当前系统架构的薄弱环节在哪？哪些 Enabler 是阻塞性的（不做就无法开始 Feature 开发）？
+3. **NFRs 定义**: 强制规定非功能性需求：
+   - 性能指标（QPS、P99 延迟、吞吐量）
+   - 安全标准（加密协议、认证机制、数据脱敏）
+4. **接口契约**: 如果涉及多个子系统交互，定义关键 API 契约
+
+**SA 视角的自检问题**：
+- 每个 Enabler 是否有明确的"完成标志"（而不是开放性的研究任务）？
+- NFRs 的指标是否可量化、可自动化测试？
+- 是否过度设计？Enabler 数量是否与 Feature 复杂度匹配？
+
+#### 视角 3: 发布火车工程师 (RTE) — 决定"流程风险"与"依赖管理"
+
+在 Thought 中以 `[RTE]` 标记此视角的思考：
+
+1. **依赖分析**: 检查 PM 的 Feature 和 SA 的 Enabler 之间的时序依赖 — 哪些 Enabler 必须先完成才能开始哪些 Feature？
+2. **风险评估**: 识别潜在的交付风险：
+   - 技术风险（新技术栈、外部 API 未就绪）
+   - 资源风险（关键人员依赖、跨团队协调）
+   - 集成风险（多个 Feature 的集成点）
+3. **里程碑建议**: 基于依赖关系，给出关键里程碑的建议时间节点
+
+**RTE 视角的自检问题**：
+- 是否存在循环依赖？
+- 关键路径上的任务是否有 buffer？
+- 如果某个 Enabler 延期，哪些 Feature 会受影响？
+
+---
+
+### 阶段 2：生成标准化 PI 规划
+
+完成三个视角的推演后，**必须立即调用** `publish_pi_plan` 工具。
+
+**Action**: 调用 `publish_pi_plan`，传入推演结果的 JSON：
+
+```json
+{
+  "pi_vision": "本 PI 的核心业务愿景（一两句话概括）",
+  "features": [
+    {
+      "feature_id": "FEAT_XXX_001",
+      "title": "动宾结构的简洁标题",
+      "benefit_hypothesis": "如果实现此功能，将带来 XXX 业务收益",
+      "acceptance_criteria": ["AC1: ...", "AC2: ..."]
+    }
+  ],
+  "enablers": [
+    {
+      "enabler_id": "ENAB_XXX_001",
+      "title": "技术任务名称",
+      "architectural_purpose": "为什么需要这个底层改造"
+    }
+  ],
+  "nfrs": {
+    "performance": "具体的性能指标约束",
+    "security": "具体的安全与合规约束"
+  },
+  "dependencies": [
+    {
+      "source_id": "ENAB_XXX_001",
+      "target_id": "FEAT_XXX_001",
+      "reason": "依赖原因的一句话描述"
+    }
+  ]
+}
+```
+
+**命名规范**：
+- Feature ID: `FEAT_<领域缩写>_<序号>`，如 `FEAT_OTA_001`
+- Enabler ID: `ENAB_<技术栈缩写>_<序号>`，如 `ENAB_KAFKA_001`
+
+**Observation**: 获取渲染后的 Markdown 架构蓝图，包含愿景、特性清单、Enabler 表、NFRs、依赖关系、执行顺序和质量门禁检查清单。
+
+将此蓝图**完整展示给用户**，征求反馈。
+
+---
+
+### 阶段 3：任务下发到 Gitea（用户确认后）
+
+当用户确认规划方案后，将 Feature 和 Enabler **逐一拆解为 User Story**，通过 `create_gitea_ticket` 在 Gitea 创建工单。
+
+#### 拆解原则
+
+- 每个 Feature 拆解为 1-3 个 User Story（每个 Story 应在 1-3 天内可完成）
+- 每个 Enabler 拆解为 1-2 个技术任务 Story
+- Story 的标题采用**动宾结构**（如"实现固件版本解析 API"，"部署 Kafka 基础镜像"）
+
+#### 创建工单
+
+对每个 Story，**Action**: 调用 `create_gitea_ticket`：
+
+```json
+{
+  "title": "动宾结构的任务标题",
+  "body": "## 溯源\n- Parent: FEAT_XXX_001\n\n## 任务上下文\n<描述做什么、为什么>\n\n## 验收标准\n- [ ] AC-1: ...\n- [ ] AC-2: ...\n\n## NFRs\n- 性能: ...\n- 安全: ...\n\n## 技术实现思路\n<基于架构师视角补充>",
+  "labels": ["type/story", "domain/<领域>", "priority/<high|medium|low>", "status/todo"],
+  "parent_reference_id": "FEAT_XXX_001",
+  "estimated_hours": 8
+}
+```
+
+**标签约定**：
+- `type/story` | `type/enabler` | `type/spike` — 任务类型
+- `domain/<领域>` — 业务领域（如 `domain/cloud`、`domain/vehicle`、`domain/infra`）
+- `priority/high` | `priority/medium` | `priority/low` — 优先级
+- `status/todo` — 初始状态
+
+**工时估算参考**：
+- 简单 CRUD / 配置变更: 4-8 小时
+- 标准 API 开发 + 单元测试: 8-16 小时
+- 基础设施搭建 / 中间件部署: 16-24 小时
+- 跨系统集成 + 联调: 16-32 小时
+
+创建完成后，汇总所有工单链接，输出执行看板。
+
+## 4. Thought 内部推演示范
+
+以下示范 Thought 过程的结构（非固定内容，仅展示格式）：
+
+```
+Thought:
+用户提交了一个 Epic："实现面向欧洲市场的车云 OTA 升级平台"。
+我需要以铁三角三个视角依次推演。
+
+[PM] 产品管理视角：
+- 核心交付价值：让欧洲车主能安全、快速地接收 OTA 升级
+- Feature 拆解：
+  1. FEAT_OTA_001: 云端固件版本依赖检查 — 防止下载不兼容固件
+  2. FEAT_OTA_002: 端侧断点续传 — 弱网环境下提升下载成功率
+  3. FEAT_OTA_003: 升级状态实时推送 — 车主可在 App 中看到升级进度
+- 各 Feature 的 AC 已确定，均可测试化
+
+[SA] 系统架构师视角：
+- FEAT_OTA_001 需要异步版本校验，当前没有合适的消息中间件 → ENAB_KAFKA_001
+- FEAT_OTA_002 需要支持 Range 请求的对象存储 → ENAB_S3_001
+- NFRs：API P99 < 200ms, TLS 1.3, VIN 脱敏
+- 当前架构可支撑 FEAT_OTA_003，无需新增 Enabler
+
+[RTE] 发布火车工程师视角：
+- ENAB_KAFKA_001 → FEAT_OTA_001（阻塞依赖）
+- ENAB_S3_001 → FEAT_OTA_002（阻塞依赖）
+- FEAT_OTA_003 无依赖，可并行开发
+- 风险：Kafka 集群搭建预计 2 周，是关键路径
+
+推演完成，下一步调用 publish_pi_plan 输出标准化蓝图。
+
+Action: publish_pi_plan
+Action Input: {"pi_vision": "...", "features": [...], ...}
+```
+
+## 5. 输出规范
+
+最终交付给用户的内容必须包含：
+
+1. **PI 蓝图**：`publish_pi_plan` 生成的完整 Markdown 报告
+2. **Gitea 工单汇总**（如已执行阶段 3）：
+   - 工单编号与链接列表
+   - 按执行顺序排列
+   - 标注关键路径上的任务
+3. **风险与建议**：RTE 视角识别的关键风险及缓解措施
+
+## 6. 质量约束
+
+- **三视角不可跳过**: 即使需求看起来简单，也必须走完 PM → SA → RTE 三步推演
+- **ID 必须唯一**: Feature ID 和 Enabler ID 在同一个 PI 内不可重复
+- **AC 必须可测试**: 每条验收标准都应能被转化为自动化测试用例
+- **NFRs 必须可量化**: 不接受"性能要好"这样的模糊描述，必须有具体数字
+- **依赖必须有理由**: 每条依赖关系都要说明为什么 A 必须先于 B
+
+## 7. 失败回退策略
+
+| 失败场景 | 处理方式 |
+|----------|----------|
+| 文档提取失败（file_id 无效） | 提示用户重新上传或提供文本版需求描述 |
+| 需求信息不足以拆解 Feature | 列出缺失信息清单，请求用户补充后重新推演 |
+| Gitea 配置缺失（token/repo 为空） | 仅输出 PI 蓝图，跳过工单创建，提示用户配置 Gitea 环境变量 |
+| Gitea API 调用失败 | 输出 PI 蓝图，附上拟创建工单的 JSON 列表供用户手动创建 |
+| 用户对规划有异议 | 记录反馈，调整对应视角的推演，重新生成蓝图 |

tools/filedoc/filedoc.go

@@ -0,0 +1,208 @@
+package filedoc
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+	"time"
+
+	"laodingbot/internal/logger"
+
+	openai "github.com/openai/openai-go"
+	"github.com/openai/openai-go/option"
+	"github.com/openai/openai-go/shared"
+)
+
+type Config struct {
+	APIKey  string
+	BaseURL string
+	Model   string
+	Timeout time.Duration
+}
+
+type Tool struct {
+	client         openai.Client
+	model          string
+	maxOutputChars int
+	log            *logger.Logger
+}
+
+func New(cfg Config, maxOutputChars int, log *logger.Logger) *Tool {
+	if strings.TrimSpace(cfg.Model) == "" {
+		cfg.Model = "gpt-4o-mini"
+	}
+	if cfg.Timeout <= 0 {
+		cfg.Timeout = 60 * time.Second
+	}
+	if maxOutputChars <= 0 {
+		maxOutputChars = 12000
+	}
+
+	opts := []option.RequestOption{
+		option.WithAPIKey(strings.TrimSpace(cfg.APIKey)),
+		option.WithRequestTimeout(cfg.Timeout),
+	}
+	if strings.TrimSpace(cfg.BaseURL) != "" {
+		opts = append(opts, option.WithBaseURL(strings.TrimSpace(cfg.BaseURL)))
+	}
+
+	return &Tool{
+		client:         openai.NewClient(opts...),
+		model:          cfg.Model,
+		maxOutputChars: maxOutputChars,
+		log:            log,
+	}
+}
+
+func (t *Tool) Name() string { return "extract_file_document" }
+
+func (t *Tool) Description() string {
+	return "Extract full document details from a file ID via OpenAI. Input: file_id (supports plain ID, fileid://ID, or JSON {\"file_id\":\"...\"})."
+}
+
+func (t *Tool) Call(ctx context.Context, input string) (string, error) {
+	fileID, userFocus, err := parseInput(input)
+	if err != nil {
+		return "", err
+	}
+
+	prompt := buildExtractionPrompt(fileID, userFocus)
+	messages := []openai.ChatCompletionMessageParamUnion{
+		openai.SystemMessage("fileid://" + fileID),
+		openai.UserMessage([]openai.ChatCompletionContentPartUnionParam{
+			openai.TextContentPart(prompt),
+		}),
+	}
+
+	params := openai.ChatCompletionNewParams{
+		Model:    shared.ChatModel(t.model),
+		Messages: messages,
+	}
+
+	if t.log != nil {
+		t.log.Infof("filedoc tool request model=%s file_id=%s", t.model, fileID)
+	}
+
+	resp, err := t.client.Chat.Completions.New(ctx, params)
+	if err != nil {
+		return "", fmt.Errorf("filedoc request failed: %w", err)
+	}
+	if len(resp.Choices) == 0 {
+		return "", fmt.Errorf("filedoc returned empty choices")
+	}
+
+	out := strings.TrimSpace(resp.Choices[0].Message.Content)
+	if out == "" {
+		out = "未提取到可读的文档内容。请确认 file_id 是否有效以及模型是否支持文件解析。"
+	}
+	if len(out) > t.maxOutputChars {
+		out = out[:t.maxOutputChars]
+	}
+	return out, nil
+}
+
+func buildExtractionPrompt(fileID, userFocus string) string {
+	focus := strings.TrimSpace(userFocus)
+	if focus == "" {
+		focus = "请输出完整文档信息，包括标题、主题、核心观点、结构大纲、关键术语、重要结论、风险点与后续建议。"
+	}
+
+	return strings.Join([]string{
+		"请基于所附文件输出完整文档信息。",
+		"file_id: " + fileID,
+		"",
+		"输出要求：",
+		"1) 文档基本信息：标题、文档类型、语言、可能作者/组织（若可判断）、时间线索（若可判断）。",
+		"2) 结构化摘要：按章节或逻辑段落给出要点，尽量保持原文顺序。",
+		"3) 关键数据与事实：列出关键数字、术语、专有名词、约束条件。",
+		"4) 风险与不确定性：明确哪些信息来源于文档，哪些是无法确认。",
+		"5) 面向执行的建议：给出可落地的后续行动项。",
+		"",
+		"补充关注点：",
+		focus,
+	}, "\n")
+}
+
+func parseInput(input string) (fileID string, userFocus string, err error) {
+	raw := strings.TrimSpace(input)
+	if raw == "" {
+		return "", "", fmt.Errorf("empty input: expected file_id")
+	}
+
+	if strings.HasPrefix(raw, "{") {
+		var payload map[string]any
+		if jsonErr := json.Unmarshal([]byte(raw), &payload); jsonErr == nil {
+			if id := firstNonEmptyString(payload, "file_id", "fileid", "id", "fileID"); id != "" {
+				return normalizeFileID(id), firstNonEmptyString(payload, "focus", "query", "instruction", "prompt"), nil
+			}
+		}
+	}
+
+	lines := strings.Split(raw, "\n")
+	for _, line := range lines {
+		candidate := extractFileIDToken(line)
+		if candidate != "" {
+			focus := strings.TrimSpace(strings.ReplaceAll(raw, line, ""))
+			return normalizeFileID(candidate), focus, nil
+		}
+	}
+
+	candidate := extractFileIDToken(raw)
+	if candidate == "" {
+		return "", "", fmt.Errorf("no file_id found in input")
+	}
+	return normalizeFileID(candidate), "", nil
+}
+
+func extractFileIDToken(s string) string {
+	fields := strings.FieldsFunc(s, func(r rune) bool {
+		switch r {
+		case ' ', '\t', '\n', '\r', ',', ';', '|':
+			return true
+		default:
+			return false
+		}
+	})
+
+	for _, f := range fields {
+		tok := strings.TrimSpace(strings.Trim(f, "\"'()[]{}"))
+		if tok == "" {
+			continue
+		}
+		lower := strings.ToLower(tok)
+		if strings.HasPrefix(lower, "fileid://") {
+			return tok[len("fileid://"):]
+		}
+		if strings.HasPrefix(lower, "file_id=") || strings.HasPrefix(lower, "fileid=") {
+			idx := strings.Index(tok, "=")
+			if idx >= 0 && idx+1 < len(tok) {
+				return tok[idx+1:]
+			}
+		}
+		if strings.HasPrefix(lower, "file_") || strings.HasPrefix(lower, "file-") {
+			return tok
+		}
+	}
+
+	return ""
+}
+
+func normalizeFileID(id string) string {
+	id = strings.TrimSpace(strings.Trim(id, "\"'"))
+	if strings.HasPrefix(strings.ToLower(id), "fileid://") {
+		return strings.TrimSpace(id[len("fileid://"):])
+	}
+	return id
+}
+
+func firstNonEmptyString(m map[string]any, keys ...string) string {
+	for _, k := range keys {
+		if v, ok := m[k]; ok {
+			if s, ok := v.(string); ok && strings.TrimSpace(s) != "" {
+				return strings.TrimSpace(s)
+			}
+		}
+	}
+	return ""
+}

tools/filedoc/filedoc_test.go

@@ -0,0 +1,74 @@
+package filedoc
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestNameAndDescription(t *testing.T) {
+	tool := New(Config{APIKey: "k", Model: "gpt-4o-mini"}, 5000, nil)
+	if tool.Name() != "extract_file_document" {
+		t.Fatalf("unexpected tool name: %s", tool.Name())
+	}
+	if tool.Description() == "" {
+		t.Fatal("description should not be empty")
+	}
+}
+
+func TestParseInputPlainFileID(t *testing.T) {
+	id, focus, err := parseInput("file_ec_452e96aad38940229058f193f5c5b9c6_12553222")
+	if err != nil {
+		t.Fatalf("parseInput returned error: %v", err)
+	}
+	if id != "file_ec_452e96aad38940229058f193f5c5b9c6_12553222" {
+		t.Fatalf("unexpected id: %s", id)
+	}
+	if focus != "" {
+		t.Fatalf("expected empty focus, got: %q", focus)
+	}
+}
+
+func TestParseInputFileIDSchemeWithFocus(t *testing.T) {
+	input := "fileid://file_ec_12345\n重点关注风险与建议"
+	id, focus, err := parseInput(input)
+	if err != nil {
+		t.Fatalf("parseInput returned error: %v", err)
+	}
+	if id != "file_ec_12345" {
+		t.Fatalf("unexpected id: %s", id)
+	}
+	if focus == "" {
+		t.Fatal("expected non-empty focus")
+	}
+}
+
+func TestParseInputJSON(t *testing.T) {
+	input := `{"file_id":"file_ec_888", "focus":"提取关键结论"}`
+	id, focus, err := parseInput(input)
+	if err != nil {
+		t.Fatalf("parseInput returned error: %v", err)
+	}
+	if id != "file_ec_888" {
+		t.Fatalf("unexpected id: %s", id)
+	}
+	if focus != "提取关键结论" {
+		t.Fatalf("unexpected focus: %s", focus)
+	}
+}
+
+func TestParseInputInvalid(t *testing.T) {
+	_, _, err := parseInput("hello world")
+	if err == nil {
+		t.Fatal("expected parse error")
+	}
+}
+
+func TestBuildExtractionPrompt(t *testing.T) {
+	p := buildExtractionPrompt("file_ec_abc", "关注测试条目")
+	if p == "" {
+		t.Fatal("prompt should not be empty")
+	}
+	if p != "" && !(strings.Contains(p, "file_ec_abc") && strings.Contains(p, "关注测试条目")) {
+		t.Fatalf("unexpected prompt content: %s", p)
+	}
+}

tools/giteaticket/giteaticket.go

@@ -0,0 +1,328 @@
+package giteaticket
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"laodingbot/internal/logger"
+)
+
+// Config Gitea 工单工具的配置。
+type Config struct {
+	BaseURL string        // Gitea 实例地址，例如 https://gitea.example.com
+	Token   string        // Gitea Personal Access Token
+	Owner   string        // 仓库所有者（用户名或组织名）
+	Repo    string        // 仓库名称
+	Timeout time.Duration // HTTP 请求超时
+}
+
+// TicketInput 对应 tool schema 定义的输入参数。
+type TicketInput struct {
+	Title             string   `json:"title"`
+	Body              string   `json:"body"`
+	Labels            []string `json:"labels"`
+	ParentReferenceID string   `json:"parent_reference_id"`
+	EstimatedHours    int      `json:"estimated_hours,omitempty"`
+}
+
+// giteaCreateIssueReq Gitea Create Issue API 请求体。
+type giteaCreateIssueReq struct {
+	Title  string  `json:"title"`
+	Body   string  `json:"body"`
+	Labels []int64 `json:"labels,omitempty"`
+}
+
+// giteaIssueResp Gitea Issue API 响应（仅用到的字段）。
+type giteaIssueResp struct {
+	ID        int64  `json:"id"`
+	Number    int64  `json:"number"`
+	HTMLURL   string `json:"html_url"`
+	Title     string `json:"title"`
+	State     string `json:"state"`
+	CreatedAt string `json:"created_at"`
+}
+
+// giteaLabelResp Gitea Label API 响应。
+type giteaLabelResp struct {
+	ID   int64  `json:"id"`
+	Name string `json:"name"`
+}
+
+// Tool 实现 create_gitea_ticket 工具。
+type Tool struct {
+	baseURL    string
+	token      string
+	owner      string
+	repo       string
+	httpClient *http.Client
+	log        *logger.Logger
+}
+
+// New 创建一个新的 create_gitea_ticket 工具实例。
+func New(cfg Config, log *logger.Logger) *Tool {
+	baseURL := strings.TrimRight(strings.TrimSpace(cfg.BaseURL), "/")
+	if cfg.Timeout <= 0 {
+		cfg.Timeout = 30 * time.Second
+	}
+	return &Tool{
+		baseURL:    baseURL,
+		token:      strings.TrimSpace(cfg.Token),
+		owner:      strings.TrimSpace(cfg.Owner),
+		repo:       strings.TrimSpace(cfg.Repo),
+		httpClient: &http.Client{Timeout: cfg.Timeout},
+		log:        log,
+	}
+}
+
+func (t *Tool) Name() string { return "create_gitea_ticket" }
+
+func (t *Tool) Description() string {
+	return `DevOps Agent 使用此工具，将 PI Plan 中的 Feature/Enabler 拆解为可执行 User Story，并在 Gitea 创建 Issue。输入 JSON: {"title":"...","body":"...","labels":[...],"parent_reference_id":"...","estimated_hours":8}`
+}
+
+func (t *Tool) Call(ctx context.Context, input string) (string, error) {
+	ticket, err := parseInput(input)
+	if err != nil {
+		return "", fmt.Errorf("create_gitea_ticket: invalid input: %w", err)
+	}
+	if err := validate(ticket); err != nil {
+		return "", fmt.Errorf("create_gitea_ticket: validation failed: %w", err)
+	}
+	if t.baseURL == "" || t.token == "" || t.owner == "" || t.repo == "" {
+		return "", fmt.Errorf("create_gitea_ticket: missing Gitea configuration (base_url, token, owner, repo)")
+	}
+
+	body := buildIssueBody(ticket)
+
+	labelIDs, err := t.resolveLabels(ctx, ticket.Labels)
+	if err != nil {
+		return "", fmt.Errorf("create_gitea_ticket: label resolution failed: %w", err)
+	}
+
+	if t.log != nil {
+		t.log.Infof("create_gitea_ticket: creating issue title=%q parent=%s labels=%v",
+			ticket.Title, ticket.ParentReferenceID, ticket.Labels)
+	}
+
+	issue, err := t.createIssue(ctx, ticket.Title, body, labelIDs)
+	if err != nil {
+		return "", fmt.Errorf("create_gitea_ticket: create issue failed: %w", err)
+	}
+
+	result := formatResult(issue, ticket)
+	return result, nil
+}
+
+func parseInput(input string) (*TicketInput, error) {
+	raw := strings.TrimSpace(input)
+	if raw == "" {
+		return nil, fmt.Errorf("empty input")
+	}
+	var ticket TicketInput
+	if err := json.Unmarshal([]byte(raw), &ticket); err != nil {
+		return nil, fmt.Errorf("JSON parse error: %w", err)
+	}
+	return &ticket, nil
+}
+
+func validate(t *TicketInput) error {
+	if strings.TrimSpace(t.Title) == "" {
+		return fmt.Errorf("title is required")
+	}
+	if strings.TrimSpace(t.Body) == "" {
+		return fmt.Errorf("body is required")
+	}
+	if len(t.Labels) == 0 {
+		return fmt.Errorf("labels is required and must contain at least one label")
+	}
+	if strings.TrimSpace(t.ParentReferenceID) == "" {
+		return fmt.Errorf("parent_reference_id is required")
+	}
+	return nil
+}
+
+// buildIssueBody 在原始 body 上追加溯源元数据和工时估算。
+func buildIssueBody(ticket *TicketInput) string {
+	var b strings.Builder
+	b.WriteString(ticket.Body)
+	b.WriteString("\n\n---\n\n")
+	b.WriteString("## 📋 SAFe 元数据\n\n")
+	b.WriteString(fmt.Sprintf("- **溯源 (Parent Reference)**: `%s`\n", ticket.ParentReferenceID))
+	if ticket.EstimatedHours > 0 {
+		b.WriteString(fmt.Sprintf("- **预估工时**: %d 小时\n", ticket.EstimatedHours))
+	}
+	b.WriteString(fmt.Sprintf("- **标签**: %s\n", strings.Join(ticket.Labels, ", ")))
+	return b.String()
+}
+
+// resolveLabels 通过 Gitea API 查询已有标签，将标签名映射为 ID。
+// 如果标签不存在则自动创建。
+func (t *Tool) resolveLabels(ctx context.Context, labelNames []string) ([]int64, error) {
+	existing, err := t.listLabels(ctx)
+	if err != nil {
+		return nil, err
+	}
+
+	nameToID := make(map[string]int64, len(existing))
+	for _, l := range existing {
+		nameToID[l.Name] = l.ID
+	}
+
+	ids := make([]int64, 0, len(labelNames))
+	for _, name := range labelNames {
+		name = strings.TrimSpace(name)
+		if name == "" {
+			continue
+		}
+		if id, ok := nameToID[name]; ok {
+			ids = append(ids, id)
+		} else {
+			id, err := t.createLabel(ctx, name)
+			if err != nil {
+				return nil, fmt.Errorf("create label %q: %w", name, err)
+			}
+			ids = append(ids, id)
+		}
+	}
+	return ids, nil
+}
+
+func (t *Tool) listLabels(ctx context.Context) ([]giteaLabelResp, error) {
+	url := fmt.Sprintf("%s/api/v1/repos/%s/%s/labels?limit=50", t.baseURL, t.owner, t.repo)
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return nil, err
+	}
+	t.setAuth(req)
+
+	resp, err := t.httpClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("list labels request failed: %w", err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		body, _ := io.ReadAll(resp.Body)
+		return nil, fmt.Errorf("list labels returned %d: %s", resp.StatusCode, string(body))
+	}
+
+	var labels []giteaLabelResp
+	if err := json.NewDecoder(resp.Body).Decode(&labels); err != nil {
+		return nil, fmt.Errorf("decode labels response: %w", err)
+	}
+	return labels, nil
+}
+
+func (t *Tool) createLabel(ctx context.Context, name string) (int64, error) {
+	url := fmt.Sprintf("%s/api/v1/repos/%s/%s/labels", t.baseURL, t.owner, t.repo)
+	payload := map[string]string{
+		"name":  name,
+		"color": labelColor(name),
+	}
+	body, _ := json.Marshal(payload)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		return 0, err
+	}
+	t.setAuth(req)
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := t.httpClient.Do(req)
+	if err != nil {
+		return 0, fmt.Errorf("create label request failed: %w", err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusCreated {
+		respBody, _ := io.ReadAll(resp.Body)
+		return 0, fmt.Errorf("create label returned %d: %s", resp.StatusCode, string(respBody))
+	}
+
+	var label giteaLabelResp
+	if err := json.NewDecoder(resp.Body).Decode(&label); err != nil {
+		return 0, fmt.Errorf("decode create label response: %w", err)
+	}
+	return label.ID, nil
+}
+
+func (t *Tool) createIssue(ctx context.Context, title, body string, labelIDs []int64) (*giteaIssueResp, error) {
+	url := fmt.Sprintf("%s/api/v1/repos/%s/%s/issues", t.baseURL, t.owner, t.repo)
+	payload := giteaCreateIssueReq{
+		Title:  title,
+		Body:   body,
+		Labels: labelIDs,
+	}
+	jsonBody, _ := json.Marshal(payload)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(jsonBody))
+	if err != nil {
+		return nil, err
+	}
+	t.setAuth(req)
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := t.httpClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("create issue request failed: %w", err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusCreated {
+		respBody, _ := io.ReadAll(resp.Body)
+		return nil, fmt.Errorf("create issue returned %d: %s", resp.StatusCode, string(respBody))
+	}
+
+	var issue giteaIssueResp
+	if err := json.NewDecoder(resp.Body).Decode(&issue); err != nil {
+		return nil, fmt.Errorf("decode create issue response: %w", err)
+	}
+	return &issue, nil
+}
+
+func (t *Tool) setAuth(req *http.Request) {
+	if t.token != "" {
+		req.Header.Set("Authorization", "token "+t.token)
+	}
+}
+
+func formatResult(issue *giteaIssueResp, ticket *TicketInput) string {
+	var b strings.Builder
+	b.WriteString("✅ Gitea Issue 创建成功\n\n")
+	b.WriteString(fmt.Sprintf("- **Issue 编号**: #%d\n", issue.Number))
+	b.WriteString(fmt.Sprintf("- **标题**: %s\n", issue.Title))
+	b.WriteString(fmt.Sprintf("- **状态**: %s\n", issue.State))
+	b.WriteString(fmt.Sprintf("- **链接**: %s\n", issue.HTMLURL))
+	b.WriteString(fmt.Sprintf("- **溯源 (Parent)**: %s\n", ticket.ParentReferenceID))
+	if ticket.EstimatedHours > 0 {
+		b.WriteString(fmt.Sprintf("- **预估工时**: %d 小时\n", ticket.EstimatedHours))
+	}
+	b.WriteString(fmt.Sprintf("- **标签**: %s\n", strings.Join(ticket.Labels, ", ")))
+	return b.String()
+}
+
+// labelColor 根据标签前缀返回一个辨识度高的颜色。
+func labelColor(name string) string {
+	lower := strings.ToLower(name)
+	switch {
+	case strings.HasPrefix(lower, "type/"):
+		return "#0075ca"
+	case strings.HasPrefix(lower, "domain/"):
+		return "#7057ff"
+	case strings.HasPrefix(lower, "priority/"):
+		if strings.Contains(lower, "high") || strings.Contains(lower, "critical") {
+			return "#d73a4a"
+		}
+		return "#e4e669"
+	case strings.HasPrefix(lower, "status/"):
+		return "#0e8a16"
+	default:
+		return "#ededed"
+	}
+}

tools/giteaticket/giteaticket_test.go

@@ -0,0 +1,403 @@
+package giteaticket
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"testing"
+)
+
+func validTicketInput() TicketInput {
+	return TicketInput{
+		Title: "实现云端固件版本解析 API",
+		Body: `## 溯源
+- Parent: FEAT_OTA_001
+
+## 任务上下文
+云端需要能够解析上传的固件包，提取版本号与依赖关系。
+
+## 验收标准
+- [ ] 上传 .bin 固件文件后返回解析结果 JSON
+- [ ] 解析结果包含 version, dependencies 字段
+
+## NFRs
+- 响应时间 P99 < 500ms
+
+## 技术实现思路
+使用 Go 读取固件文件头部 metadata。`,
+		Labels:            []string{"type/story", "domain/cloud", "priority/high", "status/todo"},
+		ParentReferenceID: "FEAT_OTA_001",
+		EstimatedHours:    8,
+	}
+}
+
+// ── parseInput tests ──
+
+func TestParseInputValid(t *testing.T) {
+	in := validTicketInput()
+	data, _ := json.Marshal(in)
+	ticket, err := parseInput(string(data))
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if ticket.Title != in.Title {
+		t.Errorf("title mismatch: got %q", ticket.Title)
+	}
+	if len(ticket.Labels) != 4 {
+		t.Errorf("expected 4 labels, got %d", len(ticket.Labels))
+	}
+	if ticket.ParentReferenceID != "FEAT_OTA_001" {
+		t.Errorf("parent_reference_id mismatch: got %q", ticket.ParentReferenceID)
+	}
+	if ticket.EstimatedHours != 8 {
+		t.Errorf("estimated_hours mismatch: got %d", ticket.EstimatedHours)
+	}
+}
+
+func TestParseInputEmpty(t *testing.T) {
+	_, err := parseInput("")
+	if err == nil {
+		t.Fatal("expected error for empty input")
+	}
+}
+
+func TestParseInputInvalidJSON(t *testing.T) {
+	_, err := parseInput("{bad json}")
+	if err == nil {
+		t.Fatal("expected error for invalid JSON")
+	}
+}
+
+// ── validate tests ──
+
+func TestValidateMissingTitle(t *testing.T) {
+	in := validTicketInput()
+	in.Title = ""
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "title") {
+		t.Fatalf("expected title error, got: %v", err)
+	}
+}
+
+func TestValidateMissingBody(t *testing.T) {
+	in := validTicketInput()
+	in.Body = ""
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "body") {
+		t.Fatalf("expected body error, got: %v", err)
+	}
+}
+
+func TestValidateMissingLabels(t *testing.T) {
+	in := validTicketInput()
+	in.Labels = nil
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "labels") {
+		t.Fatalf("expected labels error, got: %v", err)
+	}
+}
+
+func TestValidateMissingParentRef(t *testing.T) {
+	in := validTicketInput()
+	in.ParentReferenceID = ""
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "parent_reference_id") {
+		t.Fatalf("expected parent_reference_id error, got: %v", err)
+	}
+}
+
+func TestValidateOptionalEstimatedHours(t *testing.T) {
+	in := validTicketInput()
+	in.EstimatedHours = 0
+	err := validate(&in)
+	if err != nil {
+		t.Fatalf("estimated_hours should be optional, got: %v", err)
+	}
+}
+
+// ── buildIssueBody tests ──
+
+func TestBuildIssueBodyContainsMetadata(t *testing.T) {
+	in := validTicketInput()
+	body := buildIssueBody(&in)
+
+	if !strings.Contains(body, "FEAT_OTA_001") {
+		t.Error("body missing parent_reference_id")
+	}
+	if !strings.Contains(body, "8 小时") {
+		t.Error("body missing estimated_hours")
+	}
+	if !strings.Contains(body, "type/story") {
+		t.Error("body missing labels")
+	}
+	if !strings.Contains(body, "SAFe 元数据") {
+		t.Error("body missing metadata section header")
+	}
+}
+
+func TestBuildIssueBodyNoHours(t *testing.T) {
+	in := validTicketInput()
+	in.EstimatedHours = 0
+	body := buildIssueBody(&in)
+	if strings.Contains(body, "预估工时") {
+		t.Error("body should not contain estimated hours when 0")
+	}
+}
+
+// ── labelColor tests ──
+
+func TestLabelColor(t *testing.T) {
+	tests := []struct {
+		name     string
+		expected string
+	}{
+		{"type/story", "#0075ca"},
+		{"domain/cloud", "#7057ff"},
+		{"priority/high", "#d73a4a"},
+		{"priority/low", "#e4e669"},
+		{"status/todo", "#0e8a16"},
+		{"custom-label", "#ededed"},
+	}
+	for _, tc := range tests {
+		got := labelColor(tc.name)
+		if got != tc.expected {
+			t.Errorf("labelColor(%q) = %q, want %q", tc.name, got, tc.expected)
+		}
+	}
+}
+
+// ── Name / Description tests ──
+
+func TestNameAndDescription(t *testing.T) {
+	tool := New(Config{}, nil)
+	if tool.Name() != "create_gitea_ticket" {
+		t.Errorf("unexpected name: %s", tool.Name())
+	}
+	if tool.Description() == "" {
+		t.Error("description should not be empty")
+	}
+}
+
+// ── Integration test with mock Gitea server ──
+
+func TestCallWithMockGitea(t *testing.T) {
+	var mu sync.Mutex
+	var createdIssue map[string]interface{}
+	labelIDCounter := int64(100)
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// Verify auth header
+		auth := r.Header.Get("Authorization")
+		if auth != "token test-token" {
+			http.Error(w, "unauthorized", http.StatusUnauthorized)
+			return
+		}
+
+		switch {
+		// GET /api/v1/repos/owner/repo/labels
+		case r.Method == http.MethodGet && strings.Contains(r.URL.Path, "/labels"):
+			w.Header().Set("Content-Type", "application/json")
+			// Return one existing label
+			json.NewEncoder(w).Encode([]giteaLabelResp{
+				{ID: 1, Name: "type/story"},
+			})
+
+		// POST /api/v1/repos/owner/repo/labels
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/labels"):
+			var payload map[string]string
+			json.NewDecoder(r.Body).Decode(&payload)
+			mu.Lock()
+			labelIDCounter++
+			id := labelIDCounter
+			mu.Unlock()
+			w.WriteHeader(http.StatusCreated)
+			json.NewEncoder(w).Encode(giteaLabelResp{
+				ID:   id,
+				Name: payload["name"],
+			})
+
+		// POST /api/v1/repos/owner/repo/issues
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/issues"):
+			json.NewDecoder(r.Body).Decode(&createdIssue)
+			w.WriteHeader(http.StatusCreated)
+			json.NewEncoder(w).Encode(giteaIssueResp{
+				ID:        42,
+				Number:    42,
+				HTMLURL:   "https://gitea.example.com/owner/repo/issues/42",
+				Title:     createdIssue["title"].(string),
+				State:     "open",
+				CreatedAt: "2026-03-11T10:00:00Z",
+			})
+
+		default:
+			http.Error(w, "not found", http.StatusNotFound)
+		}
+	}))
+	defer server.Close()
+
+	tool := New(Config{
+		BaseURL: server.URL,
+		Token:   "test-token",
+		Owner:   "owner",
+		Repo:    "repo",
+	}, nil)
+
+	in := validTicketInput()
+	data, _ := json.Marshal(in)
+
+	result, err := tool.Call(context.Background(), string(data))
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if !strings.Contains(result, "#42") {
+		t.Error("result missing issue number")
+	}
+	if !strings.Contains(result, "创建成功") {
+		t.Error("result missing success message")
+	}
+	if !strings.Contains(result, "FEAT_OTA_001") {
+		t.Error("result missing parent reference")
+	}
+	if !strings.Contains(result, "8 小时") {
+		t.Error("result missing estimated hours")
+	}
+
+	// Verify the issue body sent to Gitea contains SAFe metadata
+	if body, ok := createdIssue["body"].(string); ok {
+		if !strings.Contains(body, "SAFe 元数据") {
+			t.Error("issue body missing SAFe metadata")
+		}
+		if !strings.Contains(body, "FEAT_OTA_001") {
+			t.Error("issue body missing parent reference")
+		}
+	} else {
+		t.Error("createdIssue body not captured")
+	}
+}
+
+func TestCallMissingConfig(t *testing.T) {
+	tool := New(Config{}, nil)
+	in := validTicketInput()
+	data, _ := json.Marshal(in)
+
+	_, err := tool.Call(context.Background(), string(data))
+	if err == nil || !strings.Contains(err.Error(), "missing Gitea configuration") {
+		t.Fatalf("expected config error, got: %v", err)
+	}
+}
+
+func TestCallInvalidInput(t *testing.T) {
+	tool := New(Config{
+		BaseURL: "http://localhost",
+		Token:   "t",
+		Owner:   "o",
+		Repo:    "r",
+	}, nil)
+	_, err := tool.Call(context.Background(), "not json")
+	if err == nil {
+		t.Fatal("expected error for invalid input")
+	}
+}
+
+func TestCallValidationError(t *testing.T) {
+	tool := New(Config{
+		BaseURL: "http://localhost",
+		Token:   "t",
+		Owner:   "o",
+		Repo:    "r",
+	}, nil)
+	in := validTicketInput()
+	in.Title = ""
+	data, _ := json.Marshal(in)
+
+	_, err := tool.Call(context.Background(), string(data))
+	if err == nil || !strings.Contains(err.Error(), "title") {
+		t.Fatalf("expected title validation error, got: %v", err)
+	}
+}
+
+// Test that labels resolution creates missing labels
+func TestCallCreatesNewLabels(t *testing.T) {
+	createdLabels := []string{}
+	var mu sync.Mutex
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodGet && strings.Contains(r.URL.Path, "/labels"):
+			w.Header().Set("Content-Type", "application/json")
+			// No existing labels
+			json.NewEncoder(w).Encode([]giteaLabelResp{})
+
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/labels"):
+			var payload map[string]string
+			json.NewDecoder(r.Body).Decode(&payload)
+			mu.Lock()
+			createdLabels = append(createdLabels, payload["name"])
+			mu.Unlock()
+			w.WriteHeader(http.StatusCreated)
+			json.NewEncoder(w).Encode(giteaLabelResp{
+				ID:   int64(len(createdLabels)),
+				Name: payload["name"],
+			})
+
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/issues"):
+			w.WriteHeader(http.StatusCreated)
+			json.NewEncoder(w).Encode(giteaIssueResp{
+				ID: 1, Number: 1, HTMLURL: "http://test/1", Title: "t", State: "open",
+			})
+
+		default:
+			http.Error(w, "not found", http.StatusNotFound)
+		}
+	}))
+	defer server.Close()
+
+	tool := New(Config{
+		BaseURL: server.URL,
+		Token:   "tok",
+		Owner:   "o",
+		Repo:    "r",
+	}, nil)
+
+	in := validTicketInput()
+	data, _ := json.Marshal(in)
+	_, err := tool.Call(context.Background(), string(data))
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	mu.Lock()
+	defer mu.Unlock()
+	if len(createdLabels) != 4 {
+		t.Errorf("expected 4 labels created, got %d: %v", len(createdLabels), createdLabels)
+	}
+}
+
+func TestFormatResult(t *testing.T) {
+	issue := &giteaIssueResp{
+		Number:  99,
+		Title:   "测试任务",
+		State:   "open",
+		HTMLURL: "https://gitea.example.com/issues/99",
+	}
+	ticket := &TicketInput{
+		ParentReferenceID: "ENAB_KAFKA_001",
+		Labels:            []string{"type/enabler"},
+		EstimatedHours:    4,
+	}
+	result := formatResult(issue, ticket)
+
+	if !strings.Contains(result, "#99") {
+		t.Error("missing issue number")
+	}
+	if !strings.Contains(result, "ENAB_KAFKA_001") {
+		t.Error("missing parent reference")
+	}
+	if !strings.Contains(result, "4 小时") {
+		t.Error("missing estimated hours")
+	}
+}

tools/piplan/piplan.go

@@ -0,0 +1,309 @@
+package piplan
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"laodingbot/internal/logger"
+)
+
+// Feature 产品经理视角输出的业务特性。
+type Feature struct {
+	FeatureID          string   `json:"feature_id"`
+	Title              string   `json:"title"`
+	BenefitHypothesis  string   `json:"benefit_hypothesis"`
+	AcceptanceCriteria []string `json:"acceptance_criteria"`
+}
+
+// Enabler 系统架构师视角输出的技术赋能特性（架构跑道）。
+type Enabler struct {
+	EnablerID            string `json:"enabler_id"`
+	Title                string `json:"title"`
+	ArchitecturalPurpose string `json:"architectural_purpose"`
+}
+
+// NFRs 非功能性需求。
+type NFRs struct {
+	Performance string `json:"performance"`
+	Security    string `json:"security"`
+}
+
+// Dependency RTE 梳理的任务依赖关系。
+type Dependency struct {
+	SourceID string `json:"source_id"`
+	TargetID string `json:"target_id"`
+	Reason   string `json:"reason"`
+}
+
+// PIPlanInput publish_pi_plan 工具的完整输入结构。
+type PIPlanInput struct {
+	PIVision     string       `json:"pi_vision"`
+	Features     []Feature    `json:"features"`
+	Enablers     []Enabler    `json:"enablers"`
+	NFRs         NFRs         `json:"nfrs"`
+	Dependencies []Dependency `json:"dependencies"`
+}
+
+// Tool 实现 SAFe PI 规划发布工具。
+type Tool struct {
+	maxOutputChars int
+	log            *logger.Logger
+}
+
+// New 创建一个新的 publish_pi_plan 工具实例。
+func New(maxOutputChars int, log *logger.Logger) *Tool {
+	if maxOutputChars <= 0 {
+		maxOutputChars = 20000
+	}
+	return &Tool{
+		maxOutputChars: maxOutputChars,
+		log:            log,
+	}
+}
+
+func (t *Tool) Name() string { return "publish_pi_plan" }
+
+func (t *Tool) Description() string {
+	return `当铁三角（PM, 架构师, RTE）完成 PI 规划推演后，调用此工具输出标准化的架构蓝图与任务清单。输入为 JSON，包含 pi_vision, features, enablers, nfrs, dependencies 字段。`
+}
+
+func (t *Tool) Call(ctx context.Context, input string) (string, error) {
+	plan, err := parseInput(input)
+	if err != nil {
+		return "", fmt.Errorf("publish_pi_plan: invalid input: %w", err)
+	}
+
+	if err := validate(plan); err != nil {
+		return "", fmt.Errorf("publish_pi_plan: validation failed: %w", err)
+	}
+
+	if t.log != nil {
+		t.log.Infof("publish_pi_plan: features=%d enablers=%d deps=%d",
+			len(plan.Features), len(plan.Enablers), len(plan.Dependencies))
+	}
+
+	output := render(plan)
+
+	if len(output) > t.maxOutputChars {
+		output = output[:t.maxOutputChars]
+	}
+	return output, nil
+}
+
+func parseInput(input string) (*PIPlanInput, error) {
+	raw := strings.TrimSpace(input)
+	if raw == "" {
+		return nil, fmt.Errorf("empty input")
+	}
+
+	var plan PIPlanInput
+	if err := json.Unmarshal([]byte(raw), &plan); err != nil {
+		return nil, fmt.Errorf("JSON parse error: %w", err)
+	}
+	return &plan, nil
+}
+
+func validate(p *PIPlanInput) error {
+	if strings.TrimSpace(p.PIVision) == "" {
+		return fmt.Errorf("pi_vision is required")
+	}
+	if len(p.Features) == 0 {
+		return fmt.Errorf("at least one feature is required")
+	}
+	for i, f := range p.Features {
+		if strings.TrimSpace(f.FeatureID) == "" {
+			return fmt.Errorf("features[%d].feature_id is required", i)
+		}
+		if strings.TrimSpace(f.Title) == "" {
+			return fmt.Errorf("features[%d].title is required", i)
+		}
+		if strings.TrimSpace(f.BenefitHypothesis) == "" {
+			return fmt.Errorf("features[%d].benefit_hypothesis is required", i)
+		}
+		if len(f.AcceptanceCriteria) == 0 {
+			return fmt.Errorf("features[%d].acceptance_criteria requires at least one item", i)
+		}
+	}
+	for i, e := range p.Enablers {
+		if strings.TrimSpace(e.EnablerID) == "" {
+			return fmt.Errorf("enablers[%d].enabler_id is required", i)
+		}
+		if strings.TrimSpace(e.Title) == "" {
+			return fmt.Errorf("enablers[%d].title is required", i)
+		}
+		if strings.TrimSpace(e.ArchitecturalPurpose) == "" {
+			return fmt.Errorf("enablers[%d].architectural_purpose is required", i)
+		}
+	}
+	if strings.TrimSpace(p.NFRs.Performance) == "" {
+		return fmt.Errorf("nfrs.performance is required")
+	}
+	if strings.TrimSpace(p.NFRs.Security) == "" {
+		return fmt.Errorf("nfrs.security is required")
+	}
+	for i, d := range p.Dependencies {
+		if strings.TrimSpace(d.SourceID) == "" {
+			return fmt.Errorf("dependencies[%d].source_id is required", i)
+		}
+		if strings.TrimSpace(d.TargetID) == "" {
+			return fmt.Errorf("dependencies[%d].target_id is required", i)
+		}
+	}
+	return nil
+}
+
+// render 将 PI 规划输入渲染为标准化的 Markdown 架构蓝图与任务清单。
+func render(p *PIPlanInput) string {
+	var b strings.Builder
+
+	// ── 标题 ──
+	b.WriteString("# PI 规划架构蓝图与任务清单\n\n")
+
+	// ── 1. PI 愿景 ──
+	b.WriteString("## 1. PI 愿景\n\n")
+	b.WriteString(strings.TrimSpace(p.PIVision))
+	b.WriteString("\n\n")
+
+	// ── 2. 业务特性清单 (Features) ──
+	b.WriteString("## 2. 业务特性清单 (Features)\n\n")
+	for _, f := range p.Features {
+		b.WriteString(fmt.Sprintf("### %s — %s\n\n", f.FeatureID, f.Title))
+		b.WriteString(fmt.Sprintf("**业务价值假设**: %s\n\n", f.BenefitHypothesis))
+		b.WriteString("**验收标准 (AC)**:\n\n")
+		for j, ac := range f.AcceptanceCriteria {
+			b.WriteString(fmt.Sprintf("- [ ] AC-%d: %s\n", j+1, ac))
+		}
+		b.WriteString("\n")
+	}
+
+	// ── 3. 技术赋能特性 (Enablers / 架构跑道) ──
+	b.WriteString("## 3. 技术赋能特性 (Enablers / 架构跑道)\n\n")
+	if len(p.Enablers) == 0 {
+		b.WriteString("_无技术赋能特性。_\n\n")
+	} else {
+		b.WriteString("| Enabler ID | 名称 | 架构意图 |\n")
+		b.WriteString("|------------|------|----------|\n")
+		for _, e := range p.Enablers {
+			b.WriteString(fmt.Sprintf("| %s | %s | %s |\n",
+				e.EnablerID, e.Title, e.ArchitecturalPurpose))
+		}
+		b.WriteString("\n")
+	}
+
+	// ── 4. 非功能性需求 (NFRs) ──
+	b.WriteString("## 4. 非功能性需求 (NFRs)\n\n")
+	b.WriteString(fmt.Sprintf("- **性能**: %s\n", p.NFRs.Performance))
+	b.WriteString(fmt.Sprintf("- **安全与合规**: %s\n", p.NFRs.Security))
+	b.WriteString("\n")
+
+	// ── 5. 依赖关系图 ──
+	b.WriteString("## 5. 依赖关系\n\n")
+	if len(p.Dependencies) == 0 {
+		b.WriteString("_无跨任务依赖。_\n\n")
+	} else {
+		b.WriteString("| 前置任务 (Source) | 后续任务 (Target) | 依赖原因 |\n")
+		b.WriteString("|-------------------|-------------------|----------|\n")
+		for _, d := range p.Dependencies {
+			reason := d.Reason
+			if reason == "" {
+				reason = "—"
+			}
+			b.WriteString(fmt.Sprintf("| %s | %s | %s |\n",
+				d.SourceID, d.TargetID, reason))
+		}
+		b.WriteString("\n")
+	}
+
+	// ── 6. 建议执行顺序 ──
+	b.WriteString("## 6. 建议执行顺序\n\n")
+	order := computeExecutionOrder(p)
+	for i, id := range order {
+		b.WriteString(fmt.Sprintf("%d. %s\n", i+1, id))
+	}
+	b.WriteString("\n")
+
+	// ── 7. 质量门禁检查清单 ──
+	b.WriteString("## 7. 质量门禁检查清单\n\n")
+	b.WriteString("### 业务验收测试用例\n\n")
+	for _, f := range p.Features {
+		for j, ac := range f.AcceptanceCriteria {
+			b.WriteString(fmt.Sprintf("- [ ] [%s] AC-%d: %s\n", f.FeatureID, j+1, ac))
+		}
+	}
+	b.WriteString("\n### 非功能性验证\n\n")
+	b.WriteString(fmt.Sprintf("- [ ] 性能压测: %s\n", p.NFRs.Performance))
+	b.WriteString(fmt.Sprintf("- [ ] 安全扫描: %s\n", p.NFRs.Security))
+	b.WriteString("\n")
+
+	return b.String()
+}
+
+// computeExecutionOrder 根据依赖关系计算拓扑排序的执行顺序。
+// 先排 Enabler，再排 Feature；无依赖的排在前面。
+func computeExecutionOrder(p *PIPlanInput) []string {
+	// 收集所有 ID
+	allIDs := make([]string, 0, len(p.Enablers)+len(p.Features))
+	idSet := make(map[string]bool)
+	for _, e := range p.Enablers {
+		allIDs = append(allIDs, e.EnablerID)
+		idSet[e.EnablerID] = true
+	}
+	for _, f := range p.Features {
+		allIDs = append(allIDs, f.FeatureID)
+		idSet[f.FeatureID] = true
+	}
+
+	// 构建入度表和邻接表
+	inDegree := make(map[string]int)
+	adj := make(map[string][]string)
+	for _, id := range allIDs {
+		inDegree[id] = 0
+	}
+	for _, d := range p.Dependencies {
+		if !idSet[d.SourceID] || !idSet[d.TargetID] {
+			continue
+		}
+		adj[d.SourceID] = append(adj[d.SourceID], d.TargetID)
+		inDegree[d.TargetID]++
+	}
+
+	// Kahn 拓扑排序
+	queue := make([]string, 0)
+	// 先加入度为 0 的 Enabler，再加入度为 0 的 Feature，保持稳定顺序
+	for _, e := range p.Enablers {
+		if inDegree[e.EnablerID] == 0 {
+			queue = append(queue, e.EnablerID)
+		}
+	}
+	for _, f := range p.Features {
+		if inDegree[f.FeatureID] == 0 {
+			queue = append(queue, f.FeatureID)
+		}
+	}
+
+	var result []string
+	for len(queue) > 0 {
+		curr := queue[0]
+		queue = queue[1:]
+		result = append(result, curr)
+		for _, next := range adj[curr] {
+			inDegree[next]--
+			if inDegree[next] == 0 {
+				queue = append(queue, next)
+			}
+		}
+	}
+
+	// 如果存在环，将未排序的节点追加到末尾并标记
+	if len(result) < len(allIDs) {
+		for _, id := range allIDs {
+			if inDegree[id] > 0 {
+				result = append(result, id+" ⚠️(循环依赖)")
+			}
+		}
+	}
+
+	return result
+}

tools/piplan/piplan_test.go

@@ -0,0 +1,334 @@
+package piplan
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+func validInput() PIPlanInput {
+	return PIPlanInput{
+		PIVision: "实现车云一体化 OTA 系统，支撑百万级终端设备的安全固件升级",
+		Features: []Feature{
+			{
+				FeatureID:         "FEAT_OTA_001",
+				Title:             "云端固件版本依赖检查",
+				BenefitHypothesis: "上线后减少 30% 的固件回退率",
+				AcceptanceCriteria: []string{
+					"上传固件时自动解析并记录版本依赖关系",
+					"下发升级任务时自动校验设备当前版本是否满足依赖",
+					"不满足依赖时返回明确的错误提示及所需前置版本",
+				},
+			},
+			{
+				FeatureID:         "FEAT_OTA_002",
+				Title:             "端侧断点续传",
+				BenefitHypothesis: "弱网环境下固件下载成功率提升至 99.5%",
+				AcceptanceCriteria: []string{
+					"支持分片下载与本地缓存校验",
+					"网络恢复后自动续传，无需用户干预",
+				},
+			},
+		},
+		Enablers: []Enabler{
+			{
+				EnablerID:            "ENAB_KAFKA_001",
+				Title:                "搭建跨可用区的高可用 Kafka 集群",
+				ArchitecturalPurpose: "为高并发 OTA 状态机提供可靠消息管道",
+			},
+			{
+				EnablerID:            "ENAB_S3_001",
+				Title:                "对象存储多区域同步",
+				ArchitecturalPurpose: "保证固件文件在多区域的低延迟分发",
+			},
+		},
+		NFRs: NFRs{
+			Performance: "API 响应时间 P99 < 200ms，吞吐量 > 10000 QPS",
+			Security:    "车云通信必须使用 TLS 1.3，敏感数据必须脱敏",
+		},
+		Dependencies: []Dependency{
+			{
+				SourceID: "ENAB_KAFKA_001",
+				TargetID: "FEAT_OTA_001",
+				Reason:   "版本检查服务依赖 Kafka 进行异步事件通知",
+			},
+			{
+				SourceID: "ENAB_S3_001",
+				TargetID: "FEAT_OTA_002",
+				Reason:   "断点续传需要对象存储支持 Range 请求",
+			},
+		},
+	}
+}
+
+func TestParseInputValid(t *testing.T) {
+	in := validInput()
+	data, _ := json.Marshal(in)
+	plan, err := parseInput(string(data))
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if plan.PIVision != in.PIVision {
+		t.Errorf("pi_vision mismatch: got %q", plan.PIVision)
+	}
+	if len(plan.Features) != 2 {
+		t.Errorf("expected 2 features, got %d", len(plan.Features))
+	}
+	if len(plan.Enablers) != 2 {
+		t.Errorf("expected 2 enablers, got %d", len(plan.Enablers))
+	}
+	if len(plan.Dependencies) != 2 {
+		t.Errorf("expected 2 dependencies, got %d", len(plan.Dependencies))
+	}
+}
+
+func TestParseInputEmpty(t *testing.T) {
+	_, err := parseInput("")
+	if err == nil {
+		t.Fatal("expected error for empty input")
+	}
+}
+
+func TestParseInputInvalidJSON(t *testing.T) {
+	_, err := parseInput("{not json}")
+	if err == nil {
+		t.Fatal("expected error for invalid JSON")
+	}
+}
+
+func TestValidateMissingVision(t *testing.T) {
+	in := validInput()
+	in.PIVision = ""
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "pi_vision") {
+		t.Fatalf("expected pi_vision error, got: %v", err)
+	}
+}
+
+func TestValidateNoFeatures(t *testing.T) {
+	in := validInput()
+	in.Features = nil
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "feature") {
+		t.Fatalf("expected feature error, got: %v", err)
+	}
+}
+
+func TestValidateFeatureMissingID(t *testing.T) {
+	in := validInput()
+	in.Features[0].FeatureID = ""
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "feature_id") {
+		t.Fatalf("expected feature_id error, got: %v", err)
+	}
+}
+
+func TestValidateFeatureMissingAC(t *testing.T) {
+	in := validInput()
+	in.Features[0].AcceptanceCriteria = nil
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "acceptance_criteria") {
+		t.Fatalf("expected acceptance_criteria error, got: %v", err)
+	}
+}
+
+func TestValidateEnablerMissingPurpose(t *testing.T) {
+	in := validInput()
+	in.Enablers[0].ArchitecturalPurpose = ""
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "architectural_purpose") {
+		t.Fatalf("expected architectural_purpose error, got: %v", err)
+	}
+}
+
+func TestValidateNFRsMissingPerformance(t *testing.T) {
+	in := validInput()
+	in.NFRs.Performance = ""
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "performance") {
+		t.Fatalf("expected performance error, got: %v", err)
+	}
+}
+
+func TestValidateNFRsMissingSecurity(t *testing.T) {
+	in := validInput()
+	in.NFRs.Security = ""
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "security") {
+		t.Fatalf("expected security error, got: %v", err)
+	}
+}
+
+func TestValidateDepMissingSourceID(t *testing.T) {
+	in := validInput()
+	in.Dependencies[0].SourceID = ""
+	err := validate(&in)
+	if err == nil || !strings.Contains(err.Error(), "source_id") {
+		t.Fatalf("expected source_id error, got: %v", err)
+	}
+}
+
+func TestRenderContainsSections(t *testing.T) {
+	in := validInput()
+	out := render(&in)
+
+	sections := []string{
+		"# PI 规划架构蓝图与任务清单",
+		"## 1. PI 愿景",
+		"## 2. 业务特性清单 (Features)",
+		"## 3. 技术赋能特性 (Enablers / 架构跑道)",
+		"## 4. 非功能性需求 (NFRs)",
+		"## 5. 依赖关系",
+		"## 6. 建议执行顺序",
+		"## 7. 质量门禁检查清单",
+	}
+	for _, s := range sections {
+		if !strings.Contains(out, s) {
+			t.Errorf("output missing section: %s", s)
+		}
+	}
+}
+
+func TestRenderContainsFeatureDetails(t *testing.T) {
+	in := validInput()
+	out := render(&in)
+
+	if !strings.Contains(out, "FEAT_OTA_001") {
+		t.Error("output missing FEAT_OTA_001")
+	}
+	if !strings.Contains(out, "云端固件版本依赖检查") {
+		t.Error("output missing feature title")
+	}
+	if !strings.Contains(out, "AC-1") {
+		t.Error("output missing acceptance criteria numbering")
+	}
+}
+
+func TestRenderContainsEnablerTable(t *testing.T) {
+	in := validInput()
+	out := render(&in)
+
+	if !strings.Contains(out, "ENAB_KAFKA_001") {
+		t.Error("output missing ENAB_KAFKA_001")
+	}
+	if !strings.Contains(out, "ENAB_S3_001") {
+		t.Error("output missing ENAB_S3_001")
+	}
+}
+
+func TestRenderContainsNFRs(t *testing.T) {
+	in := validInput()
+	out := render(&in)
+
+	if !strings.Contains(out, "P99 < 200ms") {
+		t.Error("output missing performance NFR")
+	}
+	if !strings.Contains(out, "TLS 1.3") {
+		t.Error("output missing security NFR")
+	}
+}
+
+func TestRenderContainsDependencies(t *testing.T) {
+	in := validInput()
+	out := render(&in)
+
+	if !strings.Contains(out, "ENAB_KAFKA_001") || !strings.Contains(out, "FEAT_OTA_001") {
+		t.Error("output missing dependency pair")
+	}
+}
+
+func TestComputeExecutionOrder(t *testing.T) {
+	in := validInput()
+	order := computeExecutionOrder(&in)
+
+	// Enablers should come before their dependent Features
+	enablerIdx := map[string]int{}
+	featureIdx := map[string]int{}
+	for i, id := range order {
+		if strings.HasPrefix(id, "ENAB_") {
+			enablerIdx[id] = i
+		} else if strings.HasPrefix(id, "FEAT_") {
+			featureIdx[id] = i
+		}
+	}
+
+	if enablerIdx["ENAB_KAFKA_001"] >= featureIdx["FEAT_OTA_001"] {
+		t.Error("ENAB_KAFKA_001 should come before FEAT_OTA_001")
+	}
+	if enablerIdx["ENAB_S3_001"] >= featureIdx["FEAT_OTA_002"] {
+		t.Error("ENAB_S3_001 should come before FEAT_OTA_002")
+	}
+}
+
+func TestComputeExecutionOrderNoDeps(t *testing.T) {
+	in := validInput()
+	in.Dependencies = nil
+	order := computeExecutionOrder(&in)
+
+	if len(order) != 4 {
+		t.Errorf("expected 4 items, got %d", len(order))
+	}
+	// Enablers first, then Features (stable order)
+	if order[0] != "ENAB_KAFKA_001" || order[1] != "ENAB_S3_001" {
+		t.Errorf("enablers should come first, got: %v", order)
+	}
+}
+
+func TestCallEndToEnd(t *testing.T) {
+	tool := New(0, nil)
+	in := validInput()
+	data, _ := json.Marshal(in)
+
+	result, err := tool.Call(context.Background(), string(data))
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !strings.Contains(result, "PI 规划架构蓝图") {
+		t.Error("output missing title")
+	}
+}
+
+func TestCallInvalidInput(t *testing.T) {
+	tool := New(0, nil)
+	_, err := tool.Call(context.Background(), "not json")
+	if err == nil {
+		t.Fatal("expected error for invalid input")
+	}
+}
+
+func TestCallMissingRequiredField(t *testing.T) {
+	tool := New(0, nil)
+	in := validInput()
+	in.PIVision = ""
+	data, _ := json.Marshal(in)
+
+	_, err := tool.Call(context.Background(), string(data))
+	if err == nil {
+		t.Fatal("expected validation error")
+	}
+}
+
+func TestNameAndDescription(t *testing.T) {
+	tool := New(0, nil)
+	if tool.Name() != "publish_pi_plan" {
+		t.Errorf("unexpected name: %s", tool.Name())
+	}
+	if tool.Description() == "" {
+		t.Error("description should not be empty")
+	}
+}
+
+func TestMaxOutputTruncation(t *testing.T) {
+	tool := New(100, nil)
+	in := validInput()
+	data, _ := json.Marshal(in)
+
+	result, err := tool.Call(context.Background(), string(data))
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(result) > 100 {
+		t.Errorf("output should be truncated to 100 chars, got %d", len(result))
+	}
+}