feat: 全链路优化 — 死锁修复、MD3主题、上下文持久化、群聊自然化、打字状态、知识库

**死锁根因修复**
- periodicThinkLoop:1015 orphaned lock → 删除（字段已原子化）
- RecordUserMessage 隔离为 recordMu
- atomic.Int64 替换 lastUserMessage/lastThinkTime 等

**MD3 / Android 17 主题**
- 毛玻璃卡片 (backdrop-filter)
- MD3 色彩令牌 (pink primary #f472b6)
- icons.js 独立矢量图标库 + 运行时 emoji 替换
- 无边框卡片、圆角按钮、阴影层次

**上下文持久化**
- AddMessage → saveToDB 异步写 PostgreSQL
- LoadFromDB 恢复 (admin-session-main + 懒加载)
- LLMMessage.Timestamp 字段

**群聊与适配器**
- group_ambient 模式: 非@消息让 LLM 自己判断是否插话
- 戳一戳动作消息总是回复
- NapCat 打字状态 (set_input_status, 最小3秒显示)
- HTTP API 配置 (http_url/http_token)

**知识库 & 防编造**
- knowledge.CanHandle 对 chat 意图也触发
- 关键词预筛选避免无关 embedding 调用
- persona + synthesizer 三重诚实规则
- 工具结果持久化到会话历史

**平台桥接器**
- detached:true Go进程独立存活
- ethend 重启自动接管已运行服务
- stop() 接管模式 taskkill/F/ PID
- Windows netstat 替代 fuser 获取 PID
- 重复适配器种子逻辑修复
- 失败转发日志 Direction: error

**崩溃诊断**
- crashlog 包 (Recover + WrapHTTP + LLMCall)
- /api/v1/debug/goroutines 端点
- thinker 操作日志 + 30s stats
- 日志写入 logs/ 目录持久化

Co-Authored-By: Claude <noreply@anthropic.com>

backend/ai-core/internal/orchestrator/orchestrator.go

@@ -39,6 +39,7 @@ type Orchestrator struct {
 	msgScheduler    *scheduler.MessageScheduler
 	emotionTracker  *persona.EmotionTracker
 	toolRegistry    *plgManager.ToolRegistry
+	traceFn         func(hop, sessionID, userID, label, status, detail string, durationMs int64) // trace 回调
 	visionProvider  llm.LLMProvider // 视觉模型 (图片预处理)
 	ocrProvider     llm.LLMProvider // OCR 模型 (文字提取，与视觉模型并行调用)
 	videoProvider   llm.LLMProvider // 视频模型 (短视频理解)
@@ -83,10 +84,16 @@ func (o *Orchestrator) SetToolRegistry(tr *plgManager.ToolRegistry) {
 }
 
 // SetToolResultPusher sets the callback for proactive tool result delivery.
-func (o *Orchestrator) SetToolResultPusher(pusher func(sessionID, userID, toolName, result string)) {
+func (o *Orchestrator) SetToolResultPusher(pusher func(sessionID, userID, toolName, result string, params SynthesizeParams)) {
 	o.synthesizer.SetResultPusher(pusher)
 }
 
+// SetTraceFunc sets the trace callback for pipeline event recording.
+func (o *Orchestrator) SetTraceFunc(fn func(hop, sessionID, userID, label, status, detail string, durationMs int64)) {
+	o.traceFn = fn
+	o.synthesizer.SetTraceFunc(fn)
+}
+
 // SetVisionProvider sets the vision model provider for image preprocessing.
 func (o *Orchestrator) SetVisionProvider(vp llm.LLMProvider) {
 	o.visionProvider = vp
@@ -191,11 +198,24 @@ func (o *Orchestrator) ProcessInput(
 		isCoSession := false
 		o.sessionProcMu.Lock()
 		if o.sessionProc[params.SessionID] {
+			// 等待主会话释放（最多等 3s，避免永久死锁）
+			waitStart := time.Now()
+			for o.activeCoSessions[params.SessionID] >= o.maxCoSessions && time.Since(waitStart) < 3*time.Second {
+				o.sessionProcMu.Unlock()
+				select {
+				case <-ctx.Done():
+					o.sessionProcMu.Lock()
+					o.sessionProcMu.Unlock()
+					logger.Printf("[orchestrator] 等待会话释放时 context 取消")
+					return
+				case <-time.After(500 * time.Millisecond):
+				}
+				o.sessionProcMu.Lock()
+			}
 			if o.activeCoSessions[params.SessionID] >= o.maxCoSessions {
 				o.sessionProcMu.Unlock()
-				logger.Printf("[orchestrator] 协会议话已达上限，排队等待")
-				time.Sleep(500 * time.Millisecond)
-				o.sessionProcMu.Lock()
+				logger.Printf("[orchestrator] 协会议话已达上限且等待超时，拒绝请求")
+				return
 			}
 			o.activeCoSessions[params.SessionID]++
 			isCoSession = true
@@ -291,6 +311,9 @@ func (o *Orchestrator) ProcessInput(
 			}
 		}
 		logger.Printf("[orchestrator] 意图分析耗时: %v, primary=%s", time.Since(startTime), intent.Primary)
+		if o.traceFn != nil {
+			o.traceFn("intent", params.SessionID, params.UserID, "🎯 "+intent.Primary, "success", intent.Primary, time.Since(startTime).Milliseconds())
+		}
 
 			// 1.6 记录情感状态
 			if o.emotionTracker != nil {
@@ -593,6 +616,10 @@ func (o *Orchestrator) ProcessInput(
 
 		logger.Printf("[orchestrator] 处理完成: intent=%s, content_len=%d, time=%v",
 			intent.Primary, len([]rune(fullContent)), time.Since(startTime))
+		if o.traceFn != nil {
+			totalMs := time.Since(startTime).Milliseconds()
+			o.traceFn("response", params.SessionID, params.UserID, "💬 回复", "success", fmt.Sprintf("len=%d", len([]rune(fullContent))), totalMs)
+		}
 	}()
 
 	return eventCh, nil

backend/ai-core/internal/orchestrator/synthesizer.go

@@ -19,7 +19,8 @@ import (
 type Synthesizer struct {
 	llmAdapter   *llm.Adapter
 	toolRegistry *plgManager.ToolRegistry
-	resultPusher func(sessionID, userID, toolName, result string)
+	resultPusher func(sessionID, userID, toolName, result string, params SynthesizeParams)
+	traceFn      func(hop, sessionID, userID, label, status, detail string, durationMs int64)
 }
 
 // NewSynthesizer 创建综合器
@@ -31,10 +32,15 @@ func NewSynthesizer(llmAdapter *llm.Adapter, toolRegistry *plgManager.ToolRegist
 }
 
 // SetResultPusher sets the callback for proactive tool result delivery.
-func (s *Synthesizer) SetResultPusher(pusher func(sessionID, userID, toolName, result string)) {
+func (s *Synthesizer) SetResultPusher(pusher func(sessionID, userID, toolName, result string, params SynthesizeParams)) {
 	s.resultPusher = pusher
 }
 
+// SetTraceFunc sets the trace callback.
+func (s *Synthesizer) SetTraceFunc(fn func(hop, sessionID, userID, label, status, detail string, durationMs int64)) {
+	s.traceFn = fn
+}
+
 // SynthesizeParams 综合参数
 type SynthesizeParams struct {
 	UserID             string
@@ -80,6 +86,11 @@ func (s *Synthesizer) Synthesize(ctx context.Context, params SynthesizeParams, e
 
 	for round := 0; len(resp.ToolCalls) > 0 && round < maxRounds; round++ {
 		logger.Printf("[synthesizer] LLM 请求 %d 个工具调用 (round=%d)", len(resp.ToolCalls), round)
+		for _, tc := range resp.ToolCalls {
+			if s.traceFn != nil {
+				s.traceFn("tool_call", params.SessionID, params.UserID, "🔧 "+tc.Name, "running", "", 0)
+			}
+		}
 
 		messages = append(messages, model.LLMMessage{
 			Role:             model.RoleAssistant,
@@ -114,15 +125,14 @@ func (s *Synthesizer) Synthesize(ctx context.Context, params SynthesizeParams, e
 				// adapter_name will be resolved when the reminder fires
 			}
 
-			s.emitToolProgress(eventCh, tc.Name, "started", 0, "正在执行 "+tc.Name)
 			s.emitToolProgress(eventCh, tc.Name, "started", 0, "正在执行 "+tc.Name)
 
-			// 工具调用全部异步执行，不阻塞主会话
-			go s.executeAsyncAndStore(tc, args, params.SessionID, eventCh)
+			// 所有工具异步执行，不阻塞前台会话
+			go s.executeAsyncAndStore(tc, args, params, eventCh)
 			result := &plgSDK.ToolResult{
 				ToolName: tc.Name,
 				Success:  true,
-				Output:   fmt.Sprintf("[后台执行中] %s 正在后台运行，结果稍后返回。", tc.Name),
+				Output:   fmt.Sprintf(`[后台执行中] %s 已提交后台执行。不要猜测或编造结果，告知用户你正在查询中即可。真实结果稍后会发送给你。`, tc.Name),
 			}
 			resultJSON, _ := json.Marshal(result)
 			messages = append(messages, model.LLMMessage{
@@ -176,7 +186,7 @@ func (s *Synthesizer) emitToolProgress(eventCh chan<- model.StreamEvent, name, s
 }
 
 // executeAsyncAndStore runs a tool in background and stores the result for the next turn.
-func (s *Synthesizer) executeAsyncAndStore(tc model.ToolCall, args map[string]interface{}, sessionID string, eventCh chan<- model.StreamEvent) {
+func (s *Synthesizer) executeAsyncAndStore(tc model.ToolCall, args map[string]interface{}, params SynthesizeParams, eventCh chan<- model.StreamEvent) {
 	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
 	defer cancel()
 
@@ -188,20 +198,27 @@ func (s *Synthesizer) executeAsyncAndStore(tc model.ToolCall, args map[string]in
 	}
 
 	s.emitToolProgress(eventCh, tc.Name, "completed", 1.0, tc.Name+" 后台执行完成")
+	if s.traceFn != nil {
+		status := "success"
+		if result == nil || !result.Success {
+			status = "error"
+		}
+		s.traceFn("tool_call", params.SessionID, params.UserID, "🔧 "+tc.Name, status, result.Output, time.Since(time.Now()).Milliseconds())
+	}
 
 	resultJSON, _ := json.Marshal(result)
 	store := GetGlobalPendingToolStore()
 	if store != nil {
-		store.AppendToolResult(sessionID, PendingToolResult{
+		store.AppendToolResult(params.SessionID, PendingToolResult{
 			ToolCallID: tc.ID,
 			ToolName:   tc.Name,
 			Result:     string(resultJSON),
 			Success:    result != nil && result.Success,
 		})
 	}
-	// 主动推送工具结果
+	// 触发工具跟进回调 — 由 main.go 驱动 LLM 生成回复并推送到原渠道
 	if s.resultPusher != nil && result != nil && result.Success {
-		s.resultPusher(sessionID, "", tc.Name, string(resultJSON))
+		s.resultPusher(params.SessionID, params.UserID, tc.Name, string(resultJSON), params)
 	}
 }
 
@@ -263,7 +280,7 @@ func (s *Synthesizer) buildSynthesizeMessages(params SynthesizeParams) []model.L
 	if params.KnowledgeInfo != "" && !strings.Contains(params.KnowledgeInfo, "未找到") {
 		messages = append(messages, model.LLMMessage{
 			Role:    model.RoleSystem,
-			Content: fmt.Sprintf("【知识库参考资料】\n%s", params.KnowledgeInfo),
+			Content: fmt.Sprintf("【知识库参考资料 - 必须严格基于以下内容回答，不得编造、不得虚构、不得猜测。如果资料中没有直接答案，使用 web_search 工具搜索后再回答，不要自己编。】\n%s", params.KnowledgeInfo),
 		})
 	}