fix: 后台思考使用深度思考模型 + 主动消息推送冷却优化

- thinker.go: Round 0 优先调用 llmAdapter(deepseek-v4-pro)，失败回退 toolAdapter
- thinker.go: RecordUserMessage 重置 lastProactiveMsgTime，活跃对话中允许推送
- proactive_decision.go: MinGap low 30→15min, medium 10→5min, high 2→1min; 小时上限 3→5

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/ai-core/internal/background/proactive_decision.go

@@ -36,11 +36,11 @@ func DefaultProactiveGuard() *ProactiveGuard {
 		QuietHoursStart: 23,
 		QuietHoursEnd:   7,
 		MinGapByUrgency: map[string]time.Duration{
-			"low":    30 * time.Minute,
-			"medium": 10 * time.Minute,
-			"high":   2 * time.Minute,
+			"low":    15 * time.Minute,
+			"medium": 5 * time.Minute,
+			"high":   1 * time.Minute,
 		},
-		MaxMessagesPerHour: 3,
+		MaxMessagesPerHour: 5,
 	}
 }
 

backend/ai-core/internal/background/thinker.go

@@ -323,6 +323,8 @@ func (t *Thinker) RecordUserMessage(sessionID string) {
 	if sessionID != "" {
 		t.activeSessionID = sessionID
 	}
+	// 用户主动发消息时重置主动消息推送冷却——活跃对话中应允许昔涟回复
+	t.lastProactiveMsgTime = time.Time{}
 	t.mu.Unlock()
 
 	// 重置静默检测定时器
@@ -594,69 +596,88 @@ func (t *Thinker) performThink(triggerReason string) {
 	// 6. 准备工具定义（通过自主工具策略过滤）
 	openAITools := t.filterToolsByPolicy(t.buildOpenAITools())
 
-	// 7. 调用 LLM（支持工具调用，策略限制轮数）
+	// 7. 调用 LLM — 优先使用深度思考模型，工具阶段回退到工具模型
 	maxToolRounds := t.autoToolPolicy.MaxToolCallsPerRound
 	var finalContent string
 	var totalToolCalls int
 	var toolCallRecords []map[string]interface{}
 
-	for round := 0; round <= maxToolRounds; round++ {
-		resp, err := t.toolAdapter.ChatWithTools(ctx, messages, openAITools)
-		if err != nil {
-			log.Printf("[后台思考] LLM调用失败 (round=%d): %v", round, err)
-			return
-		}
+	// Round 0: 深度思考模型（优先），失败时回退到工具模型
+	resp, err := t.llmAdapter.ChatWithTools(ctx, messages, openAITools)
+	if err != nil {
+		log.Printf("[后台思考] 深度思考模型调用失败，回退到工具模型: %v", err)
+		resp, err = t.toolAdapter.ChatWithTools(ctx, messages, openAITools)
+	}
+	if err != nil {
+		log.Printf("[后台思考] LLM调用失败: %v", err)
+		return
+	}
 
-		if len(resp.ToolCalls) == 0 {
-			finalContent = resp.Content
-			break
-		}
-
-		log.Printf("[后台思考] LLM 请求 %d 个工具调用 (round=%d)", len(resp.ToolCalls), round)
-
-		assistantMsg := model.LLMMessage{
-			Role:             model.RoleAssistant,
-			Content:          resp.Content,
-			ToolCalls:        resp.ToolCalls,
-			ReasoningContent: resp.ReasoningContent,
-		}
-		messages = append(messages, assistantMsg)
-
-		for _, tc := range resp.ToolCalls {
-			var args map[string]interface{}
-			if err := json.Unmarshal([]byte(tc.Arguments), &args); err != nil {
-				log.Printf("[后台思考] 工具 %s 参数解析失败: %v", tc.Name, err)
-				args = make(map[string]interface{})
+	if len(resp.ToolCalls) == 0 {
+		finalContent = resp.Content
+	} else {
+		// 深度思考模型请求了工具调用，进入工具执行循环
+		for round := 0; round <= maxToolRounds; round++ {
+			if round > 0 {
+				// 后续轮次使用工具模型
+				resp, err = t.toolAdapter.ChatWithTools(ctx, messages, openAITools)
+				if err != nil {
+					log.Printf("[后台思考] 工具模型调用失败 (round=%d): %v", round, err)
+					return
+				}
 			}
 
-			result, execErr := t.toolRegistry.Execute(ctx, tc.Name, args)
-			if execErr != nil {
-				log.Printf("[后台思考] 工具 %s 执行失败: %v", tc.Name, execErr)
-			}
-
-			resultJSON, _ := json.Marshal(result)
-			messages = append(messages, model.LLMMessage{
-				Role:       model.RoleTool,
-				Content:    string(resultJSON),
-				ToolCallID: tc.ID,
-			})
-
-			totalToolCalls++
-			toolCallRecords = append(toolCallRecords, map[string]interface{}{
-				"name": tc.Name,
-				"args": args,
-			})
-		}
-
-		if round == maxToolRounds {
-			finalResp, finalErr := t.llmAdapter.Chat(ctx, messages)
-			if finalErr != nil {
-				log.Printf("[后台思考] 最终总结调用失败: %v", finalErr)
+			if round > 0 && len(resp.ToolCalls) == 0 {
 				finalContent = resp.Content
-			} else {
-				finalContent = finalResp.Content
+				break
+			}
+
+			log.Printf("[后台思考] LLM 请求 %d 个工具调用 (round=%d)", len(resp.ToolCalls), round)
+
+			assistantMsg := model.LLMMessage{
+				Role:             model.RoleAssistant,
+				Content:          resp.Content,
+				ToolCalls:        resp.ToolCalls,
+				ReasoningContent: resp.ReasoningContent,
+			}
+			messages = append(messages, assistantMsg)
+
+			for _, tc := range resp.ToolCalls {
+				var args map[string]interface{}
+				if err := json.Unmarshal([]byte(tc.Arguments), &args); err != nil {
+					log.Printf("[后台思考] 工具 %s 参数解析失败: %v", tc.Name, err)
+					args = make(map[string]interface{})
+				}
+
+				result, execErr := t.toolRegistry.Execute(ctx, tc.Name, args)
+				if execErr != nil {
+					log.Printf("[后台思考] 工具 %s 执行失败: %v", tc.Name, execErr)
+				}
+
+				resultJSON, _ := json.Marshal(result)
+				messages = append(messages, model.LLMMessage{
+					Role:       model.RoleTool,
+					Content:    string(resultJSON),
+					ToolCallID: tc.ID,
+				})
+
+				totalToolCalls++
+				toolCallRecords = append(toolCallRecords, map[string]interface{}{
+					"name": tc.Name,
+					"args": args,
+				})
+			}
+
+			if round == maxToolRounds {
+				finalResp, finalErr := t.llmAdapter.Chat(ctx, messages)
+				if finalErr != nil {
+					log.Printf("[后台思考] 最终总结调用失败: %v", finalErr)
+					finalContent = resp.Content
+				} else {
+					finalContent = finalResp.Content
+				}
+				break
 			}
-			break
 		}
 	}