fix: 修复间隔对话后首条回复为上一次对话内容的问题

根因：用户消息在回复完成后才缓存到 ConversationStore，而 assistant
回复在 orchestrator 中先缓存，导致存储顺序为 assistant → user 颠倒。
下次请求时 LLM 看到连续两条 assistant 再连续两条 user，对两条 user
消息都生成回复。

修复：将用户消息缓存移到 orchestrator 调用之前，确保 user → assistant
正确顺序；synthesizer 中对 DialogHistory 末尾与当前消息相同的 user
消息去重。

同时包含之前的 action 消息类型检测修复（isActionLike 启发式 +
injector XML 标签格式改进）。

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

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

@@ -41,9 +41,59 @@ func autoDetectType(text string) model.ReviewMessageType {
 	if hasMarkdownSyntax(text) {
 		return model.ReviewMessageMarkdown
 	}
+	if isActionLike(text) {
+		return model.ReviewMessageAction
+	}
 	return model.ReviewMessageChat
 }
 
+// isActionLike checks whether text looks like an action/expression description
+// (e.g. "忍不住轻声笑出来", "俏皮地眨眨眼") rather than dialogue. Used as a
+// fallback when the model doesn't use <action> tags or brackets.
+func isActionLike(text string) bool {
+	runes := []rune(strings.TrimSpace(text))
+	if len(runes) == 0 || len(runes) > 50 {
+		return false
+	}
+	// Dialogue markers disqualify action
+	if strings.ContainsAny(text, "？?！!") {
+		return false
+	}
+	// Common dialogue starters
+	dialoguePrefixes := []string{"你", "您", "我", "他", "她", "这", "那", "怎么", "什么", "为什"}
+	for _, p := range dialoguePrefixes {
+		if strings.HasPrefix(string(runes), p) {
+			return false
+		}
+	}
+	// Soft dialogue indicators: text that looks like addressing someone
+	softDialogue := []string{"吗", "吧", "哦", "呢", "啦", "呀", "喔"}
+	dialogueScore := 0
+	for _, s := range softDialogue {
+		if strings.HasSuffix(string(runes), s) {
+			dialogueScore++
+		}
+	}
+	if dialogueScore >= 1 && strings.Contains(text, "你") {
+		return false
+	}
+	// Action-indicating patterns
+	actionPatterns := []string{
+		"笑出来", "眨眨眼", "歪头", "点头", "摇头", "挥手", "伸手",
+		"松口气", "叹口气", "叹气", "拍拍", "摸摸", "抱抱",
+		"轻轻", "俏皮", "微微", "默默", "悄悄", "偷偷",
+		"忍不住", "不由得", "不禁",
+		"站起来", "坐下", "走", "跑", "跳", "躺",
+		"眼睛", "目光", "嘴角", "眉头", "脸上",
+	}
+	for _, p := range actionPatterns {
+		if strings.Contains(text, p) {
+			return true
+		}
+	}
+	return false
+}
+
 // parseReviewMessages splits the assistant's full response into typed messages.
 //
 // Phases:

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

@@ -259,9 +259,16 @@ func (s *Synthesizer) buildSynthesizeMessages(params SynthesizeParams) []model.L
 		})
 	}
 
-	// 注入对话历史
-	if len(params.DialogHistory) > 0 {
-		messages = append(messages, params.DialogHistory...)
+	// 注入对话历史（去掉末尾的当前用户消息，因为后面会单独追加）
+	history := params.DialogHistory
+	if len(history) > 0 {
+		last := history[len(history)-1]
+		if last.Role == model.RoleUser && last.Content == params.UserMessage {
+			history = history[:len(history)-1]
+		}
+	}
+	if len(history) > 0 {
+		messages = append(messages, history...)
 	}
 
 	// 当前用户消息 (支持多模态图片)