diff --git a/backend/internal/eino/graph.go b/backend/internal/eino/graph.go
index f8a2449..8bf3a13 100644
--- a/backend/internal/eino/graph.go
+++ b/backend/internal/eino/graph.go
@@ -18,8 +18,9 @@ import (
 const (
 	nodeSTT      = "stt"
 	nodeHistory  = "history"
-	nodeLLM      = "llm"
-	nodeSplitter = "splitter"
+	nodeLLM            = "llm"
+	nodeMessageToString = "msg2str"
+	nodeSplitter       = "splitter"
 	nodeTTS      = "tts"
 	nodeDone     = "done"
 )
@@ -69,6 +70,7 @@ func NewPipelineGraph(
 	_ = g.AddLambdaNode(nodeSTT, NewSTTLambda(sttService))
 	_ = g.AddLambdaNode(nodeHistory, NewHistoryLambda(sessionMgr.GetHistory, maxHistory))
 	_ = g.AddChatModelNode(nodeLLM, chatModel)
+	_ = g.AddLambdaNode(nodeMessageToString, NewMessageToStringLambda())
 	_ = g.AddLambdaNode(nodeSplitter, NewSplitterLambda())
 	_ = g.AddLambdaNode(nodeTTS, NewTTSLambda(
 		ttsService,
@@ -83,7 +85,8 @@ func NewPipelineGraph(
 	_ = g.AddEdge(compose.START, nodeSTT)
 	_ = g.AddEdge(nodeSTT, nodeHistory)
 	_ = g.AddEdge(nodeHistory, nodeLLM)
-	_ = g.AddEdge(nodeLLM, nodeSplitter)
+	_ = g.AddEdge(nodeLLM, nodeMessageToString)
+	_ = g.AddEdge(nodeMessageToString, nodeSplitter)
 	_ = g.AddEdge(nodeSplitter, nodeTTS)
 	_ = g.AddEdge(nodeTTS, nodeDone)
 	_ = g.AddEdge(nodeDone, compose.END)
diff --git a/backend/internal/eino/nodes_history.go b/backend/internal/eino/nodes_history.go
index a558a6b..c48b488 100644
--- a/backend/internal/eino/nodes_history.go
+++ b/backend/internal/eino/nodes_history.go
@@ -40,30 +40,12 @@ func NewHistoryLambda(historyFetcher func(ctx context.Context, sessionID string,
 		scenarioPrompt := llm.GetScenarioPrompt(scenario, language)
 		systemPrompt := llm.BuildSystemPrompt(language, detailLevel, scenarioPrompt)
 
-		// 构建 system message（含图片）
+		// 构建 system message（仅文本，多模态内容只能放在 user 角色）
 		systemMsg := &schema.Message{
 			Role:    schema.System,
 			Content: systemPrompt,
 		}
 
-		// 如果有图片，添加到 system message 的多模态内容中
-		if len(imageData) > 0 {
-			base64Str := base64.StdEncoding.EncodeToString(imageData)
-			mimeType := detectImageMimeType(imageData)
-			systemMsg.UserInputMultiContent = []schema.MessageInputPart{
-				{
-					Type: schema.ChatMessagePartTypeImageURL,
-					Image: &schema.MessageInputImage{
-						MessagePartCommon: schema.MessagePartCommon{
-							Base64Data: &base64Str,
-							MIMEType:   mimeType,
-						},
-						Detail: schema.ImageURLDetailAuto,
-					},
-				},
-			}
-		}
-
 		messages := []*schema.Message{systemMsg}
 
 		// 获取并追加历史消息
@@ -81,11 +63,37 @@ func NewHistoryLambda(historyFetcher func(ctx context.Context, sessionID string,
 			}
 		}
 
-		// 追加当前用户输入
-		messages = append(messages, &schema.Message{
-			Role:    schema.User,
-			Content: sttOut.Text,
-		})
+		// 追加当前用户输入（含图片，多模态内容只能放在 user 角色）
+		// 注意：不能同时设置 Content 和 UserInputMultiContent，需要统一放到 MultiContent 中
+		if len(imageData) > 0 {
+			base64Str := base64.StdEncoding.EncodeToString(imageData)
+			mimeType := detectImageMimeType(imageData)
+			parts := []schema.MessageInputPart{
+				{
+					Type: schema.ChatMessagePartTypeText,
+					Text: sttOut.Text,
+				},
+				{
+					Type: schema.ChatMessagePartTypeImageURL,
+					Image: &schema.MessageInputImage{
+						MessagePartCommon: schema.MessagePartCommon{
+							Base64Data: &base64Str,
+							MIMEType:   mimeType,
+						},
+						Detail: schema.ImageURLDetailAuto,
+					},
+				},
+			}
+			messages = append(messages, &schema.Message{
+				Role:                schema.User,
+				UserInputMultiContent: parts,
+			})
+		} else {
+			messages = append(messages, &schema.Message{
+				Role:    schema.User,
+				Content: sttOut.Text,
+			})
+		}
 
 		log.Infow("历史组装完成",
 			"request_id", requestID,
diff --git a/backend/internal/eino/nodes_splitter.go b/backend/internal/eino/nodes_splitter.go
index 6f3de8b..e3cfafb 100644
--- a/backend/internal/eino/nodes_splitter.go
+++ b/backend/internal/eino/nodes_splitter.go
@@ -20,18 +20,49 @@ var sentenceDelimiters = map[rune]bool{
 	'?':  true,
 }
 
-// NewSplitterLambda 创建句子分割 Transform Lambda 节点。
-// 输入: StreamReader[string]（LLM 完整文本的单帧流）→ 输出: StreamReader[[]string]（句子数组流）
+// NewMessageToStringLambda 创建 Message → String 转换 Lambda 节点。
+// 输入: *schema.Message → 输出: string
 //
-// 在 Stream 模式下，框架自动将 ChatModel 的 StreamReader[*schema.Message]
-// concat 为 string 后传入此节点。此节点将文本按句子边界切分，
-// 每切出一个句子就输出一次，供 TTS 节点实时合成。
-func NewSplitterLambda() *compose.Lambda {
-	return compose.TransformableLambda(func(ctx context.Context, input *schema.StreamReader[string]) (*schema.StreamReader[[]string], error) {
-		sr, sw := schema.Pipe[[]string](8)
+// 提取 Message.Content 文本，供 Splitter 节点消费。
+func NewMessageToStringLambda() *compose.Lambda {
+	return compose.TransformableLambda(func(ctx context.Context, input *schema.StreamReader[*schema.Message]) (*schema.StreamReader[string], error) {
+		sr, sw := schema.Pipe[string](8)
 
 		go func() {
 			defer sw.Close()
+			defer input.Close()
+
+			for {
+				msg, err := input.Recv()
+				if err != nil {
+					if err == io.EOF {
+						return
+					}
+					sw.Send("", err)
+					return
+				}
+				if msg != nil && msg.Content != "" {
+					sw.Send(msg.Content, nil)
+				}
+			}
+		}()
+
+		return sr, nil
+	})
+}
+
+// NewSplitterLambda 创建句子分割 Transform Lambda 节点。
+// 输入: StreamReader[string]（LLM token 流）→ 输出: StreamReader[string]（完整句子流）
+//
+// 逐字符累积，按句子分隔符切分。每切出一个完整句子就输出一次，
+// 供下游 TTS 节点实时合成。
+func NewSplitterLambda() *compose.Lambda {
+	return compose.TransformableLambda(func(ctx context.Context, input *schema.StreamReader[string]) (*schema.StreamReader[string], error) {
+		sr, sw := schema.Pipe[string](8)
+
+		go func() {
+			defer sw.Close()
+			defer input.Close()
 
 			var buffer strings.Builder
 
@@ -43,23 +74,22 @@ func NewSplitterLambda() *compose.Lambda {
 						if buffer.Len() > 0 {
 							text := strings.TrimSpace(buffer.String())
 							if text != "" {
-								sw.Send([]string{text}, nil)
+								sw.Send(text, nil)
 							}
 						}
 						return
 					}
-					sw.Send(nil, err)
+					sw.Send("", err)
 					return
 				}
 
-				// chunk 是 concat 后的完整文本（单帧流）
 				// 逐字符累积，按句子分隔符切分
 				for _, r := range chunk {
 					buffer.WriteRune(r)
 					if sentenceDelimiters[r] {
 						text := strings.TrimSpace(buffer.String())
 						if text != "" {
-							sw.Send([]string{text}, nil)
+							sw.Send(text, nil)
 						}
 						buffer.Reset()
 					}
diff --git a/backend/internal/eino/nodes_tts.go b/backend/internal/eino/nodes_tts.go
index 5d4e7b7..09c86d1 100644
--- a/backend/internal/eino/nodes_tts.go
+++ b/backend/internal/eino/nodes_tts.go
@@ -3,90 +3,114 @@ package eino
 import (
 	"context"
 	"encoding/base64"
+	"io"
 
 	"github.com/cloudwego/eino/compose"
+	"github.com/cloudwego/eino/schema"
 
 	"github.com/hhs/camtalk/internal/ai/tts"
 	"github.com/hhs/camtalk/internal/logger"
 	"github.com/hhs/camtalk/internal/models"
 )
 
-// NewTTSLambda 创建 TTS Lambda 节点。
-// 输入: []string（句子数组，框架自动从 StreamReader concat）→ 输出: struct{}
+// NewTTSLambda 创建 TTS Transform Lambda 节点。
+// 输入: StreamReader[string]（句子流）→ 输出: StreamReader[struct{}]（结果流）
 //
-// 将句子数组转为 channel，调用 ttsService.SynthesizeStream() 流式合成，
+// 流式消费每个句子，调用 ttsService.SynthesizeStream() 合成，
 // 逐 chunk 推送 tts_audio 到客户端。TTS 失败静默跳过。
 func NewTTSLambda(ttsService tts.Service, ttsVoice string, ttsSpeed float64, ttsOutputFmt string, ttsSampleRate int) *compose.Lambda {
-	return compose.InvokableLambda(func(ctx context.Context, sentences []string) (struct{}, error) {
-		log := logger.Log
-		sender := senderFromCtx(ctx)
-		requestID := requestIDFromCtx(ctx)
-		state := stateFromCtx(ctx)
+	return compose.TransformableLambda(func(ctx context.Context, input *schema.StreamReader[string]) (*schema.StreamReader[struct{}], error) {
+		sr, sw := schema.Pipe[struct{}](8)
 
-		// 检查 TTS 是否启用（从 State 或 context 获取）
-		// TTSEnabled 信息在 PipelineInput 中，通过 State 传递
-		if state != nil {
-			state.mu.Lock()
-			ttsEnabled := true // 默认启用，由适配器通过 State 设置
-			state.mu.Unlock()
-			if !ttsEnabled {
-				return struct{}{}, nil
-			}
-		}
+		go func() {
+			defer sw.Close()
+			defer input.Close()
 
-		if len(sentences) == 0 {
-			return struct{}{}, nil
-		}
+			log := logger.Log
+			sender := senderFromCtx(ctx)
+			requestID := requestIDFromCtx(ctx)
 
-		if sender == nil || requestID == "" {
-			return struct{}{}, nil
-		}
-
-		log.Infow("开始 TTS 合成", "request_id", requestID, "sentence_count", len(sentences))
-
-		// 将句子数组转为 channel（ttsService.SynthesizeStream 需要 <-chan string）
-		sentenceCh := make(chan string, len(sentences))
-		for _, s := range sentences {
-			sentenceCh <- s
-		}
-		close(sentenceCh)
-
-		// 调用 TTS 服务
-		ttsStream, err := ttsService.SynthesizeStream(ctx, sentenceCh, tts.Options{
-			Voice:      ttsVoice,
-			Speed:      ttsSpeed,
-			OutputFmt:  ttsOutputFmt,
-			SampleRate: ttsSampleRate,
-		})
-		if err != nil {
-			log.Errorw("TTS 合成启动失败（已跳过）", "error", err, "request_id", requestID)
-			return struct{}{}, nil // TTS 失败不中断流程
-		}
-
-		// 消费 TTS 音频流，推送到客户端
-		for chunk := range ttsStream {
-			select {
-			case <-ctx.Done():
-				log.Infow("TTS 流被中断", "request_id", requestID)
-				return struct{}{}, ctx.Err()
-			default:
+			if sender == nil || requestID == "" {
+				// 消费并丢弃流
+				for {
+					_, err := input.Recv()
+					if err != nil {
+						return
+					}
+				}
 			}
 
-			audioBase64 := base64.StdEncoding.EncodeToString(chunk.Audio)
-
-			if err := sender.SendTTSAudio(models.WsTTSAudio{
-				Type:      "tts_audio",
-				RequestID: requestID,
-				Audio:     audioBase64,
-				MimeType:  "audio/mp3",
-				IsLast:    chunk.IsLast,
-				Final:     chunk.Final,
-			}); err != nil {
-				log.Errorw("发送 tts_audio 失败", "error", err)
+			// 收集句子，按批次合成 TTS
+			var sentences []string
+			for {
+				sentence, err := input.Recv()
+				if err != nil {
+					if err == io.EOF {
+						break
+					}
+					log.Errorw("TTS: stream recv error", "error", err, "request_id", requestID)
+					break
+				}
+				if sentence != "" {
+					sentences = append(sentences, sentence)
+				}
 			}
-		}
 
-		log.Infow("TTS 合成完成", "request_id", requestID)
-		return struct{}{}, nil
+			if len(sentences) == 0 {
+				sw.Send(struct{}{}, nil)
+				return
+			}
+
+			log.Infow("开始 TTS 合成", "request_id", requestID, "sentence_count", len(sentences))
+
+			// 将句子数组转为 channel
+			sentenceCh := make(chan string, len(sentences))
+			for _, s := range sentences {
+				sentenceCh <- s
+			}
+			close(sentenceCh)
+
+			// 调用 TTS 服务
+			ttsStream, err := ttsService.SynthesizeStream(ctx, sentenceCh, tts.Options{
+				Voice:      ttsVoice,
+				Speed:      ttsSpeed,
+				OutputFmt:  ttsOutputFmt,
+				SampleRate: ttsSampleRate,
+			})
+			if err != nil {
+				log.Errorw("TTS 合成启动失败（已跳过）", "error", err, "request_id", requestID)
+				sw.Send(struct{}{}, nil)
+				return
+			}
+
+			// 消费 TTS 音频流，推送到客户端
+			for chunk := range ttsStream {
+				select {
+				case <-ctx.Done():
+					log.Infow("TTS 流被中断", "request_id", requestID)
+					sw.Send(struct{}{}, ctx.Err())
+					return
+				default:
+				}
+
+				audioBase64 := base64.StdEncoding.EncodeToString(chunk.Audio)
+
+				if err := sender.SendTTSAudio(models.WsTTSAudio{
+					Type:      "tts_audio",
+					RequestID: requestID,
+					Audio:     audioBase64,
+					MimeType:  "audio/mp3",
+					IsLast:    chunk.IsLast,
+					Final:     chunk.Final,
+				}); err != nil {
+					log.Errorw("发送 tts_audio 失败", "error", err)
+				}
+			}
+
+			log.Infow("TTS 合成完成", "request_id", requestID)
+			sw.Send(struct{}{}, nil)
+		}()
+
+		return sr, nil
 	})
 }