diff --git a/runtime/llm/anthropic.go b/runtime/llm/anthropic.go
index 931156a..fa2f2ac 100644
--- a/runtime/llm/anthropic.go
+++ b/runtime/llm/anthropic.go
@@ -284,7 +284,7 @@ func parseAnthropicSSE(ctx context.Context, r io.Reader, out chan<- Event) {
 				}
 			case "content_block_start":
 				var v struct {
-					Index       int `json:"index"`
+					Index        int `json:"index"`
 					ContentBlock struct {
 						Type string `json:"type"`
 						ID   string `json:"id"`
diff --git a/runtime/llm/client.go b/runtime/llm/client.go
index 54c29b5..9783bbc 100644
--- a/runtime/llm/client.go
+++ b/runtime/llm/client.go
@@ -121,6 +121,18 @@ func (c *Client) run(ctx context.Context, model string, messages []Message, tool
 	}
 }
 
+// replayThink 拼接一条 assistant 消息捕获到的全部思考块。
+func replayThink(m Message) string {
+	var b strings.Builder
+	for _, t := range m.Thinking {
+		if t.Thinking != "" {
+			b.WriteString(t.Thinking)
+			b.WriteString("\n")
+		}
+	}
+	return strings.TrimRight(b.String(), "\n")
+}
+
 func toProviderMessages(messages []Message) []map[string]any {
 	out := make([]map[string]any, 0, len(messages))
 	for _, m := range messages {
@@ -132,7 +144,14 @@ func toProviderMessages(messages []Message) []map[string]any {
 				"content":      m.Content,
 			})
 		case RoleAssistant:
-			item := map[string]any{"role": "assistant", "content": m.Content}
+			// 回放捕获的思考（Message.Thinking 由 SSE 的 reasoning_content 捕获而来）：
+			// 交错思考模型要求历史 assistant 原样带回 <think> 块。没有思考块的
+			// 消息零改动——非思考模型的上下文不受影响。
+			content := m.Content
+			if think := replayThink(m); think != "" {
+				content = "<think>\n" + think + "\n</think>\n\n" + content
+			}
+			item := map[string]any{"role": "assistant", "content": content}
 			if len(m.ToolCalls) > 0 {
 				calls := make([]map[string]any, 0, len(m.ToolCalls))
 				for _, tc := range m.ToolCalls {
diff --git a/runtime/llm/sse.go b/runtime/llm/sse.go
index 92a367b..82ee100 100644
--- a/runtime/llm/sse.go
+++ b/runtime/llm/sse.go
@@ -52,6 +52,7 @@ func ParseSSE(ctx context.Context, r io.Reader) <-chan Event {
 
 func emitParsed(ctx context.Context, r io.Reader, out chan<- Event) {
 	var content strings.Builder
+	var reasoning strings.Builder
 	calls := map[int]*ToolCall{}
 	order := []int{}
 	var usage Usage
@@ -108,8 +109,12 @@ func emitParsed(ctx context.Context, r io.Reader, out chan<- Event) {
 		}
 		// reasoning_content（国产 OpenAI 兼容模型的思考链）与正文分流：
 		// 只作为 reasoning_delta 事件外发，不计入 Message.Content（与 runner
-		// stripThinkBlocks"正文无 think"语义对齐）。
+		// stripThinkBlocks"正文无 think"语义对齐）。同时累计进 reasoning——
+		// 交错思考模型（MiniMax/DeepSeek R1 等）要求历史 assistant 原样带回
+		// 思考内容，丢弃会降低后续轮质量；捕获进 Message.Thinking 随会话
+		// 转录持久化，请求序列化侧按需回放。
 		if choice.Delta.Reasoning != "" {
+			reasoning.WriteString(choice.Delta.Reasoning)
 			select {
 			case <-ctx.Done():
 				emitAbort(out, content.String(), ctx.Err())
@@ -166,6 +171,9 @@ func emitParsed(ctx context.Context, r io.Reader, out chan<- Event) {
 		order = nil
 	}
 	msg := Message{Role: RoleAssistant, Content: content.String(), StopReason: mapFinish(finished, len(order) > 0)}
+	if reasoning.Len() > 0 {
+		msg.Thinking = []ThinkingBlock{{Thinking: reasoning.String()}}
+	}
 	for _, idx := range order {
 		msg.ToolCalls = append(msg.ToolCalls, *calls[idx])
 	}
diff --git a/runtime/llm/think_replay_test.go b/runtime/llm/think_replay_test.go
new file mode 100644
index 0000000..b8c4cc7
--- /dev/null
+++ b/runtime/llm/think_replay_test.go
@@ -0,0 +1,45 @@
+package llm
+
+import (
+	"strings"
+	"testing"
+)
+
+// 交错思考模型（MiniMax 等）要求历史 assistant 原样带回思考内容：
+// SSE 把 reasoning_content 捕获进 Message.Thinking；序列化把它作为
+// <think> 前缀回放。没有思考块的消息保持零改动。
+func TestReasoningCaptureAndReplay(t *testing.T) {
+	// 回放侧。
+	msgs := []Message{
+		{Role: RoleUser, Content: "总结"},
+		{Role: RoleAssistant, Content: "结论。", Thinking: []ThinkingBlock{{Thinking: "先读材料。"}}},
+		{Role: RoleAssistant, Content: "没有思考的轮次。"},
+	}
+	out := toProviderMessages(msgs)
+	if got := out[1]["content"].(string); !strings.HasPrefix(got, "<think>\n先读材料。\n</think>") || !strings.HasSuffix(got, "结论。") {
+		t.Fatalf("assistant replay malformed: %q", got)
+	}
+	if got := out[2]["content"].(string); got != "没有思考的轮次。" {
+		t.Fatalf("non-thinking assistant must be untouched: %q", got)
+	}
+	if _, ok := out[0]["tool_calls"]; ok {
+		t.Fatal("user messages carry no tool_calls")
+	}
+	// 捕获侧：直接构造 SSE 流。
+	sse := "data: {\"choices\":[{\"delta\":{\"reasoning_content\":\"思考\"}}]}\n\n" +
+		"data: {\"choices\":[{\"delta\":{\"content\":\"正文\"}}]}\n\n" +
+		"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n" +
+		"data: [DONE]\n\n"
+	var end Event
+	for ev := range ParseSSE(t.Context(), strings.NewReader(sse)) {
+		if ev.Kind == KindMessageEnd {
+			end = ev
+		}
+	}
+	if end.Message.Content != "正文" {
+		t.Fatalf("content polluted by reasoning: %q", end.Message.Content)
+	}
+	if len(end.Message.Thinking) != 1 || end.Message.Thinking[0].Thinking != "思考" {
+		t.Fatalf("reasoning not captured: %+v", end.Message.Thinking)
+	}
+}
diff --git a/runtime/llm/types.go b/runtime/llm/types.go
index 113bb8b..8d5ea83 100644
--- a/runtime/llm/types.go
+++ b/runtime/llm/types.go
@@ -52,9 +52,9 @@ type Message struct {
 	Images     []Image    `json:"images,omitempty"`
 	ToolCallID string     `json:"tool_call_id,omitempty"`
 	ToolCalls  []ToolCall `json:"tool_calls,omitempty"`
-	// Thinking 是原生 Anthropic 路径的思考块（含签名），随 assistant 消息
-	// 在请求里原样往返（Anthropic 要求思考块在内容块最前且签名不失）。
-	// OpenAI 路径不写这个字段（reasoning 不落盘，见 sse.go 注释）。
+	// Thinking 是思考块：Anthropic 原生路径含签名、在请求里原样往返；
+	// OpenAI 兼容路径由 SSE 的 reasoning_content 捕获（不含签名），随会话
+	// 转录持久化，请求序列化时按 <think> 块回放（见 client.go replayThink）。
 	Thinking   []ThinkingBlock `json:"thinking,omitempty"`
 	StopReason StopReason      `json:"stop_reason,omitempty"`
 	IsError    bool            `json:"is_error,omitempty"`