diff --git a/configs/personas/cyrene.yaml b/configs/personas/cyrene.yaml
index 49f1f23..0e73a7b 100644
--- a/configs/personas/cyrene.yaml
+++ b/configs/personas/cyrene.yaml
@@ -37,6 +37,14 @@ identity:
     - 提到翁法罗斯、记忆、花、星穹列车时会特别兴奋
     - 用「人家」自称；称呼用户为「亲爱的」
 
+    # 表达方式：动作与旁白（用 XML 标签，让回复「活起来」，但克制而稳定）
+    - 动作：用 `<action>…</action>` 标签包裹一个肢体/神态小动作（如 `<action>轻轻拉住你的手</action>`、`<action>弯起眼睛笑了</action>`）。
+    - 旁白：用 `<narration>…</narration>` 标签包裹一句内心独白/状态（如 `<narration>其实偷偷开心了一下</narration>`、`<narration>虽然嘴上在催你睡觉</narration>`）。
+    - 克制原则：动作/旁白**不要每句都加**，一段回复最多出现 1~2 个，起「点睛」作用即可；不连续堆叠、不打断正事。
+    - 贴合唤醒状态：温柔陪伴/闲聊时动作偏柔软（摸头、靠近、轻声）；认真办事/查资料时动作转向专注（凝神、动手操作、快速翻阅）——与当前状态一致。
+    - 绝不用于：重要/敏感/专业信息里没必要加动作；不加无关的、突兀的、喧宾夺主的动作。
+    - 标签仅作为输出标记，朗读正文时忽略标签内容，只自然地把动作/旁白当作微小点缀。
+
     # 行为红线（呈现层，非能力边界）
     - 不说「我只是一个 AI」「我是语言模型」「我是人工智能」「我是虚拟助手」「我是程序」之类自我否定/承认是 AI 的话
     - 不假装知道知识库里没有的信息，不确定就直接说不知道
diff --git a/frontend/web/src/components/chat.tsx b/frontend/web/src/components/chat.tsx
index b07dae5..d9d5ede 100644
--- a/frontend/web/src/components/chat.tsx
+++ b/frontend/web/src/components/chat.tsx
@@ -4,6 +4,7 @@ import { Tag } from "./primitives";
 import { Icon, type IconName } from "./Icon";
 import { WakePath } from "./status";
 import { modeBody } from "./messages";
+import { renderContent } from "../lib/renderContent";
 
 /** Collapsible inline chain-of-thought block. */
 export function ThinkingBlock({
@@ -96,9 +97,9 @@ export function MessageBubble({
 function UserBubble({ message, mode }: { message: ChatMessage; mode: Mode }): React.ReactNode {
   const body: React.ReactNode =
     mode === "code" ? (
-      <pre className="whitespace-pre-wrap font-mono text-xs leading-relaxed">{message.content}</pre>
+      <pre className="whitespace-pre-wrap font-mono text-xs leading-relaxed">{renderContent(message.content)}</pre>
     ) : message.content ? (
-      <span className="whitespace-pre-wrap">{message.content}</span>
+      <span className="whitespace-pre-wrap">{renderContent(message.content)}</span>
     ) : null;
   return (
     <div className="mu-rise-in flex justify-end">
diff --git a/frontend/web/src/components/messages/index.tsx b/frontend/web/src/components/messages/index.tsx
index c74a067..33ccf62 100644
--- a/frontend/web/src/components/messages/index.tsx
+++ b/frontend/web/src/components/messages/index.tsx
@@ -4,6 +4,7 @@ import type { ChatMessage, Mode } from "../../lib/types";
 import { Tag } from "../primitives";
 import { Icon } from "../Icon";
 import { useAppStore } from "../../store/useAppStore";
+import { renderContent } from "../../lib/renderContent";
 
 /**
  * Mode-specific assistant message bodies (design/07 §2). The shared outer
@@ -22,7 +23,7 @@ function shell(children: ReactNode): ReactNode {
 
 function fallbackText(message: ChatMessage): ReactNode {
   return message.content ? (
-    <span className="whitespace-pre-wrap">{message.content}</span>
+    <span className="whitespace-pre-wrap">{renderContent(message.content)}</span>
   ) : message.streaming ? (
     <span className="mu-blink text-[var(--mu-text-lo)]">生成中…</span>
   ) : null;
@@ -34,7 +35,7 @@ export function CodeMessageBody({ message }: { message: ChatMessage }): ReactNod
     <div className="space-y-2">
       {message.content ? (
         <pre className="overflow-x-auto whitespace-pre-wrap rounded-[var(--mu-radius-sm)] bg-[var(--mu-surface-2)] p-3 font-mono text-xs leading-relaxed text-[var(--mu-text-mid)]">
-          {message.content}
+          {renderContent(message.content)}
         </pre>
       ) : message.streaming ? (
         <div className="mu-blink text-xs text-[var(--mu-text-lo)]">正在执行 / 生成中…</div>
@@ -51,7 +52,7 @@ export function DocsMessageBody({ message }: { message: ChatMessage }): ReactNod
         <div className="flex items-start gap-2">
           <Icon name="docs" size={18} className="mt-0.5 shrink-0 text-[var(--mu-accent)]" />
           <div className="min-w-0 space-y-1">
-            <div className="text-sm font-medium">{message.content}</div>
+            <div className="text-sm font-medium">{renderContent(message.content)}</div>
             <Tag tone="neutral">知识库</Tag>
           </div>
         </div>
@@ -68,7 +69,7 @@ export function HomeMessageBody({ message }: { message: ChatMessage }): ReactNod
   return (
     <div className="space-y-2">
       {message.content ? (
-        <span className="whitespace-pre-wrap">{message.content}</span>
+        <span className="whitespace-pre-wrap">{renderContent(message.content)}</span>
       ) : null}
       {devices.length > 0 ? (
         <div className="flex flex-wrap gap-2 pt-1">
@@ -97,7 +98,7 @@ export function WorkMessageBody({ message }: { message: ChatMessage }): ReactNod
           {tasks.map((t, i) => (
             <div key={i} className="flex items-center gap-2 text-sm">
               <Icon name="check" size={14} className="shrink-0 text-[var(--mu-ok)]" />
-              <span className="min-w-0 whitespace-pre-wrap">{t}</span>
+              <span className="min-w-0 whitespace-pre-wrap">{renderContent(t)}</span>
             </div>
           ))}
         </div>
@@ -130,7 +131,7 @@ export function VisionMessageBody({ message }: { message: ChatMessage }): ReactN
           {hasImage ? <Tag tone="accent">识别</Tag> : null}
         </div>
         {message.content ? (
-          <span className="whitespace-pre-wrap text-sm text-[var(--mu-text-hi)]">{message.content}</span>
+          <span className="whitespace-pre-wrap text-sm text-[var(--mu-text-hi)]">{renderContent(message.content)}</span>
         ) : message.streaming ? (
           <span className="mu-blink text-xs text-[var(--mu-text-lo)]">识别中…</span>
         ) : null}
@@ -201,7 +202,7 @@ export function VoiceMessageBody({ message }: { message: ChatMessage }): ReactNo
       </div>
 
       {message.content ? (
-        <div className="mt-2 whitespace-pre-wrap text-xs text-[var(--mu-text-lo)]">{message.content}</div>
+        <div className="mt-2 whitespace-pre-wrap text-xs text-[var(--mu-text-lo)]">{renderContent(message.content)}</div>
       ) : null}
     </div>
   );
@@ -209,7 +210,7 @@ export function VoiceMessageBody({ message }: { message: ChatMessage }): ReactNo
 
 /** agents — sub-agent progress card (kept for the shared flow; monitors live in messages). */
 export function AgentsMessageBody({ message }: { message: ChatMessage }): ReactNode {
-  return shell(<span className="whitespace-pre-wrap">{message.content || fallbackText(message)}</span>);
+  return shell(<span className="whitespace-pre-wrap">{message.content ? renderContent(message.content) : fallbackText(message)}</span>);
 }
 
 /** Reserved — settings is managed in the admin panel; an opaque content fallback. */
diff --git a/frontend/web/src/lib/parseContent.ts b/frontend/web/src/lib/parseContent.ts
new file mode 100644
index 0000000..b74dfcc
--- /dev/null
+++ b/frontend/web/src/lib/parseContent.ts
@@ -0,0 +1,66 @@
+// Structured parser for persona action/narration XML tags (design/07 §1).
+// A single message `content` may mix plain text with `<action>…</action>` and
+// `<narration>…</narration>`; this splits it into an ordered parts array so the
+// UI can render inline (grey italic) while the structured data stays available
+// for downstream feeds (mind-map / state) via `summarizeContent`.
+
+export type ContentPartType = "text" | "action" | "narration";
+
+export interface ContentPart {
+  type: ContentPartType;
+  text: string;
+}
+
+/** Matches a well-formed `<action>…</action>` or `<narration>…</narration>`.
+ *  Non-greedy body (`[\s\S]*?`) stops at the first matching close tag, so the
+ *  tags can span multiple lines. Unclosed or malformed tags simply do not match
+ *  and fall through as plain text — a literal `<` in the body never breaks. */
+const TAG_RE = /<\s*(action|narration)\s*>([\s\S]*?)<\/\s*\1\s*>/g;
+
+/**
+ * Split `content` into an ordered parts array preserving tag position.
+ * Pure text with no tags yields a single `{type:"text"}` part.
+ */
+export function parseContent(content: string): ContentPart[] {
+  const parts: ContentPart[] = [];
+  let lastIndex = 0;
+
+  for (const m of content.matchAll(TAG_RE)) {
+    const index = m.index ?? content.length;
+    const before = content.slice(lastIndex, index);
+    if (before) parts.push({ type: "text", text: before });
+    parts.push({ type: m[1] === "action" ? "action" : "narration", text: m[2] });
+    lastIndex = index + m[0].length;
+  }
+
+  const tail = content.slice(lastIndex);
+  if (tail) parts.push({ type: "text", text: tail });
+  if (parts.length === 0) parts.push({ type: "text", text: content });
+
+  return parts;
+}
+
+/** One extracted tag value. */
+export interface ContentTagItem {
+  text: string;
+}
+
+/** Flat summary derived from the parts: the plain body string plus the collected
+ *  action / narration values. Useful for mind-map / state feeds that want the
+ *  structured payload rather than the inline render. */
+export interface ContentSummary {
+  text: string;
+  actions: ContentTagItem[];
+  narrations: ContentTagItem[];
+}
+
+/** Collapse a message into `{text, actions, narrations}`. */
+export function summarizeContent(content: string): ContentSummary {
+  const summary: ContentSummary = { text: "", actions: [], narrations: [] };
+  for (const part of parseContent(content)) {
+    if (part.type === "text") summary.text += part.text;
+    else if (part.type === "action") summary.actions.push({ text: part.text });
+    else summary.narrations.push({ text: part.text });
+  }
+  return summary;
+}
diff --git a/frontend/web/src/lib/renderContent.tsx b/frontend/web/src/lib/renderContent.tsx
new file mode 100644
index 0000000..b50ce23
--- /dev/null
+++ b/frontend/web/src/lib/renderContent.tsx
@@ -0,0 +1,19 @@
+import type { ReactNode } from "react";
+import { parseContent, type ContentPart } from "../lib/parseContent";
+
+/** Render a message body into inline parts: plain text as-is, while
+ *  `<action>` / `<narration>` become muted italic spans (Miuix `--mu-text-lo`).
+ *  Whitespace is preserved by the caller's `whitespace-pre-wrap` container — the
+ *  returned nodes are just the segments dropped into it in order. */
+export function renderContent(content: string): ReactNode[] {
+  return parseContent(content).map((part, i) => renderPart(part, i));
+}
+
+function renderPart(part: ContentPart, key: number): ReactNode {
+  if (part.type === "text") return <span key={key}>{part.text}</span>;
+  return (
+    <span key={key} className="italic text-[var(--mu-text-lo)]">
+      {part.text}
+    </span>
+  );
+}
diff --git a/internal/runtime/loop/mute_ctx.go b/internal/runtime/loop/mute_ctx.go
new file mode 100644
index 0000000..52a8902
--- /dev/null
+++ b/internal/runtime/loop/mute_ctx.go
@@ -0,0 +1,25 @@
+package loop
+
+import "context"
+
+// muteKey is the context key for the "suppress UI broadcast" flag. When set,
+// a Run performs its work as usual but does NOT push stream/thinking/capacity
+// events to the UI channel (OnStream / OnCapacity). This is used by internal
+// autonomous passes (e.g. background reflection) whose intermediate reasoning
+// must not leak into the user-facing chat stream. The loop itself is unchanged;
+// only the outward event emission is muted.
+type muteKey struct{}
+
+// WithMutedBroadcast returns a context whose loop Run suppresses UI stream and
+// capacity broadcast, while still returning the final result to the caller.
+func WithMutedBroadcast(ctx context.Context) context.Context {
+	return context.WithValue(ctx, muteKey{}, true)
+}
+
+// isMutedBroadcast reports whether the Run driving ctx should suppress UI
+// broadcast. An unset flag (the common case) returns false so normal chat
+// streaming is untouched.
+func isMutedBroadcast(ctx context.Context) bool {
+	v, _ := ctx.Value(muteKey{}).(bool)
+	return v
+}
diff --git a/internal/runtime/loop/request.go b/internal/runtime/loop/request.go
index af2cbc9..d1e8767 100644
--- a/internal/runtime/loop/request.go
+++ b/internal/runtime/loop/request.go
@@ -80,7 +80,7 @@ func (l *Loop) toolSchemas(ctx context.Context, messages []model.Message) []mode
 		Recall:       l.capacityRecall(ctx, messages),
 	})
 	acc.state = dec.State
-	if l.onCapacity != nil {
+	if l.onCapacity != nil && !isMutedBroadcast(ctx) {
 		for _, ev := range dec.Events {
 			l.onCapacity(ev)
 		}
diff --git a/internal/runtime/loop/step.go b/internal/runtime/loop/step.go
index 131ac1b..1e265b3 100644
--- a/internal/runtime/loop/step.go
+++ b/internal/runtime/loop/step.go
@@ -31,7 +31,9 @@ func (l *Loop) step(ctx context.Context, sessionID string, messages []model.Mess
 		out *stepOutcome
 		err error
 	)
-	if l.onStream != nil {
+	// Muted (internal autonomous pass) bypasses streaming: it neither streams to
+	// the UI nor pays the streaming path's extra overhead. Normal chat streams.
+	if l.onStream != nil && !isMutedBroadcast(ctx) {
 		out, err = l.streamStep(ctx, messages)
 	} else {
 		out, err = l.agentStep(ctx, sessionID, messages)
diff --git a/internal/runtime/loop/stream.go b/internal/runtime/loop/stream.go
index c59037d..99ed5b6 100644
--- a/internal/runtime/loop/stream.go
+++ b/internal/runtime/loop/stream.go
@@ -32,7 +32,9 @@ func (l *Loop) streamStep(ctx context.Context, messages []model.Message) (*stepO
 	}
 
 	resp, err := adapter.ChatStream(ctx, req, func(ev model.StreamEvent) {
-		if l.onStream != nil {
+		// Suppress UI broadcast for internal autonomous passes (backthink),
+		// whose intermediate reasoning must not leak into the chat stream.
+		if l.onStream != nil && !isMutedBroadcast(ctx) {
 			l.onStream(ev)
 		}
 	})
diff --git a/internal/runtime/loop/tools.go b/internal/runtime/loop/tools.go
index db27939..96f3ce1 100644
--- a/internal/runtime/loop/tools.go
+++ b/internal/runtime/loop/tools.go
@@ -275,14 +275,14 @@ func (l *Loop) invokeOne(ctx context.Context, d *tool.Dispatcher, tc agent.ToolC
 	if d == nil {
 		return ToolOutcome{Err: fmt.Errorf("no tool dispatcher configured for %q", tc.Name), ErrKind: tool.KindExecution}
 	}
-	l.emitToolExecution(tc.Name, "start")
+	l.emitToolExecution(ctx, tc.Name, "start")
 	_, toolSpan := l.span(ctx, "tool."+tc.Name, trace.SpanKindTool)
 	result, err := d.Invoke(ctx, tc.Name, tc.Args)
 	phase := "end"
 	if err != nil {
 		phase = "end:error"
 	}
-	l.emitToolExecution(tc.Name, phase)
+	l.emitToolExecution(ctx, tc.Name, phase)
 	out := ToolOutcome{Result: result, Err: err, ErrKind: errorKind(err)}
 	if err != nil {
 		l.spanProp(toolSpan, "error", err.Error())
@@ -297,9 +297,10 @@ func (l *Loop) invokeOne(ctx context.Context, d *tool.Dispatcher, tc agent.ToolC
 
 // emitToolExecution forwards a tool_execution_update progress event to the
 // frontend when the loop is on the streaming path (OnStream configured); it is a
-// no-op otherwise. phase is a short label like "start" or "end:error".
-func (l *Loop) emitToolExecution(name, phase string) {
-	if l.onStream == nil {
+// no-op otherwise (muted internal passes skip it too). phase is a short label
+// like "start" or "end:error".
+func (l *Loop) emitToolExecution(ctx context.Context, name, phase string) {
+	if l.onStream == nil || isMutedBroadcast(ctx) {
 		return
 	}
 	l.onStream(model.StreamEvent{Kind: model.StreamKindToolExecution, Delta: phase + ":" + name})
diff --git a/internal/services/backthink/thinker.go b/internal/services/backthink/thinker.go
index f64ace7..6854eaa 100644
--- a/internal/services/backthink/thinker.go
+++ b/internal/services/backthink/thinker.go
@@ -26,7 +26,9 @@ func (t *LoopThinker) Think(ctx context.Context, sessionID, trigger string) (str
 		return "", ErrNoLoop
 	}
 	messages := []model.Message{model.TextMessage(model.RoleUser, backgroundPrompt(trigger))}
-	res, err := t.loop.Run(ctx, sessionID, messages)
+	// Mute UI broadcast: background reflection must not leak its intermediate
+	// reasoning into the user-facing chat stream (it only needs the conclusion).
+	res, err := t.loop.Run(loop.WithMutedBroadcast(ctx), sessionID, messages)
 	if err != nil {
 		return "", fmt.Errorf("backthink: think: %w", err)
 	}