diff --git a/src/agentkit/experts/orchestrator.py b/src/agentkit/experts/orchestrator.py
index 7f15a6f..c3d894d 100644
--- a/src/agentkit/experts/orchestrator.py
+++ b/src/agentkit/experts/orchestrator.py
@@ -589,9 +589,21 @@ class TeamOrchestrator(
             f"where natural."
         )
         # Temporarily wrap task with hint — call _decompose_task with augmented task.
+        # U4/KTD5: re-decomposition is a re-planning signal — pass trigger_reason
+        # so retrieve_prompt_reflection can inform the new plan with historical
+        # reflection. "loop_detection" is the closest KTD5 trigger: the original
+        # plan hit a structural limit (too many independent subtasks) and needs
+        # re-planning, analogous to detecting a planning loop.
+        # ponytail: verify_retry / schema_validation triggers belong to the
+        # verify-failed / schema-failed retry paths (plan_exec_engine.py /
+        # verification_loop.py), which do not call _decompose_task today;
+        # wiring those is deferred until those paths are integrated with
+        # team re-planning.
         augmented_task = f"{task}\n\n[Re-decomposition hint]: {original_hint}"
         try:
-            new_phases = await self._decompose_task(lead, augmented_task)
+            new_phases = await self._decompose_task(
+                lead, augmented_task, trigger_reason="loop_detection"
+            )
         except (LLMProviderError, asyncio.TimeoutError, ConnectionError, ValueError) as e:
             logger.warning(f"U4 re-decompose failed: {e}, keeping original phases")
             return phases
diff --git a/src/agentkit/server/frontend/src/stores/chatStore.ts b/src/agentkit/server/frontend/src/stores/chatStore.ts
index 4fe56b0..0dfdf33 100644
--- a/src/agentkit/server/frontend/src/stores/chatStore.ts
+++ b/src/agentkit/server/frontend/src/stores/chatStore.ts
@@ -5,7 +5,7 @@ import { useTeamStore } from "@/stores/team";
 import { useDocumentsStore } from "@/stores/documents";
 import { useCalendarStore } from "@/stores/calendar";
 import { useChatSocket } from "@/stores/chatSocket";
-import { useChatStream } from "@/stores/chatStream";
+import { useChatStream, updateLastStep } from "@/stores/chatStream";
 import type { BoardState } from "@/stores/chatStream";
 import { pickExpertIdentity } from "@/components/chat/helpers/expertIdentity";
 import type {
@@ -225,6 +225,16 @@ export const useChatStore = defineStore("chat", () => {
     // 会话隔离 — 切换会话时重置三个 transient state，避免跨会话数据泄漏。
     // force-reload 同一会话时不重置，防止误清空。
     if (prevConvId !== id) {
+      // U1/P2-1: if the previous conversation had an autonomy-paused execution
+      // waiting for resume, cancel it server-side so the engine's
+      // _resume_handler future does not hang for up to 30 minutes.
+      if (stream.autonomyPaused.value && socket.ws.value?.readyState === WebSocket.OPEN) {
+        try {
+          socket.ws.value.send(JSON.stringify({ type: "cancel" } as WsClientMessage));
+        } catch {
+          // swallow — socket teardown will cancel the future server-side
+        }
+      }
       stream.boardState.value = null;
       stream.debateState.value = null;
       stream.collaborationState.value = null;
@@ -546,6 +556,17 @@ export const useChatStore = defineStore("chat", () => {
    */
   function resumeAutonomy(resumeToken: string): void {
     stream.autonomyPaused.value = null;
+    // U1/P2-2: mark the "自主执行已暂停" milestone as completed so the
+    // streaming step list does not leave a perpetual "running" entry.
+    const cid = currentConversationId.value;
+    if (cid) {
+      updateLastStep(
+        stream.streamingStepsByConv,
+        (s) => s.type === "milestone" && s.label === "自主执行已暂停",
+        { status: "success" },
+        cid,
+      );
+    }
     if (!socket.ws.value || socket.ws.value.readyState !== WebSocket.OPEN) {
       return;
     }