diff --git a/docs/superpowers/plans/2026-08-06-problem-reaction.md b/docs/superpowers/plans/2026-08-06-problem-reaction.md
index dce6b10..4bb681c 100644
--- a/docs/superpowers/plans/2026-08-06-problem-reaction.md
+++ b/docs/superpowers/plans/2026-08-06-problem-reaction.md
@@ -2,6 +2,8 @@
 
 > 2026-08-06 规则更新：当前实现已改为单选、点击即提交、提交后不可修改，并通过 `(problem, user)` 数据库唯一约束保证一人一题一条。新接口使用单值 `type` / `mine_type`；为支持前后端错序部署，过渡期仍接受单元素 `types` 并返回数组 `mine`。统计直接查询数据库，不再使用 reaction 缓存。本文中的多选及缓存步骤是早期实施记录。
 
+> 2026-08-06 后台形态变更：Task 3（管理接口 `/api/admin/reaction`）与 Task 6（后台反馈统计页）的产物已整体删除。后台反馈现在只是**题目列表的一列**：显示票数最高的那个表情图标，tooltip 出「标签名 + 人数」，无评价留空，并列取 `ReactionType` 定义序靠前的一个。数据由 `GET /api/admin/problem` 的 `top_reaction` 字段下发，聚合逻辑在 `reaction/services.py:get_top_reactions()`。详见 spec 顶部同日说明。
+
 > **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
 
 **Goal:** 把题目点评从「三维评分 + 文字」重写为「一排七个表情按钮，点击即表态」，并把后台从逐条评论管理改为按题目聚合的反馈统计表。
@@ -1090,6 +1092,8 @@ Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>"
 
 实现全部完成并通过终审。以下是评审过程中判定「可以带着合并」的次要项，都不阻塞上线，记在这里免得日后重新发现一遍。
 
+其中涉及后台统计页与统计缓存的四条（题号筛选返回空结果、缓存快照少算一票、`ReactionStatsRow` 缺 `_ratio`、后台筛选不重置页码与 `listStats` 无序号守卫）随后台页面下线与缓存移除一并作废，代码已不存在，保留仅作记录。
+
 **后端**
 
 - `problem_id` 传非数字（如 `?problem_id=abc`）会走兜底 except 返回泛化 500 并打一条 stack trace。前端始终传数字，实际影响只是日志噪音。与原 `comment/views/oj.py` 的写法一致。
diff --git a/docs/superpowers/specs/2026-08-06-problem-reaction-design.md b/docs/superpowers/specs/2026-08-06-problem-reaction-design.md
index 64151b5..4859462 100644
--- a/docs/superpowers/specs/2026-08-06-problem-reaction-design.md
+++ b/docs/superpowers/specs/2026-08-06-problem-reaction-design.md
@@ -2,6 +2,8 @@
 
 > 2026-08-06 规则更新：评价已改为单选，点击一个表情后立即提交，且提交后不可修改。一人一题在数据库中最多保留一条记录。新请求字段为 `type`，新响应字段为 `mine_type: ReactionKey | null`；为支持前后端错序部署，过渡期仍接受单元素 `types` 并返回数组 `mine`。计数直接查询数据库，不使用 Redis 缓存。本文后续关于“最多选 3 个、可取消、整份覆盖、统计缓存”的内容仅保留为早期设计记录，不再代表当前行为。
 
+> 2026-08-06 后台形态变更：独立的后台反馈统计页与 `GET /api/admin/reaction` 已全部下线（`reaction/views/admin.py`、`reaction/urls/admin.py`、`src/admin/reaction/list.vue`、路由与菜单项均已删除）。取而代之，**后台题目列表**（`src/admin/problem/list.vue`）多一列「反馈」，只显示该题票数最高的那一个表情图标，tooltip 出「标签名 + 人数」；无人评价则留空，并列第一时按 `ReactionType` 的定义顺序取靠前的一个。数据来自 `GET /api/admin/problem` 新增的 `top_reaction: {type, count} | null` 字段，由 `reaction/services.py:get_top_reactions()` 对当页题目做一次聚合查询算出（比赛题目列表不返回该字段，前端也不显示该列）。本文「管理接口」「后台页面」两节仅为早期设计记录，占比排序、七列计数表、按题号筛选均已不存在。
+
 日期：2026-08-06
 涉及仓库：`OnlineJudge`（后端）、`ojnext`（前端）
 
diff --git a/oj/urls.py b/oj/urls.py
index 2a244a7..96b6aea 100644
--- a/oj/urls.py
+++ b/oj/urls.py
@@ -16,7 +16,6 @@ urlpatterns = [
     path("api/admin/", include("utils.urls")),
     path("api/", include("message.urls.oj")),
     path("api/", include("reaction.urls.oj")),
-    path("api/admin/", include("reaction.urls.admin")),
     path("api/", include("tutorial.urls.tutorial")),
     path("api/admin/", include("tutorial.urls.admin")),
     path("api/", include("ai.urls.oj")),
diff --git a/problem/views/admin.py b/problem/views/admin.py
index dd2cd08..1e56409 100644
--- a/problem/views/admin.py
+++ b/problem/views/admin.py
@@ -14,6 +14,7 @@ from django.http import StreamingHttpResponse
 from account.decorators import ensure_created_by, problem_permission_required, teacher_admin_required
 from contest.models import Contest, ContestStatus
 from judge.sql_runner import SQLCaseError, build_display
+from reaction.services import get_top_reactions
 from submission.models import Submission
 from utils.api import APIError, APIView, CSRFExemptAPIView, validate_serializer
 from utils.openai import get_ai_client
@@ -283,7 +284,12 @@ class ProblemAPI(ProblemBase):
 
         if not user.can_mgmt_all_problem():
             problems = problems.filter(created_by=user)
-        return self.success(self.paginate_data(request, problems, ProblemAdminListSerializer))
+
+        data = self.paginate_data(request, problems, ProblemAdminListSerializer)
+        top_reactions = get_top_reactions([row["id"] for row in data["results"]])
+        for row in data["results"]:
+            row["top_reaction"] = top_reactions.get(row["id"])
+        return self.success(data)
 
     @problem_permission_required
     @validate_serializer(EditProblemSerializer)
diff --git a/reaction/services.py b/reaction/services.py
new file mode 100644
index 0000000..f2e1a47
--- /dev/null
+++ b/reaction/services.py
@@ -0,0 +1,19 @@
+from django.db.models import Count
+
+from reaction.models import Reaction, ReactionType
+
+# 并列票数时按 ReactionType 的定义顺序取靠前的那个，跟前端 REACTIONS 的顺序一致
+TYPE_ORDER = {t.value: index for index, t in enumerate(ReactionType)}
+
+
+def get_top_reactions(problem_ids):
+    """返回 {problem_id: {"type": ..., "count": ...}}，只包含有评价的题目。"""
+    if not problem_ids:
+        return {}
+    top = {}
+    rows = Reaction.objects.filter(problem_id__in=problem_ids).values("problem_id", "type").annotate(count=Count("id"))
+    for row in rows:
+        current = top.get(row["problem_id"])
+        if current is None or (-row["count"], TYPE_ORDER[row["type"]]) < (-current["count"], TYPE_ORDER[current["type"]]):
+            top[row["problem_id"]] = {"type": row["type"], "count": row["count"]}
+    return top
diff --git a/reaction/urls/admin.py b/reaction/urls/admin.py
deleted file mode 100644
index 9bf6fb8..0000000
--- a/reaction/urls/admin.py
+++ /dev/null
@@ -1,7 +0,0 @@
-from django.urls import path
-
-from ..views.admin import ReactionStatsAPI
-
-urlpatterns = [
-    path("reaction", ReactionStatsAPI.as_view()),
-]
diff --git a/reaction/views/admin.py b/reaction/views/admin.py
deleted file mode 100644
index ac90513..0000000
--- a/reaction/views/admin.py
+++ /dev/null
@@ -1,25 +0,0 @@
-from django.db.models import Count, F, Max, Q
-
-from account.decorators import super_admin_required
-from reaction.models import Reaction, ReactionType
-from utils.api import APIView
-
-
-class ReactionStatsAPI(APIView):
-    @super_admin_required
-    def get(self, request):
-        queryset = (
-            Reaction.objects.values(pid=F("problem___id"), title=F("problem__title"))
-            .annotate(users=Count("user", distinct=True))
-            .annotate(last_time=Max("create_time"))
-            .annotate(**{t.value: Count("id", filter=Q(type=t.value)) for t in ReactionType})
-            .order_by("-users")
-        )
-
-        problem_id = request.GET.get("problem")
-        if problem_id:
-            queryset = queryset.filter(problem___id__iexact=problem_id, problem__contest_id__isnull=True)
-
-        data = self.paginate_data(request, queryset)
-        data["results"] = list(data["results"])
-        return self.success(data)