diff --git a/ci/verify-backend.sh b/ci/verify-backend.sh
index 6e221ad..46ca219 100644
--- a/ci/verify-backend.sh
+++ b/ci/verify-backend.sh
@@ -61,7 +61,7 @@ status=0
 (cd "$src/backend" && dotnet test --solution MoShiro.slnx -c Release --no-build) 2>&1 | tee "$log" || status=${PIPESTATUS[0]}
 [[ "$status" == 0 ]] && exit 0
 
-# 个别端到端测试在机器忙的时候会偶发失败（比如探针 5 秒内没连上）：失败的测试单独重跑一次，重跑通过就当通过（日志里留着第一次的失败）；
+# 个别端到端测试会偶发失败（比如 Hy2 阶梯测试里探针掉线）：失败的测试单独重跑（最多两次），重跑通过就当通过（日志里留着第一次的失败）；
 # 失败太多说明是真的坏了，不重跑
 mapfile -t failed < <(grep -E '^(failed|失败) ' "$log" | awk '{print $2}' | awk -F. '{print $NF}' | sort -u)
 if ((${#failed[@]} == 0 || ${#failed[@]} > 3)); then
@@ -69,10 +69,16 @@ if ((${#failed[@]} == 0 || ${#failed[@]} > 3)); then
   exit "$status"
 fi
 for name in "${failed[@]}"; do
-  echo "retrying $name once"
-  (cd "$src/backend" && dotnet test --solution MoShiro.slnx -c Release --no-build --filter-method "$name") || {
-    echo "$name failed again" >&2
-    exit 1
-  }
+  passed=false
+  # 最多重跑两次：这个端到端测试在 CI 主机上不稳，一次重跑不一定够
+  for attempt in 1 2; do
+    echo "retrying $name (attempt $attempt of 2)"
+    # 过滤要带通配符；别的程序集里没有匹配的测试，会因为「零个测试」返回 8，忽略它
+    if (cd "$src/backend" && dotnet test --solution MoShiro.slnx -c Release --no-build --filter-method "*$name*" --ignore-exit-code 8); then
+      passed=true
+      break
+    fi
+  done
+  [[ "$passed" == true ]] || { echo "$name failed again" >&2; exit 1; }
 done
 echo "the failed tests passed on retry"