diff --git a/include/condy/runtime.hpp b/include/condy/runtime.hpp
index db20affd..645f7a3b 100644
--- a/include/condy/runtime.hpp
+++ b/include/condy/runtime.hpp
@@ -346,6 +346,11 @@ class Runtime {
         schedule_msg_ring_(
             curr_runtime,
             detail::encode_work(nullptr, detail::WorkType::Ignore));
+        if (curr_runtime != nullptr) {
+            // Ensure the wakeup msg is submitted. Since user may call
+            // thread.join() after allow_exit()
+            curr_runtime->ring_.submit();
+        }
     }
 
     void flush_global_queue_() noexcept {
diff --git a/tests/test_runtime.cpp b/tests/test_runtime.cpp
index 45a45cfb..db6f1538 100644
--- a/tests/test_runtime.cpp
+++ b/tests/test_runtime.cpp
@@ -5,6 +5,7 @@
 #include "condy/runtime_options.hpp"
 #include "condy/task.hpp"
 #include <atomic>
+#include <chrono>
 #include <doctest.h>
 #include <thread>
 
@@ -322,3 +323,21 @@ TEST_CASE("test runtime - runtime destroyed on another runtime") {
     runtime_a.allow_exit();
     runtime_a.run();
 }
+
+TEST_CASE("test runtime - runtime join on another runtime") {
+    condy::Runtime runtime_a(options);
+
+    auto func = [&]() -> condy::Coro<void> {
+        auto runtime_b = std::make_unique<condy::Runtime>(options);
+        std::thread thread_b([&]() { runtime_b->run(); });
+        // wait runtime_b call submit_and_wait
+        std::this_thread::sleep_for(std::chrono::milliseconds(1));
+        runtime_b->allow_exit();
+        thread_b.join();
+        co_return;
+    };
+
+    condy::co_spawn(runtime_a, func()).detach();
+    runtime_a.allow_exit();
+    runtime_a.run();
+}
\ No newline at end of file