diff --git a/cpp/logos_api_client.cpp b/cpp/logos_api_client.cpp
index 0a4bcad..d3db4d7 100644
--- a/cpp/logos_api_client.cpp
+++ b/cpp/logos_api_client.cpp
@@ -144,6 +144,21 @@ void LogosAPIClient::invokeRemoteMethodAsync(const QString& objectName, const QS
 {
     if (!callback) return;
 
+    // The async path acquires a replica too, so it must also run on the owner
+    // thread. Unlike the sync path we post non-blocking (QueuedConnection): the
+    // worker caller returns immediately and the result callback fires on the
+    // owner thread when the reply arrives.
+    if (QThread::currentThread() != this->thread()) {
+        QMetaObject::invokeMethod(this,
+            [this, objectName, methodName, args,
+             callback = std::move(callback), timeout]() mutable {
+                invokeRemoteMethodAsync(objectName, methodName, args,
+                                        std::move(callback), timeout);
+            },
+            Qt::QueuedConnection);
+        return;
+    }
+
     QString token = getToken(objectName);
 
     if (token.isEmpty() && objectName != "capability_module" && m_capability_consumer) {
diff --git a/cpp/logos_thread_marshal.h b/cpp/logos_thread_marshal.h
index de50830..55a5f18 100644
--- a/cpp/logos_thread_marshal.h
+++ b/cpp/logos_thread_marshal.h
@@ -19,13 +19,20 @@ namespace logos {
 // thread) without the module touching Qt — the SDK transparently marshals the
 // call onto the owner thread.
 //
-// Requirements: `obj`'s thread must be running an event loop (it is — the
-// module's main thread runs QCoreApplication::exec()). The same-thread guard
-// avoids the BlockingQueuedConnection self-deadlock.
+// Requirements:
+//   - `obj`'s thread must be running an event loop (it is — the module's main
+//     thread runs QCoreApplication::exec()). The same-thread guard avoids the
+//     BlockingQueuedConnection self-deadlock.
+//   - The return type must be void or default-constructible (the marshaled
+//     branch holds the result in a local before assigning it), and must not be
+//     a reference (there'd be nothing to bind the local to). Both are satisfied
+//     by the SDK's uses here (void, QVariant, LogosObject*, LogosAPIClient*).
 template <typename Fn>
 auto runOnOwnerThread(QObject* obj, Fn&& fn) -> decltype(fn())
 {
     using Ret = decltype(fn());
+    static_assert(!std::is_reference_v<Ret>,
+                  "runOnOwnerThread does not support reference return types");
     if (QThread::currentThread() == obj->thread()) {
         return fn();
     }
diff --git a/tests/sdk/test_worker_thread_ipc.cpp b/tests/sdk/test_worker_thread_ipc.cpp
index 6ef31cd..ca0495a 100644
--- a/tests/sdk/test_worker_thread_ipc.cpp
+++ b/tests/sdk/test_worker_thread_ipc.cpp
@@ -76,10 +76,15 @@ TEST_F(WorkerThreadIpcTest, InvokeFromWorkerThreadRunsOnOwnerThread)
 {
     QThread* const ownerThread = QThread::currentThread();
 
+    // The ModuleProxy created by registerObject (owned by providerApi) keeps a
+    // raw pointer to the provider while published, so the provider must outlive
+    // it. Declaring `provider` before `providerApi` ensures that: at end of
+    // scope, providerApi (and its proxy) is destroyed first, then the provider.
+    ThreadProbeProvider provider;
+
     // Provider registered on the owner thread.
     LogosAPI providerApi("thread_probe");
-    auto* provider = new ThreadProbeProvider();
-    ASSERT_TRUE(providerApi.getProvider()->registerObject("thread_probe", provider));
+    ASSERT_TRUE(providerApi.getProvider()->registerObject("thread_probe", &provider));
     // Authorize the token the consumer will present, so the call reaches the
     // provider instead of being rejected by the authz check.
     providerApi.getProvider()->saveToken("caller", "tok");
@@ -117,10 +122,6 @@ TEST_F(WorkerThreadIpcTest, InvokeFromWorkerThreadRunsOnOwnerThread)
 
     // The crux: the inter-module call must execute on the owner thread, not the
     // worker thread. Without the marshaling fix it runs on the worker thread.
-    EXPECT_EQ(provider->calledThread.load(), ownerThread)
+    EXPECT_EQ(provider.calledThread.load(), ownerThread)
         << "inter-module call executed on the worker thread instead of the owner thread";
-
-    // provider is intentionally leaked: the ModuleProxy created by
-    // registerObject (owned by providerApi) references it for the lifetime of
-    // the LogosAPI, which outlives this scope.
 }