diff --git a/src/runtime/nodes/mqtt_shared.rs b/src/runtime/nodes/mqtt_shared.rs
index a5b0e3e..ca7d5e0 100644
--- a/src/runtime/nodes/mqtt_shared.rs
+++ b/src/runtime/nodes/mqtt_shared.rs
@@ -164,16 +164,26 @@ async fn drive(
     loop {
         match poller.poll().await {
             Ok(PollEvent::Connected) => {
+                // Re-subscribe and publish birth, but NOT inline: each client
+                // request enqueues into a bounded channel that only this driver
+                // drains by polling. Awaiting many requests here (e.g. dozens of
+                // mqtt-in filters) fills the queue while the loop is not polling,
+                // which deadlocks the connection. Hand the work to a detached
+                // task so the event loop keeps being polled and the queue drains.
                 let filters: Vec<(String, u8)> = subs
                     .lock()
                     .unwrap()
                     .iter()
                     .map(|s| (s.filter.clone(), s.qos))
                     .collect();
-                for (filter, qos) in filters {
-                    let _ = client.subscribe(&filter, qos).await;
-                }
-                birth.publish(client.as_ref()).await;
+                let client = client.clone();
+                let birth = birth.clone();
+                tokio::spawn(async move {
+                    for (filter, qos) in filters {
+                        let _ = client.subscribe(&filter, qos).await;
+                    }
+                    birth.publish(client.as_ref()).await;
+                });
             }
             Ok(PollEvent::Message(m)) => {
                 let subs = subs.lock().unwrap();