diff --git a/examples/atmega32u4-device-arduino/src/ci_responder.c b/examples/atmega32u4-device-arduino/src/ci_responder.c
index d2e912b..ceb716f 100644
--- a/examples/atmega32u4-device-arduino/src/ci_responder.c
+++ b/examples/atmega32u4-device-arduino/src/ci_responder.c
@@ -43,10 +43,15 @@ static uint32_t ci_rng(void *ctx) {
 
 static uint32_t ci_write_fn(const uint32_t *words, uint32_t count, void *ctx) {
     (void)ctx;
-    for (uint32_t i = 0; i < count; i++)
-        if (!midi2duino_write_word(words[i]))
-            return i;                 /* ring full: partial write reported */
-    return count;
+    /* Whole packet or nothing: a split would leave half a UMP in the ring and
+     * corrupt the USB output. Pump the transport while the ring is full, with
+     * a bound so a host that stops reading cannot hang the responder. */
+    for (uint16_t tries = 0; tries < 1000; tries++) {
+        if (midi2duino_write(words, (uint8_t)count))
+            return count;
+        midi2duino_task();
+    }
+    return 0;
 }
 
 void ci_responder_init(void) {
diff --git a/examples/atmega32u4-device-arduino/src/midi2_usb.cpp b/examples/atmega32u4-device-arduino/src/midi2_usb.cpp
index b0f56af..c0a9d38 100644
--- a/examples/atmega32u4-device-arduino/src/midi2_usb.cpp
+++ b/examples/atmega32u4-device-arduino/src/midi2_usb.cpp
@@ -17,12 +17,12 @@ static midi2_dispatch   g_dp;
 static midi2_proc_state g_proc;
 static uint8_t          g_sysex7_buf[160];   /* MIDI-CI reassembly */
 
-/* Stream messages (MT 0xF) drive the responder; everything else echoes back.
+/* Stream messages drive the responder; everything else echoes back.
  * SysEx7 never lands here: the processor reassembles it and delivers it
  * through on_sysex7 to the MIDI-CI responder. */
 static void on_ump(const uint32_t *words, uint8_t word_count, void *ctx) {
     (void)ctx;
-    if ((words[0] >> 28) == 0xF)
+    if (midi2_msg_get_mt(words) == MIDI2_MT_STREAM)
         midi2_dispatch_feed(words, word_count, &g_dp);
     else
         midi2duino_write(words, word_count);
@@ -56,13 +56,19 @@ void Midi2Usb::task() {
                 return;
             have = 1;
         }
-        uint8_t need = midi2_msg_word_count((uint8_t)(msg[0] >> 28));
+        uint8_t mt   = midi2_msg_get_mt(msg);
+        uint8_t need = midi2_msg_word_count(mt);
         while (have < need) {
             if (!midi2duino_read(&w))
                 return;                 /* rest of the message next round */
             msg[have++] = w;
         }
-        midi2_proc_feed(&g_proc, msg, need);
+        /* Data128 (SysEx8 / MDS) is echoed raw: midi2_proc consumes it for
+         * reassembly, which is off here, so it never reaches on_ump. */
+        if (mt == MIDI2_MT_DATA128)
+            midi2duino_write(msg, need);
+        else
+            midi2_proc_feed(&g_proc, msg, need);
         have = 0;
     }
 }
diff --git a/examples/atmega32u4-device-arduino/src/stream_responder.c b/examples/atmega32u4-device-arduino/src/stream_responder.c
index 0e945d9..5897c56 100644
--- a/examples/atmega32u4-device-arduino/src/stream_responder.c
+++ b/examples/atmega32u4-device-arduino/src/stream_responder.c
@@ -11,8 +11,15 @@
 #define IDENT_VERSION       0x00010000UL   /* 1.0 */
 #define FB_NUM_MAIN         0
 
+/* A dropped Stream reply breaks host discovery, so pump the transport until
+ * the packet fits. Bounded so a host that stops reading cannot hang the loop
+ * (the pump only moves rings to endpoints; it never re-enters the processor). */
 static void send4(const uint32_t *w) {
-    midi2duino_write(w, 4);
+    for (uint16_t tries = 0; tries < 1000; tries++) {
+        if (midi2duino_write(w, 4))
+            return;
+        midi2duino_task();
+    }
 }
 
 /* multi-packet text notification, 14 bytes per UMP (13 for FB name) */