diff --git a/backend/open_webui/socket/main.py b/backend/open_webui/socket/main.py
index 1b7f67ba50..95d7ca962d 100644
--- a/backend/open_webui/socket/main.py
+++ b/backend/open_webui/socket/main.py
@@ -789,7 +789,9 @@ async def resume_stream(sid, data):
 
     # Cap payload bytes. Keep the newest entries that fit; older ones
     # are either already in the DB-backed content or will arrive via the
-    # final done:True checkpoint.
+    # final done:True checkpoint. A single envelope larger than the cap
+    # is skipped rather than allowed through — forcing it past the cap
+    # would still exceed Socket.IO's buffer.
     total_bytes = 0
     capped = []
     for env in reversed(envelopes):
@@ -797,7 +799,9 @@ async def resume_stream(sid, data):
             size = len(json.dumps(env))
         except Exception:
             continue
-        if total_bytes + size > RESUME_STREAM_REPLAY_MAX_BYTES and capped:
+        if size > RESUME_STREAM_REPLAY_MAX_BYTES:
+            continue
+        if total_bytes + size > RESUME_STREAM_REPLAY_MAX_BYTES:
             break
         capped.append(env)
         total_bytes += size
diff --git a/src/lib/components/chat/Chat.svelte b/src/lib/components/chat/Chat.svelte
index c5dc02b012..68053ccd4e 100644
--- a/src/lib/components/chat/Chat.svelte
+++ b/src/lib/components/chat/Chat.svelte
@@ -451,12 +451,15 @@
 			let message = history.messages[event.message_id];
 
 			if (message) {
-				// Buffer live frames during replay; _replayed frames skip the fence.
-				// Store the ack callback alongside the event so Socket.IO
-				// call-style events (confirmation/execute/input) don't lose
-				// their response path when buffered and later replayed.
+				// Buffer live frames during replay; _replayed frames skip
+				// the fence. Call-style events that carry an ack callback
+				// ALSO skip the fence — the server is waiting on the ack
+				// and buffering could exceed WEBSOCKET_EVENT_CALLER_TIMEOUT
+				// (up to RESUME_FENCE_TIMEOUT_MS of delay). Ack events
+				// don't carry seq and don't mutate streamed content, so
+				// bypassing is safe against the original replay race.
 				const queue = resumeQueueByMessageId.get(event.message_id);
-				if (queue && !event?._replayed) {
+				if (queue && !event?._replayed && !cb) {
 					queue.push({ event, cb });
 					return;
 				}
@@ -701,7 +704,9 @@
 			message.id,
 			setTimeout(() => {
 				console.warn('resume-stream fence timed out for', message.id);
-				clearResumeFence(message.id);
+				clearResumeFence(message.id).catch((e) =>
+					console.error('resume fence timeout flush failed', e)
+				);
 			}, RESUME_FENCE_TIMEOUT_MS)
 		);
 		const requestId =