diff --git a/plugins/commando.c b/plugins/commando.c
index 0e4a8b3ba..177860603 100644
--- a/plugins/commando.c
+++ b/plugins/commando.c
@@ -152,10 +152,6 @@ static struct command_result *send_response(struct command *command UNUSED,
 	if (msglen > 65000) {
 		msglen = 65000;
 		msgtype = COMMANDO_MSG_REPLY_CONTINUES;
-		/* We need to make a copy first time before we call back, since
-		 * plugin will reuse it! */
-		if (!result)
-			reply->buf = tal_dup_talarr(reply, char, reply->buf);
 	} else {
 		if (msglen == 0) {
 			tal_free(reply);
@@ -188,12 +184,32 @@ static struct command_result *cmd_done(struct command *command,
 {
 	struct reply *reply = tal(plugin, struct reply);
 	reply->incoming = tal_steal(reply, incoming);
-	reply->buf = (char *)buf;
 
-	/* result is contents of "error" or "response": we want top-leve
-	 * object */
-	reply->off = obj->start;
-	reply->len = obj->end;
+	/* We make a copy, but substititing the original id! */
+	if (incoming->json_id) {
+		const char *id_start, *id_end;
+		const jsmntok_t *id = json_get_member(buf, obj, "id");
+		size_t off;
+
+		/* Old id we're going to omit */
+		id_start = json_tok_full(buf, id);
+		id_end = id_start + json_tok_full_len(id);
+
+		reply->len = obj->end - obj->start
+			- (id_end - id_start)
+			+ strlen(incoming->json_id);
+		reply->buf = tal_arr(reply, char, reply->len);
+		memcpy(reply->buf, buf + obj->start,
+		       id_start - (buf + obj->start));
+		off = id_start - (buf + obj->start);
+		memcpy(reply->buf + off, incoming->json_id, strlen(incoming->json_id));
+		off += strlen(incoming->json_id);
+		memcpy(reply->buf + off, id_end, (buf + obj->end) - id_end);
+	} else {
+		reply->len = obj->end - obj->start;
+		reply->buf = tal_strndup(reply, buf + obj->start, reply->len);
+	}
+	reply->off = 0;
 
 	return send_response(command, NULL, NULL, reply);
 }
diff --git a/tests/test_plugin.py b/tests/test_plugin.py
index 822744329..c2d06580c 100644
--- a/tests/test_plugin.py
+++ b/tests/test_plugin.py
@@ -2609,9 +2609,7 @@ def test_plugin_shutdown(node_factory):
 
 def test_commando(node_factory, executor):
     l1, l2 = node_factory.line_graph(2, fundchannel=False,
-                                     opts={'log-level': 'io',
-                                           # FIXME: Currently, our JSON ids in replies are wrong, hence BROKEN!
-                                           'allow_broken_log': True})
+                                     opts={'log-level': 'io'})
 
     # Nothing works until we've issued a rune.
     fut = executor.submit(l2.rpc.call, method='commando',
@@ -2706,8 +2704,7 @@ def test_commando(node_factory, executor):
 
 
 def test_commando_rune(node_factory):
-    # FIXME: Currently, our JSON ids in replies are wrong, hence BROKEN!
-    l1, l2 = node_factory.get_nodes(2, opts={'allow_broken_log': True})
+    l1, l2 = node_factory.get_nodes(2)
 
     # Force l1's commando secret
     l1.rpc.datastore(key=['commando', 'secret'], hex='1241faef85297127c2ac9bde95421b2c51e5218498ae4901dc670c974af4284b')
@@ -2942,8 +2939,7 @@ def test_commando_rune(node_factory):
 
 def test_commando_stress(node_factory, executor):
     """Stress test to slam commando with many large queries"""
-    # FIXME: Currently, our JSON ids in replies are wrong, hence BROKEN!
-    nodes = node_factory.get_nodes(5, opts={'allow_broken_log': True})
+    nodes = node_factory.get_nodes(5)
 
     rune = nodes[0].rpc.commando_rune()['rune']
     for n in nodes[1:]: