We show that Repair is Nearly Generation! By adding symbolic information to Large Language Models trained on code (Codex), we can build a competitive repair engine in days!