QuoteBench: How Matched Scores Can Hide Command-Path Failures

Shangao Li, Yao Zhang, Volker Tresp +1

LLM coding agents issue Bash commands through interfaces that may serialize, wrap, and reparse model output. Matched execution scores alone cannot distinguish command-generation er…