caveman-setup

SUSPICIOUS. The skill's core function is coherent, but it reroutes all LLM traffic and gateway credentials through a Caveman-controlled proxy, and in BYOK mode it forwards provider keys to that proxy. That third-party interception may be the product's purpose, but it materially changes data flow and trust boundaries; combined with an autonomous verification request, this makes the skill medium-high risk rather than benign. The scanner's command-injection hits appear to be documentation/template-literal false positives, not active execution.