sandcastle

SUSPICIOUS. The package and stated capabilities are largely coherent for an agent-orchestration skill, and the npm install path appears legitimate. The main issue is that the skill body and reference docs include literal !`command` pre-execution syntax; in a Claude skill, those examples become real load-time shell commands, causing unintended command execution and potential exposure of repo state before approval. Broad hooks and secret mounts are powerful but proportionate to the orchestration purpose when used carefully.