parallel-vibe

该技能目的与能力基本一致：为同一任务创建多工作区并调用外部 AI CLI 比较方案。主要风险不在明显恶意，而在高权限编排：它读取项目内容、批量生成工作区、调用外部 runner 执行命令，并允许自定义 shell runner；同时仅提供“软护栏”，不是真正沙箱。整体更像高风险自动化编排工具而非恶意技能，判定为 SUSPICIOUS。

This code appears to be a legitimate orchestration tool for running multiple 'runners' in isolated per-thread workspaces. It is not obfuscated and contains no obvious hardcoded credentials or hidden backdoors. However, it intentionally allows execution of arbitrary external commands and gives those external processes access to copied workspace files. The main security risks are: (1) arbitrary command execution via 'shell' runner or malicious/compromised runner CLIs; (2) potential data exfiltration because runners receive workspace contents and can send them externally; (3) destructive operations via shutil.rmtree of destination workspaces; (4) symlink policies that, if set to 'keep', can lead to workspace boundary escapes. These are expected features for this kind of tool but make it potentially dangerous if untrusted plans, config, or source directories are used. I assess low probability that this module itself is malware, but medium-high risk that it can be abused for supply-chain or data-exfiltration attacks depending on how it's configured and what runners are invoked.