The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill recommends and provides instructions for installing third-party agent skills from GitHub repositories (specifically mattpocock/skills and joeseesun/qiaomu-goal-meta-skill) which are not listed as trusted or well-known organizations. This introduces unverified external code and prompts into the agent environment.
[REMOTE_CODE_EXECUTION]: The core mechanism of the skill involves an 'unattended delivery loop' where agents generate implementation code ('goals') and subsequently execute them in headless terminal environments. This automated implementation of generated logic represents a significant security risk if the generation process is compromised.
[COMMAND_EXECUTION]: The skill relies on an extensive set of shell scripts (control-plane kernel) to manage state, detect hosts, and launch runners. These scripts interact with environment variables and user-supplied paths in the .adr/ directory, increasing the potential attack surface for local command manipulation.
[PRIVILEGE_ESCALATION]: Templates provided in scripts/launch-runner-template.sh and documentation in references/orca-host.md explicitly include and suggest the use of high-privilege flags for agents, such as --permission-mode bypassPermissions and --sandbox danger. While the skill notes that these should only be used with user authorization, providing ready-to-use configurations lowers the barrier for executing agents with minimal safety constraints.

devloop