The Agent Skills Directory

[COMMAND_EXECUTION]: The skill mentions various commands (e.g., $linter-setup, $project-init, $sync-codex) and platform-specific syntax (e.g., !command). These are used as cross-references to other skills or standard development workflows within the platform's ecosystem and do not represent arbitrary or dangerous shell execution.
[PROMPT_INJECTION]: While the skill contains directive language such as 'Ignore Claude-specific mode-switch instructions' and 'Strict execution contract', these are context-setting instructions for the AI agent to maintain consistency within its specific platform (Codex) rather than malicious attempts to bypass safety filters or exfiltrate data.
[REMOTE_CODE_EXECUTION]: The skill references external tools like linters (ESLint), mutation testing tools (Stryker, PITest), and CI providers. However, it explicitly states it does not install these tools itself (delegating that to $linter-setup) and requires user confirmation for tool selection, mitigating risk of unauthorized remote code execution.
[DATA_EXFILTRATION]: There are no patterns detected that attempt to read sensitive files (e.g., .ssh, .aws/credentials) or send data to unauthorized external domains. The network operations referenced are for standard development tasks (e.g., git, gh pr) within a trusted workflow context.

harness-setup