The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs local shell command execution via git status --short to ensure the repository is in a clean state before applying context updates.
[DATA_EXFILTRATION]: The skill includes robust privacy guards that explicitly forbid making outbound network calls and prevent the automatic collection of sensitive identity information from the environment (such as userEmail or OS environment variables).
[PROMPT_INJECTION]: The skill handles untrusted user input through an interactive interview process, creating a surface for indirect prompt injection. This is mitigated by a mandatory human-in-the-loop diff preview, where the user must approve any text before it is saved to the context files.
[CREDENTIALS_UNSAFE]: The skill incorporates a security scanner that identifies credential patterns (e.g., API keys, Bearer tokens, JWTs) in user responses, refusing to save them and requiring redaction to prevent accidental exposure in public repositories.

onboard