The Agent Skills Directory

[COMMAND_EXECUTION] (LOW): The skill automatically executes project-specific build and test commands (e.g., npm install, npm test, cargo build, pytest, go test). This is standard for development automation but means the agent will execute code defined in the project's configuration files.
[EXTERNAL_DOWNLOADS] (LOW): The use of package managers like npm, pip, poetry, and go triggers downloads from external registries. Per [TRUST-SCOPE-RULE], these are considered low risk as they target trusted standard registries, but the packages themselves are defined by the untrusted project files.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill ingests data from CLAUDE.md to determine directory preferences.
Ingestion points: CLAUDE.md via grep.
Boundary markers: None used during the grep operation.
Capability inventory: File system access, command execution (npm/pip/git).
Sanitization: The skill uses a case statement to validate the $LOCATION variable against a whitelist of allowed paths (.worktrees, worktrees, or a specific global path), which significantly mitigates path traversal or injection risks from the CLAUDE.md input.
[DATA_EXFILTRATION] (SAFE): The skill includes an explicit 'Safety Verification' step that uses git check-ignore to ensure isolated worktrees are not accidentally tracked or committed to the repository, which is a security best practice for data leakage prevention.

using-git-worktrees