The Agent Skills Directory

[REMOTE_CODE_EXECUTION] (HIGH): The skill automatically executes package installation and test commands based on files detected in the repository.
Evidence: The 'Run Project Setup' and 'Verify Clean Baseline' sections in SKILL.md invoke npm install, pip install -r requirements.txt, poetry install, cargo build, npm test, cargo test, pytest, and go test without a human-in-the-loop confirmation.
Risk: An attacker can include malicious preinstall or postinstall scripts in package.json, or embed malicious code within the test suite that executes when the agent attempts to verify the 'clean' baseline.
[COMMAND_EXECUTION] (HIGH): The skill performs shell-level directory creation and navigation based on variables derived from the repository environment.
Evidence: git worktree add "$path" -b "$BRANCH_NAME" and subsequent cd "$path" commands.
[EXTERNAL_DOWNLOADS] (MEDIUM): The skill initiates downloads of external dependencies from public registries.
Evidence: Execution of npm install, pip install, and go mod download triggers network requests to external package repositories.
[PROMPT_INJECTION] (MEDIUM): The skill parses CLAUDE.md for instructional preferences, creating a surface for indirect prompt injection.
Evidence: grep -i "worktree.*director" CLAUDE.md is used to determine directory selection logic from an untrusted repository file.

using-git-worktrees