The Agent Skills Directory

COMMAND_EXECUTION (MEDIUM): The skill instructs the agent to run npm install or pnpm install after checking out lock files from an external source during conflict resolution.
Evidence: git checkout --theirs package-lock.json && npm install.
Risk: If an attacker provides a malicious package.json with post-install scripts in a branch or PR, the agent will execute that code when following these instructions.
EXTERNAL_DOWNLOADS (LOW): The use of package managers like npm, pnpm, and cargo triggers the download of external dependencies.
Evidence: npm install, pnpm install, and cargo generate-lockfile.
PROMPT_INJECTION (LOW): The skill is vulnerable to indirect prompt injection via untrusted external data in Git branches.
Ingestion points: git checkout --theirs and repository archaeology commands.
Boundary markers: None.
Capability inventory: npm install, pnpm install, git push, and sops execution.
Sanitization: None.

git