The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install a binary from an external third-party source using Homebrew.
Evidence: brew install --cask vectal-labs/tap/repo-sync in SKILL.md.
[COMMAND_EXECUTION]: The skill makes extensive use of system commands to manage services, modify plist files, and execute the third-party binary.
Evidence: Calls to repo-sync, /usr/bin/plutil, and /bin/launchctl are used to manage the lifecycle and configuration of the sync service.
[PERSISTENCE]: The skill configures a macOS LaunchAgent to ensure the repo-sync service runs automatically at login.
Evidence: Instructions for managing ~/Library/LaunchAgents/com.vectal-labs.repo-sync.plist and using launchctl bootstrap to ensure background execution across sessions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and synchronize the contents of Git repositories, which could contain instructions intended to influence the agent's behavior.
Ingestion points: Git repositories added to the sync service (configured in SKILL.md and references/operations.md).
Boundary markers: None present to delimit repository content from agent instructions.
Capability inventory: File system read/write, network operations (git push/pull), and execution of shell commands via the repo-sync binary.
Sanitization: The skill mentions blocking specific filenames known to contain secrets (e.g., .env, private keys) but does not validate the content of the files themselves.
[METADATA_POISONING]: The repo-sync skill install command performs lateral movement by automatically copying the skill into multiple agent-specific directories.
Evidence: The skill attempts to install itself into ~/.agents/skills/repo-sync, ~/.claude/skills, and ~/.hermes/skills, which modifies the configuration environment of other AI agents.

repo-sync