The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the user or agent to download and immediately execute content from a remote source.
Fetches an installation script from the vendor's GitHub repository (https://raw.githubusercontent.com/av/harbor/refs/heads/main/install.sh) and pipes it directly to bash in SKILL.md.
Fetches markdown instructions from GitHub (https://raw.githubusercontent.com/av/harbor/refs/heads/main/install.md) and pipes them directly into AI agents (claude and codex) in SKILL.md.
[COMMAND_EXECUTION]: The skill facilitates the execution of various host tools and services.
The harbor launch command in SKILL.md and references/launch.md executes host-native coding tools (like claude, grok, vscode) with arbitrary user-supplied arguments.
Instructs the agent to run complex Docker and CLI commands (harbor, docker, nvidia-smi) for service management and troubleshooting.
[PRIVILEGE_ESCALATION]: The setup instructions include commands that require administrative privileges.
Suggests using sudo usermod -aG docker $USER to modify system group memberships in SKILL.md.
[CREDENTIALS_UNSAFE]: Hardcoded default configuration values are present.
references/stack.md explicitly lists a default secret h@rb0r for the Open WebUI service.
[PERSISTENCE]: The skill involves modifications to environment and configuration files that persist across sessions.
Recommends sourcing shell initialization files (~/.bashrc, ~/.zshrc) after installation in SKILL.md.
Temporarily modifies the Grok tool's global configuration file (~/.grok/config.toml) during session launch in references/launch.md.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect injection via processed external data.
Ingestion points: Ingests model specifications, tool arguments, and Hugging Face Hub metadata in SKILL.md and references/models.md.
Boundary markers: None identified in the prompt interpolation for harbor launch or harbor pull commands.
Capability inventory: Includes broad shell execution (bash), Docker service management, and remote code fetching capabilities across reference files.
Sanitization: No evidence of input validation or escaping for the arguments passed to the harbor or tool CLI calls.

run-llms