The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to search for and install external tools using the command npx skills add <owner/repo@skill> -g -y. The use of the -y flag is high-risk as it explicitly bypasses human confirmation for installing and running third-party code. While the skill recommends checking installation counts and GitHub stars, these metrics are not reliable security indicators and can be manipulated by attackers.
[COMMAND_EXECUTION]: The skill frequently invokes shell commands to manage technical stacks, including uv, conda, poetry, and npm. It also includes a custom script scripts/detect_stack.py that is executed via the command line to parse project files and determine which tools to recommend, which could be exploited if run against a maliciously crafted project.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading and executing remote code via the Model Context Protocol (MCP) and the 'agent-skills' ecosystem. Commands like npx -y @modelcontextprotocol/server-* and uvx mcp-server-* download and run packages from npm and PyPI registries at runtime.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It processes untrusted external data from project manifest files (e.g., package.json, pyproject.toml) using scripts/detect_stack.py. This data is then used to influence the agent's tool selection and installation behavior.
Ingestion points: The scripts/detect_stack.py script reads dependency lists from user-provided project directories.
Boundary markers: No explicit delimiters or instructions are used to separate untrusted project data from the agent's internal reasoning.
Capability inventory: The skill possesses extensive capabilities including global tool installation (npx skills add -g), shell command execution, and file system access.
Sanitization: The script performs basic parsing (JSON/TOML) but does not validate or sanitize dependency names before they are used in recommendations or potentially interpolated into further commands.

light-tool-selection