The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's installation script (scripts/check_install_cli.sh) automatically downloads and installs the 'swiftopenai-cli' package from the NPM registry using 'npm install -g' if it is missing or outdated. While this is necessary for the skill's functionality, it involves executing code from an external repository at runtime.- [COMMAND_EXECUTION]: The skill performs multiple command-line operations, including global package management with NPM and environment configuration via the 'swiftopenai' CLI tool.- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because user instructions are passed directly to the 'swiftopenai agent' command. This CLI tool supports the Model Context Protocol (MCP), allowing it to interact with the filesystem and other services. A malicious prompt could potentially lead to unauthorized tool execution if the underlying agent is configured with active MCP servers.
Ingestion points: User task input in the core workflow and usage patterns.
Boundary markers: Input is wrapped in double quotes in shell templates, but no specific delimiters or system-level instructions are used to distinguish user input from agent commands.
Capability inventory: The CLI tool explicitly supports 'mcp-servers' (e.g., filesystem, github) and 'allowed-tools' flags, enabling significant system interaction.
Sanitization: There is no evidence of input sanitization or output validation to prevent the agent from executing unintended commands based on model responses.

llm-router