The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to interact with the host system by executing a local command-line interface (CLI) named router. The instructions in SKILL.md direct the agent to run various subcommands such as router run, router effort, and router status to manage model sessions and quotas.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input by interpolating it into shell commands.
- Ingestion points: User-provided strings are captured as <task> and <sub-step> in SKILL.md (e.g., router run "<task>").
- Boundary markers: The instructions use double quotes around the interpolated variables in the command templates, which serves as a basic syntactic boundary but does not prevent shell metacharacter injection if the agent fails to escape the content.
- Capability inventory: The skill has the capability to execute arbitrary shell commands via the router CLI utility as specified in SKILL.md and references/cli.md.
- Sanitization: There are no explicit instructions for the agent to sanitize, escape, or validate the user-provided task descriptions before they are passed to the shell environment.

model-router