The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes an internal platform script at .agents/skills/platform/skill-lab.mjs to execute Node.js commands. The documentation specifies that these commands are run without a shell and within a minimal environment to reduce the risk of command injection.
[INDIRECT_PROMPT_INJECTION]: The skill provides an interface to process untrusted data through fixture definitions and command arrays, which constitutes an attack surface for indirect injection.
Ingestion points: Untrusted data enters via the fixture-definition and command inputs defined in SKILL.md.
Boundary markers: The skill restricts output to the system temporary directory and avoids shell-based command interpolation.
Capability inventory: The skill possesses file writing capabilities (write-files) and the ability to execute sandboxed processes (run-sandboxed-command).
Sanitization: Is dependent on the skill-sandbox component to enforce execution boundaries and environment isolation.

integration-playground