The Agent Skills Directory

[COMMAND_EXECUTION]: Python scripts within the evaluation directory (eval/run_harness.py and eval/target_agent_runner.py) utilize subprocess.run to execute commands. This functionality is used to run a marketplace agent within a Docker container for behavioral testing. The harness enforces security by requiring digest-pinned images and utilizing restrictive execution flags such as --network none and --read-only.- [EXTERNAL_DOWNLOADS]: The SKILL.md instructions recommend using the npx skills@latest command. This facilitates fetching and executing tools from the standard NPM registry for repository setup and management.- [PROMPT_INJECTION]: The skill implements an evaluation framework that processes fixture repositories, creating a surface for indirect prompt injection. This is addressed by the following architecture: (1) Ingestion points: Untrusted fixture content is read from eval/fixtures/repos/ into the agent workspace. (2) Boundary markers: The harness utilizes Docker with --network none and --read-only root to isolate the agent from the host and network. (3) Capability inventory: The agent's capabilities are restricted to generating JSON action artifacts within a confined /tmp directory, without repository write access. (4) Sanitization: Artifacts are validated by validate-harness-results.py against defined expectations to ensure compliant and safe behavior.

skills-marketplace-readiness