The Agent Skills Directory

[COMMAND_EXECUTION]: The test file tests/test_results.py uses subprocess.run to launch benchmarking scripts during offline validation. This is a standard testing pattern and is confined to local script execution.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to user-defined LLM provider endpoints to perform evaluations. These operations are essential to the skill's primary function and require the user to provide their own API keys and base URLs.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text returned from third-party LLM endpoints. It mitigates risks by parsing this data structurally (JSON) for performance metrics rather than executing it as instructions.
Ingestion points: LLM response bodies are read into memory in availability_probe.py, fidelity_probe.py, speed_probe.py, usecase_runner.py, and protocol_probe.py.
Boundary markers: Not explicitly implemented for model responses.
Capability inventory: The skill uses aiohttp and urllib for network requests and writes result artifacts to local storage.
Sanitization: Responses are processed via structured JSON parsing and specific behavioral pattern matching (e.g., canary echo detection).

llm-eval-harness