The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The file grid_eval.py contains a hardcoded authentication token sk-dev-proxy-123 within the _agent_judge function, which is used to authenticate requests to the local evaluation proxy at http://localhost:4001.
[DATA_EXFILTRATION]: The shell script run.sh contains logic to source environment variables from a hardcoded absolute path: ${HOME}/workspace/experiments/sparta/.env. This exposes sensitive configurations and credentials from external directories to the skill's execution environment.
[COMMAND_EXECUTION]: The eval_app.py module uses subprocess.run to dynamically execute pip install commands at runtime if the typer or rich packages are missing, bypassing standard package management practices.
[EXTERNAL_DOWNLOADS]: The pyproject.toml file defines a dependency on the scillm package, which is retrieved directly from a GitHub repository (github.com/grahama1970/scillm).
[PROMPT_INJECTION]: The skill facilitates automated LLM evaluation by loading ground-truth JSON files and interpolating their contents into model prompts and automated judge instructions, creating a surface for indirect prompt injection.
Ingestion points: External ground-truth JSON files loaded in find_minimum.py and grid_eval.py.
Boundary markers: None present; data is interpolated directly into message arrays.
Capability inventory: Shell command execution via subprocess.run and network requests via httpx to local service endpoints.
Sanitization: Input content is stripped but not sanitized for instruction-carrying patterns.

llm-eval-lab