The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions and usage patterns involve significant use of command-line tools. The compute_reward method in SKILL.md utilizes ctx.terminal("pytest /workspace/test.py") to execute shell commands for verifying agent outputs. Additionally, usage-patterns.md provides multiple CLI examples for running the environments in serve, process, and evaluate modes, which involve executing local Python scripts and managing virtual environments.
[EXTERNAL_DOWNLOADS]: The skill describes using the datasets library in the setup() method of environments to fetch data from Hugging Face. It also integrates with Weights & Biases (wandb) for logging training and evaluation metrics, which are standard practices and utilize well-known services.
[DATA_EXFILTRATION]: The skill provides clear guidance on handling sensitive information, specifically instructing users to use environment variables for the OPENROUTER_API_KEY rather than hardcoding credentials. This is a positive security practice for managing external API access.
[INDIRECT_PROMPT_INJECTION]: The skill defines a framework where untrusted data is ingested from external datasets and subsequently used to generate prompts for an AI agent.
Ingestion points: External data is loaded via datasets.load_dataset in the setup() method (SKILL.md).
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands are implemented in the format_prompt logic.
Capability inventory: The environment includes capabilities to execute shell commands via ctx.terminal in the compute_reward function and runs a multi-turn agent loop in the evaluate method (SKILL.md).
Sanitization: There is no mention of sanitizing or escaping the content retrieved from external datasets before it is presented to the agent.

hermes-atropos-environments