The Agent Skills Directory

[Command Execution]: The skill uses subprocess.run() to verify the agent's capability to execute shell commands. This is an opt-in check (active-safe) and is limited to running a benign Python command: python -c "print('ok')". It does not execute arbitrary user input.
Evidence found in: scripts/inspect_runtime.py
[External Downloads / Network Operations]: The skill performs a single network request to https://pypi.org to verify if the agent has internet access. This is an opt-in check (active-safe) using a well-known, high-availability service and does not download executable content or transmit sensitive data.
Evidence found in: scripts/inspect_runtime.py
[Indirect Prompt Injection Surface]: The skill ingests external data in the form of observations.json to document available tools and skills. While this represents a potential injection surface, the skill implements specific safety boundaries, uses structured data (JSON), and performs slugification on inputs to minimize risk. As per the security analysis protocol, this is classified as a low-severity surface.
Evidence found in: scripts/inspect_tools.py and references/safety-boundaries.md
[Filesystem Operations]: The skill includes an opt-in check to verify filesystem write permissions by creating and immediately deleting a single temporary file. This process is self-cleaning and adheres to the skill's documented safety protocol.
Evidence found in: scripts/inspect_runtime.py

agent-harness-explorer