The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to use Docker CLI commands (docker run, docker build, docker create, docker exec) to execute code within isolated containers. It uses volume mounting (-v "$(pwd)":/workspace) to allow file sharing between the host and the container for research purposes.
[EXTERNAL_DOWNLOADS]: The skill references several official and well-known base images from Docker Hub, including python:3.11, pytorch/pytorch:latest, tensorflow/tensorflow:latest-gpu, node:20, and ubuntu:24.04. These images provide the runtime environment for research tasks.
[PROMPT_INJECTION]: This skill possesses an indirect prompt injection surface as it is specifically designed to ingest and execute untrusted code from external research repositories.
Ingestion points: The host's current working directory is mounted directly into the container, exposing the project files to the execution environment.
Boundary markers: Isolation is enforced through standard Docker containerization primitives (namespaces and control groups).
Capability inventory: The skill uses the Bash(docker:*) tool to install dependencies via pip and execute arbitrary scripts like train.py within the container.
Sanitization: No specific sanitization of the research code or its output is documented, as the container itself serves as the security boundary.

docker