The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides a recipe in references/recipes/click-cli.md for auto-discovering commands by dynamically compiling and executing Python files in a plugin directory using exec(compile(f.read(), f.name, "exec"), ns, ns).
[REMOTE_CODE_EXECUTION]: Recipes in references/recipes/service-patterns.md and references/recipes/ray-training.md demonstrate dynamic module loading and execution using importlib.import_module and getattr to resolve and run functions from string-based module paths.
[COMMAND_EXECUTION]: The references/recipes/ray-training.md file includes patterns for executing shell commands via subprocess.run, including calls to the AWS CLI (aws s3 sync) and a custom SubprocessExecutor that launches tasks using uv run.
[EXTERNAL_DOWNLOADS]: The documentation frequently instructs users to install external Python dependencies (such as ray, mlflow, and warp-lang) and utilize runtime_env.pip to provision environments on distributed worker nodes.
[PROMPT_INJECTION]: The RayExecutor and SubprocessExecutor implementations in references/recipes/ray-training.md ingest structured JobSpec data which is passed into dynamically loaded functions, creating an indirect prompt injection surface if inputs are derived from untrusted sources.

python