The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The installation instructions in SKILL.md direct users to clone a repository (https://github.com/radixark/miles.git) and execute pip install -e .. This procedure involves downloading and running the setup.py or pyproject.toml configuration files from an unverified third-party source.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to pull and run a Docker image (radixark/miles:latest) and references several GitHub repositories belonging to users or organizations not recognized as trusted vendors or well-known services (e.g., radixark/miles, THUDM/slime, sgl-project/sglang, and zhaochenyang20/Awesome-ML-SYS-Tutorial). These represent unverified external dependencies.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its ingestion of untrusted external content.
Ingestion points: The skill processes model checkpoints via the --hf-checkpoint flag and training data via the --prompt-data flag (documented in SKILL.md).
Boundary markers: No delimiters, XML tags, or instructions to ignore embedded commands are present in the training workflows to protect the agent from malicious content within these files.
Capability inventory: The training framework executes complex shell commands (e.g., python train.py) and performs extensive file system and hardware (GPU) operations.
Sanitization: There is no evidence of validation or sanitization of the training data or checkpoint metadata before they are processed by the training logic.

miles-rl-training