The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download a Docker image (radixark/miles:latest) and clone a Git repository (https://github.com/radixark/miles.git) from an unverified user account. While the skill author is listed as Orchestra Research, the software originates from a different, unauthenticated namespace.
[DYNAMIC_EXECUTION]: The skill documentation describes the ability to load and execute arbitrary Python scripts via specific command-line arguments.
Evidence: references/api-reference.md includes flags --custom-generate-function-path and --custom-rm-path which allow the execution of user-specified local scripts during the training process.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest external data for reinforcement learning training, creating a potential surface for instructions embedded in the data to influence the agent or the resulting model.
Ingestion points: SKILL.md (Workflow 1 and 2) references the ingestion of JSONL files via the --prompt-data flag.
Boundary markers: The instructions do not specify any delimiters or safety prompts to isolate the training content from the framework's operational logic.
Capability inventory: The skill executes a training framework (train.py) with capabilities including file system access for checkpoints and high-privilege hardware access (GPU IPC/shared memory).
Sanitization: There is no mention of validation, filtering, or sanitization of the training prompts before they are processed by the framework.

miles-rl-training