The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in run_pipeline.py to execute a local demonstration script (agent_demo.py). While this is a common pattern for integrated workflows, it involves shell execution of local files.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied datasets (.jsonl files) for model training and evaluation. Maliciously crafted prompts within these datasets could influence the agent's behavior during the evaluation phase if the dataset contains instructions aimed at the agent itself rather than the model being trained.
Ingestion points: run_pipeline.py reads data from paths provided in --train and --eval arguments.
Boundary markers: The prompt construction in run_eval adds a /no-think tag but lacks robust boundary markers to separate untrusted data from system instructions.
Capability inventory: The skill can execute local scripts via subprocess.run, write to the filesystem, and perform network requests to Fireworks AI and OpenRouter APIs.
Sanitization: The script performs standard JSON parsing but does not sanitize the content of the messages for potential injection patterns before sending them to the API.
[CREDENTIALS_UNSAFE]: The skill instructions and code (run_pipeline.py, agent_demo.py) explicitly reference loading sensitive API keys (FIREWORKS_API_KEY, OPENROUTER_API_KEY) from a local .env file. While storing secrets in .env is a standard practice, the skill's reliance on these environmental secrets for core functionality is noted.

grpo-finetune