The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the py-openjudge package from PyPI, which is the core framework for the reward logic.
[DYNAMIC_EXECUTION]: The skill utilizes CodeExecutionGrader (e.g., in pointwise.md) to score programming tasks by executing the model's generated response. This involves running untrusted code generated by an AI model, which is a standard requirement for RL code-generation tasks but represents an execution surface.
[INDIRECT_PROMPT_INJECTION]: The skill implements a judge-based reward system that is susceptible to indirect prompt injection.
Ingestion points: Untrusted model rollouts are ingested via the rollouts list in pointwise.md and the pairs_dataset in pairwise.md.
Boundary markers: Prompt templates use headers like Query:, Response 1:, and Response 2: to separate fields, but they lack explicit instructions for the judge LLM to disregard any commands or overrides contained within those response fields.
Capability inventory: The system possesses capabilities including arbitrary code execution (via CodeExecutionGrader) and network access for LLM judging (via OpenAIChatModel targeting the Alibaba DashScope API).
Sanitization: There is no evidence of input sanitization or escaping to prevent adversarial content in a rollout from influencing the reward score.

02-rl-reward