The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the nemo_gym package directly from NVIDIA's official GitHub repository (https://github.com/NVIDIA-NeMo/Gym). This is a legitimate dependency for the described framework.- [INDIRECT_PROMPT_INJECTION]: The generated code for the verify endpoint creates a surface for indirect prompt injection by processing model-generated trajectories to determine rewards.
Ingestion points: The verify method in server.py and the architecture described in references/architecture.md ingest untrusted model data via body.response.output.
Boundary markers: No specific delimiters or safety instructions are included in the generated boilerplate to distinguish between data and embedded instructions.
Capability inventory: The scaffolded server is capable of exposing arbitrary tool endpoints (app.post) and interacting with session state.
Sanitization: The provided reward grading logic (e.g., substring matching) does not sanitize or validate model outputs before processing.- [COMMAND_EXECUTION]: The skill utilizes uv run for executing local Python scripts and validation commands, and ng_run for launching the environment via the NeMo Gym CLI. These are standard operations for environment development.

generate-nemo-gym-env