The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a framework for processing external environment observations which serves as a potential surface for indirect instructions.\n
Ingestion points: Environment observation buffers in scripts/env_template.py and references/environments.md ingest data from simulated environments.\n
Boundary markers: Absent. The skill does not implement specific boundary markers or safety instructions for environment-generated data.\n
Capability inventory: The skill utilizes subprocess execution for training via torchrun and the puffer CLI (mentioned in SKILL.md and references/training.md), and performs file system operations for saving model checkpoints in scripts/train_template.py.\n
Sanitization: Absent. No explicit sanitization or filtering of environment observations is implemented before processing.\n- [DYNAMIC_EXECUTION]: The skill facilitates the use of runtime compilation and optimization techniques to achieve high throughput.\n
Evidence: references/environments.md provides a guide and code template for implementing and compiling C extensions as Python modules for performance.\n
Evidence: references/policies.md encourages the use of torch.compile for neural network model optimization.\n- [COMMAND_EXECUTION]: The skill utilizes shell commands for distributed training and task execution.\n
Evidence: SKILL.md and references/training.md provide examples of using torchrun for multi-GPU training and the puffer CLI for managing reinforcement learning tasks.\n- [EXTERNAL_DOWNLOADS]: The skill instructs on the installation of its core library and integration with external monitoring platforms.\n
Evidence: SKILL.md contains instructions to install the pufferlib package.\n
Evidence: references/training.md documents integration with well-known logging services such as Weights & Biases and Neptune.

pufferlib