The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to fetch model implementations and source code from Hugging Face Hub and GitHub for research and integration purposes. These references target well-known services and are intrinsic to the skill's primary purpose.
[COMMAND_EXECUTION]: The workflow involves executing standard development and profiling tools including docker, bash, nvidia-smi, nsys, and ncu. These commands are used for environment setup and performance benchmarking within a controlled containerized environment.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted code from external repositories to perform autonomous kernel creation.
Ingestion points: The agent reads transformers library source code and remote model code from the Hugging Face Hub (e.g., modeling_*.py files).
Boundary markers: The instructions mandate the use of precise GitHub permalinks with line anchors for code references to ensure provenance.
Capability inventory: The skill uses filesystem operations to create new kernel files and executes shell commands for testing and benchmarking within Docker containers.
Sanitization: The skill does not explicitly document sanitization of the ingested code, relying on the agent's research logic and Docker isolation to manage risks.

tilegym-monkey-patch-kernels-to-transformers