The Agent Skills Directory

[COMMAND_EXECUTION]: Runtime compilation of CUDA source files. The evaluation script evals/files/cudf-native-stream-handoff-boundary/code/run_smoke.sh invokes nvcc to compile a local C++ file (threaded_handoff.cu). This behavior is associated with the primary purpose of the skill for GPU development.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands for library activation and system monitoring, such as python -m cudf.pandas and nvidia-smi.
[PROMPT_INJECTION]: Indirect prompt injection surface identified through data ingestion. The skill processes external data via read_csv and read_parquet functions, which could ingest malicious instructions if the data is attacker-controlled.
Ingestion points: read_csv and read_parquet in SKILL.md, references/dask-cudf-patterns.md, and evaluation scripts like evals/files/cudf-csv-etl/code/etl_pipeline.py.
Boundary markers: No explicit boundary markers or warnings about embedded instructions in data are present.
Capability inventory: Shell command execution (python -m, nvcc) and file system writes (to_csv, to_parquet).
Sanitization: No explicit sanitization or validation of external data content is described.

accelerated-computing-cudf