The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation framework included in the skill (specifically evals/grader.py and run_checks.py) utilizes subprocess.Popen to execute system tools such as the C++ compiler (g++ or clang++), Python interpreters, and compiled test binaries. These commands are essential for the skill's primary purpose of verifying code repairs and proof-of-concept implementations.
[DYNAMIC_EXECUTION]: The core functionality of the skill involves the generation, building, and execution of C++ or Java code to implement cuDF paths in query engines. This dynamic execution is a functional requirement and is documented for users. The evaluation tools automate this process to establish correctness on the host integration boundary.
[INDIRECT_PROMPT_INJECTION]: The skill operates as a source-code repair tool, which involves ingesting untrusted C++ code (e.g., adapter.hpp) and processing it through compilation and execution. This creates a surface for indirect prompt injection.
Ingestion points: External C++ source files (adapter.hpp) processed during evaluation tasks.
Boundary markers: The SKILL.md and TASK.md files provide instructions that define the scope of work and identify protected fixture files.
Capability inventory: The skill uses subprocess execution to compile and run binaries, as seen in grader.py and the various run_checks.py scripts.
Sanitization: The evaluation logic in run_checks.py implements resource constraints using the Python resource module, setting limits on CPU time, address space, and file output size to mitigate the risks of executing untrusted code.

cudf-query-engine-builder