The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts scripts/verify.py and scripts/benchmark.py utilize importlib.import_module to dynamically load and execute Python classes from a provided verification directory. This allows the skill to test various operator implementations against reference versions.
[COMMAND_EXECUTION]: scripts/verify.py uses subprocess.Popen to execute a copy of itself as a sub-process. This architecture is used to enforce execution timeouts and provide a layer of process isolation during kernel verification.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and execute code files which may be produced by other agents or influenced by untrusted user inputs. * Ingestion points: Code and task files are read from paths provided in the generated-code-path and task-file-path arguments. * Boundary markers: The skill does not implement explicit prompt boundary markers. * Capability inventory: The skill has the capability to execute code via importlib and manage processes via subprocess. * Sanitization: The scripts/validate_triton_impl.py script performs AST analysis to verify that the generated code contains expected Triton kernel decorators and does not contain prohibited PyTorch functional calls.

external-cannbot-ops-triton-op-verifier