The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for using a specialized CLI (cinefoundry) and a Python one-liner to manage system memory. The Python script performs a filesystem walk and uses posix_fadvise with the POSIX_FADV_DONTNEED flag to evict specific files from the kernel page cache. This is a legitimate performance optimization for loading large AI models on shared memory systems.
[SAFE]: The skill monitors system memory state by reading from /proc/meminfo, which is a standard method for obtaining diagnostic information about free and available memory. The logic specifically addresses the behavior of NVIDIA GB10 hardware where CUDA allocations do not count towards container cgroups.
[INDIRECT_PROMPT_INJECTION]: The skill provides a workflow for planning memory usage where user-provided requirements (e.g., model names and memory sizes) are passed to a planning tool. While this represents a data ingestion surface, the skill instructs the agent on appropriate measurement and planning steps rather than promoting unsafe data handling.

spark-model-scheduler