The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of hardware diagnostic tools (e.g., mthreads-gmi, musaInfo) and distributed training launchers (e.g., torchrun) as part of its core development workflow for MUSA accelerators.
[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest external repository data such as AGENTS.md and environment documentation, creating a potential surface for indirect prompt injection. However, it includes explicit constraints to prevent bypassing repository-level security or usage restrictions.
Ingestion points: Repository-level AGENTS.md, environment documentation, and experiment rules (SKILL.md).
Boundary markers: Present; the instructions explicitly state that repository documentation remains authoritative and forbid the agent from bypassing restrictions on GPU runs, data preparation, or artifact paths.
Capability inventory: Shell command execution for hardware status, build system invocation for MUSA extensions, and distributed process management via torchrun.
Sanitization: No explicit sanitization or filtering of external repository content is documented.
[DYNAMIC_EXECUTION]: The skill provides templates for compiling native MUSA extensions (using .mu and .cpp sources) via setuptools and torch_musa build utilities. It also describes runtime API adaptation using torchada to maintain compatibility with legacy CUDA codebases.

musa-pytorch-code-porting