The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands based on user-provided inputs, specifically git clone <repo_url> and ls -la <local_path> during the initial code acquisition phase. Additionally, it runs generated verification scripts and testing tools such as verify_npu.sh and pytest during the validation phase.
[EXTERNAL_DOWNLOADS]: The skill references and provides instructions to download the CANN Toolkit from official Huawei cloud infrastructure (e.g., ascend-repo.obs.cn-east-2.myhuaweicloud.com). These downloads target a well-known technology service provider.
[REMOTE_CODE_EXECUTION]: The reference documentation provided with the skill includes instructions for downloading and executing installation scripts (e.g., bash Ascend-cann-toolkit_...run --install). These execution patterns are associated with official vendor resources from Huawei.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it analyzes external, potentially untrusted code repositories while possessing high-privilege capabilities.
Ingestion points: Source code files obtained via git clone or local directory scanning as described in Phase 1.
Boundary markers: The skill instructions do not specify the use of delimiters or provide directives for the agent to ignore instructions embedded within the analyzed source code.
Capability inventory: The skill allows the agent to execute shell commands, run Python scripts, and write new files to the local environment.
Sanitization: There are no documented mechanisms for sanitizing or validating the content of the analyzed files before they are processed by the agent.

npu-adapter-reviewer