The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze and import code from user-provided repositories, creating an attack surface for malicious source code.
Ingestion points: Workflow 1 (Static Analysis) and Workflow 4 (End-to-End Adaptation) in SKILL.md read and import source code from user-defined project paths.
Boundary markers: The instructions lack explicit delimiters or warnings to ignore potentially malicious instructions embedded in the analyzed code.
Capability inventory: The skill utilizes grep for searching, import for module loading, and torch.load for loading model weights from user projects.
Sanitization: No sanitization or sandboxing is performed on the user-provided code before it is imported into the agent's environment.
[EXTERNAL_DOWNLOADS]: The skill downloads and installs dependencies from official vendor and well-known repositories.
Fetches Ascend developer tools from the official Gitee repository: https://gitee.com/ascend/tools.git.
Downloads Python wheels for aclruntime and ais_bench from Huawei's Open Buckets Service at https://aisbench.obs.myhuaweicloud.com.
[DYNAMIC_EXECUTION]: The skill instructions (SKILL.md, Workflow 1, Phase 2) require the agent to generate and execute temporary Python scripts to probe the metadata and runtime behavior of user models.

atc-model-converter