The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs the agent to download binary wheel files (.whl) from a community GitHub repository (github.com/woct0rdho/SageAttention). This repository and its owner are not part of the verified or trusted organizations list, representing an unverified third-party source.\n- [REMOTE_CODE_EXECUTION]: The installation process utilizes pip install on remote binary packages. Executing these pre-compiled binaries can potentially allow for arbitrary code execution with the permissions of the local user.\n- [COMMAND_EXECUTION]: The skill provides instructions for the agent to construct and execute shell commands by interpolating variables discovered at runtime, such as the path to the Python executable and system cache paths (e.g., rmdir /s /q "%USERPROFILE%\.triton").\n- [INDIRECT_PROMPT_INJECTION]: The skill relies on environment data retrieved via tool calls (such as get_system_stats) to determine paths and versions for its execution logic. If this data is manipulated by an external actor, it could lead to the execution of unintended commands.\n
Ingestion points: Environment statistics and python paths provided by install_comfyui or get_system_stats (SKILL.md).\n
Boundary markers: No explicit boundary markers or ignore instructions are provided for the interpolated environment strings.\n
Capability inventory: The skill uses subprocess calls for pip installation, version checking, and file system deletion (SKILL.md).\n
Sanitization: The skill does not perform validation, escaping, or filtering on the strings derived from the environment before they are passed to shell commands.

triton-sageattention