The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill architecture facilitates the installation and execution of arbitrary code from external sources. The 'Extension System' section describes a workflow where users install extensions by providing a GitHub URL (e.g., via installExtension). These extensions contain Python logic (generator.py) that the local backend executes to perform 3D generation tasks.
[DYNAMIC_EXECUTION]: The application dynamically loads and executes Python code from computed file paths within the local extensions directory. The generate function in the extension's generator.py acts as a dynamic entry point for executing script logic that is not part of the core application distribution.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading content from external repositories and services. This includes cloning GitHub repositories for extensions and downloading large model weights (e.g., weights.safetensors) from Hugging Face based on metadata provided in an extension's manifest.json file. While Hugging Face is a well-known service, the execution of accompanying scripts from arbitrary repos presents a risk.
[COMMAND_EXECUTION]: The installation guide instructs users to execute shell commands to modify file permissions (chmod +x launcher.sh) and install dependencies using package managers (npm install, pip install -r requirements.txt), establishing a pattern of local command execution.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection via untrusted data ingestion.
Ingestion points: The repoUrl parameter in the installExtension IPC call and URLs within the manifest.json file.
Boundary markers: None identified in the provided documentation.
Capability inventory: Includes arbitrary Python code execution, file system writes for mesh export, and network access for downloading models.
Sanitization: No evidence of sanitization or validation for the content of the external GitHub repositories or model files before execution.

modly-image-to-3d