The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill extensively uses shell-piped execution patterns to install critical development tools, a method that is susceptible to supply chain attacks. This includes the inference.sh CLI installer (curl -fsSL https://cli.inference.sh | sh), the astral.sh 'uv' installer (curl -LsSf https://astral.sh/uv/install.sh | sh), and the 'nvm' installer (curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.0/install.sh | bash). A similar pattern is used for Windows via PowerShell: powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex".
[EXTERNAL_DOWNLOADS]: Fetches scripts and binary installers from various remote sources, including the 'fnm' installer from Vercel's domain (https://fnm.vercel.app/install) and model assets via the Hugging Face Hub.
[COMMAND_EXECUTION]: The skill relies on the 'belt' CLI to perform high-privilege operations such as scaffold creation, local code testing, and cloud deployment of the application.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process agent-generated logic and configuration files for deployment, creating a significant indirect injection surface.
Ingestion points: Reads and processes code logic in inference.py and inference.js, as well as configuration in inf.yml.
Capability inventory: The skill has the capability to execute the generated code locally via belt app test and deploy it to a live environment using belt app deploy.
Boundary markers: None. The instructions do not provide delimiters or warnings to the agent regarding the handling of potentially untrusted logic or instructions embedded within the application code.
Sanitization: No validation, escaping, or security scanning is performed on the generated logic before it is executed or deployed.

building-inferencesh-apps