The Agent Skills Directory

Privilege Escalation (HIGH): The script scripts/init-artifact.sh executes npm install -g pnpm if the package manager is missing. Installing global packages typically requires administrative or root privileges and modifies the global system environment.
Unverifiable Dependencies (MEDIUM): The skill installs approximately 60 Node.js packages across two scripts (init-artifact.sh and bundle-artifact.sh). While many are standard libraries (e.g., React, Tailwind), the use of unpinned versions or 'latest' tags creates a large attack surface for supply chain compromises.
Command Execution & Opaque Data (HIGH): In scripts/init-artifact.sh, the command tar -xzf "$COMPONENTS_TARBALL" -C src/ extracts a local archive (shadcn-components.tar.gz) into the source directory. The contents of this archive are not provided for auditing, meaning it could contain malicious scripts or obfuscated code that is subsequently executed during the pnpm exec parcel build step.
Dynamic Execution (MEDIUM): The skill uses node -e to programmatically modify tsconfig.json and tsconfig.app.json via inline JavaScript. This involves parsing and rewriting JSON files at runtime, which can be risky if input variables (like $PROJECT_NAME) are manipulated.
Indirect Prompt Injection (LOW): As a generator for HTML 'artifacts', the skill creates a surface where malicious instructions could be embedded in the processed data.
Ingestion points: User-edited code in the src/ directory.
Boundary markers: None identified in the bundling script.
Capability inventory: Subprocess calls for pnpm, parcel, and html-inline in scripts/bundle-artifact.sh.
Sanitization: None; the script inlines all assets into a single HTML file without validation.

web-artifacts-builder