The Agent Skills Directory

[COMMAND_EXECUTION]: The skill starts a local Node.js server (server.cjs) to power the visual companion and uses platform-specific commands (open, rundll32.exe, xdg-open) to automatically launch the user's browser.
[DATA_EXFILTRATION]: The visual companion server fetches a branding image from primeradiant.com and appends the SUPERPOWERS_VERSION as a query parameter, which serves as a basic version-tracking telemetry mechanism.
[INDIRECT_PROMPT_INJECTION]: The skill processes user interactions from the browser tab (such as clicking on design choices) by reading a local JSON-formatted events file (state_dir/events). The agent incorporates this feedback into the conversation. The attack surface is limited as the agent generates the browser content, and interactions are stored in a structured format.
[CREDENTIALS_UNSAFE]: The skill correctly implements security best practices by generating a 32-byte session token for each server instance. This token is required via a query parameter or cookie for all HTTP and WebSocket requests, preventing unauthorized local or network access to the companion server. Tokens are stored in owner-only files with restricted permissions.
[REMOTE_CODE_EXECUTION]: The server implementation includes robust security checks when serving local assets, including resolving real paths to prevent directory traversal and checking for symbolic or hard links to ensure only regular files are accessed.

brainstorming