The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses local shell scripts (start-server.sh, stop-server.sh) and Node.js (server.cjs) to manage a visual companion server. It uses child_process to open the system browser when screens are ready. The process management logic includes verification of process identity (via instance IDs) before termination to avoid affecting unrelated processes.
[DATA_EXPOSURE]: The companion server implements robust authorization via a session-specific hex token generated at startup. This token is passed in the URL and stored in a secure cookie with HttpOnly, SameSite=Strict, and Path=/ flags. The server also implements path traversal protection using fs.realpathSync to ensure served files are strictly within the designated content directory.
[INDIRECT_PROMPT_INJECTION]: The skill processes structured interaction data from the visual companion (clicks recorded in a JSONL events file). While this creates a surface for indirect instructions, the data is handled as structured values rather than raw prompts. Furthermore, the workflow includes human approval gates and a subagent verification step ("builder check") for the final design document, which mitigates the risk of malicious instructions influencing the final output.
[DYNAMIC_EXECUTION]: The server implementation manually handles WebSocket frames (RFC 6455) and HTTP requests using only Node.js built-in modules, significantly reducing the supply chain risk associated with external dependencies.

brainstorming