The Agent Skills Directory

[SAFE]: The skill implements a local HTTP and WebSocket server (scripts/server.cjs) to support browser-based visual aids. The implementation follows security best practices, including per-session token authentication (32-byte hex token), binding to the loopback interface (127.0.0.1), and utilizing fs.realpathSync to prevent directory traversal when serving files. The server also employs security headers like Content-Security-Policy, X-Frame-Options: DENY, and Cross-Origin-Resource-Policy to mitigate common web vulnerabilities.
[COMMAND_EXECUTION]: The server process uses child_process.execFile to invoke the system's default browser. It utilizes hardcoded platform-specific binaries (such as open, xdg-open, or rundll32.exe) and passes the session URL as a single argument, which prevents shell command injection.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest existing project data and user interaction events to inform its brainstorming process.
Ingestion points: Existing project files, documentation, and git commits are read to establish context (SKILL.md). User interaction events (clicks and choices) are read from a JSONL file in the session state directory (scripts/server.cjs).
Boundary markers: The instructions explicitly forbid the agent from writing implementation code or building scaffolding until the design specification has been formally approved by the user.
Capability inventory: The skill can start a local Node.js server, write files to the project directory (under docs/specs/), and invoke subagents for document review.
Sanitization: Browser-generated interaction events are parsed and validated as structured JSON to ensure data integrity.

brainstorming