The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements WebSocket message handlers that process data from external servers, which represents a potential surface for indirect prompt injection if the received data is subsequently interpolated into an LLM prompt context.\n
Ingestion points: socket.onmessage event handlers located in examples/core.md, examples/presence.md, and examples/binary.md.\n
Boundary markers: The code utilizes try-catch blocks for JSON.parse and recommends discriminated unions for exhaustive message type handling, providing basic structural validation to ensure messages conform to expected schemas.\n
Capability inventory: The skill's scripts are restricted to browser-based operations, including React state management, DOM event listeners, and network communication via the native WebSocket API. No sensitive file system writes or system-level command executions are present.\n
Sanitization: Data is validated against expected types and structures (e.g., instanceof ArrayBuffer checks, type narrowing in switch statements, and string-only checks in presence parsing).

web-realtime-websockets