The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses local shell scripts (start-server.sh, stop-server.sh) and a Node.js server (server.cjs) to manage visual mockup sessions. These scripts include robust lifecycle management, automatically shutting down the server if the parent process exits or after 30 minutes of inactivity.
[PROMPT_INJECTION]: The agent processes browser interaction data, which creates an indirect prompt injection surface.
Ingestion points: Interaction logs (clicks and text) are recorded in $STATE_DIR/events and read by the agent to refine designs, as described in references/visual-companion.md.
Boundary markers: No specific delimiters or safety instructions are used when interpolating this interaction data into the prompt.
Capability inventory: The skill can write design specifications to the local file system and transition to implementation planning skills.
Sanitization: Browser event data is merged into the agent's context without explicit validation or filtering.
[DATA_EXFILTRATION]: The visual companion server exposes files from a session-specific directory over HTTP/WebSocket. While restricted to localhost by default, the server includes path validation using path.basename to prevent directory traversal and unauthorized file access outside the intended session directory.

brainstorming