The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests structured interaction data from a local file (state_dir/events) that is populated by user actions in a browser via a WebSocket server.
Ingestion points: The agent reads event logs from the session state directory as described in visual-companion.md.
Boundary markers: The logs are stored in JSONL format without specific delimiters or explicit instructions to ignore embedded commands within the event fields.
Capability inventory: The skill can execute bash scripts for server management, run a Node.js server, and perform file system operations such as writing HTML designs and reading event logs.
Sanitization: The server performs standard JSON parsing on incoming WebSocket messages. The agent is instructed to interpret these events as structured interaction data to inform the brainstorming dialogue.
[COMMAND_EXECUTION]: The skill uses shell scripts (start-server.sh, stop-server.sh) to manage a local Node.js server (server.cjs). It utilizes system utilities like ps and kill for process monitoring and uses platform-specific commands such as open or xdg-open to launch the web browser.
[SAFE]: The implementation of the local visualization server incorporates several security best practices. It uses session-based authentication tokens to mitigate DNS rebinding and unauthorized access, employs directory traversal protections when serving files, and uses owner-only file permissions for sensitive session metadata. It also includes an automatic shutdown watchdog that terminates the server if the parent process dies or after a period of inactivity.

brainstorming