The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell scripts (scripts/start-server.sh, scripts/stop-server.sh) and a Node.js server (scripts/server.cjs) to manage local process lifecycles. These scripts facilitate starting a background server, resolving parent process IDs for lifecycle management, and launching system web browsers using platform-native utilities like open or xdg-open via child_process.execFile.
[INDIRECT_PROMPT_INJECTION]: The visual companion feature creates a data feedback loop where the agent reads user interactions from the local file system.
Ingestion points: The agent is instructed in visual-companion.md to read the $STATE_DIR/events file, which contains JSON-formatted interaction data from the browser.
Boundary markers: The instructions lack explicit delimiting or validation steps for the content of the events file, relying on the agent to merge it with terminal dialogue.
Capability inventory: The skill includes file system manipulation (read/write), process management, and local networking capabilities.
Sanitization: There are no explicit instructions to sanitize or validate the JSON events ingested from the browser, which could allow a malicious user to inject instructions via the interaction channel.
[DYNAMIC_EXECUTION]: The agent is designed to generate and write HTML content fragments at runtime to screen_dir, which are then served and executed within the user's web browser environment. The server protects against path traversal by using realpathSync and verifying that all served files reside within the intended content directory.

sp-brainstorming