The Agent Skills Directory

[COMMAND_EXECUTION]: The skill launches a local Node.js server using the scripts/start-server.sh script. This server is part of the 'Visual Companion' feature and is used to render HTML mockups and diagrams for the user to view in a browser.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the browser companion. The agent is instructed to read structured interaction data from a local file (STATE_DIR/events). This file is populated by the Node.js server based on events received from the browser. If the server is exposed to the network (e.g., by using the --host 0.0.0.0 option), an attacker could potentially send malicious events that influence the agent's behavior during the brainstorming process.
Ingestion points: The agent reads user feedback and interaction data from STATE_DIR/events in the session directory.
Boundary markers: Absent. The instructions do not specify delimiters or warnings for the data read from the events file, relying on the agent to merge it with terminal text.
Capability inventory: The skill has the authority to write files to the project's documentation directory, commit changes to git, and trigger the writing-plans implementation skill.
Sanitization: No explicit sanitization or validation of the JSON-formatted interaction data is performed before the agent processes it.

brainstorming