The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell scripts and Node.js process management to handle a local brainstorming server. Specifically, start-server.sh and stop-server.sh manage background tasks using nohup, ps, and kill.
[DYNAMIC_EXECUTION]: The companion server script (server.cjs) dynamically generates HTML documents at runtime by injecting content fragments into a template. It also serves local files (images, scripts, etc.) from a temporary session directory. While this involves dynamic code generation, it is central to the skill's purpose and uses templates rather than arbitrary execution of untrusted input.
[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface by ingesting untrusted browser interaction data through a WebSocket connection.
Ingestion points: Browser events (clicks and selections) are recorded as JSON in the state_dir/events file and then read by the agent as part of the brainstorming flow (SKILL.md).
Boundary markers: The interaction is technically constrained to specific data-choice attributes and structured JSON, but the agent integrates this external data into its reasoning context.
Capability inventory: The server script (server.cjs) has file system access (read/write) restricted to the session directory via path.basename and basic process control capabilities through shell scripts.
Sanitization: The instructions explicitly mention redacting sensitive information (secrets/PII) before sharing context with subagents, providing a procedural safeguard.

brainstorming