The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses npx playwright screenshot to capture the dashboard state. The target URL is dynamically determined by reading dashboard/vite.config.ts or dashboard/package.json. If these files contain shell metacharacters (e.g., in the port number or script flags), it could lead to command injection when the agent executes the Playwright command.
[DYNAMIC_EXECUTION]: In Phases 2 and 7.2, the instructions state that if the Playwright CLI is unavailable, the agent should 'use a small inline Playwright script.' This directs the agent to generate and execute code at runtime to interact with the browser, which is a form of dynamic code execution.
[INDIRECT_PROMPT_INJECTION]: The skill has a vulnerability surface for indirect prompt injection through dashboard screenshots and project files.
Ingestion points: Screenshots of the dashboard URL (Phase 2, 5, 7.2, 7.6) and content of RoomBuilder.ts/AgentSprite.ts (Phase 7.1).
Boundary markers: None specified for the analysis of screenshots or code files.
Capability inventory: Playwright screenshot execution, file system writes (copying assets), and file modifications to source code.
Sanitization: No explicit sanitization or validation of the dashboard content or configuration values is mentioned.
[REMOTE_CODE_EXECUTION]: The workflow in Phases 4 and 7.5 involves 'dispatching subagents' in parallel to implement changes. While this is an agent platform feature, the coordination and passing of instructions between agents for file-system modifications represents a significant capability that relies on the underlying platform's security boundaries.

opensquad-dashboard-design