The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions specify the use of npx debug-bridge-feedback-mcp to initialize an MCP server, which involves downloading a package from the NPM registry at runtime.
[REMOTE_CODE_EXECUTION]: The execution of the debug-bridge-feedback-mcp package via npx constitutes a remote code execution pattern as it fetches and runs external code to provide tool capabilities to the agent.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a loop where the agent is instructed to arm a watcher and wait for 'change requests' from a sidecar dock and a browser environment. This ingestion of untrusted data represents an indirect prompt injection surface.
Ingestion points: Browser DOM snapshots, element attributes, screenshots, and natural language descriptions provided by the user through the Design Mode sidecar dock.
Boundary markers: The runbook lacks explicit instructions for the agent to use delimiters or to disregard potentially malicious instructions embedded within the browser content or dock messages.
Capability inventory: The agent has the ability to execute CLI commands (debug-bridge), write files to local storage (/tmp), and execute arbitrary JavaScript within a browser context.
Sanitization: The skill does not describe any validation or sanitization mechanisms for the data received from the bridge before it is incorporated into the agent's reasoning or code-modification tasks.
[DYNAMIC_EXECUTION]: The skill includes an eval capability and a browser_control evaluate tool, which allow the agent to execute arbitrary JavaScript code within the target browser's execution context.
[COMMAND_EXECUTION]: The operation of the skill relies on executing multiple CLI tools, including debug-bridge for session management and tmux for process isolation.

debug-bridge