The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from existing canvases and brand kits which could contain malicious instructions.\n
Ingestion points: The agent reads data via moda canvas read (in references/reading-and-verifying.md) and moda brand show (in references/brand.md).\n
Boundary markers: The documentation includes explicit instructions to treat canvas content as data rather than instructions and defines an agent_instructions metadata field as context only.\n
Capability inventory: The agent can execute commands via Bash(moda:*), run sandboxed JavaScript with moda canvas edit, and modify the canvas with moda canvas markup.\n
Sanitization: JavaScript execution is restricted to a sandbox that blocks sensitive APIs (e.g., eval, require, fetch), and markup parsing follows a strict XML grammar.\n- [DYNAMIC_EXECUTION]: The skill implements a sandboxed JavaScript environment for manipulating canvas nodes and animations at runtime.\n
Evidence: The moda canvas edit command (detailed in references/edit-code.md) and the <generate> markup tag (in references/markup.md) execute agent-generated code. The sandbox provides a Math facade while blocking access to global objects like process, document, and console to prevent breakout.\n- [COMMAND_EXECUTION]: The skill relies on a specialized command-line interface (moda) for all operations, including media generation and file management.\n
Evidence: The Bash(moda:*) tool is used to perform subcommands such as canvas markup, media generate-video, and file upload. These interactions are restricted to the vendor's CLI namespace but allow broad manipulation of the platform's assets.

moda-video-motion