The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions specify the installation of the @aml-jsx/sdk Node.js package. This is a legitimate resource managed by the skill's author ('we-are-singular').\n- [COMMAND_EXECUTION]: The framework allows agents to execute shell commands and interact with the filesystem. These capabilities are governed by an explicit permission model (<Agent permissions={...}>) and are intended to be executed within isolated guest environments such as dockerSandbox or modalSandbox to prevent host compromise.\n- [CREDENTIALS_UNSAFE]: The documentation provides examples of managing sensitive API tokens (e.g., LINEAR_TOKEN, DAYTONA_API_KEY) using process.env and environmental variable discovery. This aligns with security best practices for avoiding hardcoded secrets.\n- [PROMPT_INJECTION]: The framework represents a platform that ingests external data, which is an inherent surface for indirect prompt injection.\n
Ingestion points: External data enters the context through <Agent> content and <Skill src="..." /> file references.\n
Boundary markers: The system uses <Sandbox> and <Workspace> components to provide execution isolation and narrow access leases.\n
Capability inventory: The framework supports shell execution, network access, and filesystem reads/writes through tools and MCP servers, all scoped to individual agents.\n
Sanitization: The SDK enforces structured validation using Zod schemas for tool inputs and outputs, providing a layer of protection against malformed data and command injection.

aml-jsx