XSky Core Architecture

This skill provides deep knowledge of XSky AI Agent framework internals.

Overview

XSky is a multi-platform AI agent framework with these core components:

XSky (Orchestrator)
├── Planner (Task → Workflow XML)
├── Chain (Execution tracking)
├── Context (State management)
├── Memory (Conversation compression)
└── Agents (Task executors)

Execution Flow

1. Task Received: User provides natural language task
2. Planning: Planner converts to Workflow XML using LLM
3. Parsing: XML parsed into agent tree with dependencies
4. Execution: Agents execute in dependency order
5. Result: Final result returned to user

Key Classes

XSky (core/xsky.ts)

Main orchestrator with these methods:

• generate(taskPrompt) → Creates workflow from task
• execute(taskId) → Runs workflow
• run(taskPrompt) → Generate + execute in one call
• pauseTask(taskId) → Pause execution
• abortTask(taskId) → Cancel execution

Planner (core/plan.ts)

Converts natural language to Workflow XML:

• Uses LLM to analyze task and select agents
• Outputs structured XML with dependencies
• Supports replanning on failure

Chain (core/chain.ts)

Tracks execution state:

• Records each agent's execution
• Stores results and errors
• Enables debugging and replay

Context (core/context.ts)

Manages task state:

• taskId - Unique task identifier
• variables - Key-value storage
• workflow - Current workflow
• controller - AbortController for cancellation
• conversation - Chat messages during execution

AgentContext

Per-agent execution context:

• Access to variables
• Tool execution
• Script injection
• Result storage

Agent Architecture

All agents extend base classes:

• Agent - Base class with LLM dialogue loop
• BaseBrowserAgent - Browser abstraction
• BaseBrowserLabelsAgent - Element labeling approach
• BaseFileAgent - File system operations
• BaseShellAgent - Command execution

Key Source Files