Defines a base Tool class and @tool decorator that automatically generates input schemas from function signatures and enforces a consistent ToolResult return type. This pattern allows developers to create new tools by simply defining typed functions, promoting consistency and reducing boilerplate across the tool ecosystem (file, shell, search, image generation, etc.).

Abstract base agent class encapsulating cross-cutting concerns: LLM invocation with exponential backoff retry, tool call execution, memory management (including context window truncation and summarization), and conversation history handling. Specific agents (Planner, ReAct) inherit and specialize the invoke method, demonstrating clean separation of concerns and enabling easy extension of agent types.

Implementation of a message queue using Redis Streams for publishing and consuming domain events (e.g., agent steps, tool results). Provides a reliable, scalable backbone for real-time communication between backend services and the frontend via SSE. The interface-based design allows swapping with other message brokers (e.g., RabbitMQ, Kafka) without affecting dependent code.

Custom React hook that establishes a Server-Sent Events (SSE) connection to the backend for real-time session updates (agent thoughts, tool executions, messages). Handles connection lifecycle, retry logic, and event parsing, providing a reusable pattern for any SSE-dependent UI component. Demonstrates effective integration of backend event streaming with frontend state management.

Defines the IUnitOfWork abstract protocol for managing database transactions across multiple repositories. The implementation (db_uow.py) uses SQLAlchemy async sessions to ensure atomic commit/rollback operations. This pattern isolates transaction logic from business logic, enabling clean separation of concerns and facilitating testing with mock UoW implementations.