@agforge/core

Agent protocol, execution skeleton, interceptor chain, message types, and ScopeContext — the foundational layer of AgForge SDK.

Overview

@agforge/core defines only interfaces and execution skeletons, with no concrete implementations. It provides:

• Agent — abstract agent protocol
• ReActAgent — zero-built-in-interceptor agent with the ReAct loop skeleton
• Interceptor chain — onion-model composition across multiple lifecycle hooks
• Message types — full message type definitions (user, assistant, tool call, tool result, etc.)
• Model interface — LLM abstraction for streaming request() returning AsyncGenerator<AssistantMessageChunk>
• Tool interface — dual-channel output (structResult + llmMessage)
• ScopeContext — WeakMap-based data isolation (private + shared)
• MessageStore — conversation history persistence interface

Install

npm install @agforge/core
# or
pnpm add @agforge/core

Usage

import { ReActAgent } from '@agforge/core';
import type { Model, Tool, Interceptor } from '@agforge/core';

For a fully-featured agent with 14 built-in interceptors, use @agforge/engine which provides DefaultAgent.

Lifecycle Hooks

Agent.execute()
├── execute           — Agent-level entry & exit
│   └── doExecute     — ReAct loop entry
│       └── eachLoop  — Single iteration
│           ├── llmRequest    — LLM call
│           └── executeTool   — Tool execution entry
│               ├── resolveTool    — Match tool, parse params
│               └── doExecuteTool  — Core tool execution
└── stop              — Graceful shutdown

Interceptor Phases

Interceptors are organized into phases to guarantee execution order:

Related Packages

License

ISC