@adhd/agent-engine-orchestrator

Orchestration engine for @adhd/agent-mcp. Implements the core agent execution loop, provider adapters, message windowing, tool calling, and delegation supervision.

Status: Core engine (shipped v0.0.1)
Package: npm install @adhd/agent-engine-orchestrator
Consumers: @adhd/agent-mcp (primary)

What it does

• Agent orchestration loop — coordinates model calls, tool execution, message history, and task completion
• Provider adapters — normalizes Anthropic, OpenAI, DeepSeek, and Gemini APIs to a unified interface
• Context management — implements append-only message history with intelligent collapse (2.0.2+)
• Token accounting — provider-neutral usage tracking with cache-aware fields
• Delegation supervision — enforces delegation depth, budget caps, and permission inheritance for child agents
• Hook system — pre/post interceptors for policy enforcement and observability

Usage

This is an internal package consumed by @adhd/agent-mcp. End users interact with it through the MCP server's tools (task, result, agent_create, etc.).

For engine integration (rare):

import { Orchestrator } from '@adhd/agent-engine-orchestrator';

const orchestrator = new Orchestrator({
  policy: policyEngine,
  db: agentDatabase,
  hooks: hookRegistry,
});

const result = await orchestrator.run({
  agentName: 'my-agent',
  messages: [{ role: 'user', content: 'Find bugs in /src' }],
  model: 'claude-opus-4-1',
  provider: 'anthropic',
});

Key files

• src/engine/orchestrator.ts — main orchestration loop
• src/providers/ — Anthropic, OpenAI, DeepSeek, Gemini adapters
• src/engine/windowing.ts — context management
• src/plugins/ — hook system and tool dispatch

Architecture

• Part of the 6-package agent framework family
• Depends on: @adhd/agent-base-types, @adhd/agent-store-runtime
• Depended on by: @adhd/agent-mcp

See /entrypoint/agent-mcp/docs/architecture-and-security.md for the full agent runtime architecture.