The agent engine

@aparte/core already runs a full agent loop: when you drive a chat with AparteClient, it streams the model, splits the reply into segments (text, thinking, tool calls, artifacts), runs the tool-calling loop, and reports usage — all inline. Core works without this package.

@aparte/engine is that loop as a headless, framework-agnostic package: zero runtime dependencies, no DOM, runs in the browser or Node. Its headline export, runStreamAgent, is the exact loop core embeds inline — extracted so a backend can run it server-side with provably identical behaviour.

It is deliberately just the loop core drives, plus the agnostic conversation compactor. Opt-in tools (ask-question, RAG, skills, code execution) belong in plugins/*; product behaviour (memory, intent orchestration) and the not-yet-wired text agent loop live elsewhere. None of that ships here.

Install

npm install @aparte/engine

@aparte/core is an optional peer: runStreamAgent and the parsers need nothing from it, so you can install @aparte/engine alone. If you wire it into core’s client (below) you already have @aparte/core; otherwise npm install @aparte/core @aparte/engine. ESM-only (like the rest of @aparte/*); CJS consumers use await import().

What’s in it

Area	Exports	Status
Structured-stream loop	runStreamAgent, StreamRunEvent, the artifact-XML parser	Ready — the seam below
Context compaction	compactConversation + token-budget / sliding-window helpers	Ready

Everything is a plain function or class — no globals, no side effects (sideEffects: false), fully tree-shakeable, so you pull in only what you use.

The primary use: the streamRunner seam

Core stays the zero-dependency leaf: it never imports @aparte/engine. Instead, AparteClient exposes an injection point, streamRunner. Give it runStreamAgent and the client delegates its loop to the engine, rendering the engine’s events through core’s createStreamAdapter:

import { AparteClient } from '@aparte/core';
import { runStreamAgent } from '@aparte/engine';

const client = new AparteClient({
  // …your transport / config…
  streamRunner: runStreamAgent,   // delegate the loop to the engine
});

With no streamRunner, the inline loop runs (the default). With one, the engine runs the loop and core renders it — same messages, same events, same DOM output.

Proven parity

The two paths aren’t “meant” to match — it’s tested. The engine’s stream-parity suite drives core’s real inline loop and runStreamAgent (through the real createStreamAdapter) against the same scripted transport and asserts an identical call sequence and usage across nine scenarios (plain text, thinking, human-in-the-loop tool approve/reject, streamed and one-shot artifacts, multi-phase pipelines, forced tool calls). So the seam is a drop-in, not an approximation.