@elyxndra/chat

The chat brain shared by Elyxndra's apps: threads, settings, the built-in system prompt, local chat tools and the send → stream → tools → continue loop — on top of @elyxndra/engine (which model is live, its wire spec and health) and @elyxndra/agent (streaming client, context budget). Host-agnostic, Node 20+.

import { createEngineRuntime } from "@elyxndra/engine";
import { ChatApp } from "@elyxndra/chat";

const rt = await createEngineRuntime();
const chat = new ChatApp({ runtime: rt });
await chat.init();
chat.on("delta", (id, content, reasoning) => process.stdout.write(content));
chat.on("change", () => {
  /* render chat.messages, chat.conversations, chat.isGenerating … */
});
chat.send("Hello");

• Threads — ConversationStore: one pretty-printed JSON file per thread under <app data>/conversations/, a stable format every Elyxndra host reads and writes, so chats carry over between them. Newest-created first, atomic writes.
• Settings — ChatSettings: profile (name, role, style…), system prompt, thinking/effort defaults, tool switches, web-research config; provider keys in the OS credential store.
• Prompt — builtinSystemPrompt(): identity + live model/engine/machine context, regenerated per request.
• Tools — get_system_info, get_model_search_results, plus the agent package's web_research / web_fetch (opt-in).
• Loop — ChatApp.send()/regenerateLastReply()/stopGeneration(), tool rounds run concurrently and keep one uninterrupted "generating" state, context auto-trim with engine-calibrated estimates, per-message stats.

Tests: npm test (offline; a fake SSE engine stands in for the model server).