OpenSIN Neural Bus

Quick Start · Features · Architecture · Usage · Deploy · Contributing

The event-driven nervous system connecting all OpenSIN agents — durable, replayable, and Ouroboros-aware.

---

Quick Start

1. Start NATS docker compose up -d nats

2. Install npm install

3. Test npm test

Tip

Full setup: npm install && docker compose up -d nats && npm test — all tests cover pub/sub, request/reply, durable resume, and replay.

---

Features

Capability	Description	Status
JetStream Integration	Stable connect/reconnect wrapper with validated event envelopes	✅
Durable Consumers	Resume from last acked message after restart — no context loss	✅
Request/Reply	Synchronous request-response pattern over NATS subjects	✅
Event Envelopes	Standardized envelope with correlation IDs and source tracking	✅
Subject Taxonomy	Documented subject hierarchy for all event types	✅
Ouroboros Bridge	Auto-bridge to memory (rememberLesson) and capability (registerCapability)	✅
Python SDK	SQLite-backed memory with apply_event_envelope()	✅
Agent Runtime	Reusable publish/consume patterns for A2A agents	✅

Core Exports (TypeScript)

import {
  OpenCodeJetStreamClient,   // Stable NATS connection wrapper
  OpenSinAgentRuntime,       // Agent runtime with pub/sub patterns
  SUBJECTS,                  // Subject taxonomy constants
  createEventEnvelope,       // Envelope factory with validation
} from "@opensin/neural-bus";

(back to top)

---

Architecture

flowchart TB
    subgraph Agents["Agent Runtimes"]
        HERMES[A2A-SIN-Hermes]
        ORCH[A2A-SIN-Orchestrator]
        CLI[OpenCode CLI]
    end

    subgraph Bus["Neural Bus (NATS JetStream)"]
        PUB[Publish]
        SUB[Subscribe]
        REQ[Request/Reply]
        DUR[Durable Consumer]
    end

    subgraph Memory["Ouroboros Memory"]
        LESSON[rememberLesson]
        CAPA[registerCapability]
        SQLITE[(SQLite Memory)]
    end

    HERMES --> PUB
    ORCH --> REQ
    CLI --> PUB
    PUB --> SUB
    REQ --> SUB
    SUB --> DUR
    DUR --> LESSON
    DUR --> CAPA
    LESSON --> SQLITE
    CAPA --> SQLITE

    classDef agentClass fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    classDef busClass fill:#fff3e0,stroke:#e65100,stroke-width:2px
    classDef memClass fill:#fce4ec,stroke:#880e4f,stroke-width:2px

    class HERMES,ORCH,CLI agentClass
    class PUB,SUB,REQ,DUR busClass
    class LESSON,CAPA,SQLITE memClass

Loading

For detailed subject taxonomy see docs/jetstream-subject-taxonomy.md.

(back to top)

---

Usage

Agent Runtime — Publish Events

const runtime = new OpenSinAgentRuntime({
  agentId: "a2a-sin-hermes",
  sessionId: "session-001",
  bus,
});

await runtime.publishObservation({
  message: "worker booted",
  branch: "feat/new-feature",
});

await runtime.publishLessonLearned({
  context: "JetStream reconnect handling",
  lesson: "Reuse durable consumer name for automatic restart recovery.",
  successRate: 1.0,
});

Durable Consumer Pattern

const worker = await runtime.consumeAssignedWork(
  {
    subject: SUBJECTS.workflowRequest,
    stream: "OPENSIN_WORKFLOW_EVENTS",
    durableName: "my-worker",
    deliverPolicy: "all",
    ackWaitMs: 500,
  },
  async (event) => {
    console.log("received work", event.payload);
  },
);

Important

Reusing the same durableName after restart resumes from the last acked message — no context resend needed!

Request / Reply

// Server side
const server = await bus.serveRequests(SUBJECTS.workflowRequest, async (request) => {
  return createEventEnvelope({
    kind: "workflow.reply",
    subject: SUBJECTS.workflowReply,
    source: { id: "a2a-sin-orchestrator", runtime: "agent-runtime" },
    correlationId: request.id,
    payload: { accepted: true },
  });
});

// Client side
const reply = await bus.request(
  createEventEnvelope({
    kind: "workflow.request",
    subject: SUBJECTS.workflowRequest,
    source: { id: "opencode-cli", runtime: "opencode-cli" },
    payload: { task: "resume durable work" },
  }),
);

Ouroboros Bridge

The bus automatically invokes bridge methods when events include ouroboros.rememberLesson or ouroboros.registerCapability:

Bridge Method	Purpose	Storage
rememberLesson(record)	Store learned lessons for future agents	SQLite
registerCapability(record)	Register new agent capabilities	SQLite

The Python SDK exposes apply_event_envelope() for mirroring JetStream envelopes into SQLite-backed memory.

---

Deploy

Methode	Target	Zweck
Local	docker compose up -d nats	Development with embedded NATS
OCI VM	92.5.60.87:4222	Production NATS JetStream server
Package	@opensin/neural-bus (npm)	Shared library for all agents

Warning

The production NATS server runs on the OCI VM. All agents must connect to nats://92.5.60.87:4222 in production.

---

Documentation

Document	Purpose
Subject Taxonomy	Complete NATS subject hierarchy
ARCHITECTURE.md	System architecture deep dive
CONTRIBUTING.md	How to contribute

---

Changelog

v1.0.0 (2026-04-14)

• JetStream integration surface with stable connect/reconnect
• Validated event envelopes with correlation IDs
• Durable consumer pattern for restart recovery
• Request/reply helpers over NATS subjects
• Ouroboros bridge points (rememberLesson, registerCapability)
• Python SDK with SQLite-backed memory
• Documented subject taxonomy

---

Contributing

1. Fork the repository
2. Create your feature branch (git checkout -b feature/amazing-feature)
3. Start NATS: docker compose up -d nats
4. Install: npm install
5. Run tests: npm test
6. Commit and push
7. Open a Pull Request

---

License

MIT. See LICENSE for details.

---

_{Built by OpenSIN-AI Fleet}

(back to top)

---

Agent Configuration System (v5)

This project is part of the OpenSIN-AI agent ecosystem and uses the unified agent configuration system:

Datei	Zweck
oh-my-sin.json	Zentrales Team Register
oh-my-openagent.json	Subagenten-Modelle
my-sin-team-infra.json	Team Infrastructure Modelle

Subagenten-Modelle

Subagent	Modell
explore	nvidia-nim/stepfun-ai/step-3.5-flash
librarian	nvidia-nim/stepfun-ai/step-3.5-flash

PARALLEL-EXPLORATION MANDATE

Bei Codebase-Analyse auf grossen Projekten MUESSEN Agenten 5-10 parallele explore + 5-10 librarian-Agenten starten.

→ Full Documentation