方案已弃用
本文记录 NATS 作为 agent 间通信方案的架构决策历史（2026-07）。 该方案已弃用 —— 现行为：gbrain（
${LOCALIP}:7842）共享知识 + Discord 唤醒人类侧 agent。 保留此文档是为了回答「为什么当初选 NATS、后来为何弃用」，不是当前运维依据。
Agent Communication Architecture
Principle
Agent-to-agent communication runs on NATS (LAN-local, no internet). Discord webhook is ONLY for waking the human-facing agent (me).
Communication Layers
Layer 1: NATS Primary (agent-to-agent)
- Transport: nats://${LANIP}:4222 (LAN, zero internet dependency)
- Subjects: agent.dm.
/ agent.broadcast / agent.rpc. - Durability: JetStream stream agent_messages (50K max)
- Reliability: No proxy, no internet, no Discord needed
- Latency: <1ms (LAN)
Layer 2: Discord Webhook (wake human agent)
- Purpose: Notify me (Alaya) when there’s a NATS message needing attention
- Transport: HTTP POST to discord.com (via proxy or direct)
- Fallback if webhook fails: Message stays in NATS JetStream
- Recovery: I check for missed messages via:
- nats_check.py (before each response)
- watchdog cron (every 5 min)
Layer 3: JetStream Durability (backstop)
- All DM/broadcast messages persist in agent_messages stream
- Even if all listeners fail, messages survive until consumed
- Each agent has dedicated consumers (alaya-dm, alaya-bc, etc.)
Why NOT to depend on Discord webhook for agent-to-agent
| Scenario | Impact on NATS | Impact on Webhook |
|---|---|---|
| Proxy down | None (no proxy needed) | Webhook still works (direct 570ms) |
| Internet down | None (LAN) | Webhook fails |
| Discord down | None | Webhook fails |
| WSL network down | Both fail | Both fail |
Architecture Decision
- NATS is the SOURCE OF TRUTH for agent messages
- Discord webhook is a NOTIFICATION overlay (convenience, not reliability)
- If webhook fails, messages are still processed on next response check
- NO agent depends on another agent’s webhook for communication
Implementation (v6)
- v6 listener uses Discord webhook for wake notifications
- NATS auto-reply for ping/status (zero internet dependency)
- All agent-to-agent traffic stays on NATS