AetherWeaver — Open-Source Learned Multi-Turn LLM Orchestration

简体中文

AetherWeaver is an orchestration platform that learns. Send it a message. It decides whether to answer directly or coordinate Thinker → Worker → Verifier across multiple models. It learns from every success and failure, improving weekly. Open-source, free to deploy on Vercel, uses your own API keys.

---

What It Does

Multi-Turn Orchestration

AetherWeaver doesn't pick one model and call it done. For complex tasks, it runs a coordinated multi-turn pipeline:

Request → Thinker (analyze & plan) → Worker (execute) → Verifier (check & accept/retry)

• Adaptive depth: simple queries = 1 turn. Complex code/math/reasoning = 2-4 turns.
• Budget protection: token, cost, and latency caps per request. Degrades gracefully on timeout — skips Verifier and returns Worker output rather than failing.
• Single endpoint: model: "auto" — the coordinator handles the rest.

Self-Evolution

AetherWeaver improves from usage. Every request generates a reward signal — code execution results, math answer matching, or explicit 👍/👎 feedback:

Feedback → Failure Diagnosis (4-dimension analysis) → Weekly Retraining (sep-CMA-ES, CPU-runnable) → updated coordinator weights (<100KB JSON)

The coordinator learns to route better over time. No labeled dataset needed — verifiable tasks provide ground truth automatically.

Intelligent Context Management

Multi-turn orchestration bloats context. AetherWeaver's context builder assembles role-optimized prompts instead of passing the full history to every model:

• Sensory filter: drops irrelevant turns
• Topic grouper: semantic clustering per conversation phase
• Per-role assembly: Thinker sees problem + plan. Worker sees plan + evidence. Verifier sees output + requirements.

4-turn context overhead = ~1.5-2x single call, not 4x.

Deploy Anywhere, Bring Your Own Keys

git clone https://github.com/inoribea/AetherWeaver.git && cd AetherWeaver
yarn install && cp .env.example .env.local
yarn dev   # localhost:3000
yarn deploy  # Vercel (free Hobby tier)

One command to production. Your OpenAI / Anthropic / Google / DeepSeek / Qwen / Hunyuan keys. Your infrastructure. Your data never touches a third-party orchestration service.

---

Key Advantages

Learned routing, not static rules. The coordinator trains via gradient-free evolution (sep-CMA-ES) on actual usage outcomes. It gets smarter without human tuning.

Multi-turn without multi-endpoint complexity. One model: "auto" call triggers Thinker → Worker → Verifier coordination. No workflow DSL, no agent framework to configure.

Full transparency. Routing decisions, model choices, confidence scores, and per-turn traces are all logged to Langfuse. Coordinator weights are a <100KB JSON file — inspectable, versioned, auditable.

Self-improving without data annotation. Code execution pass/fail and math answer matching provide clean reward signals automatically. User feedback is optional, additive.

Free and open. MIT license. Vercel Hobby tier. No seat licenses, no output-token pricing, no vendor lock-in.

---

Quick Start

git clone https://github.com/inoribea/AetherWeaver.git
cd AetherWeaver
yarn install
cp .env.example .env.local

Minimum config:

OPENAI_API_KEY=sk-...
# or GOOGLE_API_KEY=...
# or NEKO_API_KEY=... + NEKO_BASE_URL=...

yarn dev   # → http://localhost:3000

curl -X POST http://localhost:3000/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"Write a lock-free concurrent hashmap in Rust"}]}'

---

Configuration

Variable	Default	Description
ORCHESTRATION_MODE	adaptive	fast / standard / deep / adaptive
MAX_TURNS	4	Max orchestration turns per request
ORCHESTRATION_TIMEOUT_MS	55000	Timeout (5s Vercel buffer)
COORDINATOR_SIDECAR_URL	—	Local Qwen3-0.6B for hidden-state routing (100% Trinity fidelity)
ENABLE_FEEDBACK	true	Collect 👍/👎 feedback for learning

---

Observability

Every request produces a Langfuse trace tree:

Request
├── Coordinator Decision — model, confidence, reasoning
├── Thinker — model, tokens, latency
├── Worker
├── Verifier — verdict, rationale
└── Feedback — rating, auto-verification, failure diagnosis

---

Architecture

app/api/v1/chat/completions/route.ts   ← OpenAI-compatible entry point
utils/coordinator/                     ← embedder, classifier, bandit, sidecar
utils/orchestration/                   ← LangGraph state graph, roles, context builder
utils/feedback/                        ← feedback store, failure diagnosis
scripts/eval/                          ← eval harness (6 strategies + expert baseline)
scripts/train/                         ← sep-CMA-ES retraining + targeted optimization
docker/coordinator-sidecar/            ← optional ONNX Qwen3-0.6B sidecar

---

Documentation

Document	Contents
Trinity Orchestration Roadmap	Full implementation plan
Deployment Guide	Vercel deployment
API Usage	Chat API reference

---

Research Foundations

• Trinity (Sakana AI, ICLR 2026) — Multi-turn coordinator + sep-CMA-ES training. arXiv:2512.04695
• SkillForge (Alibaba, SIGIR 2026) — Self-evolving skills, failure diagnosis. arXiv:2604.08618
• GAM (BAAI) — JIT-compilation agent memory. arXiv:2511.18423
• LightMem (Zhejiang Univ) — Three-stage lightweight memory. arXiv:2510.18866

---

License

MIT © AetherWeaver