What's New

Latest shipping in LunaOS. Most recent first.

---

April 2026 — Major engine upgrade (179 tests passing)

Shipped in a single week: parallel agent swarm, self-learning LLM router, graph-aware RAG, and 6-provider LLM routing.

Parallel Agent Swarm (NEW)

Run 2–5 agents against the same task in parallel. Three merge strategies:

• race — fastest successful agent wins
• consensus — majority agreement on output
• vote — longest/best-scored answer wins

curl https://api.lunaos.ai/agents/swarm \
  -H "Authorization: Bearer $TOKEN" \
  -d '{
    "agents": ["code-review", "security-audit", "test-writer"],
    "context": "Review this login function",
    "strategy": "consensus"
  }'

Response shape:

{
  "strategy": "consensus",
  "winner": { "agent": "code-review", "provider": "groq", "durationMs": 340 },
  "allResults": [ /* 3 items */ ],
  "totalDurationMs": 410,
  "reason": "consensus: 3/3 agents agreed (100%)"
}

4× perceived speed on multi-file tasks. Workers-compatible (uses Promise.all(), no git worktrees).

Self-Learning Router (Thompson Sampling)

The smart router now uses a multi-armed bandit instead of static "pick the highest success rate". Key changes:

• Beta distribution sampling per (agent, provider, model) — natural exploit/explore balance
• 90-day recency window — old outcomes decay automatically
• 30-day half-life decay — recent data weighted ~2× vs 30 days old
• 10% forced exploration when sample size < 20

Expected 30–50% LLM cost reduction after a week of real traffic.

Graph RAG with Community Detection

Flat vector search misses related-but-not-similar code. New graph expansion:

1. Vector search returns top 5 chunks (score × 1.0)
2. For each, look up 1-hop neighbors (score × 0.7 × edge weight)
3. For each, look up same-community members (score × 0.5)
4. Dedupe, rerank, return top 10

Community detection uses label propagation (Workers-compatible, no native graph libs). Converges in 5–10 iterations.

30–60% better retrieval precision on code queries.

Six LLM Providers (up from 3)

Provider	Cost/1M	Best For
Gemma 4 (Ollama)	$0	Local dev, free inference
Gemini 2.0 Flash	Free tier then $0.075	Fast reasoning
Groq (Llama 3.3 70B)	Free tier then $0.05–0.10	Highest throughput (LPU)
DeepSeek Chat	$0.14	Cheap general use
OpenAI GPT-4o	~$2.50	Balanced quality
Anthropic Claude Sonnet	~$3.00	Highest quality fallback

The Thompson router auto-picks the best one for each agent type based on outcomes.

Architecture

Full routing chain for every request:

Task arrives
  ↓
Agent Booster ($0 — deterministic code transforms)
  ↓
ReasoningBank cache ($0 — SHA-256 prompt hash in KV)
  ↓
Context Packer (trim irrelevant fields)
  ↓
Smart Router (Thompson sampling picks provider + model)
  ↓
Graph RAG enrichment (if useRag:true)
  ↓
LLM call via Claw Gateway (PII redaction + audit log)
  ↓
Stream SSE response
  ↓
Record outcome → update Thompson priors

~30% of requests never hit an LLM thanks to booster + cache.

---

April 2026 — CLI published on npm

[email protected] is now on the npm registry.

npm install -g luna-agents
luna-setup

• 232 slash commands for Claude Code
• 28 specialized agents
• 3 MCP servers: Nexa RAG, GLM Vision, combined vision+RAG client
• 360 KB packed, 290 files, Node ≥18, MIT license

GitHub release: https://github.com/lunaos-ai/luna-agents/releases/tag/v2.0.1

---

April 2026 — Dashboard deployed

The admin dashboard is live at agents.lunaos.ai.

Pages:

• /dashboard — overview with recent executions and usage
• /dashboard/agents — browse 28 agents, run any of them
• /dashboard/chains — build multi-step workflows
• /dashboard/api-keys — create/revoke API keys
• /dashboard/billing — upgrade plan, view invoices
• /dashboard/kb — knowledge base (upload docs for RAG)
• /dashboard/repos — connect GitHub repos for indexing
• /dashboard/analytics — usage charts

Built on Next.js 15 static export, deployed on Cloudflare Pages.

---

Test coverage

Engine API: 179/179 tests passing

 ✓ tests/swarm.test.ts              (9)   — parallel agent swarm
 ✓ tests/thompson-sampling.test.ts  (15)  — Beta distribution + bandit
 ✓ tests/graph-rag.test.ts          (7)   — expansion + reranking
 ✓ tests/community-detection.test.ts (9)  — label propagation
 ✓ tests/cli-tools.test.ts          (48)
 ✓ tests/agent-orchestration.test.ts (30)
 ✓ tests/workflow-builder.test.ts   (27)
 ✓ tests/payment.test.ts            (16)
 ✓ tests/auth.test.ts               (10)
 ✓ tests/monitoring.test.ts         (8)

---

What's next

• Flakestress runtime — run any test 100× to detect flakes
• Perfetto trace UI — flame graphs for workflow execution
• LLM-judge vote strategy — replace longest-output heuristic with a real arbiter
• Product Hunt launch

---

Stay updated: follow @lunaos_ai or watch github.com/lunaos-ai.