uncfreak1255-code OpenClaw Development Skill
System Architecture
OpenClaw is a multi-agent Discord bot at ~/.openclaw/ running on 64GB Apple Silicon Mac.
6 Agents
| Agent | Role | Key Integrations |
|---|---|---|
| Doni | Coordinator — routes messages to specialists | Message bus, inter-agent |
| Cove | Guest operations — handles guest inquiries | Hostaway API, RAG |
| Vector | SEO — content and optimization | Search APIs, content gen |
| Dev | Engineering — code tasks | File system, git |
| Spark | Memory/RAG — knowledge retrieval | LanceDB, embeddings |
| Ops | Monitoring — system health | Logs, metrics |
Runtime Stack
- LLM: LM Studio at localhost:1234 (qwen2.5-7b-instruct, ~33GB RAM)
- RAG: LanceDB at ~/.openclaw/data/lancedb/ + nomic-embed-text-v1.5 (768 dims)
- Runtime: Node.js, ALL CommonJS (require/module.exports — NO ESM)
- 327 memories seeded across 5 LanceDB tables
5-Tier Context System
| Tier | Context | Cost | Use Case |
|---|---|---|---|
| 1 | 8K | FREE (local) | Simple routing, status |
| 2 | 16K | FREE (local) | Standard queries |
| 3 | 32K | ~$0.50/M | Complex queries |
| 4 | 128K | ~$3/M | Multi-doc RAG |
| 5 | 200K | ~$15-75/M (Opus) | Full analysis |
Shared Modules (at ~/.openclaw/shared/)
- vectorstore.js — LanceDB table management
- embedding-client.js — nomic-embed-text-v1.5 interface
- rag-injection.js — Query → retrieve → inject pipeline
- context-tier.js — Automatic tier selection
- message-bus.js — Inter-agent message routing
- inter-agent.js — Agent-to-agent communication protocol
- smart-compaction.js — Two-pass Gemini Flash-Lite compaction
- llm-router-patch.js — LLM provider routing
Code Standards
- CommonJS ONLY:
const x = require('./x')/module.exports = { x } - Error handling: Wrap ALL LLM calls in try/catch with fallback
- Logging:
console.log('[MODULE_NAME]', message) - Testing:
node -e "require('./path')"to verify modules load - Two-layer structure: agents//agent/ (runtime) + workspaces// (full workspace)
Current Phase
Phase 4 complete (327 memories seeded, shared modules deployed). Next: wire Phase 4 modules into llm-router.js and test end-to-end Discord message flow.