Architecture Reference — v1.4

Seven Agents. Three Machines. No Cloud.

A multi-agent AI development team distributed across three machines in the homelab. Every model in the primary lineup costs $0.00. Architecture, implementation, review — all automated.

7
AI Agents
3
Machines
11+
Models Available
$0
Per Workflow
1M
Max Context

The Team

Click any card to expand model details, fallbacks, and capabilities.

🔬
@system-analyst
System Analyst & Requirements Engineer
Tencent Hy3 295B
The first agent to run on any complex task. Takes a user's request and produces a structured requirements analysis — core goals, constraints, edge cases, risks, and clarifying questions. Feeds context to the architect.
FREE 300s → Nemotron 120B

Model Configuration

Modeltencent/hy3 (295B MoE, 21B active)
HostOpenRouter (free tier)
Context256K tokens
Max Output8,192 tokens
Primary FallbackNemotron 3 Super 120B
Local FallbackLlama 3.2 (Ollama)
click to expand ▾
🏗️
@lead-architect
Lead Software Architect
Big Pickle / GLM-4-6
The heavyweight designer. 355B MoE model with 200K context. Thinks in modules, interfaces, and data flow. Produces structured blueprints with operational concerns — deployment, monitoring, failure modes. Uses deep multi-step reasoning before output.
FREE 1200s → Qwen3.6-27B (melchior)

Model Configuration

Underlying ModelGLM-4-6 (Zhipu AI)
Architecture355B MoE (~35B active)
HostOpenCode Zen API
Context200K tokens
Max Output32,000 tokens
Timeout1200s (20 min)
Primary FallbackQwen3.6-27B (melchior)
Local FallbackQwen 2.5 Coder (Ollama)
Capabilitiesreasoning, tool_calling, structured_output
click to expand ▾
⚡
@heavy-implementer
Heavy-Duty Implementer
Qwen3.6-27B
Fast, reliable code generation running on melchior over LAN. Takes architecture blueprints and turns them into production-quality code with proper error handling, type hints, and edge case coverage. Always-on, zero latency.
FREE — LAN 300s → Big Pickle (Zen)

Model Configuration

ModelQwen3.6-27B (reasoning)
Hostmelchior (192.168.6.122:8000)
Inference EnginevLLM
Context128K tokens
Max Output16,384 tokens
Primary FallbackBig Pickle (Zen)
Local FallbackQwen 2.5 Coder (Ollama)
click to expand ▾
🔍
@qa-reviewer
QA & Code Review Specialist
DeepSeek Flash Free
Meticulous code reviewer. Finds bugs, security vulnerabilities, performance issues, and missed edge cases. Writes unit tests and suggests improvements. Thorough, skeptical, and detail-oriented.
FREE 300s → Qwen3.6-27B (melchior)

Model Configuration

ModelDeepSeek V4 Flash (Free)
HostOpenCode Zen API
Context128K tokens
Max Output8,192 tokens
Primary FallbackQwen3.6-27B (melchior)
Local FallbackLlama 3.2 (Ollama)
click to expand ▾
📡
@triage-agent
Triage & Routing Agent
Qwen3.6-27B
The traffic cop. First to touch every request — classifies task type and complexity, determines which agents are needed, and routes the work. Fast, efficient, keeps the team organized.
FREE — LAN 300s → DeepSeek Flash Free

Model Configuration

ModelQwen3.6-27B (reasoning)
Hostmelchior (192.168.6.122:8000)
Context128K tokens
Primary FallbackDeepSeek V4 Flash Free
Local FallbackLlama 3.2 (Ollama)
click to expand ▾
🧬
@code-specialist
Code Specialist — Large Scale
Nemotron 3 Ultra 550B
For massive codebases and multi-file projects. 550B MoE model with a 1 million token context window. Reads entire repositories, understands the full picture, then makes changes. Deployed when triage detects "large-scale" complexity.
FREE 300s → Tencent Hy3 295B

Model Configuration

ModelNVIDIA Nemotron 3 Ultra 550B (55B active)
HostOpenRouter (free tier)
Context1,000,000 tokens
Max Output16,384 tokens
Primary FallbackTencent Hy3 295B
Local FallbackQwen 2.5 Coder (Ollama)
click to expand ▾
📖
@documenter
Documentation & Knowledge Specialist
Nemotron 3 Super 120B
Produces thorough, well-structured documentation. API docs, READMEs, user guides, architecture notes, and knowledge bases. Can explain complex code in simple terms and create text-based architecture diagrams.
FREE 300s → Nemotron 3 Ultra 550B

Model Configuration

ModelNVIDIA Nemotron 3 Super 120B (12B active)
HostOpenRouter (free tier)
Context1,000,000 tokens
Max Output16,384 tokens
Primary FallbackNemotron 3 Ultra 550B
Local FallbackLlama 3.2 (Ollama)
click to expand ▾

The Workflow

Every task flows through the pipeline. Triage determines which phases are needed — not every task uses all agents.

Phase 1 — Triage

Classify & Route

📡 @triage-agent · Qwen3.6-27B

Receives the raw task. Returns a JSON classification: task_type, complexity, agents_needed, and reasoning. Determines which subsequent phases are required.

Phase 1.5 — System Analysis

Requirements & Constraints

🔬 @system-analyst · Tencent Hy3 295B

Structures the problem before any design begins. Identifies core requirements, constraints, edge cases, failure modes, dependencies, technical risks, and asks clarifying questions. Output feeds directly into the architect's context.

Phase 2 — Architecture

Design the System

🏗️ @lead-architect · Big Pickle / GLM-4-6

Produces a structured blueprint: component/module breakdown, data flow, technology recommendations, file structure, key interfaces, and operational concerns. Receives system analysis + original task as context. 1200s timeout allows deep reasoning.

Phase 3 — Implementation

Write the Code

⚡ @heavy-implementer · Qwen3.6-27B (melchior)

Writes complete, production-quality code following the architecture blueprint. Includes error handling, type hints, documentation, and edge case coverage. For large-scale tasks, @code-specialist (Nemotron 3 Ultra 550B, 1M context) handles the work instead.

Phase 4 — QA Review

Find the Problems

🔍 @qa-reviewer · DeepSeek Flash Free

Reviews implementation output for bugs, security vulnerabilities, performance issues, and missed edge cases. Writes unit tests and suggests improvements. Returns pass/fail/needs-work assessment.

Phase 4.5 — Documentation

Write the Docs

📖 @documenter · Nemotron 3 Super 120B

Produces API documentation, READMEs, usage instructions, architecture notes, and known limitations. Only runs when triage determines documentation is needed.

Phase 5 — Synthesize

Assemble the Report

The orchestrator combines all phase outputs (analysis, architecture, implementation, review, documentation) into a single structured report. The session is saved to sessions/ as JSON with a UUID prefix.

# Full workflow — triage routes everything automatically
devgroup "Build a REST API for managing fantasy football drafts"

# Single agent — bypass the workflow
devgroup --agent lead-architect "Design a WebSocket chat system"

# Interactive console — chat with any agent
devgroup --interactive

Infrastructure

Three machines, one hosted API layer. Each machine has a distinct role in the pipeline.

Orchestrator

brainbox

192.168.6.101
Python codebase, session management, CLI entry point, interactive console
dev_group.py — API + CLI
orchestrator.py — workflow engine
agents/client.py — LLM HTTP client
agents/models.py — model registry
sessions/ — saved session JSON
Inference

melchior

192.168.6.122
vLLM serving — Qwen3.6-27B for triage, implementation, and architect fallback
Qwen3.6-27B via vLLM
128K context · 16K output
No auth · LAN only
Reasoning model (thinking included)
GPU Compute

achiral

192.168.4.17
Image generation, TTS, ASR — separate from dev team but shares the homelab
ComfyUI (FLUX image gen)
Gemma 4 (iGPU)
Kokopop TTS
Parakeet ASR
Hosted APIs

OpenCode Zen + OpenRouter

opencode.ai · openrouter.ai
Free-tier hosted models — Big Pickle, DeepSeek, Tencent Hy3, Nemotron, gpt-oss
Big Pickle (200K · free)
DeepSeek V4 Flash Free
Nemotron 3 Ultra 550B (1M · free)
Tencent Hy3 295B (256K · free)
Nemotron 3 Super 120B (1M · free)
gpt-oss-20B (131K · free)
Gemma 4 26B (256K · free)

Complete Model Registry

Every model available to the dev group, including fallbacks and future paid options.

Model Host Context Output Timeout Cost Status
Big Pickle (GLM-4-6)
355B MoE · @lead-architect
 Zen API 200K 32K 1200s FREE ✅ Active
Qwen3.6-27B
@heavy-implementer · @triage-agent
 melchior (vLLM) 128K 16K 300s FREE ✅ Active
DeepSeek V4 Flash Free
@qa-reviewer
 Zen API 128K 8K 300s FREE ✅ Active
Tencent Hy3 295B
@system-analyst · 21B active
 OpenRouter 256K 8K 300s FREE ✅ Active
Nemotron 3 Ultra 550B
@code-specialist · 55B active
 OpenRouter 1M 16K 300s FREE ✅ Active
Nemotron 3 Super 120B
@documenter · 12B active
 OpenRouter 1M 16K 300s FREE ✅ Active
gpt-oss-20B
Extra · OpenAI, 3.6B active
 OpenRouter 131K 8K 300s FREE 🔄 Available
Gemma 4 26B
Extra · MoE, 4B active
 OpenRouter 256K 8K 300s FREE 🔄 Available
Qwen 3.5 Plus
Zen fallback
 Zen API 128K 8K 300s $0.20/M 🔄 Fallback
Qwen 3.6 Plus
Zen paid · stronger Qwen
 Zen API 128K 8K 300s $0.50/M 🔄 Fallback
DeepSeek V4 Pro
Future @qa-reviewer (paid)
 Zen API 128K 8K 300s $1.74/M Placeholder
Claude Sonnet 4.6
Future @lead-architect (paid)
 OpenRouter (planned) — — — ~$3/M Placeholder
Llama 3.2
Local fallback (Ollama)
 Localhost 8K 4K 300s FREE 🔄 Fallback
Qwen 2.5 Coder 14B
Local fallback (Ollama)
 Localhost 32K 8K 300s FREE 🔄 Fallback

CLI & API Reference

Shell commands and Python API for interacting with the dev group.

devgroup "..." Full workflow — triage routes automatically
devgroup -a lead-architect "..." Ask a specific agent directly
devgroup --interactive Chat-style REPL with all agents
devgroup --triage "..." See triage classification only
devgroup --list Show team roster
devgroup --sessions List past sessions
import sys
sys.path.insert(0, '/usr/local/share/projects/opencode')
from dev_group import DevGroup

dg = DevGroup()

# Full workflow
result = dg.ask("Build a database migration tool")

# Single agent
code = dg.implement("Write an async HTTP client")
design = dg.design("Design a caching layer")
review = dg.review(code)

# Direct question (no workflow)
answer = dg.ask("Explain this error", full_workflow=False)
# Start the interactive console
devgroup --interactive

# In the REPL:
# Ask a specific agent
@lead-architect Design a REST API for user management

# Ask everyone
/all What are your thoughts on microservices vs monoliths?

# Plain text → goes to @heavy-implementer
Write a Python script to parse JSON logs

# Commands
/list — show team members
/history — show last 10 messages
/quit — exit

Future Plans

Paid models that will be activated when K decides to invest. Models already defined in the codebase — just need a key.

Claude Sonnet 4.6 Planned

Premium upgrade for @lead-architect. Consistently rated #1-2 for system design and architecture. Would sit alongside Big Pickle as an optional "heavy architect" mode.

~$3.00 per million tokens (input) via OpenRouter

DeepSeek V4 Pro Planned

Deep upgrade for @qa-reviewer. Already defined in zen_models() as deepseek-v4-pro. Deeper code review capabilities than the free Flash tier. Just reassign the agent's model field.

$1.74/M input · $3.48/M output via Zen API
# To activate Claude Sonnet 4.6 as lead-architect:
# 1. Add to agents/models.py under openrouter_models():
# ModelConfig(name="claude-sonnet", ...)
# 2. Change DEV_GROUP['lead-architect']['model'] = 'claude-sonnet'
# 3. Add OPENROUTER_API_KEY to config/api.env

# To activate DeepSeek V4 Pro as qa-reviewer:
# 1. Model config already exists as 'deepseek-v4-pro'
# 2. Change DEV_GROUP['qa-reviewer']['model'] = 'deepseek-v4-pro'