# Boring for Gemini - System Architecture (V14)
> **Version**: V14.0.0
> **Last Updated**: 2026-01-15
> **Status**: Production
---
## 🏗️ The 4-Layer Cognitive Architecture
Boring V14 introduces a **Cognitive Architecture** that moves beyond simple tool calling to autonomous reasoning and execution. The system is organized into four distinct layers:
```mermaid
graph TD
classDef flow fill:#e1f5fe,stroke:#01579b,stroke-width:2px;
classDef cognitive fill:#fff3e0,stroke:#ff6f00,stroke-width:2px;
classDef tools fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px;
classDef infra fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px;
subgraph L4_Flow ["Layer 4: Flow Engine (Orchestration)"]
OneDragon["One Dragon Flow"]:::flow
AutoOpt["Auto Optimization"]:::flow
VibeSession["Vibe Session"]:::flow
end
subgraph L3_Cognition ["Layer 3: Cognitive Engine (System 2)"]
SeqThink["Sequential Thinking"]:::cognitive
CritThink["Critical Thinking"]:::cognitive
BrainMgr["Brain Manager (Memory)"]:::cognitive
Router["Semantic Tool Router"]:::cognitive
end
subgraph L2_Tools ["Layer 2: Tool Layer (Capabilities)"]
RAG["RAG & Search"]:::tools
Git["Git & Versioning"]:::tools
Review["Code Review"]:::tools
Security["Security Scan"]:::tools
end
subgraph L1_Infra ["Layer 1: Infrastructure (Foundation)"]
MCP["MCP Server Protocol"]:::infra
Core["Core Config & Env"]:::infra
Services["Services (Monitor/Notify)"]:::infra
Storage["SQLite Storage"]:::infra
end
L4_Flow --> L3_Cognition
L3_Cognition --> L2_Tools
L2_Tools --> L1_Infra
```
---
## 🧠 Layer 3: Cognitive Engine (The Brain)
This is the key differentiator of V14. Instead of reactively calling tools, the agent **thinks** before acting.
### 1. Sequential Thinking (System 2)
The `sequentialthinking` module allows the agent to:
- Break complex problems into atomic steps.
- Revise its own assumptions ("Wait, that approach won't work...").
- Branch into alternative hypotheses.
- Maintain context over long executions.
### 2. Semantic Tool Router
The `tool_router.py` acts as a context-aware gateway:
- **Input**: Natural language intent ("Fix the login bug").
- **Processing**: Semantic analysis + Flow Stage filtering.
- **Output**: A curated list of 3-5 relevant tools (out of 60+).
- **Benefit**: Reduces context usage by 80% and increases accuracy.
---
## 🐉 Layer 4: One Dragon Flow (Autonomous)
The **One Dragon Flow** (`boring_flow`) allows the agent to autonomously traverse the entire software development lifecycle:
| Stage | Goal | Active Mechanisms |
|-------|------|-------------------|
| **1. Design** | Architect the solution | `boring_arch_check`, `sequantialthinking`, `boring_predict_impact` |
| **2. Implement** | Write the code | `boring_code_review`, `boring_test_gen`, `context7_query` |
| **3. Polish** | Refine & Optimize | `boring_perf_tips`, `boring_security_scan` |
| **4. Verify** | Ensure Quality | `boring_verify`, `boring_integrity_score` |
---
## 🗂️ Module Structure (V14 Source Map)
The codebase reflects this layered architecture:
```bash
src/boring/
├── flow/ # [L4] Flow Engine & Nodes
│ ├── engine.py # The orchestrator
│ └── nodes/ # Architect, Builder, Critic nodes
│
├── mcp/ # [L1/L2] MCP Server & Tools
│ ├── tool_router.py # The Semantic Gateway
│ ├── tool_profiles.py # Profile Management
│ └── tools/ # Individual Tool Implementations
│
├── intelligence/ # [L3] Cognitive Modules
│ ├── brain_manager.py # Long-term Memory
│ └── thinking/ # Thinking Strategies
│
├── services/ # [L1] Core Services
│ ├── monitor.py # Web Dashboard
│ └── notifier.py # Notification System
│
├── core/ # [L1] Fundamental Metadata
│ ├── config.py # Configuration
│ └── context.py # Context Management
│
└── main.py # CLI Entry Point
```
---
## 🛡️ Security & Integrity
### Shadow Mode
A "Sandboxed Execution" environment that intercepts all tool calls.
- **Strict Mode**: Requires manual approval for *any* file modification.
- **Standard Mode**: Auto-approves safe edits, blocks heuristic risks.
### Offline-First Architecture
V14 supports complete air-gapped operation:
- **Docs**: Local embeddings via `chromadb`.
- **LLM**: Local connection to `Ollama` or `LM Studio` via `boring-adapter`.
- **Privacy**: Zero data egress guarantee.
---
## 📊 Performance Metrics
- **Context Optimization**: 120k tokens -> 4k tokens (via Semantic Router).
- **Latency**: <200ms overhead per tool call.
- **Success Rate**: 92% for autonomous E2E tasks (measured on `boring verify`).
