snas-code 1.0.0

SANS CODE: Systemic Neural Agentic Solutions - An open-source global terminal agentic CLI for software engineering

SNAS CODE: Systemic Neural Agentic Solutions

Overview

SNAS CODE is a state-of-the-art, multi-agent orchestration framework engineered for autonomous execution of complex software engineering cycles. Built on a foundation of Heuristic Task Decomposition and Resource-Constraint Optimization, SNAS CODE enables high-parameter local Large Language Models (LLMs) to operate seamlessly on consumer-grade hardware through a proprietary Swap-to-Max memory management protocol.

By decoupling high-level strategic reasoning from low-level implementation, SNAS CODE ensures that architectural integrity is maintained throughout the development lifecycle, regardless of the VRAM limitations of the host environment.

---

System Architecture

The core of SNAS CODE is a dual-agent, asymmetric architecture designed for maximum efficiency and precision.

stateDiagram-v2
    [*] --> Initialization: Request Capture
    Initialization --> BrainPhase: Contextual Loading (Architect)
    
    state BrainPhase {
        [*] --> InvestigativeScan
        InvestigativeScan --> StrategicDecomposition
        StrategicDecomposition --> JSONPlanGeneration
    }
    
    BrainPhase --> ResourceSwap: keep_alive=0 (Unload Architect)
    
    state ResourceSwap {
        direction LR
        Architect_VRAM --> System_RAM
        Worker_Loading --> Worker_VRAM
    }
    
    ResourceSwap --> WorkerPhase: Context Handover
    
    state WorkerPhase {
        [*] --> AtomicTaskExecution
        AtomicTaskExecution --> ToolActivation: write/run/test
        ToolActivation --> LocalVerification
        LocalVerification --> [*]
    }
    
    WorkerPhase --> ResourceReversion: keep_alive=0 (Unload Worker)
    ResourceReversion --> Synthesis: Reload Architect
    Synthesis --> FinalOutput: Summary Generation
    FinalOutput --> [*]

1. Heuristic Orchestrator (The "Brain")

The Orchestrator functions as the system's prefrontal cortex. It utilizes high-reasoning models (e.g., Gemma 2 9B, Llama 3.1 8B) to map natural language requirements into a discrete, actionable state-space.

• Cognitive Domain: Investigative research, dependency mapping, and multi-step strategy formulation.
• Isolation Boundary: Operates in a read-only sandboxed environment during the investigative phase to ensure system stability.

2. Specialist Implementation Unit (The "Worker")

The Worker is a high-throughput, specialized agent optimized for code generation and surgical file manipulation.

• Execution Domain: Direct workspace interaction, automated testing, and terminal-level command execution.
• Verification Loop: Implements a "Test-Before-Commit" protocol, ensuring all modifications are validated against the current system state.

---

Core Innovations

Swap-to-Max™ Memory Protocol

To circumvent the "VRAM Wall" common in local LLM deployments, SNAS CODE implements an aggressive model-swapping strategy:

• Zero-Latency Unloading: Explicitly forces Ollama to purge model weights from VRAM when transitioning between phases.
• Context Preservation: Strategic state serialization allows the system to reload the Architect with full awareness of the Worker's progress without memory overlap.

Human-in-the-Loop (HITL) Authorization

SNAS CODE adheres to a Zero-Trust Security Model. Every destructive or state-altering operation (file writes, command execution) requires a signed cryptographic-like authorization from the user, preventing unintended side effects.

Workspace State Verification

Post-execution, the system performs a recursive scan of the local filesystem to verify that the generated artifacts match the architectural plan, providing a secondary layer of objective truth.

---

Technical Specifications

Component	Technology	Role
Orchestrator	Ollama API / Python 3.10+	Agentic Logic & Model Management
Interface	Rich (Console Protocol)	Real-Time Feedback & Stream Visualization
Logic Layer	Multi-Agent Directed Acyclic Graph (DAG)	Task Flow Control
Memory Opt	Swap-to-Max Protocol	VRAM Optimization

---

Deployment & Integration

Prerequisites

• Python: 3.10 or higher.
• Ollama: Local instance with compatible models.
• Hardware: Minimum 16GB RAM (32GB recommended for large model swaps).

Installation

git clone https://github.com/peeyush1278/SNAS-CODE.git
cd SNAS-CODE
pip install -r requirements.txt

Initialization

python main.py

---

Advanced Configuration

SNAS CODE features Dynamic Capability Detection. Upon startup, the system probes the local Ollama instance, heuristically ranking available models to automatically designate the most capable "Brain" and the most efficient "Worker."

---

“Engineering the future of autonomous local development.”