Перейти к содержанию

Hybrid AI Architecture for Air-Gapped Development — ADR

Status: ACCEPTED | Date: 2026-07-18

Decision: Hybrid Core + Modular Agents

Architecture: DevoxxGenie (core) + Cline (VS Code) + Veai (RU, opt) + Tabby (completion)

LLM: llama.cpp / Ollama — CPU-optimized, OpenAI-compatible API

MCP: filesystem, git, SAST, jfr, memory — all local

Memory: Memory Bank (.md) + RAG (Qdrant/ChromaDB)

Models: Qwen3-Coder-30B, GigaChat-20B, CodeLlama:code, nomic-embed-text

Observability: Prometheus + Grafana

Component Selection (weighted analysis v2)

Component Role License Score
DevoxxGenie Core IDE agent Apache 2.0 8.55
Veai RU-market agent Proprietary 8.10
Continue.dev Archived Apache 2.0 8.20
Cline VS Code agent Apache 2.0 8.05

DevoxxGenie selected as core due to open-source + approve workflow. Veai recommended for RU organizations requiring geopolitical resilience.

Russian Ecosystem

Tool Type Air-Gapped
Veai JetBrains plugin VPC/self-hosting
GigaCode Multi-IDE plugin GitVerse on-prem
GigaIDE Standalone IDE Local install
GigaChat-20B LLM (GGUF) llama.cpp
SourceCraft JetBrains+VS Code Verify

All IDE plugins in 38-ide-plugins.sh are open-source (Apache 2.0/MIT). Proprietary tools documented but not auto-installed.