ruOS v1.1.0 — Agentic LLM Reasoning + Intel Support + Security

ruOS v1.1.0 — Agentic LLM Reasoning + Intel Support + Security

The first agentic OS now reasons with a local LLM, protects itself with AI security, and runs on Intel hardware.

What's New (since v1.0.0)

Local LLM Reasoning (Qwen2.5-3B)

• Agent decisions powered by Qwen2.5-3B-Instruct running on CUDA
• RAG from brain: searches 600+ memories for context before each decision
• 1.2s reasoning latency, rate-limited to every 15 min (or on state change)
• Falls back to rules if LLM unavailable

AIDefence Security Layer

• 32 prompt injection patterns + 6 PII detectors + unicode homoglyph protection
• Scans brain content before it enters LLM context (RAG defense)
• Validates LLM output against action allowlist
• ruos-agent --security test — 6/6 threat categories detected in <0.1ms

Intel/OpenVINO Embedder

• ruos-embedder-intel — drop-in replacement for CUDA embedder
• Runs on Intel CPU (AVX-512/AMX), Intel Arc GPU, or Intel iGPU
• Same model (bge-small-en-v1.5, 384-d), same API

DiskANN Vector Index

• Vamana graph replaces brute-force array
• Scales to 1M+ vectors with ~300 MB RAM (graph only)
• Currently in brute-force mode (<2K vectors), graph search activates at scale

ruos-agent Daemon

• Observe-reason-decide-act loop every 5 min
• Self-monitor: restart crashed services, GPU thermal protection
• Auto-profile: LLM + rules evaluate GPU util + time + presence → profile switch
• Embed backfill: 100% vector coverage achieved (was 65%)
• Nightly: export + DPO train + self-eval at 3 AM
• OTA: weekly update check from GitHub releases

Bootstrap Wizard

• ruos-bootstrap — interactive installer with 7 deployment roles
• Auto-detects hardware (GPU, arch, RAM, sensors) and recommends role
• Roles: workstation, edge, cluster-primary, cluster-secondary, agent-only, docker, minimal
• Federation: QUIC brain sync, mDNS auto-discovery, ed25519 auth

Brain Optimization

• Deduplication: 1,588 → 600+ unique memories (66% reduction)
• Content resolution: blob read/write on all memory operations
• Search dedup: no duplicate results in search output
• SQLite vacuum: compacted storage

Packages

Package	Arch	Size	Description
ruos-core	amd64	4.0 MB	MCP server (102 tools), brain, profiles, init
ruos-core	arm64	829 KB	Same for Pi/Jetson
ruos-brain-base	all	359 KB	Pre-trained brain.rvf (50 memories)
ruos-desktop	amd64	3.6 MB	Tauri desktop dashboard
ruos-embedder	amd64	68 MB	CUDA embedder (candle + bge-small-en-v1.5)
ruos-embedder-intel	all	3.6 KB	NEW Intel/OpenVINO embedder
ruos-agent	all	21 KB	NEW Agentic daemon + LLM serve + AIDefence + bootstrap

Install

# Bootstrap (recommended)
bash <(curl -s https://raw.githubusercontent.com/cognitum-one/ruOS/main/scripts/ruos-bootstrap)

# Manual
sudo dpkg -i ruos-core_*.deb ruos-brain-base_*.deb ruos-agent_*.deb
# NVIDIA GPU:
sudo dpkg -i ruos-embedder_*.deb
# Intel:
sudo dpkg -i ruos-embedder-intel_*.deb

Performance

Metric	Value
Search	23ms avg (DiskANN)
Embed	2.1ms avg (candle-cuda)
LLM reasoning	1.2s (Qwen2.5-3B, CUDA)
AIDefence scan	<0.1ms
Agent heartbeat	0.07s (rules) / 1.0s (LLM)
Search quality	0.77 avg score, zero drift
VRAM	7.6 / 16.3 GB (embedder + Qwen)

ADRs

19 architecture decision records including:

• ADR-0015: Agentic heartbeat daemon
• ADR-0016: DiskANN vector index + Intel embedder
• ADR-0017: Local LLM reasoning via Qwen2.5-3B
• ADR-0018: AIDefence security layer

Links

• Platform: cognitum-one/ruVultra
• RuView: ruvnet/RuView
• Cognitum: cognitum.one

ruOS v1.0.0 — The First Agentic Operating System

ruOS v1.0.0 — The First Agentic Operating System

ruOS doesn't just respond to commands — it observes, reasons, and acts on its own behalf. Built for the Claude Code era.

Agentic Capabilities

Capability	How	Schedule
LLM reasoning	Qwen2.5-3B on CUDA via ruos-llm-serve	Every 15 min (or on state change)
RAG from brain	Searches 544 memories for context before decisions	With each reasoning call
Self-monitor	Restart crashed services, GPU thermal protection	Every 5 min
Auto-profile	GPU utilization + time + presence → profile switch	Every 5 min
Embed backfill	Vectorize unembedded memories during idle	Every 5 min
Nightly training	Export pairs → DPO fine-tune → LoRA adapter	Daily 3 AM
Self-evaluation	Test search quality, detect drift	Hourly
OTA updates	Check GitHub releases, install new packages	Weekly

Performance

Metric	Value
Search latency	24ms avg (30x sustained)
Embed latency	2.1ms avg
LLM reasoning	1.2s per call (Qwen2.5-3B, CUDA)
Agent heartbeat	0.1s (rule-only) / 1.0s (with LLM)
Search quality	0.77 avg score, zero drift
Vector coverage	100% (544 unique memories)
VRAM usage	47% (embedder 498 MB + Qwen 6.3 GB)

Optimizations (this release)

• Deduplication: 1,588 → 544 memories (66% were duplicates)
• DiskANN Vamana graph: ready for 1M+ vectors (brute force at <2K, graph search above)
• LLM rate limiting: only reasons every 15 min or on significant state change
• Improved RAG: context-aware queries based on GPU state
• SQLite VACUUM: freed fragmented pages

Stack (all 10 levels complete)

Level	Component	Details
1	Identity	Ed25519 keys
2	Brain	RVF + SQLite, 544 unique memories
3	Embedder	CUDA bge-small-en-v1.5, 2ms
4	Search	DiskANN Vamana, avg score 0.77
5	MCP tools	102 stdio + 22 brain = 124
6	Profiles	6 profiles, auto-switch via LLM
7	Desktop	Tauri v2 + Svelte 5
8	Training data	3,265 preference pairs
9	DPO training	loss 0.081, 100% eval accuracy
10	Agent + LLM	Qwen2.5-3B reasoning + RAG + OTA

Packages

Package	Arch	Description
ruos-core	amd64	MCP server, brain, profiles, init
ruos-core	arm64	Same for Pi/Jetson
ruos-brain-base	all	Pre-trained brain.rvf
ruos-desktop	amd64	Tauri desktop dashboard
ruos-embedder-intel	all	Intel/OpenVINO embedder (drop-in for CUDA)

Install

sudo dpkg -i ruos-core_*.deb ruos-brain-base_*.deb
ruvultra-init setup

Architecture Decision Records

18 ADRs document every design choice:

• ADR-0015: Agentic heartbeat daemon
• ADR-0016: DiskANN vector index + Intel embedder
• ADR-0017: Local LLM reasoning via Qwen2.5-3B

Links

• Platform: cognitum-one/ruVultra
• RuView: ruvnet/RuView
• Cognitum: cognitum.one