v2026.10.43 — VRAM Orchestrator + NVIDIA GPU Skills

What's New

VRAM Orchestrator (v2026.10.43)

Automatic GPU VRAM management for NVIDIA GPUs (tested on RTX 5090, 32GB).

• nvidia-smi polling — periodic GPU state monitoring (VRAM usage, temp, utilization)
• Model auto-swap — when GPU services need VRAM, the orchestrator downgrades the LLM to a smaller model, then upgrades back when done
• Time-bounded leases — services reserve VRAM with auto-expiry, preventing memory hogging
• Async mutex — all VRAM operations serialized, no race conditions
• Emergency OOM — auto-unloads all models if VRAM drops below critical threshold
• 3 agent tools: vram_status, vram_acquire, vram_release
• 4 API endpoints: GET /api/vram, POST /api/vram/acquire, POST /api/vram/release, GET /api/vram/check
• Config schema: vram.* section with services budget map, thresholds, auto-upgrade

NVIDIA GPU Skills (v2026.10.42)

• cuOpt — GPU-accelerated vehicle routing optimization via NVIDIA cuOpt v26.02 async API (tested live, VRP solved in 74ms)
• AI-Q Research — Deep research via Nemotron Super 49B NIM API with citation extraction
• OpenShell Sandbox — K3s-based secure code execution with declarative YAML policies
• All gated behind TITAN_NVIDIA=1 env var

Fixes (v2026.10.41-42)

• Voice mic leak — VoiceOverlay cleanup
• Tool visibility — security.allowedTools default
• OpenAI-compat keepModelPrefix bug
• 6 TypeScript type errors
• Voice system prompt rewrite (500 tokens vs 3000+)

Stats

• 4,321 tests across 135 files (all passing)
• 0 TypeScript errors, 0 ESLint errors
• ~155 tools across 100+ skills