What problem does it solve? Running multiple local LLMs on CPU-only hardware with limited RAM risks out-of-memory crashes, concurrent 7B inference overload, and inconsistent model naming. This Skill enforces the SwarmXQ platform's model orchestration invariants so every model acquisition, eviction, and tag resolution follows a single safe contract. ## Core Features & Use Cases - SINGLE-7B Lock Enforcement: Ensures only one 7B-class model is inference-active at any time by requiring evictIncompatible() before every 7B load. - Canonical Tag Resolution: Routes all legacy aliases (phi4-fast, deepseek-reasoner, qwen-worker) through resolveCanonicalTag() so only canonical tags reach the model registry. - RAM Pressure Gating: Implements readPressure() and getRamSnapshot() against /proc/meminfo with normal/high/critical levels and protected constants like RAM_CRITICAL_MB=800. - Use Case: When modifying model-orchestrator.ts or video-runtime-config.ts to add a new pipeline stage, use this Skill to correctly acquire the Architect model, resolve its canonical tag, and degrade gracefully to the lite model on 8 GB hosts. ## Quick Start Ask the AI to review a change to model-orchestrator.ts and verify that every 7B model acquisition calls evictIncompatible() and resolveCanonicalTag() before loading.
