What problem does it solve? Hermes defines 18 configurable auxiliary model slots for side-jobs like vision, compression, title generation, and approval, but the dashboard only surfaces 11 and misconfiguring them wastes main-model tokens, breaks on deprecated Gemini models, or silently fails with thinking models on LM Studio. ## Core Features & Use Cases - Tiered slot assignment: Maps the 18 aux slots to cost-appropriate models, from gemini-3.6-flash for vision down to cheap flash-lite tiers for titles and approvals. - Thinking-model handling: Explains why Qwen3.5/LFM2.5 models return empty content at low max_tokens, why per-request thinking toggles are silent no-ops on LM Studio, and how to build non-thinking virtual models via model.yaml. - Privacy provisioning: Tiers providers from local LM Studio through TEE confidential inference (Chutes, Phala, NEAR AI) down to free tiers that train on data, for security-sensitive environments. - Use Case: A user on a 16GB Mac wants vision and web extraction on local Qwen3.5-4B, titles on a non-thinking LFM2.5 model, and compression on a TEE provider; this Skill gives the exact hermes config commands, fallback chains, and verification probes. ## Quick Start Ask the agent to configure the Hermes vision and compression auxiliary slots to use gemini-3.6-flash with a fallback chain and verify the result with hermes config get auxiliary.
