What problem does it solve? Hermes Desktop sessions can silently route tool turns to unsuitable models (Copilot ACP, local 3B models, or OpenRouter as primary) and produce fake tool-call theater instead of real terminal or browser calls, wasting tokens and breaking agent reliability. ## Core Features & Use Cases - Pin-vs-route enforcement: Pins hermes-main via the LiteLLM gateway for Desktop tool turns and blocks copilot-acp, hermes-local/qwen2.5:3b, and OpenRouter-as-primary, allowing only nous-deepseek → glm-5.3 → gpt-4o fallbacks. - Ori-style tool eval without an LLM judge: Asserts transcripts called terminal or browser directly and fails on meta tool_call without a name, nested tool_search, or babysitting prose like fake pip install output. - Local usage analytics: Parses LiteLLM traffic JSONL at ~/.hermes/litellm-logs/traffic.jsonl instead of calling the paid OpenRouter Analytics API. - Use Case: Before letting a Hermes Desktop agent run a tool turn, run the policy script with your provider, model, and intent; proceed only when the JSON output reports route.allow is true. ## Quick Start Run the router policy script with provider custom:litellm-gateway, model hermes-main, and intent tool, then proceed only if the JSON output allows the route.
