Hermes v3 — The Divine Messenger of Task Routing

Named for the swift messenger of the gods, Hermes carries each task to the model destined to serve it. It treats routing as the core optimization problem: rather than paying an LLM to decide what a task is, Hermes classifies every prompt with a zero-token weighted-heuristic scorer — regex signals across eight categories (NER, sentiment, summarization, factual QA, math, logic, debugging, codegen) fused with paraphrase-proof structural cues (code fences, digit density, passage shape, multiple-choice markers). The highest-scoring category wins; only genuinely ambiguous non-code tasks spend a few tokens on a cheap one-word LLM tie-breaker. The runner fans tasks out concurrently (semaphore-bounded), dedupes identical prompts to save calls, and retries with corrective re-prompts and model escalation on malformed output — all under a hard runtime budget with a guaranteed final flush. Output is written atomically (.tmp → validate → os.replace), so /output/results.json is always valid JSON even on partial completion or crash. Hermes ships as a lean linux/amd64 container (python:3.12-slim, just main.py + the OpenAI client), secrets injected at runtime by the harness, distributed via GitHub Container Registry. The result: accuracy where it matters — reasoning tasks get strong models and room to think — and frugality everywhere else, since routing itself costs nothing.