Wearables expose signals about a person’s physiological, cognitive, and affective state, and large language models (LLMs) can turn them into personalized guidance. Existing systems typically run in the cloud and emit free-form text. We present Syni, an agent built on a strict separation: a deterministic upstream engine produces a versioned Human State Interface (HSI); Syni is purely its consumer. This lets Syni run offline-first on-device, condition through an audited seam, guarantee schema-safe output via grammar-constrained decoding with a generate–validate–repair–fallback pipeline, and keep raw biosignals off the model—simplifying deployment, validation, and privacy. In preliminary measurements, an unversioned renderer change between training and inference silently collapses state-reflection from \(85\%\) to \(0\%\) while schema-validity stays \(100\%\) —a failure no schema check catches. Grammar-constrained decoding lifts a 1B model’s first-pass validity from \(0\%\) to \(100\%\) , and coach inference runs on an iPhone 17 Pro in ∼ 2 s—showing contract-mediated, on-device agents are practical.