The past day of Hermes Agent development reads like a checklist of the boring-but-load-bearing problems that decide whether an agent feels trustworthy. The commits cluster around three themes: gateway readiness, session-scoped model selection, and usage attribution.
Gateway readiness now means "actually running"
Two commits attack the gateway's idea of "ready." One adds authenticated runtime readiness checks, so the gateway stops reporting healthy before the runtime it fronts can actually serve. The other goes further: it grounds readiness in live runtime state instead of a cached assumption. If you have ever watched a desktop app show a green dot while the backend behind it was dead, you know why this matters — a false ready signal means the first prompt fails in a confusing way instead of at startup, where it is easy to diagnose.
Model switches stay where you put them
The model picker got two related fixes. A TUI fix preserves picker session scope across all paths, and a desktop fix keeps model picker switches session-scoped. The behavior being defended: when you switch models mid-conversation, that choice should apply to this session — not leak into every other open session or evaporate the moment you navigate away. There is also a feature commit that tracks per-model token usage for mid-session model switches, which is the accounting side of the same coin: if you bounce between a cheap and an expensive model inside one session, the token ledger now knows which is which.
Usage attribution got hardened
Three commits touch usage attribution and telemetry. One hardens per-route usage attribution, another maps usage attribution contributors, and a telemetry fix persists a first-accounted fallback route. Together they are about answering a question every team running shared agents eventually asks: who spent what, on which route, and why does the number not add up at the end of the month.
None of this is glamorous. It is the difference between an agent you can run in a team and one that only works on the maintainer's laptop — which is exactly why Hermes positions itself as a local-first autonomous assistant rather than a demo. The skill system is where those sessions actually do work, and these commits make sure the plumbing underneath it does not lie to you.
Top comments (0)