The hub can now be drained and upgraded without losing work. A draining hub refuses new mutating work while it finishes what it is running, a durable event log lets a reconnecting client replay everything it missed, and durable runs are queued rather than dropped. Replayed events are deduped by event id, so an event that arrives on both the replay and the live stream is delivered once Fixed tool calling being silently disabled for custom OpenAI-Compatible models whose capability list was synthesized from convenience flags like `supportsReasoning`. An inferred list read as an authoritative denial and stripped every tool from the request; explicitly authored capability lists still decide Langfuse traces now carry session and client identity for hub-backed and delegated-agent runs.

Hub sessions rebuild client name and version from request headers, and delegated agents group under their parent session instead of appearing as unattributed traces Refreshed the model catalog, which updates model lists and pricing across providers and changes the resolved default model for several of them (DeepSeek, Crof, CrossModel, Eden AI, Kilo, and NanoGPT). If you use one of those providers without…