dsh-response-meta

In DSH Web, every model reply carries runtime signals—model name, reasoning extent, tokens per second, time to first token, total runtime—that are buried in a hover-only native timing group, and manually interrupted replies have no usage record at all. Checking how a single call actually performed means repeated hovering, or simply not having the data.

dsh-response-meta renders a persistent runtime summary on a dedicated second row of every assistant reply, covering both completed and interrupted responses. The model name is read from the turn's request/header session event and inherited for subsequent header-less turns; reasoning extent prefers provider-reported reasoning tokens with a fallback to reasoning text length; tokens per second is computed over decode time from first token to message completion, matching DSH's native timing semantics. Interrupted replies, which usually lack a usage record, still show the model and streamed reasoning extent, and any real usage rescued by a finalization race is retained. Missing fields are omitted individually; when all are unavailable nothing is rendered; narrow screens wrap without truncation or horizontal overflow.

Well suited for heavy DSH Web users who switch between models often, care about TTFT and reasoning depth, or frequently interrupt generation mid-stream—key per-reply metrics become visible at a glance with zero extra configuration.