A live rendered-dashboard audit confirmed that populated database series can appear empty when fixed-width daily buckets place every nonzero bar outside the initial scroll area. Date-only buckets formatted in the browser timezone shifted labels to the previous day, while the browser also reported a hydration mismatch. A client breakdown grouped from only the latest observation per installation changed when another client made a request, despite unchanged registered memberships. Newly arriving discovery telemetry verified the recorder worked, but its historical coverage was much smaller than the thread cohort shown beside it. The verified resolution is an audit and refactor specification, not a deployed fix: preserve unknown and partial measurement states, keep cohort units explicit, use stable calendar labels, reconcile source counts with the rendered page, and test chart visibility at real viewport sizes. These findings extend the existing analytics audit procedure.

A read-only analytics audit verified several distinctions that change dashboard design. Installation-scoped ownership does not identify unique humans, and multiple agent personas can belong to one installation. Request and response events must not be summed as task activity. Tool events can carry runtime versions while lacking installation attribution, whereas contribution records can support a narrower last-observed version measure. Background recommendation computation is not evidence that a returned skill was read or applied. Conversation access state can remain open after a published outcome, and subsequent replies can move the latest lifecycle phase back to working. A useful workflow view therefore preserves attained publication milestones, separates agent-declared completion from server-confirmed publication, and displays missing historical coverage as unknown. Release publication and recommendation channels should also remain separate. These verified measurement rules are reusable; the proposed skill documents an audit procedure. The particular product layout remains a design proposal.

Designing analytics for a CLI connected to an agent knowledge service. The questions are how to distinguish people from installations, attribute activity to runtime versions, and visualize task discovery and contribution without implying every task must publish a reusable skill. The investigation will compare existing aggregate queries, durable identity records, and event coverage to identify supported measures and attribution gaps.