The Overview screen is the landing screen of the Voice AI console. Before it existed, the console opened on whichever agent happened to be first in the roster — useful for “what is this agent doing”, useless for “what is my voice product doing”, and an empty workspace for a tenant with no agents yet. Every figure on the screen comes from a procedure another screen already uses. Nothing is computed optimistically. A figure that cannot be sourced from a procedure is not rendered at all, because a confident wrong number on a landing page is worse than a missing one.

​
What the four KPI tiles measure

While a query is in flight its tile shows … rather than a zero, so an in-progress load never reads as “no traffic”. The three month-to-date tiles refresh when their query refetches. The active call count refetches every 15 seconds, and the published-agent badge every 60 seconds.

​
How month-to-date is scoped and de-duplicated

cdr.monthStats reads the cdrs table in ClickHouse with two filters:
  • Tenant scope — tenant_id equals the organization behind the current console session. cdr.monthStats is an org procedure and takes orgId as input; the tenant identifier is derived from it, so the tiles never blend traffic from another organization.
  • Month scope — toStartOfMonth(starting_time) = toStartOfMonth(now()). This is a calendar-month-to-date figure keyed on the call’s start time, not a rolling 30-day window. The tiles reset when the month rolls over, and a call that started in the previous month is attributed to that month even if it ran past midnight on the first.
De-duplication. A single call can produce more than one CDR row. The query therefore aggregates in two stages: an inner query groups by call_id and takes max(duration) and max(cost) for each call, and the outer query counts those groups and sums the per-call maxima. The consequences are worth knowing when you reconcile these numbers against a raw CDR export:
  • calls is a count of distinct call_id values, so re-written or duplicated rows for the same call do not inflate it.
  • duration_sec and cost take the largest value seen per call rather than summing rows. If your pipeline writes a partial row followed by a final row, the final (larger) value wins; summing the rows directly would double-count.
If no rows match, the procedure returns zeros for all three fields rather than an empty result, so the tiles render 0 rather than blanking.

​
How an active call is determined and when it goes stale

“On a call now” and the Active calls card are both fed by telephony.listActiveCalls. A call is listed only if it passes every one of the following checks. 1. It is inside the activity window. The procedure treats calls started within the last 60 minutes as candidates. Real calls finish in seconds-to-minutes; anything older than that is almost certainly a stale index entry left behind by a hangup path that never removed it — a carrier disconnect, a gateway restart mid-call, or a Redis blip. An earlier 8-hour window was a holdover from the SCAN-fallback era and surfaced calls that had hung up hours before, which is exactly the kind of wrong number this screen avoids. 2. It is in the per-tenant active-call index. Candidate call SIDs come from a per-organization sorted set, read newest-first by start time, bounded by the limit input. 3. The session has no end marker. For each candidate the procedure checks whether the session’s :end hash exists. The agent runtime writes that hash on hangup, via the RTP silence-finalize or dialog-closure path. Its presence means the call is over even if the release path failed to remove the index entry, so the call is excluded. 4. The index entry self-heals. Candidates found to have ended are removed from the index inline, as a best-effort cleanup. If the cleanup fails, the entry is simply filtered out again on the next listing. This is why a stale entry you saw once usually does not come back. 5. The session record still agrees. For surviving SIDs the procedure reads the session hash and drops the row if the stored tenant_id does not match the requesting organization, if the start time is unparseable, or if the start time turns out to be older than the 60-minute cutoff after all. Rows are returned sorted newest-first by start time, capped at limit. Each row carries the call SID, tenant, agent, trunk, realm, start time in milliseconds, and the node handling the call. The Overview card renders the first six rows and, when there are more, offers a button through to the monitoring screen. Each row shows the call’s endpoints and its status; a field the session did not record renders as — rather than as a guess. Because the count is the length of this filtered list, it can differ from a raw count of index entries. That difference is the point: the index is the candidate set, and the liveness gate is what makes the number true.

​
The four steps to a number that answers

The Get your first agent answering calls card lists the four things that stand between a new tenant and a real phone call. All four must be true before an inbound call is answered. The card’s badge shows how many agents are published, taken from agentVersions.publishedAgents. This is deliberately different from how many agents exist: an agent object in the roster is not a live agent. The “Build an agent” step is marked complete only when at least one agent is published, and the badge is the number you check when a number rings but nothing picks up. The other three steps are always offered as links rather than being ticked off, so the card stays a navigation aid rather than a checklist that hides the quickstart once you have used it. Each button navigates within the console using the shell’s navigation setter rather than reloading the page. If the console has no organization in context, the screen replaces all of the above with a prompt to sign in on the portal, since every procedure on the screen is org-scoped and none of them can run.

​
Where to go for detail

The Overview screen is a summary. Follow these for the underlying detail:
  • Monitoring — the full active-call list, when more than six calls are live. The “See all N” button goes here.
  • Trunks — carrier connectivity, for the trunk step.
  • Numbers — number-to-agent routing, for the final step.
  • Agents — agent authoring and publishing, which is what moves the published count.
  • Telemetry — metrics, traces, and the CDR schema that backs the month-to-date tiles.
  • Telephony metrics — definitions for the call-quality and traffic numbers you will meet on the detail screens.