How ReacherX turns activity into Analytics and Agent observability

features/agent-ops/ui/AgentOpsDashboard.tsx is △ Agent-operation view. Its tabs include overview, discovery, quality, memory, and activity. Users can inspect queries, memories, events, evaluator runs, and suggestions through detail views.

These are related views, not interchangeable counts. A memory event is not a new prospect. A generated query is not necessarily an activated query. Keep those meanings intact when adding a metric.

Follow a value through the backend

Start with convex/analytics.ts for dashboard assembly and convex/lib/analyticsCore.ts for shared calculations. The current response-rate calculation uses recorded responded and contacted events in the chosen window:

Read the numerator and denominator definitions around that line before changing the label. This is an event-window calculation. It is not automatically a cohort study following the exact same contacted people until every reply arrives.

Daily reporting rows contain hourly information used for windowed calculations:

Account for readiness and time

The frontend checks reporting readiness through convex/workspaceReporting.ts. When reporting is ready, it uses reactive queries. When it is not, the dashboards have snapshot-action paths rather than assuming empty results mean zero activity.

Time windows use the workspace reporting time zone. Check both the visible range and the normalized backend window when a count differs around midnight. The UTC event timestamp keeps event ordering and hour buckets consistent across the system; the local reporting day decides which day the dashboard shows as today.

The UI also changes labels for the workspace use case. A recruiting view can say candidates while the underlying record remains a prospect. Keep presentation names separate from the data semantics.

Make a detail view explain the summary

features/agent-ops/ui/AgentOpsPanel.tsx opens records behind the dashboard. A memory detail can show source, confidence, canonical instruction, status, and indexing state. An evaluator run can show what it produced.

That is useful observability: a reader can move from “something happened” to the record that explains it. New summary metrics should have a clear definition and a sensible investigation path.

For tests, start with convex/agentOpsDashboard.integration.test.ts, convex/agentOpsInventory.integration.test.ts, and convex/workspaceReportingAggregate.integration.test.ts. Include empty data, time boundaries, repeated events, paging, and a workspace with reporting not yet ready.

Before contributing a chart, write down exactly what one count means. Then prove it with a small fixture and inspect the rendered label. The contributor guide explains how to turn that focused change into a useful contribution.