Watching the mesh without reading the mail

An air traffic control tower tracks every plane: where it is, where it is heading, when it took off. It does not open the luggage. Mesh telemetry is that tower for your nodes. It shows the flight paths of questions across the mesh, but the cargo, your documents and answers, stays sealed inside the nodes.

As nodes discover peers, route a query, request a synthesis, or handle something locally, each emits a small trace: which node, which event, an optional peer, a timestamp. These are sent fire-and-forget to the Control Plane, so emitting telemetry never slows down or blocks the actual work a node is doing.

The trace records that a node routed to a peer, not what was asked or answered. Raw vectors, document text, and query content never enter the telemetry stream by design; the boundary between the Data Plane and the Control Plane is exactly where the content stops and the metadata continues.

The Admin console subscribes to this stream over Server-Sent Events and renders it as a live graph of mesh traffic. Two details make it feel immediate rather than fragile. When you connect, the Control Plane replays a short buffer of recent traces, so the graph is populated straight away instead of starting blank. And a periodic heartbeat keeps the connection warm through idle timeouts and proxies, so the stream does not quietly die during a lull.

The replay buffer is intentionally bounded: it exists to seed the current view, not to serve as a long-term audit archive, so telemetry is not a substitute for durable audit logging. Because traces are fire-and-forget, a trace that cannot be delivered is dropped rather than retried, which is the right trade for observability but means the stream is best-effort, not guaranteed-complete.