Omniscio

Control server request windows (what the local API is doing)

Omniscio's local control server measures its own traffic in short windows and reports it: how many requests, how many failed, how many were slow or abandoned, and how long they took. You can read the latest numbers on demand, and a route that keeps failing raises one inbox card until it recovers.

On this page

What it is

The control server is the local HTTP API that sessions, scripts and hotkeys call. A "request window" is a rolling bucket of that traffic: the server tallies each window's requests, its 4xx and 5xx answers, the requests that crossed the slow threshold, the ones the caller abandoned before an answer, and the average and peak latency.

It exists so throughput and latency can be charted over time, and so a slow-request threshold can be set against real history rather than a guess. It measures the server; it is not a record of what any individual caller asked for.

Where to find it

  • On demand: the control server's GET /diag/cli-metrics route answers the current totals.
  • In the log: each flush writes a summary line tagged [cli-server][metric] window: to main.log.
  • In the inbox: a route that keeps returning server errors raises one card — "control server: one route keeps failing" — which clears itself when that route recovers.

How it behaves

  • A window flushes at 200 requests or every five minutes, whichever comes first. An empty window is not flushed at all, so an idle machine writes nothing.
  • Slow means ten seconds or more. A request slower than that is counted in the window's slow tally as well as in its request count.
  • An abandoned request is not a request. A caller that disconnects before an answer is counted as abandoned, so a client that gives up does not look like server load.
  • One row per flush. Each flush persists a single telemetry row carrying only numbers — requests, 4xx, 5xx, slow, abandoned, average milliseconds and peak milliseconds. No route path, query, header or caller identity is stored with it.
  • The alarm is deduplicated and self-clearing. A single route reaching ten or more 5xx answers raises the inbox card once; repeated failures do not stack cards, and the card goes away when the route serves again.
  • Fleet escalation has a floor. A window is escalated when it carried at least twenty requests and at least a fifth of them were 5xx or slow, so one early failure in a nearly empty window does not page anyone.
  • It never breaks a call. Measurement is fail-open: an error recording a window is swallowed and logged, and a relocated instance suppresses the inbox card while still writing its log lines.

Last verified 2026-11-09