Every call attempt now leaves a diagnosis

CallPages now instruments every call attempt end to end. Each attempt stores a sanitized diagnostic record — what the browser reported, where the attempt died, how long it took — feeding a Diagnostics view and an attempt funnel, so a failed call is evidence, not a mystery.

When a call died, the record said almost nothing — ‘websocket error’ covered a multitude of sins, and identical-looking failures (some only two seconds long, over and over) couldn’t be told apart from real customers struggling. You cannot fix what you cannot see, and we could not see.

Two seconds, fifty times

Illustration: two phones with voice waves and a friendly speech bubble — a CallPages conversation

The case that forced this work: a call that lived for exactly 2,010 milliseconds and died with a one-line error — repeated until it looked like weather. Without diagnostics, each of those attempts was indistinguishable from a customer with a broken phone. With them, the pattern names itself: same duration, same stage, same signature — an infrastructure fingerprint, not fifty separate accidents.

Instrument first, argue later

CallPages’ standing rule is that failure capture ships in version one of everything. Diagnostics is that rule applied to the call itself — the product’s core event. Owners get a section that answers ‘are my calls working?’ with data; engineers get bundles that reproduce instead of rumors; and the next two-second ghost will be caught by the funnel the day it appears, not by a customer writing in.

The Learning Center version of this article is in preparation and will follow with a verified walkthrough of the Diagnostics section.

How it works

Every call attempt — connected or not — writes a diagnostic bundle: browser and device class, microphone state history, connection timing, and the exact failure stage, sanitized before storage.
Diagnostics live in their own store and their own dashboard section, so the call record itself stays clean and the funnel counts attempts honestly: started, ringing, connected, failed — and where failures cluster.
Failures raise a ladder of alerts instead of silence: an attempt that dies in two seconds with a socket error is a pattern the system now surfaces on its own.
Instrumentation shipped in the first version, not as a follow-up — the learning loop starts with the first real caller, not the first complaint.