Architecture

Three layers, two dependencies

The dashboard package contains three things that are usually discussed as one:

A read layer — contracts, drivers, the backend-neutral query IR.
An analysis layer — fingerprints, error reports, signal correlation.
A presentation layer — panels, the JSON API, the React app.

This package needs the first two and not the third. That split is the whole design:

The read layer is a hard dependency. Without it there is nothing to measure.
The presentation layer is a soft one. Every class that knows the dashboard's UI exists lives under src/Ui/, registers behind a check, and is the first thing to go when the dashboard is not served.

An architecture test enforces the boundary: nothing outside src/Ui/ may import the dashboard's panel, HTTP or facade namespaces. The one exception is RequestScope, which lives under an HTTP namespace but is a plain value object — the query layer's scope primitive, built here from config in a cron job with no request in sight.

The point is portability. If the read layer ever becomes its own package, this one follows by changing an import, not by being rewritten.

Where state lives, and where it does not


Occurrences, stack traces, spans, metrics	Your telemetry stores. Read on demand, never copied
Status, assignee, snooze, cooldown, what we paged about	Four tables here

This is why the dashboard's promise survives. WritesToBackend is about your telemetry stores — Tempo, Loki, Prometheus, ClickHouse. A local Laravel table is not one of those. The dashboard can still assert, and test, that it never writes to your data.

It also means these tables are cheap. Delete them and you lose the decisions your team made, not the telemetry.

The pass

scan ──▶ IssueLedger      reads the window, folds it into the ledger
     ├─▶ IncidentCorrelator   groups what started together, names the cause
     ├─▶ IncidentRecorder     persists it under a stable signature
     └─▶ Notifier             announces only what is news

Everything in Correlate/ is pure: it reads, it decides, it returns value objects. Persistence happens afterwards, in one place. That is what lets correlation re-run from scratch every pass and still produce a stable result — see incidents.