Laravel Telemetry Ui Laravel Package

The header's Copy link button copies a shareable URL for the current view.

A desktop host serves the dashboard from 127.0.0.1 on a port it picked, so what it copies only ever resolves on that machine, while that app is running — a button offering to share something unshareable, in the header of every page.

'copy_link' => env('TELEMETRY_UI_COPY_LINK', true),

On by default; most hosts are reachable at the address they serve from. Config rather than a host override, because the honest answer for such a host is that the feature does not apply, not that it should be restyled.

Nothing else changed. No API moved.

The problem

Every dashboard page load resolved which pages to show by asking the backend once per page with a detect pattern. There are 16, each its own PromQL round trip, in series.

Measured against a remote Grafana through its datasource proxy: the shell took 2.25 s with the detect cache cold, reproducible to ±0.05 s. The same shell against a loopback backend was ~10 ms, and the backend answered any single query in 2–40 ms. Not query cost — sixteen round trips of latency, one after another.

The per-pattern cache hid it until the TTL lapsed, which is why it presented as the app randomly becoming slow rather than as a slow page.

The fix

Contracts\EnumeratesMetricNames — a new optional driver capability:

metricNamesMatching(array $patterns, string $scope = ''): array

PrometheusSource implements it with one /api/v1/label/__name__/values call whose match[] selector is the alternation of every pattern plus the scope; MimirSource inherits it with its prefix. The detector then decides each pattern against the returned names in PHP.

Optional on purpose, following the existing idiom (ProbesConnection, AggregatesSpans, CreatesIssues): widening MetricsSource would be a fatal break for drivers outside this repo, and a driver that does not implement it keeps the old per-pattern path unchanged.

Scope batches inside the same selector, so a scoped detection stays scoped. Fail-open is unchanged: a backend that cannot answer still shows every page rather than hiding them.

Upgrading

Nothing to do. No API removed, no signature changed, and a custom metrics driver that does not implement the new contract behaves exactly as before — one call per pattern.

If you have written a metrics driver and want the speedup, implement EnumeratesMetricNames.

Still slow, and not fixed here

The shell makes 18 other sequential backend calls for its cards and fleet label lookups. Detection is no longer the bottleneck; that is.

Detection's lookback is five minutes regardless of the selected period, so a metric family that stopped emitting ten minutes ago hides its page even when the reader is looking at last week. Pre-existing, and left alone deliberately rather than smuggling a behaviour change into a performance fix.

The fix

From cboxdk/laravel-telemetry v1.2 a refused export batch is reported rather than thrown. AnnotationWriter still discarded the return value, so its catch no longer saw the failure and write() returned true regardless — telemetry-ui:annotate and the scan-versions cron announced annotations the backend had declined.

Nothing was lost: the scanner reads what is already annotated from the backend, so a refused version is seen again and retried next run. What was wrong was the output, in the one place an operator looks.

The writer's catch also stops swallowing Error. It is deliberately broad so a backend outage cannot abort a cron, and that breadth buried a TypeError for an afternoon. A transport failure is the backend's problem; a TypeError is ours.

Why this is a minor and not a patch

cboxdk/laravel-telemetry now requires ^1.2, up from ^1.0. You cannot report a refusal without the version that reports one — against 1.0, where flush() returns void, the check is unreachable code.

If you are pinned below 1.2, Composer will refuse this release rather than install something that cannot work. Upgrading the emitter is the only step; nothing in this package's own API changed.

Upgrading

composer update cboxdk/laravel-telemetry-ui cboxdk/laravel-telemetry

If you assert on AnnotationWriter::write() in your own tests and mock the emitter, note that flush() returns an ExportReport from 1.2 and the writer now reads it. A mocked flush() needs to say what it returns — new ExportReport is an empty, successful one.

Three things a reader notices, and one extension point for hosts.

The time range and refresh interval now survive navigation and reload

The range lived only in the URL. pageUrl() re-attached it to nav links, so it survived some navigation — but any link that did not go through it (a host's navLink(), a card deep link, a trace drill-in, a reload of a bare URL) silently dropped back to the default. Pick "last 7 days", click into a trace, come back, and you were looking at the last hour without being told.

Auto refresh was worse: it persisted, but the control's label was initialised from the server-rendered <option>, always "off" — so the displayed interval and the running timer disagreed.

Both are now one piece of shared state, resolved once per request and remembered in a cookie. A cookie because the cards are server-rendered Livewire components that query their backends during the first render: the window has to be in PHP before the query runs, or the dashboard spends a full round of queries on the wrong window and then jumps.

An explicit URL parameter always wins, and taking one updates what is remembered. Deep links behave exactly as before. Copy link now pins the resolved state into the URL — without that, a shared link would retarget to the recipient's saved range.

Scope (service, env) is remembered too. Enforcement lives downstream of the value, so a remembered scope is exactly as powerful as a hand-typed ?service= — that is, it cannot widen a tenancy lock — and a value the lock no longer allows is dropped rather than shown.

The page header stays on screen

.tui-header carries the title, the scope switcher and the period selector: the controls reached for most on a long page, and the hardest to find once scrolled.

TelemetryUi::connection()

A host registers its connections and which is current; the header renders a native <select> that navigates on change.

TelemetryUi::connection('prod', 'Production', route('connections.switch', $id));
TelemetryUi::currentConnection('prod');

For hosts that mount the dashboard as their whole UI, where switching backend previously meant leaving it. Nothing registered renders nothing.

TelemetryUi::viewState() exposes the window, refresh interval and scope to a host, with a ViewStateChanged event to hook. Docs: docs/extension-points/view-state.md and docs/extension-points/connection-switcher.md.

Also

↻ Refresh now re-runs the cards instead of reloading the page, which threw away client state.

A request-scoped service that captured the Request pinned one request's state under runtimes that do not flush scoped bindings — fixed.

Upgrading

No API removed and nothing renamed. If you assert on rendered dashboard HTML you may need to account for the connection <select> (only when a host registers connections) and the sticky header's wrapper styles.

TelemetryUi::navLink() registers a link at the foot of the icon rail, on every page including trace detail.

Needed wherever the dashboard is the application — a desktop shell, a kiosk window, an iframe — where the rail is the only navigation the reader has and there was previously no way out.

TelemetryUi::navLink('connections', 'Connections', route('connections.index'), 'connection');

Icons are names (back, home, settings, connection, server, database, user), not markup — the rail draws inline SVG and host strings never reach that sink. Nothing is registered by default.

Also bumps guzzle to 7.15.2 and commonmark to 2.9.0, clearing a high-severity host-check bypass (CVE-2026-69246) and several DoS advisories.

Docs: docs/extension-points/navigation.md

Fixes the TelemetryUi facade [@method](https://github.com/method) annotation for resolveConnectionsUsing($resolver, bool $needsViewer = true), which 1.4.0 left stale — static analysers flagged correct calls as passing an unknown argument.

Additive release, no BC break.

• Contracts\ProbesConnection — ConnectionManager::probe($name) returns a classified ProbeResult (reachable / TLS / unauthorized / not-found / unexpected-API). Drivers check the API shape, so a Loki URL pasted into the traces field is caught at test time rather than on every card. Never throws.
• Per-connection TLS verify — true, a custom CA bundle path, or false. Only an explicit boolean false disables verification; a stringy "false" from an env var fails closed.
• Contracts\WritesToBackend — makes a read-only posture checkable rather than claimed.
• resolveConnectionsUsing(..., needsViewer: false) — for unauthenticated single-user hosts. Default stays viewer-gated.

See CHANGELOG.md for detail.

Added

• Statamic pages carry a breakdown, not just a chart. Under each thin overview chart, a list a reader can scan: Static Cache by outcome, Glide by preset, Forms by form, Content by type — which item dominates, without per-item spans. Stache (no facet label to group by) gains a warm-build latency table instead — the p50/p75/p90/p95/p99 distribution behind its single P95 stat.
• Analytics audience tables now show visitors, not just views. The Campaigns and Sources & audience breakdowns gain a Visitors column beside Views (the distinct-visitor count was already computed but never shown), with a fixed table layout so the compact side-by-side tables can't overflow into each other.

Changed

• Navigation reorganised. Queues split into their own sidebar group, and Monitoring broken into Frontend (Analytics, Web Vitals, Users) and Infrastructure (Hosts, Logs, System) — so each area is a focused tier-1 entry instead of one long list.

Fixed

• Chart bars no longer vanish on hover. On the bar charts (e.g. a page's Traffic), ECharts' emphasis state repainted the stacked step-area with a collapsed baseline, so the fill disappeared while the pointer was over it. Emphasis is disabled on the series (the axis tooltip is the hover affordance); annotation markers keep their own emphasis.
• Detail pages light up the nav again. A hidden detail page (a page, issue, host, request, …) now highlights its list page in both the icon rail and the subnav via a declared parent, instead of leaving the whole navigation with nothing active.
• Web Vitals aggregate summary restored. The Core Web Vitals card passed its p75 LCP/CLS/INP summary to the stats component under the wrong attribute (:stats instead of :items), so the headline was silently dropped and the page rendered as a bare table.

Changed

• Require cboxdk/laravel-telemetry ^1.0 (was ^0.3.0). Aligns the dashboard with the now-stable telemetry 1.0 line — the old ^0.3.0 constraint excluded telemetry 0.4 and 1.0, so a fresh install pinned an older telemetry. No dashboard code changes; verified against telemetry 1.0.0 (326 tests green).

A visual overhaul plus New-Relic-style database dashboards. No PHP API changes — the MetricsSource / TracesSource / LogsSource contracts and every result DTO are unchanged from 1.0.0, so drivers built for 1.0 keep working untouched.

Added

• Query performance dashboards (New-Relic style): DB queries aggregated by normalised statement and ranked by the total DB time they consume (not just the single slowest run), with calls / avg / p95 / max / share, N+1 detection, a per-minute throughput chart, and a query-detail drawer.
• Searchable comboboxes replace every native <select> (scope, period, auto-refresh, and all per-card filters). Type-ahead filtering, full keyboard navigation (↑/↓/Enter/Esc), and a selected-state check. Each wraps a hidden native <select> so wire:model.live, x-model, and form navigation keep working unchanged.

Changed

• Full reskin to the Cbox design system: warm-neutral oklch tokens, a light theme by default with a dark toggle (persisted, no flash-of-theme), Plus Jakarta Sans display + JetBrains Mono for data/IDs/numbers, and clear 1px borders. Charts (ECharts) now read the design tokens, so they follow light/dark automatically.
• Two-tier app shell (Intercom model): a 56px icon rail — three states (minimised / hover-overlay-with-labels / pinned in-flow) — plus a collapsible contextual subnav (header toggle or ⌘.). Both states persist in localStorage.

Notes

• Cosmetic-only heads-up for downstream customisation: the legacy --tui-* CSS variables are now remapped onto Cbox tokens, and the filter/scope pickers no longer render a visible native <select> (a hidden one is retained for binding). If you overrode --tui-* values or styled those selects directly, review your overrides — no code migration is required.

Added

• Scope lock (tenancy). Constrain the whole dashboard to a fixed set of services / environments — dynamically per user via TelemetryUi::restrictScopeUsing(), or statically with no code via telemetry-ui.scope.lock / TELEMETRY_UI_LOCK_SERVICES / TELEMETRY_UI_LOCK_ENVIRONMENTS. The picker now reflects the lock: it offers only allowed values, drops "All" for a locked dimension, and hides a picker locked to a single value entirely. Enforcement stays at query time.
• Page-detail show pages. Analytics and Frontend rows now open a dedicated page-detail scoped to that one URL path — traffic (views, visitors, referrers, countries, devices), real-user performance (Core Web Vitals + load timings), the page's browser→backend traces, and its browser errors — instead of dumping into a pre-filtered trace search.
• Richer visitor breakdowns. The analytics "Sources & audience" card gains Sources, Regions, Cities, Operating systems and Browsers alongside Referrers, Countries and Devices, each toggleable via telemetry-ui.analytics.dimensions.
• Marketing channel — a first-class, zero-cardinality dimension. Visits are classified into Direct / Organic search / Social / Email / Paid / Referral / Internal, derived at read time from the referrer (and UTM medium + paid click-id once the emitter captures them), so it costs no ingest cardinality. Configure your own hosts with telemetry-ui.analytics.internal_hosts.
• Campaigns card. UTM campaign attribution (campaign / source / medium, and higher-cardinality content / term) from the emitter's telemetry.analytics.utm capture — with a single empty state until it's on.
• Cardinality guide in the analytics cookbook: per-dimension high/low classification, the capture-vs-surface control split, and the ClickHouse path.

Fixed

• Chart-annotation toggle no longer shows a duplicate "Cache purge" (the Statamic marker is now its own labelled, distinctly-coloured entry) and its rows/labels align in one clean column.

Changed

• Requires cboxdk/laravel-telemetry ^0.3.0 for the UTM / campaign capture the Campaigns card and channel enrichment read.

Added

• Dashboard cards drill into their pages. Cards that summarise a dedicated page (Requests activity/duration, Exceptions, Jobs, queue and autoscale cards) gain a "Requests →"-style header link when rendered on the dashboard or as an embedded widget, carrying the active period/service/env scope. On the card's own page the link is suppressed. Package cards opt in by setting protected ?string $drillPage = 'my-page'.

• Hide chart annotations per type. The header gains a ⚑ toggle listing the configured marker types (Deploy, Incident, …) with a checkbox each, plus a show/hide-all master switch — uncheck the noisy ones and every chart drops those lines instantly. Purely client-side: cards always ship the full annotation set and the charts filter marker lines by kind, so toggling costs zero backend queries. The choice sits in the URL (ann_off), so it survives navigation and deep links.

• Grafana-style relative time ranges. ?from=now-1h&to=now, now-7d, now+30m … (units s/m/h/d/w/M/y) work everywhere from/to do — evaluated at view time, so a shared relative link always shows the trailing window instead of a frozen one. Plain unix seconds still work, and the header shows relative expressions verbatim.

• Livewire updates carry their component everywhere. With cboxdk/laravel-telemetry ≥ 0.2.1, POST /livewire/update is named livewire:{component} (batched updates: livewire:batch), so the routes table groups per component instead of lumping thousands of opaque updates into one row. The request log shows the component(s) behind each update inline, and the Livewire page gains the Requests page's grouping/live-tail pair: a per-component table and a scoped live request log.

• Collapsible sidebar navigation. The nav groups (Activity, Monitoring, Statamic, …) collapse to chevrons so a long page list fits on one screen; only the active group opens by default and the choice persists in localStorage. Top-level items (Dashboard, Traces, Issues) stay visible.

• Frontend page rows drill into their traces. Core Web Vitals and Page performance rows open the browser→backend traces for that URL path, matched on span.url.path, carrying the active scope.

Fixed

• Drill links no longer appear on a card's own page. Cards lazy-load in a follow-up Livewire request where the page route param is gone, so every card thought it was on the dashboard and grew a self-referencing "Autoscale →" link. The page view now passes the page slug into each card at mount.
• Sparse counters no longer read as zero. Scaling actions and SLA breaches born inside the selected window were invisible to increase() (Prometheus never sees the 0→first-value jump). Cards now count series births too, so "Scale down: 1" shows up the moment the first scale-down ever happens.
• The Cluster card explains itself on single-host installs instead of showing misleading "0 managers / 0%" stats — the queue_autoscale_cluster_* gauges only exist in cluster mode, and the card now says so.
• Routes ⇄ Request log toggle no longer reloads the page. The toggle was a plain link (full page load); it is now a Livewire event both sibling cards listen to, so the swap is instant and keeps scroll/filter state.
• Request log renders full-width. Its live-poll wrapper <div> was the card's grid item, so the span 2 on the inner card never reached the grid and the log was squeezed into one column. The wrapper now uses display: contents.
• Live tail is on by default on the request log.
• Scaling actions card matched nothing on real data. The autoscaler's direction label carries up / down (the WorkersScaled action), not scale_up / scale_down; the card now groups by the label instead of filtering on guessed values. Verified against live production series — as are the rest of the autoscale names, including queue_autoscale_sla_breach_ratio and queue_autoscale_sla_predicted_pickup_seconds.
• Environment scope now works on every log-based card. The scope put deployment_environment_name in the Loki stream selector, but backends that index only service_name as a stream label (e.g. otel-lgtm) carry the environment as structured metadata — so a selected environment silently matched nothing and Analytics, the log viewer, unified errors and annotations all returned zero. The environment is now a pipeline label filter ({service_name="…"} | deployment_environment_name="…"), correct whether the backend indexes it as a stream label or not. The analytics Countries/Devices empty states now name the emitter flags (TELEMETRY_ANALYTICS_GEO / TELEMETRY_ANALYTICS_UA) that populate them.
• Rootless traces no longer read as a perpetual "Loading…". An orphaned browser page-view trace (no backend root span) showed "Loading…" forever in the drawer; it is now labelled (the span name, "Unnamed trace" or "Trace not found") since the fetch is already complete by render time.

Fixed

• Queue throughput cards queried queue_metrics_queue_throughput_per_minute; the OTLP collector translates the {jobs}/min unit to a _per_min suffix (verified against live production data), so the Throughput card, the queues table's Jobs/min column and the queue-detail header matched nothing. Now they query queue_metrics_queue_throughput_per_min. All other derived metric names and labels verified correct against live data.

Full Changelog: https://github.com/cboxdk/laravel-telemetry-ui/compare/v0.2.0...v0.2.1

Added

• Queues page (autodetected via queue_metrics_.*) for fleets running cboxdk/laravel-queue-metrics' OpenTelemetry integration (v3.2.0+): backlog by state (pending/scheduled/reserved), per-queue throughput, oldest-job age, busy/idle worker fleet with utilization, and a per-queue table with backlog-trend sparklines. Each queue drills into a queue-detail page: headline numbers, backlog and throughput for that queue, the autoscaler's target-vs-active steering, and the job classes running on it (each linking on to its job-detail page).
• Autoscale page (autodetected via queue_autoscale_.*) for cboxdk/laravel-queue-autoscale v3.11+: target vs active workers, executed scaling actions by direction, SLA health (predicted pickup, queues in breach, breach transitions) and cluster capacity (workers/required/capacity, managers, utilization, recommended hosts).

Changed

• Require cboxdk/laravel-telemetry ^0.2.0 — its first stable release (was ^0.1.0-alpha.3).

Full Changelog: https://github.com/cboxdk/laravel-telemetry-ui/compare/v0.1.0-alpha.6...v0.2.0

Added

• Copy an issue as Markdown for an LLM. The issue page's Actions & context sidebar gains a ⧉ Copy for LLM button that puts a self-contained Markdown brief on the clipboard — exception, message, location, occurrences/users, environment/release/host, the request that hit it, the suspect change, the releases it rode in on and the freshest stacktrace — ready to paste into a model with "help me fix this". No tracker write access required.

Changed

• Optional sidebar groups follow the selected service. Schema detection now scopes to the active service/environment, so a group like Statamic (or Horizon, Reverb, …) only appears when that service emits its metrics — not because some other service in the fleet does. "All services" still detects fleet-wide, and the tenancy lock still bounds what a viewer can see.

Added

• Issue rows go straight to the issue page — no drawer detour — and the page gains a Sentry-style Actions & context sidebar next to the trend: a prefilled "+ Create ticket" button, the tracker tickets that already mention this exception (searched live), the key facts (first/ last seen, users, env, release, host → its page, throw site) and the suspect change event.
• Request log with live tail. The Requests page's Routes card gains a toggle sibling: a request LOG — individual requests, newest first, with time, method+path, status badge, user, client IP and duration. Filter by user id, client IP, path or status class (each user/IP cell is click-to-tail), hit ● Live and the list re-polls every few seconds — production debugging for "what is this user hitting right now?". Every row opens the readable request story in the pane.
• Traces read like requests now — the waterfall is the last resort. Opening a trace (pane or full page) tells the story instead of dumping spans: request facts (method, route, path+query, status, client IP, user, user agent, body sizes), cost totals off the root tallies (queries + query time, N+1 count, cache ops, redis commands, models hydrated, views, CPU time, memory), then one readable section per concern — Database (statements slowest-first, N+1 called out), Upstream calls (URL + status + duration), Cache (hit/miss/write summary + keys), Redis, Dispatched jobs, Views, Storage, captured Headers (request/response) and Logs written during the request (trace-correlated). The raw span waterfall collapses behind a "Raw trace — N spans" toggle. Job/command traces adapt automatically.
• Purpose-built trace filters: status-code class (2xx–5xx), path contains and client IP join status/source/route/name/duration — all URL-synced, all composing scoped TraceQL under the hood.
• Full issue page (Sentry-style show view). Every error group now has its own page (error-detail?group=…, the drawer's "Full page" button): header with events / users / first seen / last seen, an events trend chart with the deploy/change markers drawn on top (release markers, Sentry-style), tag distributions (host, environment, release, service, user — "is it one box, one release, one customer?"), and the full deep-dive: request strip, root-cause hints, source context, stacktrace and recent occurrences. Drawer and page share one per-request-memoized ErrorGroupReport, so the page's four cards cost one set of backend queries.
• Users affected. Error groups now count distinct users (from the enduser.id laravel-telemetry ≥ alpha.18 stamps on exception records): a Users column on the errors list and a "users affected" fact on the group panel.
• Errors list filters. Free-text filter over type/message and a source selector (server / web / full-stack), both URL-synced.
• Scope moved to the top header. Service and environment now sit next to the period picker on every page — Sentry's top-bar pattern — and the sidebar is pure navigation.
• Full-color card borders. The context tiles and root-cause box wear their accent color as a full border instead of a left edge.
• Sentry-style errors list. Every group row now carries its in-period trend sparkline, first seen and last seen, a NEW badge for groups born within 24 hours, and a sort control (events / last seen / first seen). First-seen looks beyond the page period (min 7 days) so it means what it says; counts and trends stay period-scoped.
• Root-cause hints on the error-group panel. The change event (deploy, migration, feature flag, …) closest before the group's first occurrence — within 48h — is named as the suspect ("Deploy v9.1.0 at 14:02 — first seen 18 minutes later"), and the panel shows which releases the sampled occurrences carry, with an "only this release" badge when every one points at a single release.

Fixed

• Annotation callout is anchored to its marker line again. The deploy/ change callout now sits hard against the line it describes — flipping to the line's other side near the chart edge (where deploys cluster) instead of detaching to a fixed corner — with a caret pointing back at the line. The raw axis tooltip no longer stacks on top of it: the series readout is suppressed while the callout is open, and the callout renders above it regardless. Hovering into the callout keeps it open to reach Open trace, so the click-to-pin affordance is gone.

Highlights

The drawer is now a docked properties pane. On wide screens it pushes the page aside instead of covering it — no backdrop, the page stays interactive, and selecting another row swaps the pane's content (links inside the pane still stack with back-navigation). It also opens instantly with a skeleton while the data loads.

Interactive, cluster-aware annotations. Hovering a marker line opens a callout anchored to the line (the pointer can move into it); clicking pins the same callout in the same place. Horizontal rollouts fold into one marker: 200 servers deploying = one line with ×200, the rollout span (first → last host) and the covered hosts.

Exception groups link back to the request. Env / release / host facts (host → its detail page) plus a "Latest occurrence" strip off the trace root: method + route (→ route detail page), status, user and the request trace.

Host detail page. Headline CPU/memory/load/requests, host-scoped system charts, and a "Services on this host" card fed by the services' own Prometheus exporters (mysqld/redis/postgres/node probes as defaults) — plus exporter-less app-side Database and Redis sections built on laravel-telemetry's db.queries (alpha.17) and redis.commands counters, honestly badged observed.

Issues page = errors + tickets. The unified error groups sit above the tracker list, and issue/PR bodies render as formatted markdown through a strict-allowlist sanitizer (external content, never executable).

Context strip names its scope — the exact host (linked) and service the tiles describe, or "all hosts" when that's the truth.

Fixed

• Charts no longer stick at the width they measured mid-render (ECharts instance out of Alpine's reactive proxy + a ResizeObserver).

Requires cboxdk/laravel-telemetry ≥ 0.1.0-alpha.17 for the app-side Database host section. See CHANGELOG.md for the full list.

Highlights

Sentry-style error-group drawer. Click any row on the unified Errors card for the full issue view: message, occurrence stats (count, first/last seen, source), the latest stacktrace with highlighted source context, a prefilled "+ ticket" compose button, and a recent-occurrences table whose trace links stack onto the drawer. Deep-linkable via ?exception=<group>, always inside the tenancy scope lock.

The Errors card now works against real data. It previously searched Tempo for span.exception.group, an attribute laravel-telemetry never puts on spans (backend exceptions are span events + structured Loki records) — so the card was permanently empty in production. It now reads the Loki exception records and merges in browser exception spans, fingerprinted read-side with the backend's own algorithm.

laravel-telemetry v0.1.0-alpha.16 support. Five auto-detected pages — Horizon, Reverb, Feature Flags (Pennant), Storage, Livewire — plus Rate limiting on Requests, real-user Core Web Vitals (p75 LCP/CLS/INP) on Frontend, Duplicate queries (N+1) on Queries, a CPU-profile strip on the trace view, and OTel span links (queue retries) as clickable linked-trace rows.

Mobile-friendly. Off-canvas sidebar behind a topbar hamburger, touch-sized controls, full-width trace drawer, no iOS focus-zoom.

Cache purge annotations. cache_purge (emit via telemetry-ui:annotate) and statamic_cache_purge, matching the events cboxdk/statamic-telemetry emits on every stache/static/glide clear.

Fixed

• Assets are exempt from the dashboard throttle — a rate-limited JS bundle no longer takes every chart down.

See CHANGELOG.md for the full list.

Added

• Analytics page (visit analytics) — a privacy-first traffic dashboard built on the emitter's unsampled analytics.page_view stream: a page-views trend chart (with deploy annotations), unique visitors (the cookieless daily session hash — no cookies, no stored IP), views-per-visit, bounce rate (single-page-view sessions) and average engagement time (from analytics.engagement events), top pages with distinct visitors, and a sources/audience breakdown (referrers, and — when the emitter's geo/User-Agent enrichment is on — countries and devices). Trace/Loki-sourced so it's exact for low-traffic sites and a bounded sample at scale (the eventual answer being a ClickHouse sink behind the same cards). Also: the Loki driver now surfaces per-entry structured metadata, so high-cardinality OTLP log attributes (the visit dimensions) are readable instead of dropped.

• Frontend page (RUM) — a new Monitoring page for real-user browser data: Page performance (navigation timings per page — loads, avg load, TTFB, DOM-interactive, from the document.load spans) and Failed browser requests (fetch/XHR calls that 5xx'd or errored, grouped by URL, each row opening a representative trace where a same-origin failure continues into the backend span that caused it). Trace-sourced (no RUM metric exists), bounded sample.

• Unified errors list (frontend + backend) — a new lead card on the Exceptions page groups every error by exception.group, the Sentry-style fingerprint (class + top in-app frame) that both the backend handler and the browser SDK stamp with the same algorithm. A JS TypeError and a PHP exception that are "the same issue" collapse into one row tagged web/server/full-stack, with an occurrence count and last-seen; clicking a row opens a representative trace (→ waterfall + host context), and "all" jumps to every trace for that fingerprint. Trace-sourced (metrics can't unify — frontend errors exist only as spans), so counts are over a bounded recent sample.

• Frontend / RUM spans in the unified trace — browser spans emitted by cboxdk/laravel-telemetry's frontend proxy (alpha.6/7) now read as first-class frontend rows. They share the backend's service.name, so the per-span server-stamped browser=true attribute is the marker: browser spans get a web badge in the waterfall, document.load shows its RUM timings (TTFB, DOM), and browser fetch spans render their URL + status. Trace search gains a Source filter (frontend/backend) that scopes on span.browser. Because the browser continues the backend's traceparent, a page load, its fetches and the server spans they trigger already nest into one waterfall — end-to-end frontend→backend on open data.

• Dimensional drill-down / filtering (Grafana-style) — every span/resource attribute in the trace properties window (host, user, team, client IP, deployment, method — whatever the app emits) is a click-to-filter link that scopes Traces to { .key = "value" }. Plus a new Hosts page listing every host/server reporting telemetry (request volume, errors, CPU, memory), each row filtering requests to that host.

• Purpose-built detail pages with progressive drill-down — clicking a row opens a dedicated detail page instead of a pre-filtered trace search, à la Nightwatch, for routes, jobs and exceptions. Each shows the entity's own numbers scoped to it, and drills deeper: a route detail goes throughput → latency → exact status codes → its individual traces (→ waterfall + host context). Built on a "hidden page" concept (routable + rendered, out of the sidebar), a scopeMatchers() card hook and per-entity ScopesTo* traits, so the overview cards are reused scoped to one entity — the pattern extends to hosts, queries, etc. cheaply.

• MCP server — php artisan mcp:start telemetry-ui serves metrics, traces, logs and the correlation/analysis tools over the Model Context Protocol, built on the first-party laravel/mcp package, so an agent (Claude Desktop, Cursor, …) can query the stack directly for incident RCA. Six read-only Server\Tools, including trace_context (a trace plus the host/runtime signals around it, flagged against normal). Same read drivers the dashboard uses.

• Remote MCP over HTTP with OAuth + DCR — set TELEMETRY_UI_MCP_WEB=true (and install laravel/passport) to expose the server over HTTP behind auth:api, with the OAuth 2.1 authorization server and Dynamic Client Registration endpoint that laravel/mcp provides — no custom OAuth code. Off by default; Passport stays optional.

• Signal correlation — a trace now shows the host and runtime signals recorded around it (CPU, load, memory, network, process RSS) in a context strip beside the waterfall, scoped by service + host and the trace's time window. This is the thing an app-only monitor can't do: the same Prometheus scrapes system_*/process_* — and node_exporter, mysqld_exporter, … when present — right next to the app. Config-driven and fail-open per signal (telemetry-ui.context.signals); a new headless Analysis\SignalContext is the reusable foundation.

• "What was different" — each context signal also carries its baseline (the typical value for that scope over a longer lookback), so a tile reads "Host CPU 95% (typical 30%)" and flags outliers. Answers the "was the box busted?" question at a glance, without ML — just an honest comparison to normal.

• php artisan telemetry-ui:check — probes each configured connection with its cheapest read and reports OK/FAIL/not-configured; exits non-zero on failure so it doubles as a deploy healthcheck.

• Annotation writing — php artisan telemetry-ui:annotate <marker> emits a marker (deploy, incident, scaling, migration, feature, version — or your own) through the telemetry pipeline into Loki, where it renders as a vertical line on every chart. No local state: the same store the dashboard already reads. cboxdk/laravel-telemetry is now a hard dependency (it provides the write path, and the dashboard instruments its own stack).

• Proactive auto-version annotations — php artisan telemetry-ui:scan-versions (schedule it) detects a laravel_version that's live in the metrics but un-annotated and marks it, so an un-announced deploy still lands on the charts. Stateless: it dedups against the version annotations already in Loki.

• Whole-row click targets on the routes, jobs, facet, slow-query, trace-search, outgoing and exceptions tables — the entire row drills into the matching traces (or opens the trace drawer / matching issues), not just the small link. cmd/ctrl-click opens in a new tab. Outgoing rows filter traces by server.address; exception rows jump to their matching issues (or the scoped error traces when no tracker is configured).

Changed

• Dashboard cards now stream in (lazy on-load) instead of rendering eagerly, so the page shell paints instantly and a slow backend query on one card no longer blocks the whole page; each card loads in its own parallel request.

Fixed

• Chart hover tooltips and drag-to-zoom both work now. The dataZoom brush was kept permanently armed for drag-select, which put every chart in select-mode and suppressed hover tooltips — so "no data on hover" and "zoom broken" were the same bug. Replaced with a raw zrender drag-select (own selection band), so hovering shows values (trigger: 'axis') and dragging realigns the range.
• Linear now surfaces GraphQL errors (auth/permission/query failures, which Linear returns as HTTP 200 with an errors array) as a SourceException instead of silently returning an empty issue list.
• Prometheus/Mimir non-finite values: NaN/+Inf/-Inf (which Prometheus serializes as strings) were cast to a misleading 0.0. They are now dropped so gauges/ratios show a gap instead of a false zero, and the scalar result branch no longer risks a raw TypeError past the SourceException boundary.
• Fleet (sidebar service/environment) cache TTL is now config-wired (telemetry-ui.fleet.ttl / TELEMETRY_UI_FLEET_TTL), matching the other cache TTLs.

Changed

• ConnectionManager::client() is now public so custom drivers registered via extend() can reuse the configured ApiClient (auth, tenancy, cache, retries) instead of building one by hand.

Security

• Finer-grained authorization. The viewTelemetryUi gate is now re-run on Livewire updates (card/drawer actions POST to /livewire/update, which previously skipped it — the gate was only enforced at page load). The gate also receives the page slug, so an app can restrict individual pages (e.g. the PII-heavy Logs/Users) without closing the whole dashboard — denied pages 403 and drop from the sidebar/palette. And a new manageTelemetryUi ability gates write actions (creating tracker issues), checked server-side and hiding the compose UI, so a read-only viewer can't file tickets (it falls back to the view gate, so existing setups are unchanged). See the new authorization doc.
• Backend failures no longer leak the internal endpoint, query string or raw response body to the dashboard. SourceException now carries a generic user-facing message and a separate detail; ApiClient logs the full detail server-side (the dashboard gate may be opened to semi-trusted operators).
• The MCP web transport throws at boot when OAuth is enabled but laravel/passport is absent, instead of registering a half-configured authorization server. mcp.web.middleware documents that auth:api is the only guard on that endpoint.
• MCP tools are bounded (row/series/window/limit caps + a dedicated throttle), tagValues lookups carry a time window + limit, and the annotation writer can no longer crash a command or the scan-versions cron on an emit failure.

Documentation

• New configuration reference (every key
  • env var), signal correlation and custom detail pages guides; README and roadmap rewritten for the correlation/MCP/annotations/drill-down surface.

Verified

• GitHub and Linear issue read and create paths exercised end-to-end against the live APIs. Sentry remains fixture-tested only — see the verification-status table in the issue-trackers docs.