# Capture gallery

<!-- Generated by python3 -m scripts.gallery.render --write from scripts/gallery/data.py and
     docs/images/capture.json. Do not edit. -->

37 plates across 7 stations, captured 16 November 2026 UTC from 100,158 recorded requests with identifiers replaced and content removed.

Every plate is one state photographed on both primary surfaces and split along the diagonal from the bottom-left corner to the top-right. The upper-left half is Instrument Black, the lower-right half is Instrument Clear. Nothing else changes between the halves: the same data, the same scroll, the same open surface.

Values are a static sample from recorded history, not a benchmark. The offline indicator describes the capture session.

## The whole bench

The default view at 1920 px and at phone width; every plate carries the Black and Clear surfaces across one diagonal.

<figure>
<img src="images/overview/dashboard.png" alt="The millivolt dashboard at 1920 pixels: nine global readings above a provider explorer, a traffic timeline and a request log of recorded requests.">
<figcaption><strong>01 · The whole bench</strong> Global totals stay fixed while the explorer scopes the timeline and log; upper-left Black, lower-right Clear.</figcaption>
</figure>

<figure>
<img src="images/overview/mobile.png" alt="The dashboard at phone width, with the KPI band stacked and the explorer reduced to a dimension picker.">
<figcaption><strong>02 · 390 pixels wide</strong> Below 1200 pixels the locked bench becomes a scrolling page and the nine-item rail becomes one picker.</figcaption>
</figure>

## Six readings of one ruler

The plotted presets, the summary tiles, and the window picker that bounds them.

<figure>
<img src="images/charts/overview.png" alt="The Overview preset: five tiles for volume, tokens, cost, health and speed, each with a share line and a sparkline.">
<figcaption><strong>03 · Summary tiles</strong> The overview preset answers what the window contains before any series is drawn.</figcaption>
</figure>

<figure>
<img src="images/charts/traffic.png" alt="Traffic timeline with request and error counts from recorded history.">
<figcaption><strong>04 · Traffic</strong> Requests and errors share one time axis; the totals strip carries the window rate.</figcaption>
</figure>

<figure>
<img src="images/charts/errors.png" alt="Error timeline with affected-request bars and a separate percentage axis.">
<figcaption><strong>05 · Errors</strong> A 429 is a rate limit, not an error; the two series stay separate.</figcaption>
</figure>

<figure>
<img src="images/charts/tokens.png" alt="Token timeline with input, output, reasoning and cached usage.">
<figcaption><strong>06 · Tokens</strong> Cached reads are drawn beside billed input, so a high cache share is visible rather than inferred.</figcaption>
</figure>

<figure>
<img src="images/charts/speed-latency.png" alt="Speed and latency timeline with independent axes and the p95 percentile selected.">
<figcaption><strong>07 · Speed and latency</strong> The percentile control applies to every measured rate in the preset.</figcaption>
</figure>

<figure>
<img src="images/charts/cost.png" alt="Cost timeline with spend and request-count axes.">
<figcaption><strong>08 · Provider-reported cost</strong> Spend is whatever the upstream reported, plotted beside the request volume that produced it.</figcaption>
</figure>

<figure>
<img src="images/charts/window.png" alt="The timeframe panel open over the timeline: a coverage rail, a zoom ladder from one minute to one year, and calendar presets.">
<figcaption><strong>09 · The window picker</strong> Windows wider than the stored history are disabled, so the picker cannot promise history that is not there.</figcaption>
</figure>

## Nine ways into the traffic

The dimension rail, its cards, and one scoped reading.

<figure>
<img src="images/explorer/providers.png" alt="Provider explorer with pseudonymous groups, recorded costs and health signals.">
<figcaption><strong>10 · Providers</strong> The rail counts distinct values; each card shows its share and the signals that matter for that dimension.</figcaption>
</figure>

<figure>
<img src="images/explorer/models.png" alt="Model explorer with pseudonymous model groups and recorded usage metrics.">
<figcaption><strong>11 · Models</strong> The model dimension applies the configured grouping rules before it counts.</figcaption>
</figure>

<figure>
<img src="images/explorer/clients.png" alt="Client explorer with pseudonymous client groups and their recorded traffic.">
<figcaption><strong>12 · Clients</strong> The client dimension answers which caller produced a share of the window.</figcaption>
</figure>

<figure>
<img src="images/explorer/conversations.png" alt="Conversation explorer with main and sub counts and recorded request totals.">
<figcaption><strong>13 · Conversations</strong> Conversation filters select exact leaves; the rail reports main and sub conversations separately.</figcaption>
</figure>

<figure>
<img src="images/explorer/tools.png" alt="Tool explorer with recorded call counts and the traffic behind them.">
<figcaption><strong>14 · Tools</strong> The tool dimension counts calls the relay observed, not calls a client claimed.</figcaption>
</figure>

<figure>
<img src="images/explorer/time.png" alt="Time explorer with recorded buckets and their request totals.">
<figcaption><strong>15 · Time</strong> Time is a grouping like any other, so a bucket can be filtered and compared with the rest of the window.</figcaption>
</figure>

<figure>
<img src="images/explorer/status.png" alt="Status explorer with final HTTP classes and their recorded counts.">
<figcaption><strong>16 · Status</strong> The status dimension separates what the client received from what the upstream first returned.</figcaption>
</figure>

<figure>
<img src="images/explorer/errors.png" alt="Error explorer with canonical error classes and their affected request counts.">
<figcaption><strong>17 · Error class</strong> The error dimension groups by the proxy&#x27;s canonical class, not by provider wording.</figcaption>
</figure>

<figure>
<img src="images/explorer/keys.png" alt="API key explorer with pseudonymous key groups and their recorded traffic.">
<figcaption><strong>18 · API keys</strong> Keys are hashed in the copy, so a group is stable without exposing a credential.</figcaption>
</figure>

<figure>
<img src="images/explorer/scoped.png" alt="The explorer filtered to one provider, with the scope chip above the rail and the selected card marked.">
<figcaption><strong>19 · A scoped reading</strong> A card toggles an exact filter; the chip removes it, and the timeline and log follow the same scope.</figcaption>
</figure>

## Every crossing

The request log, an absorbed retry, and the drawer that explains a single crossing.

<figure>
<img src="images/requests/table.png" alt="Request table with recorded statuses, token usage, latency, throughput and cost.">
<figcaption><strong>20 · The request log</strong> Each row is one crossing; the columns are the facts a later review needs first.</figcaption>
</figure>

<figure>
<img src="images/requests/retries.png" alt="The request log with one row expanded into its absorbed retry attempts, each carrying its own status and timing.">
<figcaption><strong>21 · An absorbed retry</strong> A retry the client never saw is still a recorded upstream attempt.</figcaption>
</figure>

<figure>
<img src="images/requests/details.png" alt="Request drawer showing pseudonymous identity, the summary readout and recorded request parameters.">
<figcaption><strong>22 · The request drawer</strong> The drawer opens on the record, not on a reconstruction of it.</figcaption>
</figure>

<figure>
<img src="images/requests/performance.png" alt="The same request drawer showing token and performance measurements.">
<figcaption><strong>23 · Tokens and performance</strong> Usage is grouped by the question it answers, so timing and tokens do not compete in one list.</figcaption>
</figure>

<figure>
<img src="images/requests/diagnostics.png" alt="The request drawer&#x27;s diagnostics group, with queue waits, rate-limit handling and absorbed attempts.">
<figcaption><strong>24 · Diagnostics</strong> The diagnostics group is where a transparent retry stops being invisible.</figcaption>
</figure>

## The operator plane

The five policy sections and the composer behind them.

<figure>
<img src="images/console/overview.png" alt="The operator console overview with the live line, section rail and the change ledger.">
<figcaption><strong>25 · Overview</strong> The overview is the register of what is in force, not a second settings form.</figcaption>
</figure>

<figure>
<img src="images/console/limits.png" alt="The limits section with no policy set and an empty rule register.">
<figcaption><strong>26 · Limits</strong> An empty policy surface states the state instead of hiding it.</figcaption>
</figure>

<figure>
<img src="images/console/limits-composer.png" alt="The limit composer open with its scope chips, budgets, window, schedule and live preview.">
<figcaption><strong>27 · The limit composer</strong> A draft is local until it is committed; the composer previews the effect before a write.</figcaption>
</figure>

<figure>
<img src="images/console/pause.png" alt="The pause section before any hold exists.">
<figcaption><strong>28 · Pause</strong> A hold parks new work for a scope without changing a route.</figcaption>
</figure>

<figure>
<img src="images/console/capture.png" alt="The capture section with no session running.">
<figcaption><strong>29 · Capture</strong> A capture session retains bodies for a bounded scope and time; the panel states both.</figcaption>
</figure>

<figure>
<img src="images/console/export.png" alt="The export section previewing the whole-history record count and the download receipt.">
<figcaption><strong>30 · Export</strong> The receipt counts what the chosen mode will write before the file is produced.</figcaption>
</figure>

<figure>
<img src="images/console/system.png" alt="The system section with restart available and the four restart steps.">
<figcaption><strong>31 · System</strong> A restart is a state machine with named steps, not a button with a spinner.</figcaption>
</figure>

## The patch panel

Five categories of the public example configuration.

<figure>
<img src="images/settings/dashboard.png" alt="Settings showing the public dashboard configuration fields.">
<figcaption><strong>32 · Dashboard</strong> Every field is generated from the schema, including its help and the restart badge.</figcaption>
</figure>

<figure>
<img src="images/settings/queue.png" alt="Queue and retry Settings with their public values and field descriptions.">
<figcaption><strong>33 · Queue and retry</strong> Retry policy, error classes and pacing sit together because they act on the same wait.</figcaption>
</figure>

<figure>
<img src="images/settings/storage.png" alt="Storage Settings with the bounded writer, queue, persistence and deletion controls.">
<figcaption><strong>34 · Storage</strong> Deletion is scoped and previewed; the storage facts come from the running store.</figcaption>
</figure>

<figure>
<img src="images/settings/providers.png" alt="Provider Settings showing the public Cursor compatibility profile.">
<figcaption><strong>35 · Providers</strong> Provider data is a field map in configuration, not a branch in the router.</figcaption>
</figure>

<figure>
<img src="images/settings/backup.png" alt="Backup Settings with the download and restore rows and their scope choices.">
<figcaption><strong>36 · Backup and restore</strong> A restore is inspected before it is applied; the sheet says what will be replaced.</figcaption>
</figure>

## One bench, two surfaces

The shared theme tray on Black and on Clear.

<figure>
<img src="images/theme/tray.png" alt="The shared theme tray in both surfaces: two comparison panes, a surface switch and eight palettes.">
<figcaption><strong>37 · The theme tray</strong> Palette, surface and order come from the same registry the site publishes; the seam is the surface switch made visible.</figcaption>
</figure>
