Collapsible Thinking Trace
The "thought for 5s" panel: a native details/summary outside, individually expandable steps inside — buttons rather than nested disclosures, which are announced as a disclosure within a disclosure.
What one run cost, per step, so an expensive answer is attributable rather than a monthly surprise.
components/agent-cost-breakdown.tsxnpx hoverlab add agent-cost-breakdown
Or over MCP, from your editor's agent — no account needed.
Free to read, copy and install, for personal and non-commercial projects. Shipping it in client work or a paid product needs Pro ($79 once). The source lands in your repo and stops being ours — no attribution, nothing to upgrade.
Start a page with this section — add more, order them, and leave with the page source.
Per step, not per month. A bill that only arrives in aggregate cannot be acted on — the useful version says which step was expensive and whether it needed to be.
Rendered live in your current theme — this is the same component whose source is below, not a screenshot of it.
'use client'
/**
* <AgentCostBreakdown> — What one run cost, per step, so an expensive answer is attributable rather than a monthly surprise.
*
* A bill that arrives only in aggregate cannot be acted on. The layout
* problem is attribution: a run costs one number, and the useful version is
* that number split by step, so an expensive answer is traceable to the step
* that made it expensive.
*
* The obvious wrong answer is a token count. Tokens are a unit the person
* paying does not think in, and two models price them differently — the row
* that matters is money, with tokens as the detail behind it.
*
* The reasoning row is tone `warning` while being the correct and necessary
* majority of the cost. That is deliberate: warning here means "this is
* where the money is", not "this is broken", and the detail says the useful
* thing — three steps would have reached the same answer. A cost panel that
* flags nothing teaches nothing.
*
* The retry inside the tool-calls row is named rather than hidden, because a
* retried call is billed like any other call and a reader reconciling this
* against an invoice will otherwise be one call short.
*
* Rows are buttons with `aria-pressed`, since drilling into a step is the
* obvious next action, and `status` carries the currency figure separately
* from `tone` so the amounts are legible without colour. The demo selects
* the first row rather than the expensive one, because the panel reads top
* to bottom as a sequence.
*/
import * as React from 'react'
type Tone = 'neutral' | 'positive' | 'warning' | 'critical'
export interface AgentCostBreakdownRow {
id: string
label: string
detail?: string
tone?: Tone
/** What the badge says. Falls back to the tone name when absent. */
status?: string
}
export interface AgentCostBreakdownProps {
heading?: string
intro?: string
rows?: AgentCostBreakdownRow[]
className?: string
}
/*
Tones as complete utility classes, never assembled from fragments.
Tailwind scans source text, so `text-${tone}-foreground` produces no
class at all — the same failure as an undefined token, and just as
invisible in review.
*/
const TONE_CLASS: Record<Tone, string> = {
neutral: 'bg-muted text-muted-foreground',
positive: 'bg-primary/10 text-primary',
warning: 'bg-accent text-accent-foreground',
critical: 'bg-destructive/10 text-destructive',
}
const ROWS: AgentCostBreakdownRow[] = [
{ id: "row-1", label: "Retrieval — 12 passages", detail: "Cheap and worth it. Cutting this is what makes the model guess.", tone: "positive", status: "$0.004" },
{ id: "row-2", label: "Reasoning — 4 steps", detail: "The bulk of it. Three steps would have reached the same answer.", tone: "warning", status: "$0.112" },
{ id: "row-3", label: "Tool calls — 6 invocations", detail: "One retry included, which is billed like any other call.", tone: "neutral", status: "$0.019" },
{ id: "row-4", label: "Final answer", detail: "Output tokens only. The cheapest part of almost every run.", tone: "positive", status: "$0.007" },
]
export function AgentCostBreakdown({
heading = "What this run cost",
intro = "Per step, not per month. A bill that only arrives in aggregate cannot be acted on — the useful version says which step was expensive and whether it needed to be.",
rows = ROWS,
className,
}: AgentCostBreakdownProps) {
/*
Per-instance prefix for every id this block emits.
The literals these replaced were a latent duplicate the moment the
block appeared twice on one document, and `aria-labelledby` on a
duplicated id resolves to the first match -- so the second copy was
labelled by the first copy's heading. Client component, so `useId` is
the right tool.
*/
const uid = React.useId()
const [selected, setSelected] = React.useState<string | null>(rows[0]?.id ?? null)
return (
<section
aria-labelledby={`${uid}-agent-cost-breakdown-heading`}
className={`w-full bg-background px-6 py-16 ${className ?? ''}`}
>
<div className="mx-auto max-w-3xl">
<h2
id={`${uid}-agent-cost-breakdown-heading`}
className="text-2xl font-semibold tracking-tight text-foreground"
>
{heading}
</h2>
<p className="mt-2 text-sm text-muted-foreground">{intro}</p>
<ul className="mt-8 divide-y divide-border overflow-hidden rounded-xl border border-border bg-card">
{rows.map((row) => {
const isSelected = row.id === selected
return (
<li key={row.id}>
{/*
A button, not a div with onClick. The row is operable, so
it has to be reachable by keyboard and announce its
selected state — aria-pressed is what carries that.
*/}
<button
type="button"
aria-pressed={isSelected}
onClick={() => setSelected(row.id)}
className={`flex w-full items-center justify-between gap-4 px-5 py-4 text-start transition-colors hover:bg-muted/60 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset ${
isSelected ? 'bg-muted/40' : ''
}`}
>
<span className="min-w-0">
<span className="block truncate text-sm font-medium text-foreground">
{row.label}
</span>
{row.detail ? (
<span className="mt-0.5 block text-sm text-muted-foreground">
{row.detail}
</span>
) : null}
</span>
<span
className={`shrink-0 rounded-full px-2.5 py-1 text-xs font-medium ${
TONE_CLASS[row.tone ?? 'neutral']
}`}
>
{row.status ?? row.tone ?? 'neutral'}
</span>
</button>
</li>
)
})}
</ul>
</div>
</section>
)
}
bg-card, text-muted-foreground) — it inherits your theme instead of overriding it.Drop it at components/agent-cost-breakdown.tsx and import it where you need the section:
import { AgentCostBreakdown } from '@/components/agent-cost-breakdown'1 of this block’s props are simple enough to drive from here. Change them and the block below re-renders — it is the same component whose source is above, not a mock of it. Everything else it accepts is in the table underneath.
Read out of the component’s own type and signature, so this cannot drift from the source below. Every prop has a default — the component renders standalone before you pass it anything.
| Prop | Type | Default |
|---|---|---|
heading | string | "What this run cost" |
intro | string | — |
rows | AgentCostBreakdownRow[] | ROWS |
className | string | — |
The same block rendered once to markup, wrapped as a file your framework compiles. Tailwind classes are framework-agnostic, so the design transfers intact — the behaviour does not.
This block is interactive. The markup below is its initial state with the event handlers stripped — you will need to re-wire the behaviour in your framework.
<!--
Agent Cost Breakdown — markup from the Hoverlab catalog.
This is the block rendered once to HTML and wrapped as a component
file. It is not a port of the React source: the Tailwind classes carry
the design, which is the part that took the work, and they are the same
in every framework.
This block is interactive in React and the handlers are NOT here.
Buttons, toggles and menus render in their initial state and do
nothing until you wire them up.
-->
<section aria-labelledby="_R_0_-agent-cost-breakdown-heading" class="w-full bg-background px-6 py-16 ">
<div class="mx-auto max-w-3xl">
<h2 id="_R_0_-agent-cost-breakdown-heading" class="text-2xl font-semibold tracking-tight text-foreground">What this run cost</h2>
<p class="mt-2 text-sm text-muted-foreground">Per step, not per month. A bill that only arrives in aggregate cannot be acted on — the useful version says which step was expensive and whether it needed to be.</p>
<ul class="mt-8 divide-y divide-border overflow-hidden rounded-xl border border-border bg-card">
<li>
<button type="button" aria-pressed="true" class="flex w-full items-center justify-between gap-4 px-5 py-4 text-start transition-colors hover:bg-muted/60 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset bg-muted/40">
<span class="min-w-0">
<span class="block truncate text-sm font-medium text-foreground">Retrieval — 12 passages</span>
<span class="mt-0.5 block text-sm text-muted-foreground">Cheap and worth it. Cutting this is what makes the model guess.</span>
</span>
<span class="shrink-0 rounded-full px-2.5 py-1 text-xs font-medium bg-primary/10 text-primary">$0.004</span>
</button>
</li>
<li>
<button type="button" aria-pressed="false" class="flex w-full items-center justify-between gap-4 px-5 py-4 text-start transition-colors hover:bg-muted/60 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset ">
<span class="min-w-0">
<span class="block truncate text-sm font-medium text-foreground">Reasoning — 4 steps</span>
<span class="mt-0.5 block text-sm text-muted-foreground">The bulk of it. Three steps would have reached the same answer.</span>
</span>
<span class="shrink-0 rounded-full px-2.5 py-1 text-xs font-medium bg-accent text-accent-foreground">$0.112</span>
</button>
</li>
<li>
<button type="button" aria-pressed="false" class="flex w-full items-center justify-between gap-4 px-5 py-4 text-start transition-colors hover:bg-muted/60 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset ">
<span class="min-w-0">
<span class="block truncate text-sm font-medium text-foreground">Tool calls — 6 invocations</span>
<span class="mt-0.5 block text-sm text-muted-foreground">One retry included, which is billed like any other call.</span>
</span>
<span class="shrink-0 rounded-full px-2.5 py-1 text-xs font-medium bg-muted text-muted-foreground">$0.019</span>
</button>
</li>
<li>
<button type="button" aria-pressed="false" class="flex w-full items-center justify-between gap-4 px-5 py-4 text-start transition-colors hover:bg-muted/60 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset ">
<span class="min-w-0">
<span class="block truncate text-sm font-medium text-foreground">Final answer</span>
<span class="mt-0.5 block text-sm text-muted-foreground">Output tokens only. The cheapest part of almost every run.</span>
</span>
<span class="shrink-0 rounded-full px-2.5 py-1 text-xs font-medium bg-primary/10 text-primary">$0.007</span>
</button>
</li>
</ul>
</div>
</section>
What each framework gets across the whole catalog — effects convert properly; this rung is markup.
The component, its props, the design tokens it expects and the command that installs it — as one prompt. Paste it into Claude, Cursor, v0 or ChatGPT and what they build around it will match the rest of the catalog instead of inventing its own system.
Want the whole screen instead of this one section? Open a page and copy it entire.
All of them free to read, copy and install — no account, no locked tiles, no watermarked preview. The whole catalog is open, and so are the API and the CLI.
Browse Agent ReasoningCopying the code is free. Putting it in client work or a paid product is what Pro is for — the licence, not the access.
The "thought for 5s" panel: a native details/summary outside, individually expandable steps inside — buttons rather than nested disclosures, which are announced as a disclosure within a disclosure.
Dense tool invocations that open into their arguments and result, with status carried by icon and word rather than colour, and errors marked as alerts only once the user opens them.
Run status as rows — done, running, failed, queued — with native progress elements, per-task retry buttons that name their task, and progress announced from one small status region instead of every row.
The gap before the first token: a shimmer grid, a phase label that advances as the wait lengthens, and a running counter — which is the only honest answer to "is it stuck". The timer is hidden from the live region on purpose.
The trace after something broke: retries counted in the open with what each one cost, a step that was silently worked around drawn in amber rather than green, and the raw error kept apart from the agent’s reading of it.
What the agent tried, what failed, and what it did about it — the run history a user needs before they trust a second attempt.