From dbd1c0cd6fa813327b0e243623a1bade1a0ea6ec Mon Sep 17 00:00:00 2001
From: Steven Enamakel <31011319+senamakel@users.noreply.github.com>
Date: Thu, 25 Jun 2026 15:04:39 -0700
Subject: [PATCH] feat(agent): track session token costs (#4136)

---
 .../chat/ComposerTokenStats.test.tsx          | 146 ++++++++++
 .../components/chat/ComposerTokenStats.tsx    | 263 ++++++++++++++----
 app/src/lib/i18n/ar.ts                        |  16 ++
 app/src/lib/i18n/bn.ts                        |  16 ++
 app/src/lib/i18n/de.ts                        |  18 ++
 app/src/lib/i18n/en.ts                        |  17 ++
 app/src/lib/i18n/es.ts                        |  18 ++
 app/src/lib/i18n/fr.ts                        |  18 ++
 app/src/lib/i18n/hi.ts                        |  16 ++
 app/src/lib/i18n/id.ts                        |  18 ++
 app/src/lib/i18n/it.ts                        |  18 ++
 app/src/lib/i18n/ko.ts                        |  16 ++
 app/src/lib/i18n/pl.ts                        |  17 ++
 app/src/lib/i18n/pt.ts                        |  18 ++
 app/src/lib/i18n/ru.ts                        |  17 ++
 app/src/lib/i18n/zh-CN.ts                     |  16 ++
 app/src/pages/Conversations.tsx               |  37 ++-
 .../components/PlanReviewCard.tsx             |   2 +-
 app/src/providers/ChatRuntimeProvider.tsx     |  58 +++-
 app/src/services/api/threadUsageApi.ts        |  92 ++++++
 app/src/services/chatService.ts               |  38 ++-
 app/src/store/chatRuntimeSlice.test.ts        | 164 +++++++++++
 app/src/store/chatRuntimeSlice.ts             | 205 ++++++++++++--
 src/core/socketio.rs                          |  38 +++
 src/openhuman/agent/cost.rs                   |  75 ++---
 src/openhuman/agent/harness/mod.rs            |   1 +
 .../agent/harness/session/builder/setters.rs  |   1 +
 .../agent/harness/session/runtime.rs          |   9 +
 .../agent/harness/session/transcript.rs       | 202 +++++++++++++-
 .../agent/harness/session/transcript_tests.rs | 155 +++++++++++
 .../agent/harness/session/turn/core.rs        |  59 +++-
 .../harness/session/turn_engine_adapter.rs    |  11 +-
 src/openhuman/agent/harness/session/types.rs  |   6 +
 .../agent/harness/subagent_runner/mod.rs      |   2 +-
 .../harness/subagent_runner/ops/observer.rs   |  25 +-
 .../harness/subagent_runner/ops/runner.rs     |  19 +-
 .../agent/harness/subagent_runner/types.rs    |  18 ++
 .../agent/harness/turn_subagent_usage.rs      | 174 ++++++++++++
 .../agent/task_dispatcher/executor.rs         |   3 +
 .../tools/spawn_subagent.rs                   |   1 +
 src/openhuman/channels/proactive.rs           |   1 +
 .../channels/providers/presentation.rs        |  37 ++-
 src/openhuman/channels/providers/web/ops.rs   |   2 +
 .../channels/providers/web/run_task.rs        |   5 +
 src/openhuman/channels/providers/web/types.rs |   4 +
 src/openhuman/inference/model_context.rs      |   7 +-
 src/openhuman/threads/ops.rs                  | 158 +++++++++++
 src/openhuman/threads/schemas.rs              |  29 ++
 tests/json_rpc_e2e.rs                         | 124 +++++++++
 49 files changed, 2255 insertions(+), 155 deletions(-)
 create mode 100644 app/src/components/chat/ComposerTokenStats.test.tsx
 create mode 100644 app/src/services/api/threadUsageApi.ts
 create mode 100644 src/openhuman/agent/harness/turn_subagent_usage.rs

diff --git a/app/src/components/chat/ComposerTokenStats.test.tsx b/app/src/components/chat/ComposerTokenStats.test.tsx
new file mode 100644
index 000000000..1550c9100
--- /dev/null
+++ b/app/src/components/chat/ComposerTokenStats.test.tsx
@@ -0,0 +1,146 @@
+import { configureStore } from '@reduxjs/toolkit';
+import { fireEvent, render, screen, within } from '@testing-library/react';
+import { Provider } from 'react-redux';
+import { describe, expect, it, vi } from 'vitest';
+
+import chatRuntimeReducer, { recordChatTurnUsage } from '../../store/chatRuntimeSlice';
+import ComposerTokenStats from './ComposerTokenStats';
+
+vi.mock('../../lib/i18n/I18nContext', () => ({ useT: () => ({ t: (k: string) => k }) }));
+
+function renderWithUsage(
+  payloads: Array<Parameters<typeof recordChatTurnUsage>[0]>,
+  props?: { model?: string | null; threadId?: string | null }
+) {
+  const store = configureStore({ reducer: { chatRuntime: chatRuntimeReducer } });
+  for (const p of payloads) store.dispatch(recordChatTurnUsage(p));
+  return render(
+    <Provider store={store}>
+      <ComposerTokenStats {...props} />
+    </Provider>
+  );
+}
+
+const oneTurn = [
+  {
+    inputTokens: 1200,
+    outputTokens: 300,
+    cachedTokens: 50,
+    costUsd: 0.0123,
+    contextWindow: 200_000,
+  },
+];
+
+describe('<ComposerTokenStats />', () => {
+  it('renders nothing before any turn when no model is known', () => {
+    const { container } = renderWithUsage([]);
+    expect(container).toBeEmptyDOMElement();
+  });
+
+  it('renders the clickable context row before the first turn when a model is known', () => {
+    renderWithUsage([], { model: 'reasoning-v1' });
+    const row = screen.getByRole('button');
+    expect(row).toHaveTextContent('token.ctxLabel');
+  });
+
+  it('keeps the inline row minimal: context window and cost only', () => {
+    renderWithUsage(oneTurn, { model: 'reasoning-v1' });
+    const row = screen.getByRole('button');
+    // Context window uses the real reported window (200K), not just a default.
+    expect(row).toHaveTextContent('token.ctxLabel');
+    expect(row).toHaveTextContent('200K');
+    // Cost inline.
+    expect(row).toHaveTextContent('$0.012');
+    // Tokens (in/out) are NOT inline — they live in the popover.
+    expect(row).not.toHaveTextContent('token.inLabel');
+    expect(row).not.toHaveTextContent('token.outLabel');
+    // The model id is NOT inline (it lives in the popover).
+    expect(row).not.toHaveTextContent('reasoning-v1');
+  });
+
+  it('toggles the breakdown on click and shows explicit labelled rows + tooltips + model', () => {
+    renderWithUsage(oneTurn, { model: 'reasoning-v1' });
+    expect(screen.queryByTestId('composer-token-breakdown')).not.toBeInTheDocument();
+
+    fireEvent.click(screen.getByRole('button'));
+    const bd = screen.getByTestId('composer-token-breakdown');
+    // Explicit, spelled-out labels with explanatory tooltips.
+    expect(within(bd).getByText('token.popInput')).toHaveAttribute('title', 'token.tipInput');
+    expect(within(bd).getByText('token.popOutput')).toHaveAttribute('title', 'token.tipOutput');
+    expect(within(bd).getByText('token.popCacheHit')).toHaveAttribute('title', 'token.tipCacheHit');
+    // Cache hit shows a hit-rate percentage: 50 cached / 1200 input ≈ 4%.
+    expect(within(bd).getByText(/50 \(4%\)/)).toBeInTheDocument();
+    // Model id surfaced inside the popover.
+    expect(within(bd).getByText('reasoning-v1')).toBeInTheDocument();
+
+    // Clicking again closes it.
+    fireEvent.click(screen.getByRole('button'));
+    expect(screen.queryByTestId('composer-token-breakdown')).not.toBeInTheDocument();
+  });
+
+  it('highlights the context segment while the breakdown is open', () => {
+    renderWithUsage(oneTurn);
+    const ctx = screen.getByText(/token\.ctxLabel/);
+    expect(ctx.className).not.toMatch(/bg-primary/);
+    fireEvent.click(screen.getByRole('button'));
+    expect(ctx.className).toMatch(/bg-primary/);
+  });
+
+  it('closes on Escape and on an outside click', () => {
+    renderWithUsage(oneTurn);
+    fireEvent.click(screen.getByRole('button'));
+    expect(screen.getByTestId('composer-token-breakdown')).toBeInTheDocument();
+    fireEvent.keyDown(document, { key: 'Escape' });
+    expect(screen.queryByTestId('composer-token-breakdown')).not.toBeInTheDocument();
+
+    fireEvent.click(screen.getByRole('button'));
+    expect(screen.getByTestId('composer-token-breakdown')).toBeInTheDocument();
+    fireEvent.mouseDown(document.body);
+    expect(screen.queryByTestId('composer-token-breakdown')).not.toBeInTheDocument();
+  });
+
+  it('breaks down spend per agent: orchestrator (derived) + sub-agents', () => {
+    renderWithUsage([
+      {
+        inputTokens: 500,
+        outputTokens: 100,
+        costUsd: 0.01,
+        subAgents: [{ agentId: 'researcher', inputTokens: 200, outputTokens: 40, costUsd: 0.004 }],
+      },
+    ]);
+    fireEvent.click(screen.getByRole('button'));
+    const bd = screen.getByTestId('composer-token-breakdown');
+    // Orchestrator = totals − sub-agents: tokens 600−240=360, cost 0.01−0.004=0.006.
+    expect(within(bd).getByText('token.orchestrator')).toBeInTheDocument();
+    expect(within(bd).getByText(/360/)).toBeInTheDocument();
+    expect(within(bd).getByText(/\$0\.006/)).toBeInTheDocument();
+    // Sub-agent row: 200 + 40 = 240 combined tokens, its own cost.
+    expect(within(bd).getByText('researcher')).toBeInTheDocument();
+    expect(within(bd).getByText(/240/)).toBeInTheDocument();
+    expect(within(bd).getByText(/\$0\.004/)).toBeInTheDocument();
+  });
+
+  it('reads the active thread bucket when a threadId is provided', () => {
+    // Two threads with different usage; the footer must reflect the selected one.
+    renderWithUsage(
+      [
+        { inputTokens: 999, outputTokens: 999, costUsd: 0.5, threadId: 'thr-other' },
+        { inputTokens: 1200, outputTokens: 300, costUsd: 0.0123, threadId: 'thr-active' },
+      ],
+      { threadId: 'thr-active' }
+    );
+    fireEvent.click(screen.getByRole('button'));
+    const bd = screen.getByTestId('composer-token-breakdown');
+    // Active thread's input tokens (1.2K), not the other thread's 999.
+    expect(within(bd).getByText('1.2K')).toBeInTheDocument();
+    expect(within(bd).queryByText('999')).not.toBeInTheDocument();
+  });
+
+  it('shows the orchestrator row and a no-sub-agents note when none ran', () => {
+    renderWithUsage([{ inputTokens: 100, outputTokens: 20, costUsd: 0.001 }]);
+    fireEvent.click(screen.getByRole('button'));
+    const bd = screen.getByTestId('composer-token-breakdown');
+    expect(within(bd).getByText('token.orchestrator')).toBeInTheDocument();
+    expect(within(bd).getByText('token.noSubAgents')).toBeInTheDocument();
+  });
+});
diff --git a/app/src/components/chat/ComposerTokenStats.tsx b/app/src/components/chat/ComposerTokenStats.tsx
index 8fb99ae83..4ecf15668 100644
--- a/app/src/components/chat/ComposerTokenStats.tsx
+++ b/app/src/components/chat/ComposerTokenStats.tsx
@@ -1,6 +1,11 @@
-import { useT } from '../../lib/i18n/I18nContext';
-import { useAppSelector } from '../../store/hooks';
+import { useEffect, useRef, useState } from 'react';
 
+import { useT } from '../../lib/i18n/I18nContext';
+import { emptySessionTokenUsage, type SubAgentUsage } from '../../store/chatRuntimeSlice';
+import { useAppSelector } from '../../store/hooks';
+import Tooltip from '../ui/Tooltip';
+
+/** Fallback context window when the core hasn't reported a real one yet. */
 const DEFAULT_CONTEXT_WINDOW = 200_000;
 
 function fmt(n: number): string {
@@ -10,75 +15,150 @@ function fmt(n: number): string {
   return `${(n / 1_000_000).toFixed(1)}M`;
 }
 
+/** Format a USD cost compactly: sub-cent values keep more precision. */
+function fmtUsd(n: number): string {
+  if (!Number.isFinite(n) || n <= 0) return '$0.00';
+  if (n < 0.01) return `$${n.toFixed(4)}`;
+  if (n < 1) return `$${n.toFixed(3)}`;
+  return `$${n.toFixed(2)}`;
+}
+
 function ok(n: number): boolean {
   return Number.isFinite(n) && n > 0;
 }
 
-function dot() {
-  return <span className="text-content-faint dark:text-neutral-700">·</span>;
+/**
+ * One labelled row in the hover breakdown. The label carries a `title` tooltip
+ * explaining the metric, hinted with a dotted underline + help cursor.
+ */
+function UsageRow({ label, tip, value }: { label: string; tip: string; value: string }) {
+  return (
+    <div className="flex justify-between gap-3">
+      <dt
+        title={tip}
+        className="cursor-help underline decoration-dotted decoration-content-faint underline-offset-2">
+        {label}
+      </dt>
+      <dd className="font-mono text-content">{value}</dd>
+    </div>
+  );
+}
+
+/** One agent's line in the per-agent breakdown: combined tokens · cost · runs. */
+function AgentLine({
+  name,
+  tokens,
+  costUsd,
+  runs,
+}: {
+  name: string;
+  tokens: number;
+  costUsd: number;
+  runs: number;
+}) {
+  return (
+    <li className="flex items-center justify-between gap-3">
+      <span className="truncate text-content-secondary" title={name}>
+        {name}
+      </span>
+      <span className="whitespace-nowrap font-mono text-content">
+        {fmt(tokens)} · {fmtUsd(costUsd)} · {runs}×
+      </span>
+    </li>
+  );
 }
 
 interface ComposerTokenStatsProps {
-  /** Resolved model id, shown as the leading stat when present. */
+  /** Resolved model id, surfaced inside the breakdown popover. */
   model?: string | null;
+  /**
+   * Active thread id. When set, the footer shows that thread's usage bucket
+   * (seeded from persisted transcripts + live turns); otherwise it falls back
+   * to the global app-session aggregate.
+   */
+  threadId?: string | null;
 }
 
-export default function ComposerTokenStats({ model }: ComposerTokenStatsProps = {}) {
+const EMPTY_USAGE = emptySessionTokenUsage();
+
+export default function ComposerTokenStats({ model, threadId }: ComposerTokenStatsProps = {}) {
   const { t } = useT();
-  const usage = useAppSelector(state => state.chatRuntime.sessionTokenUsage);
+  const usage = useAppSelector(state =>
+    threadId
+      ? (state.chatRuntime.usageByThread[threadId] ?? EMPTY_USAGE)
+      : state.chatRuntime.sessionTokenUsage
+  );
+  const [open, setOpen] = useState(false);
+  const rootRef = useRef<HTMLDivElement>(null);
+
+  // The breakdown is click-toggled (not hover). Dismiss on an outside click or
+  // Escape so it behaves like a popover rather than a sticky panel.
+  useEffect(() => {
+    if (!open) return;
+    const onPointerDown = (e: MouseEvent) => {
+      if (rootRef.current && !rootRef.current.contains(e.target as Node)) setOpen(false);
+    };
+    const onKey = (e: KeyboardEvent) => {
+      if (e.key === 'Escape') setOpen(false);
+    };
+    document.addEventListener('mousedown', onPointerDown);
+    document.addEventListener('keydown', onKey);
+    return () => {
+      document.removeEventListener('mousedown', onPointerDown);
+      document.removeEventListener('keydown', onKey);
+    };
+  }, [open]);
 
   const inTok = usage.inputTokens || 0;
   const outTok = usage.outputTokens || 0;
+  const cachedTok = usage.cachedTokens || 0;
   const turns = usage.turns || 0;
-  const lastIn = usage.lastTurnInputTokens || 0;
-  const lastOut = usage.lastTurnOutputTokens || 0;
+  const costUsd = usage.costUsd || 0;
+  const subAgents: SubAgentUsage[] = Object.values(usage.subAgents ?? {});
 
-  // Still render when only the model is known (no turns yet) so the resolved
-  // model stays visible in the composer footer.
+  // Render as soon as a model is resolved (even before the first turn) so the
+  // clickable usage row is always present in the footer; the context segment
+  // shows 0% until the first turn reports usage.
   if (turns === 0 && !model) return null;
 
-  const showIn = ok(inTok);
-  const showOut = ok(outTok);
-  const contextUsed = lastIn + lastOut;
-  const showCtx = ok(contextUsed);
-  const contextPct = showCtx
-    ? Math.min(100, Math.round((contextUsed / DEFAULT_CONTEXT_WINDOW) * 100))
-    : 0;
+  const contextWindow = ok(usage.contextWindow) ? usage.contextWindow : DEFAULT_CONTEXT_WINDOW;
+  const contextUsed = usage.lastTurnContextUsed || 0;
+  const contextPct = Math.min(100, Math.round((contextUsed / contextWindow) * 100));
+
+  const showCost = ok(costUsd);
+
+  // Orchestrator (the parent/main agent) spend = session totals minus everything
+  // attributed to sub-agents. Derived here so no extra backend data is needed.
+  const subTotals = subAgents.reduce(
+    (acc, s) => ({
+      tokens: acc.tokens + s.inputTokens + s.outputTokens,
+      cost: acc.cost + s.costUsd,
+    }),
+    { tokens: 0, cost: 0 }
+  );
+  const orchestratorTokens = Math.max(0, inTok + outTok - subTotals.tokens);
+  const orchestratorCost = Math.max(0, costUsd - subTotals.cost);
 
   const parts: React.ReactNode[] = [];
 
-  if (model) {
+  // Inline footer is intentionally minimal: just context usage · cost. The full
+  // token breakdown (in/out/cached, per-agent) lives in the click-open popover.
+  // The context counter is always shown (primary metric + toggle hint) and is
+  // highlighted while the breakdown is open.
+  parts.push(
+    <span
+      key="ctx"
+      title={t('token.contextWindow')}
+      className={
+        open ? 'rounded bg-primary-500/15 px-1 text-primary-700 dark:text-primary-300' : undefined
+      }>
+      {t('token.ctxLabel')} {contextPct}% ({fmt(contextUsed)}/{fmt(contextWindow)})
+    </span>
+  );
+  if (showCost) {
     parts.push(
-      <span key="model" className="truncate" title={model}>
-        {model}
-      </span>
-    );
-  }
-  if (showIn) {
-    parts.push(
-      <span key="in" title={t('token.inputTokens')}>
-        {t('token.inLabel')} {fmt(inTok)}
-      </span>
-    );
-  }
-  if (showOut) {
-    parts.push(
-      <span key="out" title={t('token.outputTokens')}>
-        {t('token.outLabel')} {fmt(outTok)}
-      </span>
-    );
-  }
-  if (turns > 0) {
-    parts.push(
-      <span key="turns" title={t('token.turnsCount')}>
-        {turns} {turns === 1 ? t('token.turn') : t('token.turns')}
-      </span>
-    );
-  }
-  if (showCtx) {
-    parts.push(
-      <span key="ctx" title={t('token.contextWindow')}>
-        {t('token.ctxLabel')} {contextPct}% ({fmt(contextUsed)}/{fmt(DEFAULT_CONTEXT_WINDOW)})
+      <span key="cost" title={t('token.costTitle')}>
+        {fmtUsd(costUsd)}
       </span>
     );
   }
@@ -86,13 +166,84 @@ export default function ComposerTokenStats({ model }: ComposerTokenStatsProps =
   if (parts.length === 0) return null;
 
   return (
-    <div className="flex min-w-0 flex-wrap items-center gap-2.5 text-[10px] font-mono text-content-faint select-none">
-      {parts.map((part, i) => (
-        <span key={i} className="contents">
-          {i > 0 && dot()}
-          {part}
-        </span>
-      ))}
+    <div ref={rootRef} className="relative flex min-w-0 items-center">
+      {/* Hover hint that the compact row is interactive; click opens the full
+          breakdown. The hint is suppressed while the popover is already open. */}
+      <Tooltip label={open ? '' : t('token.clickForDetails')} side="top">
+        <button
+          type="button"
+          onClick={() => setOpen(o => !o)}
+          aria-expanded={open}
+          aria-label={t('token.sessionUsageTitle')}
+          className="flex min-w-0 cursor-pointer flex-wrap items-center gap-1.5 border-0 bg-transparent p-0 text-[10px] font-mono text-content-faint select-none">
+          {parts.map((part, i) => (
+            <span key={i} className="contents">
+              {part}
+            </span>
+          ))}
+        </button>
+      </Tooltip>
+      {open && (
+        <div
+          data-testid="composer-token-breakdown"
+          role="dialog"
+          aria-label={t('token.sessionUsageTitle')}
+          className="absolute bottom-full left-0 z-50 mb-1.5 w-64 rounded-md border border-line-strong bg-surface p-2.5 text-[11px] shadow-lg">
+          <div className="mb-1.5 font-semibold text-content">{t('token.sessionUsageTitle')}</div>
+          {model && (
+            <div className="mb-1.5 truncate font-mono text-content-faint" title={model}>
+              {model}
+            </div>
+          )}
+          <dl className="space-y-0.5 text-content-secondary">
+            <UsageRow label={t('token.popInput')} tip={t('token.tipInput')} value={fmt(inTok)} />
+            <UsageRow label={t('token.popOutput')} tip={t('token.tipOutput')} value={fmt(outTok)} />
+            {ok(cachedTok) && (
+              <UsageRow
+                label={t('token.popCacheHit')}
+                tip={t('token.tipCacheHit')}
+                value={`${fmt(cachedTok)} (${
+                  inTok > 0 ? Math.min(100, Math.round((cachedTok / inTok) * 100)) : 0
+                }%)`}
+              />
+            )}
+            <UsageRow
+              label={t('token.popContext')}
+              tip={t('token.contextWindow')}
+              value={`${contextPct}% (${fmt(contextUsed)}/${fmt(contextWindow)})`}
+            />
+            <UsageRow
+              label={t('token.costLabel')}
+              tip={t('token.costTitle')}
+              value={fmtUsd(costUsd)}
+            />
+          </dl>
+          <div className="mt-2 border-t border-line-subtle pt-1.5">
+            <div className="mb-1 font-semibold text-content">{t('token.byAgentHeading')}</div>
+            <ul className="space-y-0.5 text-content-secondary">
+              {/* Orchestrator first, then each sub-agent archetype. */}
+              <AgentLine
+                name={t('token.orchestrator')}
+                tokens={orchestratorTokens}
+                costUsd={orchestratorCost}
+                runs={turns}
+              />
+              {subAgents.map(sub => (
+                <AgentLine
+                  key={sub.agentId}
+                  name={sub.agentId}
+                  tokens={sub.inputTokens + sub.outputTokens}
+                  costUsd={sub.costUsd}
+                  runs={sub.runs}
+                />
+              ))}
+            </ul>
+            {subAgents.length === 0 && (
+              <div className="mt-0.5 text-content-faint">{t('token.noSubAgents')}</div>
+            )}
+          </div>
+        </div>
+      )}
     </div>
   );
 }
diff --git a/app/src/lib/i18n/ar.ts b/app/src/lib/i18n/ar.ts
index d98c92e65..e423c5dec 100644
--- a/app/src/lib/i18n/ar.ts
+++ b/app/src/lib/i18n/ar.ts
@@ -2268,6 +2268,22 @@ const messages: TranslationMap = {
   'token.turns': 'دورات',
   'token.ctxLabel': 'سياق',
   'token.contextWindow': 'استخدام نافذة السياق (آخر دورة)',
+  'token.cachedLabel': 'مخزّن مؤقتًا',
+  'token.costLabel': 'التكلفة',
+  'token.costTitle': 'التكلفة التقديرية لهذه الجلسة (دولار أمريكي)',
+  'token.sessionUsageTitle': 'استخدام الجلسة',
+  'token.subAgentsHeading': 'استخدام الوكلاء الفرعيين',
+  'token.byAgentHeading': 'حسب الوكيل',
+  'token.orchestrator': 'المنسّق',
+  'token.noSubAgents': 'لا يوجد استخدام للوكلاء الفرعيين بعد',
+  'token.popInput': 'رموز الإدخال',
+  'token.popOutput': 'رموز الإخراج',
+  'token.popCacheHit': 'إصابة ذاكرة التخزين المؤقت',
+  'token.popContext': 'نافذة السياق',
+  'token.tipInput': 'الرموز المُرسلة إلى النموذج في هذه الجلسة — مطالباتك إضافةً إلى سجل المحادثة.',
+  'token.tipOutput': 'الرموز التي ولّدها النموذج في هذه الجلسة.',
+  'token.tipCacheHit': 'رموز إدخال تُقدَّم من بادئة مطالبة مخزّنة مؤقتًا، وتُحتسب بسعر مخفّض.',
+  'token.clickForDetails': 'انقر لعرض تفاصيل استخدام الجلسة',
   'catalog.noCapabilityBinding': 'لا يوجد ربط بالإمكانات',
   'catalog.downloadFailed': 'فشل التنزيل',
   'catalog.active': 'نشط',
diff --git a/app/src/lib/i18n/bn.ts b/app/src/lib/i18n/bn.ts
index a4ab5ef54..d78c723df 100644
--- a/app/src/lib/i18n/bn.ts
+++ b/app/src/lib/i18n/bn.ts
@@ -2319,6 +2319,22 @@ const messages: TranslationMap = {
   'token.turns': 'টার্ন',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'কনটেক্সট উইন্ডো ব্যবহার (শেষ টার্ন)',
+  'token.cachedLabel': 'ক্যাশড',
+  'token.costLabel': 'খরচ',
+  'token.costTitle': 'এই সেশনের আনুমানিক খরচ (USD)',
+  'token.sessionUsageTitle': 'সেশন ব্যবহার',
+  'token.subAgentsHeading': 'সাব-এজেন্ট ব্যবহার',
+  'token.byAgentHeading': 'এজেন্ট অনুযায়ী',
+  'token.orchestrator': 'অর্কেস্ট্রেটর',
+  'token.noSubAgents': 'এখনও কোনো সাব-এজেন্ট ব্যবহার নেই',
+  'token.popInput': 'ইনপুট টোকেন',
+  'token.popOutput': 'আউটপুট টোকেন',
+  'token.popCacheHit': 'ক্যাশ হিট',
+  'token.popContext': 'কনটেক্সট উইন্ডো',
+  'token.tipInput': 'এই সেশনে মডেলে পাঠানো টোকেন — আপনার প্রম্পট এবং কথোপকথনের ইতিহাস।',
+  'token.tipOutput': 'এই সেশনে মডেল দ্বারা তৈরি টোকেন।',
+  'token.tipCacheHit': 'ক্যাশড প্রম্পট প্রিফিক্স থেকে দেওয়া ইনপুট টোকেন, কম হারে বিল করা হয়।',
+  'token.clickForDetails': 'সেশন ব্যবহারের বিবরণের জন্য ক্লিক করুন',
   'catalog.noCapabilityBinding': 'কোনো ক্যাপাবিলিটি বাইন্ডিং নেই',
   'catalog.downloadFailed': 'ডাউনলোড ব্যর্থ',
   'catalog.active': 'সক্রিয়',
diff --git a/app/src/lib/i18n/de.ts b/app/src/lib/i18n/de.ts
index b4e673e99..9f8c94d95 100644
--- a/app/src/lib/i18n/de.ts
+++ b/app/src/lib/i18n/de.ts
@@ -2373,6 +2373,24 @@ const messages: TranslationMap = {
   'token.turns': 'Runden',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'Kontextfenster-Nutzung (letzte Runde)',
+  'token.cachedLabel': 'Zwischengespeichert',
+  'token.costLabel': 'Kosten',
+  'token.costTitle': 'Geschätzte Kosten dieser Sitzung (USD)',
+  'token.sessionUsageTitle': 'Sitzungsnutzung',
+  'token.subAgentsHeading': 'Subagenten-Nutzung',
+  'token.byAgentHeading': 'Nach Agent',
+  'token.orchestrator': 'Orchestrator',
+  'token.noSubAgents': 'Noch keine Subagenten-Nutzung',
+  'token.popInput': 'Eingabe-Tokens',
+  'token.popOutput': 'Ausgabe-Tokens',
+  'token.popCacheHit': 'Cache-Treffer',
+  'token.popContext': 'Kontextfenster',
+  'token.tipInput':
+    'In dieser Sitzung an das Modell gesendete Tokens — deine Prompts plus der Gesprächsverlauf.',
+  'token.tipOutput': 'In dieser Sitzung vom Modell erzeugte Tokens.',
+  'token.tipCacheHit':
+    'Eingabe-Tokens aus einem zwischengespeicherten Prompt-Präfix, zu einem reduzierten Tarif abgerechnet.',
+  'token.clickForDetails': 'Für Details zur Sitzungsnutzung klicken',
   'catalog.noCapabilityBinding': 'Keine Fähigkeitsbindung',
   'catalog.downloadFailed': 'Der Download ist fehlgeschlagen',
   'catalog.active': 'Aktiv',
diff --git a/app/src/lib/i18n/en.ts b/app/src/lib/i18n/en.ts
index c3d046150..7f643d22d 100644
--- a/app/src/lib/i18n/en.ts
+++ b/app/src/lib/i18n/en.ts
@@ -2721,6 +2721,23 @@ const en: TranslationMap = {
   'token.turns': 'turns',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'Context window usage (last turn)',
+  'token.cachedLabel': 'Cached',
+  'token.costLabel': 'Cost',
+  'token.costTitle': 'Estimated cost this session (USD)',
+  'token.sessionUsageTitle': 'Session usage',
+  'token.subAgentsHeading': 'Sub-agent usage',
+  'token.byAgentHeading': 'By agent',
+  'token.orchestrator': 'Orchestrator',
+  'token.noSubAgents': 'No sub-agent usage yet',
+  'token.popInput': 'Input tokens',
+  'token.popOutput': 'Output tokens',
+  'token.popCacheHit': 'Cache hit',
+  'token.popContext': 'Context window',
+  'token.tipInput':
+    'Tokens sent to the model this session — your prompts plus the conversation history.',
+  'token.tipOutput': 'Tokens generated by the model this session.',
+  'token.tipCacheHit': 'Input tokens served from a cached prompt prefix, billed at a reduced rate.',
+  'token.clickForDetails': 'Click for session usage details',
 
   // Catalog
   'catalog.noCapabilityBinding': 'No capability binding',
diff --git a/app/src/lib/i18n/es.ts b/app/src/lib/i18n/es.ts
index d3cac50b3..54d425ab8 100644
--- a/app/src/lib/i18n/es.ts
+++ b/app/src/lib/i18n/es.ts
@@ -2359,6 +2359,24 @@ const messages: TranslationMap = {
   'token.turns': 'turnos',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'Uso de ventana de contexto (último turno)',
+  'token.cachedLabel': 'En caché',
+  'token.costLabel': 'Costo',
+  'token.costTitle': 'Costo estimado de esta sesión (USD)',
+  'token.sessionUsageTitle': 'Uso de la sesión',
+  'token.subAgentsHeading': 'Uso de subagentes',
+  'token.byAgentHeading': 'Por agente',
+  'token.orchestrator': 'Orquestador',
+  'token.noSubAgents': 'Aún no hay uso de subagentes',
+  'token.popInput': 'Tokens de entrada',
+  'token.popOutput': 'Tokens de salida',
+  'token.popCacheHit': 'Acierto de caché',
+  'token.popContext': 'Ventana de contexto',
+  'token.tipInput':
+    'Tokens enviados al modelo en esta sesión: tus prompts más el historial de la conversación.',
+  'token.tipOutput': 'Tokens generados por el modelo en esta sesión.',
+  'token.tipCacheHit':
+    'Tokens de entrada servidos desde un prefijo de prompt en caché, facturados a una tarifa reducida.',
+  'token.clickForDetails': 'Haz clic para ver los detalles de uso de la sesión',
   'catalog.noCapabilityBinding': 'Sin vinculación de capacidad',
   'catalog.downloadFailed': 'Descarga fallida',
   'catalog.active': 'Activo',
diff --git a/app/src/lib/i18n/fr.ts b/app/src/lib/i18n/fr.ts
index ae63137aa..79da3501b 100644
--- a/app/src/lib/i18n/fr.ts
+++ b/app/src/lib/i18n/fr.ts
@@ -2372,6 +2372,24 @@ const messages: TranslationMap = {
   'token.turns': 'tours',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'Utilisation de la fenêtre de contexte (dernier tour)',
+  'token.cachedLabel': 'En cache',
+  'token.costLabel': 'Coût',
+  'token.costTitle': 'Coût estimé de cette session (USD)',
+  'token.sessionUsageTitle': 'Utilisation de la session',
+  'token.subAgentsHeading': 'Utilisation des sous-agents',
+  'token.byAgentHeading': 'Par agent',
+  'token.orchestrator': 'Orchestrateur',
+  'token.noSubAgents': 'Aucune utilisation de sous-agent pour le moment',
+  'token.popInput': 'Jetons d’entrée',
+  'token.popOutput': 'Jetons de sortie',
+  'token.popCacheHit': 'Succès de cache',
+  'token.popContext': 'Fenêtre de contexte',
+  'token.tipInput':
+    'Jetons envoyés au modèle durant cette session — vos invites plus l’historique de la conversation.',
+  'token.tipOutput': 'Jetons générés par le modèle durant cette session.',
+  'token.tipCacheHit':
+    'Jetons d’entrée servis depuis un préfixe d’invite en cache, facturés à un tarif réduit.',
+  'token.clickForDetails': 'Cliquez pour les détails d’utilisation de la session',
   'catalog.noCapabilityBinding': 'Aucune liaison de fonctionnalité',
   'catalog.downloadFailed': 'Échec du téléchargement',
   'catalog.active': 'Actif',
diff --git a/app/src/lib/i18n/hi.ts b/app/src/lib/i18n/hi.ts
index 8740dbac5..d760158c4 100644
--- a/app/src/lib/i18n/hi.ts
+++ b/app/src/lib/i18n/hi.ts
@@ -2315,6 +2315,22 @@ const messages: TranslationMap = {
   'token.turns': 'चक्र',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'संदर्भ विंडो उपयोग (अंतिम चक्र)',
+  'token.cachedLabel': 'कैश्ड',
+  'token.costLabel': 'लागत',
+  'token.costTitle': 'इस सत्र की अनुमानित लागत (USD)',
+  'token.sessionUsageTitle': 'सत्र उपयोग',
+  'token.subAgentsHeading': 'उप-एजेंट उपयोग',
+  'token.byAgentHeading': 'एजेंट अनुसार',
+  'token.orchestrator': 'ऑर्केस्ट्रेटर',
+  'token.noSubAgents': 'अभी तक कोई उप-एजेंट उपयोग नहीं',
+  'token.popInput': 'इनपुट टोकन',
+  'token.popOutput': 'आउटपुट टोकन',
+  'token.popCacheHit': 'कैश हिट',
+  'token.popContext': 'संदर्भ विंडो',
+  'token.tipInput': 'इस सत्र में मॉडल को भेजे गए टोकन — आपके प्रॉम्प्ट और बातचीत का इतिहास।',
+  'token.tipOutput': 'इस सत्र में मॉडल द्वारा उत्पन्न टोकन।',
+  'token.tipCacheHit': 'कैश किए गए प्रॉम्प्ट प्रिफ़िक्स से दिए गए इनपुट टोकन, कम दर पर बिल किए गए।',
+  'token.clickForDetails': 'सत्र उपयोग विवरण के लिए क्लिक करें',
   'catalog.noCapabilityBinding': 'कोई कैपेबिलिटी बाइंडिंग नहीं',
   'catalog.downloadFailed': 'डाउनलोड विफल',
   'catalog.active': 'एक्टिव',
diff --git a/app/src/lib/i18n/id.ts b/app/src/lib/i18n/id.ts
index d3dad2c49..0318f0604 100644
--- a/app/src/lib/i18n/id.ts
+++ b/app/src/lib/i18n/id.ts
@@ -2320,6 +2320,24 @@ const messages: TranslationMap = {
   'token.turns': 'putaran',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'Penggunaan jendela konteks (putaran terakhir)',
+  'token.cachedLabel': 'Tersimpan',
+  'token.costLabel': 'Biaya',
+  'token.costTitle': 'Perkiraan biaya sesi ini (USD)',
+  'token.sessionUsageTitle': 'Penggunaan sesi',
+  'token.subAgentsHeading': 'Penggunaan sub-agen',
+  'token.byAgentHeading': 'Per agen',
+  'token.orchestrator': 'Orkestrator',
+  'token.noSubAgents': 'Belum ada penggunaan sub-agen',
+  'token.popInput': 'Token masukan',
+  'token.popOutput': 'Token keluaran',
+  'token.popCacheHit': 'Cache hit',
+  'token.popContext': 'Jendela konteks',
+  'token.tipInput':
+    'Token yang dikirim ke model pada sesi ini — prompt Anda ditambah riwayat percakapan.',
+  'token.tipOutput': 'Token yang dihasilkan model pada sesi ini.',
+  'token.tipCacheHit':
+    'Token masukan yang dilayani dari prefiks prompt yang di-cache, ditagih dengan tarif lebih rendah.',
+  'token.clickForDetails': 'Klik untuk detail penggunaan sesi',
   'catalog.noCapabilityBinding': 'Tidak ada binding kemampuan',
   'catalog.downloadFailed': 'Unduhan gagal',
   'catalog.active': 'Aktif',
diff --git a/app/src/lib/i18n/it.ts b/app/src/lib/i18n/it.ts
index 82d9579e2..9aafd277a 100644
--- a/app/src/lib/i18n/it.ts
+++ b/app/src/lib/i18n/it.ts
@@ -2353,6 +2353,24 @@ const messages: TranslationMap = {
   'token.turns': 'turni',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'Utilizzo finestra di contesto (ultimo turno)',
+  'token.cachedLabel': 'In cache',
+  'token.costLabel': 'Costo',
+  'token.costTitle': 'Costo stimato di questa sessione (USD)',
+  'token.sessionUsageTitle': 'Utilizzo della sessione',
+  'token.subAgentsHeading': 'Utilizzo dei sub-agenti',
+  'token.byAgentHeading': 'Per agente',
+  'token.orchestrator': 'Orchestratore',
+  'token.noSubAgents': 'Nessun utilizzo di sub-agenti ancora',
+  'token.popInput': 'Token di input',
+  'token.popOutput': 'Token di output',
+  'token.popCacheHit': 'Cache hit',
+  'token.popContext': 'Finestra di contesto',
+  'token.tipInput':
+    'Token inviati al modello in questa sessione — i tuoi prompt più la cronologia della conversazione.',
+  'token.tipOutput': 'Token generati dal modello in questa sessione.',
+  'token.tipCacheHit':
+    'Token di input serviti da un prefisso di prompt in cache, fatturati a tariffa ridotta.',
+  'token.clickForDetails': 'Fai clic per i dettagli di utilizzo della sessione',
   'catalog.noCapabilityBinding': 'Nessun binding di capacità',
   'catalog.downloadFailed': 'Download fallito',
   'catalog.active': 'Attivo',
diff --git a/app/src/lib/i18n/ko.ts b/app/src/lib/i18n/ko.ts
index f8b3700b6..c64bcae8d 100644
--- a/app/src/lib/i18n/ko.ts
+++ b/app/src/lib/i18n/ko.ts
@@ -2294,6 +2294,22 @@ const messages: TranslationMap = {
   'token.turns': '턴',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': '컨텍스트 윈도우 사용량 (마지막 턴)',
+  'token.cachedLabel': '캐시됨',
+  'token.costLabel': '비용',
+  'token.costTitle': '이 세션의 예상 비용 (USD)',
+  'token.sessionUsageTitle': '세션 사용량',
+  'token.subAgentsHeading': '하위 에이전트 사용량',
+  'token.byAgentHeading': '에이전트별',
+  'token.orchestrator': '오케스트레이터',
+  'token.noSubAgents': '아직 하위 에이전트 사용량이 없습니다',
+  'token.popInput': '입력 토큰',
+  'token.popOutput': '출력 토큰',
+  'token.popCacheHit': '캐시 적중',
+  'token.popContext': '컨텍스트 윈도우',
+  'token.tipInput': '이 세션에서 모델로 보낸 토큰 — 사용자 프롬프트와 대화 기록.',
+  'token.tipOutput': '이 세션에서 모델이 생성한 토큰.',
+  'token.tipCacheHit': '캐시된 프롬프트 접두부에서 제공된 입력 토큰으로, 낮은 요금으로 청구됩니다.',
+  'token.clickForDetails': '세션 사용량 세부정보를 보려면 클릭하세요',
   'catalog.noCapabilityBinding': '기능 바인딩 없음',
   'catalog.downloadFailed': '다운로드 실패',
   'catalog.active': '활성',
diff --git a/app/src/lib/i18n/pl.ts b/app/src/lib/i18n/pl.ts
index 3a117cd32..09087fa9e 100644
--- a/app/src/lib/i18n/pl.ts
+++ b/app/src/lib/i18n/pl.ts
@@ -2339,6 +2339,23 @@ const messages: TranslationMap = {
   'token.turns': 'tury',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'Użycie okna kontekstu (ostatnia tura)',
+  'token.cachedLabel': 'Z pamięci',
+  'token.costLabel': 'Koszt',
+  'token.costTitle': 'Szacowany koszt tej sesji (USD)',
+  'token.sessionUsageTitle': 'Użycie sesji',
+  'token.subAgentsHeading': 'Użycie podagentów',
+  'token.byAgentHeading': 'Wg agenta',
+  'token.orchestrator': 'Orkiestrator',
+  'token.noSubAgents': 'Brak użycia podagentów',
+  'token.popInput': 'Tokeny wejściowe',
+  'token.popOutput': 'Tokeny wyjściowe',
+  'token.popCacheHit': 'Trafienie w pamięci podręcznej',
+  'token.popContext': 'Okno kontekstu',
+  'token.tipInput': 'Tokeny wysłane do modelu w tej sesji — Twoje prompty plus historia rozmowy.',
+  'token.tipOutput': 'Tokeny wygenerowane przez model w tej sesji.',
+  'token.tipCacheHit':
+    'Tokeny wejściowe podane z buforowanego prefiksu promptu, rozliczane po niższej stawce.',
+  'token.clickForDetails': 'Kliknij, aby zobaczyć szczegóły użycia sesji',
   'catalog.noCapabilityBinding': 'Brak powiązanej możliwości',
   'catalog.downloadFailed': 'Pobieranie nie powiodło się',
   'catalog.active': 'Aktywny',
diff --git a/app/src/lib/i18n/pt.ts b/app/src/lib/i18n/pt.ts
index ffe6ad0a6..00e4c583e 100644
--- a/app/src/lib/i18n/pt.ts
+++ b/app/src/lib/i18n/pt.ts
@@ -2360,6 +2360,24 @@ const messages: TranslationMap = {
   'token.turns': 'turnos',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'Uso da janela de contexto (último turno)',
+  'token.cachedLabel': 'Em cache',
+  'token.costLabel': 'Custo',
+  'token.costTitle': 'Custo estimado desta sessão (USD)',
+  'token.sessionUsageTitle': 'Uso da sessão',
+  'token.subAgentsHeading': 'Uso de subagentes',
+  'token.byAgentHeading': 'Por agente',
+  'token.orchestrator': 'Orquestrador',
+  'token.noSubAgents': 'Ainda sem uso de subagentes',
+  'token.popInput': 'Tokens de entrada',
+  'token.popOutput': 'Tokens de saída',
+  'token.popCacheHit': 'Acerto de cache',
+  'token.popContext': 'Janela de contexto',
+  'token.tipInput':
+    'Tokens enviados ao modelo nesta sessão — seus prompts mais o histórico da conversa.',
+  'token.tipOutput': 'Tokens gerados pelo modelo nesta sessão.',
+  'token.tipCacheHit':
+    'Tokens de entrada servidos de um prefixo de prompt em cache, cobrados a uma taxa reduzida.',
+  'token.clickForDetails': 'Clique para ver os detalhes de uso da sessão',
   'catalog.noCapabilityBinding': 'Sem vinculação de funcionalidade',
   'catalog.downloadFailed': 'Falha no download',
   'catalog.active': 'Ativo',
diff --git a/app/src/lib/i18n/ru.ts b/app/src/lib/i18n/ru.ts
index 6f15e1bf4..55d2f418b 100644
--- a/app/src/lib/i18n/ru.ts
+++ b/app/src/lib/i18n/ru.ts
@@ -2334,6 +2334,23 @@ const messages: TranslationMap = {
   'token.turns': 'циклов',
   'token.ctxLabel': 'CTX',
   'token.contextWindow': 'Использование окна контекста (последний цикл)',
+  'token.cachedLabel': 'Из кэша',
+  'token.costLabel': 'Стоимость',
+  'token.costTitle': 'Оценочная стоимость этой сессии (USD)',
+  'token.sessionUsageTitle': 'Использование сессии',
+  'token.subAgentsHeading': 'Использование субагентов',
+  'token.byAgentHeading': 'По агенту',
+  'token.orchestrator': 'Оркестратор',
+  'token.noSubAgents': 'Пока нет использования субагентов',
+  'token.popInput': 'Входные токены',
+  'token.popOutput': 'Выходные токены',
+  'token.popCacheHit': 'Попадание в кэш',
+  'token.popContext': 'Окно контекста',
+  'token.tipInput': 'Токены, отправленные модели в этой сессии — ваши запросы и история разговора.',
+  'token.tipOutput': 'Токены, сгенерированные моделью в этой сессии.',
+  'token.tipCacheHit':
+    'Входные токены, выданные из кэшированного префикса запроса и тарифицируемые по сниженной ставке.',
+  'token.clickForDetails': 'Нажмите, чтобы увидеть детали использования сессии',
   'catalog.noCapabilityBinding': 'Нет привязки возможностей',
   'catalog.downloadFailed': 'Загрузка не удалась',
   'catalog.active': 'Активно',
diff --git a/app/src/lib/i18n/zh-CN.ts b/app/src/lib/i18n/zh-CN.ts
index 48c785003..68c058d03 100644
--- a/app/src/lib/i18n/zh-CN.ts
+++ b/app/src/lib/i18n/zh-CN.ts
@@ -2197,6 +2197,22 @@ const messages: TranslationMap = {
   'token.turns': '轮',
   'token.ctxLabel': '上下文',
   'token.contextWindow': '上下文窗口使用量（上一轮）',
+  'token.cachedLabel': '已缓存',
+  'token.costLabel': '费用',
+  'token.costTitle': '本次会话的预估费用（美元）',
+  'token.sessionUsageTitle': '会话用量',
+  'token.subAgentsHeading': '子代理用量',
+  'token.byAgentHeading': '按代理',
+  'token.orchestrator': '编排器',
+  'token.noSubAgents': '暂无子代理用量',
+  'token.popInput': '输入令牌',
+  'token.popOutput': '输出令牌',
+  'token.popCacheHit': '缓存命中',
+  'token.popContext': '上下文窗口',
+  'token.tipInput': '本次会话发送给模型的令牌——你的提示词加上对话历史。',
+  'token.tipOutput': '本次会话由模型生成的令牌。',
+  'token.tipCacheHit': '由缓存的提示词前缀提供的输入令牌，按较低费率计费。',
+  'token.clickForDetails': '点击查看会话用量详情',
   'catalog.noCapabilityBinding': '无能力绑定',
   'catalog.downloadFailed': '下载失败',
   'catalog.active': '活跃',
diff --git a/app/src/pages/Conversations.tsx b/app/src/pages/Conversations.tsx
index f2547bd88..1b8aeaf86 100644
--- a/app/src/pages/Conversations.tsx
+++ b/app/src/pages/Conversations.tsx
@@ -36,6 +36,7 @@ import { trackEvent } from '../services/analytics';
 import { applyOpenRouterFreeModels } from '../services/api/openrouterFreeModels';
 import { subagentApi } from '../services/api/subagentApi';
 import { threadApi } from '../services/api/threadApi';
+import { fetchThreadTokenUsage } from '../services/api/threadUsageApi';
 import {
   aiRegenerate,
   chatCancel,
@@ -57,6 +58,7 @@ import {
   clearThreadSendPending,
   enqueueFollowup,
   fetchAndHydrateTurnState,
+  hydrateThreadUsage,
   markSubagentCancelled,
   markThreadSendPending,
   type QueuedFollowup,
@@ -553,6 +555,39 @@ const Conversations = ({
     }
   };
 
+  // Seed the composer footer with the selected thread's persisted token/cost
+  // usage (read back from its session transcripts) so the totals reflect prior
+  // turns instead of starting at zero. Best-effort; live turns accumulate on top
+  // via recordChatTurnUsage and a brand-new thread (hasUsage=false) is left as-is.
+  useEffect(() => {
+    if (!selectedThreadId) return;
+    let cancelled = false;
+    void fetchThreadTokenUsage(selectedThreadId)
+      .then(u => {
+        if (cancelled || !u.hasUsage) return;
+        dispatch(
+          hydrateThreadUsage({
+            threadId: u.threadId,
+            inputTokens: u.inputTokens,
+            outputTokens: u.outputTokens,
+            cachedTokens: u.cachedInputTokens,
+            costUsd: u.costUsd,
+            turns: u.turnCount,
+            contextWindow: u.contextWindow,
+            lastTurnInputTokens: u.lastTurnInputTokens,
+            lastTurnOutputTokens: u.lastTurnOutputTokens,
+            subAgents: u.subagents,
+          })
+        );
+      })
+      .catch(() => {
+        /* best-effort seed; the footer still fills from live turns */
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [selectedThreadId, dispatch]);
+
   useEffect(() => {
     let cancelled = false;
 
@@ -2881,7 +2916,7 @@ const Conversations = ({
           className="mt-2 flex items-center justify-between gap-2"
           data-walkthrough="chat-agent-panel">
           <div className="flex min-w-0 items-center gap-2">
-            <ComposerTokenStats model={resolvedModel} />
+            <ComposerTokenStats model={resolvedModel} threadId={selectedThreadId} />
             {/* Set/show the thread goal; click opens the editor above the composer. */}
             <ThreadGoalFooterTrigger ctl={threadGoal} />
           </div>
diff --git a/app/src/pages/conversations/components/PlanReviewCard.tsx b/app/src/pages/conversations/components/PlanReviewCard.tsx
index 8013ef27d..117bcd3fe 100644
--- a/app/src/pages/conversations/components/PlanReviewCard.tsx
+++ b/app/src/pages/conversations/components/PlanReviewCard.tsx
@@ -70,7 +70,7 @@ export const PlanReviewCard: React.FC<Props> = ({ threadId, review }) => {
       role="alertdialog"
       aria-label={t('conversations.planReview.title')}
       data-testid="plan-review-card"
-      className="mb-2 rounded-xl border border-ocean-300 bg-ocean-50 p-3 text-sm shadow-sm dark:border-ocean-700 dark:bg-ocean-950">
+      className="mb-2 rounded-xl border border-ocean-300 bg-surface p-3 text-sm shadow-md dark:border-ocean-700">
       <div className="flex items-start gap-2">
         <span aria-hidden className="text-base leading-none text-ocean-700 dark:text-ocean-200">
           🗺️
diff --git a/app/src/providers/ChatRuntimeProvider.tsx b/app/src/providers/ChatRuntimeProvider.tsx
index f206b3686..0d8675479 100644
--- a/app/src/providers/ChatRuntimeProvider.tsx
+++ b/app/src/providers/ChatRuntimeProvider.tsx
@@ -200,6 +200,50 @@ function chatDoneExtraMetadata(event: ChatDoneEvent): Record<string, unknown> |
   return event.citations?.length ? { citations: event.citations } : undefined;
 }
 
+/**
+ * Map a `chat_done` event's holistic usage onto the `recordChatTurnUsage`
+ * payload. Prefers the structured `usage` object (tokens + cost + context window
+ * + per-sub-agent breakdown); falls back to the deprecated flat token fields for
+ * any older core that still emits them.
+ */
+function chatTurnUsagePayload(event: ChatDoneEvent): {
+  inputTokens: number;
+  outputTokens: number;
+  cachedTokens?: number;
+  costUsd?: number;
+  contextWindow?: number;
+  threadId?: string;
+  subAgents?: Array<{
+    agentId: string;
+    inputTokens: number;
+    outputTokens: number;
+    costUsd: number;
+  }>;
+} {
+  const u = event.usage;
+  if (u) {
+    return {
+      inputTokens: u.input_tokens,
+      outputTokens: u.output_tokens,
+      cachedTokens: u.cached_input_tokens,
+      costUsd: u.cost_usd,
+      contextWindow: u.context_window,
+      threadId: event.thread_id,
+      subAgents: (u.subagents ?? []).map(s => ({
+        agentId: s.agent_id,
+        inputTokens: s.input_tokens,
+        outputTokens: s.output_tokens,
+        costUsd: s.cost_usd,
+      })),
+    };
+  }
+  return {
+    inputTokens: event.total_input_tokens ?? 0,
+    outputTokens: event.total_output_tokens ?? 0,
+    threadId: event.thread_id,
+  };
+}
+
 export function findPendingDelegationContext(
   entries: ToolTimelineEntry[],
   round: number
@@ -1087,12 +1131,7 @@ const ChatRuntimeProvider = ({ children }: { children: React.ReactNode }) => {
           store.getState().chatRuntime.parallelRequestThreads[event.request_id] !== undefined
         ) {
           const parallelRequestId = event.request_id;
-          dispatch(
-            recordChatTurnUsage({
-              inputTokens: event.total_input_tokens,
-              outputTokens: event.total_output_tokens,
-            })
-          );
+          dispatch(recordChatTurnUsage(chatTurnUsagePayload(event)));
           if (!event.segment_total && event.full_response.length > 0) {
             void (async () => {
               try {
@@ -1127,12 +1166,7 @@ const ChatRuntimeProvider = ({ children }: { children: React.ReactNode }) => {
         const segmentDelivery = takeSegmentDelivery(segmentDeliveriesRef.current, deliveryKey);
         const completeSegmentDelivery = hasCompleteSegmentDelivery(event, segmentDelivery);
 
-        dispatch(
-          recordChatTurnUsage({
-            inputTokens: event.total_input_tokens,
-            outputTokens: event.total_output_tokens,
-          })
-        );
+        dispatch(recordChatTurnUsage(chatTurnUsagePayload(event)));
         dispatch(clearInferenceStatusForThread({ threadId: event.thread_id }));
         dispatch(clearStreamingAssistantForThread({ threadId: event.thread_id }));
         dispatch(clearPendingApprovalForThread({ threadId: event.thread_id }));
diff --git a/app/src/services/api/threadUsageApi.ts b/app/src/services/api/threadUsageApi.ts
new file mode 100644
index 000000000..10a113439
--- /dev/null
+++ b/app/src/services/api/threadUsageApi.ts
@@ -0,0 +1,92 @@
+import { callCoreRpc } from '../coreRpcClient';
+
+/** One sub-agent archetype's contribution within a thread. */
+export interface ThreadSubagentUsage {
+  agentId: string;
+  inputTokens: number;
+  outputTokens: number;
+  costUsd: number;
+  runs: number;
+}
+
+/** Camel-cased per-thread usage totals consumed by the composer footer. */
+export interface ThreadTokenUsage {
+  threadId: string;
+  inputTokens: number;
+  outputTokens: number;
+  cachedInputTokens: number;
+  costUsd: number;
+  turnCount: number;
+  lastTurnInputTokens: number;
+  lastTurnOutputTokens: number;
+  contextWindow: number;
+  model: string | null;
+  updated: string | null;
+  hasUsage: boolean;
+  /** Per-archetype sub-agent breakdown (already included in the totals above). */
+  subagents: ThreadSubagentUsage[];
+}
+
+/** Wire shape returned by `openhuman.threads_token_usage` (snake_case). */
+interface ThreadSubagentUsageWire {
+  agent_id: string;
+  input_tokens: number;
+  output_tokens: number;
+  cost_usd: number;
+  runs: number;
+}
+
+interface ThreadTokenUsageWire {
+  thread_id: string;
+  input_tokens: number;
+  output_tokens: number;
+  cached_input_tokens: number;
+  cost_usd: number;
+  turn_count: number;
+  last_turn_input_tokens: number;
+  last_turn_output_tokens: number;
+  context_window: number;
+  model: string | null;
+  updated: string | null;
+  has_usage: boolean;
+  subagents?: ThreadSubagentUsageWire[];
+}
+
+interface Envelope<T> {
+  data?: T;
+}
+
+/**
+ * Fetch a thread's persisted token/cost totals from the core (read back from
+ * its session transcripts). Returns zeros with `hasUsage: false` for a thread
+ * that has no completed turns yet.
+ */
+export async function fetchThreadTokenUsage(threadId: string): Promise<ThreadTokenUsage> {
+  const response = await callCoreRpc<Envelope<ThreadTokenUsageWire>>({
+    method: 'openhuman.threads_token_usage',
+    params: { thread_id: threadId },
+  });
+  const d = response?.data;
+  if (!d) throw new Error('threads_token_usage returned an empty envelope');
+  return {
+    threadId: d.thread_id,
+    inputTokens: d.input_tokens,
+    outputTokens: d.output_tokens,
+    cachedInputTokens: d.cached_input_tokens,
+    costUsd: d.cost_usd,
+    turnCount: d.turn_count,
+    lastTurnInputTokens: d.last_turn_input_tokens,
+    lastTurnOutputTokens: d.last_turn_output_tokens,
+    contextWindow: d.context_window,
+    model: d.model,
+    updated: d.updated,
+    hasUsage: d.has_usage,
+    subagents: (d.subagents ?? []).map(s => ({
+      agentId: s.agent_id,
+      inputTokens: s.input_tokens,
+      outputTokens: s.output_tokens,
+      costUsd: s.cost_usd,
+      runs: s.runs,
+    })),
+  };
+}
diff --git a/app/src/services/chatService.ts b/app/src/services/chatService.ts
index 3f27fb01f..60c0b0aef 100644
--- a/app/src/services/chatService.ts
+++ b/app/src/services/chatService.ts
@@ -51,13 +51,47 @@ export interface ChatToolResultEvent {
   tool_call_id?: string;
 }
 
+/** One sub-agent's token/cost contribution within a turn (hover breakdown). */
+export interface SubagentUsageWire {
+  task_id: string;
+  agent_id: string;
+  input_tokens: number;
+  output_tokens: number;
+  cost_usd: number;
+}
+
+/**
+ * Holistic token/cost/context totals for a completed turn, carried on
+ * `chat_done`. Every numeric is a turn total (parent agent + any sub-agents);
+ * `subagents` breaks the same spend down per child. `context_window` is `0` when
+ * the core couldn't resolve the model's window.
+ */
+export interface TurnUsageWire {
+  input_tokens: number;
+  output_tokens: number;
+  cached_input_tokens: number;
+  cost_usd: number;
+  context_window: number;
+  subagents?: SubagentUsageWire[];
+}
+
 export interface ChatDoneEvent {
   thread_id: string;
   request_id?: string;
   full_response: string;
   rounds_used: number;
-  total_input_tokens: number;
-  total_output_tokens: number;
+  /**
+   * @deprecated Superseded by {@link ChatDoneEvent.usage}. The core no longer
+   * populates these flat fields; read token totals from `usage` instead.
+   */
+  total_input_tokens?: number;
+  /** @deprecated See {@link ChatDoneEvent.total_input_tokens}. */
+  total_output_tokens?: number;
+  /**
+   * Holistic token/cost/context usage for the turn (parent + sub-agents).
+   * Absent on synthetic done events that never ran a real turn.
+   */
+  usage?: TurnUsageWire | null;
   /** Emoji reaction decided by the local model (if any). */
   reaction_emoji?: string | null;
   /** Total segments when the response was split into bubbles by Rust. */
diff --git a/app/src/store/chatRuntimeSlice.test.ts b/app/src/store/chatRuntimeSlice.test.ts
index 3674b912b..33b607164 100644
--- a/app/src/store/chatRuntimeSlice.test.ts
+++ b/app/src/store/chatRuntimeSlice.test.ts
@@ -9,7 +9,10 @@ import chatRuntimeReducer, {
   clearRuntimeForThread,
   hydrateRuntimeFromRunLedger,
   hydrateRuntimeFromSnapshot,
+  hydrateThreadUsage,
   type QueueStatus,
+  recordChatTurnUsage,
+  resetSessionTokenUsage,
   setQueueStatusForThread,
   setToolTimelineForThread,
 } from './chatRuntimeSlice';
@@ -48,6 +51,167 @@ function makeStore() {
   return configureStore({ reducer: { chatRuntime: chatRuntimeReducer } });
 }
 
+describe('chatRuntimeSlice recordChatTurnUsage', () => {
+  it('accumulates tokens, cost, and context window across turns', () => {
+    const store = makeStore();
+    store.dispatch(
+      recordChatTurnUsage({
+        inputTokens: 1000,
+        outputTokens: 200,
+        cachedTokens: 50,
+        costUsd: 0.012,
+        contextWindow: 200_000,
+      })
+    );
+    store.dispatch(
+      recordChatTurnUsage({
+        inputTokens: 500,
+        outputTokens: 100,
+        cachedTokens: 10,
+        costUsd: 0.008,
+        contextWindow: 200_000,
+      })
+    );
+    const usage = store.getState().chatRuntime.sessionTokenUsage;
+    expect(usage.inputTokens).toBe(1500);
+    expect(usage.outputTokens).toBe(300);
+    expect(usage.cachedTokens).toBe(60);
+    expect(usage.costUsd).toBeCloseTo(0.02, 6);
+    expect(usage.turns).toBe(2);
+    expect(usage.contextWindow).toBe(200_000);
+    // Context gauge tracks the latest turn's input+output, not the running sum.
+    expect(usage.lastTurnContextUsed).toBe(600);
+  });
+
+  it('rolls sub-agent spend into a per-archetype breakdown keyed by agentId', () => {
+    const store = makeStore();
+    store.dispatch(
+      recordChatTurnUsage({
+        inputTokens: 100,
+        outputTokens: 20,
+        subAgents: [
+          { agentId: 'researcher', inputTokens: 40, outputTokens: 10, costUsd: 0.001 },
+          { agentId: 'coder', inputTokens: 80, outputTokens: 30, costUsd: 0.003 },
+        ],
+      })
+    );
+    store.dispatch(
+      recordChatTurnUsage({
+        inputTokens: 50,
+        outputTokens: 10,
+        subAgents: [{ agentId: 'researcher', inputTokens: 60, outputTokens: 5, costUsd: 0.002 }],
+      })
+    );
+    const subs = store.getState().chatRuntime.sessionTokenUsage.subAgents;
+    expect(subs.researcher).toEqual({
+      agentId: 'researcher',
+      inputTokens: 100,
+      outputTokens: 15,
+      costUsd: 0.003,
+      runs: 2,
+    });
+    expect(subs.coder.runs).toBe(1);
+    expect(subs.coder.inputTokens).toBe(80);
+  });
+
+  it('keeps the prior context window when a turn reports an unknown (0) window', () => {
+    const store = makeStore();
+    store.dispatch(
+      recordChatTurnUsage({ inputTokens: 10, outputTokens: 5, contextWindow: 128_000 })
+    );
+    store.dispatch(recordChatTurnUsage({ inputTokens: 10, outputTokens: 5, contextWindow: 0 }));
+    expect(store.getState().chatRuntime.sessionTokenUsage.contextWindow).toBe(128_000);
+  });
+
+  it('coerces non-finite / negative inputs to zero', () => {
+    const store = makeStore();
+    store.dispatch(
+      recordChatTurnUsage({ inputTokens: Number.NaN, outputTokens: -50, costUsd: -1 })
+    );
+    const usage = store.getState().chatRuntime.sessionTokenUsage;
+    expect(usage.inputTokens).toBe(0);
+    expect(usage.outputTokens).toBe(0);
+    expect(usage.costUsd).toBe(0);
+    expect(usage.turns).toBe(1);
+  });
+
+  it('resetSessionTokenUsage clears all accumulated usage', () => {
+    const store = makeStore();
+    store.dispatch(
+      recordChatTurnUsage({
+        inputTokens: 100,
+        outputTokens: 20,
+        costUsd: 0.01,
+        subAgents: [{ agentId: 'researcher', inputTokens: 1, outputTokens: 1, costUsd: 0.001 }],
+      })
+    );
+    store.dispatch(resetSessionTokenUsage());
+    const usage = store.getState().chatRuntime.sessionTokenUsage;
+    expect(usage.inputTokens).toBe(0);
+    expect(usage.costUsd).toBe(0);
+    expect(usage.turns).toBe(0);
+    expect(usage.subAgents).toEqual({});
+  });
+
+  it('routes a turn with a threadId into that thread bucket (and the global)', () => {
+    const store = makeStore();
+    store.dispatch(
+      recordChatTurnUsage({ inputTokens: 100, outputTokens: 20, costUsd: 0.01, threadId: 'thr-a' })
+    );
+    store.dispatch(
+      recordChatTurnUsage({ inputTokens: 50, outputTokens: 10, costUsd: 0.005, threadId: 'thr-b' })
+    );
+    const { usageByThread, sessionTokenUsage } = store.getState().chatRuntime;
+    expect(usageByThread['thr-a'].inputTokens).toBe(100);
+    expect(usageByThread['thr-a'].costUsd).toBeCloseTo(0.01, 6);
+    expect(usageByThread['thr-b'].inputTokens).toBe(50);
+    // Global aggregate still sums both threads.
+    expect(sessionTokenUsage.inputTokens).toBe(150);
+  });
+
+  it('hydrateThreadUsage seeds a thread bucket and live turns accumulate on top', () => {
+    const store = makeStore();
+    store.dispatch(
+      hydrateThreadUsage({
+        threadId: 'thr-a',
+        inputTokens: 1000,
+        outputTokens: 300,
+        cachedTokens: 40,
+        costUsd: 0.02,
+        turns: 3,
+        contextWindow: 1_000_000,
+        lastTurnInputTokens: 400,
+        lastTurnOutputTokens: 120,
+        subAgents: [
+          { agentId: 'coder', inputTokens: 300, outputTokens: 80, costUsd: 0.006, runs: 2 },
+        ],
+      })
+    );
+    let bucket = store.getState().chatRuntime.usageByThread['thr-a'];
+    expect(bucket.inputTokens).toBe(1000);
+    expect(bucket.turns).toBe(3);
+    expect(bucket.contextWindow).toBe(1_000_000);
+    expect(bucket.lastTurnContextUsed).toBe(520);
+    // Sub-agent breakdown reconstructed from persisted transcripts.
+    expect(bucket.subAgents.coder).toEqual({
+      agentId: 'coder',
+      inputTokens: 300,
+      outputTokens: 80,
+      costUsd: 0.006,
+      runs: 2,
+    });
+
+    // A live turn for the same thread adds on top of the seeded base.
+    store.dispatch(
+      recordChatTurnUsage({ inputTokens: 200, outputTokens: 50, costUsd: 0.004, threadId: 'thr-a' })
+    );
+    bucket = store.getState().chatRuntime.usageByThread['thr-a'];
+    expect(bucket.inputTokens).toBe(1200);
+    expect(bucket.turns).toBe(4);
+    expect(bucket.costUsd).toBeCloseTo(0.024, 6);
+  });
+});
+
 describe('chatRuntimeSlice queue status', () => {
   it('sets queue status for a thread', () => {
     const store = makeStore();
diff --git a/app/src/store/chatRuntimeSlice.ts b/app/src/store/chatRuntimeSlice.ts
index d22af566f..e4197ce7a 100644
--- a/app/src/store/chatRuntimeSlice.ts
+++ b/app/src/store/chatRuntimeSlice.ts
@@ -214,6 +214,20 @@ export interface StreamingAssistantState {
  */
 export type InferenceTurnLifecycle = 'started' | 'streaming' | 'interrupted';
 
+/**
+ * Per-sub-agent token/cost contribution, accumulated across the session and
+ * keyed by the sub-agent archetype id (e.g. `researcher`). Drives the hover
+ * breakdown under the composer footer's cost/context cluster.
+ */
+export interface SubAgentUsage {
+  agentId: string;
+  inputTokens: number;
+  outputTokens: number;
+  costUsd: number;
+  /** How many times this archetype was spawned across the session. */
+  runs: number;
+}
+
 /** Running per-session totals accumulated from `chat:done` events (#703). */
 export interface SessionTokenUsage {
   inputTokens: number;
@@ -222,6 +236,90 @@ export interface SessionTokenUsage {
   lastUpdated: number;
   lastTurnInputTokens: number;
   lastTurnOutputTokens: number;
+  /** Cached-input tokens accumulated across the session. */
+  cachedTokens: number;
+  /** Total USD cost accumulated across the session (parent + sub-agents). */
+  costUsd: number;
+  /**
+   * Most recent known model context window (tokens). `0` until a turn reports a
+   * real value; the UI falls back to a default when unknown.
+   */
+  contextWindow: number;
+  /** Last turn's input+output tokens — the context-window gauge numerator. */
+  lastTurnContextUsed: number;
+  /** Per-sub-agent spend for the session, keyed by archetype id. */
+  subAgents: Record<string, SubAgentUsage>;
+}
+
+/** A zeroed [SessionTokenUsage] bucket. */
+export function emptySessionTokenUsage(): SessionTokenUsage {
+  return {
+    inputTokens: 0,
+    outputTokens: 0,
+    turns: 0,
+    lastUpdated: 0,
+    lastTurnInputTokens: 0,
+    lastTurnOutputTokens: 0,
+    cachedTokens: 0,
+    costUsd: 0,
+    contextWindow: 0,
+    lastTurnContextUsed: 0,
+    subAgents: {},
+  };
+}
+
+/** Payload accepted by `recordChatTurnUsage` (and applied per turn). */
+export interface ChatTurnUsagePayload {
+  inputTokens: number;
+  outputTokens: number;
+  cachedTokens?: number;
+  costUsd?: number;
+  contextWindow?: number;
+  /** Thread the turn belongs to; routes the delta to that thread's bucket. */
+  threadId?: string;
+  subAgents?: Array<{
+    agentId: string;
+    inputTokens: number;
+    outputTokens: number;
+    costUsd: number;
+  }>;
+}
+
+const nonNeg = (n: number | undefined): number =>
+  typeof n === 'number' && Number.isFinite(n) ? Math.max(0, n) : 0;
+
+/** Fold one turn's usage delta into a bucket (mutates in place). */
+function applyTurnUsage(usage: SessionTokenUsage, payload: ChatTurnUsagePayload): void {
+  const inTok = nonNeg(payload.inputTokens);
+  const outTok = nonNeg(payload.outputTokens);
+  usage.inputTokens += inTok;
+  usage.outputTokens += outTok;
+  usage.cachedTokens += nonNeg(payload.cachedTokens);
+  usage.costUsd += nonNeg(payload.costUsd);
+  usage.turns += 1;
+  usage.lastUpdated = Date.now();
+  usage.lastTurnInputTokens = inTok;
+  usage.lastTurnOutputTokens = outTok;
+  usage.lastTurnContextUsed = inTok + outTok;
+  // Only overwrite the known context window when the turn reported a real value
+  // (>0); an unknown-window turn leaves the prior value intact.
+  const ctxWindow = nonNeg(payload.contextWindow);
+  if (ctxWindow > 0) usage.contextWindow = ctxWindow;
+  for (const sub of payload.subAgents ?? []) {
+    if (!sub || typeof sub.agentId !== 'string' || sub.agentId.length === 0) continue;
+    const existing = usage.subAgents[sub.agentId] ?? {
+      agentId: sub.agentId,
+      inputTokens: 0,
+      outputTokens: 0,
+      costUsd: 0,
+      runs: 0,
+    };
+    existing.inputTokens += nonNeg(sub.inputTokens);
+    existing.outputTokens += nonNeg(sub.outputTokens);
+    existing.costUsd += nonNeg(sub.costUsd);
+    existing.runs += 1;
+    usage.subAgents[sub.agentId] = existing;
+  }
 }
 
 /**
@@ -356,7 +454,15 @@ interface ChatRuntimeState {
    * download / retry affordances (#2779).
    */
   artifactsByThread: Record<string, ArtifactSnapshot[]>;
+  /** Global, app-session-wide token usage (legacy aggregate). */
   sessionTokenUsage: SessionTokenUsage;
+  /**
+   * Per-thread token usage, keyed by thread id. Seeded from persisted
+   * transcripts via `hydrateThreadUsage` when a thread is opened, then kept live
+   * by `recordChatTurnUsage`. The composer footer reads the active thread's
+   * bucket so its totals reflect the selected thread, not the whole app session.
+   */
+  usageByThread: Record<string, SessionTokenUsage>;
   queueStatusByThread: Record<string, QueueStatus>;
   /**
    * Follow-up messages the user submitted while a turn was still streaming
@@ -407,14 +513,8 @@ const initialState: ChatRuntimeState = {
   pendingApprovalByThread: {},
   pendingPlanReviewByThread: {},
   artifactsByThread: {},
-  sessionTokenUsage: {
-    inputTokens: 0,
-    outputTokens: 0,
-    turns: 0,
-    lastUpdated: 0,
-    lastTurnInputTokens: 0,
-    lastTurnOutputTokens: 0,
-  },
+  sessionTokenUsage: emptySessionTokenUsage(),
+  usageByThread: {},
   queueStatusByThread: {},
   queuedFollowupsByThread: {},
 };
@@ -1121,32 +1221,76 @@ const chatRuntimeSlice = createSlice({
       state.queuedFollowupsByThread = {};
       state.pendingSendThreadIds = {};
     },
-    recordChatTurnUsage: (
+    recordChatTurnUsage: (state, action: PayloadAction<ChatTurnUsagePayload>) => {
+      // Fold into the global aggregate and, when the turn names a thread, into
+      // that thread's bucket (what the composer footer reads).
+      applyTurnUsage(state.sessionTokenUsage, action.payload);
+      const threadId = action.payload.threadId;
+      if (threadId) {
+        const bucket = state.usageByThread[threadId] ?? emptySessionTokenUsage();
+        applyTurnUsage(bucket, action.payload);
+        state.usageByThread[threadId] = bucket;
+      }
+    },
+    /**
+     * Seed a thread's usage bucket from persisted transcript totals (the
+     * `openhuman.threads_token_usage` RPC). Replaces the bucket so re-opening a
+     * thread reflects its on-disk history rather than starting at zero. Live
+     * turns then accumulate on top via `recordChatTurnUsage`.
+     */
+    hydrateThreadUsage: (
       state,
-      action: PayloadAction<{ inputTokens: number; outputTokens: number }>
+      action: PayloadAction<{
+        threadId: string;
+        inputTokens: number;
+        outputTokens: number;
+        cachedTokens: number;
+        costUsd: number;
+        turns: number;
+        contextWindow: number;
+        lastTurnInputTokens: number;
+        lastTurnOutputTokens: number;
+        subAgents?: Array<{
+          agentId: string;
+          inputTokens: number;
+          outputTokens: number;
+          costUsd: number;
+          runs: number;
+        }>;
+      }>
     ) => {
-      const inTok = Number.isFinite(action.payload.inputTokens)
-        ? Math.max(0, action.payload.inputTokens)
-        : 0;
-      const outTok = Number.isFinite(action.payload.outputTokens)
-        ? Math.max(0, action.payload.outputTokens)
-        : 0;
-      state.sessionTokenUsage.inputTokens += inTok;
-      state.sessionTokenUsage.outputTokens += outTok;
-      state.sessionTokenUsage.turns += 1;
-      state.sessionTokenUsage.lastUpdated = Date.now();
-      state.sessionTokenUsage.lastTurnInputTokens = inTok;
-      state.sessionTokenUsage.lastTurnOutputTokens = outTok;
+      const p = action.payload;
+      if (!p.threadId) return;
+      // Reconstruct the per-archetype sub-agent map from the persisted breakdown
+      // (read back from the thread's `__` sub-agent transcripts).
+      const subAgents: Record<string, SubAgentUsage> = {};
+      for (const s of p.subAgents ?? []) {
+        if (!s || typeof s.agentId !== 'string' || s.agentId.length === 0) continue;
+        subAgents[s.agentId] = {
+          agentId: s.agentId,
+          inputTokens: nonNeg(s.inputTokens),
+          outputTokens: nonNeg(s.outputTokens),
+          costUsd: nonNeg(s.costUsd),
+          runs: nonNeg(s.runs),
+        };
+      }
+      state.usageByThread[p.threadId] = {
+        inputTokens: nonNeg(p.inputTokens),
+        outputTokens: nonNeg(p.outputTokens),
+        cachedTokens: nonNeg(p.cachedTokens),
+        costUsd: nonNeg(p.costUsd),
+        turns: nonNeg(p.turns),
+        lastUpdated: Date.now(),
+        lastTurnInputTokens: nonNeg(p.lastTurnInputTokens),
+        lastTurnOutputTokens: nonNeg(p.lastTurnOutputTokens),
+        contextWindow: nonNeg(p.contextWindow),
+        lastTurnContextUsed: nonNeg(p.lastTurnInputTokens) + nonNeg(p.lastTurnOutputTokens),
+        subAgents,
+      };
     },
     resetSessionTokenUsage: state => {
-      state.sessionTokenUsage = {
-        inputTokens: 0,
-        outputTokens: 0,
-        turns: 0,
-        lastUpdated: 0,
-        lastTurnInputTokens: 0,
-        lastTurnOutputTokens: 0,
-      };
+      state.sessionTokenUsage = emptySessionTokenUsage();
+      state.usageByThread = {};
     },
     /**
      * Apply a persisted [TurnState] snapshot from the Rust core to the
@@ -1292,6 +1436,7 @@ export const {
   clearRuntimeForThread,
   clearAllChatRuntime,
   recordChatTurnUsage,
+  hydrateThreadUsage,
   resetSessionTokenUsage,
   hydrateRuntimeFromSnapshot,
   hydrateRuntimeFromRunLedger,
diff --git a/src/core/socketio.rs b/src/core/socketio.rs
index 5373cb99a..8bce55a2a 100644
--- a/src/core/socketio.rs
+++ b/src/core/socketio.rs
@@ -210,6 +210,44 @@ pub struct WebChannelEvent {
     /// shown after [`Self::tool_display_label`].
     #[serde(skip_serializing_if = "Option::is_none")]
     pub tool_display_detail: Option<String>,
+    /// Holistic token/cost/context usage for a completed turn (parent +
+    /// sub-agents), carried on `chat_done`. Lets the UI footer show session
+    /// tokens, USD cost, and real context-window utilisation, with a
+    /// per-sub-agent hover breakdown. `None` for every non-`chat_done` event and
+    /// for synthetic done events that never ran a real turn.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub usage: Option<TurnUsagePayload>,
+}
+
+/// Token/cost/context totals for one completed turn, attached to `chat_done`.
+///
+/// Every numeric is a turn total (parent agent **plus** any sub-agents spawned
+/// during the turn); the `subagents` list breaks the same spend down per child
+/// for the UI hover. `context_window` is `0` when the core couldn't resolve the
+/// model's window (e.g. an unknown cloud model) — the UI falls back to a
+/// default in that case.
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub struct TurnUsagePayload {
+    pub input_tokens: u64,
+    pub output_tokens: u64,
+    pub cached_input_tokens: u64,
+    pub cost_usd: f64,
+    pub context_window: u64,
+    /// Per-sub-agent spend, omitted from the wire when no sub-agents ran.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub subagents: Vec<SubagentUsagePayload>,
+}
+
+/// One sub-agent's token/cost contribution within a turn (hover breakdown).
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub struct SubagentUsagePayload {
+    pub task_id: String,
+    pub agent_id: String,
+    pub input_tokens: u64,
+    pub output_tokens: u64,
+    pub cost_usd: f64,
 }
 
 /// Per-event subagent progress detail attached to `WebChannelEvent`.
diff --git a/src/openhuman/agent/cost.rs b/src/openhuman/agent/cost.rs
index 0a4c5c6ff..e7bcf9b88 100644
--- a/src/openhuman/agent/cost.rs
+++ b/src/openhuman/agent/cost.rs
@@ -62,45 +62,47 @@ const FALLBACK_PRICING: ModelPricing = ModelPricing {
 /// them as best-effort estimates for cases where the backend doesn't
 /// echo `charged_amount_usd`.
 pub const PRICING_TABLE: &[ModelPricing] = &[
-    // Reasoning tier — currently maps to Claude Opus 4.x family.
+    // Reasoning tier — managed "Pro" model rates (estimate; the backend's
+    // echoed `charged_amount_usd` is authoritative when present). Shared with
+    // the coding/agentic tiers below. Update when backend pricing changes.
     ModelPricing {
         model: "reasoning-v1",
-        input_per_mtok_usd: 15.00,
-        cached_input_per_mtok_usd: 1.50,
-        output_per_mtok_usd: 75.00,
+        input_per_mtok_usd: 0.435,
+        cached_input_per_mtok_usd: 0.003625,
+        output_per_mtok_usd: 0.87,
     },
-    // Chat tier — Kimi K2.6 Turbo on Fireworks (backend PR #760).
-    // Low TTFT, 128k context, `supportsThinking: false`. Rates track
-    // Fireworks' published Kimi turbo pricing at time of writing.
+    // Chat tier — managed "Flash" model rates (estimate). Cheaper, lower-latency
+    // model used for direct conversational turns.
     ModelPricing {
         model: "chat-v1",
-        input_per_mtok_usd: 0.60,
-        cached_input_per_mtok_usd: 0.06,
-        output_per_mtok_usd: 2.50,
+        input_per_mtok_usd: 0.14,
+        cached_input_per_mtok_usd: 0.0028,
+        output_per_mtok_usd: 0.28,
     },
-    // Legacy chat tier slug retained for older transcripts/configs.
+    // Legacy chat tier slug retained for older transcripts/configs — "Flash"
+    // rates, same as `chat-v1`.
     ModelPricing {
         model: "reasoning-quick-v1",
-        input_per_mtok_usd: 0.60,
-        cached_input_per_mtok_usd: 0.06,
-        output_per_mtok_usd: 2.50,
+        input_per_mtok_usd: 0.14,
+        cached_input_per_mtok_usd: 0.0028,
+        output_per_mtok_usd: 0.28,
     },
-    // Agentic tier — maps to Sonnet-class models.
+    // Agentic tier — managed "Pro" model rates (same as reasoning).
     ModelPricing {
         model: "agentic-v1",
-        input_per_mtok_usd: 3.00,
-        cached_input_per_mtok_usd: 0.30,
-        output_per_mtok_usd: 15.00,
+        input_per_mtok_usd: 0.435,
+        cached_input_per_mtok_usd: 0.003625,
+        output_per_mtok_usd: 0.87,
     },
-    // Coding tier — Sonnet-class.
+    // Coding tier — managed "Pro" model rates (same as reasoning).
     ModelPricing {
         model: "coding-v1",
-        input_per_mtok_usd: 3.00,
-        cached_input_per_mtok_usd: 0.30,
-        output_per_mtok_usd: 15.00,
+        input_per_mtok_usd: 0.435,
+        cached_input_per_mtok_usd: 0.003625,
+        output_per_mtok_usd: 0.87,
     },
-    // Vision tier — multimodal Sonnet-class. Estimate only; the backend's
-    // echoed `charged_amount_usd` is authoritative when present.
+    // Vision tier — multimodal; estimate only. The backend's echoed
+    // `charged_amount_usd` is authoritative when present.
     ModelPricing {
         model: "vision-v1",
         input_per_mtok_usd: 3.00,
@@ -236,8 +238,9 @@ mod tests {
 
     #[test]
     fn lookup_pricing_matches_canonical_tiers() {
-        assert_eq!(lookup_pricing("reasoning-v1").input_per_mtok_usd, 15.0);
-        assert_eq!(lookup_pricing("agentic-v1").output_per_mtok_usd, 15.0);
+        // Reasoning/agentic share the managed "Pro" rates.
+        assert_eq!(lookup_pricing("reasoning-v1").input_per_mtok_usd, 0.435);
+        assert_eq!(lookup_pricing("agentic-v1").output_per_mtok_usd, 0.87);
     }
 
     #[test]
@@ -257,7 +260,9 @@ mod tests {
 
     #[test]
     fn lookup_pricing_handles_concrete_vendor_names() {
-        assert_eq!(lookup_pricing("claude-opus-4.7").input_per_mtok_usd, 15.0);
+        // `claude-opus-4.7` (dotted, not a catalog id) resolves via the `opus`
+        // vendor heuristic to the reasoning tier ("Pro" rates).
+        assert_eq!(lookup_pricing("claude-opus-4.7").input_per_mtok_usd, 0.435);
         assert_eq!(
             lookup_pricing("claude-sonnet-4-6").output_per_mtok_usd,
             15.0
@@ -275,11 +280,11 @@ mod tests {
 
     #[test]
     fn estimate_call_cost_subtracts_cached_input() {
-        // 1M standard input + 1M cached input + 1M output on agentic-v1.
+        // 1M standard input + 1M cached input + 1M output on agentic-v1 ("Pro").
         let u = usage(2_000_000, 1_000_000, 1_000_000, 0.0);
         let est = estimate_call_cost_usd("agentic-v1", &u);
-        // 1M * 3 + 1M * 0.3 + 1M * 15 = 18.3
-        assert!((est - 18.3).abs() < 1e-6, "got {est}");
+        // 1M*0.435 + 1M*0.003625 + 1M*0.87 = 1.308625
+        assert!((est - 1.308625).abs() < 1e-6, "got {est}");
     }
 
     #[test]
@@ -291,19 +296,19 @@ mod tests {
     #[test]
     fn call_cost_falls_back_to_estimate_when_charged_zero() {
         let u = usage(1_000_000, 0, 0, 0.0);
-        // 1M input * 3 = 3
-        assert!((call_cost_usd("agentic-v1", &u) - 3.0).abs() < 1e-6);
+        // 1M input * 0.435 = 0.435
+        assert!((call_cost_usd("agentic-v1", &u) - 0.435).abs() < 1e-6);
     }
 
     #[test]
     fn turn_cost_accumulates_charged_and_estimated_separately() {
         let mut tc = TurnCost::new();
         tc.add_call("reasoning-v1", &usage(0, 0, 0, 0.10));
-        tc.add_call("agentic-v1", &usage(1_000_000, 0, 0, 0.0)); // est: 3.00
+        tc.add_call("agentic-v1", &usage(1_000_000, 0, 0, 0.0)); // est: 0.435
         assert_eq!(tc.call_count, 2);
         assert!((tc.charged_usd - 0.10).abs() < 1e-6);
-        assert!((tc.estimated_usd - 3.0).abs() < 1e-6);
-        assert!((tc.total_usd() - 3.10).abs() < 1e-6);
+        assert!((tc.estimated_usd - 0.435).abs() < 1e-6);
+        assert!((tc.total_usd() - 0.535).abs() < 1e-6);
     }
 
     #[test]
diff --git a/src/openhuman/agent/harness/mod.rs b/src/openhuman/agent/harness/mod.rs
index fb9ef20e4..f772ea4a5 100644
--- a/src/openhuman/agent/harness/mod.rs
+++ b/src/openhuman/agent/harness/mod.rs
@@ -47,6 +47,7 @@ pub(crate) mod tool_filter;
 mod tool_loop;
 pub(crate) mod tool_result_artifacts;
 pub mod turn_attachments_context;
+pub mod turn_subagent_usage;
 pub mod worktree_context;
 
 pub use definition::{
diff --git a/src/openhuman/agent/harness/session/builder/setters.rs b/src/openhuman/agent/harness/session/builder/setters.rs
index 996e05cce..61d8be429 100644
--- a/src/openhuman/agent/harness/session/builder/setters.rs
+++ b/src/openhuman/agent/harness/session/builder/setters.rs
@@ -531,6 +531,7 @@ impl AgentBuilder {
             auto_save: self.auto_save.unwrap_or(false),
             last_memory_context: None,
             last_turn_citations: Vec::new(),
+            last_turn_usage_totals: None,
             history: Vec::new(),
             post_turn_hooks: self.post_turn_hooks,
             learning_enabled: self.learning_enabled,
diff --git a/src/openhuman/agent/harness/session/runtime.rs b/src/openhuman/agent/harness/session/runtime.rs
index a625bfcd4..9cd793508 100644
--- a/src/openhuman/agent/harness/session/runtime.rs
+++ b/src/openhuman/agent/harness/session/runtime.rs
@@ -347,6 +347,15 @@ impl Agent {
         std::mem::take(&mut self.last_turn_citations)
     }
 
+    /// Drain and return the holistic token/cost/context totals for the latest
+    /// completed turn (parent + sub-agents). `None` until a turn has run.
+    /// Consumed by web-channel delivery to populate the `chat_done` usage fields.
+    pub fn take_last_turn_usage_totals(
+        &mut self,
+    ) -> Option<crate::openhuman::agent::harness::turn_subagent_usage::LastTurnUsage> {
+        self.last_turn_usage_totals.take()
+    }
+
     // ─────────────────────────────────────────────────────────────────
     // Static helpers for turn parsing + telemetry
     // ─────────────────────────────────────────────────────────────────
diff --git a/src/openhuman/agent/harness/session/transcript.rs b/src/openhuman/agent/harness/session/transcript.rs
index 9612344d8..4e293e382 100644
--- a/src/openhuman/agent/harness/session/transcript.rs
+++ b/src/openhuman/agent/harness/session/transcript.rs
@@ -54,7 +54,7 @@ use crate::openhuman::inference::provider::ChatMessage;
 use crate::openhuman::inference::provider::ToolCall;
 use anyhow::{Context, Result};
 use serde::{Deserialize, Serialize};
-use std::collections::HashMap;
+use std::collections::{BTreeMap, HashMap};
 use std::fmt::Write as FmtWrite;
 use std::fs;
 use std::path::{Path, PathBuf};
@@ -620,6 +620,206 @@ pub fn find_root_transcript_for_thread(workspace_dir: &Path, thread_id: &str) ->
     matches.pop()
 }
 
+/// Aggregated token/cost usage for a chat thread, summed across **all** of the
+/// thread's root session transcripts (a thread reopened across days/restarts
+/// produces several files). `last_turn_*`, `model`, and `updated` come from the
+/// newest transcript so the UI can render a context-window gauge for the most
+/// recent turn. Returns `None` when no transcript exists yet (a brand-new
+/// thread with no completed turns).
+#[derive(Debug, Clone, Default, PartialEq)]
+pub struct ThreadUsageSummary {
+    /// Orchestrator (parent) token totals — the root transcript(s) only. Root
+    /// transcripts never include sub-agent calls (those go to a separate
+    /// observer + their own `__` transcript files); see [`Self::subagents`].
+    pub input_tokens: u64,
+    pub output_tokens: u64,
+    pub cached_input_tokens: u64,
+    pub cost_usd: f64,
+    pub turn_count: usize,
+    /// Input/output tokens of the most recent assistant turn (context gauge).
+    pub last_turn_input_tokens: u64,
+    pub last_turn_output_tokens: u64,
+    /// Model that served the most recent turn, if recorded.
+    pub model: Option<String>,
+    /// RFC-3339 `updated` of the newest transcript.
+    pub updated: String,
+    /// Per-archetype sub-agent spend, reconstructed from the thread's `__`
+    /// sub-agent transcripts (grouped by `agent_name`).
+    pub subagents: Vec<SubagentArchetypeUsage>,
+}
+
+/// One sub-agent archetype's summed spend within a thread (e.g. all `coder`
+/// runs). `model` is the model that served one of its runs, used to price it.
+#[derive(Debug, Clone, Default, PartialEq)]
+pub struct SubagentArchetypeUsage {
+    pub agent_id: String,
+    pub input_tokens: u64,
+    pub output_tokens: u64,
+    pub cached_input_tokens: u64,
+    /// How many sub-agent runs of this archetype contributed.
+    pub runs: usize,
+    pub model: Option<String>,
+}
+
+/// Parse just the `_meta` header of a root transcript JSONL (cheap — stops at
+/// the first non-empty line).
+fn read_transcript_meta_only(path: &Path) -> Option<TranscriptMeta> {
+    let raw = fs::read_to_string(path).ok()?;
+    for line in raw.lines() {
+        let line = line.trim();
+        if line.is_empty() {
+            continue;
+        }
+        let ml: MetaLine = serde_json::from_str(line).ok()?;
+        let mp = ml.meta;
+        return Some(TranscriptMeta {
+            agent_name: mp.agent,
+            agent_id: mp.agent_id,
+            agent_type: mp.agent_type,
+            dispatcher: mp.dispatcher,
+            provider: mp.provider,
+            model: mp.model,
+            created: mp.created,
+            updated: mp.updated,
+            turn_count: mp.turn_count,
+            input_tokens: mp.input_tokens,
+            output_tokens: mp.output_tokens,
+            cached_input_tokens: mp.cached_input_tokens,
+            charged_amount_usd: mp.charged_amount_usd,
+            thread_id: mp.thread_id,
+            task_id: mp.task_id,
+        });
+    }
+    None
+}
+
+/// Extract the last assistant message's usage + model from a transcript JSONL.
+/// Only the final assistant message of a turn carries these (see the JSONL
+/// format docs at the top of this module).
+fn read_last_assistant_usage(path: &Path) -> Option<(MessageUsage, Option<String>)> {
+    let raw = fs::read_to_string(path).ok()?;
+    let mut result = None;
+    let mut seen_meta = false;
+    for line in raw.lines() {
+        let line = line.trim();
+        if line.is_empty() {
+            continue;
+        }
+        if !seen_meta {
+            seen_meta = true; // first non-empty line is the `_meta` header
+            continue;
+        }
+        if let Ok(ml) = serde_json::from_str::<MessageLine>(line) {
+            if ml.role == "assistant" {
+                if let Some(usage) = ml.usage {
+                    result = Some((usage, ml.model));
+                }
+            }
+        }
+    }
+    result
+}
+
+/// Summed token/cost usage for `thread_id` across its root transcripts, or
+/// `None` when the thread has no persisted turns yet.
+pub fn read_thread_usage_summary(
+    workspace_dir: &Path,
+    thread_id: &str,
+) -> Option<ThreadUsageSummary> {
+    let thread_id = thread_id.trim();
+    if thread_id.is_empty() {
+        return None;
+    }
+
+    let raw_dir = raw_session_dir(workspace_dir);
+    let entries = fs::read_dir(&raw_dir).ok()?;
+
+    // Single scan: split the thread's transcripts into root (orchestrator) and
+    // `__` sub-agent files. Root totals stay the parent's; sub-agent files are
+    // grouped by archetype for the per-agent breakdown.
+    let mut root_matches: Vec<PathBuf> = Vec::new();
+    let mut sub_matches: Vec<PathBuf> = Vec::new();
+    for path in entries.flatten().map(|entry| entry.path()) {
+        if path.extension().and_then(|s| s.to_str()) != Some("jsonl") {
+            continue;
+        }
+        let Some(stem) = path.file_stem().and_then(|s| s.to_str()) else {
+            continue;
+        };
+        let is_subagent = stem.contains("__");
+        let matches_thread = read_transcript_meta_only(&path)
+            .map(|m| m.thread_id.as_deref() == Some(thread_id))
+            .unwrap_or(false);
+        if !matches_thread {
+            continue;
+        }
+        if is_subagent {
+            sub_matches.push(path);
+        } else {
+            root_matches.push(path);
+        }
+    }
+
+    if root_matches.is_empty() && sub_matches.is_empty() {
+        return None;
+    }
+    root_matches.sort();
+
+    let mut summary = ThreadUsageSummary::default();
+    for path in &root_matches {
+        if let Some(meta) = read_transcript_meta_only(path) {
+            summary.input_tokens = summary.input_tokens.saturating_add(meta.input_tokens);
+            summary.output_tokens = summary.output_tokens.saturating_add(meta.output_tokens);
+            summary.cached_input_tokens = summary
+                .cached_input_tokens
+                .saturating_add(meta.cached_input_tokens);
+            summary.cost_usd += meta.charged_amount_usd;
+            summary.turn_count = summary.turn_count.saturating_add(meta.turn_count);
+        }
+    }
+
+    // Newest root transcript drives the last-turn gauge + model + updated stamp.
+    if let Some(newest) = root_matches.last() {
+        if let Some(meta) = read_transcript_meta_only(newest) {
+            summary.updated = meta.updated;
+        }
+        if let Some((usage, model)) = read_last_assistant_usage(newest) {
+            summary.last_turn_input_tokens = usage.input;
+            summary.last_turn_output_tokens = usage.output;
+            summary.model = model;
+        }
+    }
+
+    // Group sub-agent transcripts by archetype (`agent_name`).
+    let mut groups: BTreeMap<String, SubagentArchetypeUsage> = BTreeMap::new();
+    for path in &sub_matches {
+        let Some(meta) = read_transcript_meta_only(path) else {
+            continue;
+        };
+        let group =
+            groups
+                .entry(meta.agent_name.clone())
+                .or_insert_with(|| SubagentArchetypeUsage {
+                    agent_id: meta.agent_name.clone(),
+                    ..Default::default()
+                });
+        group.input_tokens = group.input_tokens.saturating_add(meta.input_tokens);
+        group.output_tokens = group.output_tokens.saturating_add(meta.output_tokens);
+        group.cached_input_tokens = group
+            .cached_input_tokens
+            .saturating_add(meta.cached_input_tokens);
+        group.runs = group.runs.saturating_add(1);
+        if group.model.is_none() {
+            if let Some((_, model)) = read_last_assistant_usage(path) {
+                group.model = model;
+            }
+        }
+    }
+    summary.subagents = groups.into_values().collect();
+
+    Some(summary)
+}
+
 // ── Path resolution ──────────────────────────────────────────────────
 
 /// Resolve a transcript path under `session_raw/{stem}.jsonl` — a
diff --git a/src/openhuman/agent/harness/session/transcript_tests.rs b/src/openhuman/agent/harness/session/transcript_tests.rs
index f01d1cb6e..cb35c7f89 100644
--- a/src/openhuman/agent/harness/session/transcript_tests.rs
+++ b/src/openhuman/agent/harness/session/transcript_tests.rs
@@ -821,3 +821,158 @@ fn find_root_transcript_for_thread_excludes_subagent_files() {
         found.display()
     );
 }
+
+#[test]
+fn read_thread_usage_summary_totals_last_turn_and_model() {
+    let ws = TempDir::new().unwrap();
+    let raw = raw_session_dir(ws.path());
+    std::fs::create_dir_all(&raw).unwrap();
+
+    let mut meta = sample_meta();
+    meta.thread_id = Some("thr-xyz".into());
+    meta.input_tokens = 5000;
+    meta.output_tokens = 1200;
+    meta.cached_input_tokens = 800;
+    meta.charged_amount_usd = 0.0045;
+    meta.turn_count = 3;
+
+    let tu = TurnUsage {
+        provider: "openhuman-backend".into(),
+        model: "reasoning-v1".into(),
+        usage: MessageUsage {
+            input: 400,
+            output: 120,
+            cached_input: 50,
+            context_window: 1_000_000,
+            cost_usd: 0.0009,
+        },
+        ts: "2026-04-11T14:35:22Z".into(),
+        reasoning_content: None,
+        tool_calls: Vec::new(),
+        iteration: 0,
+    };
+    let path = raw.join("1700000000_main.jsonl");
+    write_transcript(&path, &sample_messages(), &meta, Some(&tu)).unwrap();
+
+    let summary = read_thread_usage_summary(ws.path(), "thr-xyz").expect("summary present");
+    assert_eq!(summary.input_tokens, 5000);
+    assert_eq!(summary.output_tokens, 1200);
+    assert_eq!(summary.cached_input_tokens, 800);
+    assert!((summary.cost_usd - 0.0045).abs() < 1e-9);
+    assert_eq!(summary.turn_count, 3);
+    assert_eq!(summary.last_turn_input_tokens, 400);
+    assert_eq!(summary.last_turn_output_tokens, 120);
+    assert_eq!(summary.model.as_deref(), Some("reasoning-v1"));
+}
+
+#[test]
+fn read_thread_usage_summary_sums_multiple_transcripts() {
+    let ws = TempDir::new().unwrap();
+    let raw = raw_session_dir(ws.path());
+    std::fs::create_dir_all(&raw).unwrap();
+
+    let mut mk = |stem: &str, input: u64, cost: f64| {
+        let mut meta = sample_meta();
+        meta.thread_id = Some("thr-multi".into());
+        meta.input_tokens = input;
+        meta.output_tokens = 0;
+        meta.cached_input_tokens = 0;
+        meta.charged_amount_usd = cost;
+        meta.turn_count = 1;
+        write_transcript(
+            &raw.join(format!("{stem}.jsonl")),
+            &sample_messages(),
+            &meta,
+            None,
+        )
+        .unwrap();
+    };
+    mk("1700000000_main", 100, 0.01);
+    mk("1700000100_main", 250, 0.02);
+
+    let s = read_thread_usage_summary(ws.path(), "thr-multi").expect("summary present");
+    assert_eq!(s.input_tokens, 350);
+    assert!((s.cost_usd - 0.03).abs() < 1e-9);
+    assert_eq!(s.turn_count, 2);
+}
+
+#[test]
+fn read_thread_usage_summary_none_for_unknown_thread() {
+    let ws = TempDir::new().unwrap();
+    assert!(read_thread_usage_summary(ws.path(), "no-such-thread").is_none());
+    // Empty thread id is rejected too.
+    assert!(read_thread_usage_summary(ws.path(), "   ").is_none());
+}
+
+#[test]
+fn read_thread_usage_summary_groups_subagents_by_archetype() {
+    let ws = TempDir::new().unwrap();
+    let raw = raw_session_dir(ws.path());
+    std::fs::create_dir_all(&raw).unwrap();
+
+    // Root (orchestrator) transcript — never includes sub-agent calls.
+    let mut root = sample_meta();
+    root.thread_id = Some("thr-sub".into());
+    root.agent_name = "main".into();
+    root.input_tokens = 1000;
+    root.output_tokens = 200;
+    root.cached_input_tokens = 0;
+    root.charged_amount_usd = 0.0;
+    root.turn_count = 2;
+    write_transcript(
+        &raw.join("1700000000_main.jsonl"),
+        &sample_messages(),
+        &root,
+        None,
+    )
+    .unwrap();
+
+    // Sub-agent transcripts (stems contain `__`): coder x2 + researcher x1.
+    let mut sub = |stem: &str, agent: &str, input: u64, output: u64| {
+        let mut m = sample_meta();
+        m.thread_id = Some("thr-sub".into());
+        m.agent_name = agent.into();
+        m.input_tokens = input;
+        m.output_tokens = output;
+        m.cached_input_tokens = 0;
+        m.charged_amount_usd = 0.0;
+        m.turn_count = 1;
+        write_transcript(
+            &raw.join(format!("{stem}.jsonl")),
+            &sample_messages(),
+            &m,
+            None,
+        )
+        .unwrap();
+    };
+    sub("1700000000_main__1700000001_coder", "coder", 300, 60);
+    sub("1700000000_main__1700000002_coder", "coder", 100, 20);
+    sub(
+        "1700000000_main__1700000003_researcher",
+        "researcher",
+        500,
+        90,
+    );
+
+    let s = read_thread_usage_summary(ws.path(), "thr-sub").expect("summary present");
+    // Root totals are orchestrator-only (sub-agents are separate).
+    assert_eq!(s.input_tokens, 1000);
+    assert_eq!(s.output_tokens, 200);
+    // Grouped by archetype.
+    assert_eq!(s.subagents.len(), 2);
+    let coder = s
+        .subagents
+        .iter()
+        .find(|g| g.agent_id == "coder")
+        .expect("coder group");
+    assert_eq!(coder.input_tokens, 400);
+    assert_eq!(coder.output_tokens, 80);
+    assert_eq!(coder.runs, 2);
+    let researcher = s
+        .subagents
+        .iter()
+        .find(|g| g.agent_id == "researcher")
+        .expect("researcher group");
+    assert_eq!(researcher.input_tokens, 500);
+    assert_eq!(researcher.runs, 1);
+}
diff --git a/src/openhuman/agent/harness/session/turn/core.rs b/src/openhuman/agent/harness/session/turn/core.rs
index 6a02c18cc..35318e395 100644
--- a/src/openhuman/agent/harness/session/turn/core.rs
+++ b/src/openhuman/agent/harness/session/turn/core.rs
@@ -803,7 +803,8 @@ impl Agent {
                 crate::openhuman::agent::multimodal::extract_image_placeholders_in_text(
                     user_message,
                 );
-            let outcome =
+            let (outcome_result, subagent_usage_entries) =
+                super::super::super::turn_subagent_usage::with_turn_collector(
                 super::super::super::turn_attachments_context::with_current_turn_image_placeholders(
                     turn_image_placeholders,
                     super::super::super::model_vision_context::with_current_model_vision(
@@ -829,18 +830,64 @@ impl Agent {
                     None, // main agent compacts via its ContextManager in before_dispatch
                         )),
                     ),
+                ),
                 )
-                .await?;
+                .await;
+            let outcome = outcome_result?;
 
             // Pull the observer's accounting out, then drop it to release the
             // `&mut self` borrow so the epilogue can use `self`.
             let did_push_final = observer.did_push_final;
-            let cumulative_input = observer.cumulative_input;
-            let cumulative_output = observer.cumulative_output;
-            let cumulative_cached = observer.cumulative_cached;
-            let cumulative_charged = observer.cumulative_charged;
+            let mut cumulative_input = observer.cumulative_input;
+            let mut cumulative_output = observer.cumulative_output;
+            let mut cumulative_cached = observer.cumulative_cached;
+            let mut cumulative_charged = observer.cumulative_charged;
             let last_turn_usage = observer.last_turn_usage.take();
             drop(observer);
+
+            // Roll any sub-agent spend gathered during this turn into the
+            // session-level token/cost meters so the UI footer reflects the
+            // *holistic* cost (parent + delegated children). The global cost
+            // tracker is fed separately, per provider call, by each sub-agent's
+            // observer. `subagent_usage_entries` is also forwarded to the
+            // `chat_done` event for the per-child hover breakdown.
+            if !subagent_usage_entries.is_empty() {
+                let mut sub_input = 0u64;
+                let mut sub_output = 0u64;
+                let mut sub_cached = 0u64;
+                let mut sub_charged = 0.0f64;
+                for entry in &subagent_usage_entries {
+                    sub_input += entry.usage.input_tokens;
+                    sub_output += entry.usage.output_tokens;
+                    sub_cached += entry.usage.cached_input_tokens;
+                    sub_charged += entry.usage.charged_amount_usd;
+                }
+                tracing::debug!(
+                    subagents = subagent_usage_entries.len(),
+                    sub_input,
+                    sub_output,
+                    sub_charged,
+                    "[agent_loop] folding sub-agent spend into turn totals"
+                );
+                cumulative_input += sub_input;
+                cumulative_output += sub_output;
+                cumulative_cached += sub_cached;
+                cumulative_charged += sub_charged;
+            }
+
+            // Capture the turn's holistic totals (parent + sub-agents) so the
+            // web-channel delivery layer can forward them on `chat_done` for the
+            // UI footer's session token / cost / context meters.
+            self.last_turn_usage_totals = Some(
+                crate::openhuman::agent::harness::turn_subagent_usage::LastTurnUsage {
+                    input_tokens: cumulative_input,
+                    output_tokens: cumulative_output,
+                    cached_input_tokens: cumulative_cached,
+                    cost_usd: cumulative_charged,
+                    context_window: turn_context_window.unwrap_or(0),
+                    subagents: subagent_usage_entries,
+                },
+            );
             let records = std::mem::take(&mut tool_source.records);
 
             self.context.record_tool_calls(records.len());
diff --git a/src/openhuman/agent/harness/session/turn_engine_adapter.rs b/src/openhuman/agent/harness/session/turn_engine_adapter.rs
index 5ecf274cf..8cd8c5511 100644
--- a/src/openhuman/agent/harness/session/turn_engine_adapter.rs
+++ b/src/openhuman/agent/harness/session/turn_engine_adapter.rs
@@ -309,10 +309,17 @@ impl TurnObserver for AgentObserver<'_> {
     fn record_usage(&mut self, provider: &str, model: &str, usage: &UsageInfo) {
         self.agent.context.record_usage(usage);
         crate::openhuman::cost::record_provider_usage(model, usage);
+        // Effective per-call cost: the backend-charged amount when the provider
+        // echoes one, else the per-model catalog estimate (#4124). Using this
+        // instead of raw `charged_amount_usd` means BYO/local providers that
+        // never bill a charge still contribute a priced cost to the session
+        // total. `cumulative_charged` is therefore the *net cost*, not strictly
+        // the backend charge.
+        let call_cost = crate::openhuman::agent::cost::call_cost_usd(model, usage);
         self.cumulative_input += usage.input_tokens;
         self.cumulative_output += usage.output_tokens;
         self.cumulative_cached += usage.cached_input_tokens;
-        self.cumulative_charged += usage.charged_amount_usd;
+        self.cumulative_charged += call_cost;
         self.last_turn_usage = Some(transcript::TurnUsage {
             provider: provider.to_string(),
             model: model.to_string(),
@@ -321,7 +328,7 @@ impl TurnObserver for AgentObserver<'_> {
                 output: usage.output_tokens,
                 cached_input: usage.cached_input_tokens,
                 context_window: usage.context_window,
-                cost_usd: usage.charged_amount_usd,
+                cost_usd: call_cost,
             },
             ts: chrono::Utc::now().to_rfc3339(),
             reasoning_content: None,
diff --git a/src/openhuman/agent/harness/session/types.rs b/src/openhuman/agent/harness/session/types.rs
index c950ff879..82de2ee30 100644
--- a/src/openhuman/agent/harness/session/types.rs
+++ b/src/openhuman/agent/harness/session/types.rs
@@ -70,6 +70,12 @@ pub struct Agent {
     /// Citation metadata collected from memory recall for the most recent turn.
     /// Consumed by web-channel delivery to render source chips in the UI.
     pub(super) last_turn_citations: Vec<crate::openhuman::agent::memory_loader::MemoryCitation>,
+    /// Holistic token/cost/context accounting for the most recent turn (parent +
+    /// any sub-agents spawned during it). Consumed by web-channel delivery to
+    /// surface session token/cost/context meters in the UI footer. `None` until
+    /// the first turn completes.
+    pub(super) last_turn_usage_totals:
+        Option<crate::openhuman::agent::harness::turn_subagent_usage::LastTurnUsage>,
     pub(super) history: Vec<ConversationMessage>,
     pub(super) post_turn_hooks: Vec<Arc<dyn PostTurnHook>>,
     pub(super) learning_enabled: bool,
diff --git a/src/openhuman/agent/harness/subagent_runner/mod.rs b/src/openhuman/agent/harness/subagent_runner/mod.rs
index 734e6f9f4..3c3a7c423 100644
--- a/src/openhuman/agent/harness/subagent_runner/mod.rs
+++ b/src/openhuman/agent/harness/subagent_runner/mod.rs
@@ -42,7 +42,7 @@ pub use autonomous::{autonomous_iter_cap, with_autonomous_iter_cap};
 pub use ops::run_subagent;
 pub use types::{
     SubagentCheckpointData, SubagentMode, SubagentRunError, SubagentRunOptions, SubagentRunOutcome,
-    SubagentRunStatus,
+    SubagentRunStatus, SubagentUsage,
 };
 
 // Crate-internal re-exports: `agent::debug` calls the text-mode protocol
diff --git a/src/openhuman/agent/harness/subagent_runner/ops/observer.rs b/src/openhuman/agent/harness/subagent_runner/ops/observer.rs
index f8cf5b016..4fcfec396 100644
--- a/src/openhuman/agent/harness/subagent_runner/ops/observer.rs
+++ b/src/openhuman/agent/harness/subagent_runner/ops/observer.rs
@@ -18,6 +18,9 @@ pub(super) struct SubagentObserver {
     pub(super) task_id: String,
     pub(super) force_text_mode: bool,
     pub(super) usage: AggregatedUsage,
+    /// Provenance + usage of the sub-agent's most recent provider call. Persisted
+    /// onto the transcript's last assistant message (carries provider + model so
+    /// per-thread usage reads can price the sub-agent at its actual model).
     pub(super) last_turn_usage: Option<transcript::TurnUsage>,
 }
 
@@ -109,22 +112,36 @@ impl super::super::super::engine::TurnObserver for SubagentObserver {
     fn record_usage(
         &mut self,
         provider: &str,
-        _model: &str,
+        model: &str,
         usage: &crate::openhuman::inference::provider::UsageInfo,
     ) {
         self.usage.input_tokens += usage.input_tokens;
         self.usage.output_tokens += usage.output_tokens;
         self.usage.cached_input_tokens += usage.cached_input_tokens;
-        self.usage.charged_amount_usd += usage.charged_amount_usd;
+        // Effective per-call cost: backend-charged when present, else the
+        // per-model catalog estimate (#4124) — so a sub-agent on a BYO/local
+        // provider that bills no charge still contributes a priced cost to the
+        // parent turn's net total. `charged_amount_usd` here is the *net cost*,
+        // matching the parent adapter's convention.
+        let call_cost = crate::openhuman::agent::cost::call_cost_usd(model, usage);
+        self.usage.charged_amount_usd += call_cost;
+        // Mirror the parent adapter: feed every sub-agent provider call into the
+        // global cost tracker so the cost dashboard/summary reflects delegated
+        // spend (previously sub-agent tokens were invisible to it). The per-turn
+        // rollup into the UI footer happens separately via `turn_subagent_usage`.
+        crate::openhuman::cost::record_provider_usage(model, usage);
+        // Provenance for the transcript's last assistant message (#4134): keeps
+        // the provider + model so per-thread usage reads can price the sub-agent
+        // at its actual model.
         self.last_turn_usage = Some(transcript::TurnUsage {
             provider: provider.to_string(),
-            model: _model.to_string(),
+            model: model.to_string(),
             usage: transcript::MessageUsage {
                 input: usage.input_tokens,
                 output: usage.output_tokens,
                 cached_input: usage.cached_input_tokens,
                 context_window: usage.context_window,
-                cost_usd: usage.charged_amount_usd,
+                cost_usd: call_cost,
             },
             ts: chrono::Utc::now().to_rfc3339(),
             reasoning_content: None,
diff --git a/src/openhuman/agent/harness/subagent_runner/ops/runner.rs b/src/openhuman/agent/harness/subagent_runner/ops/runner.rs
index 957433eac..0bfe3dd50 100644
--- a/src/openhuman/agent/harness/subagent_runner/ops/runner.rs
+++ b/src/openhuman/agent/harness/subagent_runner/ops/runner.rs
@@ -794,7 +794,7 @@ async fn run_typed_mode(
         model_vision,
         "[subagent_runner] resolved sub-agent model vision capability"
     );
-    let (output, iterations, _agg_usage, early_exit_tool) = Box::pin(run_inner_loop(
+    let (output, iterations, agg_usage, early_exit_tool) = Box::pin(run_inner_loop(
         subagent_provider.as_ref(),
         &mut history,
         &parent.all_tools,
@@ -884,6 +884,22 @@ async fn run_typed_mode(
         crate::openhuman::agent::harness::subagent_runner::types::SubagentRunStatus::Completed
     };
 
+    // Surface this run's token/cost totals so the parent turn can roll them
+    // into the session-level meters and the global cost tracker. Also push the
+    // breakdown into any active turn-scoped collector (see
+    // `turn_subagent_usage`) so a delegating parent attributes per-child spend.
+    let usage = crate::openhuman::agent::harness::subagent_runner::types::SubagentUsage {
+        input_tokens: agg_usage.input_tokens,
+        output_tokens: agg_usage.output_tokens,
+        cached_input_tokens: agg_usage.cached_input_tokens,
+        charged_amount_usd: agg_usage.charged_amount_usd,
+    };
+    crate::openhuman::agent::harness::turn_subagent_usage::record_subagent_usage(
+        task_id,
+        &definition.id,
+        usage,
+    );
+
     Ok(SubagentRunOutcome {
         task_id: task_id.to_string(),
         agent_id: definition.id.clone(),
@@ -893,5 +909,6 @@ async fn run_typed_mode(
         mode: SubagentMode::Typed,
         status,
         final_history: history,
+        usage,
     })
 }
diff --git a/src/openhuman/agent/harness/subagent_runner/types.rs b/src/openhuman/agent/harness/subagent_runner/types.rs
index f29073a55..c03a4cb5c 100644
--- a/src/openhuman/agent/harness/subagent_runner/types.rs
+++ b/src/openhuman/agent/harness/subagent_runner/types.rs
@@ -112,6 +112,24 @@ pub struct SubagentRunOutcome {
     /// sessions persist this so an idle worker can resume without rebuilding
     /// its context from only the parent transcript.
     pub final_history: Vec<ChatMessage>,
+    /// Token + cost accounting accumulated across every provider call this
+    /// sub-agent made. Surfaced so the parent turn can roll child spend into
+    /// the session totals (tokens + USD) and the global cost tracker. See
+    /// [`SubagentUsage`].
+    pub usage: SubagentUsage,
+}
+
+/// Token + cost totals for a single sub-agent run.
+///
+/// Mirrors the inner-loop `AggregatedUsage`, lifted into the public outcome so
+/// the parent turn can fold sub-agent spend into the session-level token/cost
+/// meters surfaced in the UI footer.
+#[derive(Debug, Clone, Copy, Default, PartialEq)]
+pub struct SubagentUsage {
+    pub input_tokens: u64,
+    pub output_tokens: u64,
+    pub cached_input_tokens: u64,
+    pub charged_amount_usd: f64,
 }
 
 /// Which prompt-construction path the runner took for a sub-agent.
diff --git a/src/openhuman/agent/harness/turn_subagent_usage.rs b/src/openhuman/agent/harness/turn_subagent_usage.rs
new file mode 100644
index 000000000..ba32b9832
--- /dev/null
+++ b/src/openhuman/agent/harness/turn_subagent_usage.rs
@@ -0,0 +1,174 @@
+//! Task-local collector for **sub-agent token/cost spend within the current
+//! turn**.
+//!
+//! When a parent agent delegates to a sub-agent mid-turn, the sub-agent's
+//! provider calls are accounted by its own [`SubagentObserver`] and never reach
+//! the parent's `AgentObserver` cumulative totals (the two run on separate
+//! observers). Without a bridge, sub-agent spend is invisible to the
+//! session-level token/cost meters surfaced in the UI footer, and to the global
+//! cost tracker.
+//!
+//! This module installs an [`Arc<Mutex<Vec<SubagentUsageEntry>>>`] as a
+//! task-local around the parent's `run_turn_engine` call. Synchronous
+//! delegations (`spawn_subagent`) run inline on the same tokio task, so the
+//! sub-agent runner can [`record_subagent_usage`] its totals into the active
+//! collector. After the engine returns, the parent [`drain`]s the collector to:
+//!
+//! 1. fold child tokens + USD into the turn's cumulative meters, and
+//! 2. attribute per-child spend for the `chat_done` breakdown (hover detail).
+//!
+//! Background / async sub-agents (`spawn_async_subagent`) run on detached tasks
+//! that do **not** inherit this task-local; their spend completes after the
+//! parent turn's `chat_done` and is captured by the global cost tracker
+//! instead. [`current_collector`] returns `None` outside any scope (CLI /
+//! direct invocation / tests), so recording is strictly additive.
+
+use std::sync::{Arc, Mutex};
+
+use super::subagent_runner::SubagentUsage;
+
+/// One sub-agent's spend, tagged with its identity for the per-child breakdown.
+#[derive(Debug, Clone, PartialEq)]
+pub struct SubagentUsageEntry {
+    pub task_id: String,
+    pub agent_id: String,
+    pub usage: SubagentUsage,
+}
+
+/// Holistic token/cost accounting for a single completed turn, including any
+/// sub-agent spend rolled in. Captured on the session at turn end and consumed
+/// by the web-channel delivery layer, which forwards it on the `chat_done`
+/// event so the UI footer can show session tokens, context-window utilisation,
+/// USD cost, and a per-sub-agent hover breakdown.
+#[derive(Debug, Clone, Default, PartialEq)]
+pub struct LastTurnUsage {
+    /// Input (prompt) tokens for the turn, parent + sub-agents.
+    pub input_tokens: u64,
+    /// Output (completion) tokens for the turn, parent + sub-agents.
+    pub output_tokens: u64,
+    /// Cached-input tokens for the turn, parent + sub-agents.
+    pub cached_input_tokens: u64,
+    /// USD cost for the turn (backend-charged where available, else estimated),
+    /// parent + sub-agents.
+    pub cost_usd: f64,
+    /// The model's context window for this turn (`0` when unknown, e.g. a cloud
+    /// model whose window the core couldn't resolve). Lets the UI show real
+    /// context utilisation instead of a hard-coded default.
+    pub context_window: u64,
+    /// Per-sub-agent spend gathered during the turn, for the hover breakdown.
+    pub subagents: Vec<SubagentUsageEntry>,
+}
+
+/// Shared, mutable list of sub-agent spend gathered during one parent turn.
+pub type TurnSubagentUsage = Arc<Mutex<Vec<SubagentUsageEntry>>>;
+
+tokio::task_local! {
+    /// Active per-turn sub-agent usage collector, installed around the parent's
+    /// `run_turn_engine` call. Absent outside a turn scope.
+    static TURN_SUBAGENT_USAGE: TurnSubagentUsage;
+}
+
+/// The collector active for the current turn, or `None` when no scope is in
+/// effect (e.g. a sub-agent running on a detached background task, or a direct
+/// CLI invocation).
+pub fn current_collector() -> Option<TurnSubagentUsage> {
+    TURN_SUBAGENT_USAGE.try_with(|c| c.clone()).ok()
+}
+
+/// Record a finished sub-agent's token/cost totals into the active turn
+/// collector. No-op when there is no active scope. Called by the sub-agent
+/// runner once it has the run's aggregated usage.
+pub fn record_subagent_usage(task_id: &str, agent_id: &str, usage: SubagentUsage) {
+    let Some(collector) = current_collector() else {
+        tracing::trace!(
+            task_id,
+            agent_id,
+            "[turn_subagent_usage] no active collector — sub-agent spend not rolled into parent turn"
+        );
+        return;
+    };
+    let entry = SubagentUsageEntry {
+        task_id: task_id.to_string(),
+        agent_id: agent_id.to_string(),
+        usage,
+    };
+    // Recover the inner vec on poison (a panic in another sub-agent) rather than
+    // dropping this turn's accounting entirely.
+    let mut guard = collector.lock().unwrap_or_else(|poisoned| {
+        tracing::warn!(
+            task_id,
+            agent_id,
+            "[turn_subagent_usage] collector mutex poisoned — recovering"
+        );
+        poisoned.into_inner()
+    });
+    tracing::debug!(
+        task_id,
+        agent_id,
+        input_tokens = usage.input_tokens,
+        output_tokens = usage.output_tokens,
+        charged_usd = usage.charged_amount_usd,
+        "[turn_subagent_usage] recorded sub-agent spend into parent turn"
+    );
+    guard.push(entry);
+}
+
+/// Run `future` with a fresh sub-agent usage collector installed, returning both
+/// the future's output and the gathered per-child entries. Intended call site is
+/// around the parent agent's `run_turn_engine` invocation.
+pub async fn with_turn_collector<F, R>(future: F) -> (R, Vec<SubagentUsageEntry>)
+where
+    F: std::future::Future<Output = R>,
+{
+    let collector: TurnSubagentUsage = Arc::new(Mutex::new(Vec::new()));
+    let out = TURN_SUBAGENT_USAGE.scope(collector.clone(), future).await;
+    let entries = collector
+        .lock()
+        .map(|g| g.clone())
+        .unwrap_or_else(|p| p.into_inner().clone());
+    (out, entries)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn usage(input: u64, output: u64, usd: f64) -> SubagentUsage {
+        SubagentUsage {
+            input_tokens: input,
+            output_tokens: output,
+            cached_input_tokens: 0,
+            charged_amount_usd: usd,
+        }
+    }
+
+    #[tokio::test]
+    async fn no_collector_outside_scope() {
+        assert!(current_collector().is_none());
+        // Must not panic when there is nothing to record into.
+        record_subagent_usage("t1", "researcher", usage(10, 5, 0.01));
+    }
+
+    #[tokio::test]
+    async fn collects_entries_within_scope() {
+        let ((), entries) = with_turn_collector(async {
+            record_subagent_usage("t1", "researcher", usage(10, 5, 0.01));
+            record_subagent_usage("t2", "coder", usage(20, 8, 0.02));
+        })
+        .await;
+        assert_eq!(entries.len(), 2);
+        assert_eq!(entries[0].task_id, "t1");
+        assert_eq!(entries[0].usage.input_tokens, 10);
+        assert_eq!(entries[1].agent_id, "coder");
+        assert_eq!(entries[1].usage.charged_amount_usd, 0.02);
+    }
+
+    #[tokio::test]
+    async fn scope_does_not_leak() {
+        let _ = with_turn_collector(async {
+            record_subagent_usage("t1", "researcher", usage(1, 1, 0.0));
+        })
+        .await;
+        assert!(current_collector().is_none());
+    }
+}
diff --git a/src/openhuman/agent/task_dispatcher/executor.rs b/src/openhuman/agent/task_dispatcher/executor.rs
index 46c28816e..9a478e3e2 100644
--- a/src/openhuman/agent/task_dispatcher/executor.rs
+++ b/src/openhuman/agent/task_dispatcher/executor.rs
@@ -227,6 +227,9 @@ pub(super) async fn run_autonomous(
                     response,
                     prompt,
                     &[],
+                    // Background/cron turns don't surface in the chat footer; their
+                    // token/cost spend is still captured by the global cost tracker.
+                    None,
                 )
                 .await;
             }
diff --git a/src/openhuman/agent_orchestration/tools/spawn_subagent.rs b/src/openhuman/agent_orchestration/tools/spawn_subagent.rs
index 492b1db76..edf7e69c6 100644
--- a/src/openhuman/agent_orchestration/tools/spawn_subagent.rs
+++ b/src/openhuman/agent_orchestration/tools/spawn_subagent.rs
@@ -859,6 +859,7 @@ mod tests {
             mode: SubagentMode::Typed,
             status: SubagentRunStatus::Completed,
             final_history: Vec::new(),
+            usage: Default::default(),
         }
     }
 
diff --git a/src/openhuman/channels/proactive.rs b/src/openhuman/channels/proactive.rs
index d94d4481c..f8d7c56f8 100644
--- a/src/openhuman/channels/proactive.rs
+++ b/src/openhuman/channels/proactive.rs
@@ -215,6 +215,7 @@ impl EventHandler for ProactiveMessageSubscriber {
             task_board: None,
             tool_display_label: None,
             tool_display_detail: None,
+            usage: None,
         });
 
         // 2. If an active external channel is configured, deliver there too.
diff --git a/src/openhuman/channels/providers/presentation.rs b/src/openhuman/channels/providers/presentation.rs
index a5d475095..385b0178e 100644
--- a/src/openhuman/channels/providers/presentation.rs
+++ b/src/openhuman/channels/providers/presentation.rs
@@ -9,7 +9,8 @@
 //! 2. **Emoji reactions** — decide whether the assistant should react to the
 //!    user's message with an emoji.
 
-use crate::core::socketio::WebChannelEvent;
+use crate::core::socketio::{SubagentUsagePayload, TurnUsagePayload, WebChannelEvent};
+use crate::openhuman::agent::harness::turn_subagent_usage::LastTurnUsage;
 use crate::openhuman::config::rpc as config_rpc;
 
 use super::web::publish_web_channel_event;
@@ -17,6 +18,32 @@ use super::web::publish_web_channel_event;
 const MIN_SEGMENT_CHARS: usize = 40;
 const MAX_SEGMENTS: usize = 5;
 
+/// Convert a turn's [`LastTurnUsage`] into the wire payload carried on
+/// `chat_done`. Returns `None` for a turn that recorded no spend at all (e.g. a
+/// synthetic budget-exhausted placeholder) so the event stays compact.
+fn usage_payload(usage: Option<&LastTurnUsage>) -> Option<TurnUsagePayload> {
+    let usage = usage?;
+    let subagents = usage
+        .subagents
+        .iter()
+        .map(|s| SubagentUsagePayload {
+            task_id: s.task_id.clone(),
+            agent_id: s.agent_id.clone(),
+            input_tokens: s.usage.input_tokens,
+            output_tokens: s.usage.output_tokens,
+            cost_usd: s.usage.charged_amount_usd,
+        })
+        .collect();
+    Some(TurnUsagePayload {
+        input_tokens: usage.input_tokens,
+        output_tokens: usage.output_tokens,
+        cached_input_tokens: usage.cached_input_tokens,
+        cost_usd: usage.cost_usd,
+        context_window: usage.context_window,
+        subagents,
+    })
+}
+
 /// Deliver an agent response to the frontend, applying local-model
 /// presentation (segmentation + reaction) when the model is available.
 ///
@@ -30,7 +57,10 @@ pub async fn deliver_response(
     full_response: &str,
     user_message: &str,
     citations: &[crate::openhuman::agent::memory_loader::MemoryCitation],
+    usage: Option<&LastTurnUsage>,
 ) {
+    let usage_payload = usage_payload(usage);
+
     // Spawn reaction decision in parallel — it runs on the local model and
     // shouldn't block segmentation or delivery.
     let user_msg_owned = user_message.to_string();
@@ -78,6 +108,7 @@ pub async fn deliver_response(
             } else {
                 Some(serde_json::json!(citations))
             },
+            usage: usage_payload,
         });
         return;
     }
@@ -126,6 +157,8 @@ pub async fn deliver_response(
             } else {
                 None
             },
+            // Usage is attached only to the terminal `chat_done`, never segments.
+            usage: None,
         });
     }
 
@@ -164,6 +197,7 @@ pub async fn deliver_response(
         } else {
             Some(serde_json::json!(citations))
         },
+        usage: usage_payload,
     });
 }
 
@@ -466,6 +500,7 @@ pub mod test_support {
             full_response,
             user_message,
             citations,
+            None,
         )
         .await;
     }
diff --git a/src/openhuman/channels/providers/web/ops.rs b/src/openhuman/channels/providers/web/ops.rs
index 69a098bf0..3b2626800 100644
--- a/src/openhuman/channels/providers/web/ops.rs
+++ b/src/openhuman/channels/providers/web/ops.rs
@@ -609,6 +609,7 @@ pub async fn start_chat(
                     &chat_result.full_response,
                     &user_message,
                     &chat_result.citations,
+                    chat_result.usage.as_ref(),
                 )
                 .await;
             }
@@ -813,6 +814,7 @@ async fn spawn_parallel_turn(
                     &chat_result.full_response,
                     &user_message,
                     &chat_result.citations,
+                    chat_result.usage.as_ref(),
                 )
                 .await;
             }
diff --git a/src/openhuman/channels/providers/web/run_task.rs b/src/openhuman/channels/providers/web/run_task.rs
index 6b5b23fed..c7dd277df 100644
--- a/src/openhuman/channels/providers/web/run_task.rs
+++ b/src/openhuman/channels/providers/web/run_task.rs
@@ -237,9 +237,11 @@ pub(crate) async fn run_chat_task(
             // later genuine empty response. See #3386.
             super::ops::clear_budget_signal(thread_id).await;
             let citations = agent.take_last_turn_citations();
+            let usage = agent.take_last_turn_usage_totals();
             Ok(WebChatTaskResult {
                 full_response: response,
                 citations,
+                usage,
             })
         }
         Err(err) => {
@@ -269,6 +271,7 @@ pub(crate) async fn run_chat_task(
                     Ok(WebChatTaskResult {
                         full_response: inference_budget_exceeded_user_message().to_string(),
                         citations: Vec::new(),
+                        usage: None,
                     })
                 }
                 BudgetCorrelation::UpgradeEmptyToBudget => {
@@ -286,6 +289,7 @@ pub(crate) async fn run_chat_task(
                     Ok(WebChatTaskResult {
                         full_response: inference_budget_exceeded_user_message().to_string(),
                         citations: Vec::new(),
+                        usage: None,
                     })
                 }
                 BudgetCorrelation::PassThrough => Err(err_message),
@@ -408,6 +412,7 @@ mod tests {
         Ok(WebChatTaskResult {
             full_response: "hello".to_string(),
             citations: Vec::new(),
+            usage: None,
         })
     }
 
diff --git a/src/openhuman/channels/providers/web/types.rs b/src/openhuman/channels/providers/web/types.rs
index e8f8b936a..3cc666687 100644
--- a/src/openhuman/channels/providers/web/types.rs
+++ b/src/openhuman/channels/providers/web/types.rs
@@ -68,6 +68,10 @@ pub(super) struct ParallelEntry {
 pub(super) struct WebChatTaskResult {
     pub(super) full_response: String,
     pub(super) citations: Vec<crate::openhuman::agent::memory_loader::MemoryCitation>,
+    /// Holistic token/cost/context totals for the turn (parent + sub-agents),
+    /// forwarded to the frontend on `chat_done`. `None` for synthetic results
+    /// (e.g. budget-exhausted placeholders) that never ran a real turn.
+    pub(super) usage: Option<crate::openhuman::agent::harness::turn_subagent_usage::LastTurnUsage>,
 }
 
 /// Per-request metadata carried alongside a chat send. Currently used by the
diff --git a/src/openhuman/inference/model_context.rs b/src/openhuman/inference/model_context.rs
index de9af34b8..4da1b814a 100644
--- a/src/openhuman/inference/model_context.rs
+++ b/src/openhuman/inference/model_context.rs
@@ -11,6 +11,8 @@ use crate::openhuman::config::{
 
 /// Conservative default for OpenHuman abstract tier models (tokens).
 const TIER_LARGE_CONTEXT: u64 = 200_000;
+/// Reasoning tier — backed by a 1M-context model.
+const TIER_REASONING_CONTEXT: u64 = 1_000_000;
 const TIER_STANDARD_CONTEXT: u64 = 128_000;
 const TIER_LOCAL_CONTEXT: u64 = 8_192;
 
@@ -121,7 +123,8 @@ pub fn context_window_for_model(model: &str) -> Option<u64> {
 
 fn tier_context_window(model: &str) -> Option<u64> {
     match model {
-        MODEL_REASONING_V1 | MODEL_AGENTIC_V1 | MODEL_CODING_V1 => Some(TIER_LARGE_CONTEXT),
+        MODEL_REASONING_V1 => Some(TIER_REASONING_CONTEXT),
+        MODEL_AGENTIC_V1 | MODEL_CODING_V1 => Some(TIER_LARGE_CONTEXT),
         "summarization-v1" => Some(TIER_SUMMARIZATION_CONTEXT),
         MODEL_CHAT_V1 | MODEL_REASONING_QUICK_V1 | "chat" => Some(TIER_STANDARD_CONTEXT),
         m if m.starts_with("gemma") || m.contains(":1b") || m.contains("270m") => {
@@ -271,7 +274,7 @@ mod tests {
 
     #[test]
     fn tier_aliases_resolve() {
-        assert_eq!(context_window_for_model("reasoning-v1"), Some(200_000));
+        assert_eq!(context_window_for_model("reasoning-v1"), Some(1_000_000));
         assert_eq!(context_window_for_model("agentic-v1"), Some(200_000));
         assert_eq!(context_window_for_model("chat-v1"), Some(128_000));
         assert_eq!(
diff --git a/src/openhuman/threads/ops.rs b/src/openhuman/threads/ops.rs
index 156de2762..b74aac19e 100644
--- a/src/openhuman/threads/ops.rs
+++ b/src/openhuman/threads/ops.rs
@@ -664,6 +664,164 @@ pub async fn turn_state_clear(
     Ok(envelope(ClearTurnStateResponse { cleared }, None, None))
 }
 
+/// Request for [`token_usage`]: the thread whose persisted usage to total.
+#[derive(Debug, Clone, serde::Deserialize)]
+pub struct ThreadTokenUsageRequest {
+    pub thread_id: String,
+}
+
+/// Aggregated token/cost usage for one thread, read back from its persisted
+/// session transcripts. Seeds the UI footer when the user selects a thread so
+/// the totals reflect prior turns instead of starting at zero.
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct ThreadTokenUsageResponse {
+    pub thread_id: String,
+    pub input_tokens: u64,
+    pub output_tokens: u64,
+    pub cached_input_tokens: u64,
+    pub cost_usd: f64,
+    pub turn_count: usize,
+    /// Tokens of the most recent turn — numerator for the context-window gauge.
+    pub last_turn_input_tokens: u64,
+    pub last_turn_output_tokens: u64,
+    /// Context window (tokens) inferred from the last model; `0` when unknown.
+    pub context_window: u64,
+    pub model: Option<String>,
+    pub updated: Option<String>,
+    /// `false` when the thread has no persisted turns yet (all zeros).
+    pub has_usage: bool,
+    /// Per-archetype sub-agent spend (re-audited at current pricing). The
+    /// top-level totals already include this; it's broken out for the UI's
+    /// per-agent footer rows.
+    pub subagents: Vec<SubagentUsageDto>,
+}
+
+/// One sub-agent archetype's contribution within a thread.
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct SubagentUsageDto {
+    pub agent_id: String,
+    pub input_tokens: u64,
+    pub output_tokens: u64,
+    pub cost_usd: f64,
+    pub runs: usize,
+}
+
+/// Total a thread's persisted token/cost usage across its root transcripts.
+pub async fn token_usage(
+    request: ThreadTokenUsageRequest,
+) -> Result<RpcOutcome<ApiEnvelope<ThreadTokenUsageResponse>>, String> {
+    let dir = workspace_dir().await?;
+    let summary = crate::openhuman::agent::harness::session::transcript::read_thread_usage_summary(
+        &dir,
+        &request.thread_id,
+    );
+
+    // Re-audit cost at CURRENT pricing rather than trusting the
+    // `charged_amount_usd` persisted in the transcript: those values were
+    // stamped at turn time and don't reflect later tier-pricing corrections.
+    // Recompute from the persisted token counts using the last-known model's
+    // rates; falls back to `fallback` only when the model is unknown.
+    let audit_cost =
+        |model: Option<&str>, input: u64, output: u64, cached: u64, fallback: f64| match model {
+            Some(m) => crate::openhuman::agent::cost::estimate_call_cost_usd(
+                m,
+                &crate::openhuman::inference::provider::UsageInfo {
+                    input_tokens: input,
+                    output_tokens: output,
+                    cached_input_tokens: cached,
+                    ..Default::default()
+                },
+            ),
+            None => fallback,
+        };
+
+    let response = match summary {
+        Some(s) => {
+            let context_window = s
+                .model
+                .as_deref()
+                .and_then(crate::openhuman::inference::model_context::context_window_for_model)
+                .unwrap_or(0);
+
+            // Orchestrator (root) spend, re-audited.
+            let orchestrator_cost = audit_cost(
+                s.model.as_deref(),
+                s.input_tokens,
+                s.output_tokens,
+                s.cached_input_tokens,
+                s.cost_usd,
+            );
+
+            // Sub-agent archetypes, each re-audited with its own model. Older
+            // sub-agent transcripts didn't persist a model on their messages, so
+            // fall back to the thread's (root) model rather than pricing them at
+            // $0 — sub-agents usually run on the same managed tier as the parent.
+            let mut subagents = Vec::with_capacity(s.subagents.len());
+            let (mut sub_in, mut sub_out, mut sub_cached, mut sub_cost) = (0u64, 0u64, 0u64, 0.0);
+            for g in &s.subagents {
+                let sub_model = g.model.as_deref().or(s.model.as_deref());
+                let cost = audit_cost(
+                    sub_model,
+                    g.input_tokens,
+                    g.output_tokens,
+                    g.cached_input_tokens,
+                    0.0,
+                );
+                sub_in = sub_in.saturating_add(g.input_tokens);
+                sub_out = sub_out.saturating_add(g.output_tokens);
+                sub_cached = sub_cached.saturating_add(g.cached_input_tokens);
+                sub_cost += cost;
+                subagents.push(SubagentUsageDto {
+                    agent_id: g.agent_id.clone(),
+                    input_tokens: g.input_tokens,
+                    output_tokens: g.output_tokens,
+                    cost_usd: cost,
+                    runs: g.runs,
+                });
+            }
+
+            // Top-level totals = orchestrator + all sub-agents.
+            ThreadTokenUsageResponse {
+                thread_id: request.thread_id.clone(),
+                input_tokens: s.input_tokens.saturating_add(sub_in),
+                output_tokens: s.output_tokens.saturating_add(sub_out),
+                cached_input_tokens: s.cached_input_tokens.saturating_add(sub_cached),
+                cost_usd: orchestrator_cost + sub_cost,
+                turn_count: s.turn_count,
+                last_turn_input_tokens: s.last_turn_input_tokens,
+                last_turn_output_tokens: s.last_turn_output_tokens,
+                context_window,
+                model: s.model,
+                updated: Some(s.updated),
+                has_usage: true,
+                subagents,
+            }
+        }
+        None => ThreadTokenUsageResponse {
+            thread_id: request.thread_id.clone(),
+            input_tokens: 0,
+            output_tokens: 0,
+            cached_input_tokens: 0,
+            cost_usd: 0.0,
+            turn_count: 0,
+            last_turn_input_tokens: 0,
+            last_turn_output_tokens: 0,
+            context_window: 0,
+            model: None,
+            updated: None,
+            has_usage: false,
+            subagents: Vec::new(),
+        },
+    };
+
+    let has_usage = response.has_usage;
+    Ok(envelope(
+        response,
+        Some(counts([("has_usage", usize::from(has_usage))])),
+        None,
+    ))
+}
+
 #[cfg(test)]
 #[path = "ops_tests.rs"]
 mod tests;
diff --git a/src/openhuman/threads/schemas.rs b/src/openhuman/threads/schemas.rs
index 8a5c20e20..47b866674 100644
--- a/src/openhuman/threads/schemas.rs
+++ b/src/openhuman/threads/schemas.rs
@@ -34,6 +34,7 @@ pub fn all_controller_schemas() -> Vec<ControllerSchema> {
         schemas("turn_state_clear"),
         schemas("task_board_get"),
         schemas("task_board_put"),
+        schemas("token_usage"),
     ]
 }
 
@@ -103,6 +104,10 @@ pub fn all_registered_controllers() -> Vec<RegisteredController> {
             schema: schemas("task_board_put"),
             handler: handle_task_board_put,
         },
+        RegisteredController {
+            schema: schemas("token_usage"),
+            handler: handle_token_usage,
+        },
     ]
 }
 
@@ -450,6 +455,23 @@ pub fn schemas(function: &str) -> ControllerSchema {
                 required: true,
             }],
         },
+        "token_usage" => ControllerSchema {
+            namespace: "threads",
+            function: "token_usage",
+            description: "Total a thread's persisted token/cost usage from its session transcripts.",
+            inputs: vec![FieldSchema {
+                name: "thread_id",
+                ty: TypeSchema::String,
+                comment: "Thread identifier.",
+                required: true,
+            }],
+            outputs: vec![FieldSchema {
+                name: "result",
+                ty: TypeSchema::Json,
+                comment: "Envelope with the thread's token/cost totals (zeros when no turns yet).",
+                required: true,
+            }],
+        },
         _other => ControllerSchema {
             namespace: "threads",
             function: "unknown",
@@ -647,6 +669,13 @@ fn handle_task_board_put(params: Map<String, Value>) -> ControllerFuture {
     })
 }
 
+fn handle_token_usage(params: Map<String, Value>) -> ControllerFuture {
+    Box::pin(async move {
+        let p = parse::<ops::ThreadTokenUsageRequest>(params)?;
+        to_json(ops::token_usage(p).await?)
+    })
+}
+
 // ── Helpers ──────────────────────────────────────────────────────────
 
 fn parse<T: DeserializeOwned>(params: Map<String, Value>) -> Result<T, String> {
diff --git a/tests/json_rpc_e2e.rs b/tests/json_rpc_e2e.rs
index 8e8a18ba2..67949d71e 100644
--- a/tests/json_rpc_e2e.rs
+++ b/tests/json_rpc_e2e.rs
@@ -13264,3 +13264,127 @@ async fn poll_team_task_status(rpc_base: &str, team_id: &str, task_id: &str, wan
     }
     false
 }
+
+/// End-to-end: plant a thread's session transcript on disk, then verify the
+/// `openhuman.threads_token_usage` RPC reads back the correct cumulative token
+/// totals, cost, last-turn usage, model, and inferred context window — the data
+/// the composer footer seeds itself from when a thread is selected.
+#[tokio::test]
+async fn json_rpc_threads_token_usage_reads_persisted_thread_totals() {
+    let _env_lock = json_rpc_e2e_env_lock();
+    let tmp = tempdir().expect("tempdir");
+    let home = tmp.path();
+    let openhuman_home = home.join(".openhuman");
+    let workspace = home.join("workspace");
+
+    let _home_guard = EnvVarGuard::set_to_path("HOME", home);
+    let _workspace_guard = EnvVarGuard::set_to_path("OPENHUMAN_WORKSPACE", &workspace);
+    let _backend_url_guard = EnvVarGuard::unset("BACKEND_URL");
+    let _vite_backend_guard = EnvVarGuard::unset("VITE_BACKEND_URL");
+
+    write_min_config(&openhuman_home, "http://127.0.0.1:9");
+
+    // Plant a root session transcript for thread "thr-e2e": a `_meta` header with
+    // the cumulative totals + an assistant message carrying last-turn usage/model.
+    let raw = workspace.join("session_raw");
+    std::fs::create_dir_all(&raw).expect("mkdir session_raw");
+    let jsonl = format!(
+        "{}\n{}\n{}\n",
+        json!({"_meta": {
+            "agent": "main", "dispatcher": "native",
+            "created": "2026-04-11T14:30:00Z", "updated": "2026-04-11T14:35:22Z",
+            "turn_count": 2, "input_tokens": 4200, "output_tokens": 900,
+            "cached_input_tokens": 600, "charged_amount_usd": 0.0123, "thread_id": "thr-e2e"
+        }}),
+        json!({"role": "user", "content": "hi"}),
+        json!({"role": "assistant", "content": "hello", "model": "reasoning-v1",
+            "usage": {"input": 350, "output": 80, "cached_input": 40, "cost_usd": 0.0009},
+            "ts": "2026-04-11T14:35:22Z"}),
+    );
+    std::fs::write(raw.join("1700000000_main.jsonl"), jsonl).expect("write transcript");
+
+    // A sub-agent transcript (stem contains `__`) for the SAME thread: a `coder`
+    // archetype. Its message carries NO model (mirroring how sub-agent
+    // transcripts were historically written), so pricing must fall back to the
+    // thread's model rather than $0. Its spend is grouped under `coder` and
+    // folded into the thread totals — not collapsed into the orchestrator.
+    let sub_jsonl = format!(
+        "{}\n{}\n",
+        json!({"_meta": {
+            "agent": "coder", "dispatcher": "native",
+            "created": "2026-04-11T14:33:00Z", "updated": "2026-04-11T14:34:00Z",
+            "turn_count": 1, "input_tokens": 1000, "output_tokens": 200,
+            "cached_input_tokens": 0, "charged_amount_usd": 0.0, "thread_id": "thr-e2e"
+        }}),
+        json!({"role": "assistant", "content": "done"}),
+    );
+    std::fs::write(
+        raw.join("1700000000_main__1700000050_coder.jsonl"),
+        sub_jsonl,
+    )
+    .expect("write subagent transcript");
+
+    let (rpc_addr, rpc_join) = serve_on_ephemeral(build_core_http_router(false)).await;
+    let rpc_base = format!("http://{rpc_addr}");
+
+    // Known thread → totals read back from the transcripts (orchestrator + coder).
+    let resp = post_json_rpc(
+        &rpc_base,
+        1,
+        "openhuman.threads_token_usage",
+        json!({ "thread_id": "thr-e2e" }),
+    )
+    .await;
+    let envelope = assert_no_jsonrpc_error(&resp, "threads_token_usage");
+    let data = envelope
+        .get("data")
+        .unwrap_or_else(|| panic!("missing data envelope: {envelope}"));
+    // Top-level totals include the sub-agent: 4200+1000 in, 900+200 out.
+    assert_eq!(data["input_tokens"], 5200);
+    assert_eq!(data["output_tokens"], 1100);
+    assert_eq!(data["cached_input_tokens"], 600);
+    // Cost is RE-AUDITED at current pricing, NOT the stale persisted charge.
+    // The coder sub-agent has no model, so it's priced at the thread's model
+    // (reasoning-v1 = "Pro"), NOT $0.
+    // orchestrator: (4200-600)*0.435 + 600*0.003625 + 900*0.87 = 0.002351175
+    // coder:        1000*0.435 + 0 + 200*0.87                  = 0.000609
+    // total                                                     = 0.002960175
+    let cost = data["cost_usd"].as_f64().expect("cost_usd");
+    assert!(
+        (cost - 0.002_960_175).abs() < 1e-9,
+        "re-audited total cost should be ~0.00296, got {cost}"
+    );
+    assert_eq!(data["turn_count"], 2);
+    assert_eq!(data["last_turn_input_tokens"], 350);
+    assert_eq!(data["last_turn_output_tokens"], 80);
+    assert_eq!(data["model"], "reasoning-v1");
+    // reasoning-v1 resolves to a 1M context window.
+    assert_eq!(data["context_window"], 1_000_000);
+    assert_eq!(data["has_usage"], true);
+    // Sub-agent breakdown grouped by archetype.
+    let subs = data["subagents"].as_array().expect("subagents array");
+    assert_eq!(subs.len(), 1);
+    assert_eq!(subs[0]["agent_id"], "coder");
+    assert_eq!(subs[0]["input_tokens"], 1000);
+    assert_eq!(subs[0]["output_tokens"], 200);
+    assert_eq!(subs[0]["runs"], 1);
+    assert!((subs[0]["cost_usd"].as_f64().expect("sub cost") - 0.000_609).abs() < 1e-9);
+
+    // Unknown thread → all-zero totals with has_usage=false (brand-new thread).
+    let resp_unknown = post_json_rpc(
+        &rpc_base,
+        2,
+        "openhuman.threads_token_usage",
+        json!({ "thread_id": "thr-does-not-exist" }),
+    )
+    .await;
+    let unknown = assert_no_jsonrpc_error(&resp_unknown, "threads_token_usage unknown");
+    let unknown_data = unknown
+        .get("data")
+        .unwrap_or_else(|| panic!("missing data envelope: {unknown}"));
+    assert_eq!(unknown_data["input_tokens"], 0);
+    assert_eq!(unknown_data["cost_usd"], 0.0);
+    assert_eq!(unknown_data["has_usage"], false);
+
+    rpc_join.abort();
+}