diff --git a/apps/desktop/src/features/chat/transcript/ActivityGroup.tsx b/apps/desktop/src/features/chat/transcript/ActivityGroup.tsx
index 8b725b69c..2d614a75b 100644
--- a/apps/desktop/src/features/chat/transcript/ActivityGroup.tsx
+++ b/apps/desktop/src/features/chat/transcript/ActivityGroup.tsx
@@ -52,6 +52,7 @@ import {
   useAutomaticDisclosure,
 } from "./shared";
 import { TranscriptSearchContext } from "../../../lib/transcript-search-context";
+import type { LiveTokenRate } from "./hooks/useLiveTokenRate";
 import { useAppStore } from "../../../stores/app-store";
 import { resolveThinkingDisplayMode } from "../../../lib/turn-process";
 
@@ -421,8 +422,39 @@ export const ActivityGroup = memo(function ActivityGroup({
   );
 }, activityGroupPropsEqual);
 
+function formatLiveTokenCount(value: number): string {
+  if (value >= 1_000_000) return `${(value / 1_000_000).toFixed(value >= 10_000_000 ? 0 : 1)}M`;
+  if (value >= 10_000) return `${Math.round(value / 1000)}k`;
+  if (value >= 1000) return `${(value / 1000).toFixed(1)}k`;
+  return String(value);
+}
+
+/** Compact tok/s chip for the run-activity / working status strip. */
+export function LiveTokenRateLabel({ rate }: { rate: LiveTokenRate }) {
+  const { t } = useTranslation();
+  if (rate.tokensPerSecond === undefined) return null;
+  return (
+    <span
+      className="working-token-rate"
+      data-testid="live-token-rate"
+      title={t("chat.usageThroughputLabel")}
+      aria-hidden="true"
+    >
+      {t(
+        rate.estimated
+          ? "chat.usageLiveThroughputEstimated"
+          : "chat.usageLiveThroughput",
+        { count: formatLiveTokenCount(rate.tokensPerSecond) },
+      )}
+    </span>
+  );
+}
+
 /** Keep the running turn visible when no more specific runtime phase is known. */
-export function WorkingIndicator({ startedAt }: { startedAt?: number } = {}) {
+export function WorkingIndicator({
+  startedAt,
+  tokenRate,
+}: { startedAt?: number; tokenRate?: LiveTokenRate } = {}) {
   const { t } = useTranslation();
   const [elapsed, setElapsed] = useState(0);
   const startedAtRef = useRef(startedAt ?? Date.now());
@@ -450,6 +482,9 @@ export function WorkingIndicator({ startedAt }: { startedAt?: number } = {}) {
         <span />
       </span>
       <span className="working-indicator-label">{t("chat.running")}</span>
+      {tokenRate != null && tokenRate.tokensPerSecond !== undefined ? (
+        <LiveTokenRateLabel rate={tokenRate} />
+      ) : null}
       {elapsed > 0 ? (
         <span className="working-elapsed" aria-hidden="true">
           {formatToolDuration(elapsed)}
@@ -459,7 +494,13 @@ export function WorkingIndicator({ startedAt }: { startedAt?: number } = {}) {
   );
 }
 
-export function RunActivityIndicator({ activity }: { activity: AgentActivity }) {
+export function RunActivityIndicator({
+  activity,
+  tokenRate,
+}: {
+  activity: AgentActivity;
+  tokenRate?: LiveTokenRate;
+}) {
   const { t } = useTranslation();
   const [now, setNow] = useState(Date.now);
   const retryErrorDetailsId = useId();
@@ -604,6 +645,9 @@ export function RunActivityIndicator({ activity }: { activity: AgentActivity })
         <span />
       </span>
       {labelContent}
+      {tokenRate != null && tokenRate.tokensPerSecond !== undefined ? (
+        <LiveTokenRateLabel rate={tokenRate} />
+      ) : null}
       <span className="working-elapsed" aria-hidden="true">
         {elapsed}
       </span>
diff --git a/apps/desktop/src/features/chat/transcript/ChatTranscript.tsx b/apps/desktop/src/features/chat/transcript/ChatTranscript.tsx
index a8398de83..458e41d4d 100644
--- a/apps/desktop/src/features/chat/transcript/ChatTranscript.tsx
+++ b/apps/desktop/src/features/chat/transcript/ChatTranscript.tsx
@@ -1,4 +1,4 @@
-import { memo, type MouseEvent as ReactMouseEvent } from "react";
+import { memo, useMemo, type MouseEvent as ReactMouseEvent } from "react";
 import { useTranslation } from "react-i18next";
 import type { PlanningState, UiMessage } from "@pi-desktop/shared";
 import { proposalKindForMode } from "@pi-desktop/shared";
@@ -15,6 +15,8 @@ import {
   RunActivityIndicator,
   WorkingIndicator,
 } from "./ActivityGroup";
+import { useLiveTokenRate } from "./hooks/useLiveTokenRate";
+import { assistantTurnMessages } from "../../../lib/assistant-turns";
 import { TranscriptHistory, TranscriptTail } from "./AssistantTurn";
 import { useTranscriptScroll } from "./hooks/useTranscriptScroll";
 import type { TranscriptSearchTarget } from "../../../lib/transcript-reading";
@@ -171,6 +173,24 @@ function TranscriptBody({
     planningState === "planning" &&
     !hasSpecializedActivity;
 
+  const streamingAssistant = useMemo(() => {
+    if (!transcriptRunning || tailEntry?.kind !== "assistant-turn") {
+      return undefined;
+    }
+    return [...assistantTurnMessages(tailEntry)]
+      .reverse()
+      .find((message) => message.status === "streaming");
+  }, [transcriptRunning, tailEntry]);
+
+  const liveTokenRate = useLiveTokenRate({
+    active: showStatus,
+    content: streamingAssistant?.content,
+    thinking: streamingAssistant?.thinking,
+    outputTokens:
+      streamingAssistant?.usage?.outputTokens ??
+      streamingAssistant?.responseOutputTokens,
+  });
+
   // The tail status lane is part of the layout for the whole running turn: the
   // indicators below mount and clear with the turn's phase, and a lane that
   // came and went with them would resize `.thread-content` and push the rows
@@ -304,10 +324,15 @@ function TranscriptBody({
           {runtimeStatusLane ? (
             <div className="transcript-runtime-status">
               {showRunActivity && specializedActivity ? (
-                <RunActivityIndicator activity={specializedActivity} />
+                <RunActivityIndicator
+                  activity={specializedActivity}
+                  tokenRate={liveTokenRate}
+                />
               ) : null}
               {showPlanning ? <PlanningIndicator kind={planningKind} /> : null}
-              {showWorking ? <WorkingIndicator /> : null}
+              {showWorking ? (
+                <WorkingIndicator tokenRate={liveTokenRate} />
+              ) : null}
             </div>
           ) : null}
         </div>
diff --git a/apps/desktop/src/features/chat/transcript/hooks/useLiveTokenRate.ts b/apps/desktop/src/features/chat/transcript/hooks/useLiveTokenRate.ts
new file mode 100644
index 000000000..6554b079b
--- /dev/null
+++ b/apps/desktop/src/features/chat/transcript/hooks/useLiveTokenRate.ts
@@ -0,0 +1,88 @@
+import { useEffect, useRef, useState } from "react";
+import {
+  appendTokenSample,
+  calculateWindowedTokenRate,
+  resolveStreamingOutputTokens,
+  shouldResetTokenRateWindow,
+  type TokenRateSample,
+} from "../../../../lib/streaming-token-rate";
+
+export type LiveTokenRate = {
+  tokensPerSecond: number | undefined;
+  estimated: boolean;
+};
+
+const IDLE_RATE: LiveTokenRate = {
+  tokensPerSecond: undefined,
+  estimated: false,
+};
+
+/**
+ * Sample the active turn's output tokens on a short interval and expose a
+ * sliding-window tok/s reading for the transcript stream-health strip.
+ *
+ * Latest content / thinking / outputTokens live in refs so the interval is
+ * not torn down on every stream delta — only `active` / `tickMs` restart it.
+ */
+export function useLiveTokenRate(input: {
+  active: boolean;
+  content?: string;
+  thinking?: string;
+  outputTokens?: number;
+  tickMs?: number;
+}): LiveTokenRate {
+  const [rate, setRate] = useState<LiveTokenRate>(IDLE_RATE);
+  const samplesRef = useRef<TokenRateSample[]>([]);
+  const estimatedRef = useRef(false);
+  const inputRef = useRef(input);
+  inputRef.current = input;
+
+  useEffect(() => {
+    if (!input.active) {
+      samplesRef.current = [];
+      estimatedRef.current = false;
+      setRate(IDLE_RATE);
+      return;
+    }
+
+    const sample = () => {
+      const current = inputRef.current;
+      const nowMs = Date.now();
+      const resolved = resolveStreamingOutputTokens({
+        outputTokens: current.outputTokens,
+        content: current.content,
+        thinking: current.thinking,
+      });
+      const last = samplesRef.current[samplesRef.current.length - 1];
+      if (
+        shouldResetTokenRateWindow({
+          wasEstimated: estimatedRef.current,
+          nowEstimated: resolved.estimated,
+          previousTokens: last?.tokens,
+          nextTokens: resolved.tokens,
+        })
+      ) {
+        samplesRef.current = [];
+      }
+      estimatedRef.current = resolved.estimated;
+      samplesRef.current = appendTokenSample(
+        samplesRef.current,
+        { atMs: nowMs, tokens: resolved.tokens },
+        nowMs,
+      );
+      setRate({
+        tokensPerSecond: calculateWindowedTokenRate(
+          samplesRef.current,
+          nowMs,
+        ),
+        estimated: resolved.estimated,
+      });
+    };
+
+    sample();
+    const timer = window.setInterval(sample, input.tickMs ?? 250);
+    return () => window.clearInterval(timer);
+  }, [input.active, input.tickMs]);
+
+  return input.active ? rate : IDLE_RATE;
+}
diff --git a/apps/desktop/src/lib/streaming-token-rate.ts b/apps/desktop/src/lib/streaming-token-rate.ts
new file mode 100644
index 000000000..1570b0d5d
--- /dev/null
+++ b/apps/desktop/src/lib/streaming-token-rate.ts
@@ -0,0 +1,181 @@
+/**
+ * Live conversation output rate (tok/s) for stream-health display.
+ *
+ * Prefer provider `outputTokens` when the runtime reports them mid-stream;
+ * otherwise estimate from visible thinking+answer text with the same
+ * four-code-points-per-token heuristic used for stopped-turn throughput.
+ */
+
+export type TokenRateSample = {
+  atMs: number;
+  /** Cumulative output tokens observed by this sample. */
+  tokens: number;
+};
+
+export type WindowedTokenRateOptions = {
+  /** Sliding window length. Default 2000ms. */
+  windowMs?: number;
+  /** Minimum elapsed time before a rate is reported. Default 200ms. */
+  minDurationMs?: number;
+};
+
+export type StreamingOutputTokens = {
+  tokens: number;
+  /** True when tokens came from text estimate rather than provider usage. */
+  estimated: boolean;
+};
+
+const DEFAULT_WINDOW_MS = 2_000;
+const DEFAULT_MIN_DURATION_MS = 200;
+const DEFAULT_RETAIN_MS = 12_000;
+
+/**
+ * Resolve the best available cumulative output token count for the live turn.
+ */
+export function resolveStreamingOutputTokens(input: {
+  outputTokens?: number;
+  content?: string;
+  thinking?: string;
+}): StreamingOutputTokens {
+  const reported = input.outputTokens;
+  if (
+    typeof reported === "number" &&
+    Number.isFinite(reported) &&
+    reported > 0
+  ) {
+    return { tokens: Math.round(reported), estimated: false };
+  }
+
+  const visible = `${input.thinking ?? ""}${input.content ?? ""}`.trim();
+  if (!visible) return { tokens: 0, estimated: true };
+  return {
+    tokens: Math.max(1, Math.ceil(Array.from(visible).length / 4)),
+    estimated: true,
+  };
+}
+
+/**
+ * True when mid-stream usage flips from a text estimate to provider
+ * `outputTokens` and the provider count is lower than the estimate. Callers
+ * must reset the sample window so monotonic clamping cannot freeze the rate
+ * at a hard `0 tok/s` without `≈`.
+ */
+export function shouldResetTokenRateWindow(input: {
+  wasEstimated: boolean;
+  nowEstimated: boolean;
+  previousTokens?: number;
+  nextTokens: number;
+}): boolean {
+  if (!input.wasEstimated || input.nowEstimated) return false;
+  if (
+    typeof input.previousTokens !== "number" ||
+    !Number.isFinite(input.previousTokens) ||
+    !Number.isFinite(input.nextTokens)
+  ) {
+    return false;
+  }
+  return input.nextTokens < input.previousTokens;
+}
+
+/**
+ * Append a cumulative sample, dropping entries older than the retain horizon.
+ * Tokens are forced monotonic so a noisy estimate cannot go backwards.
+ */
+export function appendTokenSample(
+  samples: readonly TokenRateSample[],
+  next: TokenRateSample,
+  nowMs = next.atMs,
+  retainMs = DEFAULT_RETAIN_MS,
+): TokenRateSample[] {
+  if (
+    !Number.isFinite(next.atMs) ||
+    !Number.isFinite(next.tokens) ||
+    next.tokens < 0
+  ) {
+    return samples.slice();
+  }
+
+  const retainBefore = nowMs - retainMs;
+  const kept = samples.filter(
+    (sample) =>
+      Number.isFinite(sample.atMs) &&
+      Number.isFinite(sample.tokens) &&
+      sample.atMs >= retainBefore,
+  );
+  const last = kept[kept.length - 1];
+  if (last && last.tokens === next.tokens && next.atMs - last.atMs < 40) {
+    return kept;
+  }
+  const tokens = last ? Math.max(last.tokens, Math.round(next.tokens)) : Math.round(next.tokens);
+  return [...kept, { atMs: next.atMs, tokens }];
+}
+
+/**
+ * Sliding-window tokens/s from cumulative samples ending at `nowMs`.
+ *
+ * - Returns `undefined` when no positive output has been observed yet (hide
+ *   the chip through TTFT), or when elapsed time is zero / below
+ *   `minDurationMs` (not enough signal yet).
+ * - Returns `0` only after positive output was seen and the window advanced
+ *   without token growth (stall / reconnect), so the UI can show the stream
+ *   is no longer producing.
+ * - Concentrates a late burst into a high instantaneous rate by ignoring
+ *   earlier growth outside the window.
+ */
+export function calculateWindowedTokenRate(
+  samples: readonly TokenRateSample[],
+  nowMs: number,
+  options: WindowedTokenRateOptions = {},
+): number | undefined {
+  const windowMs = options.windowMs ?? DEFAULT_WINDOW_MS;
+  const minDurationMs = options.minDurationMs ?? DEFAULT_MIN_DURATION_MS;
+  if (
+    !Number.isFinite(nowMs) ||
+    !Number.isFinite(windowMs) ||
+    windowMs <= 0 ||
+    !Number.isFinite(minDurationMs) ||
+    minDurationMs < 0
+  ) {
+    return undefined;
+  }
+
+  const ordered = samples
+    .filter(
+      (sample) =>
+        Number.isFinite(sample.atMs) &&
+        Number.isFinite(sample.tokens) &&
+        sample.atMs <= nowMs &&
+        sample.tokens >= 0,
+    )
+    .slice()
+    .sort((left, right) => left.atMs - right.atMs);
+  if (ordered.length === 0) return undefined;
+
+  // Until some positive output exists, never report a stall `0` (false TTFT).
+  if (!ordered.some((sample) => sample.tokens > 0)) return undefined;
+
+  const first = ordered[0];
+  const latest = ordered[ordered.length - 1];
+  const windowStart = nowMs - windowMs;
+
+  let startTokens = first.tokens;
+  let startAt = first.atMs;
+  let foundBeforeWindow = false;
+  for (const sample of ordered) {
+    if (sample.atMs <= windowStart) {
+      startTokens = sample.tokens;
+      startAt = windowStart;
+      foundBeforeWindow = true;
+    }
+  }
+  if (!foundBeforeWindow) {
+    startTokens = first.tokens;
+    startAt = first.atMs;
+  }
+
+  const durationMs = nowMs - startAt;
+  if (durationMs <= 0 || durationMs < minDurationMs) return undefined;
+
+  const deltaTokens = Math.max(0, latest.tokens - startTokens);
+  return Math.round(deltaTokens / (durationMs / 1000));
+}
diff --git a/apps/desktop/src/styles/messages.css b/apps/desktop/src/styles/messages.css
index d32271f99..41ebf6226 100644
--- a/apps/desktop/src/styles/messages.css
+++ b/apps/desktop/src/styles/messages.css
@@ -576,6 +576,13 @@
   font-variant-numeric: tabular-nums;
 }
 
+.working-token-rate {
+  color: var(--ds-text-secondary);
+  font-size: var(--text-xs-plus);
+  font-variant-numeric: tabular-nums;
+  white-space: nowrap;
+}
+
 .planning-state-indicator {
   box-sizing: border-box;
   display: flex;
diff --git a/apps/desktop/test/active-turn-surface.test.mjs b/apps/desktop/test/active-turn-surface.test.mjs
index 656a8fb17..a85e14c81 100644
--- a/apps/desktop/test/active-turn-surface.test.mjs
+++ b/apps/desktop/test/active-turn-surface.test.mjs
@@ -42,7 +42,20 @@ test("active turns show immediate and phase-specific feedback without a progress
   assert.match(transcript, /const specializedActivity = agentActivity/);
   assert.match(transcript, /!hasSpecializedActivity/);
   assert.match(transcript, /const showWorking =/);
-  assert.match(transcript, /\{showWorking \? <WorkingIndicator \/> : null\}/);
+  assert.match(
+    transcript,
+    /\{showWorking \? \(\s*<WorkingIndicator tokenRate=\{liveTokenRate\} \/>\s*\) : null\}/,
+  );
+  assert.match(transcript, /useLiveTokenRate\(/);
+  assert.match(transcript, /function LiveTokenRateLabel\(/);
+  assert.match(transcript, /data-testid="live-token-rate"/);
+  assert.match(transcript, /data-testid="live-token-rate"[\s\S]*?aria-hidden="true"/);
+  assert.match(
+    transcript,
+    /tokenRate != null && tokenRate\.tokensPerSecond !== undefined/,
+  );
+  assert.match(transcript, /usageLiveThroughput/);
+  assert.match(messagesStyles, /\.working-token-rate\s*\{/);
   assert.match(transcript, /function PlanningIndicator\(/);
   assert.match(transcript, /data-testid="planning-indicator"/);
   assert.match(transcript, /const showPlanning =/);
diff --git a/apps/desktop/test/streaming-token-rate.test.mjs b/apps/desktop/test/streaming-token-rate.test.mjs
new file mode 100644
index 000000000..c7085dd3f
--- /dev/null
+++ b/apps/desktop/test/streaming-token-rate.test.mjs
@@ -0,0 +1,240 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { readFile } from "node:fs/promises";
+import {
+  appendTokenSample,
+  calculateWindowedTokenRate,
+  resolveStreamingOutputTokens,
+  shouldResetTokenRateWindow,
+} from "../src/lib/streaming-token-rate.ts";
+
+test("windowed rate ignores growth outside the sliding window", () => {
+  const samples = [
+    { atMs: 0, tokens: 0 },
+    { atMs: 1_000, tokens: 100 },
+    { atMs: 3_000, tokens: 120 },
+  ];
+  // Window [1000, 3000]: start tokens at t=1000 → 100, end → 120 ⇒ 20 / 2s = 10.
+  assert.equal(
+    calculateWindowedTokenRate(samples, 3_000, {
+      windowMs: 2_000,
+      minDurationMs: 200,
+    }),
+    10,
+  );
+});
+
+test("zero duration returns undefined", () => {
+  assert.equal(
+    calculateWindowedTokenRate([{ atMs: 1_000, tokens: 50 }], 1_000, {
+      minDurationMs: 200,
+    }),
+    undefined,
+  );
+  assert.equal(
+    calculateWindowedTokenRate(
+      [
+        { atMs: 1_000, tokens: 0 },
+        { atMs: 1_000, tokens: 80 },
+      ],
+      1_000,
+      { minDurationMs: 0 },
+    ),
+    undefined,
+  );
+});
+
+test("burst concentrates rate inside a short recent window", () => {
+  const samples = [
+    { atMs: 0, tokens: 0 },
+    { atMs: 1_800, tokens: 0 },
+    { atMs: 2_000, tokens: 400 },
+  ];
+  // Window [0, 2000] but growth is only in the last 200ms: 400 / 2s = 200
+  // when measuring against now with a 2s window that includes the quiet prefix.
+  assert.equal(
+    calculateWindowedTokenRate(samples, 2_000, {
+      windowMs: 2_000,
+      minDurationMs: 100,
+    }),
+    200,
+  );
+  // Narrower window that starts at the quiet sample: 400 / 0.2s = 2000.
+  assert.equal(
+    calculateWindowedTokenRate(samples, 2_000, {
+      windowMs: 200,
+      minDurationMs: 100,
+    }),
+    2_000,
+  );
+});
+
+test("stall reports zero after the window advances without growth", () => {
+  const samples = [
+    { atMs: 0, tokens: 0 },
+    { atMs: 500, tokens: 100 },
+  ];
+  assert.equal(
+    calculateWindowedTokenRate(samples, 3_000, {
+      windowMs: 2_000,
+      minDurationMs: 200,
+    }),
+    0,
+  );
+});
+
+test("never-produced output stays undefined instead of false 0 tok/s", () => {
+  const samples = [
+    { atMs: 0, tokens: 0 },
+    { atMs: 500, tokens: 0 },
+    { atMs: 2_000, tokens: 0 },
+  ];
+  assert.equal(
+    calculateWindowedTokenRate(samples, 3_000, {
+      windowMs: 2_000,
+      minDurationMs: 200,
+    }),
+    undefined,
+  );
+});
+
+test("below min duration stays undefined even with tokens", () => {
+  assert.equal(
+    calculateWindowedTokenRate(
+      [
+        { atMs: 0, tokens: 0 },
+        { atMs: 50, tokens: 40 },
+      ],
+      50,
+      { windowMs: 2_000, minDurationMs: 200 },
+    ),
+    undefined,
+  );
+});
+
+test("appendTokenSample keeps tokens monotonic and drops stale entries", () => {
+  const first = appendTokenSample([], { atMs: 1_000, tokens: 10 }, 1_000, 1_000);
+  assert.deepEqual(first, [{ atMs: 1_000, tokens: 10 }]);
+  const second = appendTokenSample(
+    first,
+    { atMs: 1_500, tokens: 8 },
+    1_500,
+    1_000,
+  );
+  assert.deepEqual(second, [
+    { atMs: 1_000, tokens: 10 },
+    { atMs: 1_500, tokens: 10 },
+  ]);
+  const trimmed = appendTokenSample(
+    second,
+    { atMs: 3_000, tokens: 20 },
+    3_000,
+    1_000,
+  );
+  assert.deepEqual(trimmed, [{ atMs: 3_000, tokens: 20 }]);
+});
+
+test("estimate→provider flip detects lower provider count for window reset", () => {
+  assert.equal(
+    shouldResetTokenRateWindow({
+      wasEstimated: true,
+      nowEstimated: false,
+      previousTokens: 120,
+      nextTokens: 40,
+    }),
+    true,
+  );
+  assert.equal(
+    shouldResetTokenRateWindow({
+      wasEstimated: true,
+      nowEstimated: false,
+      previousTokens: 40,
+      nextTokens: 80,
+    }),
+    false,
+  );
+  assert.equal(
+    shouldResetTokenRateWindow({
+      wasEstimated: false,
+      nowEstimated: false,
+      previousTokens: 120,
+      nextTokens: 40,
+    }),
+    false,
+  );
+
+  // Without reset, monotonic clamp freezes growth and the window drains to 0.
+  const estimated = [
+    { atMs: 0, tokens: 0 },
+    { atMs: 500, tokens: 100 },
+    { atMs: 1_000, tokens: 120 },
+  ];
+  const clamped = appendTokenSample(
+    estimated,
+    { atMs: 1_200, tokens: 40 },
+    1_200,
+  );
+  assert.equal(clamped[clamped.length - 1].tokens, 120);
+  assert.equal(
+    calculateWindowedTokenRate(clamped, 3_500, {
+      windowMs: 2_000,
+      minDurationMs: 200,
+    }),
+    0,
+  );
+
+  // Resetting the window on handoff keeps the chip hidden until fresh signal.
+  assert.equal(
+    calculateWindowedTokenRate([{ atMs: 1_200, tokens: 40 }], 1_250, {
+      windowMs: 2_000,
+      minDurationMs: 200,
+    }),
+    undefined,
+  );
+  assert.equal(
+    calculateWindowedTokenRate(
+      [
+        { atMs: 1_200, tokens: 40 },
+        { atMs: 1_700, tokens: 90 },
+      ],
+      1_700,
+      { windowMs: 2_000, minDurationMs: 200 },
+    ),
+    100,
+  );
+});
+
+test("resolveStreamingOutputTokens prefers provider usage over text estimate", () => {
+  assert.deepEqual(
+    resolveStreamingOutputTokens({
+      outputTokens: 42,
+      content: "a long answer that would estimate differently",
+    }),
+    { tokens: 42, estimated: false },
+  );
+  assert.deepEqual(
+    resolveStreamingOutputTokens({ content: "abcd" }),
+    { tokens: 1, estimated: true },
+  );
+  assert.deepEqual(resolveStreamingOutputTokens({}), {
+    tokens: 0,
+    estimated: true,
+  });
+});
+
+test("useLiveTokenRate interval depends only on active/tickMs via input refs", async () => {
+  const source = await readFile(
+    new URL("../src/features/chat/transcript/hooks/useLiveTokenRate.ts", import.meta.url),
+    "utf8",
+  );
+  assert.match(source, /inputRef\.current = input/);
+  assert.match(source, /shouldResetTokenRateWindow\(/);
+  assert.match(
+    source,
+    /}, \[input\.active, input\.tickMs\]\);/,
+  );
+  assert.doesNotMatch(
+    source,
+    /}, \[\s*input\.active,\s*input\.content,\s*input\.thinking,\s*input\.outputTokens,\s*input\.tickMs,\s*\]\)/,
+  );
+});
diff --git a/docs/spec/04-ux/08-component-spec.md b/docs/spec/04-ux/08-component-spec.md
index a378a05ee..7be41d030 100644
--- a/docs/spec/04-ux/08-component-spec.md
+++ b/docs/spec/04-ux/08-component-spec.md
@@ -1872,8 +1872,13 @@ Single message render — either user (plaintext) or assistant (markdown streami
   estimate note are intentionally omitted from the default view. Rows below
   the heading share one muted-label / tabular-value rhythm separated by
   spacing; the popover keeps its floating-layer edge and draws no inner
-  section rules (D297). Generation speed is a completed-turn value in tokens
-  per second and is not updated while a response is streaming. The
+  section rules (D297). Generation speed inside the inspector remains a
+  completed-turn snapshot in tokens per second. While a turn is running, the
+  transcript stream-health strip (the working / run-activity indicator, which
+  stays visible for the whole running turn) shows a live
+  sliding-window `tok/s` reading so reconnects and stalls are visible (D639).
+  Live rate prefers provider output usage when present and otherwise estimates
+  from visible thinking+answer text, marked as approximate. The
   context-window total uses the same effective model window as the agent
   sidecar: a published models.dev `limit.context` replaces a legacy 128k
   generic binding seed, while a non-default per-model Advanced value remains
diff --git a/docs/spec/04-ux/09-interaction-patterns.md b/docs/spec/04-ux/09-interaction-patterns.md
index 75b32eca4..3c23ab5ca 100644
--- a/docs/spec/04-ux/09-interaction-patterns.md
+++ b/docs/spec/04-ux/09-interaction-patterns.md
@@ -759,6 +759,16 @@ may be retained while exactly one workspace supplies the visible shell context.
 6. Composer re-activates (unblocked)
 7. Abort is idempotent — pressing abort when already aborting does nothing
 
+### 3.2a Live token rate during a run
+
+While a turn is running, the transcript status strip shows a live
+sliding-window token rate (`tok/s`) next to the working / run-activity
+indicator, which stays visible for the whole running turn. Prefer
+provider output usage when present; otherwise estimate from visible
+thinking+answer text and label as approximate. The rate clears when the
+turn becomes idle; the context inspector keeps the completed-turn
+generation speed (D639).
+
 ### 3.3 Abort UX
 
 - Abort button changes to "Aborting..." briefly (100ms), then disappears
diff --git a/docs/spec/06-delivery/04-e2e-test-plan.md b/docs/spec/06-delivery/04-e2e-test-plan.md
index eaa650762..d1a4e3103 100644
--- a/docs/spec/06-delivery/04-e2e-test-plan.md
+++ b/docs/spec/06-delivery/04-e2e-test-plan.md
@@ -7769,7 +7769,7 @@ must keep splitting are covered by `markdown-blocks.test.mjs`.
     not restore or duplicate text or chips.
   - Scratch bytes remain under the existing session lifecycle.
 - **Specs linked**: `04-ux/08-component-spec.md` §11.5/§11.8,
-  `04-ux/09-interaction-patterns.md` §3.2/§8a.2,
+  `04-ux/09-interaction-patterns.md` §3.2a/§8a.2,
   `03-runtime/10-session-state-machine.md`, `08-meta/decisions-log.md` (D209),
   ADR 0070
 - **Acceptance**: C (conversation & stream), F (persistence), Quality
@@ -8284,6 +8284,36 @@ must keep splitting are covered by `markdown-blocks.test.mjs`.
 - **Status**: Unit/source-contract covered; full cross-platform UI journey Draft
   (run only in a capable environment when this surface changes)
 
+
+#### E2E-CHAT-live-token-rate-shows-during-stream
+
+- **Preconditions**: A session with a configured provider that streams answer
+  tokens (or a test double that emits text deltas on a known cadence).
+- **Steps**:
+  1. Send a prompt that produces a multi-second streamed answer.
+  2. While the working / run-activity indicator is visible (before answer text),
+     observe the status strip.
+  3. Once answer tokens begin streaming, confirm the working indicator stays
+     visible and keeps showing the live rate.
+  4. Optionally pause or throttle the upstream stream long enough for the
+     sliding window to drain, then resume.
+  5. Let the turn finish and confirm the live rate clears; open the context
+     usage inspector and confirm completed-turn Generation speed is still shown.
+- **Expected**:
+  - Live rate appears as `N tok/s` (or the locale equivalent) next to the
+    stream-health indicator, updating during the turn.
+  - When provider output usage is absent mid-stream, the reading is labeled as
+    approximate (`≈`).
+  - A stall drops the live rate toward `0 tok/s` without clearing the chip
+    while the turn remains active.
+  - Idle clears the live chip; inspector Generation speed remains the
+    completed-turn snapshot (D639).
+- **Specs linked**: `04-ux/08-component-spec.md` §8.3,
+  `04-ux/09-interaction-patterns.md` §3.2a, D639
+- **Acceptance**: C (conversation & stream)
+- **Milestone**: M6
+- **Status**: Unit-covered for windowed rate math; UI journey Draft
+
 #### E2E-121: Goal approval resumes autonomous acceptance-criteria execution
 
 - **Preconditions**: A project-bound session has a configured provider and is
@@ -9096,7 +9126,7 @@ must keep splitting are covered by `markdown-blocks.test.mjs`.
 | A — App startup | E2E-001, E2E-002, E2E-003, E2E-004, E2E-067, E2E-076, E2E-079, E2E-092, E2E-097, E2E-143, E2E-150, E2E-168, E2E-204 |
 | A / C / F / Quality — Tray session navigation | E2E-TRAY-bounded-session-navigation |
 | B — Model config | E2E-005, E2E-006, E2E-007, E2E-038, E2E-050, E2E-052, E2E-055, E2E-066, E2E-080, E2E-082, E2E-102c, E2E-102d, E2E-102e, E2E-151, E2E-154, E2E-163, E2E-166, E2E-172, E2E-174, E2E-197, E2E-005G, E2E-005J, E2E-199, E2E-201, E2E-202, E2E-203, E2E-205, E2E-206, E2E-209 |
-| C — Conversation & stream | E2E-CHAT-long-history-stream-keeps-responsive, E2E-CHAT-running-status-survives-output-pauses, E2E-008, E2E-008d, E2E-008e, E2E-008a, E2E-009, E2E-010, E2E-011, E2E-011a, E2E-011b, E2E-011d, E2E-011e, E2E-011g, E2E-031, E2E-040, E2E-047, E2E-048, E2E-048A, E2E-049, E2E-052, E2E-053, E2E-054, E2E-055, E2E-059, E2E-059a, E2E-060c, E2E-060d, E2E-061, E2E-061a, E2E-062, E2E-064, E2E-065, E2E-068, E2E-071, E2E-073, E2E-074, E2E-075, E2E-081, E2E-083, E2E-084, E2E-086, E2E-087, E2E-088, E2E-088b, E2E-089, E2E-090, E2E-COMPOSER-narrow-controls, E2E-094, E2E-095, E2E-096, E2E-097, E2E-098, E2E-099, E2E-102, E2E-102a, E2E-102b, E2E-102c, E2E-102d, E2E-102g, E2E-106, E2E-109, E2E-111, E2E-114, E2E-116, E2E-117, E2E-118, E2E-119, E2E-120, E2E-121, E2E-218, E2E-259, E2E-219, E2E-AGENTS-001, E2E-142, E2E-144, E2E-145, E2E-146, E2E-146a, E2E-147, E2E-151, E2E-154, E2E-155, E2E-158, E2E-159, E2E-161, E2E-162, E2E-166, E2E-172, E2E-173, E2E-174, E2E-177, E2E-178, E2E-179, E2E-180, E2E-182, E2E-183, E2E-187, E2E-198, E2E-199, E2E-202, E2E-203, E2E-207, E2E-208, E2E-CHAT-content-width-handles, E2E-250, E2E-102i, E2E-PLUGIN-session-orchestrator-real-workers, E2E-SUBAGENT-settlement-updates-before-parent-poll, E2E-SUBAGENT-resume-a-settled-delegation |
+| C — Conversation & stream | E2E-CHAT-long-history-stream-keeps-responsive, E2E-CHAT-running-status-survives-output-pauses, E2E-008, E2E-008d, E2E-008e, E2E-008a, E2E-009, E2E-010, E2E-011, E2E-011a, E2E-011b, E2E-011d, E2E-011e, E2E-011g, E2E-031, E2E-040, E2E-047, E2E-048, E2E-048A, E2E-049, E2E-052, E2E-053, E2E-054, E2E-055, E2E-059, E2E-059a, E2E-060c, E2E-060d, E2E-061, E2E-061a, E2E-062, E2E-064, E2E-065, E2E-068, E2E-071, E2E-073, E2E-074, E2E-075, E2E-081, E2E-083, E2E-084, E2E-086, E2E-087, E2E-088, E2E-088b, E2E-089, E2E-090, E2E-COMPOSER-narrow-controls, E2E-094, E2E-095, E2E-096, E2E-097, E2E-098, E2E-099, E2E-102, E2E-102a, E2E-102b, E2E-102c, E2E-102d, E2E-102g, E2E-106, E2E-109, E2E-111, E2E-114, E2E-116, E2E-117, E2E-118, E2E-119, E2E-120, E2E-CHAT-live-token-rate-shows-during-stream, E2E-121, E2E-218, E2E-259, E2E-219, E2E-AGENTS-001, E2E-142, E2E-144, E2E-145, E2E-146, E2E-146a, E2E-147, E2E-151, E2E-154, E2E-155, E2E-158, E2E-159, E2E-161, E2E-162, E2E-166, E2E-172, E2E-173, E2E-174, E2E-177, E2E-178, E2E-179, E2E-180, E2E-182, E2E-183, E2E-187, E2E-198, E2E-199, E2E-202, E2E-203, E2E-207, E2E-208, E2E-CHAT-content-width-handles, E2E-250, E2E-102i, E2E-PLUGIN-session-orchestrator-real-workers, E2E-SUBAGENT-settlement-updates-before-parent-poll, E2E-SUBAGENT-resume-a-settled-delegation |
 | C — Conversation & stream (composer drafts) | E2E-011c, E2E-011c-1 |
 | D — Workspace | E2E-012, E2E-013, E2E-022B, E2E-024I, E2E-047, E2E-049, E2E-057, E2E-058, E2E-060, E2E-068, E2E-075, E2E-078, E2E-153, E2E-158, E2E-182, E2E-187, E2E-252 |
 | D — Workspace (project ordering) | E2E-253 |
diff --git a/docs/spec/08-meta/decisions-log.md b/docs/spec/08-meta/decisions-log.md
index cb529ea7d..f2aac73b2 100644
--- a/docs/spec/08-meta/decisions-log.md
+++ b/docs/spec/08-meta/decisions-log.md
@@ -110,6 +110,7 @@ This log freezes previously open questions into concrete decisions.
 
 
 | D244 | Compact context usage summary | **Amend D103 / D184 / ADR 0047: keep the context inspector's remaining-capacity trigger, used/window counts, turn total, completed-turn speed, exact provider values, aggregate tool types/calls/tokens, and checkpoint summary, but render them as a short summary. Remove the per-tool rows, share bars, source badges, explanatory estimate paragraph, and used-capacity meter from the default panel. No protocol, storage, runtime accounting, or model metadata changes.** *(Amended by D347: the trigger moves to the composer toolbar.)* | The prior diagnostic layout made a routine capacity check tall and visually dense. Keeping the aggregate signal while removing drill-down chrome makes the default status surface scannable without changing the underlying usage data. See ADR 0103 and E2E-060d / US-UI-61. |
+| D639 | Live conversation token rate | **Amend D212 / D244: while a turn is running, the transcript stream-health strip shows a live sliding-window output rate in `tok/s` (working / run-activity indicators, which stay visible for the whole running turn). Prefer provider `outputTokens` when available; otherwise estimate from visible thinking+answer text and label as approximate. Clear when idle. The context inspector's Generation speed remains the completed-turn snapshot. Renderer only; no protocol or persistence change.** | Users could not tell whether a stream was healthy or stalled/reconnecting from the completed-turn-only speed. |
 | D347 | Composer-docked context usage inspector | **Amend D103 / D184 / D244 / ADR 0047 / ADR 0103: the compact context inspector lives in the composer right toolbar, immediately left of the model × reasoning chip, and always mirrors the newest assistant turn that reported usage. The trigger keeps the remaining-capacity ring and percentage and drops the redundant Context label. The popover heading is remaining tokens plus percentage; rows below share one label/value rhythm separated by spacing, with no inner section rules (D297). Assistant meta keeps the model badge. Renderer only.** | The inspector under the newest answer scrolled out of reach. One composer entry is the single authority for the latest snapshot, and the doubled heading rule is fixed by dropping the extra caption rather than adding hairlines. See ADR 0184 and E2E-060d / US-UI-61. |
 | D355 | Last-request occupancy in the context inspector | **Amend D103 / D184 / D244 / D347 / ADR 0047 / ADR 0103 / ADR 0184: remaining capacity, used/window counts, turn total, and provider input/output/cache/reasoning/hit-rate are the newest usage-bearing assistant message (the last model request). Occupancy is `input + output + reasoning + cacheRead + cacheWrite` on that message. They are not the sum of every model call in the visual tool-loop. Completed-turn speed and the aggregate tool row still describe that visual turn. Renderer only; host turn rollups and Token Insights stay additive billing.** | Summing cache reads across a tool loop put 367k cache read next to a 55k window. OpenCode's context widget uses only the last assistant message. See ADR 0193 and E2E-060d. |
 | D356 | Optional subagent thinking override | **Amend ADR 0062 / ADR 0144: subagent frontmatter accepts `thinkingLevel: omit` in addition to inherit and the seven canonical levels. `omit` keeps the agent bookkeeping state at `off` but uses the low-level provider stream so no thinking override is synthesized. Explicit `off` still disables thinking. Model-configuration thinking chips use an accent/inverted-text selected state in both themes. No storage schema or protocol version change. See ADR 0194 and E2E-203.** | Users need inherit, explicit off, and leave-the-provider-default as three distinct choices, and selected thinking chips were unreadable in dark mode. |
@@ -1215,9 +1216,10 @@ section mirrors only marketplace/catalog items still blocking nothing.
   duration. Runtime estimates use pi-agent-core's existing four-characters-
   per-token heuristic; provider-reported usage remains the authoritative total
   and the UI labels tool rows as estimates.
-- Generation speed is a completed-turn snapshot from provider output and final
-  stream duration; active assistant streams do not show a live token-rate
-  counter.
+- Generation speed inside the inspector is a completed-turn snapshot from
+  provider output and final stream duration *(amended by D639: while a turn is
+  running, the transcript stream-health strip shows a live sliding-window
+  `tok/s` reading)*.
 - The context-window total comes from the matching `pi-ai` model metadata used
   by the agent sidecar; provider metadata and the 128K default remain fallbacks
   for unknown models.
@@ -7384,3 +7386,16 @@ must keep splitting are covered by `markdown-blocks.test.mjs`.
   knows only Raspberry Pi CPU parts for Linux arm64. Speech-to-text and the
   rest of the app have no architecture-specific dependency. See ADR 0318,
   issue #1281, and E2E-192a.
+
+## 2026-10-01 — Live conversation token rate (D639)
+
+**Amend D212 / D244: while a turn is running, the transcript stream-health
+strip shows a live sliding-window output rate in `tok/s` next to the working
+/ run-activity indicator, which stays visible for the whole running turn.
+Prefer provider output usage when present; otherwise estimate
+from visible thinking+answer text and mark the reading as approximate. The
+rate clears when the turn becomes idle. The context inspector's Generation
+speed remains the completed-turn snapshot. Renderer only.**
+
+See `04-ux/08-component-spec.md` §8.3, `04-ux/09-interaction-patterns.md`
+§3.2a, and E2E-CHAT-live-token-rate-shows-during-stream.
diff --git a/docs/zh-CN/spec/04-ux/08-component-spec.md b/docs/zh-CN/spec/04-ux/08-component-spec.md
index 0cbae32c1..dc177332b 100644
--- a/docs/zh-CN/spec/04-ux/08-component-spec.md
+++ b/docs/zh-CN/spec/04-ux/08-component-spec.md
@@ -1302,7 +1302,7 @@ row 而不是在转录本中添加树镶边。
   一条聚合摘要，展示工具种类、调用次数和带 `~` 的估算令牌数；默认视图
   不再显示逐工具行、占比条、来源徽章或解释性提示。标题下方的行共用
   左标签/右数值节奏，只用留白分隔，弹层保留浮层描边、不画内部分隔线
-  （D297）。生成速度只代表已完成回合，流式响应期间不更新。上下文窗口
+  （D297）。检查器内的生成速度仍是已完成回合的快照（tokens/s）。回合进行中时，转录流健康条（工作中 / 运行活动指示器，整个运行回合期间保持可见）显示滑动窗口的实时 `tok/s`，以便观察重连与停顿（D639）。实时速率优先使用提供商输出用量；否则根据可见思考+答案文本估算，并标为约数。上下文窗口
   总计使用与 agent sidecar 相同的 effective model window：已发布的
   `models.dev` `limit.context` 会替换遗留的 128k 通用 binding seed，
   非默认的按模型 Advanced 值仍然优先；未知模型回退到提供商目录或默认
diff --git a/docs/zh-CN/spec/04-ux/09-interaction-patterns.md b/docs/zh-CN/spec/04-ux/09-interaction-patterns.md
index e704924c6..d62ea2324 100644
--- a/docs/zh-CN/spec/04-ux/09-interaction-patterns.md
+++ b/docs/zh-CN/spec/04-ux/09-interaction-patterns.md
@@ -608,6 +608,10 @@
 6. Composer重新激活（解锁）
 7. 中止是幂等的——当已经中止时按中止不会执行任何操作
 
+### 3.2a 运行中的实时令牌速率
+
+回合进行中时，转录状态条在工作中 / 运行活动指示器旁显示滑动窗口的实时令牌速率（`tok/s`）（该指示器在整个运行回合期间保持可见）。优先使用提供商输出用量；否则根据可见思考+答案文本估算，并标为约数。回合空闲后清除该速率；上下文检查器仍保留已完成回合的生成速度（D639）。
+
 ### 3. 3 中止用户体验
 
 - 中止按钮短暂更改为“正在中止...”（100 毫秒），然后消失
diff --git a/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md b/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
index 4c558434a..2ea68b4a1 100644
--- a/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
+++ b/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
@@ -2599,7 +2599,7 @@ Markdown 源码，不是 `text/html` 负载；对禁用行内 HTML 的外部编
   上下文窗口同量级。没有缓存读取元数据时省略缓存命中率而不是推算。
   发布的 1m 级窗口（例如 `gpt-5.6-luna` 的 1,050,000 tokens）显示为
   实际窗口而不是 128k，且与 agent runtime 使用同一个 effective window；
-  非默认 Advanced 覆盖仍然生效。生成速度只代表已完成回合。重试重新
+  非默认 Advanced 覆盖仍然生效。检查器内的生成速度仍是已完成回合快照；回合进行中时流健康条另有实时 `tok/s`（D639）。重试重新
   发送最近的用户提示，并在回合运行时禁用；body 级面板留在会话面板内，因此
   工作面板的原生浏览器或插件表面永远不会盖住它；会话面板较窄时它会随面板
   收窄，而不是越过该边界，并跟随滚动或调整大小；复制仍不包含思考文本。
@@ -5161,6 +5161,27 @@ eleven-tool-round desktop paths are verified by
   `apps/desktop/test/subagent-models.test.mjs` 和 host-core 能力测试覆盖源码/单元；
   原生选择器、渲染模态框、项目切换和运行时完整旅程仍为 Draft（除非明确要求，不在本地运行 E2E）
 
+
+#### E2E-CHAT-live-token-rate-shows-during-stream
+
+- **前提**：已配置会流式输出答案令牌的提供商会话（或按已知节拍发出文本增量的测试替身）。
+- **步骤**：
+  1. 发送一条会持续数秒流式作答的提示。
+  2. 在答案文本出现前、工作中 / 运行活动指示器可见时，观察状态条。
+  3. 答案令牌开始流式输出后，确认工作中指示器保持可见并持续显示实时速率。
+  4. 可选：暂停或节流上游流足够久，让滑动窗口排空，然后再恢复。
+  5. 让回合结束，确认实时速率清除；打开上下文用量检查器，确认已完成回合的生成速度仍显示。
+- **预期**：
+  - 实时速率以 `N tok/s`（或对应本地化文案）出现在流健康指示器旁，并在回合中更新。
+  - 流中若尚无提供商输出用量，读数标为约数（`≈` / 「约」）。
+  - 停顿会把实时速率降向 `0 tok/s`，但在回合仍活动时不清除芯片。
+  - 空闲清除实时芯片；检查器生成速度仍是已完成回合快照（D639）。
+- **链接规格**：`04-ux/08-component-spec.md` §8.3、
+  `04-ux/09-interaction-patterns.md` §3.2a、D639
+- **验收**：C（对话和直播）
+- **里程碑**：M6
+- **状态**：窗口速率数学已有单元覆盖；UI 旅程 Draft
+
 #### E2E-120：全局插件启动、下一回合编辑和停止吞吐量
 
 - **前提**：安装并启用中文显示的面板插件
@@ -5512,7 +5533,7 @@ eleven-tool-round desktop paths are verified by
 | B / F / Quality — 已选模型顺序 | E2E-MODEL-selected-order-persists |
 | A — 应用程序启动 | E2E-001、E2E-002、E2E-003、E2E-004、E2E-067、E2E-076、E2E-079、E2E-092、E2E-097、E2E-143、E2E-150、E2E-168、E2E-204、E2E-217 |
 | B——模型配置 | E2E-005、E2E-005G、E2E-006、E2E-007、E2E-038、E2E-050、E2E-052、E2E-055、E2E-066、E2E-080、E2E-082、E2E-151、E2E-005J、E2E-199、E2E-201、E2E-202、E2E-203、E2E-209、E2E-166 |
-| C — 对话和直播 | E2E-CHAT-running-status-survives-output-pauses、E2E-008、E2E-008d、E2E-008a、E2E-009、E2E-010、E2E-011、E2E-011a、E2E-011b、E2E-031、E2E-040、E2E-047、E2E-048、E2E-048A、E2E-049、E2E-052、 E2E-053、E2E-054、E2E-055、E2E-059、E2E-059a、E2E-060c、E2E-060d、E2E-061、E2E-061a、E2E-062、E2E-064、E2E-065、E2E-068、E2E-071、 E2E-073、E2E-074、E2E-075、E2E-081、E2E-083、E2E-084、E2E-086、E2E-087、E2E-088、E2E-088b、E2E-089、E2E-090、E2E-094、E2E-095、E2E-096、 E2E-097、E2E-098、E2E-099、E2E-102、E2E-102a、E2E-102b、E2E-106、E2E-109、E2E-111、E2E-114、E2E-116、E2E-117、E2E-118、E2E-119、 E2E-120、E2E-121、E2E-代理-001、E2E-142、E2E-144、E2E-145、E2E-146、E2E-147、E2E-151、E2E-199、E2E-250、E2E-166、E2E-SUBAGENT-resume-a-settled-delegation |
+| C — 对话和直播 | E2E-CHAT-running-status-survives-output-pauses、E2E-008、E2E-008d、E2E-008a、E2E-009、E2E-010、E2E-011、E2E-011a、E2E-011b、E2E-031、E2E-040、E2E-047、E2E-048、E2E-048A、E2E-049、E2E-052、 E2E-053、E2E-054、E2E-055、E2E-059、E2E-059a、E2E-060c、E2E-060d、E2E-061、E2E-061a、E2E-062、E2E-064、E2E-065、E2E-068、E2E-071、 E2E-073、E2E-074、E2E-075、E2E-081、E2E-083、E2E-084、E2E-086、E2E-087、E2E-088、E2E-088b、E2E-089、E2E-090、E2E-094、E2E-095、E2E-096、 E2E-097、E2E-098、E2E-099、E2E-102、E2E-102a、E2E-102b、E2E-106、E2E-109、E2E-111、E2E-114、E2E-116、E2E-117、E2E-118、E2E-119、 E2E-120、E2E-CHAT-live-token-rate-shows-during-stream、E2E-121、E2E-代理-001、E2E-142、E2E-144、E2E-145、E2E-146、E2E-147、E2E-151、E2E-199、E2E-250、E2E-166、E2E-SUBAGENT-resume-a-settled-delegation |
 | C — 对话和直播（输入框草稿） | E2E-011c、E2E-011c-1 |
 | A / C / F / Quality — Tray session navigation | E2E-TRAY-bounded-session-navigation |
 | D——工作区 | E2E-012、E2E-013、E2E-022B、E2E-024I、E2E-047、E2E-049、E2E-057、E2E-058、E2E-060、E2E-068、E2E-075、E2E-078、E2E-153 |
diff --git a/docs/zh-CN/spec/08-meta/decisions-log.md b/docs/zh-CN/spec/08-meta/decisions-log.md
index a128ee8ec..74485d94c 100644
--- a/docs/zh-CN/spec/08-meta/decisions-log.md
+++ b/docs/zh-CN/spec/08-meta/decisions-log.md
@@ -110,6 +110,7 @@
 | D420 | 结构化、有界且脱敏的进程日志 | **修订 ADR 0046 / ADR 0212：每条 app/host/agent NDJSON 记录都有稳定 event 和顶层关联字段。正常工具调用只产生一条完成/失败记录；sidecar 意外退出时为活动工具产生中断记录；工具协议和成绩单保持不变。中心日志会脱敏凭据与本机路径，将结构化数据限制为 8 KiB，用工具结果摘要替代原始输出，并把同一份已脱敏记录镜像到开发控制台。** | 旧的 `tool start` / `tool end` 行重复且不清楚，自由文本的子进程/错误详情还可能泄露秘密或无限增长。 |
 | D422 | 插件的宿主回合结束事件 | **`session:turnEnded` 是宿主事件，载荷为 `{ sessionId, turnId, reason }`（`completed` / `aborted` / `error`），对 `session.beginTurn` 真正开始的每个回合，在回合拆除结束时、持久化的 `session.endTurn` 尝试之后广播一次。携带的 `turnId` 是终止运行时事件本身所标识的身份，而不是恰好活动的那个回合；插件工具上下文中的 `turnId` 也会填充同一值。没有 ack，也没有重放：存活且已订阅的插件只收到一次；与崩溃、重载或宿主退出竞态的投递不作保证；收到该事件也不代表该回合的所有在途工具都已退出，因此清理必须按 `turnId` 串行化或限定作用域。不需要新权限，目前尚无任何已发布宿主发出该事件（0.14.8 也尚未包含）。见 ADR 0252。** | 驱动 GUI 的插件此前只能用空闲计时器猜测回合是否结束，而计时器会在回合中途触发、在回合结束后再次触发。宿主拥有的「每个回合一次」终止事件加上明确的回合身份，让插件可以精确结算一次，工具上下文中的同一身份还能用来关联迟到的工具结果。 |
 | D423 | 移除子智能体轮次上限 | **修订 D328 / ADR 0062 / ADR 0063 / ADR 0119 / ADR 0126 / ADR 0166 / ADR 0210：`maxTurns` 与 `MAX_SUBAGENT_MAX_TURNS` 从定义类型、frontmatter 解析器及其钳制/非法告警、`UserSubagentRecord` / `UserSubagentInput`、host-core 注册表（记录、输入、frontmatter 解析、文档渲染、`MAX_TURNS_CEILING`）、五个内置文档、`SUBAGENT_PRESETS` 以及子智能体编辑器中移除。委托只在完成、父级调用 `TaskStop`、用户 Stop，或父级终态错误中止它时结束（ADR 0189）。既有文档中的 `maxTurns` / `max-turns` / `max_turns` 现在是无法识别的 frontmatter 键，会像其他未知键一样被忽略：不报错、不告警、不让定义加载失败，也不改写用户文件。`truncated` 从 `SubagentRunStatus`、渲染器 `SubagentOutcome` 联合类型、各语言 `chat.subagentStatus` 目录项以及委托拓扑的警告计数中移除；`timed_out` 保留。不改协议版本、schema 版本或存储。见 ADR 0253、E2E-155 与 E2E-SUBAGENT-legacy-turn-limit-frontmatter-is-ignored。** | 父级看不到委托的实时工作，因此无法给出合理的轮次上限，而 60 / 50 / 40 / 80 这些内置兜底值也没有推导依据。该上限的唯一效果是把委托中途杀死，并以带部分报告的 `truncated` 呈现 —— 用户和父级模型都无法从中恢复。 |
+| D639 | 对话实时令牌速率 | **修订 D212 / D244：回合进行中时，转录流健康条显示滑动窗口的实时输出速率（`tok/s`）（工作中 / 运行活动指示器，整个运行回合期间保持可见）。优先使用提供商 `outputTokens`；否则根据可见思考+答案文本估算并标为约数。空闲时清除。上下文检查器的生成速度仍是已完成回合快照。仅渲染器改动；不改协议或持久化。** | 仅有已完成回合速度时，用户无法判断流是否健康，还是已经停顿/重连。 |
 | D244 | 紧凑的上下文用量摘要 | **修订 D103 / D184 / ADR 0047：保留上下文检查器的剩余容量触发器、已用/窗口计数、回合合计、已完成回合速度、精确的提供商数值、聚合的工具类型/调用数/令牌数以及检查点摘要，但把它们渲染为一段简短摘要。从默认面板中移除逐工具行、占比条、来源徽章、解释性估算段落和已用容量计量条。不改动协议、存储、运行时计费或模型元数据。** *（由 D347 修订：触发器移到输入框工具栏。）* | 之前的诊断式布局让一次例行的容量检查变得又高又密。保留聚合信号、移除下钻装饰，使默认状态界面可以快速浏览，同时不改变底层用量数据。参见 ADR 0103 与 E2E-060d / US-UI-61。 |
 | D347 | 输入框工具栏中的上下文用量检查器 | **修订 D103 / D184 / D244 / ADR 0047 / ADR 0103：紧凑上下文检查器放在输入框右侧工具栏、模型 × 推理芯片左侧，始终对应当前最新一条已报告用量的助手回合。触发器保留剩余容量圆环和百分比，去掉重复的 Context 文字。弹层标题为剩余 tokens + 百分比；下方行用同一套左标签/右数值节奏，只用留白分隔，不画内部分隔线（D297）。答案下方的助理元只保留模型徽章。仅渲染器改动。** | 挂在最新答案下方的检查器会随记录滚出视野。输入框只保留一个入口作为最新快照的权威位置；标题双线通过去掉多余说明文字解决，而不是加分隔线。参见 ADR 0184 与 E2E-060d / US-UI-61。 |
 | D355 | 上下文检查器按最后一次请求计算占用 | **修订 D103 / D184 / D244 / D347 / ADR 0047 / ADR 0103 / ADR 0184：剩余容量、已用/窗口计数、本轮合计，以及模型 input/output/cache/reasoning/命中率，都取最新一条已报告用量的助手消息（最后一次模型请求）。占用为该消息的 `input + output + reasoning + cacheRead + cacheWrite`。它们不是视觉工具循环里每一次请求的加总。已完成回合速度和聚合工具行仍描述该视觉回合。仅渲染器改动；宿主回合汇总和 Token Insights 仍做账单累加。** | 把工具循环里的缓存读取加总后，367k 缓存读取会紧挨着 55k 窗口。OpenCode 的上下文组件只用最后一条助手消息。参见 ADR 0193 与 E2E-060d。 |
@@ -1087,9 +1088,8 @@ project/group 层，而主要操作和页脚标识仍保留在
   持续时间。运行时间估计使用 pi-agent-core 现有的四个字符 -
   每个令牌启发式；提供商报告的使用量仍然是权威的总量
   UI 将工具行标记为估计值。
-- 生成速度是提供商输出和最终结果的完整回合快照
-  流持续时间；活动助理流不显示实时令牌率
-  柜台。
+- 检查器内的生成速度是提供商输出与最终流持续时间的已完成回合快照
+  *（由 D639 修订：回合进行中时，转录流健康条显示滑动窗口的实时 `tok/s`）*。
 - 上下文窗口总计来自所使用的匹配 `pi-ai` 模型元数据
   由Agent sidecar 提供；提供商元数据和 128K 默认值仍然是后备方案
   对于未知模型。
@@ -5192,3 +5192,9 @@ Markdown 源码，不是 `text/html` 负载；对禁用行内 HTML 的外部编
 - 已知限制：除 Raspberry Pi 板卡之外的 arm64 Linux 设备仍无法采集麦克风，因为
   `@picovoice/pvrecorder-node` 对 Linux arm64 只认识 Raspberry Pi 的 CPU part。
   语音转写和应用的其余部分没有架构相关的依赖。见 ADR 0318、issue #1281 与 E2E-192a。
+
+## 2026-10-01 — 对话实时令牌速率（D639）
+
+**修订 D212 / D244：回合进行中时，转录流健康条在工作中 / 运行活动指示器旁显示滑动窗口的实时输出速率（`tok/s`）（该指示器在整个运行回合期间保持可见）。优先使用提供商输出用量；否则根据可见思考+答案文本估算，并标为约数。回合空闲后清除该速率。上下文检查器的生成速度仍是已完成回合快照。仅渲染器改动。**
+
+见 `04-ux/08-component-spec.md` §8.3、`04-ux/09-interaction-patterns.md` §3.2a，以及 E2E-CHAT-live-token-rate-shows-during-stream。
diff --git a/packages/i18n/src/locales/de/index.ts b/packages/i18n/src/locales/de/index.ts
index e81087131..0846622fb 100644
--- a/packages/i18n/src/locales/de/index.ts
+++ b/packages/i18n/src/locales/de/index.ts
@@ -372,6 +372,8 @@ export const de = {
     "usageThroughput": "{{count}} Tokens/s",
     "usageThroughputEstimated": "≈ {{count}} Tokens/s",
     "usageThroughputUnavailable": "–",
+    "usageLiveThroughput": "{{count}} Tok/s",
+    "usageLiveThroughputEstimated": "≈ {{count}} Tok/s",
     "usageProviderUsage": "Anbieternutzung",
     "usageInput": "Eingabe",
     "usageOutput": "Ausgabe",
diff --git a/packages/i18n/src/locales/en/index.ts b/packages/i18n/src/locales/en/index.ts
index 73d5536f6..dbb3b341a 100644
--- a/packages/i18n/src/locales/en/index.ts
+++ b/packages/i18n/src/locales/en/index.ts
@@ -387,6 +387,8 @@ export const en = {
     usageThroughput: "{{count}} tokens/s",
     usageThroughputEstimated: "≈ {{count}} tokens/s",
     usageThroughputUnavailable: "—",
+    usageLiveThroughput: "{{count}} tok/s",
+    usageLiveThroughputEstimated: "≈ {{count}} tok/s",
     usageProviderUsage: "Provider usage",
     usageInput: "Input",
     usageOutput: "Output",
diff --git a/packages/i18n/src/locales/es/index.ts b/packages/i18n/src/locales/es/index.ts
index 67de6a28e..035d13ab6 100644
--- a/packages/i18n/src/locales/es/index.ts
+++ b/packages/i18n/src/locales/es/index.ts
@@ -372,6 +372,8 @@ export const es = {
     "usageThroughput": "{{count}} tokens/s",
     "usageThroughputEstimated": "≈ {{count}} tokens/s",
     "usageThroughputUnavailable": "—",
+    "usageLiveThroughput": "{{count}} tok/s",
+    "usageLiveThroughputEstimated": "≈ {{count}} tok/s",
     "usageProviderUsage": "Uso del proveedor",
     "usageInput": "Entrada",
     "usageOutput": "Salida",
diff --git a/packages/i18n/src/locales/fr/index.ts b/packages/i18n/src/locales/fr/index.ts
index 0d62e9e76..e2446bc4d 100644
--- a/packages/i18n/src/locales/fr/index.ts
+++ b/packages/i18n/src/locales/fr/index.ts
@@ -372,6 +372,8 @@ export const fr = {
     "usageThroughput": "{{count}} jetons/s",
     "usageThroughputEstimated": "≈ {{count}} jetons/s",
     "usageThroughputUnavailable": "—",
+    "usageLiveThroughput": "{{count}} jetons/s",
+    "usageLiveThroughputEstimated": "≈ {{count}} jetons/s",
     "usageProviderUsage": "Utilisation du fournisseur",
     "usageInput": "Entrée",
     "usageOutput": "Sortie",
diff --git a/packages/i18n/src/locales/ko/index.ts b/packages/i18n/src/locales/ko/index.ts
index 14ad6c0f8..8baa45792 100644
--- a/packages/i18n/src/locales/ko/index.ts
+++ b/packages/i18n/src/locales/ko/index.ts
@@ -384,6 +384,8 @@ export const ko = {
     usageThroughput: "{{count}} tokens/s",
     usageThroughputEstimated: "≈ {{count}} tokens/s",
     usageThroughputUnavailable: "—",
+    usageLiveThroughput: "{{count}} tok/s",
+    usageLiveThroughputEstimated: "≈ {{count}} tok/s",
     usageProviderUsage: "프로바이더 사용량",
     usageInput: "입력",
     usageOutput: "출력",
diff --git a/packages/i18n/src/locales/pt-BR/index.ts b/packages/i18n/src/locales/pt-BR/index.ts
index eba81c34f..5f10f30c5 100644
--- a/packages/i18n/src/locales/pt-BR/index.ts
+++ b/packages/i18n/src/locales/pt-BR/index.ts
@@ -372,6 +372,8 @@ export const ptBR = {
     usageThroughputLabel: "Velocidade de geração",
     usageThroughput: "{{count}} tokens/s",
     usageThroughputEstimated: "≈ {{count}} tokens/s",
+    usageLiveThroughput: "{{count}} tok/s",
+    usageLiveThroughputEstimated: "≈ {{count}} tok/s",
     usageThroughputUnavailable: "—",
     usageProviderUsage: "Uso do provedor",
     usageInput: "Entrada",
diff --git a/packages/i18n/src/locales/tr/index.ts b/packages/i18n/src/locales/tr/index.ts
index 84fd5b68a..6eef6c1b0 100644
--- a/packages/i18n/src/locales/tr/index.ts
+++ b/packages/i18n/src/locales/tr/index.ts
@@ -384,6 +384,8 @@ export const tr = {
     usageThroughput: "{{count}} token/sn",
     usageThroughputEstimated: "≈ {{count}} token/sn",
     usageThroughputUnavailable: "—",
+    usageLiveThroughput: "{{count}} tok/sn",
+    usageLiveThroughputEstimated: "≈ {{count}} tok/sn",
     usageProviderUsage: "Servis kullanımı",
     usageInput: "Girdi",
     usageOutput: "Çıktı",
diff --git a/packages/i18n/src/locales/zh-CN/index.ts b/packages/i18n/src/locales/zh-CN/index.ts
index c0f1141b1..5580d49a4 100644
--- a/packages/i18n/src/locales/zh-CN/index.ts
+++ b/packages/i18n/src/locales/zh-CN/index.ts
@@ -376,6 +376,8 @@ export const zhCN = {
     usageThroughput: "{{count}} tokens/s",
     usageThroughputEstimated: "约 {{count}} tokens/s",
     usageThroughputUnavailable: "—",
+    usageLiveThroughput: "{{count}} tok/s",
+    usageLiveThroughputEstimated: "约 {{count}} tok/s",
     usageProviderUsage: "模型用量",
     usageInput: "输入",
     usageOutput: "输出",
diff --git a/packages/i18n/src/locales/zh-TW/index.ts b/packages/i18n/src/locales/zh-TW/index.ts
index 46afa84c3..361d0ae28 100644
--- a/packages/i18n/src/locales/zh-TW/index.ts
+++ b/packages/i18n/src/locales/zh-TW/index.ts
@@ -376,6 +376,8 @@ export const zhTW = {
     usageThroughput: "{{count}} tokens/s",
     usageThroughputEstimated: "約 {{count}} tokens/s",
     usageThroughputUnavailable: "—",
+    usageLiveThroughput: "{{count}} tok/s",
+    usageLiveThroughputEstimated: "約 {{count}} tok/s",
     usageProviderUsage: "模型用量",
     usageInput: "輸入",
     usageOutput: "輸出",