feat(web): live header statistics while a task runs (#75)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

packages/web/e2e/chat.spec.mjs

@@ -59,6 +59,14 @@ test("chat + tool approval + stats/cost/copy + traces + files", async ({ page })
   await expect(page.getByText("exec_command").first()).toBeVisible();
   // Thinking + tool calls are wrapped in a work group; header shows running/done status.
   await expect(page.getByText("运行中").first()).toBeVisible();
+
+  // Live header statistics: the elapsed chip ticks once per second while the task runs (the
+  // pending approval below keeps it running), so its text must advance with no further server
+  // event. Scoped to the header stats container (div.hidden …): the per-reply footer reuses
+  // the same 用时 ("elapsed") label once the turn's stats line lands.
+  const headerElapsed = page.locator('div.hidden span[title="用时"]');
+  const elapsedBefore = await headerElapsed.textContent();
+  await expect(headerElapsed).not.toHaveText(elapsedBefore);
   // The user takes control of the running work group (toggle = userToggled), keeps it open, and
   // opens the exec_command card to watch the arguments. Both must survive the end of the turn.
   //

packages/web/src/features/chat/chat-page.tsx

@@ -12,6 +12,7 @@
  * created. The Session list and the new-chat entry point live in the global sidebar.
  */
 import { useCallback, useEffect, useRef, useState } from "react";
+import type { ReactNode } from "react";
 import { useNavigate, useParams } from "react-router";
 import type {
   AgentSummary,
@@ -27,9 +28,18 @@ import { ApiError } from "../../api/client";
 import { S } from "../../lib/strings";
 import { apiErrorText } from "../../lib/api-error";
 import { useDocumentTitle } from "../../lib/use-document-title";
-import { formatDateTime, formatMoney, humanizeDuration, humanizeTokens } from "../../lib/format";
+import {
+  formatDateTime,
+  formatMoney,
+  humanizeDuration,
+  humanizeDurationLive,
+  humanizeTokens,
+} from "../../lib/format";
 import { latestConversation } from "../../lib/session-grouping";
 import { approvalKey, isModelAuthDead } from "../../lib/omni/stream-model";
+import type { StreamModel } from "../../lib/omni/stream-model";
+import { bucketCostUsd, liveSessionElapsedMs } from "../../lib/omni/task-stats";
+import type { BucketPricing, TaskStatsTracker } from "../../lib/omni/task-stats";
 import { useTheme } from "../../state/theme";
 import { useProject } from "../../state/project";
 import { useSessions } from "../../state/sessions";
@@ -65,7 +75,7 @@ const STAT_ICONS = {
 } as const;
 
 /** Iconized stat item: a symbol + a value, with the title giving the full meaning. */
-function StatChip({ icon, value, label }: { icon: string; value: string; label: string }) {
+function StatChip({ icon, value, label }: { icon: string; value: ReactNode; label: string }) {
   return (
     <span
       title={label}
@@ -89,6 +99,101 @@ function StatChip({ icon, value, label }: { icon: string; value: string; label:
   );
 }
 
+/**
+ * Elapsed value for the header statistics: while a Task runs it ticks once per second over the
+ * live cumulative (settled cross-Task total + the running Task's wall clock so far, see
+ * liveSessionElapsedMs); when idle no timer runs and it renders exactly the settled total.
+ * Whole seconds while ticking, decimals only on the settled value — same convention as
+ * LiveDuration on running tool/thinking cards.
+ */
+function SessionElapsed({
+  stats,
+  taskOpen,
+  taskStartLocalMs,
+}: {
+  stats: TaskStatsTracker;
+  taskOpen: boolean;
+  taskStartLocalMs: number;
+}) {
+  const [now, setNow] = useState(() => Date.now());
+  useEffect(() => {
+    if (!taskOpen) return;
+    // Load-bearing, not redundant: `now` still holds whatever the state last saw (mount time,
+    // or the final tick of a previous Task), and the first interval callback is a full second
+    // away. The first live render after a Task starts must not compute from that stale clock,
+    // so re-anchor immediately on entering the running state.
+    setNow(Date.now());
+    const id = setInterval(() => setNow(Date.now()), 1000);
+    return () => clearInterval(id);
+  }, [taskOpen]);
+  if (!taskOpen) return <>{humanizeDuration(stats.sessionElapsedMs)}</>;
+  return <>{humanizeDurationLive(liveSessionElapsedMs(stats, taskOpen, taskStartLocalMs, now))}</>;
+}
+
+/** The header's three statistics — the chip row and the info dropdown render these verbatim. */
+interface HeaderStats {
+  tokensText: string;
+  /** Formatted session cost; null = nothing to show (no recorded figure and no live estimate). */
+  costText: string | null;
+  /** Server-reported "some usage had no pricing" flag from the last idle refetch (the chip's `*`). */
+  costUncosted: boolean;
+  elapsedNode: ReactNode;
+}
+
+/**
+ * Computes the header statistics, live while a Task runs:
+ *   - Tokens: session cumulative (main + subagents), already advancing per completed request;
+ *   - Cost: the last idle-refetched session cost, plus — while a Task is open and the session
+ *     Model has pricing — a live estimate converted from this Task's usage buckets. Subagents
+ *     may run on different models, but the estimate applies the main Model's pricing to all
+ *     live buckets: the estimate may be slightly off mid-task, and the idle getUsage refetch
+ *     reconciles to the server-recorded value (kept deliberately simple). Without pricing the
+ *     live addition is skipped (bucketCostUsd returns null, mirroring taskCost's uncosted
+ *     signal) and the value stays exactly as when idle. costText stays null until there is an
+ *     actual figure — no recorded cost plus a still-zero estimate renders nothing, not $0.00;
+ *   - Elapsed: ticking cumulative while running, settled cumulative when idle (SessionElapsed).
+ */
+function headerStats(
+  model: StreamModel,
+  sessionCost: number | null,
+  costUncosted: boolean,
+  pricing: BucketPricing | undefined,
+  currency: "USD" | "CNY",
+): HeaderStats {
+  const stats = model.stats;
+  const liveCost = model.taskOpen
+    ? bucketCostUsd(
+        {
+          cacheRead: stats.taskCacheRead,
+          cacheWrite: stats.taskCacheWrite,
+          output: stats.taskOutput,
+        },
+        pricing,
+      )
+    : null;
+  // Only take the live path once it has something to say — a positive estimate, or a recorded
+  // session cost to keep showing from the Task's first instant. On a brand-new session,
+  // sessionCost is still null and liveCost is 0 until the first token_usage lands; blindly
+  // summing would flash a formatted $0.00 the moment the Task starts — exactly the "cost is
+  // zero or something's broken" reading the chip's render conditional exists to avoid.
+  const costUsd =
+    liveCost != null && (liveCost > 0 || sessionCost != null)
+      ? (sessionCost ?? 0) + liveCost
+      : sessionCost;
+  return {
+    tokensText: humanizeTokens(stats.sessionTotal + stats.subagentTotal),
+    costText: costUsd != null ? formatMoney(costUsd, currency) : null,
+    costUncosted,
+    elapsedNode: (
+      <SessionElapsed
+        stats={stats}
+        taskOpen={model.taskOpen}
+        taskStartLocalMs={model.taskStartLocalMs}
+      />
+    ),
+  };
+}
+
 /**
  * Route id for a draft chat (`/chat/new`): the Session hasn't been persisted yet — the user may
  * still want to change the model or configure a key first. The actual Session is only created
@@ -611,16 +716,7 @@ export function ChatPage() {
     // happen mid-turn, and if only running were checked, the trailing group would flash
     // "finished running" during compaction before flipping back to "running".
     taskRunning: stream.taskState !== "idle",
-    taskCost: (stats) => {
-      if (!modelPricing) return null;
-      const b = stats.tokensByBucket;
-      return (
-        (b.cacheRead * modelPricing.cacheRead +
-          b.cacheWrite * modelPricing.cacheWrite +
-          b.output * modelPricing.output) /
-        1e6
-      );
-    },
+    taskCost: (stats) => bucketCostUsd(stats.tokensByBucket, modelPricing),
     // Reconnect countdown controls (live waiting state only): retry-now skips the
     // remaining backoff server-side (benign no-op on timing races), give-up is the
     // ordinary session abort — the engine's abort-during-backoff path ends the turn.
@@ -649,7 +745,9 @@ export function ChatPage() {
     );
   }
 
-  const totalTokens = stream.model.stats.sessionTotal + stream.model.stats.subagentTotal;
+  // Header statistics (chip row + info dropdown), live while a Task runs; recomputed every
+  // stream version bump, so the in-place-mutated model stats always read fresh.
+  const hs = headerStats(stream.model, sessionCost, costUncosted, modelPricing, currency);
   const modelInfo = models?.models.find((m) => sameModelRef(m, activeModelRef));
   const contextWindow = modelInfo?.contextWindow;
   // Assumed supported by default: only models explicitly marked vision=false show a blocking hint when adding images.
@@ -742,24 +840,20 @@ export function ChatPage() {
           <div className="hidden items-center gap-3 sm:flex">
             <StatChip
               icon={STAT_ICONS.tokens}
-              value={humanizeTokens(totalTokens)}
+              value={hs.tokensText}
               label={`${S.chat.statTokens}（Token）`}
             />
             {/* When there's no cost (the Model has no pricing configured), don't render this stat
                 at all, rather than showing a "—" — that would take up space while saying
                 nothing, only making people think the cost is zero or something's broken. */}
-            {sessionCost != null && (
+            {hs.costText != null && (
               <StatChip
                 icon={STAT_ICONS.cost}
-                value={`${formatMoney(sessionCost, currency)}${costUncosted ? " *" : ""}`}
-                label={`${S.common.cost}（${currency}）${costUncosted ? ` · ${S.usage.uncostedNote}` : ""}`}
+                value={`${hs.costText}${hs.costUncosted ? " *" : ""}`}
+                label={`${S.common.cost}（${currency}）${hs.costUncosted ? ` · ${S.usage.uncostedNote}` : ""}`}
               />
             )}
-            <StatChip
-              icon={STAT_ICONS.elapsed}
-              value={humanizeDuration(stream.model.stats.sessionElapsedMs)}
-              label={S.chat.statElapsed}
-            />
+            <StatChip icon={STAT_ICONS.elapsed} value={hs.elapsedNode} label={S.chat.statElapsed} />
           </div>
 
           {/* Files panel toggle: docks on the right of the chat instead of replacing it full-screen (use-files-panel.ts). */}
@@ -848,10 +942,9 @@ export function ChatPage() {
                 </p>
                 {/* Same as above: if there's no cost, the whole item is omitted, not left as "Cost —". */}
                 <p className="font-mono text-xs">
-                  {S.chat.statTokens} {humanizeTokens(totalTokens)}
-                  {sessionCost != null &&
-                    ` · ${S.common.cost} ${formatMoney(sessionCost, currency)}`}{" "}
-                  · {S.chat.statElapsed} {humanizeDuration(stream.model.stats.sessionElapsedMs)}
+                  {S.chat.statTokens} {hs.tokensText}
+                  {hs.costText != null && ` · ${S.common.cost} ${hs.costText}`} ·{" "}
+                  {S.chat.statElapsed} {hs.elapsedNode}
                 </p>
               </div>
             </div>

packages/web/src/lib/omni/task-stats.ts

@@ -42,6 +42,32 @@ export interface TokenBucketCounts {
   output: number;
 }
 
+/** Three pricing buckets in USD per million tokens (mirrors the server's ModelPricingDto shape). */
+export interface BucketPricing {
+  cacheRead: number;
+  cacheWrite: number;
+  output: number;
+}
+
+/**
+ * Converts a three-bucket Token count to USD at per-million-token pricing; null when no pricing
+ * is configured (uncosted — callers omit the figure rather than fabricating a $0). Shared by the
+ * per-reply stats row's turn cost and the header's live session cost, so the two conversions can
+ * never drift apart.
+ */
+export function bucketCostUsd(
+  buckets: TokenBucketCounts,
+  pricing: BucketPricing | null | undefined,
+): number | null {
+  if (!pricing) return null;
+  return (
+    (buckets.cacheRead * pricing.cacheRead +
+      buckets.cacheWrite * pricing.cacheWrite +
+      buckets.output * pricing.output) /
+    1e6
+  );
+}
+
 /** Stats tracker: Session-level cumulative counts that persist across Tasks + this-Task delta counts. */
 export interface TaskStatsTracker {
   /** Current context occupancy = total from the most recent main-session request (not updated by compaction requests). */
@@ -287,6 +313,25 @@ export function endTask(t: TaskStatsTracker, elapsedMs: number): TaskStats | nul
   return stats;
 }
 
+/**
+ * Session elapsed time for live display (the header chip): the settled cross-Task cumulative
+ * plus the currently running Task's wall clock so far. While no Task is open this is exactly
+ * `sessionElapsedMs`, so the idle rendering equals the settled value; the running addition
+ * counts from the model's task-start local clock (`taskStartLocalMs`), clamped so a clock
+ * anomaly never makes the sum go backwards. At task end {@link endTask} folds the Task's
+ * elapsed into `sessionElapsedMs` in the same model update that flips `taskOpen` off, so the
+ * live addition never double-counts across the boundary.
+ */
+export function liveSessionElapsedMs(
+  t: TaskStatsTracker,
+  taskOpen: boolean,
+  taskStartLocalMs: number | null,
+  nowMs: number,
+): number {
+  const running = taskOpen && taskStartLocalMs != null ? Math.max(0, nowMs - taskStartLocalMs) : 0;
+  return t.sessionElapsedMs + running;
+}
+
 /** Localized labels for {@link formatTaskStats} (supplied by the view layer's active dictionary). */
 export interface TaskStatsLabels {
   /** Row prefix, e.g. "Stats" / "统计信息". */

packages/web/test/task-stats.test.ts

@@ -7,11 +7,13 @@ import type { TokenUsagePayload } from "@prismshadow/penguin-core/omnimessage";
 import {
   addLlmDuration,
   beginCompaction,
+  bucketCostUsd,
   commitPendingCompaction,
   createTaskStatsTracker,
   endCompaction,
   endTask,
   formatTaskStats,
+  liveSessionElapsedMs,
   resetTaskCounters,
   trackMainUsage,
   trackSubagentUsage,
@@ -194,6 +196,67 @@ describe("TaskStatsTracker", () => {
   });
 });
 
+describe("liveSessionElapsedMs", () => {
+  it("idle (no open Task): exactly the settled cumulative — the header renders the same value as before", () => {
+    const t = createTaskStatsTracker();
+    t.sessionElapsedMs = 2300;
+    expect(liveSessionElapsedMs(t, false, 1000, 99_999)).toBe(2300);
+    // Open but without a recorded start clock: nothing to add either.
+    expect(liveSessionElapsedMs(t, true, null, 99_999)).toBe(2300);
+  });
+
+  it("running: settled cumulative + wall clock since the Task started, never going backwards", () => {
+    const t = createTaskStatsTracker();
+    t.sessionElapsedMs = 2000;
+    expect(liveSessionElapsedMs(t, true, 5000, 8000)).toBe(5000); // 2000 + 3000
+    // A clock anomaly (now before the recorded start) adds nothing instead of subtracting.
+    expect(liveSessionElapsedMs(t, true, 5000, 4000)).toBe(2000);
+  });
+
+  it("no double count across the Task boundary: endTask folds the Task in as taskOpen flips off", () => {
+    const t = createTaskStatsTracker();
+    // Task 1 settled earlier.
+    endTask(t, 1000);
+    // Task 2 runs: started at 10_000, now 14_000 -> 1000 settled + 4000 live.
+    expect(liveSessionElapsedMs(t, true, 10_000, 14_000)).toBe(5000);
+    // Task 2 ends: the same model update folds its elapsed into the cumulative AND flips
+    // taskOpen off — the live view continues from the settled value without re-adding.
+    endTask(t, 4000);
+    expect(liveSessionElapsedMs(t, false, 10_000, 15_000)).toBe(5000);
+  });
+});
+
+describe("bucketCostUsd", () => {
+  it("converts the three buckets at per-million-token pricing", () => {
+    expect(
+      bucketCostUsd(
+        { cacheRead: 2_000_000, cacheWrite: 1_000_000, output: 500_000 },
+        { cacheRead: 0.5, cacheWrite: 2, output: 10 },
+      ),
+    ).toBe(8); // 2M×$0.5/M + 1M×$2/M + 0.5M×$10/M = 1 + 2 + 5
+  });
+
+  it("no pricing -> null (uncosted; callers keep the value as-is instead of fabricating $0)", () => {
+    const buckets = { cacheRead: 100, cacheWrite: 100, output: 100 };
+    expect(bucketCostUsd(buckets, undefined)).toBeNull();
+    expect(bucketCostUsd(buckets, null)).toBeNull();
+  });
+
+  it("the live Task buckets equal the settled stats row's buckets (the mid-task estimate lands on the final turn cost)", () => {
+    const t = createTaskStatsTracker();
+    trackMainUsage(t, req(300, 100, 200));
+    trackSubagentUsage(t, req(50, 0, 3));
+    const pricing = { cacheRead: 1, cacheWrite: 2, output: 3 };
+    const live = bucketCostUsd(
+      { cacheRead: t.taskCacheRead, cacheWrite: t.taskCacheWrite, output: t.taskOutput },
+      pricing,
+    );
+    const s = endTask(t, 100);
+    expect(live).not.toBeNull();
+    expect(live).toBe(bucketCostUsd(s!.tokensByBucket, pricing));
+  });
+});
+
 describe("formatTaskStats", () => {
   const EN_LABELS = {
     stats: "Stats",