Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

revert: restore pre-#5734 response speed experience

Chris5a9437e6cc86
diff --git a/apps/desktop/src/main/__tests__/makerEventHotPathOrdering.test.ts b/apps/desktop/src/main/__tests__/makerEventHotPathOrdering.test.ts
index 60f4c7332e..f556c5229e 100644
--- a/apps/desktop/src/main/__tests__/makerEventHotPathOrdering.test.ts
+++ b/apps/desktop/src/main/__tests__/makerEventHotPathOrdering.test.ts
@@ -1060,8 +1060,8 @@ describe('maker:event hot path ordering', () => {
const claudeCostFallback = claudeDoneSource.slice(
claudeDoneSource.indexOf("} else if (typeof cumulative === 'number' && cumulative >= 0)"),
);
- expect(claudeCostFallback).not.toMatch(
- /buildClaudeTurnUsageDetails\(\s*doneData\??\.usage/,
+ expect(claudeCostFallback).toMatch(
+ /buildClaudeTurnUsageDetails\(\s*undefined,\s*undefined,\s*resolvedModel,/,
);
expect(claudeCostFallback).toContain(
"if (route !== 'provider-api' || turnContext.accessKind === 'managed')",
diff --git a/apps/desktop/src/main/maker-ipc/__tests__/sessionClaudeTurnUsage.test.ts b/apps/desktop/src/main/maker-ipc/__tests__/sessionClaudeTurnUsage.test.ts
deleted file mode 100644
index 87b0ded265..0000000000
--- a/apps/desktop/src/main/maker-ipc/__tests__/sessionClaudeTurnUsage.test.ts
+++ /dev/null
@@ -1,107 +0,0 @@
-import { readFileSync } from 'node:fs';
-import { resolve } from 'node:path';
-import { createSourceFile, isFunctionDeclaration, ScriptKind, ScriptTarget, transpileModule } from 'typescript';
-import { describe, expect, it, vi } from 'vitest';
-import { calibratedResponseDuration } from '@cindy/maker-shared/usage-format';
-import type { AgentEvent, Session } from '@cindy/maker-core';
-import type { RecordSessionClaudeTurnUsageDeps } from '../sessionClaudeTurnUsage.js';
-import { buildClaudeTurnUsageDetails, normalizeTurnUsageSegments } from '../../usage/turnCostCalculator.js';
-import { ClaudeOutputLagTimingGuard, computeModelUsageDeltas } from '../../usage/modelUsageDelta.js';
-
-// Execute the production sink with database/network effects isolated, retaining
-// real usage normalization, timing validation and message-detail construction.
-const source = readFileSync(resolve(__dirname, '../sessionClaudeTurnUsage.ts'), 'utf8').replace(/\r\n?/g, '\n');
-const ast = createSourceFile('sessionClaudeTurnUsage.ts', source, ScriptTarget.Latest, true, ScriptKind.TS);
-const declaration = ast.statements.find(node => isFunctionDeclaration(node) && node.name?.text === 'recordSessionClaudeTurnUsage');
-if (!declaration) throw new Error('recordSessionClaudeTurnUsage not found');
-const compiled = transpileModule(declaration.getText(ast), {
- compilerOptions: { target: ScriptTarget.ES2022 },
-}).outputText.replace('export function ', 'function ');
-
-function harness() {
- const record = vi.fn<(value: { turnUsageDetails: unknown }) => Promise<void>>().mockResolvedValue(undefined);
- const guard = new ClaudeOutputLagTimingGuard();
- const recordSpend = vi.fn();
- const runtime = {
- calibratedResponseDuration, normalizeTurnUsageSegments, buildClaudeTurnUsageDetails, computeModelUsageDeltas,
- captureTurnUsageContext: () => ({ providerId: null }),
- recordTurnUsageOnMessage: record,
- triggerClaudeAccountUsageRefresh: vi.fn(), triggerClaudeSubscriptionUsageRefresh: vi.fn(),
- getSessionProvider: () => undefined,
- CHATGPT_MODEL_PREFIX: 'chatgpt/', isExclusiveXaiModelId: () => false,
- detectClaudeModelMismatch: () => null,
- billingRouteForExplicitProvider: () => null, readClaudeSessionRoute: () => 'subscription',
- readClaudeApiKey: () => undefined, getReferenceModelPricing: () => null,
- CURRENT_CINDY_REGION: 'global',
- resolveClaudeTurnCostSinks: () => ({ turnMoney: null, estimatedTurnMoney: null, perModel: [] }),
- recordTurnSpend: recordSpend,
- };
- const wire = new Function('runtime', `const {${Object.keys(runtime).join(',')}} = runtime; ${compiled}; return recordSessionClaudeTurnUsage;`)(runtime) as
- typeof import('../sessionClaudeTurnUsage.js').recordSessionClaudeTurnUsage;
- const deps: RecordSessionClaudeTurnUsageDeps = {
- turnUsageContextBySession: new Map(), turnModelPromiseBySession: new Map(),
- readSessionModelForUsage: async () => 'claude-sonnet-4.5',
- lastReportedModelUsageBySession: new Map(), claudeOutputLagTimingGuard: guard,
- lastReportedCostUsdBySession: new Map(), log: { warn: vi.fn() },
- unpricedSubscriptionValueMarker: () => ({ amount: 0, currency: 'USD', kind: 'value-estimate', approximate: true }),
- };
- const send = (data: unknown) => wire(deps, { id: 'task' } as Session,
- { type: 'done', source: 'claude-code', data } as AgentEvent, 'assistant', 9_000, false);
- return { send, record, guard, deps, recordSpend };
-}
-
-function done(output = 60, cumulative = 160, durationMs = 3_000) {
- return {
- total_cost_usd: 0, usage: { input_tokens: 500, output_tokens: cumulative, cache_read_input_tokens: 200 },
- turnUsage: { input_tokens: 15, output_tokens: output, cache_read_input_tokens: 5, cache_creation_input_tokens: 2 },
- responseSpeed: { phase: 'complete', waitOrigin: 'stream', firstResponseMs: 1000, waitingMs: 0,
- durationMs, outputTokens: output, estimated: false, averageRate: output * 1000 / durationMs,
- recentRate: output * 1000 / durationMs, samples: [], sampledAt: 10_000 },
- };
-}
-
-describe('Claude completed usage without modelUsage', () => {
- it('persists current-turn TPS across cumulative results without charging or copying aggregate tokens', async () => {
- const h = harness();
- h.send(done(100, 100, 2_000));
- await vi.waitFor(() => expect(h.record).toHaveBeenCalledTimes(1));
- h.send(done());
- await vi.waitFor(() => expect(h.record).toHaveBeenCalledTimes(2));
- const first = h.record.mock.calls[0][0].turnUsageDetails;
- const second = h.record.mock.calls[1][0].turnUsageDetails;
- expect(first).toMatchObject({ outputTokens: 100, durationMs: 2_000 });
- expect(second).toMatchObject({ inputTokens: 15, outputTokens: 60, cacheReadTokens: 5,
- cacheCreateTokens: 2, durationMs: 3_000, turnDurationMs: 9_000 });
- expect(h.deps.lastReportedCostUsdBySession.get('task')).toBe(0);
- expect(h.recordSpend).not.toHaveBeenCalled();
- });
-
- it('uses the normalized delta when partial request metadata has no priced model rows', async () => {
- const h = harness();
- h.send({ ...done(), usageSegments: [{ model: 'claude-sonnet-4.5', inputTokens: 5, outputTokens: 10 }],
- usageSegmentsComplete: false });
- await vi.waitFor(() => expect(h.record).toHaveBeenCalledTimes(1));
- expect(h.record.mock.calls[0][0].turnUsageDetails).toMatchObject({ inputTokens: 15,
- outputTokens: 60, cacheReadTokens: 5, cacheCreateTokens: 2, durationMs: 3_000 });
- expect(h.recordSpend).not.toHaveBeenCalled();
- });
-
- it.each([undefined, { output_tokens: 160 }, { output_tokens: Number.NaN }, { output_tokens: -1 }])(
- 'omits timing when the normalized provider count is unavailable or mismatched (%j)', async (turnUsage) => {
- const h = harness(); h.send({ ...done(), turnUsage });
- await vi.waitFor(() => expect(h.record).toHaveBeenCalledTimes(1));
- expect(h.record.mock.calls[0][0].turnUsageDetails).toMatchObject({ outputTokens: 0 });
- expect(h.record.mock.calls[0][0].turnUsageDetails).not.toHaveProperty('durationMs');
- },
- );
-
- it('keeps a real zero and respects output-lag suppression', async () => {
- const h = harness(); h.send(done(0, 160));
- await vi.waitFor(() => expect(h.record).toHaveBeenCalledTimes(1));
- expect(h.record.mock.calls[0][0].turnUsageDetails).toMatchObject({ outputTokens: 0, durationMs: 3_000 });
- vi.spyOn(h.guard, 'evaluate').mockReturnValue({ detected: false, suppressTiming: true });
- h.send(done());
- await vi.waitFor(() => expect(h.record).toHaveBeenCalledTimes(2));
- expect(h.record.mock.calls[1][0].turnUsageDetails).not.toHaveProperty('durationMs');
- });
-});
diff --git a/apps/desktop/src/main/maker-ipc/sessionClaudeTurnUsage.ts b/apps/desktop/src/main/maker-ipc/sessionClaudeTurnUsage.ts
index af54783afa..a9621fd9ce 100644
--- a/apps/desktop/src/main/maker-ipc/sessionClaudeTurnUsage.ts
+++ b/apps/desktop/src/main/maker-ipc/sessionClaudeTurnUsage.ts
@@ -1,4 +1,3 @@
-import { calibratedResponseDuration } from '@cindy/maker-shared/usage-format';
import { captureTurnUsageContext, type TurnUsageContext } from './turnUsageContext.js';
import { isOpenAiSubscriptionProviderId } from '../maker-host/codex-account-auth.js';
import type { AgentEvent, Session } from '@cindy/maker-core';
@@ -100,13 +99,6 @@ export function recordSessionClaudeTurnUsage(
total_cost_usd?: unknown;
duration_ms?: unknown;
duration_api_ms?: unknown;
- responseSpeed?: unknown;
- turnUsage?: {
- input_tokens?: number;
- output_tokens?: number;
- cache_read_input_tokens?: number;
- cache_creation_input_tokens?: number;
- };
usage?: {
input_tokens?: number;
output_tokens?: number;
@@ -175,18 +167,11 @@ export function recordSessionClaudeTurnUsage(
: undefined,
doneData?.is_error !== true,
);
- const claudeTurnOutputTokens = modelUsageDeltas?.length
- ? modelUsageDeltas.reduce((sum, delta) => sum + delta.outputTokensDelta, 0)
- : typeof doneData?.turnUsage?.output_tokens === 'number' ? doneData.turnUsage.output_tokens : undefined;
const claudeGenerationDurationMs = outputLagTiming.suppressTiming
? undefined
- : doneData?.responseSpeed !== undefined
- ? claudeTurnOutputTokens !== undefined
- ? calibratedResponseDuration(doneData.responseSpeed, claudeTurnOutputTokens)
- : undefined
- : typeof doneData?.duration_api_ms === 'number'
- ? doneData.duration_api_ms
- : undefined;
+ : typeof doneData?.duration_api_ms === 'number'
+ ? doneData.duration_api_ms
+ : undefined;
// total_cost_usd 累计基线: 主路径不靠它算钱, 但仍跟住, 以便万一某轮缺 modelUsage
// 走兜底时累计差才准。先取"更新前"基线给兜底用, 再写入本轮累计。
const prevReportedCost = deps.lastReportedCostUsdBySession.get(session.id);
@@ -344,7 +329,7 @@ export function recordSessionClaudeTurnUsage(
// deltas 非空 → buildClaudeTurnUsageDetails 用 deltas 里的 model, fallbackModel 不取用。
// 传 perModel → 落「按模型成本明细」(含 subagent 跑的模型, 如 Haiku)。
const turnUsageDetails = buildClaudeTurnUsageDetails(
- doneData?.turnUsage ?? doneData?.usage,
+ doneData?.usage,
resolvedUsageDeltas,
'unknown',
perModel,
@@ -378,7 +363,7 @@ export function recordSessionClaudeTurnUsage(
const turnEstimatedValue =
estimatedValues.length > 0 ? addRegionalMoney(estimatedValues) : null;
const turnUsageDetails = buildClaudeTurnUsageDetails(
- doneData?.turnUsage ?? doneData?.usage,
+ doneData?.usage,
resolvedUsageDeltas,
'unknown',
perModel,
@@ -423,11 +408,9 @@ export function recordSessionClaudeTurnUsage(
}
// `doneData.usage` can be the same process-lifetime cumulative snapshot as
// `total_cost_usd`. Without model deltas or request segments it is not a reliable
- // per-turn token fact. Persist the normalized provider delta only when
- // its output matches the calibrated measurement; never the aggregate.
+ // per-turn token fact, so retain only model/timing metadata in this fallback.
const turnUsageDetails = buildClaudeTurnUsageDetails(
- claudeGenerationDurationMs !== undefined && claudeTurnOutputTokens !== undefined
- ? doneData?.turnUsage : undefined,
+ undefined,
undefined,
resolvedModel,
undefined,
@@ -435,7 +418,7 @@ export function recordSessionClaudeTurnUsage(
claudeTurnDurationMs,
);
// 本分支有三个"记不了钱"的出口(本轮 cost 未增长 / 订阅直连 / 非明确
- // provider-api 路由)。只保留可证明的模型、输出与时长;进程累计 usage 不能冒充本轮 token。
+ // provider-api 路由)。只保留可证明的模型与时长;进程累计 usage 不能冒充本轮 token。
const recordUsageOnly = async () => {
if (!turnAssistantPersistId) return;
await recordTurnUsageOnMessage({
diff --git a/apps/desktop/src/renderer/__tests__/controlledBannerPlacement.test.ts b/apps/desktop/src/renderer/__tests__/controlledBannerPlacement.test.ts
index fe492d3175..82edb97886 100644
--- a/apps/desktop/src/renderer/__tests__/controlledBannerPlacement.test.ts
+++ b/apps/desktop/src/renderer/__tests__/controlledBannerPlacement.test.ts
@@ -56,7 +56,7 @@ describe('controlled banner placement', () => {
'const controlledBannerMaxWidth = `min(${inputHalfWidth}, ${CONTROLLED_BANNER_MAX_WIDTH}px)`;',
);
expect(sessionViewSource).toContain(
- 'if (!rightLeadingSlot && (suppressContent || (isHidden && !ratePanelPinned && !completedSpeed))) return null;',
+ 'if (!rightLeadingSlot && (suppressContent || (isHidden && !ratePanelPinned))) return null;',
);
expect(controlledBannerSource).toContain("placement?: 'floating' | 'inline' | 'composer';");
expect(controlledBannerSource).toContain(
diff --git a/apps/desktop/src/renderer/__tests__/makerChatStoreLiveGeneration.test.ts b/apps/desktop/src/renderer/__tests__/makerChatStoreLiveGeneration.test.ts
index 458d6b3cdf..27ec716536 100644
--- a/apps/desktop/src/renderer/__tests__/makerChatStoreLiveGeneration.test.ts
+++ b/apps/desktop/src/renderer/__tests__/makerChatStoreLiveGeneration.test.ts
@@ -1,11 +1,6 @@
-import { describe, expect, it, vi } from 'vitest';
-
-vi.mock('@/lib/makerTransport', () => ({
- makerApiFor: () => ({ input: { stop: vi.fn(async () => ({ queue: [], paused: false })) } }),
-}));
+import { describe, expect, it } from 'vitest';
import { makerChatStore } from '@/lib/makerChatStore';
-import type { ResponseSpeedSnapshot } from '@cindy/maker-shared/usage-format';
function applyStatus(
sessionId: string,
@@ -17,7 +12,6 @@ function applyStatus(
generationDurationMs?: number;
generationActive?: boolean;
generationReliable?: boolean;
- responseSpeed?: ResponseSpeedSnapshot;
},
): void {
makerChatStore.__applyStatusUpdateForTest(sessionId, {
@@ -27,7 +21,6 @@ function applyStatus(
contextTokens: 0,
contextWindow: 0,
isRunning: partial.isRunning,
- ...(partial.responseSpeed ? { responseSpeed: partial.responseSpeed } : {}),
...(partial.outputTokens !== undefined ? { outputTokens: partial.outputTokens } : {}),
...(partial.generationDurationMs !== undefined
? { generationDurationMs: partial.generationDurationMs }
@@ -42,26 +35,6 @@ function applyStatus(
}
describe('makerChatStore live generation at turn start', () => {
- it('retains an interrupted measurement across stop and task switching, then resets on the next turn', () => {
- const sessionId = `live-gen-stop-${Math.random().toString(36).slice(2, 8)}`;
- const speed: ResponseSpeedSnapshot = { phase: 'generating', waitOrigin: 'turn', firstResponseMs: 1000,
- waitingMs: 0, outputTokens: 50, durationMs: 2000, estimated: true, recentRate: 30,
- averageRate: 25, samples: [{ durationMs: 2000, outputTokens: 50, rate: 30 }], sampledAt: Date.now() };
- try {
- applyStatus(sessionId, { isRunning: true, responseSpeed: speed });
- makerChatStore.stopSession(sessionId);
- makerChatStore.getSnapshot('another-task');
- expect(makerChatStore.getSnapshot(sessionId).agentStatus).toMatchObject({
- isRunning: false, generationActive: false,
- responseSpeed: { phase: 'complete', estimated: true, recentRate: null, outcome: 'cancelled', samples: speed.samples },
- });
- applyStatus(sessionId, { isRunning: true });
- expect(makerChatStore.getSnapshot(sessionId).agentStatus.responseSpeed).toBeUndefined();
- } finally {
- makerChatStore.purgeSession(sessionId);
- makerChatStore.purgeSession('another-task');
- }
- });
it('keeps live fields when the first running status already carries them', () => {
const sessionId = `live-gen-keep-${Math.random().toString(36).slice(2, 8)}`;
try {
@@ -108,35 +81,3 @@ describe('makerChatStore live generation at turn start', () => {
}
});
});
-
-it('keeps a response speed failure through done/status tails and uses only actual retry events', () => {
- const id = 'speed-error-retry-lifecycle';
- const speed: ResponseSpeedSnapshot = { phase: 'generating', waitOrigin: 'turn', firstResponseMs: 1000,
- waitingMs: 0, outputTokens: 50, durationMs: 2000, estimated: true, recentRate: 30,
- averageRate: 25, samples: [{ durationMs: 2000, outputTokens: 50, rate: 30 }], sampledAt: Date.now() };
- const event = (type: 'error' | 'done' | 'text', data: Record<string, unknown>) =>
- makerChatStore.__applyStreamEventForTest(id, { sessionId: id, type, source: 'pi', data });
- try {
- applyStatus(id, { isRunning: true, responseSpeed: speed });
- event('error', { message: 'API unavailable', reason: 'api-error', isTerminal: false });
- expect(makerChatStore.getSnapshot(id).agentStatus.responseSpeed?.retrying).not.toBe(true);
- event('error', { message: 'API unavailable; retrying', reason: 'api-error', willRetry: true, isTerminal: false });
- expect(makerChatStore.getSnapshot(id).agentStatus.responseSpeed).toMatchObject({ retrying: true, recentRate: null });
- applyStatus(id, { isRunning: true, responseSpeed: speed });
- expect(makerChatStore.getSnapshot(id).agentStatus.responseSpeed?.retrying).toBe(true);
- event('text', { text: 'recovered', isFinal: false });
- expect(makerChatStore.getSnapshot(id).agentStatus.responseSpeed?.retrying).toBe(false);
- event('error', { message: 'API unavailable', reason: 'api-error', isTerminal: true });
- expect(makerChatStore.getSnapshot(id)).toMatchObject({ error: 'API unavailable', errorReason: 'api-error',
- agentStatus: { isRunning: false, responseSpeed: { phase: 'complete', outcome: 'failed', recentRate: null, samples: speed.samples } } });
- event('done', {});
- applyStatus(id, { isRunning: false, responseSpeed: { ...speed, phase: 'complete', estimated: false } });
- expect(makerChatStore.getSnapshot(id).agentStatus.responseSpeed?.outcome).toBe('failed');
- applyStatus(id, { isRunning: true });
- expect(makerChatStore.getSnapshot(id).agentStatus.responseSpeed).toBeUndefined();
- applyStatus(id, { isRunning: true, responseSpeed: speed });
- event('done', { type: 'pi/agent_settled', status: 'cancelled' });
- expect(makerChatStore.getSnapshot(id)).toMatchObject({ error: null,
- agentStatus: { responseSpeed: { outcome: 'cancelled', recentRate: null } } });
- } finally { makerChatStore.purgeSession(id); }
-});
diff --git a/apps/desktop/src/renderer/__tests__/runningStatusBarSuppression.test.tsx b/apps/desktop/src/renderer/__tests__/runningStatusBarSuppression.test.tsx
index cf1c222146..0857964e55 100644
--- a/apps/desktop/src/renderer/__tests__/runningStatusBarSuppression.test.tsx
+++ b/apps/desktop/src/renderer/__tests__/runningStatusBarSuppression.test.tsx
@@ -2,9 +2,8 @@
import { readFileSync } from 'node:fs';
import { resolve } from 'node:path';
import React, { useCallback, useEffect, useRef, useState } from 'react';
-import { act, cleanup, fireEvent, render, screen } from '@testing-library/react';
+import { cleanup, fireEvent, render, screen } from '@testing-library/react';
import ts from 'typescript';
-import { responseSpeedActivity } from '@cindy/maker-shared/usage-format';
import { afterEach, expect, it, vi } from 'vitest';
import { formatSessionDuration } from '@/lib/sessionDurationFormat';
import {
@@ -17,9 +16,7 @@ import {
resolveRunningUsageMeta,
} from '@/features/cc-agent/lib/runningTokenUsage';
-vi.mock('react-i18next', () => ({ useTranslation: () => ({
- t: (key: string, values?: { value?: unknown; rate?: unknown }) => `${key}${values?.value ?? values?.rate ?? ''}`,
-}) }));
+vi.mock('react-i18next', () => ({ useTranslation: () => ({ t: (key: string) => key }) }));
// Execute the actual status bar without loading the entire session view's IPC/store graph.
const source = readFileSync(
@@ -43,14 +40,11 @@ const compiled = ts.transpileModule(component.getText(ast), {
const Icon = () => null;
const deps = {
React,
- responseSpeedActivity,
useCallback,
useEffect,
useRef,
useState,
- useTranslation: () => ({
- t: (key: string, values?: { value?: unknown; rate?: unknown }) => `${key}${values?.value ?? values?.rate ?? ''}`,
- }),
+ useTranslation: () => ({ t: (key: string) => key }),
useReducedMotion: () => true,
useAnimatedNumber: (value: number) => value,
localizeAgentStatus: (status: string) => status,
@@ -75,40 +69,19 @@ const RunningStatusBar = new Function(
`const {${Object.keys(deps).join(',')}} = deps; ${compiled}; return RunningStatusBar;`,
)(deps) as React.ComponentType<{
visible: boolean;
- responseSpeed?: import('@cindy/maker-shared/usage-format').ResponseSpeedSnapshot;
suppressContent?: boolean;
rightLeadingSlot?: React.ReactNode;
status: string;
reconnectStatus?: string | null;
- startedAt: number | null;
+ startedAt: number;
tokenUsage: number;
outputTokens: number;
generationDurationMs: number;
- generationActive?: boolean;
generationReliable?: boolean;
}>;
afterEach(cleanup);
-it('clears an older host recent interval immediately when native generation pauses, keeping history', () => {
- const props = { visible: true, status: 'Generating...', startedAt: Date.now(), tokenUsage: 0,
- outputTokens: 0, generationDurationMs: 0, generationReliable: true, generationActive: true };
- const view = render(<RunningStatusBar {...props} />);
- const measured = { ...props, outputTokens: 370, generationDurationMs: 1000 };
- view.rerender(<RunningStatusBar {...measured} />);
- const trigger = () => view.container.querySelector('[data-running-status-meta] button')!;
- expect(trigger().textContent).toContain('chat.runningStatus.tokenRate');
- fireEvent.click(trigger());
- view.rerender(<RunningStatusBar {...measured} generationActive={false} status="Running bash…" />);
- expect(trigger().textContent).not.toContain('chat.runningStatus.tokenRate');
- expect(screen.getByRole('dialog').textContent).toContain('370');
- expect(screen.getByRole('dialog').textContent).toContain('chat.runningStatus.observedPeak');
- view.rerender(<RunningStatusBar {...measured} generationActive />);
- expect(trigger().textContent).not.toContain('chat.runningStatus.tokenRate');
- view.rerender(<RunningStatusBar {...measured} generationActive outputTokens={470} generationDurationMs={2000} />);
- expect(trigger().textContent).toContain('chat.runningStatus.tokenRate');
-});
-
it.each([false, true])('reconnect hides stale speed and pinned history, then waits for fresh samples (pinned=%s)', (pinned) => {
const props = {
visible: true, status: 'Generating...', startedAt: 1, tokenUsage: 1000,
@@ -183,15 +156,9 @@ it('opens measured zero history and restores fallback across reliability and tur
expect(trigger()).not.toBeNull();
});
-it.each(['running', 'stopped', 'completed'] as const)(
- 'plan review preserves the collapsed indicator and suppresses a pinned panel (%s)',
- (state) => {
- const visible = state === 'running';
- const responseSpeed = state === 'completed' ? {
- phase: 'complete' as const, waitOrigin: 'turn' as const, firstResponseMs: 1000, waitingMs: 0,
- durationMs: 1000, outputTokens: 100, estimated: false, averageRate: 100, recentRate: 100,
- samples: [{ durationMs: 1000, outputTokens: 100, rate: 100 }], sampledAt: Date.now(),
- } : undefined;
+it.each([true, false])(
+ 'plan review preserves the collapsed indicator and suppresses a pinned panel (running=%s)',
+ (visible) => {
const props = {
visible: true,
status: 'Thinking',
@@ -211,7 +178,6 @@ it.each(['running', 'stopped', 'completed'] as const)(
<RunningStatusBar
{...props}
visible={visible}
- responseSpeed={responseSpeed}
suppressContent
rightLeadingSlot={indicator}
/>,
@@ -221,102 +187,7 @@ it.each(['running', 'stopped', 'completed'] as const)(
expect(screen.queryByRole('button', { name: 'titleBar.close' })).toBeNull();
// Explicit suppression without an independent indicator leaves no row behind.
- rerender(<RunningStatusBar {...props} visible={visible} responseSpeed={responseSpeed} suppressContent />);
+ rerender(<RunningStatusBar {...props} visible={visible} suppressContent />);
expect(container.childElementCount).toBe(0);
},
);
-
-it('keeps the control indicator before completed speed metadata but leaves an idle row empty', () => {
- const props = { visible: false, status: 'Done', startedAt: null,
- tokenUsage: 0, outputTokens: 0, generationDurationMs: 0 };
- const view = render(<RunningStatusBar {...props} />);
- expect(view.container.childElementCount).toBe(0);
- const indicator = <button aria-label="Controlled session">Device</button>;
- view.rerender(<RunningStatusBar {...props} rightLeadingSlot={indicator} />);
- const control = screen.getByRole('button', { name: 'Controlled session' });
- expect(view.container.querySelector('[data-running-status-meta]')).toBeNull();
- view.rerender(<RunningStatusBar {...props} rightLeadingSlot={indicator}
- responseSpeed={{ phase: 'complete', waitOrigin: 'turn', firstResponseMs: 1000, waitingMs: 0,
- durationMs: 1000, outputTokens: 100, estimated: false, averageRate: 100, recentRate: 100,
- samples: [{ durationMs: 1000, outputTokens: 100, rate: 100 }], sampledAt: Date.now() }} />);
- const meta = view.container.querySelector('[data-running-status-meta]')!;
- expect(meta.textContent).toContain('chat.runningStatus.lastGeneration');
- expect(control.compareDocumentPosition(meta) & Node.DOCUMENT_POSITION_FOLLOWING).toBeTruthy();
- fireEvent.click(meta.querySelector('button')!);
- expect(screen.getByRole('dialog').textContent).toContain('chat.runningStatus.finalAverage');
-});
-
-it('retains a completed response after fade, navigation/remount, then replaces it at the next turn', () => {
- vi.useFakeTimers();
- try {
- const speed = { phase: 'complete', waitOrigin: 'turn', firstResponseMs: 2000, waitingMs: 0,
- durationMs: 4000, outputTokens: 300, estimated: false, averageRate: 75, recentRate: 75,
- samples: [{ durationMs: 1000, outputTokens: 100, rate: 100 }], sampledAt: Date.now() } as const;
- const props = { visible: false, status: 'Done', startedAt: null, tokenUsage: 500,
- outputTokens: 300, generationDurationMs: 4000, responseSpeed: { ...speed, samples: [...speed.samples] } };
- const view = render(<RunningStatusBar {...props} />);
- act(() => vi.advanceTimersByTime(60_000));
- const trigger = () => view.container.querySelector('[data-running-status-meta] button');
- expect(trigger()?.textContent).toContain('chat.runningStatus.lastGeneration');
- fireEvent.click(trigger()!);
- expect(screen.getByRole('dialog').textContent).toContain('chat.runningStatus.finalAverage');
- expect(screen.getByRole('dialog').textContent).toContain('75');
- view.unmount();
- const restored = render(<RunningStatusBar {...props} />);
- expect(restored.container.textContent).toContain('chat.runningStatus.lastGeneration');
- restored.rerender(<RunningStatusBar {...props} visible startedAt={Date.now()}
- responseSpeed={{ ...props.responseSpeed, phase: 'waiting', firstResponseMs: null, waitingMs: 0,
- outputTokens: 0, durationMs: 0, recentRate: null, averageRate: null, samples: [] }} />);
- expect(restored.container.textContent).not.toContain('chat.runningStatus.lastGeneration');
- expect(restored.container.textContent).toContain('chat.runningStatus.responseWaiting');
- } finally { vi.useRealTimers(); }
-});
-
-it('expires a stalled live rate while preserving an inspectable curve, and distinguishes tool pause', () => {
- vi.useFakeTimers();
- try {
- const speed = { phase: 'generating', waitOrigin: 'turn', firstResponseMs: 1000, waitingMs: 0,
- durationMs: 2000, outputTokens: 80, estimated: true, averageRate: 40, recentRate: 370,
- samples: [{ durationMs: 2000, outputTokens: 80, rate: 370 }], sampledAt: Date.now() } as const;
- const props = { visible: true, status: 'Generating...', startedAt: Date.now(), tokenUsage: 500,
- outputTokens: 0, generationDurationMs: 0, responseSpeed: { ...speed, samples: [...speed.samples] } };
- const view = render(<RunningStatusBar {...props} />);
- fireEvent.click(view.container.querySelector('[data-running-status-meta] button')!);
- expect(screen.getByRole('dialog').textContent).toContain('chat.runningStatus.currentRate');
- act(() => vi.advanceTimersByTime(1000));
- expect(screen.getByRole('dialog').textContent).toContain('chat.runningStatus.responsePending');
- expect(view.container.querySelector('[data-running-status-meta] button')?.textContent)
- .not.toContain('chat.runningStatus.estimatedTokenRate');
- view.rerender(<RunningStatusBar {...props} status="Tool running"
- responseSpeed={{ ...props.responseSpeed, phase: 'paused', recentRate: null }} />);
- expect(screen.getByRole('dialog').textContent).toContain('chat.runningStatus.generationPaused');
- view.rerender(<RunningStatusBar {...props} status="Running bash…"
- responseSpeed={{ ...props.responseSpeed, phase: 'paused', toolActive: true, recentRate: null }} />);
- expect(view.container.textContent).toContain('chat.runningStatus.toolRunning');
- expect(screen.getByRole('dialog').textContent).toContain('chat.runningStatus.toolRunning');
- view.rerender(<RunningStatusBar {...props}
- responseSpeed={{ ...props.responseSpeed, phase: 'waiting', waitingMs: 1000, recentRate: null }} />);
- expect(view.container.textContent).toContain('chat.runningStatus.responsePending');
- expect(screen.getByRole('dialog').textContent).toContain('chat.runningStatus.responsePending');
- } finally { vi.useRealTimers(); }
-});
-
-it.each(['failed', 'cancelled', 'retrying'] as const)('shows %s consistently in the retained entry and card', (activity) => {
- const terminal = activity !== 'retrying';
- const key = activity === 'failed' ? 'responseFailed' : activity === 'cancelled' ? 'responseCancelled' : 'responseRetrying';
- const speed: import('@cindy/maker-shared/usage-format').ResponseSpeedSnapshot = {
- phase: terminal ? 'complete' : 'paused', waitOrigin: 'turn', firstResponseMs: 1000, waitingMs: 0,
- durationMs: 2000, outputTokens: 80, estimated: true, recentRate: null, averageRate: 40,
- samples: [{ durationMs: 2000, outputTokens: 80, rate: 40 }], sampledAt: Date.now(),
- ...(terminal ? { outcome: activity as 'failed' | 'cancelled' } : { retrying: true }),
- };
- const view = render(<RunningStatusBar visible={!terminal} status="Done" startedAt={null}
- tokenUsage={500} outputTokens={80} generationDurationMs={2000} responseSpeed={speed} />);
- const trigger = view.container.querySelector('[data-running-status-meta] button')!;
- expect(trigger.textContent).toContain(`chat.runningStatus.${key}`);
- expect(trigger.textContent).not.toContain('chat.runningStatus.tokenRate');
- fireEvent.click(trigger);
- expect(screen.getByRole('dialog').textContent).toContain(`chat.runningStatus.${key}`);
- expect(screen.getByRole('dialog').textContent).not.toContain('chat.runningStatus.finalAverage');
- expect(screen.getByRole('dialog').textContent).toContain('40');
-});
diff --git a/apps/desktop/src/renderer/__tests__/runningTokenRatePopover.test.tsx b/apps/desktop/src/renderer/__tests__/runningTokenRatePopover.test.tsx
index 8d8176257f..258f619711 100644
--- a/apps/desktop/src/renderer/__tests__/runningTokenRatePopover.test.tsx
+++ b/apps/desktop/src/renderer/__tests__/runningTokenRatePopover.test.tsx
@@ -32,7 +32,7 @@ it('expires the displayed rate without losing history, including after remount',
const view = render(<Probe tokens={0} />);
view.rerender(<Probe tokens={100} />);
expect(screen.getByTestId('rate').textContent).toBe('100:1');
- act(() => vi.advanceTimersByTime(999));
+ act(() => vi.advanceTimersByTime(59_999));
expect(screen.getByTestId('rate').textContent).toBe('100:1');
act(() => vi.advanceTimersByTime(1));
expect(screen.getByTestId('rate').textContent).toBe('waiting:1');
diff --git a/apps/desktop/src/renderer/features/cc-agent/CCAgentSessionView.tsx b/apps/desktop/src/renderer/features/cc-agent/CCAgentSessionView.tsx
index f05b923b61..ad46527ed2 100644
--- a/apps/desktop/src/renderer/features/cc-agent/CCAgentSessionView.tsx
+++ b/apps/desktop/src/renderer/features/cc-agent/CCAgentSessionView.tsx
@@ -1,4 +1,3 @@
-import { responseSpeedActivity, type ResponseSpeedSnapshot } from "@cindy/maker-shared/usage-format";
import { formatSessionDuration } from '@/lib/sessionDurationFormat';
import { shouldShowOpenPathError } from '../../../shared/openPathResult';
import { shouldShowFailedScheduleNotice } from '@cindy/maker-shared/schedule-model';
@@ -5154,9 +5153,7 @@ export function CCAgentSessionView({
reconnectStatus={reconnectStatus}
tokenUsage={agentStatus.tokenUsage}
outputTokens={agentStatus.outputTokens ?? 0}
- responseSpeed={agentStatus.responseSpeed}
generationDurationMs={agentStatus.generationDurationMs ?? 0}
- generationActive={agentStatus.generationActive}
generationReliable={agentStatus.generationReliable ?? true}
startedAt={agentStatus.startedAt}
visible={composerRuntimeVisible || (!pendingPlanReview && activeReconnect !== null)}
@@ -6041,9 +6038,7 @@ function RunningStatusBar({
reconnectStatus = null,
tokenUsage,
outputTokens = 0,
- responseSpeed,
generationDurationMs = 0,
- generationActive,
generationReliable = true,
startedAt,
visible,
@@ -6064,9 +6059,7 @@ function RunningStatusBar({
reconnectStatus?: string | null;
tokenUsage: number;
outputTokens?: number;
- responseSpeed?: ResponseSpeedSnapshot;
generationDurationMs?: number;
- generationActive?: boolean;
generationReliable?: boolean;
startedAt: number | null;
visible: boolean;
@@ -6153,23 +6146,13 @@ function RunningStatusBar({
const isHidden = suppressContent || (!showContent && !visible);
const workflowWaiting = workflowStatus !== undefined;
- const completedSpeed = !visible && responseSpeed?.phase === 'complete'
- && !reconnecting && !sideTaskRunning && !backgroundTasksRunning && !workflowWaiting;
// side-task / 后台子任务运行中永远当成进行态 (即便上一轮 LLM 留下的 status 文案
// 是 "Done", 此时任务还在跑, 显示 ✓ 完成图标会让用户以为已经做完)。
- const isDone = status === 'Done' && !responseSpeed?.outcome && !responseSpeed?.retrying && !reconnecting && !sideTaskRunning && !backgroundTasksRunning;
+ const isDone = status === 'Done' && !reconnecting && !sideTaskRunning && !backgroundTasksRunning;
// 后台子任务模式的左段文案:上一轮残留的 status(多半是 "Done")在此语义下是
// 误导信息,整体替换为后台运行提示。仅后台 Bash 时用带数量的专属文案 ——
// 「模型用量仍在消耗」对不调模型的 bash 任务是错误陈述。
- const speedActivity = responseSpeed ? responseSpeedActivity(responseSpeed) : null;
- const speedStatusKey = speedActivity === 'failed' ? 'responseFailed'
- : speedActivity === 'cancelled' ? 'responseCancelled'
- : speedActivity === 'retrying' ? 'responseRetrying'
- : speedActivity === 'waiting' || speedActivity === 'quiet' ? 'responsePending'
- : speedActivity === 'tool' ? 'toolRunning'
- : speedActivity === 'paused' ? 'generationPaused'
- : speedActivity === 'generating' ? 'responseGenerating' : null;
const displayStatus =
reconnectStatus ??
workflowStatus ??
@@ -6177,8 +6160,7 @@ function RunningStatusBar({
? backgroundBashOnlyCount > 0
? t('chat.backgroundActivity.bashStatus', { count: backgroundBashOnlyCount })
: t('chat.backgroundActivity.status')
- : speedStatusKey && !status.toLowerCase().startsWith('compact')
- ? t(`chat.runningStatus.${speedStatusKey}`) : localizeAgentStatus(status, t));
+ : localizeAgentStatus(status, t));
// F-COMPACT-1: when SDK is auto-summarizing the conversation, give the
// status bar a distinct icon so the user can tell "Compacting..." apart
// from "Thinking..." — both share the shimmer animation by design, but
@@ -6242,37 +6224,30 @@ function RunningStatusBar({
tokens: formatRunningTokenCount(animatedTokens),
});
const rateHistory = useRunningTokenRateHistory({
- responseSpeed,
- generationActive,
sessionKey,
startedAt,
outputTokens,
generationDurationMs,
generationReliable:
- !responseSpeed && generationReliable &&
+ generationReliable &&
!reconnecting &&
!sideTaskRunning &&
!backgroundTasksRunning &&
!workflowWaiting,
});
- const latestRate = !responseSpeed && generationActive === false ? null : rateHistory.latestRate;
- const legacyUsageMeta = resolveRunningUsageMeta({
+ const latestRate = rateHistory.latestRate;
+ const usageMeta = resolveRunningUsageMeta({
outputTokens,
generationDurationMs,
generationReliable,
tokenUsage,
latestRate,
});
- const usageMeta = responseSpeed
- ? responseSpeed.averageRate !== null
- ? { kind: 'rate' as const, rate: formatRecentOutputTokenRate(responseSpeed.averageRate) ?? '0' }
- : { kind: 'none' as const }
- : legacyUsageMeta;
const latestRateText = latestRate !== null ? formatRecentOutputTokenRate(latestRate) : null;
const rateText =
- (!isHidden || completedSpeed) && !reconnecting && usageMeta.kind === 'rate'
+ !isHidden && !reconnecting && usageMeta.kind === 'rate'
? latestRateText !== null
- ? t(responseSpeed && (responseSpeed.phase !== 'complete' || responseSpeed.estimated) ? 'chat.runningStatus.estimatedTokenRate' : 'chat.runningStatus.tokenRate', { rate: latestRateText })
+ ? t('chat.runningStatus.tokenRate', { rate: latestRateText })
: null
: null;
@@ -6285,19 +6260,17 @@ function RunningStatusBar({
transition: isHidden ? 'none' : `opacity ${STATUS_BAR_FADE_MS}ms ease-out`,
pointerEvents: isHidden ? 'none' : 'auto',
};
- const waitingText = responseSpeed?.phase === 'waiting'
- ? t('chat.runningStatus.responseWaiting', { seconds: ((responseSpeed.waitingMs + Math.max(0, Date.now() - responseSpeed.sampledAt)) / 1000).toFixed(1) }) : null;
const showRatePanel =
!reconnecting &&
(ratePanelPinned ||
(!workflowWaiting &&
!sideTaskRunning &&
!backgroundTasksRunning &&
- (Boolean(responseSpeed) || usageMeta.kind === 'rate')));
+ usageMeta.kind === 'rate'));
// A pinned panel keeps its anchor mounted through idle and subsequent turns.
- // 已完成测量保留原速度入口;没有测量的空闲任务仍收起。overlay 的 ResizeObserver 会在
+ // 空闲后真正收起,不再给输入框上方留下固定空行。overlay 的 ResizeObserver 会在
// DOM 尺寸变化后补齐 MessageStream 的 bottomPadding,因此不靠硬编码高度制造跳变。
- if (!rightLeadingSlot && (suppressContent || (isHidden && !ratePanelPinned && !completedSpeed))) return null;
+ if (!rightLeadingSlot && (suppressContent || (isHidden && !ratePanelPinned))) return null;
// 两段式布局:左(运行状态) / 右(elapsed·tokens)。
// - 左段 min-w-0(可收缩):status 并非短枚举 —— turn-start 文案带用户名(可含中文长句)、
@@ -6359,21 +6332,16 @@ function RunningStatusBar({
走 LLM, 显示残留 token 计数会误导用户以为也耗了 token。 */}
<div className="flex min-w-0 items-center justify-self-end gap-2">
{rightLeadingSlot}
- {(!suppressContent && (!isHidden || ratePanelPinned || completedSpeed)) && (
+ {(!suppressContent && (!isHidden || ratePanelPinned)) && (
<div
data-running-status-meta="true"
className="flex min-w-0 items-center gap-[6px]"
- style={ratePanelPinned || completedSpeed ? undefined : fadeStyle}
- aria-hidden={isHidden && !ratePanelPinned && !completedSpeed}
+ style={ratePanelPinned ? undefined : fadeStyle}
+ aria-hidden={isHidden && !ratePanelPinned}
>
{showRatePanel && (
<RunningTokenRatePopover
- responseSpeed={responseSpeed}
- elapsedText={completedSpeed
- ? t(responseSpeed?.outcome === 'failed' ? 'chat.runningStatus.responseFailed'
- : responseSpeed?.outcome === 'cancelled' ? 'chat.runningStatus.responseCancelled' : 'chat.runningStatus.lastGeneration')
- : speedActivity === 'retrying' ? t('chat.runningStatus.responseRetrying')
- : waitingText ?? (speedActivity === 'quiet' ? t('chat.runningStatus.responsePending') : elapsedText)}
+ elapsedText={elapsedText}
rate={latestRateText}
rateText={
workflowWaiting || sideTaskRunning || backgroundTasksRunning
@@ -6384,8 +6352,8 @@ function RunningStatusBar({
}
isTokenCount={usageMeta.kind === 'tokens'}
averageRate={usageMeta.kind === 'rate' ? usageMeta.rate : null}
- outputTokens={responseSpeed?.outputTokens ?? outputTokens}
- history={latestRate === rateHistory.latestRate ? rateHistory : { ...rateHistory, latestRate }}
+ outputTokens={outputTokens}
+ history={rateHistory}
onPinnedChange={setRatePanelPinned}
/>
)}
diff --git a/apps/desktop/src/renderer/features/cc-agent/RunningTokenRatePopover.tsx b/apps/desktop/src/renderer/features/cc-agent/RunningTokenRatePopover.tsx
index b9e431524e..665e3ddf13 100644
--- a/apps/desktop/src/renderer/features/cc-agent/RunningTokenRatePopover.tsx
+++ b/apps/desktop/src/renderer/features/cc-agent/RunningTokenRatePopover.tsx
@@ -1,5 +1,3 @@
-import type { ResponseSpeedSnapshot } from "@cindy/maker-shared/usage-format";
-import { responseSpeedActivity, responseSpeedHistory } from "@cindy/maker-shared/usage-format";
import { useEffect, useState } from 'react';
import { useTranslation } from 'react-i18next';
import { Activity, ArrowDown, X } from 'lucide-react';
@@ -17,15 +15,13 @@ import {
export function useRunningTokenRateHistory(input: {
/** 会话身份:用于进程内缓存速度历史,切任务再切回不清零(重启应用清零)。 */
- responseSpeed?: ResponseSpeedSnapshot;
sessionKey: string | null;
startedAt: number | null;
outputTokens: number;
generationDurationMs: number;
generationReliable: boolean;
- generationActive?: boolean;
}) {
- const { sessionKey, startedAt, outputTokens, generationDurationMs, generationReliable, generationActive } = input;
+ const { sessionKey, startedAt, outputTokens, generationDurationMs, generationReliable } = input;
// 挂载时从按会话的进程内缓存播种:RunningStatusBar 以 sessionId 为 key,
// 切走再切回是全新挂载,历史从缓存恢复而不是从零开始。
const [history, setHistory] = useState<RateHistory>(() => {
@@ -36,39 +32,36 @@ export function useRunningTokenRateHistory(input: {
return startedAt === null ? { ...cached, baseline: null, lastReport: null } : cached;
});
useEffect(() => {
- setHistory((previous) => {
- const recorded = recordRunningTokenRate(previous, {
+ setHistory((previous) =>
+ recordRunningTokenRate(previous, {
startedAt,
outputTokens,
generationDurationMs,
generationReliable,
- });
- return generationActive === false ? { ...recorded, latestRate: null, latestSampleAt: undefined } : recorded;
- });
- }, [startedAt, outputTokens, generationDurationMs, generationReliable, generationActive]);
+ }),
+ );
+ }, [startedAt, outputTokens, generationDurationMs, generationReliable]);
useEffect(() => {
if (sessionKey) saveCachedRateHistory(sessionKey, history);
}, [sessionKey, history]);
const [now, setNow] = useState(Date.now);
useEffect(() => {
- const timestamp = input.responseSpeed?.sampledAt ?? history.latestSampleAt;
- if (timestamp === undefined || (input.responseSpeed ? input.responseSpeed.phase !== 'generating' : history.latestRate === null)) return;
+ const timestamp = history.latestSampleAt;
+ if (timestamp === undefined || history.latestRate === null) return;
setNow(Date.now());
const timer = setTimeout(() => setNow(Date.now()),
Math.max(0, timestamp + RATE_SAMPLE_FRESH_MS - Date.now()));
return () => clearTimeout(timer);
- }, [history.latestSampleAt, history.latestRate, input.responseSpeed]);
+ }, [history.latestSampleAt, history.latestRate]);
const visibleHistory = history.latestSampleAt === undefined ||
Math.max(now, Date.now()) - history.latestSampleAt >= RATE_SAMPLE_FRESH_MS
? { ...history, latestRate: null } : history;
- if (input.responseSpeed) return responseSpeedHistory(input.responseSpeed, Math.max(now, Date.now()));
return startedAt === null || history.startedAt === startedAt
? visibleHistory
: { ...history, startedAt, baseline: null, latestRate: null };
}
export function RunningTokenRatePopover({
- responseSpeed,
elapsedText,
rate,
rateText,
@@ -78,7 +71,6 @@ export function RunningTokenRatePopover({
history,
onPinnedChange,
}: {
- responseSpeed?: ResponseSpeedSnapshot;
elapsedText: string;
rate: string | null;
rateText: string | null;
@@ -92,8 +84,6 @@ export function RunningTokenRatePopover({
const [mode, setMode] = useState<'idle' | 'hover' | 'pinned' | 'dismissed'>('idle');
const open = mode === 'pinned';
useEffect(() => onPinnedChange?.(open), [open, onPinnedChange]);
- const approximate = (value: string | null, estimated = true) => value && estimated
- ? t('chat.runningStatus.estimatedValue', { value }) : value;
const samples = history.samples;
const firstTime = samples[0]?.durationMs ?? 0;
const span = (samples.at(-1)?.durationMs ?? 0) - firstTime;
@@ -104,7 +94,6 @@ export function RunningTokenRatePopover({
}));
const line = points.map((point, index) => `${index ? 'L' : 'M'}${point.x},${point.y}`).join(' ');
const last = points.at(-1);
- const activity = responseSpeed ? responseSpeedActivity(responseSpeed) : null;
const card = (
<div
aria-description={t('chat.runningStatus.tokenRateDescription')}
@@ -114,17 +103,10 @@ export function RunningTokenRatePopover({
<div className="col-start-1 row-start-1 min-w-0 self-center">
<div className="mb-1 flex items-center gap-1.5 text-12 text-[var(--text-secondary)]">
<Activity size={14} aria-hidden="true" />
- {t(activity === 'failed' ? 'chat.runningStatus.responseFailed'
- : activity === 'cancelled' ? 'chat.runningStatus.responseCancelled'
- : activity === 'retrying' ? 'chat.runningStatus.responseRetrying'
- : activity === 'complete' ? 'chat.runningStatus.finalAverage'
- : activity === 'waiting' ? 'chat.runningStatus.responsePending'
- : activity === 'tool' ? 'chat.runningStatus.toolRunning'
- : activity === 'paused' ? 'chat.runningStatus.generationPaused'
- : activity === 'quiet' ? 'chat.runningStatus.responsePending' : 'chat.runningStatus.currentRate')}
+ {t('chat.runningStatus.currentRate')}
</div>
<div className="flex items-baseline gap-1.5 tabular-nums">
- <span className="text-28 font-medium leading-none">{approximate(rate, Boolean(responseSpeed) && (responseSpeed?.phase !== 'complete' || responseSpeed.estimated)) ?? '—'}</span>
+ <span className="text-28 font-medium leading-none">{rate ?? '—'}</span>
<span className="text-12 text-[var(--text-secondary)]">
{t('chat.runningStatus.tokenRateUnit')}
</span>
@@ -155,14 +137,14 @@ export function RunningTokenRatePopover({
<div className="flex items-center gap-2">
<dt className="text-[var(--text-secondary)]">{t('chat.runningStatus.averageRate')}</dt>
<dd className="font-medium">
- {averageRate ? t('chat.runningStatus.tokenRate', { rate: approximate(averageRate, responseSpeed?.estimated ?? false) }) : '—'}
+ {averageRate ? t('chat.runningStatus.tokenRate', { rate: averageRate }) : '—'}
</dd>
</div>
<div className="flex items-center gap-2">
<dt className="text-[var(--text-secondary)]">{t('chat.runningStatus.outputTotal')}</dt>
<dd className="font-medium">
{t('chat.runningStatus.tokenCount', {
- tokens: approximate(formatRunningTokenCount(outputTokens), responseSpeed?.estimated ?? false),
+ tokens: formatRunningTokenCount(outputTokens),
})}
</dd>
</div>
@@ -171,18 +153,12 @@ export function RunningTokenRatePopover({
<dd className="font-medium">
{samples.length > 0
? t('chat.runningStatus.tokenRate', {
- rate: approximate(formatRecentOutputTokenRate(history.peak), Boolean(responseSpeed)),
+ rate: formatRecentOutputTokenRate(history.peak),
})
: '—'}
</dd>
</div>
</dl>
- {responseSpeed && (
- <dl className="col-span-2 flex items-center justify-between gap-3 text-12 tabular-nums" title={t(`chat.runningStatus.${responseSpeed.waitOrigin === 'stream' ? 'streamWaitDescription' : 'turnWaitDescription'}`)}>
- <dt className="text-[var(--text-secondary)]">{t(`chat.runningStatus.${responseSpeed.waitOrigin === 'stream' ? 'streamWait' : 'firstResponse'}`)}</dt>
- <dd>{responseSpeed.firstResponseMs === null ? '—' : t('chat.runningStatus.waitSeconds', { seconds: (responseSpeed.firstResponseMs / 1000).toFixed(1) })}</dd>
- </dl>
- )}
</div>
</div>
);
diff --git a/apps/desktop/src/renderer/i18n/locales/en/common.json b/apps/desktop/src/renderer/i18n/locales/en/common.json
index 7138415b47..1eaffb5c96 100644
--- a/apps/desktop/src/renderer/i18n/locales/en/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/en/common.json
@@ -10146,25 +10146,8 @@
"observedPeak": "Peak",
"tokenRate": "{{rate}} tok/s",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "Streaming rates and peaks are estimates. Response totals are calibrated with reported usage. Tool execution and request waits are excluded.",
- "tokenCount": "{{tokens}} tok",
- "estimatedValue": "≈{{value}}",
- "estimatedTokenRate": "≈{{rate}} tok/s",
- "responseWaiting": "Waiting {{seconds}}s",
- "waitSeconds": "{{seconds}}s",
- "firstResponse": "First response",
- "streamWait": "Stream wait",
- "turnWaitDescription": "From the local provider turn start to the first observed text, thinking or tool content. This is not network latency.",
- "streamWaitDescription": "From SDK stream start to the first content block. The earlier request wait is not observable.",
- "lastGeneration": "Last generation",
- "finalAverage": "Final average",
- "generationPaused": "Generation paused",
- "responsePending": "Waiting for response",
- "responseFailed": "Request failed",
- "responseCancelled": "Stopped",
- "responseRetrying": "Retrying",
- "responseGenerating": "Generating",
- "toolRunning": "Running tool"
+ "tokenRateDescription": "Latest generation interval. Excludes tools and waits.",
+ "tokenCount": "{{tokens}} tok"
},
"shareImage": {
"entry": "Share as Image",
diff --git a/apps/desktop/src/renderer/i18n/locales/ja/common.json b/apps/desktop/src/renderer/i18n/locales/ja/common.json
index 7dc8ba3ed9..b6979335eb 100644
--- a/apps/desktop/src/renderer/i18n/locales/ja/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/ja/common.json
@@ -10123,25 +10123,8 @@
"observedPeak": "ピーク",
"tokenRate": "{{rate}} tok/s",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "リアルタイム速度とピークは推定値です。応答後に実使用量で補正し、ツール実行とリクエスト待機は除外します。",
- "tokenCount": "{{tokens}} tok",
- "estimatedValue": "約{{value}}",
- "estimatedTokenRate": "約{{rate}} tok/s",
- "responseWaiting": "待機 {{seconds}}秒",
- "waitSeconds": "{{seconds}}秒",
- "firstResponse": "初回応答",
- "streamWait": "ストリーム内待機",
- "turnWaitDescription": "ローカルでの実行開始から最初のテキスト、思考、ツール内容まで。ネットワーク遅延とは異なります。",
- "streamWaitDescription": "SDK ストリーム開始から最初のコンテンツまで。それ以前のリクエスト待機は測定できません。",
- "lastGeneration": "前回の生成",
- "finalAverage": "最終平均",
- "generationPaused": "生成を一時停止",
- "responsePending": "応答待ち",
- "responseFailed": "リクエスト失敗",
- "responseCancelled": "停止済み",
- "responseRetrying": "再試行中",
- "responseGenerating": "生成中",
- "toolRunning": "ツール実行中"
+ "tokenRateDescription": "直近の生成区間の速度。ツール実行と待機は除きます。",
+ "tokenCount": "{{tokens}} tok"
},
"shareImage": {
"entry": "画像として共有",
diff --git a/apps/desktop/src/renderer/i18n/locales/ko/common.json b/apps/desktop/src/renderer/i18n/locales/ko/common.json
index 241029a59e..71ee7e90f9 100644
--- a/apps/desktop/src/renderer/i18n/locales/ko/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/ko/common.json
@@ -10123,25 +10123,8 @@
"observedPeak": "최고",
"tokenRate": "{{rate}} tok/s",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "실시간 속도와 최고 속도는 추정값입니다. 응답 후 실제 사용량으로 보정하며 도구 실행과 요청 대기는 제외합니다.",
- "tokenCount": "{{tokens}} tok",
- "estimatedValue": "약 {{value}}",
- "estimatedTokenRate": "약 {{rate}} tok/s",
- "responseWaiting": "대기 {{seconds}}초",
- "waitSeconds": "{{seconds}}초",
- "firstResponse": "첫 응답",
- "streamWait": "스트림 내 대기",
- "turnWaitDescription": "로컬 제공자 턴 시작부터 첫 텍스트, 사고 또는 도구 내용까지의 시간입니다. 네트워크 지연과는 다릅니다.",
- "streamWaitDescription": "SDK 스트림 시작부터 첫 콘텐츠까지의 시간입니다. 그 이전의 요청 대기는 측정할 수 없습니다.",
- "lastGeneration": "지난 생성",
- "finalAverage": "최종 평균",
- "generationPaused": "생성 일시 중지",
- "responsePending": "응답 대기 중",
- "responseFailed": "요청 실패",
- "responseCancelled": "중지됨",
- "responseRetrying": "다시 시도 중",
- "responseGenerating": "생성 중",
- "toolRunning": "도구 실행 중"
+ "tokenRateDescription": "최근 생성 구간의 속도. 도구 실행과 대기는 제외합니다.",
+ "tokenCount": "{{tokens}} tok"
},
"shareImage": {
"entry": "이미지로 공유",
diff --git a/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json b/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
index 34a8fd0348..fc37d0ddfd 100644
--- a/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
@@ -10123,25 +10123,8 @@
"observedPeak": "峰值",
"tokenRate": "{{rate}} tok/s",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "实时速度和峰值为估算。响应结束后用真实用量校准;不计工具执行和请求等待。",
- "tokenCount": "{{tokens}} tok",
- "estimatedValue": "约{{value}}",
- "estimatedTokenRate": "约{{rate}} tok/s",
- "responseWaiting": "等待 {{seconds}}秒",
- "waitSeconds": "{{seconds}}秒",
- "firstResponse": "首响应",
- "streamWait": "流内等待",
- "turnWaitDescription": "从本地收到引擎开始到首个文字、思考或工具内容,不等同于网络延迟。",
- "streamWaitDescription": "从 SDK 流开始到首个内容块,之前的请求等待不可测。",
- "lastGeneration": "上次生成",
- "finalAverage": "最终平均",
- "generationPaused": "生成暂停",
- "responsePending": "等待响应",
- "responseFailed": "请求失败",
- "responseCancelled": "已停止",
- "responseRetrying": "正在重试",
- "responseGenerating": "正在生成",
- "toolRunning": "执行工具中"
+ "tokenRateDescription": "最近一段生成速度,不含工具执行与等待。",
+ "tokenCount": "{{tokens}} tok"
},
"shareImage": {
"entry": "分享为图片",
diff --git a/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json b/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
index c8b6c70ad0..24cfcb6100 100644
--- a/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
@@ -10122,25 +10122,8 @@
"observedPeak": "峰值",
"tokenRate": "{{rate}} tok/s",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "即時速度和峰值為估算。回應結束後以實際用量校準;不計工具執行和請求等待。",
- "tokenCount": "{{tokens}} tok",
- "estimatedValue": "約{{value}}",
- "estimatedTokenRate": "約{{rate}} tok/s",
- "responseWaiting": "等待 {{seconds}}秒",
- "waitSeconds": "{{seconds}}秒",
- "firstResponse": "首回應",
- "streamWait": "流內等待",
- "turnWaitDescription": "從本機收到引擎開始到首個文字、思考或工具內容,不等同於網路延遲。",
- "streamWaitDescription": "從 SDK 串流開始到首個內容區塊,之前的請求等待無法測量。",
- "lastGeneration": "上次生成",
- "finalAverage": "最終平均",
- "generationPaused": "生成暫停",
- "responsePending": "等待回應",
- "responseFailed": "請求失敗",
- "responseCancelled": "已停止",
- "responseRetrying": "正在重試",
- "responseGenerating": "正在生成",
- "toolRunning": "執行工具中"
+ "tokenRateDescription": "最近一段生成速度,不含工具執行與等待。",
+ "tokenCount": "{{tokens}} tok"
},
"shareImage": {
"entry": "分享為圖片",
diff --git a/apps/desktop/src/renderer/lib/makerChatStore.ts b/apps/desktop/src/renderer/lib/makerChatStore.ts
index 42a97e3887..8967850772 100644
--- a/apps/desktop/src/renderer/lib/makerChatStore.ts
+++ b/apps/desktop/src/renderer/lib/makerChatStore.ts
@@ -1,4 +1,3 @@
-import { readResponseSpeedSnapshot, stopResponseSpeed, retryResponseSpeed, resumeResponseSpeed, mergeResponseSpeedStatus, type ResponseSpeedSnapshot } from '@cindy/maker-shared/usage-format';
import { readBotTaskResults } from '@cindy/maker-shared/botCollaboration';
import { remotePluginSetupErrorCode, type PluginSetupCommandError } from './pluginSetupCommandError';
export type { PluginSetupCommandError } from './pluginSetupCommandError';
@@ -731,7 +730,6 @@ export interface AgentStatus {
/** Turn-cumulative output tokens for live TPS. */
outputTokens?: number;
/** Generation-only milliseconds including any open interval at emit time. */
- responseSpeed?: ResponseSpeedSnapshot;
generationDurationMs?: number;
/** True while the model currently owns the turn. */
generationActive?: boolean;
@@ -5576,18 +5574,10 @@ export function handleStreamEvent(
const hasCodexReconnectRecoveryOutput = isCodexReconnectRecoveryOutput(event);
const shouldClearCodexReconnectPendingCard =
event.type === 'error' ? !isCodexReconnectProgress : hasCodexReconnectRecoveryOutput;
- const resumedSpeed = ((event.type === 'text' || event.type === 'thinking') &&
- typeof (event.data as { text?: unknown })?.text === 'string' &&
- Boolean((event.data as { text: string }).text)) ||
- (event.type === 'tool_use' && hasCodexReconnectRecoveryOutput)
- ? resumeResponseSpeed(inputState.agentStatus.responseSpeed) : inputState.agentStatus.responseSpeed;
- const stateBeforeReconnectCleanup = resumedSpeed === inputState.agentStatus.responseSpeed &&
- (!hasCodexReconnectRecoveryOutput || inputState.recoverableError == null) ? inputState : {
- ...inputState,
- recoverableError: hasCodexReconnectRecoveryOutput ? null : inputState.recoverableError,
- agentStatus: resumedSpeed === inputState.agentStatus.responseSpeed ? inputState.agentStatus
- : { ...inputState.agentStatus, responseSpeed: resumedSpeed },
- };
+ const stateBeforeReconnectCleanup =
+ !hasCodexReconnectRecoveryOutput || inputState.recoverableError == null
+ ? inputState
+ : { ...inputState, recoverableError: null };
const messagesAfterReconnectCleanup = shouldClearCodexReconnectPendingCard
? removeCodexReconnectPendingCard(stateBeforeReconnectCleanup.messages)
: stateBeforeReconnectCleanup.messages;
@@ -6240,15 +6230,14 @@ export function handleStreamEvent(
});
const terminalData = event.data as
- | { cancelled?: unknown; status?: unknown; reason?: unknown; plan?: unknown; raw?: { id?: unknown; status?: unknown } }
+ | { cancelled?: unknown; reason?: unknown; plan?: unknown; raw?: { id?: unknown; status?: unknown } }
| null
| undefined;
const terminalTurnId = typeof terminalData?.raw?.id === 'string' ? terminalData.raw.id : null;
const terminalTurnStatus =
typeof terminalData?.raw?.status === 'string' ? terminalData.raw.status : null;
const terminalCancelled =
- terminalData?.cancelled === true || terminalData?.status === 'cancelled' ||
- terminalData?.reason === 'send_cancelled_before_acceptance' ||
+ terminalData?.cancelled === true ||
terminalData?.reason === 'turn_continuation_cancelled' ||
terminalData?.reason === 'user_stop_unconfirmed_wake_tasks';
const doneMessages =
@@ -6305,10 +6294,6 @@ export function handleStreamEvent(
turnStoppedByUser: state.turnStoppedByUser || terminalCancelled,
agentStatus: {
...state.agentStatus,
- responseSpeed: stopResponseSpeed(state.agentStatus.responseSpeed,
- state.turnStoppedByUser || terminalCancelled ? 'cancelled'
- : terminalTurnStatus === 'failed' || terminalData?.status === 'failed' || finalized.error ? 'failed' : undefined),
- generationActive: false,
isRunning: false,
startedAt: null,
},
@@ -6425,8 +6410,6 @@ export function handleStreamEvent(
isRunning: true,
startedAt: state.agentStatus.startedAt ?? Date.now(),
}),
- responseSpeed: (event.data as { willRetry?: boolean })?.willRetry === true
- ? retryResponseSpeed(state.agentStatus.responseSpeed) : state.agentStatus.responseSpeed,
},
};
}
@@ -6446,8 +6429,6 @@ export function handleStreamEvent(
isStreaming: true,
agentStatus: {
...state.agentStatus,
- responseSpeed: (event.data as { willRetry?: boolean })?.willRetry === true
- ? retryResponseSpeed(state.agentStatus.responseSpeed) : state.agentStatus.responseSpeed,
isRunning: true,
startedAt: state.agentStatus.startedAt ?? Date.now(),
},
@@ -6540,10 +6521,6 @@ export function handleStreamEvent(
// 初始 "Let's go" 文案上 shimmer 闪个不停(done 路径有同样的复位)。
agentStatus: {
...state.agentStatus,
- responseSpeed: suppressAutoResumeBroadcastError || isPlannedUpgradeClose
- ? stopResponseSpeed(state.agentStatus.responseSpeed)
- : stopResponseSpeed(state.agentStatus.responseSpeed, state.turnStoppedByUser ? 'cancelled' : 'failed'),
- generationActive: false,
isRunning: false,
startedAt: null,
},
@@ -7056,8 +7033,6 @@ function forceFinalizeOnSessionClosed(state: SessionChatState): SessionChatState
turnStoppedByUser: false,
agentStatus: {
...finalized.agentStatus,
- responseSpeed: stopResponseSpeed(finalized.agentStatus.responseSpeed),
- generationActive: false,
isRunning: false,
startedAt: null,
},
@@ -7070,7 +7045,7 @@ function mergeLiveGenerationStatus(
previous: AgentStatus,
): Pick<
AgentStatus,
- 'outputTokens' | 'generationDurationMs' | 'generationActive' | 'generationReliable' | 'responseSpeed'
+ 'outputTokens' | 'generationDurationMs' | 'generationActive' | 'generationReliable'
> {
// Turn start drops leftover metrics from the previous turn, then keeps any
// live fields carried by this same status. A reconnect-shaped first event
@@ -7078,7 +7053,6 @@ function mergeLiveGenerationStatus(
// zero the values that just arrived.
const baseline = isTurnStart
? {
- responseSpeed: undefined,
outputTokens: 0,
generationDurationMs: 0,
generationActive: false,
@@ -7092,8 +7066,6 @@ function mergeLiveGenerationStatus(
typeof update.generationReliable === 'boolean';
if (!hasLiveFields) {
return {
- responseSpeed: mergeResponseSpeedStatus(previous.responseSpeed,
- readResponseSpeedSnapshot(update.responseSpeed, Date.now()), update.isRunning, isTurnStart),
outputTokens: baseline.outputTokens,
generationDurationMs: baseline.generationDurationMs,
generationActive: update.isRunning ? baseline.generationActive : false,
@@ -7101,8 +7073,6 @@ function mergeLiveGenerationStatus(
};
}
const merged = {
- responseSpeed: mergeResponseSpeedStatus(previous.responseSpeed,
- readResponseSpeedSnapshot(update.responseSpeed, Date.now()), update.isRunning, isTurnStart),
outputTokens:
typeof update.outputTokens === 'number' ? update.outputTokens : baseline.outputTokens,
generationDurationMs:
@@ -15639,8 +15609,6 @@ function stopSession(
costUsd: s.agentStatus.costUsd,
contextTokens: s.agentStatus.contextTokens,
contextWindow: s.agentStatus.contextWindow,
- responseSpeed: stopResponseSpeed(s.agentStatus.responseSpeed, 'cancelled'),
- generationActive: false,
isRunning: false,
startedAt: null,
},
diff --git a/apps/desktop/src/renderer/vite-env.d.ts b/apps/desktop/src/renderer/vite-env.d.ts
index 115c0a8370..b66ee6b98e 100644
--- a/apps/desktop/src/renderer/vite-env.d.ts
+++ b/apps/desktop/src/renderer/vite-env.d.ts
@@ -806,7 +806,6 @@ interface CCAgentPermissionDismissedPayload {
}
interface CCAgentStatusUpdate {
- responseSpeed?: import("@cindy/maker-shared/usage-format").ResponseSpeedSnapshot;
sessionId: string;
status: string;
tokenUsage: number;
diff --git a/apps/mobile/app/sessions/[sessionId].tsx b/apps/mobile/app/sessions/[sessionId].tsx
index 8e7534844e..b6e2080701 100644
--- a/apps/mobile/app/sessions/[sessionId].tsx
+++ b/apps/mobile/app/sessions/[sessionId].tsx
@@ -1,5 +1,3 @@
-import { responseSpeedActivity, responseSpeedHistory } from "@cindy/maker-shared/usage-format";
-import type { ResponseSpeedSnapshot } from "@cindy/maker-shared/usage-format";
import { modelNeedsReselection } from '@/session/modelReselection';
import { getCachedDeviceProviders } from '@/device-link/deviceProvidersCache';
import { MountOnFirstOpen } from '@/session/MountOnFirstOpen';
@@ -10280,7 +10278,7 @@ export default function SessionScreen() {
例外:新建任务慢发送时这一行要说明卡在哪一步,不能被同步占位藏掉。
判据必须带 messageCount:syncingWhileEmpty 只要 loading 就为真,而收口后
还会再来几轮 load(实测日志),只看它会让已有消息的会话反复熄灭活动条。 */}
- {!companionChat && (showComposerActivity || pendingRecoveryStatus || remoteSessionRunStatus.responseSpeed?.phase === 'complete') && (creationSlowPhase !== null || !(syncingWhileEmpty && !hasRenderedMessages)) ? (
+ {!companionChat && (showComposerActivity || pendingRecoveryStatus) && (creationSlowPhase !== null || !(syncingWhileEmpty && !hasRenderedMessages)) ? (
<View
style={[
styles.composerActivityFrame,
@@ -10299,7 +10297,6 @@ export default function SessionScreen() {
streaming={isSessionStreaming}
tokenUsage={composerActivityTokenUsage}
outputTokens={remoteSessionRunStatus.outputTokens}
- responseSpeed={remoteSessionRunStatus.responseSpeed}
generationDurationMs={remoteSessionRunStatus.generationDurationMs}
generationReliable={remoteSessionRunStatus.generationReliable}
generationActive={remoteSessionRunStatus.generationActive}
@@ -12192,7 +12189,6 @@ function ComposerActivityStatus({
rateStartedAt = startedAt,
streaming = false,
tokenUsage,
- responseSpeed,
outputTokens,
generationDurationMs,
generationReliable,
@@ -12210,7 +12206,6 @@ function ComposerActivityStatus({
rateStartedAt?: number | null;
streaming?: boolean;
tokenUsage: number;
- responseSpeed?: ResponseSpeedSnapshot;
outputTokens: number;
generationDurationMs: number;
generationReliable: boolean;
@@ -12241,9 +12236,8 @@ function ComposerActivityStatus({
// Elapsed uses the local fallback. Sampling keeps the raw remote start so a
// terminal null is not a new turn, and a local send is not the previous rate.
const samplerStartedAt = rateStartedAt === undefined ? startedAt : rateStartedAt;
- const legacyRateHistory = useRunningTokenRateHistory({
+ const rateHistory = useRunningTokenRateHistory({
sessionKey,
- generationActive,
startedAt: samplerStartedAt,
outputTokens,
generationDurationMs,
@@ -12251,10 +12245,6 @@ function ComposerActivityStatus({
streaming,
});
- const rateHistory = responseSpeed ? responseSpeedHistory(responseSpeed)
- : generationActive === false ? { ...legacyRateHistory, latestRate: null } : legacyRateHistory;
- const completedSpeed = !visible && responseSpeed?.phase === 'complete';
- const speedActivity = responseSpeed ? responseSpeedActivity(responseSpeed) : null;
if (pendingRecovery && !reconnectAttempt) {
// 自愈 / 换电脑进行中:只显示这一行状态,不带耗时与速率(这一轮还没有结局)。
return (
@@ -12273,29 +12263,22 @@ function ComposerActivityStatus({
</View>
);
}
- if (!visible && !completedSpeed) return null;
-
- const elapsedText = completedSpeed
- ? t(responseSpeed?.outcome === 'failed' ? 'session.screen.responseFailed'
- : responseSpeed?.outcome === 'cancelled' ? 'session.screen.responseCancelled' : 'session.screen.lastGeneration')
- : speedActivity === 'retrying' ? t('session.screen.responseRetrying')
- : responseSpeed?.phase === 'waiting'
- ? t('session.screen.responseWaiting', { seconds: ((responseSpeed.waitingMs + Math.max(0, Date.now() - responseSpeed.sampledAt)) / 1000).toFixed(1) })
- : speedActivity === 'quiet' ? t('session.screen.responsePending') : formatComposerActivityElapsed(elapsed);
+ if (!visible) return null;
+
+ const elapsedText = formatComposerActivityElapsed(elapsed);
const tokenCount = formatComposerActivityTokenCount(tokenUsage);
const tokenText = t('session.screen.tokenCount', { tokens: tokenCount });
const tokenA11yText = t('session.screen.tokenCountFull', { tokens: tokenCount });
const showElapsedOnly = sideTaskRunning || Boolean(reconnectAttempt);
const rateValue = formatComposerActivityRateValue(showElapsedOnly ? null : rateHistory.latestRate);
- const canShowRateDetails = !showElapsedOnly && (Boolean(responseSpeed) || (generationReliable
+ const canShowRateDetails = !showElapsedOnly
+ && generationReliable
&& outputTokens > 0
&& Number.isFinite(outputTokens)
&& Number.isFinite(generationDurationMs)
- && generationDurationMs > 0));
+ && generationDurationMs > 0;
const rateText = rateValue
- ? responseSpeed && (responseSpeed.phase !== 'complete' || responseSpeed.estimated)
- ? t('session.screen.estimatedTokenRate', { rate: rateValue })
- : t('session.screen.tokenRate', { rate: rateValue })
+ ? t('session.screen.tokenRate', { rate: rateValue })
: null;
const showUsageMeta = !showElapsedOnly && (Boolean(rateText) || tokenUsage > 0);
const usageMeta = (
@@ -12337,15 +12320,9 @@ function ComposerActivityStatus({
? 'session.screen.rateLimitRetrying'
: 'session.screen.networkReconnecting',
)
- // 发送等待超过阈值时,同一行先说明卡在哪一步;创建完成后回到下面的「思考中」等状态。
+ // 发送等待超过阈值时,同一行先说明卡在哪一步;创建完成后回到「思考中」。
: slowSendPhase ? t(`session.new.slowSend.${slowSendPhase}`)
- : t(speedActivity === 'failed' ? 'session.screen.responseFailed'
- : speedActivity === 'cancelled' ? 'session.screen.responseCancelled'
- : speedActivity === 'retrying' ? 'session.screen.responseRetrying'
- : speedActivity === 'waiting' || speedActivity === 'quiet' ? 'session.screen.responsePending'
- : speedActivity === 'tool' ? 'session.screen.toolRunning'
- : speedActivity === 'paused' ? 'session.screen.generationPaused'
- : speedActivity === 'generating' ? 'session.screen.responseGenerating' : 'session.screen.thinking');
+ : t('session.screen.thinking');
return (
<View
@@ -12353,15 +12330,14 @@ function ComposerActivityStatus({
style={pillStyles.composerActivityStatus}
testID="session.composerActivityStatus"
>
- {visible && <ComposerActivityPill label={activityText} live={Boolean(slowSendPhase) && !reconnectAttempt}>
+ <ComposerActivityPill label={activityText} live={Boolean(slowSendPhase) && !reconnectAttempt}>
{reconnectAttempt ? (
<Text style={[pillStyles.composerActivityStatusText, pillStyles.composerActivityProgressText]}>
{reconnectAttempt.attempt}/{reconnectAttempt.maxAttempts}
</Text>
) : null}
- </ComposerActivityPill>}
+ </ComposerActivityPill>
<RunningTokenRatePopover
- responseSpeed={responseSpeed}
key={sessionKey}
enabled={canShowRateDetails}
availableRegion={availableRegion}
diff --git a/apps/mobile/src/__tests__/remoteSessionStore.test.ts b/apps/mobile/src/__tests__/remoteSessionStore.test.ts
index fb9a0360da..07103c6503 100644
--- a/apps/mobile/src/__tests__/remoteSessionStore.test.ts
+++ b/apps/mobile/src/__tests__/remoteSessionStore.test.ts
@@ -6785,60 +6785,3 @@ describe('list message reuse', () => {
expect(remoteSessionStore.getMessages('s1')).toEqual([]);
});
});
-
-it('retains response speed per session at completion, freezes aborts and clears the next turn', () => {
- remoteSessionStore.setDeviceSessions('dev-1', 'Mac', [session('tps-one'), session('tps-two')]);
- const speed = { phase: 'generating', waitOrigin: 'turn', firstResponseMs: 2000, waitingMs: 0,
- durationMs: 3000, outputTokens: 50, estimated: true, averageRate: 16.7, recentRate: 20,
- samples: [{ durationMs: 1000, outputTokens: 20, rate: 20 }], sampledAt: 1 };
- pushMakerStatus('tps-one', { isRunning: true, responseSpeed: speed });
- remoteSessionStore.applyRemotePush('dev-1', 'maker:event', {
- sessionId: 'tps-one', event: { type: 'error', data: { message: 'cancelled', isTerminal: true } },
- });
- expect(remoteSessionStore.getSessionRunStatus('tps-one').responseSpeed)
- .toMatchObject({ phase: 'complete', estimated: true, outputTokens: 50 });
- expect(remoteSessionStore.getSessionRunStatus('tps-two').responseSpeed).toBeUndefined();
- pushMakerStatus('tps-two', { isRunning: true });
- expect(remoteSessionStore.getSessionRunStatus('tps-one').responseSpeed?.outputTokens).toBe(50);
- pushMakerStatus('tps-one', { isRunning: true, responseSpeed: { ...speed, phase: 'waiting',
- firstResponseMs: null, outputTokens: 0, durationMs: 0, samples: [] } });
- expect(remoteSessionStore.getSessionRunStatus('tps-one').responseSpeed)
- .toMatchObject({ phase: 'waiting', firstResponseMs: null, outputTokens: 0, samples: [] });
- pushMakerStatus('tps-one', { isRunning: false, responseSpeed: { ...speed, phase: 'complete',
- outputTokens: 300, estimated: false, averageRate: 100, recentRate: 100 } });
- expect(remoteSessionStore.getSessionRunStatus('tps-one').responseSpeed)
- .toMatchObject({ phase: 'complete', estimated: false, outputTokens: 300 });
-});
-
-it('keeps response speed failures through terminal tails, distinguishes cancellation and actual retries', () => {
- remoteSessionStore.clear();
- remoteSessionStore.setDeviceSessions('dev-1', 'Host', [session('speed-errors')]);
- remoteSessionStore.enterSessionMessageDetail('speed-errors');
- const speed = { phase: 'generating', waitOrigin: 'turn', firstResponseMs: 1000, waitingMs: 0,
- durationMs: 2000, outputTokens: 50, estimated: true, averageRate: 25, recentRate: 30,
- samples: [{ durationMs: 2000, outputTokens: 50, rate: 30 }], sampledAt: Date.now() };
- const event = (type: string, data: Record<string, unknown>) => remoteSessionStore.applyRemotePush('dev-1', 'maker:event', {
- sessionId: 'speed-errors', event: { type, data, source: 'pi' },
- });
- pushMakerStatus('speed-errors', { isRunning: true, responseSpeed: speed });
- event('error', { message: 'temporary warning', isTerminal: false });
- expect(remoteSessionStore.getSessionRunStatus('speed-errors').responseSpeed?.retrying).not.toBe(true);
- event('error', { message: 'API unavailable', isTerminal: false, willRetry: true });
- expect(remoteSessionStore.getSessionRunStatus('speed-errors').responseSpeed).toMatchObject({ retrying: true, recentRate: null });
- pushMakerStatus('speed-errors', { isRunning: true, responseSpeed: speed });
- expect(remoteSessionStore.getSessionRunStatus('speed-errors').responseSpeed?.retrying).toBe(true);
- event('thinking', { stage: 'delta', blockId: 'recovered-thinking', text: 'reasoning resumed' });
- expect(remoteSessionStore.getSessionRunStatus('speed-errors').responseSpeed?.retrying).toBe(false);
- event('error', { message: 'API unavailable', isTerminal: true });
- event('done', {});
- pushMakerStatus('speed-errors', { isRunning: false, responseSpeed: { ...speed, phase: 'complete', estimated: false } });
- expect(remoteSessionStore.getSessionRunStatus('speed-errors')).toMatchObject({ isRunning: false, hasTerminalError: true,
- responseSpeed: { outcome: 'failed', recentRate: null, samples: speed.samples } });
- pushMakerStatus('speed-errors', { isRunning: true, responseSpeed: speed });
- expect(remoteSessionStore.getSessionRunStatus('speed-errors').responseSpeed?.outcome).toBeUndefined();
- event('done', { type: 'pi/agent_settled', status: 'cancelled' });
- event('error', { message: 'aborted transport', isTerminal: true });
- expect(remoteSessionStore.getSessionRunStatus('speed-errors')).toMatchObject({ hasTerminalError: false,
- responseSpeed: { outcome: 'cancelled', recentRate: null } });
- remoteSessionStore.clear();
-});
diff --git a/apps/mobile/src/__tests__/runningTokenRateHistory.test.tsx b/apps/mobile/src/__tests__/runningTokenRateHistory.test.tsx
index b1d0aca26d..2f90f5e316 100644
--- a/apps/mobile/src/__tests__/runningTokenRateHistory.test.tsx
+++ b/apps/mobile/src/__tests__/runningTokenRateHistory.test.tsx
@@ -47,15 +47,15 @@ it('waits for a paired interval and shows its rate instead of the turn average',
expect(report({ outputTokens: 1100, generationDurationMs: 12_000 })).toBe('50');
// Clock ticks and time-only reports do not manufacture slower throughput.
act(() => vi.advanceTimersByTime(5000));
- expect(host.textContent).toBe('waiting');
- expect(report({ generationDurationMs: 14_000 })).toBe('waiting');
+ expect(host.textContent).toBe('50');
+ expect(report({ generationDurationMs: 14_000 })).toBe('50');
expect(report({ outputTokens: 1140 })).toBe('20');
});
-it('expires stale samples after the one-second observation window and resumes on a new paired report', () => {
+it('expires stale samples after 60 seconds and resumes on a new paired report', () => {
report();
report({ outputTokens: 1100, generationDurationMs: 12_000 });
- act(() => vi.advanceTimersByTime(999));
+ act(() => vi.advanceTimersByTime(59_999));
expect(host.textContent).toBe('50');
act(() => vi.advanceTimersByTime(1));
expect(host.textContent).toBe('waiting');
diff --git a/apps/mobile/src/__tests__/runningTokenRatePopover.test.tsx b/apps/mobile/src/__tests__/runningTokenRatePopover.test.tsx
index 4acecd2a05..23841f4542 100644
--- a/apps/mobile/src/__tests__/runningTokenRatePopover.test.tsx
+++ b/apps/mobile/src/__tests__/runningTokenRatePopover.test.tsx
@@ -15,9 +15,6 @@ import { afterEach, beforeEach, expect, it, vi } from "vitest";
import {
clearRateHistoryCache,
loadCachedRateHistory,
- responseSpeedActivity,
- responseSpeedHistory,
- type ResponseSpeedSnapshot,
} from "@cindy/maker-shared/usage-format";
import { useRunningTokenRateHistory } from "../session/useRunningTokenRateHistory";
import { RunningTokenRatePopover } from "@/session/RunningTokenRatePopover";
@@ -128,8 +125,6 @@ vi.mock("react-i18next", () => ({
t: (key: string, args?: any) =>
key.endsWith("tokenRate")
? `${args.rate} tok/s`
- : key.endsWith("estimatedValue")
- ? `≈${args.value}`
: key.endsWith("tokenCount")
? `${args.tokens} tok`
: key,
@@ -200,8 +195,6 @@ const bindings = {
iconStroke: {},
RunningTokenRatePopover,
useRunningTokenRateHistory,
- responseSpeedActivity,
- responseSpeedHistory,
formatComposerActivityElapsed: () => "1s",
formatComposerActivityTokenCount: () => "100",
};
@@ -210,76 +203,6 @@ const ActivityStatus = new Function(
`${compiledStatus}; return ComposerActivityStatus;`,
)(...Object.values(bindings));
-it.each([
- { outputTokens: 80, estimated: false, count: '80 tok' },
- { outputTokens: 80, estimated: true, count: '≈80 tok' },
- { outputTokens: 0, estimated: false, count: '0 tok' },
-])('keeps the speed card on measured output instead of whole-turn usage ($count)', async ({ outputTokens, estimated, count }) => {
- const responseSpeed: ResponseSpeedSnapshot = {
- phase: 'complete', waitOrigin: 'turn', firstResponseMs: 100,
- waitingMs: 0, durationMs: 2000, outputTokens, estimated,
- recentRate: null, averageRate: outputTokens / 2,
- samples: [{ durationMs: 2000, outputTokens, rate: outputTokens / 2 }],
- sampledAt: Date.now(),
- };
- await act(async () => root.render(createElement(ActivityStatus, {
- ...base, visible: false, tokenUsage: 800, outputTokens: 800,
- generationDurationMs: 2000, sideTaskRunning: false,
- reconnectAttempt: null, responseSpeed,
- })));
- await gesture('onPressIn');
- await gesture('onPress');
- expect(card()!.textContent).toContain(`session.screen.outputTotal${count}`);
- expect(card()!.textContent).toContain(`session.screen.averageRate${estimated ? '≈' : ''}${outputTokens / 2} tok/s`);
- expect(card()!.textContent).not.toContain('800 tok');
- expect(card()!.textContent).not.toContain('400 tok/s');
-});
-
-it('uses whole-turn output in the legacy card when the host omits the speed snapshot', async () => {
- await act(async () => root.render(createElement(ActivityStatus, {
- ...base, visible: true, tokenUsage: 800, outputTokens: 800,
- generationDurationMs: 2000, sideTaskRunning: false, reconnectAttempt: null,
- })));
- await gesture('onPressIn');
- await gesture('onPress');
- expect(card()!.textContent).toContain('session.screen.outputTotal800 tok');
- expect(card()!.textContent).toContain('session.screen.averageRate400 tok/s');
-});
-
-it('shows waiting, execution, silent output and retained completion consistently in the status and card', async () => {
- vi.useFakeTimers();
- try {
- const speed: ResponseSpeedSnapshot = { phase: 'waiting', waitOrigin: 'turn', firstResponseMs: null,
- waitingMs: 1000, durationMs: 0, outputTokens: 0, estimated: false, recentRate: null,
- averageRate: null, samples: [], sampledAt: Date.now() };
- const show = async (responseSpeed: ResponseSpeedSnapshot, visible = true) => act(async () =>
- root.render(createElement(ActivityStatus, { ...base, visible, tokenUsage: 0,
- sideTaskRunning: false, reconnectAttempt: null, responseSpeed })),
- );
- await show(speed);
- expect(host.textContent).toContain('session.screen.responsePending');
- await gesture('onPressIn'); await gesture('onPress');
- expect(card()!.textContent).toContain('session.screen.responsePending');
- const generating = { ...speed, phase: 'generating' as const, firstResponseMs: 1000,
- hasRecentOutput: true, durationMs: 2000, outputTokens: 80, estimated: true, recentRate: 40,
- averageRate: 40, samples: [{ durationMs: 2000, outputTokens: 80, rate: 40 }] };
- await show(generating);
- expect(host.textContent).toContain('session.screen.responseGenerating');
- await act(async () => vi.advanceTimersByTime(1000));
- expect(host.textContent).toContain('session.screen.responsePending');
- expect(card()!.textContent).not.toContain('40 tok/s—');
- await show({ ...generating, phase: 'paused', toolActive: true, recentRate: null });
- expect(host.textContent).toContain('session.screen.toolRunning');
- expect(card()!.textContent).toContain('session.screen.toolRunning');
- await show({ ...generating, phase: 'complete', estimated: false, recentRate: 40 }, false);
- await act(async () => vi.advanceTimersByTime(60000));
- expect(host.textContent).toContain('session.screen.lastGeneration');
- expect(card()!.textContent).toContain('session.screen.finalAverage');
- await show({ ...speed, sampledAt: Date.now() });
- expect(host.textContent).not.toContain('session.screen.lastGeneration');
- } finally { vi.useRealTimers(); }
-});
-
it.each(["onPress", "onLongPress"])(
"records the first completed interval before enabling %s, without sampling inactive gaps",
async (open) => {
@@ -312,10 +235,6 @@ it.each(["onPress", "onLongPress"])(
expect(
loadCachedRateHistory(base.sessionKey)?.samples.map((s) => s.rate),
).toEqual([100]);
- await renderStatus({ startedAt: null, outputTokens: 100, generationDurationMs: 1000, generationActive: false });
- expect(card()!.textContent).toContain('session.screen.currentRate—');
- await renderStatus({ startedAt: null, outputTokens: 100, generationDurationMs: 1000, generationActive: true });
- expect(card()!.textContent).toContain('session.screen.currentRate—');
for (const inactive of [
{ sideTaskRunning: true },
{ reconnectAttempt: { attempt: 1, maxAttempts: 3 } },
@@ -610,7 +529,7 @@ it("uses paired generation samples, expires recent speed, and isolates a differe
expect(card()!.textContent).toContain("50");
expect(card()!.textContent).toContain("75 tok/s");
expect(card()!.textContent).toContain("100 tok/s");
- await act(async () => vi.advanceTimersByTime(1_000));
+ await act(async () => vi.advanceTimersByTime(60_000));
expect(card()!.textContent).toContain("—");
await render({ key: "other", sessionKey: "account/device/other" });
expect(card()).toBeNull();
@@ -670,36 +589,13 @@ it("moves screen reader focus into the pinned card and back to the trigger, exce
expect(harness.focus).toEqual(["text:session.screen.currentRate"]);
});
-it.each(['failed', 'cancelled', 'retrying'] as const)('shows %s consistently in the mobile retained entry and card', async (activity) => {
- const terminal = activity !== 'retrying';
- const key = activity === 'failed' ? 'responseFailed' : activity === 'cancelled' ? 'responseCancelled' : 'responseRetrying';
- const speed: ResponseSpeedSnapshot = { phase: terminal ? 'complete' : 'paused', waitOrigin: 'turn',
- firstResponseMs: 1000, waitingMs: 0, durationMs: 2000, outputTokens: 80, estimated: true,
- recentRate: null, averageRate: 40, samples: [{ durationMs: 2000, outputTokens: 80, rate: 40 }], sampledAt: Date.now(),
- ...(terminal ? { outcome: activity as 'failed' | 'cancelled' } : { retrying: true }),
- };
- await act(async () => root.render(createElement(ActivityStatus, { ...base, visible: !terminal,
- tokenUsage: 500, sideTaskRunning: false, reconnectAttempt: null, responseSpeed: speed })));
- expect(host.textContent).toContain(`session.screen.${key}`);
- expect(host.textContent).not.toContain('session.screen.lastGeneration');
- await gesture('onPressIn'); await gesture('onPress');
- expect(card()!.textContent).toContain(`session.screen.${key}`);
- expect(card()!.textContent).not.toContain('session.screen.finalAverage');
- expect(card()!.textContent).toContain('40');
-});
-
it.each([
{ label: 'switching computers' },
{ label: 'reconnecting', attempt: 2, maxAttempts: 5 },
-])('keeps $label above a retained failure and slow-send status without token metadata', async (pendingRecovery) => {
- const responseSpeed: ResponseSpeedSnapshot = {
- phase: 'complete', outcome: 'failed', waitOrigin: 'turn', firstResponseMs: 1000,
- waitingMs: 0, durationMs: 2000, outputTokens: 80, estimated: true,
- recentRate: null, averageRate: 40, samples: [], sampledAt: Date.now(),
- };
+])('shows only $label, without timer or token metadata, while recovery is pending', async (pendingRecovery) => {
await act(async () => root.render(createElement(ActivityStatus, {
- ...base, visible: false, tokenUsage: 500, sideTaskRunning: false, reconnectAttempt: null,
- responseSpeed, slowSendPhase: 'sending', pendingRecovery,
+ ...base, visible: true, tokenUsage: 500, sideTaskRunning: false, reconnectAttempt: null,
+ slowSendPhase: 'sending', pendingRecovery,
})));
const progress = pendingRecovery.attempt ? `${pendingRecovery.attempt}/${pendingRecovery.maxAttempts}` : '';
expect(host.textContent).toBe(pendingRecovery.label + progress);
diff --git a/apps/mobile/src/__tests__/sessionComposerDesktopFirst.test.ts b/apps/mobile/src/__tests__/sessionComposerDesktopFirst.test.ts
index 1d8571f929..4ff106fa49 100644
--- a/apps/mobile/src/__tests__/sessionComposerDesktopFirst.test.ts
+++ b/apps/mobile/src/__tests__/sessionComposerDesktopFirst.test.ts
@@ -212,7 +212,7 @@ describe('mobile session composer desktop-first surface', () => {
expect(source).not.toContain('testID="session.attachmentPathPanel"');
expect(source).not.toContain('被控电脑上的文件路径');
expect(source).toContain('testID="session.composerActivityStatus"');
- expect(source).toContain("'session.screen.thinking'");
+ expect(source).toContain("t('session.screen.thinking')");
// 活动状态行的三种说法都要在这里出现。过载退避与传输层重连共用同一个 attempt
// 字段, 但文案分开: 说「正在重新连接」会把用户引向排查自己的网络
// (review #844 codex P1)。字符串按 i18n key 断言, 不断言三元表达式的写法。
diff --git a/apps/mobile/src/i18n/locales/en/session.json b/apps/mobile/src/i18n/locales/en/session.json
index f642add6f8..27b0b51bf7 100644
--- a/apps/mobile/src/i18n/locales/en/session.json
+++ b/apps/mobile/src/i18n/locales/en/session.json
@@ -190,7 +190,7 @@
"outputTotal": "Output",
"observedPeak": "Peak",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "Streaming rates and peaks are estimates. Response totals are calibrated with reported usage. Tool execution and request waits are excluded.",
+ "tokenRateDescription": "Latest generation interval. Excludes tools and waits.",
"networkReconnecting": "Reconnecting…",
"modelBusyRetrying": "Model service busy — retrying…",
"rateLimitRetrying": "Request rate limited — retrying…",
@@ -271,24 +271,7 @@
"forkConfirmTitle": "Start a new session from here?",
"forkConfirmDescription": "A separate session will be created using the context at this point. The original session stays unchanged, and future messages in either session won't affect the other.",
"forkConfirm": "Start new session",
- "forkCancel": "Cancel",
- "estimatedValue": "≈{{value}}",
- "estimatedTokenRate": "≈{{rate}} tok/s",
- "responseWaiting": "Waiting {{seconds}}s",
- "waitSeconds": "{{seconds}}s",
- "firstResponse": "First response",
- "streamWait": "Stream wait",
- "turnWaitDescription": "From the local provider turn start to the first observed text, thinking or tool content. This is not network latency.",
- "streamWaitDescription": "From SDK stream start to the first content block. The earlier request wait is not observable.",
- "lastGeneration": "Last generation",
- "finalAverage": "Final average",
- "generationPaused": "Generation paused",
- "responsePending": "Waiting for response",
- "responseFailed": "Request failed",
- "responseCancelled": "Stopped",
- "responseRetrying": "Retrying",
- "responseGenerating": "Generating",
- "toolRunning": "Running tool"
+ "forkCancel": "Cancel"
},
"shareImage": {
"title": "Share Conversation",
diff --git a/apps/mobile/src/i18n/locales/ja/session.json b/apps/mobile/src/i18n/locales/ja/session.json
index 411c6ff548..db86589e79 100644
--- a/apps/mobile/src/i18n/locales/ja/session.json
+++ b/apps/mobile/src/i18n/locales/ja/session.json
@@ -190,7 +190,7 @@
"outputTotal": "累計出力",
"observedPeak": "ピーク",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "リアルタイム速度とピークは推定値です。応答後に実使用量で補正し、ツール実行とリクエスト待機は除外します。",
+ "tokenRateDescription": "直近の生成区間の速度。ツール実行と待機は除きます。",
"networkReconnecting": "再接続中…",
"modelBusyRetrying": "モデルサービスが混雑、再試行中…",
"rateLimitRetrying": "リクエストがレート制限されました。再試行中…",
@@ -271,24 +271,7 @@
"forkConfirmTitle": "ここから新しいセッションを始めますか?",
"forkConfirmDescription": "この時点の会話コンテキストを使って、独立した新しいセッションを作成します。元のセッションは変更されず、以後のメッセージは互いに影響しません。",
"forkConfirm": "新しいセッションを始める",
- "forkCancel": "キャンセル",
- "estimatedValue": "約{{value}}",
- "estimatedTokenRate": "約{{rate}} tok/s",
- "responseWaiting": "待機 {{seconds}}秒",
- "waitSeconds": "{{seconds}}秒",
- "firstResponse": "初回応答",
- "streamWait": "ストリーム内待機",
- "turnWaitDescription": "ローカルでの実行開始から最初のテキスト、思考、ツール内容まで。ネットワーク遅延とは異なります。",
- "streamWaitDescription": "SDK ストリーム開始から最初のコンテンツまで。それ以前のリクエスト待機は測定できません。",
- "lastGeneration": "前回の生成",
- "finalAverage": "最終平均",
- "generationPaused": "生成を一時停止",
- "responsePending": "応答待ち",
- "responseFailed": "リクエスト失敗",
- "responseCancelled": "停止済み",
- "responseRetrying": "再試行中",
- "responseGenerating": "生成中",
- "toolRunning": "ツール実行中"
+ "forkCancel": "キャンセル"
},
"shareImage": {
"title": "会話を共有",
diff --git a/apps/mobile/src/i18n/locales/ko/session.json b/apps/mobile/src/i18n/locales/ko/session.json
index 7534ca8548..11b65d97d2 100644
--- a/apps/mobile/src/i18n/locales/ko/session.json
+++ b/apps/mobile/src/i18n/locales/ko/session.json
@@ -186,7 +186,7 @@
"outputTotal": "누적 출력",
"observedPeak": "최고",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "실시간 속도와 최고 속도는 추정값입니다. 응답 후 실제 사용량으로 보정하며 도구 실행과 요청 대기는 제외합니다.",
+ "tokenRateDescription": "최근 생성 구간의 속도. 도구 실행과 대기는 제외합니다.",
"networkReconnecting": "다시 연결하는 중…",
"modelBusyRetrying": "모델 서비스 혼잡, 다시 시도 중…",
"rateLimitRetrying": "요청이 속도 제한에 걸렸습니다. 다시 시도 중…",
@@ -267,24 +267,7 @@
"forkConfirmTitle": "여기서 새 세션을 시작할까요?",
"forkConfirmDescription": "이 지점의 대화 맥락을 사용해 독립된 새 세션을 만듭니다. 원래 세션은 변경되지 않으며, 이후 두 세션의 메시지는 서로에게 영향을 주지 않습니다.",
"forkConfirm": "새 세션 시작",
- "forkCancel": "취소",
- "estimatedValue": "약 {{value}}",
- "estimatedTokenRate": "약 {{rate}} tok/s",
- "responseWaiting": "대기 {{seconds}}초",
- "waitSeconds": "{{seconds}}초",
- "firstResponse": "첫 응답",
- "streamWait": "스트림 내 대기",
- "turnWaitDescription": "로컬 제공자 턴 시작부터 첫 텍스트, 사고 또는 도구 내용까지의 시간입니다. 네트워크 지연과는 다릅니다.",
- "streamWaitDescription": "SDK 스트림 시작부터 첫 콘텐츠까지의 시간입니다. 그 이전의 요청 대기는 측정할 수 없습니다.",
- "lastGeneration": "지난 생성",
- "finalAverage": "최종 평균",
- "generationPaused": "생성 일시 중지",
- "responsePending": "응답 대기 중",
- "responseFailed": "요청 실패",
- "responseCancelled": "중지됨",
- "responseRetrying": "다시 시도 중",
- "responseGenerating": "생성 중",
- "toolRunning": "도구 실행 중"
+ "forkCancel": "취소"
},
"shareImage": {
"title": "대화 공유",
diff --git a/apps/mobile/src/i18n/locales/zh-CN/session.json b/apps/mobile/src/i18n/locales/zh-CN/session.json
index 1a1a32fbab..99c4141196 100644
--- a/apps/mobile/src/i18n/locales/zh-CN/session.json
+++ b/apps/mobile/src/i18n/locales/zh-CN/session.json
@@ -186,7 +186,7 @@
"outputTotal": "累计输出",
"observedPeak": "峰值",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "实时速度和峰值为估算。响应结束后用真实用量校准;不计工具执行和请求等待。",
+ "tokenRateDescription": "最近一段生成速度,不含工具执行与等待。",
"networkReconnecting": "正在重新连接…",
"modelBusyRetrying": "模型服务繁忙,正在重试…",
"rateLimitRetrying": "请求受到限流,正在重试…",
@@ -267,24 +267,7 @@
"forkConfirmTitle": "从这里开启一个新任务?",
"forkConfirmDescription": "系统会根据这里的对话上下文创建一个独立的新任务。原任务不会改变,之后两边的消息互不影响。",
"forkConfirm": "开启新任务",
- "forkCancel": "取消",
- "estimatedValue": "约{{value}}",
- "estimatedTokenRate": "约{{rate}} tok/s",
- "responseWaiting": "等待 {{seconds}}秒",
- "waitSeconds": "{{seconds}}秒",
- "firstResponse": "首响应",
- "streamWait": "流内等待",
- "turnWaitDescription": "从本地收到引擎开始到首个文字、思考或工具内容,不等同于网络延迟。",
- "streamWaitDescription": "从 SDK 流开始到首个内容块,之前的请求等待不可测。",
- "lastGeneration": "上次生成",
- "finalAverage": "最终平均",
- "generationPaused": "生成暂停",
- "responsePending": "等待响应",
- "responseFailed": "请求失败",
- "responseCancelled": "已停止",
- "responseRetrying": "正在重试",
- "responseGenerating": "正在生成",
- "toolRunning": "执行工具中"
+ "forkCancel": "取消"
},
"shareImage": {
"title": "分享对话",
diff --git a/apps/mobile/src/i18n/locales/zh-TW/session.json b/apps/mobile/src/i18n/locales/zh-TW/session.json
index 2fc6cbe2af..d7d8e8d99f 100644
--- a/apps/mobile/src/i18n/locales/zh-TW/session.json
+++ b/apps/mobile/src/i18n/locales/zh-TW/session.json
@@ -190,7 +190,7 @@
"outputTotal": "累計輸出",
"observedPeak": "峰值",
"tokenRateUnit": "tok/s",
- "tokenRateDescription": "即時速度和峰值為估算。回應結束後以實際用量校準;不計工具執行和請求等待。",
+ "tokenRateDescription": "最近一段生成速度,不含工具執行與等待。",
"networkReconnecting": "正在重新連線…",
"modelBusyRetrying": "模型服務繁忙,正在重試…",
"rateLimitRetrying": "請求受到限流,正在重試…",
@@ -271,24 +271,7 @@
"forkConfirmTitle": "從這裡開啟一個新任務?",
"forkConfirmDescription": "系統會根據這裡的對話上下文建立一個獨立的新任務。原任務不會改變,之後兩邊的訊息互不影響。",
"forkConfirm": "開啟新任務",
- "forkCancel": "取消",
- "estimatedValue": "約{{value}}",
- "estimatedTokenRate": "約{{rate}} tok/s",
- "responseWaiting": "等待 {{seconds}}秒",
- "waitSeconds": "{{seconds}}秒",
- "firstResponse": "首回應",
- "streamWait": "流內等待",
- "turnWaitDescription": "從本機收到引擎開始到首個文字、思考或工具內容,不等同於網路延遲。",
- "streamWaitDescription": "從 SDK 串流開始到首個內容區塊,之前的請求等待無法測量。",
- "lastGeneration": "上次生成",
- "finalAverage": "最終平均",
- "generationPaused": "生成暫停",
- "responsePending": "等待回應",
- "responseFailed": "請求失敗",
- "responseCancelled": "已停止",
- "responseRetrying": "正在重試",
- "responseGenerating": "正在生成",
- "toolRunning": "執行工具中"
+ "forkCancel": "取消"
},
"shareImage": {
"title": "分享對話",
diff --git a/apps/mobile/src/session/RunningTokenRatePopover.tsx b/apps/mobile/src/session/RunningTokenRatePopover.tsx
index 632967c0db..9b948c1cf2 100644
--- a/apps/mobile/src/session/RunningTokenRatePopover.tsx
+++ b/apps/mobile/src/session/RunningTokenRatePopover.tsx
@@ -1,5 +1,3 @@
-import type { ResponseSpeedSnapshot } from "@cindy/maker-shared/usage-format";
-import { responseSpeedActivity, responseSpeedHistory } from "@cindy/maker-shared/usage-format";
import { useEffect, useRef, useState, type ReactNode } from "react";
import {
AccessibilityInfo,
@@ -48,7 +46,6 @@ export function formatTokenRate(rate: number | null): string {
/** Key this component by account/device/session so gestures and counters never cross tasks. */
export function RunningTokenRatePopover({
- responseSpeed,
sessionKey,
startedAt,
outputTokens,
@@ -60,7 +57,6 @@ export function RunningTokenRatePopover({
enabled = true,
history: managedHistory,
}: {
- responseSpeed?: ResponseSpeedSnapshot;
sessionKey: string;
startedAt: number | null;
outputTokens: number;
@@ -201,7 +197,7 @@ export function RunningTokenRatePopover({
: emptyRateHistory(null);
});
useEffect(() => {
- if (managedHistory || responseSpeed) return;
+ if (managedHistory) return;
setInternalHistory((previous) =>
recordRunningTokenRate(previous, {
startedAt,
@@ -212,7 +208,6 @@ export function RunningTokenRatePopover({
);
}, [
managedHistory,
- responseSpeed,
startedAt,
outputTokens,
generationDurationMs,
@@ -225,30 +220,29 @@ export function RunningTokenRatePopover({
}, [managedHistory, sessionKey, internalHistory]);
const [now, setNow] = useState(Date.now);
useEffect(() => {
- const sampledAt = responseSpeed?.sampledAt ?? history.latestSampleAt;
- if (sampledAt === undefined || responseSpeed?.phase === 'complete') return;
+ if (history.latestSampleAt === undefined) return;
setNow(Date.now());
const timer = setTimeout(
() => setNow(Date.now()),
- Math.max(0, sampledAt + RATE_SAMPLE_FRESH_MS - Date.now()),
+ Math.max(0, history.latestSampleAt + RATE_SAMPLE_FRESH_MS - Date.now()),
);
return () => clearTimeout(timer);
- }, [history.latestSampleAt, responseSpeed]);
+ }, [history.latestSampleAt]);
// Keep observing counters before the first rate is available. Only the
// interaction surface is conditional; it must not own sampling lifetime.
if (!enabled) return <View pointerEvents="none">{children}</View>;
- const recent = responseSpeed ? responseSpeedHistory(responseSpeed, Math.max(now, Date.now())).latestRate :
+ const recent =
generationReliable &&
(startedAt === null || startedAt === history.startedAt) &&
history.latestSampleAt !== undefined &&
Math.max(now, Date.now()) - history.latestSampleAt < RATE_SAMPLE_FRESH_MS
? history.latestRate
: null;
- const average = responseSpeed ? responseSpeed.averageRate :
+ const average =
generationReliable && generationDurationMs > 0 && outputTokens > 0
? (outputTokens * 1000) / generationDurationMs
: null;
- const samples = responseSpeed?.samples ?? history.samples;
+ const samples = history.samples;
const firstTime = samples[0]?.durationMs ?? 0;
const span = (samples.at(-1)?.durationMs ?? 0) - firstTime;
const ceiling = Math.max(1, ...samples.map((sample) => sample.rate));
@@ -264,10 +258,6 @@ export function RunningTokenRatePopover({
rate === null
? "—"
: t("session.screen.tokenRate", { rate: formatTokenRate(rate) });
- const approximate = (value: string | number, estimated = true) => estimated && value !== '—'
- ? t('session.screen.estimatedValue', { value }) : value;
- const displayedOutput = responseSpeed?.outputTokens ?? outputTokens;
- const activity = responseSpeed ? responseSpeedActivity(responseSpeed, Math.max(now, Date.now())) : null;
const card = (
<View
pointerEvents={mode === "held" ? "none" : "auto"}
@@ -307,17 +297,10 @@ export function RunningTokenRatePopover({
<View style={styles.top}>
<View style={styles.metric}>
<Text ref={cardFocusTarget} style={styles.label}>
- {t(activity === 'failed' ? 'session.screen.responseFailed'
- : activity === 'cancelled' ? 'session.screen.responseCancelled'
- : activity === 'retrying' ? 'session.screen.responseRetrying'
- : activity === 'complete' ? 'session.screen.finalAverage'
- : activity === 'waiting' ? 'session.screen.responsePending'
- : activity === 'tool' ? 'session.screen.toolRunning'
- : activity === 'paused' ? 'session.screen.generationPaused'
- : activity === 'quiet' ? 'session.screen.responsePending' : 'session.screen.currentRate')}
+ {t("session.screen.currentRate")}
</Text>
<Text style={styles.value}>
- {recent === null ? '—' : approximate(formatTokenRate(recent), Boolean(responseSpeed) && (responseSpeed?.phase !== 'complete' || responseSpeed.estimated))}{" "}
+ {formatTokenRate(recent)}{" "}
<Text style={styles.label}>
{t("session.screen.tokenRateUnit")}
</Text>
@@ -358,15 +341,17 @@ export function RunningTokenRatePopover({
</View>
<View style={styles.top}>
{[
- ["averageRate", approximate(rateText(average), responseSpeed?.estimated ?? false)],
+ ["averageRate", rateText(average)],
[
"outputTotal",
t("session.screen.tokenCount", {
tokens:
- approximate(displayedOutput >= 1000 ? `${(displayedOutput / 1000).toFixed(1)}k` : displayedOutput, responseSpeed?.estimated ?? false),
+ outputTokens >= 1000
+ ? `${(outputTokens / 1000).toFixed(1)}k`
+ : outputTokens,
}),
],
- ["observedPeak", approximate(rateText(samples.length ? Math.max(...samples.map(sample => sample.rate)) : null), Boolean(responseSpeed))],
+ ["observedPeak", rateText(samples.length ? history.peak : null)],
].map(([key, value]) => (
<View style={styles.metric} key={key}>
<Text style={styles.label}>{t(`session.screen.${key}`)}</Text>
@@ -374,12 +359,6 @@ export function RunningTokenRatePopover({
</View>
))}
</View>
- {responseSpeed && (
- <View style={styles.top}>
- <Text style={styles.label}>{t(`session.screen.${responseSpeed.waitOrigin === 'stream' ? 'streamWait' : 'firstResponse'}`)}</Text>
- <Text style={styles.detail}>{responseSpeed.firstResponseMs === null ? '—' : t('session.screen.waitSeconds', { seconds: (responseSpeed.firstResponseMs / 1000).toFixed(1) })}</Text>
- </View>
- )}
</ScrollView>
</View>
);
diff --git a/apps/mobile/src/session/remoteSessionStore.ts b/apps/mobile/src/session/remoteSessionStore.ts
index 9ecbc0609d..8b23ac79d6 100644
--- a/apps/mobile/src/session/remoteSessionStore.ts
+++ b/apps/mobile/src/session/remoteSessionStore.ts
@@ -1,4 +1,3 @@
-import { readResponseSpeedSnapshot, stopResponseSpeed, retryResponseSpeed, resumeResponseSpeed, mergeResponseSpeedStatus, type ResponseSpeedSnapshot } from '@cindy/maker-shared/usage-format';
import { normalizeTaskTags, reconcileTaskTags } from '@cindy/maker-shared';
import {
createContext,
@@ -149,7 +148,6 @@ export interface RemoteSessionRunStatus {
status: string;
tokenUsage: number;
outputTokens: number;
- responseSpeed?: ResponseSpeedSnapshot;
generationDurationMs: number;
generationActive: boolean;
generationReliable: boolean;
@@ -4746,15 +4744,6 @@ export const remoteSessionStore = {
? clearSessionReconnectAttempt(sessionId)
: false;
const fullMessageWriteAllowed = fromList || messageWriteAllowed(sessionId);
- const outputData = isRecord(event.data) ? event.data : null;
- if (((type === 'text' || type === 'thinking') && readString(outputData, 'text')) || type === 'tool_use') {
- const current = readSessionRunStatus(sessionId);
- const responseSpeed = resumeResponseSpeed(current.responseSpeed);
- if (responseSpeed !== current.responseSpeed) {
- writeSessionRunStatus(sessionId, { ...current, responseSpeed });
- emit();
- }
- }
if (type === 'text') {
if (!fullMessageWriteAllowed) {
if (reconnectCleared) emit();
@@ -4837,20 +4826,11 @@ export const remoteSessionStore = {
}
}
}
- const terminalData = isRecord(event.data) ? event.data : null;
- const rawTurn = isRecord(terminalData?.raw) ? terminalData.raw : null;
- const current = readSessionRunStatus(sessionId);
- const cancelled = (!current.isRunning && current.responseSpeed?.outcome === 'cancelled') ||
- terminalData?.cancelled === true || terminalData?.status === 'cancelled' ||
- terminalData?.reason === 'send_cancelled_before_acceptance' ||
- terminalData?.reason === 'turn_continuation_cancelled' ||
- terminalData?.reason === 'user_stop_unconfirmed_wake_tasks';
- const failed = !cancelled && (isTerminalMakerErrorEvent(event) || rawTurn?.status === 'failed' || terminalData?.status === 'failed');
- const terminalErrorChanged = writeSessionRunStatus(sessionId, {
- ...current,
- ...(cancelled ? { hasTerminalError: false } : failed ? { hasTerminalError: true } : {}),
- responseSpeed: stopResponseSpeed(current.responseSpeed, cancelled ? 'cancelled' : failed ? 'failed' : undefined),
- });
+ const terminalErrorChanged = isTerminalMakerErrorEvent(event)
+ && writeSessionRunStatus(sessionId, {
+ ...readSessionRunStatus(sessionId),
+ hasTerminalError: true,
+ });
this.setSessionRunning(
sessionId,
false,
@@ -4877,7 +4857,6 @@ export const remoteSessionStore = {
...current,
isRunning: true,
reconnectAttempt,
- responseSpeed: data?.willRetry === true ? retryResponseSpeed(current.responseSpeed) : current.responseSpeed,
startedAt: current.startedAt ?? Date.now(),
});
if (changed || textFlushed) emit();
@@ -5055,8 +5034,6 @@ export const remoteSessionStore = {
status: rawStatus ?? current.status,
tokenUsage,
outputTokens,
- responseSpeed: mergeResponseSpeedStatus(current.responseSpeed,
- readResponseSpeedSnapshot(data?.responseSpeed, Date.now()), isRunning, isTurnStart),
generationDurationMs,
generationActive,
generationReliable,
@@ -5638,7 +5615,6 @@ function clearLiveGenerationOnWideRunStart(
return {
...next,
outputTokens: 0,
- responseSpeed: undefined,
generationDurationMs: 0,
generationActive: false,
generationReliable: true,
@@ -5659,9 +5635,6 @@ function writeMakerTurnRunning(sessionId: string, running: boolean): boolean {
function writeSessionRunStatus(sessionId: string, next: RemoteSessionRunStatus): boolean {
const current = readSessionRunStatus(sessionId);
- if (!next.isRunning && next.responseSpeed?.phase !== 'complete') {
- next = { ...next, responseSpeed: stopResponseSpeed(next.responseSpeed), generationActive: false };
- }
if (next.isRunning && !current.isRunning && current.hasTerminalError) {
next = { ...next, hasTerminalError: false };
}
diff --git a/apps/mobile/src/session/useRunningTokenRateHistory.ts b/apps/mobile/src/session/useRunningTokenRateHistory.ts
index 447cb9a434..2fbcca018b 100644
--- a/apps/mobile/src/session/useRunningTokenRateHistory.ts
+++ b/apps/mobile/src/session/useRunningTokenRateHistory.ts
@@ -15,7 +15,6 @@ export function useRunningTokenRateHistory(input: {
outputTokens: number;
generationDurationMs: number;
generationReliable: boolean;
- generationActive?: boolean;
/** 本地发送或队列已让活动条出现,但远端本轮 startedAt 可能仍是 null。 */
streaming?: boolean;
}) {
@@ -25,7 +24,6 @@ export function useRunningTokenRateHistory(input: {
outputTokens,
generationDurationMs,
generationReliable,
- generationActive,
streaming = false,
} = input;
// 挂载时从按会话的进程内缓存播种:ComposerActivityStatus 以账号、设备和任务身份为 key,
@@ -68,9 +66,9 @@ export function useRunningTokenRateHistory(input: {
) {
return { ...recorded, baseline: null, lastReport: null, latestRate: null };
}
- return generationActive === false ? { ...recorded, latestRate: null, latestSampleAt: undefined } : recorded;
+ return recorded;
});
- }, [sessionKey, startedAt, outputTokens, generationDurationMs, generationReliable, generationActive, streaming]);
+ }, [sessionKey, startedAt, outputTokens, generationDurationMs, generationReliable, streaming]);
useEffect(() => {
if (sessionKey) saveCachedRateHistory(sessionKey, history);
}, [sessionKey, history]);
diff --git a/docs/dev-rules/assets/response-speed-completed-dark.png b/docs/dev-rules/assets/response-speed-completed-dark.png
deleted file mode 100644
index d0c3296727..0000000000
Binary files a/docs/dev-rules/assets/response-speed-completed-dark.png and /dev/null differ
diff --git a/docs/dev-rules/assets/response-speed-completed-light.png b/docs/dev-rules/assets/response-speed-completed-light.png
deleted file mode 100644
index 311432a012..0000000000
Binary files a/docs/dev-rules/assets/response-speed-completed-light.png and /dev/null differ
diff --git a/docs/dev-rules/protocol-compatibility.md b/docs/dev-rules/protocol-compatibility.md
index 3d3e9f636c..b8a61bb4c3 100644
--- a/docs/dev-rules/protocol-compatibility.md
+++ b/docs/dev-rules/protocol-compatibility.md
@@ -11,20 +11,6 @@
> **增量适用原则**:wire protocol 兼容对所有跨端改动生效,不因是小改而豁免。
-## 回复速度快照
-
-既有 maker status 可选携带 `responseSpeed`(等待边界、估算计数、最近/平均速度与最多
-60 个采样点);仅用于显示,不参与费用或上下文计数。新客户端校验该字段,旧主机省略时
-沿用原 usage 路径,旧客户端忽略新增字段。无需服务端更新、新通道、权限或数据库 migration。
-原生重试事件可为快照附加可选 `retrying`;两端亦从既有 error / done / Stop 事件
-附加可选 `outcome` 与 `retrying` 注记,
-只反映明确失败、取消或自动重试;不要求主机发送新事件,不落盘,不更改错误原因或操作契约。
-终态真实 output 与生成 duration 的匹配沿现有消息用量记录处理。详见
-[`response-speed.md`](response-speed.md)。
-Claude 的既有 done payload 可选携带 `turnUsage`,来自 SDK result 用量的本轮增量,
-供历史速度与消息用量配对。原 `usage` 保持会话累计口径;旧端忽略新字段,新端缺字段且无模型增量时
-省略该速度,不将累计用量当本轮输出。费用与上下文仍走原计量路径。
-
## 任务列表提前同步聊天正文
同账号控制端声明 `session-list-messages-v1` 后,`sessions` 订阅同时接收普通用户/助手
diff --git a/docs/dev-rules/response-speed.md b/docs/dev-rules/response-speed.md
deleted file mode 100644
index 96b5ae1bfb..0000000000
--- a/docs/dev-rules/response-speed.md
+++ /dev/null
@@ -1,67 +0,0 @@
-# 回复速度测量与展示
-
-速度是展示投影,计费、缓存命中与上下文容量仍由引擎真实 usage 管理。流式速度不能回写
-`UsageTracker`、账单或 token 余额。实现位于 `agents/shared/response-speed.ts`,两端复用
-`maker-shared/usage-format` 的快照校验与图表模型。
-
-## 计时边界
-
-| 引擎 | 等待起点 | 首响应终点 | 校准来源 |
-| --- | --- | --- | --- |
-| Pi | 本地 `agent_start` | text / thinking / toolcall 的内容开始或首个 delta | 每条 assistant `message_end.usage.output` |
-| Claude Code | SDK `message_start` | 主线 `content_block_start` 或 delta;子 Agent 独立 | 主线 `message_delta.usage.output_tokens` |
-| Codex | 本地接受 `turn/started` | reasoning/tool item 开始或已去重文字 delta | 本轮真实 usage segments 的累计 output |
-
-这些是本地可观测边界,包含思考和工具参数,不代表首个可见文字或网络 TTFT。
-Claude 显示「流内等待」,SDK 暴露事件前的请求时间不可测。Pi/Codex 显示「首响应」。
-排队发生在引擎开始前,不进入等待或生成速度。
-
-生成计时从内容开始,到响应结束或工具/审批暂停;最后一个工具/审批结束后重新显示等待,下一次内容才恢复;并行内容已恢复时不被工具结束打断。
-Codex 在 `agentMessage` 完成时关闭观测响应,迟到 usage/turn 完成只校准计数,不延长生成时间。
-Codex 工具参数、patch 和 plan 只有成品 item,没有可配对的参数生成流;含这些输出的轮次
-不能用整轮 usage 放大后续回答的采样。保留已观测内容的估算曲线与完成平均,真实账单总量
-照常保存,但省略该轮历史 TPS。下一轮重新判断;纯流式回复仍可校准。
-Pi 重试关闭失败片段,下一请求重新等待;每轮重新清空测量。零 usage 是有效零;缺少 usage
-则保留估算。仅结束事件有 output、没有可观测流,或检测到挂起/时钟异常时不提供 TPS。
-
-## 实时与终态
-
-- 字符加权估计避免把一个传输片段当一个精确 token。它跨分片可加,但仍不是 tokenizer。
-- 最近 1 秒滑窗,Pi/Claude 最多每 250ms 推进;Codex沿用 500ms 状态节流。
- 窗口按固定 4ms 时间桶聚合,边缘最多提前裁掉不足 4ms 的估算量;旧桶不会因新输出续期。
- 窗口最多 256 个聚合点,图表最多 60 个采样点,不保留文本。
-- 实时速率、峰值及未校准输出标「约」。真实 usage 到达后校准该响应的曲线及输出;
- 完成时显示真实输出除以已观测生成时间的最终平均。校准后的曲线和峰值仍是估计分配。
-- 工具期间不冒充正在生成,当前速度为空;Pi 的 tool_execution、Codex 的配对 item 执行边界显示「执行工具中」,不能从参数流推出执行已开始。Claude 缺少同等执行事件时只标「生成暂停」;并行模型 delta 优先显示实际生成。下一请求重新显示等待响应。静默超过观测窗口与首响应前统一显示「等待响应」,内部仍区分请求等待与无新采样;静默不重启首响应计时,也不把没有可见内容当真实零输出或 API 失败。远端快照的时间戳在接收端
- 归一为本地观测时间,避免用两台机器的绝对时钟差延长等待或抹掉当前速度。
-- 停止/终态错误可能没有最后一份 usage,冻结最后观测并保留估算标识。明确终态错误标「请求失败」,用户取消标「已停止」,不冒充成功完成。曲线与已观测平均保留,当前速度为空。
-- 仅原生 `api_retry` / `auto_retry_start` 或明确 `willRetry: true` 的自动重试事件标「正在重试」;沿用原错误原因、重连进度与重试操作。普通警告与无输出不触发此态。实际文字/思考/工具调用输出或新输出快照恢复后撤掉重试标记;失败/取消标记保留至下一轮,普通收尾 status 不抹成成功。
-- 原速度入口在完成后保留「上次生成」,点击/键盘可重开详情;不依赖 hover 或先 pin,
- 不固定几秒后隐藏。下一轮用本轮等待替换,切换任务不把别的任务的数据搬过来。
-
-## 回看与持久化
-
-最新快照留在任务的进程内状态,切换任务或组件重新挂载可以回看;图表不写数据库。
-应用/Renderer 重载后首响应和曲线不恢复。已有消息用量记录继续保存真实 output 与经过
-计数匹配检查的生成 duration,支持历史最终平均 TPS;缺失或计数不匹配时省略 TPS。
-不增加数据库 schema、migration 或首响应持久字段。
-Claude result 的原始 `usage` 是会话累计值;历史速度与消息详情使用同一份真实本轮
-`turnUsage` 增量(或模型增量),不能把累计输出与本轮生成时间配对。它不来自流式估算。
-
-## 兼容与验证
-
-`UsageSnapshot.responseSpeed` 是既有 status 的可选增量字段:旧主机缺省,新界面使用原
-usage 算法;旧客户端忽略新字段。无新 IPC channel、服务端协议版本或移动端原生指纹。
-
-事件序列验证见 `response-speed.test.ts`、`response-speed-adapters.test.ts`;
-两端组件/任务状态验证覆盖工具暂停、完成保留、下一轮、停止、任务切换与旧主机兼容。
-真实设备与离线受控事件截图必须分别记录,不把离线结果声称为真实模型吞吐。
-
-## 组件验收画面
-
-真实状态栏和速度卡,使用离线受控事件(2 秒首响应、300 输出、4 秒生成),并非真实模型
-吞吐测量。两图分别加载完整 Cindy Light / Dark 主题;完成超过旧淡出时间后仍可点击查看。
-
-![完成后的速度入口和详情(浅色)](assets/response-speed-completed-light.png)
-
-![完成后的速度入口和详情(深色)](assets/response-speed-completed-dark.png)
diff --git a/packages/maker-core/src/agents/claude-code/__tests__/translator-error-result.test.ts b/packages/maker-core/src/agents/claude-code/__tests__/translator-error-result.test.ts
index 960aa71abd..c2d7a224db 100644
--- a/packages/maker-core/src/agents/claude-code/__tests__/translator-error-result.test.ts
+++ b/packages/maker-core/src/agents/claude-code/__tests__/translator-error-result.test.ts
@@ -525,12 +525,6 @@ describe('Claude Code translator is_error result guard', () => {
ctx,
);
- const progressIterator = queue[Symbol.asyncIterator]();
- const progress = await progressIterator.next();
- expect(progress.value).toMatchObject({ type: 'status', data: {
- isRunning: true, responseSpeed: { retrying: true, recentRate: null },
- } });
- await progressIterator.return?.();
expect(queue.pending, 'retryable envelope must not close the turn').toBe(0);
translateSdkMessage(
diff --git a/packages/maker-core/src/agents/claude-code/generation-timing.test.ts b/packages/maker-core/src/agents/claude-code/generation-timing.test.ts
index 3a5fbf5135..95b02b92a6 100644
--- a/packages/maker-core/src/agents/claude-code/generation-timing.test.ts
+++ b/packages/maker-core/src/agents/claude-code/generation-timing.test.ts
@@ -2011,42 +2011,3 @@ describe('claude generation pause boundaries', () => {
vi.useRealTimers();
});
});
-
-describe('Claude normalized completed-turn output', () => {
- afterEach(() => { vi.useRealTimers(); vi.restoreAllMocks(); });
-
- it('preserves raw SDK aggregates while exposing real turn output across success, zero and reset', () => {
- vi.useFakeTimers();
- const ctx = createTranslatorCtx();
- const events: AgentEvent[] = [];
- const queue = { push: (event: AgentEvent) => events.push(event) } as unknown as Parameters<typeof translateSdkMessage>[1];
- const counts = [100, 60, 0, 30];
- const aggregates = [100, 160, 160, 30];
- for (const [index, output] of counts.entries()) {
- const start = 1_000 + index * 10_000;
- vi.setSystemTime(start);
- resetClaudeGenerationTiming(ctx.rt.generation);
- ctx.tracker.beginTurn();
- ctx.turn = createTurnState();
- translateSdkMessage({ type: 'stream_event', event: { type: 'message_start',
- message: { id: `m-${index}`, model: 'claude-sonnet-4.5', usage: { input_tokens: 0 } } } }, queue, ctx);
- vi.setSystemTime(start + 1_000);
- translateSdkMessage({ type: 'stream_event', event: { type: 'content_block_start',
- content_block: { type: 'text', text: '' } } }, queue, ctx);
- translateSdkMessage({ type: 'stream_event', event: { type: 'content_block_delta',
- delta: { type: 'text_delta', text: 'answer' } } }, queue, ctx);
- vi.setSystemTime(start + 3_000);
- translateSdkMessage({ type: 'stream_event', event: { type: 'message_delta',
- usage: { output_tokens: output } } }, queue, ctx);
- vi.setSystemTime(start + 5_000);
- translateSdkMessage({ type: 'result', subtype: 'success', result: 'answer', total_cost_usd: 0,
- usage: { input_tokens: 0, output_tokens: aggregates[index] } }, queue, ctx);
- const done = events.filter(event => event.type === 'done').at(-1);
- expect(done?.data).toMatchObject({ usage: { output_tokens: aggregates[index] }, turnUsage: { output_tokens: output },
- responseSpeed: { phase: 'complete', outputTokens: output, durationMs: 2_000,
- averageRate: output / 2, estimated: false } });
- expect(done?.data).not.toHaveProperty('modelUsage');
- expect(ctx.rt.lastResultUsageAggregate?.outputTokens).toBe(aggregates[index]);
- }
- });
-});
diff --git a/packages/maker-core/src/agents/claude-code/generation-timing.ts b/packages/maker-core/src/agents/claude-code/generation-timing.ts
index 8003013381..ec16960afb 100644
--- a/packages/maker-core/src/agents/claude-code/generation-timing.ts
+++ b/packages/maker-core/src/agents/claude-code/generation-timing.ts
@@ -1,9 +1,7 @@
-import { ResponseSpeedTracker } from '../shared/response-speed.js';
const CLAUDE_GENERATION_HEARTBEAT_MS = 5_000;
const CLAUDE_GENERATION_SUSPEND_GAP_MS = 30_000;
export interface ClaudeGenerationState {
- responseSpeed: ResponseSpeedTracker;
startedAt: number | null;
durationMs: number;
/** Model time sampled with the latest parent output usage. */
@@ -33,7 +31,6 @@ export interface ClaudeGenerationState {
export function newClaudeGenerationState(): ClaudeGenerationState {
return {
- responseSpeed: new ResponseSpeedTracker('stream'),
startedAt: null,
durationMs: 0,
outputDurationMs: 0,
@@ -60,7 +57,6 @@ function sampleHeartbeat(state: ClaudeGenerationState, now = Date.now()): void {
now - previous > CLAUDE_GENERATION_HEARTBEAT_MS + CLAUDE_GENERATION_SUSPEND_GAP_MS
) {
state.reliable = false;
- state.responseSpeed.invalidate();
}
state.heartbeatAt = now;
}
@@ -88,7 +84,6 @@ function closeInterval(state: ClaudeGenerationState, endedAt: number): void {
export function resetClaudeGenerationTiming(state: ClaudeGenerationState): void {
stopHeartbeat(state);
- state.responseSpeed.reset('stream');
state.startedAt = null;
state.pendingToolIds.clear();
state.settledPauseIds.clear();
@@ -119,7 +114,6 @@ export function beginClaudeGenerationAtRequestStart(
): void {
for (const pauseId of state.pendingToolIds) state.settledPauseIds.add(pauseId);
state.pendingToolIds.clear();
- state.responseSpeed.beginRequest(startedAt);
beginClaudeGeneration(state, startedAt);
}
@@ -135,7 +129,6 @@ export function pauseClaudeGeneration(
if (state.pendingToolIds.has(pauseId)) return;
if (state.settledPauseIds.has(pauseId)) return;
if (state.pendingToolIds.size === 0) {
- state.responseSpeed.pause(pausedAt);
// No open generation interval means we never saw message_start (or it was
// already closed). Pausing here would resume the clock at tool_result and
// still count earlier output tokens, so tok/s would be inflated.
diff --git a/packages/maker-core/src/agents/claude-code/translator.ts b/packages/maker-core/src/agents/claude-code/translator.ts
index a333386e39..9f55212e04 100644
--- a/packages/maker-core/src/agents/claude-code/translator.ts
+++ b/packages/maker-core/src/agents/claude-code/translator.ts
@@ -39,6 +39,7 @@ import type { UsageTracker } from '../shared/usage-tracker.js';
import { isModelAccessDenied } from './model-access-error.js';
import { attachLiveGeneration, sampleGenerationDuration } from '../shared/live-generation-snapshot.js';
import {
+ beginClaudeGeneration,
beginClaudeGenerationAtRequestStart,
finalizeClaudeGeneration,
markClaudeGenerationUnreliable,
@@ -342,7 +343,6 @@ function ccLiveStatus(
return {
status,
...attachLiveGeneration(ctx.tracker.snapshot(), {
- responseSpeed: ctx.rt.generation.responseSpeed.snapshot(),
outputTokens: mainTurnOutputTokens(ctx.tracker),
durationMs: ctx.rt.generation.outputDurationMs,
openStartedAt: ctx.rt.generation.startedAt,
@@ -1068,8 +1068,6 @@ function handleSystem(
return;
}
if (msg.subtype === 'api_retry') {
- if (ctx.turn.interruptRequested || !ctx.rt.generation.responseSpeed.beginRetry()) return;
- queue.push({ type: 'status', data: ccLiveStatus(ctx, 'Working...', true), source: 'claude-code' });
// SDKAPIRetryMessage 明确表示本次 API 失败仍在自动重试,不是 turn 终态。
// 除日志外也暂存最后一次 retry 的错误详情,覆盖 SDK 没有额外发送
// assistant.error envelope、最终 ResultMessage 又没有 result 文本的路径。
@@ -1691,18 +1689,6 @@ function handleStreamEvent(
: undefined;
// 冗余 add: 防 SDK 顺序契约变化 (stream_event 先于 assistant message yield), Set 幂等。
- if (!parentToolUseId && event.type === 'content_block_start') ctx.rt.generation.responseSpeed.content();
- if (!parentToolUseId && event.type === 'content_block_delta') {
- const delta = event.delta;
- const text = delta?.text ?? delta?.thinking ?? delta?.partial_json;
- if (typeof text === 'string' && ctx.rt.generation.responseSpeed.delta(text)) {
- queue.push({ type: 'status', data: ccLiveStatus(ctx, 'Generating...', true), source: 'claude-code' });
- }
- }
- if (!parentToolUseId && event.type === 'message_delta') {
- ctx.rt.generation.responseSpeed.reportOutput(event.usage?.output_tokens);
- ctx.rt.generation.responseSpeed.pause();
- }
if (event.type === 'content_block_start') {
const cb = event.content_block;
if (cb && cb.type === 'text') {
@@ -2210,7 +2196,6 @@ function handleResult(
resultUsage != null && segmentTotals.outputTokens === resultUsage.outputTokens,
);
const liveGeneration = ctx.rt.generation;
- liveGeneration.responseSpeed.finish();
const endSnapshot = ctx.tracker.endTurn(
resultUsage
? {
@@ -2492,7 +2477,6 @@ function handleResult(
data: {
status: 'Done',
...attachLiveGeneration(endSnapshot, {
- responseSpeed: liveGeneration.responseSpeed.snapshot(),
outputTokens: liveTurnOutput,
durationMs: liveGeneration.outputDurationMs,
openStartedAt: null,
@@ -2516,15 +2500,6 @@ function handleResult(
: msg;
const resultWithUsageSegments = {
...safeResult,
- responseSpeed: liveGeneration.responseSpeed.snapshot(),
- // Keep SDK usage intact; the normalized provider delta can pair
- // with the current turn's response-speed denominator.
- turnUsage: resultUsage ? {
- input_tokens: resultUsage.inputTokens,
- output_tokens: resultUsage.outputTokens,
- cache_read_input_tokens: resultUsage.cacheReadTokens,
- cache_creation_input_tokens: resultUsage.cacheCreateTokens,
- } : undefined,
usageSegments: turnUsageSegments,
usageSegmentsComplete,
modelUsageCumulativeStartsAtZero: ctx.modelUsageCumulativeStartsAtZero?.() === true,
diff --git a/packages/maker-core/src/agents/codex/index.test.ts b/packages/maker-core/src/agents/codex/index.test.ts
index 35abb5a8e9..69d4b61dde 100644
--- a/packages/maker-core/src/agents/codex/index.test.ts
+++ b/packages/maker-core/src/agents/codex/index.test.ts
@@ -18385,8 +18385,6 @@ describe('CodexAgent MCP thread context hooks', () => {
}
now = 5_000;
- handlers.itemStarted!({ threadId: 'start-thread-id', turnId: 'turn-ordering',
- item: { id: 'ordering-reasoning', type: 'reasoning', summary: [], content: [] } });
handlers.turnCompleted({
threadId: 'start-thread-id',
turn: { id: 'turn-ordering', status: 'completed' },
@@ -18396,8 +18394,7 @@ describe('CodexAgent MCP thread context hooks', () => {
});
const done = events.find((event) => event.type === 'done');
expect((done?.data as { usage?: { durationMs?: number } }).usage?.durationMs)
- .toBeUndefined();
- expect(events.findLast((event) => event.type === 'status')?.data.responseSpeed).toMatchObject({ outputTokens: 0, averageRate: null, firstResponseMs: expectedDurationMs });
+ .toBe(expectedDurationMs);
await handle.close();
} finally {
@@ -18406,146 +18403,6 @@ describe('CodexAgent MCP thread context hooks', () => {
},
);
- it.each(['serial', 'parallel-tools', 'parallel-output'])('publishes paired tool phases and closes response time before delayed usage/turn completion (%s)', async (mode) => {
- let now = 1_000;
- const nowSpy = vi.spyOn(Date, 'now').mockImplementation(() => now);
- const agent = new CodexAgent(createDeps());
- const host = installFakeHost(agent);
- const handle = await agent.startSession({ sessionId: 'session-speed-lifecycle', model: 'gpt-5.4', workingDir: '/repo' });
- const handlers = host.getThreadHandlers()!;
- const events: AgentEvent[] = [];
- void (async () => { for await (const event of handle.events()) events.push(event); })();
- const latestSpeed = () => events.findLast((event) => event.type === 'status')?.data.responseSpeed;
- try {
- handlers.turnStarted!({ threadId: 'start-thread-id', turn: { id: 'speed-turn' } });
- now = 2_000;
- handlers.agentMessageDelta!({ threadId: 'start-thread-id', turnId: 'speed-turn', itemId: 'answer-1', delta: 'abcd' });
- now = 3_000;
- handlers.itemStarted!({ threadId: 'start-thread-id', turnId: 'speed-turn',
- item: { id: 'tool-1', type: 'commandExecution', command: 'pwd', cwd: '/repo', status: 'inProgress' } } as never);
- await waitForExpectation(() => expect(latestSpeed()).toMatchObject({ phase: 'paused', toolActive: true, durationMs: 1_000 }));
- // A duplicate start must not reopen generation during tool execution.
- now = 3_100;
- handlers.itemStarted!({ threadId: 'start-thread-id', turnId: 'speed-turn',
- item: { id: 'tool-1', type: 'commandExecution', command: 'pwd', cwd: '/repo', status: 'inProgress' } } as never);
- await waitForExpectation(() => expect(latestSpeed()).toMatchObject({ phase: 'paused', durationMs: 1_000 }));
- if (mode === 'parallel-tools') {
- now = 4_000;
- handlers.itemStarted!({ threadId: 'start-thread-id', turnId: 'speed-turn',
- item: { id: 'tool-2', type: 'commandExecution', command: 'pwd', cwd: '/repo', status: 'inProgress' } } as never);
- now = 8_000;
- handlers.itemCompleted!({ threadId: 'start-thread-id', turnId: 'speed-turn',
- item: { id: 'tool-2', type: 'commandExecution', command: 'pwd', cwd: '/repo', status: 'completed' } } as never);
- await waitForExpectation(() => expect(latestSpeed()).toMatchObject({ phase: 'paused', toolActive: true }));
- }
- if (mode === 'parallel-output') {
- now = 8_000;
- handlers.agentMessageDelta!({ threadId: 'start-thread-id', turnId: 'speed-turn', itemId: 'parallel-answer', delta: 'ijkl' });
- }
- now = 9_000;
- handlers.itemCompleted!({ threadId: 'start-thread-id', turnId: 'speed-turn',
- item: { id: 'tool-1', type: 'commandExecution', command: 'pwd', cwd: '/repo', status: 'completed', aggregatedOutput: '/repo' } } as never);
- await waitForExpectation(() => expect(latestSpeed()).toMatchObject({ phase: mode === 'parallel-output' ? 'generating' : 'waiting', toolActive: false, waitingMs: 0 }));
- now = 11_000;
- handlers.agentMessageDelta!({ threadId: 'start-thread-id', turnId: 'speed-turn', itemId: 'answer-2', delta: 'efgh' });
- now = 13_000;
- handlers.itemCompleted!({ threadId: 'start-thread-id', turnId: 'speed-turn',
- item: { id: 'answer-2', type: 'agentMessage', text: 'efgh' } } as never);
- const observedMs = mode === 'parallel-output' ? 6_000 : 3_000;
- await waitForExpectation(() => expect(latestSpeed()).toMatchObject({ phase: 'paused', durationMs: observedMs }));
- now = 19_000;
- handlers.tokenUsageUpdated!({ threadId: 'start-thread-id', turnId: 'speed-turn', tokenUsage: {
- total: { inputTokens: 10, cachedInputTokens: 0, outputTokens: 60, reasoningOutputTokens: 0, totalTokens: 70 },
- last: { inputTokens: 10, cachedInputTokens: 0, outputTokens: 60, reasoningOutputTokens: 0, totalTokens: 70 },
- } } as never);
- now = 25_000;
- handlers.turnCompleted!({ threadId: 'start-thread-id', turn: { id: 'speed-turn', status: 'completed', durationMs: 24_000 } });
- await waitForExpectation(() => expect(events.some((event) => event.type === 'done')).toBe(true));
- // Tool arguments are not a measured stream: the whole-turn 60 output
- // tokens must not inflate these two/three observed character estimates.
- const observedTokens = mode === 'parallel-output' ? 3 : 2;
- expect(latestSpeed()).toMatchObject({ phase: 'complete', outputTokens: observedTokens,
- durationMs: observedMs, averageRate: observedTokens * 1_000 / observedMs, estimated: true });
- const usage = events.find((event) => event.type === 'done')?.data.usage;
- expect(usage).toMatchObject({ completionTokens: 60 });
- expect(usage?.durationMs).toBeUndefined();
- now = 26_000;
- handlers.turnStarted!({ threadId: 'start-thread-id', turn: { id: 'next-speed-turn' } });
- now = 26_100;
- handlers.itemStarted!({ threadId: 'start-thread-id', turnId: 'next-speed-turn', item: { id: 'reason-2', type: 'reasoning', summary: [], content: [] } } as never);
- await waitForExpectation(() => expect(latestSpeed()?.outputTokens).toBe(0));
- } finally {
- await handle.close();
- nowSpy.mockRestore();
- }
- });
-
- it.each(['commandExecution', 'mcpToolCall', 'fileChange', 'plan'])(
- 'keeps tool-first %s turns approximate and restores calibration on the next text-only turn',
- async (type) => {
- let now = 1_000;
- const nowSpy = vi.spyOn(Date, 'now').mockImplementation(() => now);
- const agent = new CodexAgent(createDeps());
- const host = installFakeHost(agent);
- const handle = await agent.startSession({ sessionId: `session-tool-first-${type}`, model: 'gpt-5.4', workingDir: '/repo' });
- const handlers = host.getThreadHandlers()!;
- const events: AgentEvent[] = [];
- void (async () => { for await (const event of handle.events()) events.push(event); })();
- const latestSpeed = () => events.findLast(event => event.type === 'status')?.data.responseSpeed;
- try {
- handlers.turnStarted!({ threadId: 'start-thread-id', turn: { id: 'tool-first' } });
- const item = type === 'commandExecution'
- ? { id: 'first-tool', type, command: 'pwd', cwd: '/repo', status: 'inProgress' }
- : type === 'mcpToolCall'
- ? { id: 'first-tool', type, server: 'test', tool: 'lookup', arguments: { query: 'where' }, status: 'inProgress' }
- : type === 'fileChange'
- ? { id: 'first-tool', type, changes: [{ path: '/repo/example', kind: { type: 'update' }, diff: '+answer' }], status: 'completed' }
- : { id: 'first-tool', type, text: 'Look up the answer' };
- const execution = type === 'commandExecution' || type === 'mcpToolCall';
- now = 2_000;
- if (execution) {
- handlers.itemStarted!({ threadId: 'start-thread-id', turnId: 'tool-first', item } as never);
- await waitForExpectation(() => expect(latestSpeed()).toMatchObject({ phase: 'paused', toolActive: true, firstResponseMs: 1_000, durationMs: 0 }));
- now = 8_000;
- }
- handlers.itemCompleted!({ threadId: 'start-thread-id', turnId: 'tool-first', item: { ...item, status: 'completed' } } as never);
- if (execution) await waitForExpectation(() => expect(latestSpeed()).toMatchObject({ phase: 'waiting', toolActive: false }));
-
- const finishAnswer = async (turnId: string, start: number, cumulativeOutput: number, lastOutput: number) => {
- now = start;
- handlers.agentMessageDelta!({ threadId: 'start-thread-id', turnId, itemId: `${turnId}-answer`, delta: 'a'.repeat(40) });
- now = start + 1_000;
- handlers.agentMessageDelta!({ threadId: 'start-thread-id', turnId, itemId: `${turnId}-answer`, delta: 'b'.repeat(40) });
- await waitForExpectation(() => expect(latestSpeed()).toMatchObject({ phase: 'generating', recentRate: 10 }));
- now = start + 2_000;
- handlers.itemCompleted!({ threadId: 'start-thread-id', turnId, item: { id: `${turnId}-answer`, type: 'agentMessage', text: 'a'.repeat(40) + 'b'.repeat(40) } } as never);
- now = start + 3_000;
- handlers.tokenUsageUpdated!({ threadId: 'start-thread-id', turnId, tokenUsage: {
- total: { inputTokens: 20, cachedInputTokens: 0, outputTokens: cumulativeOutput, reasoningOutputTokens: 0, totalTokens: cumulativeOutput + 20 },
- last: { inputTokens: 10, cachedInputTokens: 0, outputTokens: lastOutput, reasoningOutputTokens: 0, totalTokens: lastOutput + 10 },
- } } as never);
- handlers.turnCompleted!({ threadId: 'start-thread-id', turn: { id: turnId, status: 'completed' } });
- await waitForExpectation(() => expect(latestSpeed()?.phase).toBe('complete'));
- return events.findLast(event => event.type === 'done')?.data.usage;
- };
- const mixedUsage = await finishAnswer('tool-first', 10_000, 150, 150);
- expect(latestSpeed()).toMatchObject({ estimated: true, outputTokens: 20, durationMs: 2_000, averageRate: 10 });
- expect(latestSpeed()?.samples.every((sample: { rate: number }) => sample.rate <= 10)).toBe(true);
- expect(mixedUsage).toMatchObject({ completionTokens: 150 });
- expect(mixedUsage?.durationMs).toBeUndefined();
-
- now = 14_000;
- handlers.turnStarted!({ threadId: 'start-thread-id', turn: { id: 'text-only' } });
- const textUsage = await finishAnswer('text-only', 15_000, 190, 40);
- expect(latestSpeed()).toMatchObject({ estimated: false, outputTokens: 40, durationMs: 2_000, averageRate: 20 });
- expect(textUsage).toMatchObject({ completionTokens: 40, durationMs: 2_000 });
- } finally {
- await handle.close();
- nowSpy.mockRestore();
- }
- },
- );
-
it('starts response-first generation timing at turn/started and excludes an approval wait', async () => {
let now = 1_000;
const nowSpy = vi.spyOn(Date, 'now').mockImplementation(() => now);
@@ -18619,10 +18476,7 @@ describe('CodexAgent MCP thread context hooks', () => {
});
await waitForExpectation(() => expect(events.some((event) => event.type === 'done')).toBe(true));
const done = events.find((event) => event.type === 'done');
- // No root output/usage was observed: tool/interaction facts remain
- // available, but cannot become a persisted token-rate denominator.
- expect((done?.data as { usage?: { durationMs?: number } }).usage?.durationMs).toBeUndefined();
- expect(events.findLast((event) => event.type === 'status')?.data.responseSpeed).toMatchObject({ outputTokens: 0, averageRate: null, firstResponseMs: null });
+ expect((done?.data as { usage?: { durationMs?: number } }).usage?.durationMs).toBe(3_000);
await handle.close();
} finally {
@@ -18711,9 +18565,7 @@ describe('CodexAgent MCP thread context hooks', () => {
expect(events.some((event) => event.type === 'done')).toBe(true),
);
const done = events.find((event) => event.type === 'done');
- // A last-only usage notification without a cumulative ledger cursor and
- // no content stream cannot provide a calibrated generation rate.
- expect((done?.data as { usage?: { durationMs?: number } }).usage?.durationMs).toBeUndefined();
+ expect((done?.data as { usage?: { durationMs?: number } }).usage?.durationMs).toBe(1_500);
await handle.close();
} finally {
@@ -18789,10 +18641,7 @@ describe('CodexAgent MCP thread context hooks', () => {
expect(handle.getUsageSnapshot().tokenUsage).toBe(540 - cachedInputTokens);
handlers.turnCompleted!({ threadId: 'start-thread-id', turn: { id: 'turn-1', status: 'completed' } });
await waitForExpectation(() => expect(events.some((event) => event.type === 'done')).toBe(true));
- // Real output exists, but no streamed text was observed: keep accounting
- // and the legacy paired sample, while omitting an unobservable final TPS.
- expect(events.find((event) => event.type === 'done')?.data.usage.durationMs).toBeUndefined();
- expect(events.filter((event) => event.type === 'status').at(-1)?.data.responseSpeed).toMatchObject({ outputTokens: 500, averageRate: null });
+ expect(events.find((event) => event.type === 'done')?.data.usage.durationMs).toBe(5_000);
expect(events.filter((event) => event.type === 'status').at(-1)?.data).toMatchObject({
status: 'Done', outputTokens: 500, generationDurationMs: 5_000,
});
@@ -19161,10 +19010,7 @@ describe('CodexAgent MCP thread context hooks', () => {
});
await waitForExpectation(() => expect(events.some((event) => event.type === 'done')).toBe(true));
const done = events.find((event) => event.type === 'done');
- // No root output/usage was observed: tool/interaction facts remain
- // available, but cannot become a persisted token-rate denominator.
- expect((done?.data as { usage?: { durationMs?: number } }).usage?.durationMs).toBeUndefined();
- expect(events.findLast((event) => event.type === 'status')?.data.responseSpeed).toMatchObject({ outputTokens: 0, averageRate: null, firstResponseMs: 1_000 });
+ expect((done?.data as { usage?: { durationMs?: number } }).usage?.durationMs).toBe(3_000);
await handle.close();
} finally {
@@ -19242,10 +19088,7 @@ describe('CodexAgent MCP thread context hooks', () => {
});
await waitForExpectation(() => expect(events.some((event) => event.type === 'done')).toBe(true));
const done = events.find((event) => event.type === 'done');
- // No root output/usage was observed: tool/interaction facts remain
- // available, but cannot become a persisted token-rate denominator.
- expect((done?.data as { usage?: { durationMs?: number } }).usage?.durationMs).toBeUndefined();
- expect(events.findLast((event) => event.type === 'status')?.data.responseSpeed).toMatchObject({ outputTokens: 0, averageRate: null, firstResponseMs: null });
+ expect((done?.data as { usage?: { durationMs?: number } }).usage?.durationMs).toBe(3_000);
if (outcome === 'server-resolved') {
decision.resolve({ kind: 'ask_user_question', answers: {} });
diff --git a/packages/maker-core/src/agents/codex/index.ts b/packages/maker-core/src/agents/codex/index.ts
index 93b2a63e11..b3281c9911 100644
--- a/packages/maker-core/src/agents/codex/index.ts
+++ b/packages/maker-core/src/agents/codex/index.ts
@@ -131,7 +131,6 @@ import {
} from '../shared/auto-review-decision.js';
import type { ReviewableAction } from '../shared/auto-review.js';
import { UsageTracker } from '../shared/usage-tracker.js';
-import { calibratedResponseDuration } from '@cindy/maker-shared/usage-format';
import { attachLiveGeneration, sampleGenerationDuration } from '../shared/live-generation-snapshot.js';
import { getDefaultImageResizer } from '../shared/image-resizer.js';
import { formatManagedImageReferences } from '../shared/managed-image-reference.js';
@@ -172,8 +171,8 @@ import {
classifyCodexError,
translateErrorNotification,
translateItemNotification,
- observeCodexItemTiming,
beginCodexGenerationTurn,
+ codexGenerationDurationMs,
finalizeCodexGenerationTurn,
pauseCodexGeneration,
resetCodexGenerationTiming,
@@ -3709,7 +3708,6 @@ assertRouteCurrent();
const usageTracker = new UsageTracker();
const translatorRt: CodexRuntimeState = newCodexRuntimeState();
const liveUsageSnapshot = () => attachLiveGeneration(usageTracker.snapshot(), {
- responseSpeed: translatorRt.responseSpeed.snapshot(),
outputTokens: usageTracker.getTurnUsage().output,
durationMs: translatorRt.generationOutputDurationMs,
openStartedAt: translatorRt.generationStartedAt,
@@ -10825,7 +10823,6 @@ assertRouteCurrent();
text?: unknown;
} | null | undefined;
if (!candidate || candidate.type !== 'plan') return false;
- translatorRt.responseSpeedHasUnobservedOutput = true;
if (typeof candidate.text === 'string') proposedPlanText = candidate.text;
return true;
}
@@ -10997,16 +10994,7 @@ assertRouteCurrent();
const realTurnUsage = usageTracker.getTurnUsage();
const realTurnUsageSegments = usageTracker.getTurnUsageSegments();
finalizeCodexGenerationTurn(translatorRt, turn.id);
- const unmeasuredOutputOnly = translatorRt.responseSpeedHasUnobservedOutput
- && translatorRt.responseSpeed.snapshot().outputTokens === 0;
- // With no sampled units, retain the real total but expose no rate. Mixed
- // turns retain only their observed estimates, never a whole-turn scale.
- translatorRt.responseSpeed.finish(realTurnUsageSegments.length > 0
- && (!translatorRt.responseSpeedHasUnobservedOutput || unmeasuredOutputOnly)
- ? realTurnUsage.output : undefined);
- if (!translatorRt.generationTimingReliable || unmeasuredOutputOnly) translatorRt.responseSpeed.invalidate();
- const speedSnapshot = translatorRt.responseSpeed.snapshot();
- const generationDurationMs = calibratedResponseDuration(speedSnapshot, realTurnUsage.output);
+ const generationDurationMs = codexGenerationDurationMs(translatorRt);
const codexDoneUsage = {
promptTokens: realTurnUsage.input,
completionTokens: realTurnUsage.output,
@@ -11017,9 +11005,12 @@ assertRouteCurrent();
cachedTokens: realTurnUsage.cacheRead,
cacheCreationTokens: realTurnUsage.cacheCreate,
segments: realTurnUsageSegments,
- // Persist TPS timing only when observed stream time pairs with real turn output.
+ // With usage, exclude post-output finalization. Without usage, retain
+ // the measured duration metadata (zero output cannot produce a rate).
...(generationDurationMs !== undefined ? {
- durationMs: generationDurationMs,
+ durationMs: realTurnUsage.output > 0
+ ? translatorRt.generationOutputDurationMs || undefined
+ : generationDurationMs,
} : {}),
...(typeof turn.durationMs === 'number' && Number.isFinite(turn.durationMs)
? { turnDurationMs: turn.durationMs }
@@ -11223,7 +11214,6 @@ assertRouteCurrent();
data: {
status: 'Done',
...attachLiveGeneration(endSnap, {
- responseSpeed: translatorRt.responseSpeed.snapshot(),
outputTokens: realTurnUsage.output,
durationMs: translatorRt.generationOutputDurationMs,
openStartedAt: null,
@@ -12418,12 +12408,10 @@ assertRouteCurrent();
params.turnId,
'started',
);
- observeCodexItemTiming(translatorRt, 'started', translatedParams);
pushItemStatus(translatedItem);
translateItemNotification('started', translatedParams, eventQueue, {
rt: translatorRt,
log,
- timingObserved: true,
onCompactBoundary: handleCompactBoundary,
});
// 重放帧后发:translator 刚推的 running 帧不得把已重放出的终态盖回去。
@@ -12558,12 +12546,6 @@ assertRouteCurrent();
log,
onCompactBoundary: handleCompactBoundary,
});
- // Close the observed response after translating any final text that
- // was absent from deltas. A later usage/turn terminal calibrates the
- // counts without charging its delivery lag as model generation.
- if (!isLateCollabTerminal && translatedItem.type === 'agentMessage') {
- translatorRt.responseSpeed.pause();
- }
// A late V1 spawn completion is the spawn tool closing, not necessarily
// the child closing. Reassert a running compact state, or an explicit
// failed/stopped tracker state; a completed replay would only duplicate
@@ -12599,7 +12581,6 @@ assertRouteCurrent();
producedOutputTurnIds.add(params.turnId);
noteRecoveryModelWork(params.turnId);
translateAgentMessageDelta(params, eventQueue, { rt: translatorRt, log });
- maybePushUsageRefresh();
},
turnPlanUpdated: (params) => {
if (enqueueIfBufferedTurn(params.turnId, () => handlers.turnPlanUpdated?.(params))) return;
@@ -12615,7 +12596,6 @@ assertRouteCurrent();
producedOutputTurnIds.add(params.turnId);
noteRecoveryModelWork(params.turnId);
translateReasoningSummaryTextDelta(params, eventQueue, { rt: translatorRt, log });
- maybePushUsageRefresh();
},
reasoningSummaryPartAdded: (params) => {
// thinking 流同样算产出(与非缓冲路径一致)。
@@ -12632,7 +12612,6 @@ assertRouteCurrent();
producedOutputTurnIds.add(params.turnId);
noteRecoveryModelWork(params.turnId);
translateReasoningTextDelta(params, eventQueue, { rt: translatorRt, log });
- maybePushUsageRefresh();
},
accountRateLimitsUpdated: (params) =>
translateAccountRateLimitsUpdated(params, eventQueue, { rt: translatorRt, log }),
diff --git a/packages/maker-core/src/agents/codex/translator.ts b/packages/maker-core/src/agents/codex/translator.ts
index 7a22a7cd0e..9f6739c6d8 100644
--- a/packages/maker-core/src/agents/codex/translator.ts
+++ b/packages/maker-core/src/agents/codex/translator.ts
@@ -1,4 +1,3 @@
-import { ResponseSpeedTracker } from '../shared/response-speed.js';
/**
* Codex app-server v2 Notification → maker-core 语义 AgentEvent 翻译器。
*
@@ -81,9 +80,6 @@ export interface CodexRuntimeState {
/** item.id → agentMessage 已 emit 文本字符长度(citation 归一化后的空间,见 handleAgentMessage)。 */
itemTextLen: Map<string, number>;
/** Current model-active interval start; null while tools/approvals own the turn. */
- responseSpeed: ResponseSpeedTracker;
- /** Tool inputs arrive after generation, so whole-turn usage cannot calibrate only observed text. */
- responseSpeedHasUnobservedOutput: boolean;
generationStartedAt: number | null;
/** Tool/approval boundaries currently owning the turn. Generation resumes after all finish. */
generationPendingToolIds: Set<string>;
@@ -134,8 +130,6 @@ export function newCodexRuntimeState(): CodexRuntimeState {
reasoningStartedAt: new Map(),
reasoningTextLen: new Map(),
itemTextLen: new Map(),
- responseSpeed: new ResponseSpeedTracker(),
- responseSpeedHasUnobservedOutput: false,
generationStartedAt: null,
generationPendingToolIds: new Set(),
generationDurationMs: 0,
@@ -172,7 +166,6 @@ function sampleCodexGenerationHeartbeat(rt: CodexRuntimeState, now = Date.now())
now - previous > CODEX_GENERATION_HEARTBEAT_MS + CODEX_GENERATION_SUSPEND_GAP_MS
) {
rt.generationTimingReliable = false;
- rt.responseSpeed.invalidate();
}
rt.generationHeartbeatAt = now;
}
@@ -196,7 +189,6 @@ export function resetCodexGenerationTiming(rt: CodexRuntimeState): void {
rt.generationOutputDurationMs = 0;
rt.generationTurnId = null;
rt.generationTimingReliable = true;
- rt.responseSpeedHasUnobservedOutput = false;
}
function closeCodexGenerationInterval(rt: CodexRuntimeState, endedAt: number): void {
@@ -220,7 +212,6 @@ export function beginCodexGenerationTurn(
if (rt.generationTurnId !== turnId) {
resetCodexGenerationTiming(rt);
rt.generationTurnId = turnId;
- rt.responseSpeed.reset('turn', startedAt);
}
if (rt.generationStartedAt === null && rt.generationPendingToolIds.size === 0) {
rt.generationStartedAt = startedAt;
@@ -266,7 +257,6 @@ export function pauseCodexGeneration(
return;
}
if (rt.generationPendingToolIds.has(pauseId)) return;
- rt.responseSpeed.pause(pausedAt);
if (rt.generationPendingToolIds.size === 0) closeCodexGenerationInterval(rt, pausedAt);
rt.generationPendingToolIds.add(pauseId);
}
@@ -282,12 +272,6 @@ export function resumeCodexGeneration(
return;
}
if (rt.generationPendingToolIds.size === 0) rt.generationStartedAt = resumedAt;
- // Completion of the last tool/approval is a locally observed waiting
- // boundary, not evidence that the next model response has started. Preserve
- // a parallel response that already resumed on an actual output delta.
- if (rt.generationPendingToolIds.size === 0 && rt.responseSpeed.snapshot(resumedAt).phase === 'paused') {
- rt.responseSpeed.beginRequest(resumedAt);
- }
if (rt.generationPendingToolIds.size === 0) startCodexGenerationHeartbeat(rt);
}
@@ -340,12 +324,10 @@ function noteCodexGenerationBoundary(
}
const pauseId = `item:${item.id}`;
if (phase === 'started') {
- if (item.type !== 'contextCompaction') rt.responseSpeed.toolStarted(pauseId);
pauseCodexGeneration(rt, turnId, pauseId, receivedAt);
return;
}
if (phase !== 'completed') return;
- rt.responseSpeed.toolEnded(pauseId);
resumeCodexGeneration(rt, turnId, pauseId, receivedAt);
}
@@ -360,8 +342,6 @@ interface TranslatorLog {
export interface CodexTranslateContext {
rt: CodexRuntimeState;
log: TranslatorLog;
- /** Host observed the boundary before publishing its status frame. */
- timingObserved?: boolean;
/**
* Maker Memory flush 观察器 — codex contextCompaction completed 时调,
* controller 重置 fired 阈值 (compact 后 context 又有空间, 可重新触发)。
@@ -393,30 +373,6 @@ export interface CodexTranslateContext {
type ItemPhase = 'started' | 'updated' | 'completed';
-/** Observe before status publication without changing the established event order. */
-export function observeCodexItemTiming(
- rt: CodexRuntimeState,
- phase: ItemPhase,
- notification: ItemStartedNotification['params'] | ItemUpdatedNotification['params'] | ItemCompletedNotification['params'],
-): void {
- const item = notification.item;
- const itemType = item?.type;
- if (!item || typeof item !== 'object') return;
- if (phase === 'started' && CODEX_GENERATION_PAUSE_ITEM_TYPES.has(itemType)
- && rt.generationPendingToolIds.has(`item:${item.id}`)) return;
- if (phase === 'started' && (itemType === 'reasoning' || CODEX_GENERATION_PAUSE_ITEM_TYPES.has(itemType))) {
- rt.responseSpeed.content();
- }
- noteCodexGenerationBoundary(rt, phase, item, notification);
- // These inputs have no argument delta/start pair. Their output tokens may be
- // included in turn usage, but their generation time is not in our samples.
- // Retain measured streaming estimates without scaling them by that total.
- if ((CODEX_GENERATION_PAUSE_ITEM_TYPES.has(itemType) && itemType !== 'contextCompaction')
- || itemType === 'fileChange' || itemType === 'plan') {
- rt.responseSpeedHasUnobservedOutput = true;
- }
-}
-
export function translateItemNotification(
phase: ItemPhase,
notification:
@@ -436,7 +392,12 @@ export function translateItemNotification(
ctx.log.warn('item missing type field', { phase, itemKeys: Object.keys(item) });
return;
}
- if (!ctx.timingObserved) observeCodexItemTiming(ctx.rt, phase, notification);
+ noteCodexGenerationBoundary(
+ ctx.rt,
+ phase,
+ item as { id?: unknown; type?: unknown },
+ notification as { turnId?: unknown },
+ );
switch (itemType) {
case 'agentMessage':
@@ -1145,7 +1106,6 @@ function emitAgentMessageProgress(
ctx.rt.itemRawText.set(itemId, rawText);
ctx.rt.itemTextLen.set(itemId, emitted.length);
if (delta.length === 0) return;
- ctx.rt.responseSpeed.delta(delta);
queue.push({
type: 'text',
data: { text: delta, isFinal: false, agentMessageId: itemId,
@@ -1361,7 +1321,6 @@ export function translateReasoningSummaryTextDelta(
ctx: CodexTranslateContext,
): void {
if (!params.delta) return;
- ctx.rt.responseSpeed.delta(params.delta);
ensureReasoningStarted(params.itemId, queue, ctx);
const prevLen = ctx.rt.reasoningTextLen.get(params.itemId) ?? 0;
ctx.rt.reasoningTextLen.set(params.itemId, prevLen + params.delta.length);
@@ -1396,7 +1355,6 @@ export function translateReasoningTextDelta(
ctx: CodexTranslateContext,
): void {
if (!params.delta) return;
- ctx.rt.responseSpeed.delta(params.delta);
ensureReasoningStarted(params.itemId, queue, ctx);
const prevLen = ctx.rt.reasoningTextLen.get(params.itemId) ?? 0;
ctx.rt.reasoningTextLen.set(params.itemId, prevLen + params.delta.length);
diff --git a/packages/maker-core/src/agents/pi/__tests__/pi-translator.test.ts b/packages/maker-core/src/agents/pi/__tests__/pi-translator.test.ts
index 2babc441a6..b56625510c 100644
--- a/packages/maker-core/src/agents/pi/__tests__/pi-translator.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/pi-translator.test.ts
@@ -1427,8 +1427,7 @@ describe('pi translator', () => {
cacheCreateTokens: 3,
}),
]);
- // Completed-only frames cannot prove when streaming began.
- expect(usage.durationMs).toBeUndefined();
+ expect(usage.durationMs).toBeGreaterThanOrEqual(1_200);
expect(usage.turnDurationMs).toBeGreaterThanOrEqual(0);
// 快照累计 input+output。
expect(usageSnapshotOf(ctx).tokenUsage).toBe(120);
@@ -1578,7 +1577,7 @@ describe('pi translator', () => {
disposePiTranslateContext(ctx);
});
- it('keeps native timestamp timing but omits TPS without a first-content observation', () => {
+ it('reads Pi v0.83 generation duration from timestamp with a live heartbeat', () => {
const ctx = createPiTranslateContext(noopLogger);
const { queue, events } = makeQueue();
const timestamp = Date.now() - 1_200;
@@ -1599,8 +1598,8 @@ describe('pi translator', () => {
);
translatePiEvent(ev({ type: 'agent_settled' }), queue, ctx);
const usage = (events.find((e) => e.type === 'done')!.data as { usage: Record<string, unknown> }).usage;
- expect(usage.durationMs).toBeUndefined();
- expect(usageSnapshotOf(ctx).generationDurationMs).toBeGreaterThanOrEqual(1_200);
+ expect(usage.durationMs).toEqual(expect.any(Number));
+ expect(usage.durationMs).toBeGreaterThanOrEqual(1_200);
expect(usage.turnDurationMs).toEqual(expect.any(Number));
});
@@ -2240,90 +2239,3 @@ describe('pi translator', () => {
});
});
});
-
-describe('Pi tool-end response speed publication', () => {
- function withTurn(run: (harness: {
- send: (now: number, event: Record<string, unknown>) => void;
- latestStatus: () => unknown;
- events: AgentEvent[];
- ctx: ReturnType<typeof createPiTranslateContext>;
- }) => void) {
- const clock = vi.spyOn(Date, 'now').mockReturnValue(1_000);
- const ctx = createPiTranslateContext(noopLogger);
- const { queue, events } = makeQueue();
- const send = (now: number, event: Record<string, unknown>) => {
- clock.mockReturnValue(now);
- translatePiEvent(ev(event), queue, ctx);
- };
- try {
- send(1_000, { type: 'agent_start' });
- send(2_000, { type: 'message_start', message: { role: 'assistant' } });
- send(3_000, { type: 'message_update', assistantMessageEvent: { type: 'text_delta', delta: 'a'.repeat(40) } });
- run({ send, events, ctx, latestStatus: () => events.filter(event => event.type === 'status').at(-1)?.data });
- } finally {
- disposePiTranslateContext(ctx);
- clock.mockRestore();
- }
- }
-
- it('publishes remaining parallel tools and starts waiting immediately after the last result', () => {
- withTurn(({ send, latestStatus, ctx }) => {
- send(5_000, { type: 'message_end', message: { role: 'assistant', content: [], usage: { output: 80 } } });
- send(6_000, { type: 'tool_execution_start', toolCallId: 'a', toolName: 'read', args: {} });
- send(7_000, { type: 'tool_execution_start', toolCallId: 'b', toolName: 'bash', args: {} });
- send(8_000, { type: 'tool_execution_end', toolCallId: 'a', result: 'done' });
- expect(latestStatus()).toMatchObject({ isRunning: true,
- responseSpeed: { phase: 'paused', toolActive: true, recentRate: null } });
- send(9_000, { type: 'tool_execution_end', toolCallId: 'b', result: 'done' });
- expect(latestStatus()).toMatchObject({ isRunning: true,
- responseSpeed: { phase: 'waiting', toolActive: false, waitingMs: 0, recentRate: null,
- durationMs: 2_000, firstResponseMs: 2_000, outputTokens: 80 } });
- send(10_000, { type: 'message_start', message: { role: 'toolResult' } });
- expect(usageSnapshotOf(ctx).responseSpeed).toMatchObject({ phase: 'waiting', waitingMs: 1_000, durationMs: 2_000 });
- });
- });
-
- it('preserves a request already waiting while a parallel tool finishes', () => {
- withTurn(({ send, latestStatus }) => {
- send(4_000, { type: 'message_end', message: { role: 'assistant', content: [] } });
- send(5_000, { type: 'tool_execution_start', toolCallId: 'a', toolName: 'read', args: {} });
- send(6_000, { type: 'message_start', message: { role: 'assistant' } });
- send(9_000, { type: 'tool_execution_end', toolCallId: 'a', result: 'done' });
- expect(latestStatus()).toMatchObject({ responseSpeed: {
- phase: 'waiting', toolActive: false, waitingMs: 3_000, durationMs: 1_000 } });
- });
- });
-
- it('preserves parallel model output instead of reopening its generation clock', () => {
- withTurn(({ send, latestStatus }) => {
- send(3_100, { type: 'tool_execution_start', toolCallId: 'a', toolName: 'read', args: {} });
- send(4_000, { type: 'message_update', assistantMessageEvent: { type: 'text_delta', delta: 'b'.repeat(40) } });
- send(4_500, { type: 'tool_execution_end', toolCallId: 'a', result: 'done' });
- expect(latestStatus()).toMatchObject({ responseSpeed: {
- phase: 'generating', toolActive: false, durationMs: 1_500, recentRate: 10, outputTokens: 20 } });
- });
- });
-
- it('does not publish running status for orphan, duplicate or settled-turn results', () => {
- withTurn(({ send, events, latestStatus }) => {
- const countStatus = () => events.filter(event => event.type === 'status').length;
- let count = countStatus();
- send(4_000, { type: 'tool_execution_end', toolCallId: 'unknown', result: 'done' });
- expect(countStatus()).toBe(count);
- send(5_000, { type: 'message_end', message: { role: 'assistant', content: [] } });
- send(6_000, { type: 'tool_execution_start', toolCallId: 'a', toolName: 'read', args: {} });
- send(7_000, { type: 'tool_execution_end', toolCallId: 'a', result: 'done' });
- count = countStatus();
- send(8_000, { type: 'tool_execution_end', toolCallId: 'a', result: 'done' });
- expect(countStatus()).toBe(count);
- send(9_000, { type: 'tool_execution_start', toolCallId: 'b', toolName: 'read', args: {} });
- send(10_000, { type: 'agent_settled' });
- count = countStatus();
- const terminal = latestStatus();
- send(11_000, { type: 'tool_execution_end', toolCallId: 'b', result: 'done' });
- expect(countStatus()).toBe(count);
- expect(latestStatus()).toEqual(terminal);
- expect(latestStatus()).toMatchObject({ isRunning: false, responseSpeed: { phase: 'complete' } });
- });
- });
-});
diff --git a/packages/maker-core/src/agents/pi/translator.ts b/packages/maker-core/src/agents/pi/translator.ts
index 99a9af11f3..17064889fa 100644
--- a/packages/maker-core/src/agents/pi/translator.ts
+++ b/packages/maker-core/src/agents/pi/translator.ts
@@ -1,5 +1,3 @@
-import { calibratedResponseDuration } from '@cindy/maker-shared/usage-format';
-import { ResponseSpeedTracker } from '../shared/response-speed.js';
/**
* pi RPC 事件 → AgentEvent 翻译层。
*
@@ -162,7 +160,6 @@ export interface PiTranslateContext {
terminalAssistantErrorEmitted: boolean;
/** 整轮 wall-clock 起点;只用于诊断,不参与 TPS。 */
turnWallClockStartedAt: number;
- responseSpeed: ResponseSpeedTracker;
generationDurationMs: number;
/** Open generation interval start; 0 while tools/user waits own the turn. */
generationOpenAt: number;
@@ -227,7 +224,6 @@ export function createPiTranslateContext(logger: Logger): PiTranslateContext {
finalAssistantText: '',
finalAssistantStopReason: null,
turnWallClockStartedAt: 0,
- responseSpeed: new ResponseSpeedTracker(),
generationDurationMs: 0,
generationOpenAt: 0,
generationTimingReliable: true,
@@ -344,7 +340,6 @@ function samplePiGenerationHeartbeat(ctx: PiTranslateContext, now = Date.now()):
PI_GENERATION_HEARTBEAT_MS + PI_GENERATION_SUSPEND_GAP_MS
) {
ctx.generationHeartbeatReliable = false;
- ctx.responseSpeed.invalidate();
}
ctx.generationHeartbeatAt = now;
}
@@ -369,7 +364,6 @@ export function usageSnapshotOf(ctx: PiTranslateContext): UsageSnapshot {
costUsd: ctx.costUsd,
},
{
- responseSpeed: ctx.responseSpeed.snapshot(),
outputTokens: ctx.turnOutput,
// Pi reports output only at message_end, alongside the closed duration.
durationMs: ctx.generationDurationMs,
@@ -690,7 +684,6 @@ export function translatePiEvent(
ctx.pendingAssistantError = null;
ctx.terminalAssistantErrorEmitted = false;
ctx.turnWallClockStartedAt = Date.now();
- ctx.responseSpeed.reset('turn');
ctx.generationDurationMs = 0;
ctx.generationTimingReliable = true;
stopPiGenerationHeartbeat(ctx);
@@ -710,9 +703,6 @@ export function translatePiEvent(
return;
case 'message_start': {
- const role = (event.message as { role?: string } | undefined)?.role;
- if (role && role !== 'assistant') return;
- ctx.responseSpeed.beginRequest();
ctx.thinkingBlocks.clear();
ctx.streamStopTokenByIndex.clear();
const message = event.message as { usage?: PiUsage } | undefined;
@@ -727,10 +717,6 @@ export function translatePiEvent(
case 'message_update': {
const delta = event.assistantMessageEvent as Record<string, unknown> | undefined;
if (!delta || typeof delta.type !== 'string') return;
- if (['text_start', 'thinking_start', 'toolcall_start'].includes(delta.type)) ctx.responseSpeed.content();
- if (['text_delta', 'thinking_delta', 'toolcall_delta'].includes(delta.type) && typeof delta.delta === 'string') {
- if (ctx.responseSpeed.delta(delta.delta)) pushStatus(queue, ctx, 'Generating...', true);
- }
handleAssistantDelta(delta, queue, ctx);
return;
}
@@ -738,7 +724,6 @@ export function translatePiEvent(
case 'message_end': {
const message = event.message as PiAssistantMessage | undefined;
if (!message || message.role !== 'assistant') return;
- ctx.responseSpeed.endResponse(message.usage?.output);
const pendingPriceVariant = ctx.pendingPriceVariants.shift();
const reportedPriceVariant = priceVariantFromServiceTier(message.usage?.service_tier);
const executionPriceVariant = message.usage && ctx.resolveUsagePriceVariant?.({
@@ -811,7 +796,6 @@ export function translatePiEvent(
case 'tool_execution_start': {
const toolUseId = String(event.toolCallId ?? '');
- ctx.responseSpeed.toolStarted(toolUseId);
const toolName = String(event.toolName ?? 'tool');
const toolArgs = (event.args as Record<string, unknown>) ?? {};
if (toolUseId) ctx.toolNamesByCallId.set(toolUseId, toolName);
@@ -905,7 +889,6 @@ export function translatePiEvent(
case 'tool_execution_end': {
const toolUseId = String(event.toolCallId ?? '');
- const endedTrackedTool = ctx.responseSpeed.toolEnded(toolUseId);
const isError = event.isError === true;
const rawText = toolResultFullText(event.result);
const toolName = String(
@@ -955,13 +938,6 @@ export function translatePiEvent(
source: 'pi',
});
}
- if (endedTrackedTool && ctx.isStreaming) {
- const speed = ctx.responseSpeed.snapshot();
- // The last tool result is an observable wait boundary, even before
- // Pi emits the next message_start. Keep parallel output/request clocks.
- if (!speed.toolActive && speed.phase === 'paused') ctx.responseSpeed.beginRequest();
- pushStatus(queue, ctx, 'Working…', true);
- }
return;
}
@@ -980,7 +956,6 @@ export function translatePiEvent(
ctx.turnSettled = true;
ctx.isStreaming = false;
ctx.pendingPriceVariants = [];
- ctx.responseSpeed.finish();
stopPiGenerationHeartbeat(ctx);
const hostAbortRequested = isCurrentTurnHostAbortRequested(ctx);
const pendingAssistantError = hostAbortRequested ? null : ctx.pendingAssistantError;
@@ -989,7 +964,6 @@ export function translatePiEvent(
const outputLimited = !hostAbortRequested
&& !ctx.terminalAssistantErrorEmitted
&& ctx.finalAssistantStopReason === 'length';
- const speedDuration = calibratedResponseDuration(ctx.responseSpeed.snapshot(), ctx.turnOutput);
const usage = {
inputTokens: ctx.turnInput,
outputTokens: ctx.turnOutput,
@@ -997,10 +971,11 @@ export function translatePiEvent(
cacheCreationTokens: ctx.turnCacheWrite,
segments: ctx.turnUsageSegments.map((segment) => ({ ...segment })),
segmentsComplete: ctx.turnUsageSegmentsComplete,
- // Persist observed streaming time only after provider usage calibration.
- // Tools, request waiting and completed-only output have no TPS denominator.
- ...(speedDuration !== undefined
- ? { durationMs: speedDuration }
+ // durationMs is deliberately generation-only. If Pi does not report a
+ // per-assistant generation duration, omit it instead of charging tool
+ // execution / user waits to TPS.
+ ...(ctx.generationTimingReliable && ctx.generationDurationMs > 0
+ ? { durationMs: ctx.generationDurationMs }
: {}),
...(ctx.turnWallClockStartedAt > 0
? { turnDurationMs: Math.max(0, Date.now() - ctx.turnWallClockStartedAt) }
@@ -1069,14 +1044,11 @@ export function translatePiEvent(
}
case 'auto_retry_start': {
- ctx.responseSpeed.endResponse();
// A provider retry may begin without a matching message_end for the
// failed request. Drop any stale latch so the next request samples its
// own tariff at message_start.
ctx.pendingPriceVariants = [];
if (isCurrentTurnHostAbortRequested(ctx)) return;
- if (!ctx.responseSpeed.beginRetry()) return;
- pushStatus(queue, ctx, 'Working…', true);
// `(auto-retry N/M)` 只给过载用:mobile / Telegram 把这个后缀当成「模型服务繁忙」。
// 网络类改走 `Reconnecting... N/M`,未分类 5xx / LiteLLM in-stream 仍静默。
const progress = parsePiAutoRetryProgress(event);
diff --git a/packages/maker-core/src/agents/shared/live-generation-snapshot.ts b/packages/maker-core/src/agents/shared/live-generation-snapshot.ts
index 542f0e1435..be8802c8a0 100644
--- a/packages/maker-core/src/agents/shared/live-generation-snapshot.ts
+++ b/packages/maker-core/src/agents/shared/live-generation-snapshot.ts
@@ -2,7 +2,6 @@ import type { UsageSnapshot } from '../../types/events.js';
export interface LiveGenerationTiming {
/** Turn-cumulative output tokens, including reasoning. */
- responseSpeed?: import("@cindy/maker-shared/usage-format").ResponseSpeedSnapshot;
outputTokens: number;
/** Duration captured with outputTokens; sparse usage must not include later in-flight time. */
durationMs: number;
@@ -27,7 +26,7 @@ export function attachLiveGeneration(
typeof timing.outputTokens === 'number' && Number.isFinite(timing.outputTokens)
? Math.max(0, timing.outputTokens)
: 0;
- const next: UsageSnapshot = { ...snapshot, outputTokens, ...(timing.responseSpeed ? { responseSpeed: timing.responseSpeed } : {}) };
+ const next: UsageSnapshot = { ...snapshot, outputTokens };
if (!timing.reliable) {
return {
...next,
diff --git a/packages/maker-core/src/agents/shared/response-speed-adapters.test.ts b/packages/maker-core/src/agents/shared/response-speed-adapters.test.ts
deleted file mode 100644
index 9ffeeba8f5..0000000000
--- a/packages/maker-core/src/agents/shared/response-speed-adapters.test.ts
+++ /dev/null
@@ -1,158 +0,0 @@
-import { afterEach, expect, it, vi } from 'vitest';
-import type { AgentEvent } from '../../types/events.js';
-import type { AsyncQueue } from './async-queue.js';
-import type { Logger } from '../../interfaces/logger.js';
-import { UsageTracker } from './usage-tracker.js';
-import { newRuntimeState, translateSdkMessage, type TurnState } from '../claude-code/translator.js';
-import { createPiTranslateContext, disposePiTranslateContext, translatePiEvent, usageSnapshotOf } from '../pi/translator.js';
-import type { PiRpcEvent } from '../pi/rpc-client.js';
-import { beginCodexGenerationTurn, newCodexRuntimeState, translateAgentMessageDelta,
- translateItemNotification, pauseCodexGeneration } from '../codex/translator.js';
-
-afterEach(() => vi.restoreAllMocks());
-function harness() {
- const events: AgentEvent[] = [];
- const queue = { push: (event: AgentEvent) => events.push(event) } as unknown as AsyncQueue<AgentEvent>;
- return { events, queue };
-}
-const log = { debug: vi.fn(), info: vi.fn(), warn: vi.fn(), error: vi.fn() };
-
-it('Pi publishes streaming estimates with untouched billing/cache counters, then calibrates two tool rounds', () => {
- const clock = vi.spyOn(Date, 'now').mockReturnValue(1_000);
- const { queue, events } = harness();
- const ctx = createPiTranslateContext(log as unknown as Logger);
- const send = (now: number, event: Record<string, unknown>) => {
- clock.mockReturnValue(now); translatePiEvent(event as unknown as PiRpcEvent, queue, ctx);
- };
- try {
- send(1_000, { type: 'agent_start' });
- send(2_000, { type: 'message_start', message: { role: 'assistant' } });
- send(3_000, { type: 'message_update', assistantMessageEvent: { type: 'text_delta', delta: 'a'.repeat(40) } });
- send(4_000, { type: 'message_update', assistantMessageEvent: { type: 'toolcall_delta', delta: 'b'.repeat(40) } });
- expect(usageSnapshotOf(ctx)).toMatchObject({ tokenUsage: 0, outputTokens: 0,
- responseSpeed: { firstResponseMs: 2_000, estimated: true, outputTokens: 20 } });
- expect(events.some((event) => event.type === 'status' &&
- ((event.data as { responseSpeed?: { recentRate: number } }).responseSpeed?.recentRate ?? 0) > 0)).toBe(true);
- send(5_000, { type: 'message_end', message: { role: 'assistant', content: [],
- usage: { input: 10, output: 100, cacheRead: 20, cacheWrite: 5 }, duration: 3_000 } });
- send(6_000, { type: 'tool_execution_start', toolCallId: 't', toolName: 'bash', args: {} });
- expect(usageSnapshotOf(ctx).responseSpeed).toMatchObject({ phase: 'paused', toolActive: true, recentRate: null });
- send(16_000, { type: 'tool_execution_end', toolCallId: 't', toolName: 'bash', result: { content: [] } });
- expect(events.filter(event => event.type === 'status').at(-1)?.data).toMatchObject({
- responseSpeed: { phase: 'waiting', toolActive: false, waitingMs: 0, durationMs: 2_000 },
- });
- send(17_000, { type: 'message_start', message: { role: 'toolResult' } });
- expect(usageSnapshotOf(ctx).responseSpeed).toMatchObject({ phase: 'waiting', waitingMs: 1_000 });
- send(17_000, { type: 'message_start', message: { role: 'assistant' } });
- send(18_000, { type: 'message_update', assistantMessageEvent: { type: 'text_delta', delta: 'answer' } });
- send(19_000, { type: 'message_end', message: { role: 'assistant', content: [{ type: 'text', text: 'answer' }],
- usage: { input: 10, output: 50 }, duration: 2_000 } });
- send(20_000, { type: 'agent_settled' });
- expect(usageSnapshotOf(ctx).responseSpeed).toMatchObject({ phase: 'complete', estimated: false,
- outputTokens: 150, durationMs: 3_000, averageRate: 50 });
- expect((events.find((event) => event.type === 'done')!.data as { usage: unknown }).usage)
- .toMatchObject({ inputTokens: 20, outputTokens: 150, cacheReadTokens: 20,
- cacheCreationTokens: 5, durationMs: 3_000 });
- send(21_000, { type: 'agent_start' });
- expect(usageSnapshotOf(ctx).responseSpeed).toMatchObject({ firstResponseMs: null, outputTokens: 0, samples: [] });
- } finally { disposePiTranslateContext(ctx); }
-});
-
-it('Claude measures only SDK stream wait, keeps child output separate, and deduplicates message usage', () => {
- const clock = vi.spyOn(Date, 'now').mockReturnValue(1_000);
- const { queue, events } = harness();
- const turn: TurnState = { text: '', toolUses: 0, apiCalls: 0, sawCompactBoundary: false,
- hasEmittedText: false, uiEmittedText: '', pendingApiError: null, interruptRequested: false,
- generation: 0, interruptGeneration: 0, lastAssistantMsgHadSubstance: true };
- const ctx = { rt: newRuntimeState(), turn, log, getModel: () => 'test-model',
- getEffort: () => 'high', getPermissionMode: () => 'auto', onSessionId: vi.fn(),
- getSdkSessionId: () => undefined, getLogTitle: () => undefined, tracker: new UsageTracker() };
- const stream = (now: number, event: Record<string, unknown>, parent_tool_use_id?: string) => {
- clock.mockReturnValue(now); translateSdkMessage({ type: 'stream_event', event, parent_tool_use_id }, queue, ctx);
- };
- stream(1_000, { type: 'message_start', message: { id: 'm1', model: 'test-model' } });
- stream(2_000, { type: 'content_block_start', content_block: { type: 'thinking' } });
- stream(2_000, { type: 'content_block_delta', delta: { type: 'thinking_delta', thinking: 'a'.repeat(40) } });
- stream(3_000, { type: 'content_block_delta', delta: { type: 'input_json_delta', partial_json: 'b'.repeat(40) } });
- const beforeChild = ctx.rt.generation.responseSpeed.snapshot();
- stream(3_000, { type: 'content_block_delta', delta: { type: 'text_delta', text: 'child'.repeat(100) } }, 'child');
- expect(ctx.rt.generation.responseSpeed.snapshot()).toEqual(beforeChild);
- stream(4_000, { type: 'message_delta', usage: { output_tokens: 100 } });
- stream(5_000, { type: 'message_delta', usage: { output_tokens: 100 } });
- clock.mockReturnValue(20_000);
- translateSdkMessage({ type: 'result', subtype: 'success', result: 'done', usage: { output_tokens: 100 } }, queue, ctx);
- const done = events.find((event) => event.type === 'done')!.data as { responseSpeed: unknown };
- expect(done.responseSpeed).toMatchObject({ waitOrigin: 'stream', firstResponseMs: 1_000,
- outputTokens: 100, durationMs: 2_000, estimated: false, averageRate: 50 });
-});
-
-it('Codex counts the shared snapshot/delta cursor once and pauses while tools own the turn', () => {
- const clock = vi.spyOn(Date, 'now').mockReturnValue(1_000);
- const rt = newCodexRuntimeState();
- const { queue } = harness();
- const ctx = { rt, log };
- beginCodexGenerationTurn(rt, 'turn', 1_000);
- clock.mockReturnValue(2_000);
- translateAgentMessageDelta({ threadId: 'thread', turnId: 'turn', itemId: 'msg', delta: 'a'.repeat(40) }, queue, ctx);
- clock.mockReturnValue(3_000);
- translateItemNotification('updated', { threadId: 'thread', turnId: 'turn', item:
- { id: 'msg', type: 'agentMessage', text: 'a'.repeat(40) } }, queue, ctx);
- expect(rt.responseSpeed.snapshot().outputTokens).toBe(10);
- pauseCodexGeneration(rt, 'turn', 'tool', 4_000);
- clock.mockReturnValue(15_000);
- rt.responseSpeed.finish(100);
- expect(rt.responseSpeed.snapshot()).toMatchObject({ waitOrigin: 'turn', firstResponseMs: 1_000,
- durationMs: 2_000, outputTokens: 100, averageRate: 50 });
- // Reset disposes the old generation heartbeat without altering the completed speed snapshot.
- beginCodexGenerationTurn(rt, 'next', 16_000);
- pauseCodexGeneration(rt, 'next', 'cleanup', 16_000);
-});
-
-it('Pi publishes even an unclassified native retry without adding an error banner, then recovers on content', () => {
- const clock = vi.spyOn(Date, 'now').mockReturnValue(1000);
- const { queue, events } = harness();
- const ctx = createPiTranslateContext(log as unknown as Logger);
- try {
- translatePiEvent({ type: 'agent_start' } as PiRpcEvent, queue, ctx);
- clock.mockReturnValue(2000);
- translatePiEvent({ type: 'auto_retry_start', attempt: 1, maxAttempts: 3,
- errorMessage: 'unclassified upstream error' } as PiRpcEvent, queue, ctx);
- expect(events.filter(event => event.type === 'error')).toHaveLength(0);
- expect(events.filter(event => event.type === 'status').at(-1)?.data).toMatchObject({
- responseSpeed: { retrying: true, recentRate: null } });
- clock.mockReturnValue(12000);
- translatePiEvent({ type: 'message_start', message: { role: 'assistant' } } as PiRpcEvent, queue, ctx);
- translatePiEvent({ type: 'message_update', assistantMessageEvent: { type: 'text_delta', delta: 'recovered' } } as PiRpcEvent, queue, ctx);
- expect(usageSnapshotOf(ctx).responseSpeed).toMatchObject({ retrying: false, durationMs: 0 });
- ctx.responseSpeed.finish();
- const settledEvents = events.length;
- translatePiEvent({ type: 'auto_retry_start', attempt: 2, maxAttempts: 3 } as PiRpcEvent, queue, ctx);
- expect(events).toHaveLength(settledEvents);
- } finally { disposePiTranslateContext(ctx); }
-});
-
-it('Claude publishes its first native retry without an error banner and excludes backoff from generation', () => {
- const clock = vi.spyOn(Date, 'now').mockReturnValue(1000);
- const { queue, events } = harness();
- const turn: TurnState = { text: '', toolUses: 0, apiCalls: 0, sawCompactBoundary: false,
- hasEmittedText: false, uiEmittedText: '', pendingApiError: null, interruptRequested: false,
- generation: 0, interruptGeneration: 0, lastAssistantMsgHadSubstance: true };
- const ctx = { rt: newRuntimeState(), turn, log, getModel: () => 'test-model',
- getEffort: () => 'high', getPermissionMode: () => 'auto', onSessionId: vi.fn(),
- getSdkSessionId: () => undefined, getLogTitle: () => undefined, tracker: new UsageTracker() };
- translateSdkMessage({ type: 'system', subtype: 'api_retry', attempt: 1, max_retries: 3,
- retry_delay_ms: 10000, error_status: 500, error: 'api_error' }, queue, ctx);
- expect(events.filter(event => event.type === 'error')).toHaveLength(0);
- expect(events.filter(event => event.type === 'status').at(-1)?.data).toMatchObject({
- responseSpeed: { retrying: true, recentRate: null } });
- clock.mockReturnValue(12000);
- translateSdkMessage({ type: 'stream_event', event: { type: 'message_start', message: { id: 'recovered' } } }, queue, ctx);
- translateSdkMessage({ type: 'stream_event', event: { type: 'content_block_start', content_block: { type: 'text' } } }, queue, ctx);
- expect(ctx.rt.generation.responseSpeed.snapshot().retrying).not.toBe(true);
- expect(ctx.rt.generation.responseSpeed.snapshot().durationMs).toBe(0);
- ctx.turn.interruptRequested = true;
- const stoppedEvents = events.length;
- translateSdkMessage({ type: 'system', subtype: 'api_retry', attempt: 2, max_retries: 3,
- retry_delay_ms: 1000, error_status: 529, error: 'overloaded_error' }, queue, ctx);
- expect(events).toHaveLength(stoppedEvents);
-});
diff --git a/packages/maker-core/src/agents/shared/response-speed.test.ts b/packages/maker-core/src/agents/shared/response-speed.test.ts
deleted file mode 100644
index c1bea31587..0000000000
--- a/packages/maker-core/src/agents/shared/response-speed.test.ts
+++ /dev/null
@@ -1,237 +0,0 @@
-import { describe, expect, it } from 'vitest';
-import { retryResponseSpeed, resumeResponseSpeed, mergeResponseSpeedStatus, calibratedResponseDuration, readResponseSpeedSnapshot, stopResponseSpeed, responseSpeedHistory, responseSpeedActivity } from '@cindy/maker-shared/usage-format';
-import { ResponseSpeedTracker } from './response-speed.js';
-
-describe('response speed', () => {
- it('distinguishes observed output, silence, execution and subsequent request waiting', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 1000);
- expect(responseSpeedActivity(speed.snapshot(2000), 2000)).toBe('waiting');
- speed.delta('tool arguments', 2000);
- expect(speed.snapshot(2000).toolActive).toBe(false);
- expect(responseSpeedActivity(speed.snapshot(2000), 2000)).toBe('generating');
- expect(responseSpeedActivity(speed.snapshot(3500), 3500)).toBe('quiet');
- speed.endResponse(20, 4000);
- expect(responseSpeedActivity(speed.snapshot(4000), 4000)).toBe('paused');
- speed.toolStarted('command');
- expect(responseSpeedActivity(speed.snapshot(5000), 5000)).toBe('tool');
- // A concurrent model delta remains model generation, even with a tool active.
- speed.delta('parallel output', 6000);
- expect(responseSpeedActivity(speed.snapshot(6000), 6000)).toBe('generating');
- speed.toolEnded('command');
- speed.beginRequest(7000);
- expect(speed.snapshot(8500)).toMatchObject({ phase: 'waiting', waitingMs: 1500, firstResponseMs: 1000 });
- speed.finish(undefined, 9000);
- expect(responseSpeedActivity(speed.snapshot(10000), 10000)).toBe('complete');
- });
- it('separates first content wait from streaming time and calibrates the curve', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- expect(speed.snapshot(2_000)).toMatchObject({ phase: 'waiting', waitingMs: 2_000, recentRate: null });
- speed.content(2_000);
- speed.delta('a'.repeat(40), 2_000);
- speed.delta('b'.repeat(40), 3_000);
- expect(speed.snapshot(3_000)).toMatchObject({ firstResponseMs: 2_000, estimated: true, outputTokens: 20 });
- expect(speed.snapshot(3_000).samples.length).toBeGreaterThan(0);
- speed.endResponse(100, 4_000);
- speed.finish(undefined, 20_000);
- const done = speed.snapshot(20_000);
- expect(done).toMatchObject({ durationMs: 2_000, outputTokens: 100, estimated: false, averageRate: 50, recentRate: 50 });
- expect(done.samples.at(-1)?.rate).toBe(50);
- expect(calibratedResponseDuration(done, 100)).toBe(2_000);
- expect(calibratedResponseDuration(done, 101)).toBeUndefined();
- });
-
- it('updates a long response before any usage arrives and decays a silent window', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- speed.content(1_000);
- for (let now = 1_000; now <= 8_000; now += 250) speed.delta('abcdefgh', now);
- expect(speed.snapshot(8_000).samples.length).toBeGreaterThan(20);
- expect(speed.snapshot(8_000)).toMatchObject({ estimated: true, phase: 'generating' });
- expect(speed.snapshot(9_500).recentRate).toBeNull();
- });
-
- it('excludes tools and the next request wait, and calibrates each response separately', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- speed.content(1_000);
- speed.delta('a'.repeat(40), 1_000);
- speed.endResponse(40, 2_000);
- speed.beginRequest(10_000);
- speed.content(12_000);
- speed.delta('a'.repeat(40), 12_000);
- speed.endResponse(60, 13_000);
- speed.finish(undefined, 30_000);
- expect(speed.snapshot(30_000)).toMatchObject({ firstResponseMs: 1_000, durationMs: 2_000,
- outputTokens: 100, estimated: false, averageRate: 50 });
- });
-
- it('does not double-add repeated per-response usage and handles a real zero', () => {
- const speed = new ResponseSpeedTracker('stream');
- speed.beginRequest(1_000);
- speed.delta('a'.repeat(40), 2_000);
- speed.reportOutput(50);
- speed.reportOutput(50);
- speed.pause(3_000);
- speed.beginRequest(4_000);
- speed.delta('a'.repeat(40), 5_000);
- speed.endResponse(0, 6_000);
- speed.finish(undefined, 7_000);
- expect(speed.snapshot(7_000)).toMatchObject({ waitOrigin: 'stream', firstResponseMs: 1_000,
- outputTokens: 50, durationMs: 2_000, estimated: false, averageRate: 25 });
- });
-
- it('keeps missing usage approximate and suppresses completed-only or untimed output', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- speed.delta('a'.repeat(40), 1_000);
- speed.finish(undefined, 2_000);
- expect(speed.snapshot(2_000)).toMatchObject({ estimated: true, outputTokens: 10, averageRate: 10 });
- expect(calibratedResponseDuration(speed.snapshot(2_000), 10)).toBeUndefined();
- speed.reset('turn', 3_000);
- speed.endResponse(100, 4_000);
- speed.finish(undefined, 5_000);
- expect(speed.snapshot(5_000)).toMatchObject({ firstResponseMs: null, averageRate: null, samples: [] });
- });
-
- it('cumulative turn usage calibrates Codex without counting hidden reasoning twice', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- speed.delta('a'.repeat(40), 1_000);
- speed.pause(2_000);
- speed.delta('b'.repeat(40), 10_000);
- speed.finish(200, 11_000);
- expect(speed.snapshot(20_000)).toMatchObject({ outputTokens: 200, durationMs: 2_000, averageRate: 100 });
- speed.finish(300, 30_000);
- expect(speed.snapshot(30_000).outputTokens).toBe(200);
- });
-
- it('does not manufacture a current rate from a short burst, silence or a tool gap', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- speed.delta('a'.repeat(4000), 1000);
- expect(speed.snapshot(1250).recentRate).toBeNull();
- speed.delta('a'.repeat(40), 2000);
- const observed = speed.snapshot(2000);
- expect(observed.recentRate).toBe(10);
- const local = readResponseSpeedSnapshot(observed, 100000)!;
- expect(responseSpeedHistory(local, 100999).latestRate).toBe(10);
- expect(responseSpeedHistory(local, 101000).latestRate).toBeNull();
- expect(local.firstResponseMs).toBe(observed.firstResponseMs);
- expect(speed.snapshot(5000).recentRate).toBeNull();
- speed.pause(5000);
- speed.delta('resume', 20000);
- expect(speed.snapshot(20000).recentRate).toBeNull();
- speed.finish(1100, 21000);
- expect(speed.snapshot(21000)).toMatchObject({ durationMs: 5000, averageRate: 220 });
- });
-
- it('reset clears cancelled/retried turns and ignores late deltas after completion', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- speed.delta('a'.repeat(40), 1_000);
- speed.finish(undefined, 2_000);
- speed.delta('late', 3_000);
- speed.reset('turn', 4_000);
- expect(speed.snapshot(4_500)).toMatchObject({ phase: 'waiting', firstResponseMs: null,
- waitingMs: 500, outputTokens: 0, samples: [] });
- speed.content(5_000);
- speed.finish(0, 6_000);
- expect(speed.snapshot(6_000)).toMatchObject({ outputTokens: 0, estimated: false, averageRate: 0 });
- });
-
- it('expires high-frequency output instead of renewing accumulated units', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- speed.content(0);
- for (let now = 1; now <= 10_000; now++) {
- speed.delta('abcd', now); // 1 estimated token, 1000 deltas per second.
- if (now === 1_000 || now === 5_000 || now === 10_000) {
- expect(speed.snapshot(now).recentRate).toBeGreaterThanOrEqual(996);
- expect(speed.snapshot(now).recentRate).toBeLessThanOrEqual(1000);
- }
- }
- expect(speed.snapshot(10_000).samples.every(sample => sample.rate <= 1000)).toBe(true);
- expect(responseSpeedHistory(speed.snapshot(10_000), 10_000).peak).toBeLessThanOrEqual(1000);
- expect(speed.snapshot(11_000).recentRate).toBeNull();
- speed.pause(11_000);
- for (let now = 20_000; now <= 21_000; now++) speed.delta('abcd', now);
- expect(speed.snapshot(21_000).recentRate).toBeLessThanOrEqual(1000);
- expect(speed.snapshot(22_000).recentRate).toBeNull();
- speed.finish(11_001, 22_000);
- const done = speed.snapshot(22_000);
- expect(done).toMatchObject({ estimated: false, outputTokens: 11_001,
- durationMs: 13_000, averageRate: 11_001_000 / 13_000 });
- expect(responseSpeedHistory(done, 22_000).peak).toBeLessThanOrEqual(1000);
- });
-
- it('character estimates are independent of chunk boundaries and samples stay bounded', () => {
- const measure = (chunks: string[]) => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- for (const chunk of chunks) speed.delta(chunk, 1_000);
- return speed.snapshot(2_000).outputTokens;
- };
- expect(measure(['hello 世界 👋'])).toBe(measure(['h', 'ello ', '世', '界 ', '👋']));
- expect(measure(['👋'.repeat(10)])).toBe(measure('👋'.repeat(10).split('')));
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- for (let i = 1; i < 500; i++) speed.delta('a', i * 250);
- expect(speed.snapshot(125_000).samples.length).toBeLessThanOrEqual(60);
- speed.invalidate();
- expect(speed.snapshot(125_000)).toMatchObject({ averageRate: null, recentRate: null, samples: [] });
- });
-
- it('freezes an interrupted observation without claiming provider calibration', () => {
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 0);
- speed.delta('a'.repeat(40), 1_000);
- const stopped = stopResponseSpeed(speed.snapshot(2_000));
- expect(stopped).toMatchObject({ phase: 'complete', estimated: true, durationMs: 1_000 });
- expect(calibratedResponseDuration(stopped, 10)).toBeUndefined();
- expect(stopResponseSpeed(stopped)).toBe(stopped);
- });
-
- it('rejects malformed remote snapshots and supports older hosts omitting the field', () => {
- expect(readResponseSpeedSnapshot(undefined)).toBeUndefined();
- expect(readResponseSpeedSnapshot({ phase: 'generating', outputTokens: Infinity })).toBeUndefined();
- const speed = new ResponseSpeedTracker();
- speed.reset('turn', 1_000);
- expect(readResponseSpeedSnapshot(speed.snapshot(2_000))).toEqual(speed.snapshot(2_000));
- });
-});
-
-it('retains incomplete history without showing success or a live rate on failure, cancellation or retry', () => {
- const tracker = new ResponseSpeedTracker();
- tracker.reset('turn', 0);
- tracker.delta('observed output', 1000);
- tracker.delta('more observed output', 2000);
- const original = tracker.snapshot(2000);
- for (const outcome of ['failed', 'cancelled'] as const) {
- const stopped = stopResponseSpeed(original, outcome)!;
- expect(responseSpeedActivity(stopped, 3000)).toBe(outcome);
- expect(responseSpeedHistory(stopped).latestRate).toBeNull();
- expect(stopped.samples).toEqual(original.samples);
- expect(stopped.averageRate).toBe(original.averageRate);
- expect(calibratedResponseDuration(stopped, stopped.outputTokens)).toBeUndefined();
- expect(mergeResponseSpeedStatus(stopped, { ...original, phase: 'complete' }, false, false)?.outcome).toBe(outcome);
- expect(mergeResponseSpeedStatus(stopped, undefined, true, true)).toBeUndefined();
- }
- const retry = retryResponseSpeed(original)!;
- expect(responseSpeedActivity(retry)).toBe('retrying');
- expect(responseSpeedHistory(retry).latestRate).toBeNull();
- expect(mergeResponseSpeedStatus(retry, original, true, false)?.retrying).toBe(true);
- expect(mergeResponseSpeedStatus(retry, { ...original, outputTokens: 1000 }, true, false)?.retrying).toBe(true);
- expect(responseSpeedActivity(mergeResponseSpeedStatus(retry, { ...original, phase: 'complete' }, true, false)!)).toBe('retrying');
- expect(mergeResponseSpeedStatus(retry, { ...original, hasRecentOutput: true,
- samples: [...original.samples, { durationMs: original.durationMs + 1000, outputTokens: 100, rate: 40 }] }, true, false)?.retrying).not.toBe(true);
- expect(mergeResponseSpeedStatus(retry, { ...original, retrying: false }, true, false)?.retrying).toBe(false);
- expect(resumeResponseSpeed(retry)?.retrying).toBe(false);
- const parallelToolRetry = retryResponseSpeed({ ...original, phase: 'paused', toolActive: true })!;
- expect(parallelToolRetry.toolActive).toBe(true);
- expect(mergeResponseSpeedStatus(parallelToolRetry, { ...original, phase: 'paused', toolActive: true }, true, false)?.retrying).toBe(true);
- expect(responseSpeedActivity(resumeResponseSpeed(parallelToolRetry)!)).toBe('tool');
- expect(readResponseSpeedSnapshot({ ...original, outcome: 'unknown' })).toBeUndefined();
- expect(readResponseSpeedSnapshot({ ...original, retrying: 'yes' })).toBeUndefined();
-});
diff --git a/packages/maker-core/src/agents/shared/response-speed.ts b/packages/maker-core/src/agents/shared/response-speed.ts
deleted file mode 100644
index f052775a6e..0000000000
--- a/packages/maker-core/src/agents/shared/response-speed.ts
+++ /dev/null
@@ -1,206 +0,0 @@
-import type { ResponseSpeedSnapshot, RateSample } from '@cindy/maker-shared/usage-format';
-
-const WINDOW_MS = 1_000;
-const WINDOW_BUCKET_MS = 4;
-const REFRESH_MS = 250;
-const MAX_SAMPLES = 60;
-
-/**
- * Inspired by pi-token-speed's separate wait/stream clocks and usage reconciliation.
- * Original implementation: estimates characters, never treats a transport chunk as a token.
- * Only display facts leave this object; UsageTracker remains the accounting authority.
- */
-export class ResponseSpeedTracker {
- private origin: 'turn' | 'stream' = 'turn';
- private startedAt: number | null = null;
- private firstAt: number | null = null;
- private waitingAt: number | null = null;
- private tools = new Set<string>();
- private openAt: number | null = null;
- private closedMs = 0;
- private phase: ResponseSpeedSnapshot['phase'] = 'waiting';
- private retrying = false;
- private retryObserved = false;
- private settledTokens = 0;
- private units = 0;
- private reported: number | null = null;
- private allCalibrated = true;
- private measurable = true;
- private segment = 0;
- private lastSampleAt = 0;
- private lastDeltaAt: number | null = null;
- private window: { time: number; units: number }[] = [];
- private points: (RateSample & { segment: number; units: number })[] = [];
-
- constructor(origin: 'turn' | 'stream' = 'turn') { this.origin = origin; }
-
- reset(origin: 'turn' | 'stream', now = Date.now()): void {
- this.origin = origin;
- this.startedAt = now;
- this.waitingAt = now;
- this.tools.clear();
- this.firstAt = this.openAt = null;
- this.closedMs = this.settledTokens = this.units = 0;
- this.reported = null;
- this.allCalibrated = this.measurable = true;
- this.phase = 'waiting';
- this.retrying = this.retryObserved = false;
- this.segment = 0;
- this.lastSampleAt = 0;
- this.lastDeltaAt = null;
- this.window = [];
- this.points = [];
- }
-
- /** A new request starts waiting again; tools and prompt processing are not generation. */
- beginRequest(now = Date.now()): void {
- if (this.startedAt === null || this.phase === 'complete') this.reset(this.origin, now);
- this.endResponse(undefined, now);
- if (this.origin === 'stream' && this.firstAt === null) this.startedAt = now;
- this.phase = 'waiting';
- this.waitingAt = now;
- }
-
- content(now = Date.now()): void {
- if (this.startedAt === null || this.phase === 'complete') return;
- if (this.firstAt === null) this.firstAt = now;
- if (this.openAt === null) {
- this.openAt = now;
- this.window = [];
- }
- this.phase = 'generating';
- this.retrying = false;
- this.waitingAt = null;
- }
-
- delta(text: string, now = Date.now()): boolean {
- if (!text || this.startedAt === null || this.phase === 'complete') return false;
- this.content(now);
- // Additive across arbitrary chunks; no regex word-boundary or per-delta rounding.
- let units = 0;
- for (const character of text) units += character.charCodeAt(0) > 0xff ? 0.67 * character.length : 0.25;
- this.units += units;
- this.reported = null;
- this.lastDeltaAt = now;
- // Fixed bucket boundaries never move forward as new units arrive. At most
- // four milliseconds of the oldest edge expire early; old output cannot
- // keep renewing itself in a high-frequency stream.
- const bucketAt = Math.floor(now / WINDOW_BUCKET_MS) * WINDOW_BUCKET_MS;
- const bucket = this.window.at(-1);
- if (bucket?.time === bucketAt) bucket.units += units;
- else this.window.push({ time: bucketAt, units });
- this.prune(now);
- if (now - this.lastSampleAt < REFRESH_MS) return false;
- this.lastSampleAt = now;
- const rate = this.recent(now);
- if (rate !== null) {
- this.points.push({ durationMs: this.duration(now), outputTokens: this.settledTokens + this.units,
- rate, segment: this.segment, units: this.units });
- this.points = this.points.slice(-MAX_SAMPLES);
- }
- return true;
- }
-
- pause(now = Date.now()): void {
- if (this.openAt !== null) this.closedMs += Math.max(0, now - this.openAt);
- this.openAt = null;
- if (this.phase !== 'complete') this.phase = 'paused';
- this.window = [];
- this.waitingAt = null;
- }
-
- /** A native retry event proves automatic recovery, even when no error banner is emitted. */
- beginRetry(now = Date.now()): boolean {
- if (this.startedAt === null) this.reset(this.origin, now);
- if (this.phase === 'complete') return false;
- this.pause(now);
- this.retrying = this.retryObserved = true;
- return true;
- }
-
- toolStarted(id: string): void { if (id) this.tools.add(id); }
- toolEnded(id: string): boolean { return this.tools.delete(id); }
-
- reportOutput(output: number | undefined): void {
- if (output !== undefined && Number.isFinite(output) && output >= 0) this.reported = output;
- }
-
- /** Provider counts for this assistant response, including a legitimate zero. */
- endResponse(output?: number, now = Date.now()): void {
- this.pause(now);
- this.reportOutput(output);
- const count = this.reported ?? this.units;
- if (count > 0 && this.units === 0) this.measurable = false;
- if (this.units > 0) {
- const factor = count / this.units;
- this.points = this.points.map((point) => point.segment === this.segment
- ? { ...point, rate: point.rate * factor,
- outputTokens: this.settledTokens + point.units * factor } : point);
- }
- if (this.reported === null && this.units > 0) this.allCalibrated = false;
- this.settledTokens += count;
- this.units = 0;
- this.reported = null;
- this.segment += 1;
- }
-
- /** Codex exposes cumulative turn usage, rather than a stable request-end pair. */
- finish(total?: number, now = Date.now()): void {
- if (this.phase === 'complete') return;
- this.endResponse(undefined, now);
- if (total !== undefined && Number.isFinite(total) && total >= 0) {
- if (this.settledTokens > 0) {
- const factor = total / this.settledTokens;
- this.points = this.points.map((point) => ({ ...point,
- rate: point.rate * factor, outputTokens: point.outputTokens * factor }));
- } else if (total > 0) this.measurable = false;
- this.settledTokens = total;
- this.allCalibrated = true;
- }
- this.phase = 'complete';
- this.retrying = false;
- }
-
- invalidate(): void { this.measurable = false; }
-
- snapshot(now = Date.now()): ResponseSpeedSnapshot {
- this.prune(now);
- const outputTokens = Math.round(this.settledTokens + (this.reported ?? this.units));
- const durationMs = this.duration(now);
- const averageRate = this.measurable && durationMs > 0
- ? outputTokens * 1_000 / durationMs : null;
- return {
- phase: this.phase, waitOrigin: this.origin,
- firstResponseMs: this.startedAt !== null && this.firstAt !== null
- ? Math.max(0, this.firstAt - this.startedAt) : null,
- waitingMs: this.phase === 'waiting' && this.waitingAt !== null
- ? Math.max(0, now - this.waitingAt) : 0,
- outputTokens, durationMs,
- estimated: !this.allCalibrated || (this.units > 0 && this.reported === null),
- recentRate: this.phase === 'complete' ? averageRate : this.recent(now),
- averageRate,
- samples: this.measurable ? this.points.map(({ durationMs, outputTokens, rate }) =>
- ({ durationMs, outputTokens, rate })) : [],
- sampledAt: now,
- hasRecentOutput: this.phase === 'generating' && this.window.length > 0,
- toolActive: this.tools.size > 0,
- ...(this.retryObserved ? { retrying: this.retrying } : {}),
- };
- }
-
- private duration(now: number): number {
- return this.closedMs + (this.openAt === null ? 0 : Math.max(0, now - this.openAt));
- }
- private prune(now: number): void {
- this.window = this.window.filter((point) => now - point.time < WINDOW_MS);
- // A monotonic one-second window has at most 251 fixed buckets. Bound even
- // malformed/backwards clocks without renewing discarded output's lifetime.
- if (this.window.length > 256) this.window = this.window.slice(-256);
- }
- private recent(now: number): number | null {
- if (!this.measurable || this.openAt === null || this.lastDeltaAt === null) return null;
- const elapsed = Math.min(WINDOW_MS, Math.max(0, now - this.openAt));
- if (elapsed < WINDOW_MS || this.window.length === 0) return null;
- return this.window.reduce((sum, point) => sum + point.units, 0) * 1_000 / elapsed;
- }
-}
diff --git a/packages/maker-core/src/types/events.ts b/packages/maker-core/src/types/events.ts
index eeda14b5c1..d2df179d45 100644
--- a/packages/maker-core/src/types/events.ts
+++ b/packages/maker-core/src/types/events.ts
@@ -406,7 +406,6 @@ export interface InteractionDismissedEvent {
* 注: isRunning / status text 不属于 usage, 由 translator 在 emit status event 时单独拼。
*/
export interface UsageSnapshot {
- responseSpeed?: import("@cindy/maker-shared/usage-format").ResponseSpeedSnapshot;
tokenUsage: number;
contextTokens: number;
contextWindow: number;
diff --git a/packages/maker-shared/src/runningTokenRateHistory.ts b/packages/maker-shared/src/runningTokenRateHistory.ts
index f42e0c54f9..c40497e66e 100644
--- a/packages/maker-shared/src/runningTokenRateHistory.ts
+++ b/packages/maker-shared/src/runningTokenRateHistory.ts
@@ -3,115 +3,6 @@ interface RateCounters {
outputTokens: number;
}
-/** Display-only stream measurement. Never use these counters for billing or context. */
-export interface ResponseSpeedSnapshot {
- phase: 'waiting' | 'generating' | 'paused' | 'complete';
- /** Local provider-turn start, or SDK stream start (not network TTFT). */
- waitOrigin: 'turn' | 'stream';
- firstResponseMs: number | null;
- waitingMs: number;
- outputTokens: number;
- durationMs: number;
- estimated: boolean;
- recentRate: number | null;
- averageRate: number | null;
- samples: RateSample[];
- sampledAt: number;
- /** Fresh observed output, including warm-up before the first full rate window. */
- hasRecentOutput?: boolean;
- /** Explicit native tool execution, distinct from streamed tool arguments. */
- toolActive?: boolean;
- /** Display annotation from an explicit terminal event, never inferred from silence. */
- outcome?: 'failed' | 'cancelled';
- /** True requires an explicit retry event; false can acknowledge native recovery. */
- retrying?: boolean;
-}
-
-export function responseSpeedActivity(speed: ResponseSpeedSnapshot, now = Date.now()):
- 'waiting' | 'generating' | 'quiet' | 'tool' | 'paused' | 'complete' | 'failed' | 'cancelled' | 'retrying' {
- if (speed.phase === 'complete') return speed.outcome ?? 'complete';
- if (speed.retrying) return 'retrying';
- if (speed.phase === 'waiting') return 'waiting';
- if (speed.phase === 'paused') return speed.toolActive ? 'tool' : 'paused';
- return now - speed.sampledAt < RATE_SAMPLE_FRESH_MS &&
- (speed.hasRecentOutput ?? speed.recentRate !== null) ? 'generating' : 'quiet';
-}
-
-export function responseSpeedHistory(speed: ResponseSpeedSnapshot, now = Date.now()): RateHistory {
- return {
- ...emptyRateHistory(null),
- samples: speed.samples,
- peak: Math.max(0, ...speed.samples.map((sample) => sample.rate)),
- latestRate: speed.outcome || speed.retrying || (speed.phase === 'generating' && now - speed.sampledAt >= RATE_SAMPLE_FRESH_MS) ? null : speed.recentRate,
- latestSampleAt: speed.sampledAt,
- };
-}
-
-/** Optional additive wire field: old hosts simply omit it. Reject malformed peer data. */
-export function readResponseSpeedSnapshot(value: unknown, observedAt?: number): ResponseSpeedSnapshot | undefined {
- if (!value || typeof value !== 'object') return undefined;
- const speed = value as ResponseSpeedSnapshot;
- const finite = (n: unknown): n is number => typeof n === 'number' && Number.isFinite(n) && n >= 0;
- if (!['waiting', 'generating', 'paused', 'complete'].includes(speed.phase) ||
- !['turn', 'stream'].includes(speed.waitOrigin) || typeof speed.estimated !== 'boolean' ||
- !finite(speed.waitingMs) || !finite(speed.outputTokens) || !finite(speed.durationMs) ||
- !finite(speed.sampledAt) || (speed.hasRecentOutput !== undefined && typeof speed.hasRecentOutput !== 'boolean') ||
- (speed.toolActive !== undefined && typeof speed.toolActive !== 'boolean') ||
- (speed.outcome !== undefined && !['failed', 'cancelled'].includes(speed.outcome)) ||
- (speed.retrying !== undefined && typeof speed.retrying !== 'boolean') ||
- (speed.firstResponseMs !== null && !finite(speed.firstResponseMs)) ||
- (speed.recentRate !== null && !finite(speed.recentRate)) ||
- (speed.averageRate !== null && !finite(speed.averageRate)) || !Array.isArray(speed.samples)) return undefined;
- return { ...speed, sampledAt: observedAt ?? speed.sampledAt, samples: speed.samples.slice(-60).filter((point) =>
- point && finite(point.durationMs) && finite(point.outputTokens) && finite(point.rate)) };
-}
-
-/** Terminal errors/abort can lack a final provider snapshot. Freeze the last observation. */
-export function stopResponseSpeed(speed: ResponseSpeedSnapshot | undefined, outcome?: ResponseSpeedSnapshot['outcome']): ResponseSpeedSnapshot | undefined {
- if (!speed || (speed.phase === 'complete' && (!outcome || speed.outcome === outcome))) return speed;
- const terminalOutcome = outcome ?? speed.outcome;
- return { ...speed, phase: 'complete', ...(terminalOutcome ? { outcome: terminalOutcome } : {}), retrying: false,
- recentRate: terminalOutcome ? null : speed.averageRate,
- estimated: Boolean(terminalOutcome) || speed.estimated || speed.phase === 'generating' || speed.phase === 'waiting' };
-}
-
-/** Pause displayed generation only after an explicit automatic retry event. */
-export function retryResponseSpeed(speed: ResponseSpeedSnapshot | undefined): ResponseSpeedSnapshot | undefined {
- if (!speed || speed.phase === 'complete') return speed;
- return { ...speed, phase: 'paused', retrying: true, recentRate: null,
- hasRecentOutput: false };
-}
-
-/** Actual output ends the retry annotation without inventing a throughput sample. */
-export function resumeResponseSpeed(speed: ResponseSpeedSnapshot | undefined): ResponseSpeedSnapshot | undefined {
- return speed?.retrying && speed.phase !== 'complete'
- ? { ...speed, phase: speed.toolActive ? 'paused' : 'generating', retrying: false, hasRecentOutput: false, recentRate: null }
- : speed;
-}
-
-/** Preserve terminal/retry annotations across ordinary status tails; a new turn clears them. */
-export function mergeResponseSpeedStatus(previous: ResponseSpeedSnapshot | undefined,
- incoming: ResponseSpeedSnapshot | undefined, isRunning: boolean, isTurnStart: boolean): ResponseSpeedSnapshot | undefined {
- const baseline = isTurnStart ? undefined : previous;
- const next = incoming ?? baseline;
- if (!isRunning) return stopResponseSpeed(next, baseline?.outcome);
- if (baseline?.retrying && next && incoming?.retrying !== false) {
- // Usage calibration can raise tokens without any new output. Only a new
- // stream sample or explicit tool execution can recover a status-only view.
- const newSample = next.hasRecentOutput === true &&
- (next.samples.at(-1)?.durationMs ?? 0) > (baseline.samples.at(-1)?.durationMs ?? 0);
- if (!(next.toolActive && !baseline.toolActive) && !newSample) return retryResponseSpeed({ ...next, phase: 'paused' });
- }
- return next;
-}
-
-/** A finalized speed denominator may only pair with the same real output count. */
-export function calibratedResponseDuration(value: unknown, outputTokens: number): number | undefined {
- const speed = readResponseSpeedSnapshot(value);
- return speed?.phase === 'complete' && !speed.outcome && !speed.estimated && speed.averageRate !== null &&
- speed.outputTokens === outputTokens && speed.durationMs > 0 ? speed.durationMs : undefined;
-}
-
export interface RateSample extends RateCounters {
/** Cumulative measured generation time across turns, excluding unmeasured gaps. */
durationMs: number;
@@ -120,7 +11,7 @@ export interface RateSample extends RateCounters {
// Millisecond-scale usage batches are not meaningful throughput measurements.
const MIN_SAMPLE_DURATION_MS = 1000;
-export const RATE_SAMPLE_FRESH_MS = 1_000;
+export const RATE_SAMPLE_FRESH_MS = 60_000;
export interface RateHistory {
startedAt: number | null;