feat(proxy): disable memory-heavy features under high-concurrency mode

messages/en/settings/config.json

@@ -59,7 +59,8 @@
     "enableOpenaiResponsesWebsocket": "Enable OpenAI Responses WebSocket",
     "enableOpenaiResponsesWebsocketDesc": "When enabled, if a client opens a WebSocket connection to /v1/responses and the selected provider is a Codex type, CCH will attempt a sibling WebSocket to the upstream. If the upstream does not support WebSocket or the handshake fails, CCH gracefully falls back to standard HTTP Responses while keeping the client WebSocket open; the fallback is not counted toward circuit breakers. Non-WebSocket clients and non-Codex providers are unaffected.",
     "enableHighConcurrencyMode": "Enable High-Concurrency Mode",
-    "enableHighConcurrencyModeDesc": "When enabled, CCH disables part of the Redis debug snapshots and real-time session observability writes to reduce CPU and IO pressure under high RPM. Forwarding, rectifiers, fake-200 detection, billing, and quota enforcement remain unchanged, but Sessions debugging details may be reduced or delayed.",
+    "enableHighConcurrencyModeDesc": "When enabled, CCH disables memory-heavy Replay, stream gating, provider-racing loser billing, client-abort retention, and session diagnostics, in addition to Redis debug and observability writes. Forwarding, core billing, and quota enforcement remain enabled.",
+    "highConcurrencyModeWarning": "High-concurrency mode disables Replay, stream gating, racing-loser billing, client-abort retention, and session diagnostics.",
     "enableResponseFixer": "Enable Response Fixer",
     "enableResponseFixerDesc": "Automatically repairs common upstream response issues (encoding, SSE, truncated JSON). Enabled by default.",
     "enableThinkingSignatureRectifier": "Enable Thinking Signature Rectifier",

messages/ja/settings/config.json

@@ -59,7 +59,8 @@
     "enableOpenaiResponsesWebsocket": "OpenAI Responses WebSocket を有効化",
     "enableOpenaiResponsesWebsocketDesc": "有効にすると、クライアントが /v1/responses に WebSocket 接続し、かつ Codex タイプのプロバイダーが選択された場合、CCH は上流にも WebSocket 接続を試みます。上流が WebSocket をサポートしない、またはハンドシェイクに失敗した場合は、クライアント WebSocket を開いたまま通常の HTTP Responses に優雅にフォールバックします。このフォールバックはサーキットブレーカーにカウントされません。非 WebSocket クライアントと非 Codex プロバイダーの動作は変わりません。",
     "enableHighConcurrencyMode": "高並行モードを有効化",
-    "enableHighConcurrencyModeDesc": "有効にすると、高 RPM 時の CPU / IO 負荷を下げるため、Redis の一部デバッグスナップショットとリアルタイム Session 観測書き込みを停止します。転送、整流、fake 200 検知、課金、制限処理は維持されますが、Sessions のデバッグ詳細は減少または遅延する場合があります。",
+    "enableHighConcurrencyModeDesc": "有効にすると、Redis のデバッグスナップショットとリアルタイム Session 観測に加え、メモリ負荷の高い Replay、ストリームゲート、競合敗者の課金、クライアント中断保持、Session 診断を停止します。転送、基本課金、制限処理は維持されます。",
+    "highConcurrencyModeWarning": "高並行モードでは Replay、ストリームゲート、競合敗者の課金、クライアント中断保持、Session 診断を無効化します。",
     "enableResponseFixer": "レスポンス整流を有効化",
     "enableResponseFixerDesc": "上流応答の一般的な形式問題（エンコーディング、SSE、途切れた JSON）を自動修復します（既定で有効）。",
     "enableThinkingSignatureRectifier": "thinking 署名整流を有効化",

messages/ru/settings/config.json

@@ -59,7 +59,8 @@
     "enableOpenaiResponsesWebsocket": "Включить OpenAI Responses WebSocket",
     "enableOpenaiResponsesWebsocketDesc": "Если включено, то когда клиент открывает WebSocket-соединение с /v1/responses и выбирается провайдер типа Codex, CCH попытается установить WebSocket-соединение с вышестоящим сервером. Если сервер не поддерживает WebSocket или рукопожатие не удастся, CCH плавно переключится на обычный HTTP Responses, сохраняя WebSocket клиента открытым; этот fallback не учитывается в circuit breaker. Клиенты без WebSocket и провайдеры, отличные от Codex, работают без изменений.",
     "enableHighConcurrencyMode": "Включить режим высокой нагрузки",
-    "enableHighConcurrencyModeDesc": "Если включено, CCH отключит часть Redis-снимков для отладки и записи real-time Session-наблюдения, чтобы снизить нагрузку на CPU и IO при высоком RPM. Пересылка, rectifier-логика, обнаружение fake 200, биллинг и лимиты сохраняются, но детализация отладки в Sessions может уменьшиться или запаздывать.",
+    "enableHighConcurrencyModeDesc": "При включении CCH отключает Redis-снимки для отладки и real-time-наблюдение Session, а также ресурсоёмкие Replay, stream-gate, тарификацию проигравших в гонке, сохранение при отмене клиентом и диагностику Session. Пересылка, базовый биллинг и лимиты сохраняются.",
+    "highConcurrencyModeWarning": "Режим высокой нагрузки отключает Replay, stream-gate, тарификацию проигравших в гонке, сохранение при отмене клиентом и диагностику Session.",
     "enableResponseFixer": "Включить исправление ответов",
     "enableResponseFixerDesc": "Автоматически исправляет распространённые проблемы ответа у провайдеров (кодировка, SSE, обрезанный JSON). Включено по умолчанию.",
     "enableThinkingSignatureRectifier": "Включить исправление thinking-signature",

messages/zh-CN/settings/config.json

@@ -70,7 +70,8 @@
     "enableOpenaiResponsesWebsocket": "启用 OpenAI Responses WebSocket",
     "enableOpenaiResponsesWebsocketDesc": "启用后，当客户端以 WebSocket 连接 /v1/responses 且选中 Codex 类型供应商时，CCH 会尝试与上游建立 WebSocket。若上游不支持或握手失败，将优雅降级到普通 HTTP Responses，客户端 WebSocket 保持打开；降级不计入熔断。非 WebSocket 客户端与非 Codex 供应商行为不变。",
     "enableHighConcurrencyMode": "启用高并发模式",
-    "enableHighConcurrencyModeDesc": "开启后，将关闭部分 Redis 调试快照与实时会话观测写入，以降低高并发下的 CPU 与 IO 开销。不会影响转发、整流、fake 200 检测、计费与限额，但 Sessions 调试详情会减少或延后。",
+    "enableHighConcurrencyModeDesc": "开启后，除 Redis 调试快照与实时会话观测写入外，还会关闭高内存占用的 Replay、流式门禁、竞速输家计费、客户端中断保留计费和会话诊断。转发、基础计费与限额仍会执行。",
+    "highConcurrencyModeWarning": "高并发模式将关闭 Replay、流式门禁、竞速输家计费、客户端中断保留计费和会话诊断。",
     "interceptAnthropicWarmupRequests": "拦截 Warmup 请求（Anthropic）",
     "interceptAnthropicWarmupRequestsDesc": "开启后，识别到 Claude Code 的 Warmup 探测请求将由 CCH 直接抢答短响应，避免访问上游供应商；该请求会记录在日志中，但不计费、不限流、不计入统计。",
     "enableThinkingSignatureRectifier": "启用 thinking 签名整流器",

messages/zh-TW/settings/config.json

@@ -59,7 +59,8 @@
     "enableOpenaiResponsesWebsocket": "啟用 OpenAI Responses WebSocket",
     "enableOpenaiResponsesWebsocketDesc": "啟用後，當客戶端以 WebSocket 連線 /v1/responses 且命中 Codex 類型供應商時，CCH 會嘗試與上游建立 WebSocket 連線。若上游不支援或握手失敗，將優雅降級為一般 HTTP Responses，客戶端 WebSocket 保持開啟；降級不計入熔斷。非 WebSocket 客戶端與非 Codex 供應商行為不變。",
     "enableHighConcurrencyMode": "啟用高並發模式",
-    "enableHighConcurrencyModeDesc": "開啟後，將關閉部分 Redis 除錯快照與即時 Session 觀測寫入，以降低高並發下的 CPU 與 IO 開銷。轉發、整流、fake 200 偵測、計費與限額不受影響，但 Sessions 除錯詳情會減少或延後。",
+    "enableHighConcurrencyModeDesc": "開啟後，除 Redis 除錯快照與即時 Session 觀測寫入外，也會關閉高記憶體用量的 Replay、串流門控、競速輸家計費、客戶端中斷保留計費與 Session 診斷。轉發、基礎計費與限額仍會執行。",
+    "highConcurrencyModeWarning": "高並發模式將關閉 Replay、串流門控、競速輸家計費、客戶端中斷保留計費與 Session 診斷。",
     "enableResponseFixer": "啟用回應整流",
     "enableResponseFixerDesc": "自動修復上游回應中常見的編碼、SSE 與 JSON 格式問題（預設開啟）。",
     "enableThinkingSignatureRectifier": "啟用 thinking 簽名整流器",

src/app/[locale]/settings/config/_components/system-settings-form.tsx

@@ -947,7 +947,10 @@ export function SystemSettingsForm({
           <Switch
             id="enable-high-concurrency-mode"
             checked={enableHighConcurrencyMode}
-            onCheckedChange={(checked) => setEnableHighConcurrencyMode(checked)}
+            onCheckedChange={(checked) => {
+              setEnableHighConcurrencyMode(checked);
+              if (checked) toast.warning(t("highConcurrencyModeWarning"));
+            }}
             disabled={isPending}
           />
         </div>

src/app/v1/_lib/proxy/forwarder.ts

@@ -1714,7 +1714,9 @@ export class ProxyForwarder {
             let gateChainAudit: ProviderChainItem["streamGate"];
             const gateMode = resolveStreamGateMode();
             const shouldRunPrecommitGate =
-              gateMode === "enforce" || session.replayState?.role === "owner";
+              (typeof session.shouldRunStreamContentGate !== "function" ||
+                session.shouldRunStreamContentGate()) &&
+              (gateMode === "enforce" || session.replayState?.role === "owner");
             if (
               shouldRunPrecommitGate &&
               response.body &&
@@ -1891,6 +1893,8 @@ export class ProxyForwarder {
             isJson &&
             hasValidContentLength &&
             contentLengthBytes <= NON_STREAM_BODY_INSPECTION_MAX_BYTES;
+          // Fake-200 detection is a core failover guard, so it remains active in
+          // high-concurrency mode even though optional diagnostics are disabled.
           const shouldInspectBody = isHtml || !hasValidContentLength || shouldInspectJson;
           if (shouldStrictValidateReplayJson) {
             const validationLimit = getEnvConfig().REPLAY_MAX_PAYLOAD_BYTES;
@@ -4388,7 +4392,9 @@ export class ProxyForwarder {
 
     const rawCrossProviderFallbackEnabled = session.isRawCrossProviderFallbackEnabled();
     // 竞速输家计费开关：开启时落败供应商不被直接掐断，而是后台 drain 并计费。
-    const billHedgeLosers = (await getCachedSystemSettings()).billHedgeLosers === true;
+    const billHedgeLosers =
+      (typeof session.shouldBillHedgeLosers !== "function" || session.shouldBillHedgeLosers()) &&
+      (await getCachedSystemSettings()).billHedgeLosers === true;
     const launchedProviderIds = new Set<number>();
     let launchedProviderCount = 0;
     let settled = false;
@@ -4455,6 +4461,13 @@ export class ProxyForwarder {
     // 不取消连接：读到流自然结束（或超时/容量上限）后，复用赢家相同的计费链，
     // 把费用异步累加回原请求行。幂等（loserBillingStarted 守卫），失败静默。
     const startLoserBilling = (attempt: StreamingHedgeAttempt) => {
+      if (typeof session.shouldBillHedgeLosers === "function" && !session.shouldBillHedgeLosers()) {
+        const cancel = attempt.reader?.cancel("high_concurrency_loser_billing_disabled");
+        cancel?.catch(() => undefined);
+        releaseAttemptAgent(attempt);
+        return;
+      }
+
       if (attempt.loserBillingStarted) return;
       attempt.loserBillingStarted = true;
 
@@ -4779,6 +4792,8 @@ export class ProxyForwarder {
             // 「首个有效内容帧」。
             // 级联阈值计时器保持不动——内容慢的 attempt 不提交，自动触发下一候选竞速。
             const hedgeGateFamily =
+              (typeof session.shouldRunStreamContentGate !== "function" ||
+                session.shouldRunStreamContentGate()) &&
               (resolveStreamGateMode() === "enforce" || session.replayState?.role === "owner") &&
               session.getEndpointPolicy().kind !== "raw_passthrough"
                 ? mapProviderTypeToFamily(attempt.provider.providerType)
@@ -5436,7 +5451,10 @@ export class ProxyForwarder {
     // Discovery uses the same opt-in loser billing switch as legacy Hedge. The
     // attempt is only kept alive after a winner commits when it already has a
     // protocol-valid prefix and a readable response body (see cancelLosers).
-    const billHedgeLosers = settings.billHedgeLosers === true && session.messageContext?.id != null;
+    const billHedgeLosers =
+      (typeof session.shouldBillHedgeLosers !== "function" || session.shouldBillHedgeLosers()) &&
+      settings.billHedgeLosers === true &&
+      session.messageContext?.id != null;
     const coordinator = new DiscoveryCoordinator({ concurrency, maxRounds });
     const discoveryMetrics = new DiscoveryRequestMetrics(
       {

src/app/v1/_lib/proxy/replay/replay-guard.ts

@@ -39,6 +39,10 @@ const ATTACH_MAX_WAIT_MS = 10 * 60 * 1000;
 
 export class ProxyReplayGuard {
   static async ensure(session: ProxySession): Promise<Response | null> {
+    if (typeof session.shouldUseRequestReplay === "function" && !session.shouldUseRequestReplay()) {
+      return null;
+    }
+
     try {
       // guard 位于 provider 步骤之前：先刷新运行时覆写快照，管理端刚保存的
       // replayEnabled 首个请求即生效（底层系统设置缓存有 TTL，常态为缓存命中）

src/app/v1/_lib/proxy/replay/replay-spool.ts

@@ -570,6 +570,10 @@ export function createReplaySpoolIfOwner(
   delivery: ReplayDelivery = "stream",
   options: ReplaySpoolOptions = {}
 ): ReplaySpool | null {
+  if (typeof session.shouldUseRequestReplay === "function" && !session.shouldUseRequestReplay()) {
+    return null;
+  }
+
   const replayState = session.replayState;
   if (replayState?.role !== "owner") return null;
   const declineOwnership = (): null => {