ding113 · ding113 · Aug 22, 2026 · Aug 20, 2026 · Aug 21, 2026 · Aug 21, 2026
diff --git a/.env.example b/.env.example
@@ -171,6 +171,13 @@ FETCH_HEADERS_TIMEOUT=600000
 FETCH_BODY_TIMEOUT=600000
 MAX_RETRY_ATTEMPTS_DEFAULT=2                # 单供应商最大尝试次数（含首次调用），范围 1-10，留空使用默认值 2
 
+# 客户端断开后的 detached stream 共享带权进程级资源预算。
+# Replay owner 申请较重的 replay lease；预算不足时降级为轻量 metering，
+# 两者都无法准入时才终止上游并按 499 结算。
+DETACHED_STREAM_MAX_CONCURRENCY=64
+DETACHED_STREAM_BUDGET_BYTES=67108864
+DETACHED_STREAM_METERING_RESERVE_BYTES=16777216
+
 # 入站压缩请求体（content-encoding: zstd/gzip/deflate/br）解压上限（字节）
 # 功能说明：/v1、/v1beta 代理路径不受 proxyClientMaxBodySize 钳制，这两项是入站解压的内存/CPU 兜底。
 # - MAX_DECOMPRESSED_REQUEST_BYTES：解压输出上限，防御解压炸弹，超过按 413 拒绝。默认 100MB。

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -8,6 +8,12 @@
 
 ### 修复
 
+- 修复上游响应流发生 error 后 Node/Undici body 未完成销毁的问题：Node-to-Web adapter 和 demand-driven pump
+  现在在源流错误终态显式取消、销毁底层流，并为异步 destroy error 保留有界保护；同时将 raw body 的兜底错误监听改为一次性监听，
+  避免 HTTP/2 reset、客户端断开和竞速取消路径长期保留 socket 与 ArrayBuffer backing store (#1430)
+- 修复客户端断开后的后台计费 drain 继续累加完整响应正文导致的高并发内存放大：断线后切换到有界计量观察器，
+  仅保留 usage、终止标记、模型与协议错误等结算证据，拿到终态即取消上游；Replay owner 在预算内继续保存完整
+  客户端可见流，Replay 预算不足时降级到 metering，新增共享进程级并发与带权保留容量预算，覆盖通用流与 Gemini 透传路径 (#1430)
 - 修复 Replay owner 在客户端断线后保留完整流正文和 300 秒传输资源导致的内存失控：限制 Redis
   write-behind backlog，Replay 失效后按断线起点恢复 60 秒 drain，并为 Redis session response body
   增加默认 5 MiB 的可配置存储上限，避免大 SSE 正文及 before/after 快照放大内存和持久化压力；

diff --git a/messages/en/dashboard.json b/messages/en/dashboard.json
@@ -551,6 +551,11 @@
         "overridden": "Overridden by provider",
         "tooltip": "Thinking effort in the Codex request (reasoning.effort), shown verbatim."
       },
+      "reasoningEffortOpenai": {
+        "label": "Reasoning effort",
+        "overridden": "Overridden by provider",
+        "tooltip": "Thinking effort in the OpenAI chat/completions request (reasoning_effort / reasoning.effort), shown verbatim."
+      },
       "logicTrace": {
         "title": "Decision Chain",
         "singleRouteSelectionTitle": "Provider selection under single-route protection",

diff --git a/messages/en/settings/config.json b/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",

diff --git a/messages/ja/dashboard.json b/messages/ja/dashboard.json
@@ -551,6 +551,11 @@
         "overridden": "プロバイダーにより上書き",
         "tooltip": "Codex リクエストの推論強度 (reasoning.effort) をそのまま表示します。"
       },
+      "reasoningEffortOpenai": {
+        "label": "推論強度",
+        "overridden": "プロバイダーにより上書き",
+        "tooltip": "OpenAI chat/completions リクエストの推論強度 (reasoning_effort / reasoning.effort) をそのまま表示します。"
+      },
       "logicTrace": {
         "title": "決定チェーン",
         "singleRouteSelectionTitle": "単一経路保護での Provider 選択",

diff --git a/messages/ja/settings/config.json b/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",

diff --git a/messages/ru/dashboard.json b/messages/ru/dashboard.json
@@ -551,6 +551,11 @@
         "overridden": "Переопределено провайдером",
         "tooltip": "Интенсивность рассуждений в запросе Codex (reasoning.effort), показанная без изменений."
       },
+      "reasoningEffortOpenai": {
+        "label": "Интенсивность рассуждений",
+        "overridden": "Переопределено провайдером",
+        "tooltip": "Интенсивность рассуждений в запросе OpenAI chat/completions (reasoning_effort / reasoning.effort), показанная без изменений."
+      },
       "logicTrace": {
         "title": "Цепочка решений",
         "singleRouteSelectionTitle": "Выбор провайдера в режиме защиты одним маршрутом",

diff --git a/messages/ru/settings/config.json b/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",

diff --git a/messages/zh-CN/dashboard.json b/messages/zh-CN/dashboard.json
@@ -551,6 +551,11 @@
         "overridden": "已被供应商覆写",
         "tooltip": "Codex 请求中的思考强度（reasoning.effort），按原值显示。"
       },
+      "reasoningEffortOpenai": {
+        "label": "思考强度",
+        "overridden": "已被供应商覆写",
+        "tooltip": "OpenAI chat/completions 请求中的思考强度（reasoning_effort / reasoning.effort），按原值显示。"
+      },
       "logicTrace": {
         "title": "决策链",
         "singleRouteSelectionTitle": "单路保护下的供应商选择",

diff --git a/messages/zh-CN/settings/config.json b/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",

diff --git a/messages/zh-TW/dashboard.json b/messages/zh-TW/dashboard.json
@@ -551,6 +551,11 @@
         "overridden": "已被供應商覆寫",
         "tooltip": "Codex 請求中的思考強度（reasoning.effort），按原值顯示。"
       },
+      "reasoningEffortOpenai": {
+        "label": "思考強度",
+        "overridden": "已被供應商覆寫",
+        "tooltip": "OpenAI chat/completions 請求中的思考強度（reasoning_effort / reasoning.effort），按原值顯示。"
+      },
       "logicTrace": {
         "title": "決策鏈",
         "singleRouteSelectionTitle": "單路保護下的供應商選擇",

diff --git a/messages/zh-TW/settings/config.json b/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",

diff --git a/package.json b/package.json
@@ -62,6 +62,7 @@
     "@langfuse/otel": "^5.10.0",
     "@langfuse/tracing": "^5.10.0",
     "@lobehub/icons": "^5.15.0",
+    "@lobehub/ui": "^5.0.0",
     "@opentelemetry/sdk-node": "^0.221.0",
     "@radix-ui/react-alert-dialog": "^1.1.23",
     "@radix-ui/react-avatar": "^1.2.6",

diff --git a/src/app/[locale]/dashboard/logs/_components/error-details-dialog/components/SummaryTab.tsx b/src/app/[locale]/dashboard/logs/_components/error-details-dialog/components/SummaryTab.tsx
@@ -142,7 +142,12 @@ export function SummaryTab({
   const showNoSignatureBadge =
     thinkingSignatureDetection?.source === "fallback_no_signature_with_thinking";
   const thinkingEffortInfo = extractThinkingEffortInfo(specialSettings);
-  const effortMessageKey = thinkingEffortInfo?.source === "codex" ? "reasoningEffort" : "effort";
+  const effortMessageKey =
+    thinkingEffortInfo?.source === "codex"
+      ? "reasoningEffort"
+      : thinkingEffortInfo?.source === "openai"
+        ? "reasoningEffortOpenai"
+        : "effort";
   const effortDisplay = thinkingEffortInfo
     ? {
         requestedEffort: thinkingEffortInfo.requestedEffort,

diff --git a/src/app/[locale]/dashboard/logs/_components/thinking-effort-display.test.tsx b/src/app/[locale]/dashboard/logs/_components/thinking-effort-display.test.tsx
@@ -180,4 +180,51 @@ describe("ThinkingEffortDisplay", () => {
     expect(html).toContain("reasoningEffort.tooltip");
     expect(html).not.toContain(">medium<");
   });
+
+  test("显示 OpenAI chat/completions 请求中的思考强度", () => {
+    const html = renderToStaticMarkup(
+      <ThinkingEffortDisplay
+        specialSettings={[
+          {
+            type: "openai_reasoning_effort",
+            scope: "request",
+            hit: true,
+            effort: "max",
+            source: "reasoning_effort",
+          },
+        ]}
+      />
+    );
+
+    expect(html).toContain('data-slot="thinking-effort"');
+    expect(html).toContain("max");
+    expect(html).toContain("reasoningEffortOpenai.tooltip");
+    expect(html).not.toContain("overridden");
+  });
+
+  test("OpenAI 与 Codex 审计并存时优先展示 Codex 强度", () => {
+    const html = renderToStaticMarkup(
+      <ThinkingEffortDisplay
+        specialSettings={[
+          {
+            type: "openai_reasoning_effort",
+            scope: "request",
+            hit: true,
+            effort: "max",
+            source: "reasoning_effort",
+          },
+          {
+            type: "codex_reasoning_effort",
+            scope: "request",
+            hit: true,
+            effort: "high",
+          },
+        ]}
+      />
+    );
+
+    expect(html).toContain("high");
+    expect(html).toContain("reasoningEffort.tooltip");
+    expect(html).not.toContain(">max<");
+  });
 });
diff --git a/src/app/[locale]/dashboard/logs/_components/thinking-effort-display.tsx b/src/app/[locale]/dashboard/logs/_components/thinking-effort-display.tsx
@@ -14,7 +14,8 @@ interface ThinkingEffortDisplayProps {
 }
 
 /**
- * 在使用记录中展示任意模型的思考强度（Codex reasoning.effort 或 Anthropic effort）。
+ * 在使用记录中展示任意模型的思考强度（Codex / OpenAI chat/completions 的
+ * reasoning.effort，或 Anthropic 的 output_config.effort）。
  *
  * 供应商改变强度时同时展示请求值和实际转发值，避免只看到客户端参数而误判上游行为。
  */
@@ -26,7 +27,12 @@ export function ThinkingEffortDisplay({ specialSettings }: ThinkingEffortDisplay
     return <span className="text-muted-foreground">-</span>;
   }
 
-  const messageNamespace = effortInfo.source === "codex" ? "reasoningEffort" : "effort";
+  const messageNamespace =
+    effortInfo.source === "anthropic"
+      ? "effort"
+      : effortInfo.source === "openai"
+        ? "reasoningEffortOpenai"
+        : "reasoningEffort";
   const showEffectiveBadge = effortInfo.isOverridden && effortInfo.effectiveEffort != null;
 
   return (

diff --git a/src/app/[locale]/settings/config/_components/system-settings-form.tsx b/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>