fix: admit reasoning envelope and request copies before materialization

docs-site/src/content/docs/ja/reference/proxy-formats.md

@@ -131,6 +131,11 @@ WebSocket が無効になっている場合、アップグレード試行では
 
 これらのエンドポイントは、Claude Code および互換性のあるクライアントによって使用される Anthropic Messages 言語を話します。ほとんどのリクエストはレスポンスに変換され、通常どおりルーティングされてから、Anthropic JSON または Anthropic SSE に変換されます。
 
+変換される Messages リクエストでは、推論の再送もリクエスト共通の変換バジェットを使います。
+この制限にはエンコード・デコード時のコピー分も含まれます。超過時は
+`translation_buffer_limit` を伴う HTTP 413 を返し、署名や不透明な推論データを切り詰めません。
+ネイティブ Anthropic パススルーには、別の本文サイズ制限が適用されます。
+
 ネイティブ Anthropic パススルーは、次のすべてが当てはまる場合にのみ適格です。
 
 - ネイティブ パススルーはクロード コード設定で無効になっていません。

docs-site/src/content/docs/ko/reference/proxy-formats.md

@@ -174,6 +174,11 @@ SSE 객체, choice delta, `finish_reason`이 있는 종료 choice, `data: [DONE]
 이 엔드포인트는 Claude Code와 호환 클라이언트가 사용하는 Anthropic Messages 방언을 말합니다. 대부분의 요청은
 Responses로 변환되어 일반적으로 라우팅된 뒤, Anthropic JSON 또는 Anthropic SSE로 다시 변환됩니다.
 
+변환되는 Messages 요청의 reasoning 재전송은 요청 전체의 번역 예산을 공유합니다. 이 예산에는
+인코딩·디코딩 과정에서 생기는 복사본도 포함됩니다. 한도를 초과하면 `translation_buffer_limit`과
+HTTP 413을 반환하며, 한도에 맞추려고 서명이나 불투명 reasoning 데이터를 자르지 않습니다.
+네이티브 Anthropic passthrough에는 별도의 본문 크기 제한이 적용됩니다.
+
 네이티브 Anthropic passthrough는 다음이 모두 참일 때만 적용됩니다.
 
 - Claude Code 설정에서 native passthrough가 비활성화되어 있지 않습니다.

docs-site/src/content/docs/reference/adapters.md

@@ -195,6 +195,10 @@ header and does not guarantee a provider cache hit.
 **Auth:** `key` (`x-api-key` by default, or `Authorization: Bearer` with `apiKeyTransport: "bearer"`) or `oauth` (Bearer + `anthropic-beta`, for Claude Pro/Max).
 
 - Converts messages to Anthropic content blocks (text, base64 image, `tool_use`, `thinking`).
+- Translated Anthropic Messages reasoning replay shares the request translation budget, including
+  encoding/decoding copy overhead. Requests exceeding it return HTTP 413 with
+  `translation_buffer_limit`; signatures and opaque reasoning data are never truncated to fit.
+  Native Anthropic passthrough uses its separate body-size contract.
 - **Extended thinking math:** Anthropic requires `max_tokens > thinking.budget_tokens`. The adapter
   maps reasoning effort to a budget (minimal 1024 … max 32000), then computes a safe `max_tokens` with
   output headroom, and **drops `temperature`/`top_p`** when thinking is enabled (Anthropic forbids

docs-site/src/content/docs/reference/proxy-formats.md

@@ -278,6 +278,12 @@ These endpoints speak the Anthropic Messages dialect used by Claude Code and com
 Most requests are translated to Responses, routed normally, then translated back to Anthropic JSON
 or Anthropic SSE.
 
+On translated Messages requests, reasoning replay shares the request's translation budget.
+Envelope admission includes encoding/decoding copy overhead, not just the original signature
+length. Requests exceeding this budget return HTTP 413 with `translation_buffer_limit`;
+signatures and opaque reasoning data are never truncated to make a request fit. Native
+Anthropic passthrough retains its separate body-size contract.
+
 Base64 and URL image sources are translated in user messages and nested tool results. File-backed
 images (`source.type: "file"`) require native Anthropic passthrough; translated routes return a
 fixed HTTP 400 error asking for base64 or URL input. OpenCodex does not resolve another provider's

docs-site/src/content/docs/ru/reference/proxy-formats.md

@@ -179,6 +179,12 @@ adapter, вместо тихого изменения смысла вернёт
 клиенты. Большинство запросов переводится в Responses, маршрутизируется обычным образом, а затем
 обратно в Anthropic JSON или Anthropic SSE.
 
+Повторная передача reasoning в преобразуемых запросах Messages использует общий бюджет
+преобразования запроса, включая копии при кодировании и декодировании. При превышении лимита
+возвращается HTTP 413 с `translation_buffer_limit`; подписи и непрозрачные данные reasoning
+не обрезаются для соблюдения лимита. Для нативного Anthropic passthrough действует отдельный
+контракт ограничения размера тела.
+
 Нативный Anthropic passthrough допустим только когда одновременно выполняются все условия:
 
 - native passthrough не отключён в конфигурации Claude Code;

docs-site/src/content/docs/zh-cn/reference/proxy-formats.md

@@ -149,6 +149,10 @@ choice 增量、带 `finish_reason` 的终止 choice，以及 `data: [DONE]`。
 
 这些端点使用 Claude Code 和兼容客户端所采用的 Anthropic Messages 方言。大多数请求会被转换为 Responses，按常规路由，然后再转换回 Anthropic JSON 或 Anthropic SSE。
 
+转换后的 Messages 请求在重放推理数据时共享整个请求的转换预算，其中包含编码和解码产生的副本开销。
+超出预算时返回 HTTP 413 和 `translation_buffer_limit`，不会为了满足限制而截断签名或不透明推理数据。
+原生 Anthropic 透传使用独立的请求体大小限制。
+
 只有在满足以下全部条件时，原生 Anthropic 透传才有资格启用：
 
 - Claude Code 配置中尚未禁用原生透传；

src/bridge.ts

@@ -493,7 +493,7 @@ export function bridgeToResponsesSSE(
         const previousBytes = pendingSignatureBytes
           + pendingRedacted.reduce((sum, value) => sum + bytesOf(value), 0)
           + (hiddenText ? hiddenThinkingBytes : 0);
-        const encoded = encodeReasoningEnvelope(envelope);
+        const encoded = encodeReasoningEnvelope(envelope, budget);
         const reservation = budget?.reserveTransient(bytesOf(encoded), { kind: "reasoning" });
         pendingSignature = undefined;
         pendingSignatureBytes = 0;
@@ -533,7 +533,7 @@ export function bridgeToResponsesSSE(
         if (!hiddenRawReasoningText) return;
         rawReasoningForNextToolCall = hiddenRawReasoningText;
         const previousBytes = hiddenRawReasoningBytes;
-        const encrypted = encodeReasoningEnvelope({ txt: hiddenRawReasoningText });
+        const encrypted = encodeReasoningEnvelope({ txt: hiddenRawReasoningText }, budget);
         const reservation = budget?.reserveTransient(bytesOf(encrypted), { kind: "reasoning" });
         hiddenRawReasoningText = "";
         hiddenRawReasoningBytes = 0;
@@ -556,7 +556,7 @@ export function bridgeToResponsesSSE(
       const flushKiroRedactedReasoning = () => {
         if (!pendingKiroRedacted) return;
         const previousBytes = pendingKiroRedactedBytes;
-        const encrypted = encodeReasoningEnvelope({ krc: pendingKiroRedacted });
+        const encrypted = encodeReasoningEnvelope({ krc: pendingKiroRedacted }, budget);
         const reservation = budget?.reserveTransient(bytesOf(encrypted), { kind: "reasoning" });
         pendingKiroRedacted = undefined;
         pendingKiroRedactedBytes = 0;
@@ -902,6 +902,16 @@ export function bridgeToResponsesSSE(
         gated = true;
         stepping = false;
       };
+      const attemptTerminationCleanup = (action: () => void): boolean => {
+        try {
+          action();
+          return !terminated && !closed;
+        } catch (error) {
+          if (!isTranslatorBudgetExceededError(error)) throw error;
+          terminateForTranslatorOverflow(error);
+          return false;
+        }
+      };
       const step = async () => {
         if (stepping || closed) return;
         stepping = true;
@@ -1415,10 +1425,12 @@ export function bridgeToResponsesSSE(
           return;
         }
         if (!terminated) {
-          flushHiddenRawReasoning();
-          if (currentToolCall) failCurrentToolCall();
-          if (currentWebSearch) closeCurrentWebSearch("failed", []);
-          releasePendingWebSources();
+          if (!attemptTerminationCleanup(() => {
+            flushHiddenRawReasoning();
+            if (currentToolCall) failCurrentToolCall();
+            if (currentWebSearch) closeCurrentWebSearch("failed", []);
+            releasePendingWebSources();
+          })) return;
           const failure = responseError(
             500,
             "proxy_error",
@@ -1448,13 +1460,15 @@ export function bridgeToResponsesSSE(
       if (!terminated) {
         // The adapter generator ended without an explicit done/error event. Mark as incomplete
         // rather than completed so Codex can distinguish a clean finish from a truncated stream.
-        if (currentMsg) closeCurrentMessage();
-        if (currentReasoning) closeCurrentReasoning();
-        if (currentRawReasoning) closeCurrentRawReasoning();
-        flushHiddenRawReasoning();
-        if (currentToolCall) failCurrentToolCall();
-        if (currentWebSearch) closeCurrentWebSearch("failed", []);
-        releasePendingWebSources();
+        if (!attemptTerminationCleanup(() => {
+          if (currentMsg) closeCurrentMessage();
+          if (currentReasoning) closeCurrentReasoning();
+          if (currentRawReasoning) closeCurrentRawReasoning();
+          flushHiddenRawReasoning();
+          if (currentToolCall) failCurrentToolCall();
+          if (currentWebSearch) closeCurrentWebSearch("failed", []);
+          releasePendingWebSources();
+        })) return;
         options?.onUsage?.(undefined);
         await awaitThoughtSignatureDurability();
         emit("response.incomplete", {
@@ -1493,13 +1507,15 @@ export function bridgeToResponsesSSE(
             upstreamActivity = false;
             stallTicks = 0;
           } else if (++stallTicks >= maxStallTicks) {
-            if (currentMsg) closeCurrentMessage();
-            if (currentReasoning) closeCurrentReasoning();
-            if (currentRawReasoning) closeCurrentRawReasoning();
-            flushHiddenRawReasoning();
-            if (currentToolCall) failCurrentToolCall();
-            if (currentWebSearch) closeCurrentWebSearch("failed", []);
-            releasePendingWebSources();
+            if (!attemptTerminationCleanup(() => {
+              if (currentMsg) closeCurrentMessage();
+              if (currentReasoning) closeCurrentReasoning();
+              if (currentRawReasoning) closeCurrentRawReasoning();
+              flushHiddenRawReasoning();
+              if (currentToolCall) failCurrentToolCall();
+              if (currentWebSearch) closeCurrentWebSearch("failed", []);
+              releasePendingWebSources();
+            })) return;
             // #1926 gap 2 residual: this beat callback is synchronous, so the durability
             // barrier is not awaited on the stall-timeout kill path. The in-memory store is
             // already updated; only a crash between here and the queued write loses it,
@@ -1728,7 +1744,7 @@ function buildResponseJSONWithBudget(
     if (batchRedacted.length > 0) envelope.red = batchRedacted;
     const hidden = options?.hideThinkingSummary === true;
     if (hidden && currentSummaryReasoning && (envelope.sig || envelope.red)) envelope.txt = currentSummaryReasoning;
-    const encrypted = envelope.sig || envelope.red || envelope.txt ? encodeReasoningEnvelope(envelope) : undefined;
+    const encrypted = envelope.sig || envelope.red || envelope.txt ? encodeReasoningEnvelope(envelope, budget) : undefined;
     const sourceBytes = currentSummaryReasoningBytes + batchSignatureBytes + batchRedactedBytes;
     batchSignature = undefined;
     batchSignatureBytes = 0;
@@ -1756,7 +1772,7 @@ function buildResponseJSONWithBudget(
       // Same contract as the streaming path: no visible reasoning, txt-only envelope round-trip.
       pushOutput({
         type: "reasoning", id: `rs_${uuid()}`, summary: [],
-        encrypted_content: encodeReasoningEnvelope({ txt: currentRawReasoning }),
+        encrypted_content: encodeReasoningEnvelope({ txt: currentRawReasoning }, budget),
       }, currentRawReasoningBytes, "reasoning");
       currentRawReasoning = "";
       currentRawReasoningBytes = 0;
@@ -2044,7 +2060,7 @@ function buildResponseJSONWithBudget(
     // pushOutput reserves the item itself and releases the retained raw blob it replaces.
     pushOutput({
       type: "reasoning", id: `rs_${uuid()}`, summary: [],
-      encrypted_content: encodeReasoningEnvelope({ krc: batchKiroRedacted }),
+      encrypted_content: encodeReasoningEnvelope({ krc: batchKiroRedacted }, budget),
     }, batchKiroRedactedBytes, "reasoning");
     batchKiroRedacted = undefined;
     batchKiroRedactedBytes = 0;

src/claude/inbound.ts

@@ -18,6 +18,7 @@ import { AnthropicRequestError, isRec, type Rec } from "./inbound-records";
 import { resolveInboundModel, effortForThinkingBudget, effortFromOutputConfig, formatFromOutputConfig } from "./inbound-model-options";
 import { systemToInstructions, toolsToResponses, toolChoiceToResponses } from "./inbound-content-options";
 import { decodeReasoningEnvelope, encodeReasoningEnvelope, OCX_REASONING_PREFIX } from "../responses/reasoning-envelope";
+import { createTranslatorBudget, type TranslatorBudget } from "../lib/translator-budget";
 
 
 
@@ -210,7 +211,7 @@ function userMessageToItems(content: unknown, input: Rec[], elide: SkillElisionC
   pushUserMessage(input, pending);
 }
 
-function assistantMessageToItems(content: unknown, input: Rec[]): void {
+function assistantMessageToItems(content: unknown, input: Rec[], budget: TranslatorBudget): void {
   if (typeof content === "string") {
     if (content.length > 0) input.push({ type: "message", role: "assistant", content: [{ type: "output_text", text: content }] });
     return;
@@ -240,19 +241,24 @@ function assistantMessageToItems(content: unknown, input: Rec[]): void {
         const thinking = typeof raw.thinking === "string" ? raw.thinking : "";
         const signature = typeof raw.signature === "string" ? raw.signature : "";
         if (signature.startsWith(OCX_REASONING_PREFIX)) {
-          const owned = decodeReasoningEnvelope(signature);
+          const owned = decodeReasoningEnvelope(signature, budget);
           if (!owned) throw new AnthropicRequestError("malformed ocxr1 reasoning signature");
           if (Object.hasOwn(owned, "sig")) throw new AnthropicRequestError("OpenCodex reasoning continuity cannot be replayed as an Anthropic signature");
         }
-        const encrypted = signature.length === 0 ? undefined : signature.startsWith(OCX_REASONING_PREFIX) ? signature : encodeReasoningEnvelope({ sig: signature });
+        const encrypted = signature.length === 0 ? undefined : signature.startsWith(OCX_REASONING_PREFIX) ? signature : encodeReasoningEnvelope({ sig: signature }, budget);
+        if (encrypted) budget.chargeRetained(2 * encrypted.length, { kind: "reasoning" });
         if (thinking.length === 0 && !encrypted) break;
         input.push({ type: "reasoning", id: `rs_${crypto.randomUUID().replace(/-/g, "")}`, summary: thinking.length > 0 ? [{ type: "summary_text", text: thinking }] : [], ...(encrypted ? { encrypted_content: encrypted } : {}) });
         break;
       }
       case "redacted_thinking": {
         flush();
         const data = typeof raw.data === "string" ? raw.data : "";
-        if (data.length > 0) input.push({ type: "reasoning", id: `rs_${crypto.randomUUID().replace(/-/g, "")}`, summary: [], encrypted_content: encodeReasoningEnvelope({ red: [data] }) });
+        if (data.length > 0) {
+          const encrypted = encodeReasoningEnvelope({ red: [data] }, budget);
+          budget.chargeRetained(2 * encrypted.length, { kind: "reasoning" });
+          input.push({ type: "reasoning", id: `rs_${crypto.randomUUID().replace(/-/g, "")}`, summary: [], encrypted_content: encrypted });
+        }
         break;
       }
       default:
@@ -294,7 +300,16 @@ export function anthropicToResponsesBody(raw: unknown, cc?: OcxClaudeCodeConfig)
  * OUT-OF-BODY tuple (audit 133 R3#1 — an in-body marker would leak upstream through
  * the native Responses forward and 400).
  */
-export function anthropicToResponsesTranslation(raw: unknown, cc?: OcxClaudeCodeConfig): ClaudeInboundTranslation {
+export function anthropicToResponsesTranslation(raw: unknown, cc?: OcxClaudeCodeConfig, budget?: TranslatorBudget): ClaudeInboundTranslation {
+  const activeBudget = budget ?? createTranslatorBudget();
+  try {
+    return translateAnthropicRequest(raw, cc, activeBudget);
+  } finally {
+    if (!budget) activeBudget.dispose();
+  }
+}
+
+function translateAnthropicRequest(raw: unknown, cc: OcxClaudeCodeConfig | undefined, budget: TranslatorBudget): ClaudeInboundTranslation {
   if (!isRec(raw)) throw new AnthropicRequestError("request body must be a JSON object");
   if (typeof raw.model !== "string" || raw.model.length === 0) {
     throw new AnthropicRequestError("model is required");
@@ -315,7 +330,7 @@ export function anthropicToResponsesTranslation(raw: unknown, cc?: OcxClaudeCode
   for (const msg of raw.messages) {
     if (!isRec(msg)) throw new AnthropicRequestError("each message must be an object");
     if (msg.role === "user") userMessageToItems(msg.content, input, elide);
-    else if (msg.role === "assistant") assistantMessageToItems(msg.content, input);
+    else if (msg.role === "assistant") assistantMessageToItems(msg.content, input, budget);
     else if (msg.role === "system") {
       const text = systemMessageText(msg.content);
       if (text.length > 0) systemParts.push(text);