@bitkyc08/opencodex 2.81.0 → 2.82.0

package/src/server/messages-native.ts

@@ -36,11 +36,12 @@ import {
   buildAnthropicMessagesPassthroughRequest,
   type AnthropicMessagesPassthroughRequest,
 } from "../adapters/anthropic/passthrough";
-import { resolveInboundModel } from "../claude/inbound";
-import { anthropicErrorBody, anthropicErrorResponse, collectAnthropicMessage } from "../claude/outbound";
+import { effortFromOutputConfig, resolveInboundModel } from "../claude/inbound";
+import { anthropicErrorBody, anthropicErrorResponse, claudeOverflowSsePayload, claudePromptTooLongMessage, collectAnthropicMessage, isContextOverflowText, isThroughputLimitText } from "../claude/outbound";
 import type { AdmissionLease } from "../lib/admission";
 import { readBoundedResponseBody } from "../lib/bounded-body";
 import { classifyError } from "../lib/errors";
+import { relaySseWithPayloadRewrite } from "./sse-payload-rewrite";
 import { redactSecretString } from "../lib/redact";
 import { resolveClientRetryAfter } from "../lib/retry-after";
 import { isTranslatorBudgetExceededError, type TranslatorBudget } from "../lib/translator-budget";
@@ -107,6 +108,7 @@ import {
 import {
   noteProviderAttemptSend,
   recordAttemptCredentialSource,
+  recordAttemptRequestedEffort,
   recordFirstOutput,
   recordKeyAttemptFailure,
   recordKeyWireAttemptUsage,
@@ -119,6 +121,7 @@ import { workflowRefusalResponse } from "./workflow-refusal";
 import { sseFieldValue } from "../lib/sse-decoder";
 import { admissionModelDeniedResponse, type AdmissionModelScope } from "./admission-model-scope";
 import { nativeMessagesToolScopeDenial } from "./messages-native-scope";
+import { retainUpstreamMessagesRequestId } from "./messages-response-headers";
 
 export {
   isNativeMessagesRouteEligible,
@@ -317,6 +320,15 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
   logCtx.requestedModel = requestedModel;
   if (route.routeReason === "model-alias" || route.modelId !== requestedModel) logCtx.requestedAlias = requestedModel;
   logCtx.requestedServiceTier = typeof body.service_tier === "string" ? body.service_tier : undefined;
+  if (logCtx.requestedEffort === undefined) {
+    const effort = effortFromOutputConfig(body.output_config);
+    const thinking = isRec(body.thinking) ? body.thinking : undefined;
+    const budget = thinking?.budget_tokens;
+    if (effort !== undefined) logCtx.requestedEffort = effort;
+    else if (thinking?.type === "disabled") logCtx.requestedEffort = "none";
+    else if (thinking?.type === "enabled" && typeof budget === "number"
+      && Number.isSafeInteger(budget) && budget > 0) logCtx.requestedEffort = `budget:${budget}`;
+  }
   // Reserve spend the way native Chat does: an input estimate that never enters usage, and the
   // caller's own output ceiling.
   if (logCtx.usageLogInputTokens === undefined) {
@@ -332,6 +344,7 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
     adapter: "anthropic",
   });
   attemptHandle.seal(logCtx.accountLogLabel);
+  recordAttemptRequestedEffort(logCtx);
   const { attempt } = attemptHandle;
   const finalLog = createFinalRequestLog(logIds, logCtx);
   const finishLog: FinishLog = (status, message, meta = { closeReason: "non_stream" }) => {
@@ -500,7 +513,7 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
       && nativeMessagesDeclineReason({ ...route, provider: routed }, body, config, selector) === undefined;
   };
 
-  const send = async (recovery?: "rate-limit-429" | "oauth-account-403" | "key-429" | "key-401"): Promise<Response> => {
+  const send = async (recovery?: "rate-limit-429" | "oauth-401" | "oauth-account-403" | "key-429" | "key-401"): Promise<Response> => {
     const remaining = remainingTransientSends();
     if (requestTransientPolicy && remaining <= 0) {
       throw new Error("native Messages transient send budget exhausted before recovery dispatch");
@@ -672,7 +685,7 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
     const oauthRetryKey = {};
     const triedAccountIds = new Set<string>();
     let oauthFailovers = 0;
-    while (oauthBinding && (response.status === 429 || response.status === 403)) {
+    while (oauthBinding && (response.status === 429 || response.status === 403 || response.status === 401)) {
       const sendingBinding = oauthBinding;
       const expectedRecoverySelection = sendingBinding.selection;
       triedAccountIds.add(sendingBinding.snapshot.accountId);
@@ -690,7 +703,7 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
         canRetry: transientSendAvailable() && oauthFailovers < ANTHROPIC_POOL_MAX_FAILOVERS_PER_REQUEST,
       });
       if (!nextAccountId) break;
-      const recovery = response.status === 403 ? "oauth-account-403" : "rate-limit-429";
+      const recovery = response.status === 401 ? "oauth-401" : response.status === 403 ? "oauth-account-403" : "rate-limit-429";
       discard(response);
       oauthBinding = await resolveNativeOAuthBindingForInstance(nativeInstance!, config, { routeTarget,
         sessionKey: options.sessionKey, model: route.modelId, candidateAccountId: nextAccountId,
@@ -739,7 +752,7 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
       upstream.abort();
       return fail(499, "Client cancelled request", "api_error");
     }
-    return nativeMessagesErrorResponse(response, bodyText, finishLog);
+    return retainUpstreamMessagesRequestId(nativeMessagesErrorResponse(response, bodyText, finishLog), response.headers);
   }
 
   const contentType = response.headers.get("content-type")?.toLowerCase() ?? "";
@@ -764,10 +777,15 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
           releaseStreamTurn();
         }
       }, bodyGuard);
-      return new Response(relayed, {
+      // A configured Messages provider words an oversized-input refusal its own way; Claude Code
+      // compacts only on Anthropic's wording (devlog/_plan/261009_claude_1m_default/010). The
+      // rewrite buffers whole frames, so it sits after the tap: stall detection keeps timing raw
+      // upstream bytes. Real Anthropic pools refuse pre-stream in that wording and keep one relay.
+      const clientStream = nativeInstance ? relayed : relaySseWithPayloadRewrite(relayed, claudeOverflowSsePayload, translatorBudget);
+      return retainUpstreamMessagesRequestId(new Response(clientStream, {
         status: 200,
         headers: { "Content-Type": "text/event-stream; charset=utf-8", "Cache-Control": "no-cache", Connection: "keep-alive" },
-      });
+      }), response.headers);
     }
     // A non-streaming caller whose upstream streamed anyway: fold the stream into one message.
     const tapState: { closeReason?: FinalRequestLogMeta["closeReason"]; meta?: FinalRequestLogMeta } = {};
@@ -789,10 +807,11 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
         // as the streaming lane's row; a plain upstream error event keeps the non_stream row.
         const tapMeta = tapState.meta;
         if (tapMeta && (tapMeta.terminalStatus || tapMeta.closeReason !== "terminal")) finishLog(502, text, tapMeta);
+        if (isNativeOverflowError(error, text)) return fail(400, claudePromptTooLongMessage(text), "invalid_request_error");
         return fail(502, text, "api_error");
       }
       finishLog(200);
-      return Response.json(message);
+      return retainUpstreamMessagesRequestId(Response.json(message), response.headers);
     } catch (error) {
       cleanupAbort();
       upstream.abort();
@@ -850,12 +869,19 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
   const serialized = JSON.stringify(message);
   finishLog(200);
   if (requestedStream) {
-    return new Response(messageAsSse(message), {
+    return retainUpstreamMessagesRequestId(new Response(messageAsSse(message), {
       status: 200,
       headers: { "Content-Type": "text/event-stream; charset=utf-8", "Cache-Control": "no-cache" },
-    });
+    }), response.headers);
   }
-  return new Response(serialized, { status: 200, headers: { "Content-Type": "application/json" } });
+  return retainUpstreamMessagesRequestId(new Response(serialized, { status: 200, headers: { "Content-Type": "application/json" } }), response.headers);
+}
+
+/** A folded stream error that refuses an oversized input (same gate as the SSE rewrite). */
+function isNativeOverflowError(error: Rec, message: string): boolean {
+  const sized = error.type === "invalid_request_error" || error.type === "request_too_large";
+  return sized && !isThroughputLimitText(message)
+    && (error.code === "context_length_exceeded" || isContextOverflowText(message));
 }
 
 /**
@@ -865,11 +891,13 @@ export async function handleNativeMessages(options: HandleNativeMessagesOptions)
  */
 function nativeMessagesErrorResponse(response: Response, bodyText: string, finishLog: FinishLog): Response {
   let upstreamType: string | undefined;
+  let upstreamCode: string | undefined;
   let upstreamMessage: string | undefined;
   try {
     const parsed = JSON.parse(bodyText) as Rec;
     const details = isRec(parsed.error) ? parsed.error : parsed;
     if (typeof details.type === "string") upstreamType = details.type;
+    if (typeof details.code === "string") upstreamCode = details.code;
     if (typeof details.message === "string" && details.message.trim()) {
       upstreamMessage = redactSecretString(details.message.trim());
     }
@@ -898,9 +926,17 @@ function nativeMessagesErrorResponse(response: Response, bodyText: string, finis
   if (retryAfter) headers.set("Retry-After", retryAfter);
   else if (transient) headers.set("Retry-After", "2");
   if (replayRefusal) applyReplayRefusalClientHeaders(headers);
+  // The native envelope keeps its shape (no added code); only an overflow's wording changes, so
+  // Claude Code recognizes it and compacts (010).
+  // Only a 400/413 refusal: the shared classifier also files a 429 "too many tokens per minute"
+  // under context_length_exceeded, and that one must stay a rate limit.
+  const overflow = !replayRefusal && (response.status === 400 || response.status === 413)
+    && !isThroughputLimitText(safeMessage)
+    && (classified.code === "context_length_exceeded" || upstreamCode === "context_length_exceeded"
+      || isContextOverflowText(safeMessage));
   const out = new Response(JSON.stringify(anthropicErrorBody(
     status,
-    safeMessage,
+    overflow ? claudePromptTooLongMessage(safeMessage) : safeMessage,
     transient ? "overloaded_error" : upstreamType,
     replayRefusal ? UPSTREAM_RESET_REPLAY_REFUSED_CODE : undefined,
   )), { status, headers });

package/src/server/messages-response-headers.ts

@@ -0,0 +1,14 @@
+/** Only an opaque Anthropic request id is carried; arbitrary upstream headers stay private. */
+export function upstreamMessagesRequestIdHeaders(upstream: Headers): Record<string, string> {
+  const requestId = upstream.get("request-id");
+  return requestId && /^req_[A-Za-z0-9_-]{1,128}$/.test(requestId)
+    ? { "request-id": requestId }
+    : {};
+}
+
+/** The response is newly constructed by the Messages handler; preserve its body and markers. */
+export function retainUpstreamMessagesRequestId(response: Response, upstream: Headers): Response {
+  const requestId = upstreamMessagesRequestIdHeaders(upstream)["request-id"];
+  if (requestId) response.headers.set("request-id", requestId);
+  return response;
+}

package/src/server/request-log-conversation.ts

@@ -37,6 +37,47 @@ export function normalizeLogConversationId(raw: string | undefined | null): stri
   return createHash("sha256").update(trimmed).digest("hex").slice(0, LOG_CONVERSATION_ID_LEN);
 }
 
+const CODEX_THREAD_LINK_PREFIX = "codex://threads/";
+/** Real thread links stay well under this bound; larger pastes skip unwrapping so per-row matching stays cheap. */
+const LOG_CONVERSATION_LINK_MAX = 512;
+
+/**
+ * Unwrap a pasted `codex://threads/<id>` deep link to the bare thread id. Codex's
+ * "copy session link" clipboard carries that URI form while clients send the bare id
+ * on the wire, so Logs search accepts either paste without changing what is persisted.
+ * Query/fragment metadata (e.g. `?hostId=…`) belongs to the link, not the id, and is
+ * dropped. The scan is a bounded linear walk — no ambiguous regex — because this runs
+ * per log row on every filter evaluation.
+ */
+export function unwrapLogConversationQuery(query: string): string {
+  const trimmed = query.trim();
+  if (
+    trimmed.length > LOG_CONVERSATION_LINK_MAX ||
+    trimmed.length <= CODEX_THREAD_LINK_PREFIX.length ||
+    !trimmed.toLowerCase().startsWith(CODEX_THREAD_LINK_PREFIX)
+  ) {
+    return trimmed;
+  }
+  const rest = trimmed.slice(CODEX_THREAD_LINK_PREFIX.length);
+  const metaIndex = rest.search(/[?#]/);
+  const segment = (metaIndex === -1 ? rest : rest.slice(0, metaIndex)).replace(/\/+$/, "");
+  // A thread id is a single path segment; deeper paths or embedded whitespace keep the
+  // whole paste as the query, so foreign `codex://` URIs still fail to match.
+  return segment !== "" && !/[/\s]/.test(segment) ? segment : trimmed;
+}
+
+/**
+ * The candidate ids a conversation query may legitimately match. Unwrapping a pasted
+ * `codex://threads/<id>` link must add a match path, not replace one: a client can
+ * literally send a `codex://threads/…` session id, which is hashed whole for storage,
+ * so the untouched paste stays a candidate too.
+ */
+export function logConversationQueryCandidates(query: string): string[] {
+  const trimmed = query.trim();
+  const unwrapped = unwrapLogConversationQuery(trimmed);
+  return unwrapped === trimmed ? [unwrapped] : [unwrapped, trimmed];
+}
+
 /**
  * Filter match: accept either the persisted digest or the original preimage
  * (so pasting from Logs detail or the client-facing session id both work).
@@ -46,11 +87,13 @@ export function matchesLogConversationId(
   query: string | undefined | null,
 ): boolean {
   if (!stored) return false;
-  const trimmed = typeof query === "string" ? query.trim() : "";
-  if (!trimmed) return false;
-  if (stored === trimmed) return true;
-  const hashed = normalizeLogConversationId(trimmed);
-  return hashed !== undefined && stored === hashed;
+  for (const candidate of logConversationQueryCandidates(typeof query === "string" ? query : "")) {
+    if (!candidate) continue;
+    if (stored === candidate) return true;
+    const hashed = normalizeLogConversationId(candidate);
+    if (hashed !== undefined && stored === hashed) return true;
+  }
+  return false;
 }
 
 /**

package/src/server/request-log.ts

@@ -232,6 +232,8 @@ export interface RequestLogContext {
    * message) extracted from a `response.failed` SSE payload or non-streaming error body, so the
    * request log / GUI shows the actual upstream failure rather than only the HTTP-mapped code. */
   upstreamError?: string;
+  upstreamErrorCode?: string;
+  upstreamRequestId?: string;
   /** HTTP status derived from a terminal `response.failed` SSE payload (429/401/503/etc.). */
   terminalHttpStatus?: number;
   /** Recognized structured terminal code whose exact identity must survive status mapping. */
@@ -342,6 +344,8 @@ export interface RequestLogEntry {
   closeReason?: "terminal" | "client_cancel" | "non_stream" | "body_stall" | "body_overflow";
   /** Secret-redacted upstream error reason, surfaced in /api/logs and the GUI detail modal. */
   upstreamError?: string;
+  upstreamErrorCode?: string;
+  upstreamRequestId?: string;
   usageStatus: UsageStatus;
   usage?: OcxUsage;
   totalTokens?: number;
@@ -1095,6 +1099,24 @@ export function inspectResponseLogSsePayloadParsed(
  * a non-streaming JSON error body. We keep the FIRST non-empty reason (the original failure) and
  * run it through redactSecretString so secrets never reach /api/logs. Pure; safe on any text.
  */
+const UPSTREAM_DIAGNOSTIC_TOKEN = /^[A-Za-z0-9_.:-]{1,128}$/;
+
+function noteBoundedDiagnostic(
+  logCtx: RequestLogContext,
+  field: "upstreamErrorCode" | "upstreamRequestId",
+  value: unknown,
+): void {
+  if (logCtx[field] !== undefined) return;
+  if (typeof value !== "string" || !UPSTREAM_DIAGNOSTIC_TOKEN.test(value)) return;
+  logCtx[field] = value;
+  const status = logCtx.terminalHttpStatus;
+  if (status === undefined || status >= 500) console.warn(`[opencodex] upstream failure${status ? ` status=${status}` : ""}${logCtx.upstreamErrorCode ? ` code=${logCtx.upstreamErrorCode}` : ""}${logCtx.upstreamRequestId ? ` request_id=${logCtx.upstreamRequestId}` : ""}`);
+}
+
+export function noteUpstreamRequestId(logCtx: RequestLogContext, headers: Headers): void {
+  noteBoundedDiagnostic(logCtx, "upstreamRequestId", headers.get("openai-request-id") ?? headers.get("x-request-id"));
+}
+
 function captureUpstreamError(logCtx: RequestLogContext, text: string | null): void {
   if (!text) return;
   let parsed: unknown | undefined;
@@ -1114,14 +1136,19 @@ function captureUpstreamErrorParsed(
   if (parsed !== undefined && parsed !== null) {
     const json = parsed as {
       type?: unknown;
-      error?: { message?: unknown };
-      last_error?: { message?: unknown };
+      error?: { message?: unknown; code?: unknown };
+      last_error?: { message?: unknown; code?: unknown };
       response?: {
         error?: { type?: unknown; code?: unknown; message?: unknown };
         incomplete_details?: { reason?: unknown; message?: unknown };
       };
     };
     captureTerminalHttpStatus(logCtx, json);
+    noteBoundedDiagnostic(
+      logCtx,
+      "upstreamErrorCode",
+      json.error?.code ?? json.response?.error?.code ?? json.last_error?.code,
+    );
     const reason = json?.response?.incomplete_details?.reason;
     if (json.type === "response.incomplete"
       && logCtx.terminalIncompleteReason === undefined
@@ -1593,6 +1620,8 @@ export function addFinalRequestLog(
     ...(meta?.terminalStatus ? { terminalStatus: meta.terminalStatus } : {}),
     ...(closeReason ? { closeReason } : {}),
     ...(logCtx.upstreamError ? { upstreamError: logCtx.upstreamError } : {}),
+    ...(logCtx.upstreamErrorCode ? { upstreamErrorCode: logCtx.upstreamErrorCode } : {}),
+    ...(logCtx.upstreamRequestId ? { upstreamRequestId: logCtx.upstreamRequestId } : {}),
     usageStatus,
     ...(loggedUsage ? { usage: loggedUsage } : {}),
     ...(totalTokens !== undefined ? { totalTokens } : {}),

package/src/server/responses/adapter-continuation.ts

@@ -401,7 +401,7 @@ export function createAdapterContinuations(
         }
       }
      if (
-       (response.status === 429 || response.status === 403)
+       (response.status === 429 || response.status === 403 || response.status === 401)
        && anthropicInstance
        && transportState.anthropicPoolAccountId
         && !isNonReplayableResponse(response)
@@ -428,7 +428,7 @@ export function createAdapterContinuations(
             );
             sealRequestAttemptIdentity(logCtx.activeAttempt, logCtx.provider, transportState.activeAdapter.name, logCtx.accountLogLabel);
             recordAttemptCredentialSource(logCtx.activeAttempt, route.providerName, route.provider, transportState.activeAdapter.name);
-            nextContinuationRecoveryKind = "anthropic-oauth-429";
+            nextContinuationRecoveryKind = response.status === 401 ? "oauth-401" : "anthropic-oauth-429";
             continue;
           } catch {
             // fall through to emit continuation error below

package/src/server/responses/adapter-dispatch.ts

@@ -170,6 +170,9 @@ export async function prepareAdapterExchange(
     | "workflowRootId"
     | "claimAmbiguousResend"
     | "sendsUsed"
+    | "targetSendsUsed"
+    | "initialSendAllowance"
+    | "noteInitialDispatch"
   >,
 ) {
   const { options, config, logCtx, req } = requestContext;
@@ -374,24 +377,33 @@ export async function prepareAdapterExchange(
       const compactPrepaid = options.compactionRecoveryAttempted ? sendBudgetState.pendingHopPermit : undefined;
       if (compactPrepaid) sendBudgetState.pendingHopPermit = undefined;
       let compactPrepaidUsed = false;
+      // A combo-owned or compaction-prepaid send is settled by the physical-dispatch receipt
+      // when spend enforcement is off. Enforced spend and direct requests keep the shared
+      // physical-send reporter, which already owns the permit lifecycle there.
+      // Stable per request: a Combo or compaction permit comes from `reserveDispatch`, which starts
+      // (and freezes) the spend policy first, so `spendEnforced` cannot change before dispatch.
+      const receiptMode = !sendBudgetState.adapterSendBudget?.spendEnforced
+        && Boolean(options.comboInitialSend || compactPrepaid);
       // Combo and emergency compaction admission already book this target's first send.
       // Its configured initial ceiling is target-local; the shared remainder below still
       // accounts for earlier targets without deducting their sends from this target twice.
       const initialSendCap = transientPolicy || resetPolicy
+        || (options.comboAttempt && route.provider.adapter === "google")
         ? transientSendCapFor(transientPolicy?.attempts,
           (options.comboAttempt || compactPrepaid) ? 0 : sendBudgetState.sendsUsed)
         : 1;
       const fetchWithRetryPolicy = (route.provider.adapter === "google" || transientPolicy)
         ? fetchWithTransientRetry
         : fetchWithResetRetry;
-      upstreamResponse = await fetchWithRetryPolicy(
+      try { upstreamResponse = await fetchWithRetryPolicy(
         recovery => {
           // Unconfigured generic sends may omit the counting reporter. Capture policy
           // independently so a later recovery still belongs to this request's start.
           sendBudgetState.adapterSendBudget?.startRequest?.({
             poolId: logCtx.spendPoolId ?? route.providerName, identityId: logCtx.accountLogLabel,
           });
-          if (compactPrepaid && !compactPrepaidUsed) {
+          // Receipt mode books the physical send at executor admission instead of here.
+          if (!receiptMode && compactPrepaid && !compactPrepaidUsed) {
             if (!compactPrepaid.use()) throw new SendBudgetExhaustedError(safeHostLabel(builtInitialRequest.url));
             compactPrepaidUsed = true;
           }
@@ -405,6 +417,14 @@ export async function prepareAdapterExchange(
               dispatchOverride: oauthDispatch(builtInitialRequest),
               providerName: route.providerName,
               modelId: route.modelId,
+              onPhysicalDispatch: () => {
+                // Preserve the unconfigured direct-adapter accounting contract; Combo admission
+                // and explicit retry/compaction budgets require a receipt for their booked send.
+                if (receiptMode) {
+                  sendBudgetState.noteInitialDispatch(compactPrepaid && !compactPrepaidUsed ? compactPrepaid : undefined);
+                  compactPrepaidUsed = true;
+                }
+              },
               beforeDispatch: isCanonicalOpenAiForwardProvider(route.provider)
                 ? createCodexAuthDispatchGuard(admissionState.authCtx, options.codexAuthPolicy ?? config, route.modelId, options.admission, options.visionDescribeTerminal === true) : undefined,
             }));
@@ -413,22 +433,24 @@ export async function prepareAdapterExchange(
           abortSignal: upstream.signal,
           label: safeHostLabel(builtInitialRequest.url),
           claimAmbiguousResend: claimPreHeaderResend,
-          ...(sendBudgetState.adapterSendBudget?.spendEnforced || options.comboDispatchPermit || transientPolicy || resetPolicy || compactPrepaid
+          ...(sendBudgetState.adapterSendBudget?.spendEnforced || options.comboDispatchPermit || transientPolicy || resetPolicy || compactPrepaid || options.comboAttempt
             ? {
               // A pending compaction permit already booked this leg's first physical send.
-              ...(transientPolicy || resetPolicy || compactPrepaid
+              ...(transientPolicy || resetPolicy || compactPrepaid || options.comboAttempt
                 ? {
                   attempts: Math.min(initialSendCap,
-                    remainingTransientSendBudget(initialSendCap) + (compactPrepaid ? 1 : 0)),
+                    sendBudgetState.initialSendAllowance(initialSendCap) + (compactPrepaid ? 1 : 0)),
                 }
                 : {}),
-              // Keep the exact combo or compaction receipt on the helper callback. The reporter
-              // also settles the request and workflow counters for reset-only retries.
-              onSendsConsumed: transientSendReporter(compactPrepaid ?? options.comboDispatchPermit),
+              // Reporter path: direct requests and enforced spend. Receipt-mode sends are
+              // settled by `onPhysicalDispatch`, so a reporter here would count them twice.
+              ...(receiptMode
+                ? {}
+                : { onSendsConsumed: transientSendReporter(compactPrepaid ?? options.comboDispatchPermit) }),
             }
             : {}),
         },
-      );
+      ); } finally { compactPrepaid?.release(); }
     }
   } catch (err) {
     cleanupUpstreamAbort();
@@ -609,8 +631,7 @@ export async function prepareAdapterExchange(
           const helperCountsSends = sendBudgetState.adapterSendBudget?.spendEnforced === true || refetchTransientPolicy !== null || resetReplayPolicyFor(route.provider) !== null;
           const prepaid = sendBudgetState.pendingHopPermit;
           const configuredTotal = refetchTransientPolicy?.attempts;
-          const refetchCap = transientSendCapFor(configuredTotal,
-            sendBudgetState.sendsUsed - (prepaid ? 1 : 0));
+          const refetchCap = transientSendCapFor(configuredTotal, sendBudgetState.targetSendsUsed);
           // An exact total includes a booked hop, even when it consumed the last base slot.
           // Keep that funded send while forbidding the final reserve from widening the total.
           const prepaidLastSlot = helperCountsSends && configuredTotal !== undefined && prepaid
@@ -1076,7 +1097,7 @@ export async function prepareAdapterExchange(
       // Anthropic OAuth: recover a rate limit or proven account entitlement refusal
       // before output, within the shared request and account rotation limits.
       while (
-        (upstreamResponse.status === 429 || upstreamResponse.status === 403)
+        (upstreamResponse.status === 429 || upstreamResponse.status === 403 || upstreamResponse.status === 401)
         && anthropicInstance
         && transportState.anthropicPoolAccountId
       ) {
@@ -1101,7 +1122,7 @@ export async function prepareAdapterExchange(
           );
           sealRequestAttemptIdentity(logCtx.activeAttempt, logCtx.provider, transportState.activeAdapter.name, logCtx.accountLogLabel);
           recordAttemptCredentialSource(logCtx.activeAttempt, route.providerName, route.provider, transportState.activeAdapter.name);
-         const result = await rebuildAndRefetch("anthropic-oauth-429");
+         const result = await rebuildAndRefetch(upstreamResponse.status === 401 ? "oauth-401" : "anthropic-oauth-429");
          if ("failed" in result) return result.failed;
          upstreamResponse = result;
           if (isNonReplayableResponse(upstreamResponse)) continue recovery;

package/src/server/responses/codex-ws-exchange.ts

@@ -9,7 +9,7 @@ import { CODEX_RESPONSES_HTTP_URL, type PreparedCodexWsRequest } from "./codex-w
 import { CodexWsCorrelation } from "./codex-ws-correlation";
 import type { CodexWsSession } from "./codex-ws-session";
 import { UPGRADE_DEADLINE_MS, CODEX_WS_LIVENESS_PING_INTERVAL_MS, CODEX_WS_RESPONSE_PRELUDE_TIMEOUT_MS, MAX_CODEX_WS_FRAME_BYTES,
-  MAX_CODEX_WS_QUEUE_BYTES, markCodexWsResponse, normalizeResponsesWsRelayEvent, closedBeforeTerminalMessage,
+  MAX_CODEX_WS_QUEUE_BYTES, markCodexWsResponse, markCodexWsPreludeProjection, normalizeResponsesWsRelayEvent, closedBeforeTerminalMessage,
   codexWsCreateFrameExceedsLimit, codexWsFailureDetail, codexWsPreResponseFailure, markCodexWsStage, codexWsOcxVersion,
   markCodexWsSocketDeath, type CodexWsFailureStage, type CodexWsStageRecord } from "./codex-ws-wire";
 
@@ -23,6 +23,7 @@ interface ExchangeOptions {
   sseFallback: typeof globalThis.fetch;
   onQuota?: CodexWsQuotaObserver;
   beforeDispatch?: (headers: Headers) => void;
+  onPhysicalDispatch?: () => void;
   /** Bun version string the caller gated on; stamped onto the stage record. */
   bunVersion?: string;
 }
@@ -44,7 +45,8 @@ function rejectionHeaders(source: Record<string, unknown>, prelude: Headers): He
     }
   }
   // Reuse the metadata owner's count/value/family budgets and window freshness
-  // rules, without publishing quota twice. The unmarked HTTP response owns it.
+  // rules, without publishing quota twice. Pool bookkeeping consumes the HTTP
+  // projection; plain-main publication belongs only to the WS observer.
   const projected = new CodexWsMetadata();
   try {
     for (const values of [Object.fromEntries(prelude), source]) {
@@ -240,6 +242,7 @@ export function codexWsExchange(options: ExchangeOptions): Promise<Response> {
         // are left out: their channel may already have sent continuation frames on this socket, so
         // the create frame alone no longer describes the turn.
         if (socketDied && !nativeControl) markCodexWsSocketDeath(failureResponse, stage);
+        markCodexWsPreludeProjection(failureResponse);
         resolve(failureResponse);
         return;
       }
@@ -411,6 +414,17 @@ export function codexWsExchange(options: ExchangeOptions): Promise<Response> {
         failStream(error);
         return;
       }
+      try { options.onPhysicalDispatch?.(); } catch (error) {
+        // Admission refused before the frame left. This is a pre-open rejection,
+        // not a transport failure and not an authorization to replay over SSE.
+        settledPreOpen = true;
+        sent = false;
+        terminal = true;
+        cleanup();
+        session.dispose();
+        reject(error);
+        return;
+      }
       try {
         ws.send(frameText);
         sentAt = Date.now();
@@ -501,6 +515,7 @@ export function codexWsExchange(options: ExchangeOptions): Promise<Response> {
             cleanup();
             try { controller.close(); } catch { /* unused stream already closed */ }
             session.dispose();
+            markCodexWsPreludeProjection(rejection);
             resolve(rejection);
             return;
           }