@bitkyc08/opencodex 2.61.0 → 2.63.0-preview.20260923

package/src/server/responses/request-prepare.ts

@@ -20,6 +20,7 @@ import {
   sessionIdHeaderFromRequest,
   reasoningReplayConversationIdFromResponsesRequest,
 } from "../request-log-conversation";
+import { resolveContextPrincipal } from "../auth-cors";
 import {
   isShadowSourceModel,
   shadowSourceModelPrefix,
@@ -407,6 +408,17 @@ export async function prepareResponsesRequest(
       parsed._reasoningReplayScope = { clientThreadId: reasoningReplayConversationId };
     }
   }
+  if (parsed._reasoningReplayScope) {
+    // Scope replay cells to the caller principal. On loopback, admission carries no identity,
+    // so resolve it from an opencodex API key the caller volunteered (same rule as context
+    // history ownership). A caller that presents none has no principal, and none is invented:
+    // every keyless local process would otherwise share one bucket, and a client-visible cell id
+    // would become enough to read another caller's retained search result. Without a principal
+    // bridgeSearchReplayScope yields no scope, so nothing is recorded or restored for it. The
+    // field is always rewritten so an absent principal also clears one a reused holder carried.
+    const clientPrincipalId = resolveContextPrincipal(req, config, options.admission);
+    parsed._reasoningReplayScope = { ...parsed._reasoningReplayScope, clientPrincipalId };
+  }
   // Prefer a pre-populated id (routed Claude) over Responses headers that may be
   // absent or synthetically injected (session_id from prompt_cache_key).
   if (!logCtx.conversationId) {

package/src/server/responses/request-transport.ts

@@ -445,6 +445,11 @@ export async function prepareResponsesTransport(
           return response;
         }
         const nextAdapter = await refreshDispatchAdapter(requestParsed);
+        // Rebind before rebuilding: the rebuild's bridged-search restore and continuation
+        // restore key on the serving identity, which must be the refreshed route's, not the
+        // credential whose selection just lapsed.
+        bindRouteReasoningReplayScope({ parsed: requestParsed, providerName: route.providerName, provider: route.provider,
+          adapterName: nextAdapter.name, oauthCredentialSnapshot: replayOAuthCredentialSnapshot });
         const rebuilt = await nextAdapter.buildRequest(requestParsed, {
           headers: requestState.selectedForwardHeaders, translatorBudget,
           ...(imageTierBias > 0 ? { imageTierBias } : {}),
@@ -467,8 +472,6 @@ export async function prepareResponsesTransport(
         sameTargetToken = transportToken;
         destination = rebuilt.url;
         dispatchInit = { ...dispatchInit, method: rebuilt.method, headers, body: rebuilt.body };
-        bindRouteReasoningReplayScope({ parsed: requestParsed, providerName: route.providerName, provider: route.provider,
-          adapterName: nextAdapter.name, oauthCredentialSnapshot: replayOAuthCredentialSnapshot });
         // The next iteration validates synchronously and calls fetch in that same turn.
       }
       throw new Error("OAuth account selection changed repeatedly before dispatch");

package/src/server/responses/ws-upstream.ts

@@ -138,7 +138,7 @@ export function codexWsUpstreamFetch(
   // Never infer backend support from a model name or enable controls on a gateway.
   const control = nativeControl?.kind === "injection"
     ? ((prepared.canonical || url === OPENAI_API_RESPONSES_URL) && isInjectionRequest(JSON.parse(frameText)) ? nativeControl : undefined)
-    : (prepared.canonical || url === OPENAI_API_RESPONSES_URL) ? nativeControl : undefined;
+    : prepared.canonical ? nativeControl : undefined;
   if (control?.kind === "injection" && url === OPENAI_API_RESPONSES_URL) {
     const beta = headers["openai-beta"];
     if (!beta?.split(",").some(value => value.trim() === "responses_multi_agent=v1")) {

package/src/server/ws-bridge.ts

@@ -228,6 +228,22 @@ function sendProtocolError(ws: ServerWebSocket<WsData>, status: number, message:
   sendJsonFrame(ws, buildWsErrorFrame(status, protocolError(message)));
 }
 
+/**
+ * Report an upstream-pump failure to the client. Errors that carry a structured
+ * code (for example the undeclared-tool guard's undeclared_tool_call) keep it so
+ * clients see the same rejection identity as the SSE path; everything else stays
+ * a generic protocol error.
+ */
+function sendUpstreamError(ws: ServerWebSocket<WsData>, status: number, err: unknown): void {
+  const code = err != null && typeof (err as { code?: unknown }).code === "string"
+    ? (err as { code: string }).code
+    : undefined;
+  const message = err instanceof Error ? err.message : String(err);
+  sendJsonFrame(ws, buildWsErrorFrame(status, code
+    ? { type: "upstream_error", code, message }
+    : protocolError(message)));
+}
+
 export async function pumpResponsesSseToWebSocket(
   ws: ServerWebSocket<WsData>,
   sseStream: ReadableStream<Uint8Array>,
@@ -319,7 +335,7 @@ export async function pumpResponsesSseToWebSocket(
       && !(err instanceof WsSendDroppedError)) {
       reportTerminal("incomplete");
       try {
-        sendProtocolError(ws, 502, err instanceof Error ? err.message : String(err));
+        sendUpstreamError(ws, 502, err);
       } catch (sendErr) {
         // If delivery is already dropped, there is no useful error frame left
         // to send. Swallow only that expected transport signal; other failures

package/src/types/request.ts

@@ -30,6 +30,11 @@ export interface OcxReasoningReplayIdentity {
  * the holder, so late tool-call cache writes see the active physical identity.
  */
 export interface OcxReasoningReplayScopeRef {
+  /**
+   * Process-local caller principal from resolveContextPrincipal. Absent when the caller presented
+   * no identity (keyless loopback); replay state keyed by it then fails closed.
+   */
+  readonly clientPrincipalId?: string;
   /**
    * Conversation namespace for replay state. Historically this was always the Codex parent-thread
    * id; headerless Responses callers use a raw sanitized thread/Cursor/session fallback, never the

package/src/usage/expected-prices.ts

@@ -44,6 +44,13 @@ const GEMINI_31_PRO: Cost4 = { input: 2, output: 12, cacheRead: 0.2, cacheWrite:
 const GPT56_SOL: Cost4 = { input: 4, output: 20, cacheRead: 0.4, cacheWrite: 5 };
 const GPT6_ASTRA: Cost4 = { input: 10, output: 50, cacheRead: 1, cacheWrite: 12.5 };
 const ASTRA_API_PRICING = "https://developers.openai.com/api/docs/models/gpt-6-astra";
+/**
+ * GPT-6 Sol and Luna API list prices (released 2026-09-22; the changelog publishes input, cached
+ * input and output). Cache write follows the 1.25x-input convention every OpenAI row here uses.
+ */
+const GPT6_SOL: Cost4 = { input: 2, output: 10, cacheRead: 0.2, cacheWrite: 2.5 };
+const GPT6_LUNA: Cost4 = { input: 0.1, output: 0.5, cacheRead: 0.01, cacheWrite: 0.125 };
+const GPT6_API_PRICING = "https://developers.openai.com/api/docs/changelog (2026-09-22: GPT-6 Sol $2 / $0.20 cached / $10; GPT-6 Luna $0.10 / $0.01 cached / $0.50)";
 /**
  * Daybreak aliases. `daybreak-*-latest` never appears in the pricing table itself — only its
  * current snapshot does — so these tuples are the snapshot's published rates and carry
@@ -101,12 +108,17 @@ const CLAUDE_OPUS_46: Cost4 = { input: 5, output: 25, cacheRead: 0.5, cacheWrite
 // (0.25) on Fable 5.1 — NOT the 0.1x (1.00) that Fable 5 and every other family use;
 // the pricing page footnote calls this out explicitly. Verified 2026-09-02.
 const CLAUDE_FABLE_51: Cost4 = { input: 10, output: 50, cacheRead: 0.25, cacheWrite: 12.5 };
-// Opus 5 is priced from the maintainer's confirmation that it matches the previous
-// Opus, not from a published Opus 5 page. Hence `verified-derived`, and a source
-// string that states the provenance instead of pointing at ANTHROPIC_PRICING.
-const CLAUDE_OPUS_5_DERIVED_SOURCE =
-  "user-confirmed: claude-opus-5 matches Claude Opus 4.6; no separate Anthropic Opus 5 price page verified";
+// Opus 5 was first priced from the maintainer's confirmation that it matched Opus 4.6. The
+// pricing page now lists it at that same 5 / 25 / 0.50 / 6.25 tuple (re-verified 2026-09-23).
+const CLAUDE_OPUS_5 = CLAUDE_OPUS_46;
+// Claude Opus 5.5 (claude-opus-5-5, released 2026-09-22): 4 / 20, 5m cache write 5.00. Cache
+// hits are 0.05x base input (0.20), a model-specific footnote on the pricing page, NOT the
+// 0.1x most families use. 1M context and 128K output at one flat rate (no long-context tier).
+const CLAUDE_OPUS_55: Cost4 = { input: 4, output: 20, cacheRead: 0.2, cacheWrite: 5 };
 const ANTHROPIC_PRICING = "https://platform.claude.com/docs/en/about-claude/pricing (official; 5m cache-write tier)";
+const CLAUDE_OPUS_5_SOURCE = `anthropic official Claude Opus 5 ${ANTHROPIC_PRICING}`;
+const CLAUDE_OPUS_55_SOURCE = `anthropic official Claude Opus 5.5 ${ANTHROPIC_PRICING}; cache hit = 0.05x base input`;
+const CURSOR_OPUS_55_PRICING = "https://cursor.com/docs/models/claude-opus-5-5 (Cursor Other Models pool; same list rate as Anthropic, Fast Mode billed separately)";
 
 const GEMINI_PRICING = "https://ai.google.dev/gemini-api/docs/pricing (2026-07-22); cacheWrite=0: storage is billed per-hour, not per-token";
 const GEMINI_37_PRICING = "https://ai.google.dev/gemini-api/docs/pricing (2026-08-14); promotional rate through 2026-12-31, rises to 1.50/7.50 on 2027-01-01; cacheWrite=0: storage is billed per-hour, not per-token";
@@ -190,6 +202,10 @@ export const EXPECTED_PRICE_OVERLAYS: readonly ExpectedPriceOverlay[] = [
   { provider: "openai-apikey", modelId: "gpt-6-astra", cost4: GPT6_ASTRA, source: ASTRA_API_PRICING, verifiedAt: "2026-09-05", status: "verified" },
   // Display estimates use API prices for both login and API-key routes, including cache writes.
   { provider: "openai", modelId: "gpt-6-astra", cost4: GPT6_ASTRA, source: `API-reference comparison estimate: ${ASTRA_API_PRICING}`, verifiedAt: "2026-09-05", status: "verified-derived" },
+  { provider: "openai-apikey", modelId: "gpt-6-sol", cost4: GPT6_SOL, source: GPT6_API_PRICING, verifiedAt: "2026-09-23", status: "verified" },
+  { provider: "openai-apikey", modelId: "gpt-6-luna", cost4: GPT6_LUNA, source: GPT6_API_PRICING, verifiedAt: "2026-09-23", status: "verified" },
+  { provider: "openai", modelId: "gpt-6-sol", cost4: GPT6_SOL, source: `API-reference comparison estimate: ${GPT6_API_PRICING}`, verifiedAt: "2026-09-23", status: "verified-derived" },
+  { provider: "openai", modelId: "gpt-6-luna", cost4: GPT6_LUNA, source: `API-reference comparison estimate: ${GPT6_API_PRICING}`, verifiedAt: "2026-09-23", status: "verified-derived" },
   // claude-fable-5-1 now HAS a generated jawcode row, so the two Anthropic surfaces resolve
   // from it and these overlays are the fallback rather than the primary source. They stay:
   // the overlay lookup is keyed by the configured provider id, so an account-pool log label
@@ -203,9 +219,15 @@ export const EXPECTED_PRICE_OVERLAYS: readonly ExpectedPriceOverlay[] = [
   // cost resolution returned null and the Logs `~$` column rendered an em dash. The
   // model-level vendor fallback only searches jawcode metadata, never overlays, so one
   // anthropic row would not cover cursor/kiro — each exposing provider needs its own.
-  { provider: "anthropic", modelId: "claude-opus-5", cost4: CLAUDE_OPUS_46, source: CLAUDE_OPUS_5_DERIVED_SOURCE, verifiedAt: "2026-07-25", status: "verified-derived" },
-  { provider: "cursor", modelId: "claude-opus-5", cost4: CLAUDE_OPUS_46, source: CLAUDE_OPUS_5_DERIVED_SOURCE, verifiedAt: "2026-07-25", status: "verified-derived" },
-  { provider: "kiro", modelId: "claude-opus-5", cost4: CLAUDE_OPUS_46, source: CLAUDE_OPUS_5_DERIVED_SOURCE, verifiedAt: "2026-07-25", status: "verified-derived" },
+  { provider: "anthropic", modelId: "claude-opus-5", cost4: CLAUDE_OPUS_5, source: CLAUDE_OPUS_5_SOURCE, verifiedAt: "2026-09-23", status: "verified" },
+  { provider: "cursor", modelId: "claude-opus-5", cost4: CLAUDE_OPUS_5, source: `${CLAUDE_OPUS_5_SOURCE}; vendor list price applied to the Cursor surface`, verifiedAt: "2026-09-23", status: "verified-derived" },
+  { provider: "kiro", modelId: "claude-opus-5", cost4: CLAUDE_OPUS_5, source: `${CLAUDE_OPUS_5_SOURCE}; vendor list price applied to the Kiro credit surface`, verifiedAt: "2026-09-23", status: "verified-derived" },
+  // Claude Opus 5.5. The anthropic bundle row wins for the bare provider id; these overlays
+  // cover account-label namespaces. Cursor publishes the same list rate on its own model page.
+  { provider: "anthropic", modelId: "claude-opus-5-5", cost4: CLAUDE_OPUS_55, source: CLAUDE_OPUS_55_SOURCE, verifiedAt: "2026-09-23", status: "verified" },
+  { provider: "anthropic-apikey", modelId: "claude-opus-5-5", cost4: CLAUDE_OPUS_55, source: CLAUDE_OPUS_55_SOURCE, verifiedAt: "2026-09-23", status: "verified" },
+  // Cursor canonicalizes every Opus 5.5 spelling (thinking/effort/fast suffixes) onto this row.
+  { provider: "cursor", modelId: "claude-opus-5-5", cost4: CLAUDE_OPUS_55, source: CURSOR_OPUS_55_PRICING, verifiedAt: "2026-09-23", status: "verified" },
   // MiniMax M2.1 highspeed — published PAYG price (verified).
   { provider: "minimax", modelId: "MiniMax-M2.1-highspeed", cost4: MINIMAX_M21_HIGHSPEED, source: MINIMAX_PRICING, verifiedAt: "2026-07-20", status: "verified" },
   { provider: "minimax-cn", modelId: "MiniMax-M2.1-highspeed", cost4: MINIMAX_M21_HIGHSPEED, source: MINIMAX_PRICING, verifiedAt: "2026-07-20", status: "verified" },
@@ -371,7 +393,10 @@ export const EXPECTED_PRICE_OVERLAYS: readonly ExpectedPriceOverlay[] = [
   { provider: "devin-cli", modelId: "swe-1-6", cost4: DEVIN_SWE_17, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
   { provider: "devin-cli", modelId: "gpt-5-6-sol", cost4: GPT56_SOL, source: `enterprise list column (self-serve shows discounted 1.2/6); ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
   { provider: "devin-cli", modelId: "gpt-6-astra", cost4: GPT6_ASTRA, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "gpt-6-sol", cost4: GPT6_SOL, source: `derived: GPT-6 Sol/Luna added 2026-09-23 ahead of Devin's modelCostData table; OpenAI API list price ${GPT6_API_PRICING}`, verifiedAt: "2026-09-23", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "gpt-6-luna", cost4: GPT6_LUNA, source: `derived: GPT-6 Sol/Luna added 2026-09-23 ahead of Devin's modelCostData table; OpenAI API list price ${GPT6_API_PRICING}`, verifiedAt: "2026-09-23", status: "verified-derived" },
   { provider: "devin-cli", modelId: "claude-opus-5", cost4: CLAUDE_OPUS_46, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin-cli", modelId: "claude-opus-5-5", cost4: CLAUDE_OPUS_55, source: `derived: live Devin catalog lists claude-opus-5-5 but Devin's modelCostData table does not yet; Anthropic list price shown as estimate ${ANTHROPIC_PRICING}`, verifiedAt: "2026-09-23", status: "verified-derived" },
   { provider: "devin-cli", modelId: "claude-fable-5-1", cost4: CLAUDE_FABLE_51, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
   { provider: "devin-cli", modelId: "claude-sonnet-5", cost4: DEVIN_SONNET_5, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
   { provider: "devin-cli", modelId: "glm-5-3", cost4: GLM_53, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
@@ -385,7 +410,10 @@ export const EXPECTED_PRICE_OVERLAYS: readonly ExpectedPriceOverlay[] = [
   { provider: "devin", modelId: "gpt-5-6-sol", cost4: GPT56_SOL, source: `enterprise list column (self-serve shows discounted 1.2/6); ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
   { provider: "devin", modelId: "gpt-5-6-luna", cost4: GPT56_LUNA, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
   { provider: "devin", modelId: "gpt-5-6-terra", cost4: GPT56_TERRA, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "gpt-6-sol", cost4: GPT6_SOL, source: `derived: GPT-6 Sol/Luna added 2026-09-23 ahead of Devin's modelCostData table; OpenAI API list price ${GPT6_API_PRICING}`, verifiedAt: "2026-09-23", status: "verified-derived" },
+  { provider: "devin", modelId: "gpt-6-luna", cost4: GPT6_LUNA, source: `derived: GPT-6 Sol/Luna added 2026-09-23 ahead of Devin's modelCostData table; OpenAI API list price ${GPT6_API_PRICING}`, verifiedAt: "2026-09-23", status: "verified-derived" },
   { provider: "devin", modelId: "claude-opus-4-8", cost4: CLAUDE_OPUS_46, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
+  { provider: "devin", modelId: "claude-opus-5-5", cost4: CLAUDE_OPUS_55, source: `derived: live Devin catalog lists claude-opus-5-5 but Devin's modelCostData table does not yet; Anthropic list price shown as estimate ${ANTHROPIC_PRICING}`, verifiedAt: "2026-09-23", status: "verified-derived" },
   { provider: "devin", modelId: "claude-fable-5-1", cost4: CLAUDE_FABLE_51, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
   { provider: "devin", modelId: "claude-sonnet-5", cost4: DEVIN_SONNET_5, source: DEVIN_PRICING, verifiedAt: "2026-09-13", status: "verified-derived" },
   { provider: "devin", modelId: "glm-5-2", cost4: GLM_52, source: `enterprise list column (self-serve shows an unannounced 0 promo); ${DEVIN_PRICING}`, verifiedAt: "2026-09-13", status: "verified-derived" },
@@ -563,6 +591,8 @@ const UNIFORM_DOUBLE: Cost4 = { input: 2, output: 2, cacheRead: 2, cacheWrite: 2
 const OPENAI_PRICING_DOC = "https://developers.openai.com/api/docs/pricing";
 const OPENAI_CONTEXT_MODELS = [
   "gpt-6-astra",
+  "gpt-6-sol",
+  "gpt-6-luna",
   "gpt-5.6-sol",
   "gpt-5.6-terra",
   "gpt-5.6-luna",

package/src/web-search/executor.ts

@@ -48,13 +48,17 @@ export type SidecarOutcome = WebSearchResult & { error?: string };
  *
  * The forward backend throttles burst sidecar traffic, and without a replay the 429 becomes a
  * failed tool result that poisons the query for the whole turn (see failedQueries in loop.ts).
- * 1 initial send + 2 replays; Retry-After is honored as a lower bound and capped by
- * RETRY_AFTER_CEILING_MS (an instruction past the ceiling ends with the 429 instead of
- * parking the search). Each wait releases the unread 429 body first so sockets do not
- * accumulate under a rate-limit storm. Abort or timeout ends the wait through the existing
- * catch, exactly like an abort during the SSE parse.
+ * 1 initial send + 2 replays, counted as physical sends: connection-reset recovery inside each
+ * send draws from the same SIDECAR_MAX_SENDS budget, so the two layers cannot multiply into nine
+ * paid requests during a degraded period. Retry-After is honored as a lower bound and capped by
+ * RETRY_AFTER_CEILING_MS and the remaining sidecar deadline (an instruction past either
+ * ends with the 429 instead of parking the search). Each wait releases the unread 429 body first so sockets do not
+ * accumulate under a rate-limit storm. The release itself may take up to a second, so a
+ * deadline landing during release or backoff ends with the 429 already in hand rather than
+ * a timeout; a caller abort still ends the wait through the shared catch, exactly like an
+ * abort during the SSE parse. An exhausted budget likewise ends with the 429 in hand.
  */
-const SIDECAR_429_MAX_ATTEMPTS = 3;
+const SIDECAR_MAX_SENDS = 3;
 const SIDECAR_429_BASE_DELAY_MS = 1_000;
 const SIDECAR_429_MAX_DELAY_MS = 10_000;
 
@@ -98,10 +102,14 @@ export async function runWebSearch(
     stream: true,
   };
   const url = `${forwardProvider.baseUrl}/responses`;
+  // t0 precedes the deadline timer's start so the remaining-time check stays conservative.
+  const t0 = Date.now();
   const linkedSignal = signalWithTimeout(settings.timeoutMs, abortSignal);
   const sidecarExit = sidecarEnter("web-search");
-  const t0 = Date.now();
   try {
+    // One physical-send budget for the whole search. Each helper call receives only what is left
+    // and reports every send it makes, reset retries included.
+    let sendsLeft = SIDECAR_MAX_SENDS;
     const sendOnce = () => fetchWithResetRetry(
       // Recovery nests INSIDE the version helper: applyUpstreamRecoveryInit then always receives a
       // defined init, and withUpstreamHttpVersion spreads the result, so `protocol` and the
@@ -117,10 +125,18 @@ export async function runWebSearch(
         // `session_id`, and `x-codex-turn-metadata` to the redirect target.
         redirect: "manual",
       }, recovery), forwardProvider)),
-      { replaySafe: true, abortSignal: linkedSignal.signal, label: "web-search-sidecar" },
+      {
+        replaySafe: true,
+        abortSignal: linkedSignal.signal,
+        label: "web-search-sidecar",
+        attempts: sendsLeft,
+        onSendsConsumed: sends => { sendsLeft -= sends; },
+      },
     );
     let res = await sendOnce();
-    for (let attempt = 0; res.status === 429 && attempt + 1 < SIDECAR_429_MAX_ATTEMPTS; attempt++) {
+    // Checked before the 429 body is released: a budget found spent after the release could only
+    // end in a send-budget error, recorded as a connection failure instead of the quota evidence.
+    for (let attempt = 0; res.status === 429 && sendsLeft > 0; attempt++) {
       const delay = retryBackoffDelayMs(attempt, {
         baseDelayMs: SIDECAR_429_BASE_DELAY_MS,
         maxDelayMs: SIDECAR_429_MAX_DELAY_MS,
@@ -129,10 +145,19 @@ export async function runWebSearch(
       });
       // A deadline, not a clamp: an instruction past the ceiling ends the search with the
       // 429 instead of parking it at a provider that already said it would refuse.
-      if (delay > RETRY_AFTER_CEILING_MS) break;
-      console.warn(`[web-search] sidecar HTTP 429 — retrying (${attempt + 2}/${SIDECAR_429_MAX_ATTEMPTS}) after ${delay}ms`);
-      await releaseResponseBodyBestEffort(res.body, linkedSignal.signal);
-      await sleepWithAbort(delay, linkedSignal.signal);
+      if (delay > RETRY_AFTER_CEILING_MS || delay >= settings.timeoutMs - (Date.now() - t0)) break;
+      console.warn(`[web-search] sidecar HTTP 429 — retrying (send ${SIDECAR_MAX_SENDS - sendsLeft + 1}/${SIDECAR_MAX_SENDS}) after ${delay}ms`);
+      try {
+        await releaseResponseBodyBestEffort(res.body, linkedSignal.signal);
+        await sleepWithAbort(delay, linkedSignal.signal);
+      } catch (e) {
+        // The release above may consume up to 1s, so the sidecar deadline can land during
+        // cleanup or mid-backoff — before the replay is dispatched. The observed 429 is
+        // already in hand: end with it rather than laundering it into a timeout. A caller
+        // abort (or a non-deadline throw) still propagates to the shared catch below.
+        if (!linkedSignal.signal.aborted || linkedSignal.signal.reason === abortSignal?.reason) throw e;
+        break;
+      }
       res = await sendOnce();
     }
     // Attach the body guard before ANY branch reads it. The success path guarded itself below,