@@ -6522,6 +6522,59 @@ describe("ClaudeAdapterV2 background wake turns", () => {
     ),
   );
 
+  it.effect.each([
+    { cooldown: " (reset after 1h2m3s)", seconds: 3723 },
+    { cooldown: " (reset after 4m 6s) (last error: rate limited)", seconds: 246 },
+    { cooldown: "", seconds: 0 },
+  ])("keeps a gateway's reset on a direct 429 with cooldown '$cooldown'", (input) =>
+    Effect.gen(function* () {
+      const harness = yield* makeWakeHarness;
+      const now = yield* DateTime.now;
+      yield* harness.runtime.startTurn(
+        makeClaudeTestTurnInput({
+          threadId: harness.threadId,
+          providerThread: harness.providerThread,
+          now,
+          attemptId: RunAttemptId.make("attempt-direct-gateway-limit"),
+          text: "Continue.",
+          attachments: [],
+        }),
+      );
+      // An API-key session reports a gateway 429 only as text, with no rate_limit_info.
+      const result = `API Error: Request rejected (429) · All credentials for model claude-opus-5-5 are cooling down${input.cooldown}`;
+      yield* Queue.offerAll(harness.sdkMessages, [
+        makeAssistantErrorFrame({
+          uuid: "00000000-0000-4000-8000-000000000654",
+          error: "rate_limit",
+          text: result,
+        }),
+        makeResultFrame({
+          uuid: "00000000-0000-4000-8000-000000000655",
+          result,
+          terminalReason: "api_error",
+          isError: true,
+          apiErrorStatus: 429,
+        }),
+      ]);
+      const terminal = yield* Queue.take(harness.terminalReceipts);
+      assert.equal(terminal.status, "failed");
+      if (terminal.status !== "failed") return;
+      assert.equal(terminal.failure.class, "usage_limit");
+      assert.equal(
+        terminal.failure.resetAt,
+        input.seconds > 0
+          ? DateTime.formatIso(
+              DateTime.makeUnsafe(DateTime.toEpochMillis(now) + input.seconds * 1_000),
+            )
+          : null,
+      );
+    }).pipe(
+      Effect.provide(
+        Layer.mergeAll(IdAllocator.layer, McpProviderSessions.layer, NodeServices.layer),
+      ),
+    ),
+  );
+
   it.effect("keeps the measured context when a usage limit ends the turn", () =>
     Effect.gen(function* () {
       const harness = yield* makeWakeHarness;
 
@@ -3760,37 +3760,46 @@ function claudeUsageLimitResetAt(resetsAt: unknown): string | null {
     : null;
 }
 
+function gatewayCooldownResetAt(
+  cooldown: RegExpExecArray | null,
+  completedAt: DateTime.Utc,
+): string | undefined {
+  if (cooldown === null) return undefined;
+  const seconds =
+    Number(cooldown[1] ?? 0) * 3_600 + Number(cooldown[2] ?? 0) * 60 + Number(cooldown[3] ?? 0);
+  const resetMs = DateTime.toEpochMillis(completedAt) + seconds * 1_000;
+  return seconds > 0 && resetMs < 8.64e15
+    ? DateTime.formatIso(DateTime.makeUnsafe(resetMs))
+    : undefined;
+}
+
 function providerFailureFromResult(
   message: SDKResultMessage,
   completedAt: DateTime.Utc,
   failureHint?: string,
   usageLimited = false,
 ): OrchestrationV2ProviderFailure | null {
-  const gatewayLimit =
+  const gatewayResetAt =
     failureHint === undefined &&
     message.subtype === "success" &&
     message.is_error &&
     message.terminal_reason === "api_error" &&
     message.api_error_status === 503
-      ? /^API Error: 503 \[[^\]\r\n]+\] \[429\]: \{"type":"error","error":\{"type":"rate_limit_error",[^\r\n]*\} \(reset after (?:(\d+)h ?)?(?:(\d+)m ?)?(?:(\d+)s)?\)/.exec(
-          message.result,
+      ? gatewayCooldownResetAt(
+          /^API Error: 503 \[[^\]\r\n]+\] \[429\]: \{"type":"error","error":\{"type":"rate_limit_error",[^\r\n]*\} \(reset after (?:(\d+)h ?)?(?:(\d+)m ?)?(?:(\d+)s)?\)/.exec(
+            message.result,
+          ),
+          completedAt,
         )
-      : null;
-  if (gatewayLimit) {
-    const seconds =
-      Number(gatewayLimit[1] ?? 0) * 3_600 +
-      Number(gatewayLimit[2] ?? 0) * 60 +
-      Number(gatewayLimit[3] ?? 0);
-    const resetMs = DateTime.toEpochMillis(completedAt) + seconds * 1_000;
-    if (seconds > 0 && resetMs < 8.64e15) {
-      return makeProviderFailure({
-        class: "usage_limit",
-        message: "Claude API rate limit reached. Try again later.",
-        code: "api_error_429",
-        retryable: true,
-        resetAt: DateTime.formatIso(DateTime.makeUnsafe(resetMs)),
-      });
-    }
+      : undefined;
+  if (gatewayResetAt !== undefined) {
+    return makeProviderFailure({
+      class: "usage_limit",
+      message: "Claude API rate limit reached. Try again later.",
+      code: "api_error_429",
+      retryable: true,
+      resetAt: gatewayResetAt,
+    });
   }
   const failureClass =
     (message.subtype === "success" && message.api_error_status === 429) || usageLimited
@@ -3819,6 +3828,14 @@ function providerFailureFromResult(
     return null;
   }
   const apiErrorStatus = message.api_error_status ?? null;
+  // API-key sessions get no rate_limit_info, so a gateway's 429 carries its reset only as text.
+  const resetAt =
+    apiErrorStatus === 429 && message.result.startsWith("API Error:")
+      ? gatewayCooldownResetAt(
+          /\(reset after (?:(\d+)h ?)?(?:(\d+)m ?)?(?:(\d+)s)?\)/.exec(message.result),
+          completedAt,
+        )
+      : undefined;
   return makeProviderFailure({
     message: listedError ?? structuredError ?? failureHint ?? message.result,
     code:
@@ -3827,6 +3844,7 @@ function providerFailureFromResult(
         : `api_error_${apiErrorStatus}`,
     class: failureClass,
     retryable: apiErrorStatus === 429 || apiErrorStatus === 529 ? true : null,
+    resetAt: resetAt ?? null,
   });
 }