diff --git a/apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.test.ts b/apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.test.ts
index d61739f72c21..6dd2f5758193 100644
--- a/apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.test.ts
+++ b/apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.test.ts
@@ -1439,6 +1439,62 @@ describe("ProviderRuntimeIngestion", () => {
     expect(message?.streaming).toBe(false);
   });
 
+  it("keeps a usage limit through the limit notice and clears it once work resumes after the reset", async () => {
+    const harness = await createHarness();
+    const limitedAt = "2026-01-01T00:00:00.000Z";
+    const resetsAt = "2026-01-01T01:00:00.000Z";
+    const emitAssistantText = (id: string, createdAt: string) => {
+      harness.emit({
+        type: "content.delta",
+        eventId: asEventId(`evt-${id}-delta`),
+        provider: ProviderDriverKind.make("claudeAgent"),
+        createdAt,
+        threadId: asThreadId("thread-1"),
+        turnId: asTurnId("turn-2"),
+        itemId: asItemId(id),
+        payload: { streamKind: "assistant_text", delta: "text" },
+      });
+      harness.emit({
+        type: "item.completed",
+        eventId: asEventId(`evt-${id}-completed`),
+        provider: ProviderDriverKind.make("claudeAgent"),
+        createdAt,
+        threadId: asThreadId("thread-1"),
+        turnId: asTurnId("turn-2"),
+        itemId: asItemId(id),
+        payload: { itemType: "assistant_message", status: "completed" },
+      });
+    };
+
+    harness.emit({
+      type: "turn.usage-limited",
+      eventId: asEventId("evt-usage-limited"),
+      provider: ProviderDriverKind.make("claudeAgent"),
+      createdAt: limitedAt,
+      threadId: asThreadId("thread-1"),
+      turnId: asTurnId("turn-2"),
+      payload: { resetsAt },
+    });
+    // Claude streams its "You've hit your session limit" notice right after the limit.
+    emitAssistantText("limit-notice", limitedAt);
+    const limited = await waitForThread(harness.readModel, (entry) =>
+      entry.messages.some(
+        (message: ProviderRuntimeTestMessage) =>
+          message.id === "assistant:limit-notice" && !message.streaming,
+      ),
+    );
+    expect(limited.usageLimit?.resetsAt).toBe(resetsAt);
+
+    emitAssistantText("resumed-work", "2026-01-01T01:00:30.000Z");
+    const resumed = await waitForThread(harness.readModel, (entry) =>
+      entry.messages.some(
+        (message: ProviderRuntimeTestMessage) =>
+          message.id === "assistant:resumed-work" && !message.streaming,
+      ),
+    );
+    expect(resumed.usageLimit ?? null).toBeNull();
+  });
+
   it("streams reasoning deltas into a finalized reasoning message", async () => {
     const harness = await createHarness();
     const now = "2026-01-01T00:00:00.000Z";
diff --git a/apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.ts b/apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.ts
index 3542b112f1b1..046135b1d95f 100644
--- a/apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.ts
+++ b/apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.ts
@@ -1973,13 +1973,18 @@ const make = Effect.gen(function* () {
         });
       } else if (
         thread.usageLimit != null &&
-        ((event.type === "content.delta" && event.payload.streamKind === "assistant_text") ||
+        // Claude streams its limit notice as assistant text, so only output
+        // after the reset shows a parked turn carrying on.
+        ((event.type === "content.delta" &&
+          event.payload.streamKind === "assistant_text" &&
+          thread.usageLimit.resetsAt !== null &&
+          Date.parse(now) >= Date.parse(thread.usageLimit.resetsAt)) ||
           (event.type === "turn.completed" &&
             shouldApplyThreadLifecycle &&
             normalizeRuntimeTurnState(event.payload.state) === "completed"))
       ) {
-        // A parked turn the provider carried on by itself no longer needs a resume,
-        // and must not be interrupted by one.
+        // A parked turn the provider carried on by itself no longer needs a
+        // resume, and must not be interrupted by one.
         yield* orchestrationEngine.dispatch({
           type: "thread.usage-limit.set",
           commandId: yield* providerCommandId(event, "usage-limit-clear"),