diff --git a/frontend/src/features/chat/ChatPage.test.tsx b/frontend/src/features/chat/ChatPage.test.tsx
index 63ee285..10d8600 100644
--- a/frontend/src/features/chat/ChatPage.test.tsx
+++ b/frontend/src/features/chat/ChatPage.test.tsx
@@ -1215,6 +1215,56 @@ describe("ChatPage composer integration", () => {
     expect(await screen.findByRole("alert")).toHaveTextContent("Only 8 of 9 files were attached");
   });
 
+  it("does not offer to resend the prompt when forking fails", async () => {
+    // The banner shows one Retry for every failure it can display, and a fork
+    // is not a generation. With a prompt already sent and answered this
+    // session, the fork failure offered Retry, which resent a turn the model
+    // had already replied to.
+    const user = userEvent.setup();
+    let emit: ((event: unknown) => void) | null = null;
+    let resolveStream: (() => void) | undefined;
+    const generate = vi.fn().mockResolvedValue({
+      job_id: "job-fork",
+      kind: "generation",
+      status: "queued",
+      thread_id: "thread-a",
+      user_message_id: "message-user-1",
+    });
+    const api = chatApi({
+      generate,
+      forkChat: vi.fn().mockRejectedValue(new Error("fork exploded")),
+      chat: vi.fn(async (id: string) => ({
+        ...emptyChat(id),
+        messages: [
+          { id: "message-user-1", role: "user" as const, content: "Answered already" },
+          { id: "message-assistant-1", role: "assistant" as const, content: "Here you go." },
+        ],
+      })),
+      streamGeneration: vi.fn((_jobId, onEvent) => new Promise<void>((resolve) => {
+        resolveStream = resolve;
+        emit = (event) => (onEvent as (event: unknown) => void)(event);
+      })),
+    });
+    renderChat(api);
+
+    // Send a turn so `lastPrompt` is set -- without one the banner offers no
+    // Retry at all and the check below would pass vacuously.
+    await user.type(await screen.findByLabelText("Message Cortex"), "Answered already");
+    await user.click(screen.getByRole("button", { name: "Send message" }));
+    await waitFor(() => expect(emit).not.toBeNull());
+    await act(async () => {
+      emit!({ event_id: 1, event: "generation.completed", job_id: "job-fork", thread_id: "thread-a", data: {} });
+      resolveStream?.();
+    });
+    await screen.findByText("Here you go.");
+
+    await user.click((await screen.findAllByRole("button", { name: "Fork chat from this message" }))[0]);
+
+    expect(await screen.findByText("Could not fork this chat.")).toBeInTheDocument();
+    expect(screen.queryByRole("button", { name: "Retry last message" })).not.toBeInTheDocument();
+    expect(generate).toHaveBeenCalledTimes(1);
+  });
+
   it("explains the image capability mismatch before a generation request is made", async () => {
     const user = userEvent.setup();
     const attachment: ChatAttachment = {
diff --git a/frontend/src/features/chat/ChatPage.tsx b/frontend/src/features/chat/ChatPage.tsx
index fa01ddd..fb1cfb1 100644
--- a/frontend/src/features/chat/ChatPage.tsx
+++ b/frontend/src/features/chat/ChatPage.tsx
@@ -44,6 +44,12 @@ type Props = {
 type ScopedError = {
   message: string;
   threadId: string | null;
+  // Whether resending the last prompt is the remedy for this error. The
+  // banner offers one Retry button for every failure it can show, and most
+  // of them are not generation failures at all -- a failed fork, a failed
+  // stop, a reload that failed after the answer was already saved. Retrying
+  // the prompt there resubmits a turn the thread has already answered.
+  retryable: boolean;
 };
 
 type ChatLoadState = {
@@ -140,7 +146,7 @@ export function ChatPage({
   const handledClearRequestsRef = useRef(new Set<string>());
 
   const reportGenerationFailure = useCallback((failedThreadId: string, message: string) => {
-    setGenerationError({ threadId: failedThreadId, message });
+    setGenerationError({ threadId: failedThreadId, message, retryable: true });
   }, []);
 
   const loadChat = useCallback(async ({ preserveCurrent = false }: { preserveCurrent?: boolean } = {}) => {
@@ -214,9 +220,10 @@ export function ChatPage({
   const displayedThreadId = threadId ?? resolvedThreadId;
   const activeJobForCurrentThread = Boolean(generation.jobId && generation.threadId === displayedThreadId);
   const generationElsewhere = Boolean(generation.jobId && !activeJobForCurrentThread);
-  const visibleGenerationError = generationError && generationError.threadId === displayedThreadId
-    ? generationError.message
+  const visibleError = generationError && generationError.threadId === displayedThreadId
+    ? generationError
     : null;
+  const visibleGenerationError = visibleError?.message ?? null;
   const composerPhase: ComposerPhase = !runtimeReady
     ? "unavailable"
     : generation.phase === "stopping"
@@ -245,7 +252,9 @@ export function ChatPage({
       }
     } catch {
       if (!isLatestRequest()) return;
-      setGenerationError({ threadId: id, message: "Generation finished, but the saved chat could not be reloaded." });
+      // The generation succeeded; only the reload failed. Resending would
+      // ask for a second answer to a question already answered.
+      setGenerationError({ threadId: id, message: "Generation finished, but the saved chat could not be reloaded.", retryable: false });
       if (viewThreadIdRef.current !== id) return;
       // This call bumped the shared request version, so any route load still
       // in flight for this thread has already returned early as stale. If we
@@ -368,6 +377,7 @@ export function ChatPage({
       setGenerationError({
         threadId,
         message: runtimeMessage ?? "The local runtime is unavailable. Rescan local models after it is running.",
+        retryable: true,
       });
       return null;
     }
@@ -462,6 +472,7 @@ export function ChatPage({
       setGenerationError({
         threadId,
         message: requestError instanceof ApiError ? requestError.detail : "The response could not be started. Your message is still here.",
+        retryable: true,
       });
       return null;
     } finally {
@@ -568,9 +579,12 @@ export function ChatPage({
       }
     } catch (requestError) {
       useChatStore.getState().revertStopping(jobId);
+      // Stop failed, so the response is still running. Sending the prompt
+      // again would start a second one alongside it.
       setGenerationError({
         threadId: jobThreadId,
         message: requestError instanceof ApiError ? requestError.detail : "Could not stop the response.",
+        retryable: false,
       });
     } finally {
       stoppingRef.current = false;
@@ -615,9 +629,11 @@ export function ChatPage({
       const forked = await api.forkChat(threadId, message.id);
       onForked(forked);
     } catch (requestError) {
+      // Forking is not a generation. The thread is answered and unchanged.
       setGenerationError({
         threadId,
         message: requestError instanceof ApiError ? requestError.detail : "Could not fork this chat.",
+        retryable: false,
       });
     } finally {
       setForkingMessage(null);
@@ -788,7 +804,7 @@ export function ChatPage({
           onStop={cancel}
           onSelectModel={onSelectModel}
           onRescanModels={onRescanModels}
-          onRetry={lastPrompt ? retryLastPrompt : undefined}
+          onRetry={lastPrompt && visibleError?.retryable ? retryLastPrompt : undefined}
           onDismissError={() => setGenerationError(null)}
           generationOptions={threadOptions}
           generationDefaults={generationDefaults}