fix(chat): say when the context limit cut an answer short

frontend/src/features/chat/ChatPage.a11y.test.tsx

@@ -338,6 +338,18 @@ describe("ChatPage response announcements", () => {
     await waitFor(() => expect(announcer()).toHaveTextContent(/^Response complete$/));
   });
 
+  it("says a response was cut short when the context limit ended it, and shows why under the answer", async () => {
+    const user = userEvent.setup();
+    const { api, finishWith } = harness();
+    renderChat(api);
+    await sendPrompt(user);
+
+    await finishWith("generation.completed", { assistant_message_id: "assistant-1" }, savedAnswer({ eval_count: 900, stop_reason: "length" }));
+
+    await waitFor(() => expect(announcer()).toHaveTextContent(/^Response cut short by the context limit$/));
+    expect(await screen.findByText(/This answer was cut short by the context limit/)).toBeVisible();
+  });
+
   it("says a response stopped after the person stopped it and the answer was kept", async () => {
     const user = userEvent.setup();
     const { api, finishWith } = harness();

frontend/src/features/chat/MessageCard.test.tsx

@@ -92,4 +92,34 @@ describe("MessageCard", () => {
     expect(screen.queryByRole("region", { name: "Cortex suggests remembering" })).toBeNull();
     expect(screen.getByText("Noted.")).toBeInTheDocument();
   });
+
+  it("notes under an answer the context limit cut short, and says what helps", () => {
+    renderCard({
+      id: "assistant-1",
+      role: "assistant",
+      content: "The three steps are: first, open the",
+      stats: { eval_count: 512, tokens_per_second: 24.5, stop_reason: "length" },
+    });
+
+    const note = screen.getByRole("note");
+    expect(note).toHaveTextContent("This answer was cut short by the context limit.");
+    expect(note).toHaveTextContent("Raising the Context window gives answers more room.");
+    // The partial answer and its usage are still shown as they were.
+    expect(screen.getByText("The three steps are: first, open the")).toBeInTheDocument();
+    expect(screen.getByText(/512 tok/)).toBeInTheDocument();
+  });
+
+  it("still says why when the limit left only reasoning and no answer", () => {
+    renderCard({ id: "assistant-1", role: "assistant", content: "", thoughts: "Let me think about this carefully.", stats: { stop_reason: "length" } });
+
+    expect(screen.getByRole("note")).toHaveTextContent("cut short by the context limit");
+  });
+
+  it("adds no note to an answer that finished on its own, one with no reason recorded, or a person's message", () => {
+    renderCard({ id: "assistant-1", role: "assistant", content: "Done.", stats: { eval_count: 3, tokens_per_second: 30, stop_reason: "stop" } });
+    renderCard({ id: "assistant-2", role: "assistant", content: "Older answer.", stats: { eval_count: 3, tokens_per_second: 30 } });
+    renderCard({ id: "user-1", role: "user", content: "Question?", stats: { stop_reason: "length" } });
+
+    expect(screen.queryByText(/cut short by the context limit/)).toBeNull();
+  });
 });

frontend/src/features/chat/MessageCard.tsx

@@ -74,6 +74,10 @@ export const MessageCard = memo(function MessageCard({ message, index, isFinalAs
           <div className="markdown-body">{message.role === "user" ? <p>{message.content}</p> : <SafeMarkdown content={message.content} />}</div>
         )}
         {message.role === "assistant" && untranslated && <p className="muted-note" role="note">Couldn't translate this answer; showing the original.</p>}
+        {/* The runtime reports why the model stopped. "length" means it ran
+            into the context limit, and the answer ends wherever that fell:
+            without this it reads as though it finished. */}
+        {message.role === "assistant" && message.stats?.stop_reason === "length" && <p className="muted-note" role="note">This answer was cut short by the context limit. Raising the Context window gives answers more room.</p>}
         {message.sources && message.sources.length > 0 && <details className="sources"><summary><span>Sources</span><span className="disclosure-hint">{message.sources.length} {message.sources.length === 1 ? "item" : "items"}</span></summary><div className="details-content"><div className="markdown-body"><SafeMarkdown content={message.sources.map((source) => typeof source === "string" ? source : JSON.stringify(source)).join("\n\n")} /></div></div></details>}
       </div>
       {message.role === "assistant" && message.id && onSaveMemory && proposedMemories && proposedMemories.length > 0 && (

frontend/src/features/chat/ResponseAnnouncer.tsx

@@ -4,12 +4,14 @@ import { useChatStore } from "../../stores/useChatStore";
 
 /**
  * What to tell someone who cannot see the answer arrive, once it is saved:
- * whether it finished or was stopped, and how long it was when that is known.
+ * whether it finished, was stopped, or was cut off by the context limit, and
+ * how long it was when that is known.
  */
 function describeFinishedResponse(chat: ChatResponse | null): string {
   const lastAssistant = [...(chat?.messages ?? [])].reverse().find((message) => message.role === "assistant");
   const stats = lastAssistant?.stats;
   if (stats?.stopped) return "Response stopped";
+  if (stats?.stop_reason === "length") return "Response cut short by the context limit";
   const tokens = stats?.eval_count;
   if (typeof tokens === "number" && tokens > 0) {
     return `Response complete, ${tokens.toLocaleString()} ${tokens === 1 ? "token" : "tokens"}`;