diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6c2aa9c9..a301635e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -92,6 +92,16 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   no longer froze on the previous step's authoritative usage; the current step's
   estimate is combined per-component with `max`, so the count rises smoothly and
   never jumps backwards. (#163)
+- **AI chat: "New chat" pressed during the first turn's stream now resets the
+  thread instead of leaving the old turn streaming.** While a brand-new,
+  not-yet-adopted chat streamed its first turn, hitting "New chat" left
+  `activeChatId === null` (a no-op for the atom), so the reconciler never
+  remounted and the in-flight thread kept streaming behind the fresh one — and a
+  late refetch / late `onFinish` from that abandoned thread could yank the user
+  back into the chat they just left. "New chat" now forces a fresh empty thread
+  unconditionally and the finished thread's mount key is checked so a late
+  callback from an abandoned thread no longer adopts or re-arms the fallback.
+  (#161)
 
 ## [0.93.0] - 2026-06-21
 
diff --git a/apps/client/src/features/ai-chat/components/chat-thread.test.tsx b/apps/client/src/features/ai-chat/components/chat-thread.test.tsx
new file mode 100644
index 00000000..d1c71bc3
--- /dev/null
+++ b/apps/client/src/features/ai-chat/components/chat-thread.test.tsx
@@ -0,0 +1,94 @@
+import { describe, it, expect, vi, beforeEach } from "vitest";
+import { render, act } from "@testing-library/react";
+import { MantineProvider } from "@mantine/core";
+
+// Capture the options ChatThread passes to useChat so the test can drive the
+// hook's terminal callbacks (here: onError) directly, without a real stream. The
+// box is created via vi.hoisted so the hoisted vi.mock factory below can close
+// over it.
+const { useChatBox } = vi.hoisted(() => ({
+  useChatBox: { options: null as unknown as Record<string, unknown> | null },
+}));
+
+// Mock the AI SDK hook: record the options and return an inert, ready store so
+// ChatThread renders without any network/streaming machinery.
+vi.mock("@ai-sdk/react", () => ({
+  useChat: (options: Record<string, unknown>) => {
+    useChatBox.options = options;
+    return {
+      messages: [],
+      sendMessage: vi.fn(),
+      status: "ready",
+      stop: vi.fn(),
+      error: null,
+    };
+  },
+}));
+
+// Stub react-i18next so `t` returns the key (other component tests use the same
+// pattern); ChatThread's rendered chrome is irrelevant to this wiring test.
+vi.mock("react-i18next", () => ({
+  useTranslation: () => ({ t: (key: string) => key }),
+}));
+
+// Mock the heavy presentational children to trivial stubs — this test only
+// exercises the onError → onTurnFinished wiring, not their rendering.
+vi.mock("@/features/ai-chat/components/message-list.tsx", () => ({
+  default: () => null,
+}));
+vi.mock("@/features/ai-chat/components/chat-input.tsx", () => ({
+  default: () => null,
+}));
+vi.mock("@/features/ai-chat/components/role-cards.tsx", () => ({
+  default: () => null,
+}));
+vi.mock("@/features/ai-chat/components/chat-error-alert.tsx", () => ({
+  default: () => null,
+}));
+vi.mock("@/features/ai-chat/components/chat-stopped-notice.tsx", () => ({
+  default: () => null,
+}));
+
+import ChatThread from "./chat-thread";
+
+// matchMedia (read by MantineProvider) is stubbed globally in vitest.setup.ts.
+
+describe("ChatThread onError wiring (#161)", () => {
+  beforeEach(() => {
+    useChatBox.options = null;
+    vi.clearAllMocks();
+  });
+
+  it("onError calls onTurnFinished with (undefined, threadKey) so a late error from an abandoned thread is rejected", () => {
+    const onTurnFinished = vi.fn();
+    // Silence the deliberate console.error ChatThread logs for devtools.
+    const consoleError = vi
+      .spyOn(console, "error")
+      .mockImplementation(() => {});
+
+    render(
+      <MantineProvider>
+        <ChatThread
+          chatId="c1"
+          onTurnFinished={onTurnFinished}
+          threadKey="thread-key-1"
+        />
+      </MantineProvider>,
+    );
+
+    const options = useChatBox.options;
+    expect(options).not.toBeNull();
+    expect(typeof options?.onError).toBe("function");
+
+    // Drive the captured onError exactly as the AI SDK would on a stream error.
+    act(() => {
+      (options!.onError as (e: Error) => void)(new Error("stream blew up"));
+    });
+
+    // The thread's own mount key must be forwarded with NO server id, so the
+    // session hook can reject this finish if the thread has been abandoned.
+    expect(onTurnFinished).toHaveBeenCalledWith(undefined, "thread-key-1");
+
+    consoleError.mockRestore();
+  });
+});
diff --git a/apps/client/src/features/ai-chat/hooks/use-chat-session.test.tsx b/apps/client/src/features/ai-chat/hooks/use-chat-session.test.tsx
index 7588b1e4..164f3976 100644
--- a/apps/client/src/features/ai-chat/hooks/use-chat-session.test.tsx
+++ b/apps/client/src/features/ai-chat/hooks/use-chat-session.test.tsx
@@ -120,16 +120,40 @@ describe("useChatSession", () => {
     expect(setActiveChatId).not.toHaveBeenCalledWith("new");
   });
 
-  it("startNewChat while already in a new chat: cancelPendingAdoption stops a late refetch adopting the failed chat", () => {
-    // The Warning path the render-phase reconciler can't catch: pressing "New
-    // chat" while already in a new chat keeps activeChatId === null (a no-op for
-    // the atom), so only the explicit cancelPendingAdoption() disarms.
+  it("cancelPendingAdoption (selectChat) disarms a late refetch from adopting the just-failed chat", () => {
+    // cancelPendingAdoption is the explicit disarm the window calls from
+    // selectChat: switching to a chat whose id == null is a no-op for the atom, so
+    // the render-phase reconciler never fires and only this call disarms an armed
+    // error-path fallback. (startNewChat no longer routes through here — it calls
+    // startFreshThread, covered by the next test — but cancelPendingAdoption still
+    // backs selectChat, so this guard must hold.)
     const { result, rerender, setActiveChatId } = setup({
       activeChatId: null,
       chats: { items: [{ id: "x" }] },
     });
     result.current.onTurnFinished(undefined); // first turn failed → arm (before=["x"])
-    result.current.cancelPendingAdoption(); // window calls this from startNewChat
+    result.current.cancelPendingAdoption(); // window calls this from selectChat
+    // The just-failed row lands in a late refetch; it must NOT be adopted.
+    rerender({
+      activeChatId: null,
+      chats: { items: [{ id: "x" }, { id: "failed" }] },
+    });
+    expect(setActiveChatId).not.toHaveBeenCalledWith("failed");
+  });
+
+  it("#161: startFreshThread disarms the armed error-path fallback (New chat during the first turn)", () => {
+    // Pressing "New chat" while already in a not-yet-adopted new chat keeps
+    // activeChatId === null, so the render-phase reconciler never fires. The
+    // window now calls startFreshThread() (NOT cancelPendingAdoption) to force a
+    // fresh thread; this test pins the load-bearing fact that startFreshThread
+    // ALSO nulls pendingNewChatRef, so a late refetch of the just-failed row can't
+    // yank the user back into the abandoned chat.
+    const { result, rerender, setActiveChatId } = setup({
+      activeChatId: null,
+      chats: { items: [{ id: "x" }] },
+    });
+    result.current.onTurnFinished(undefined); // first turn failed → arm (before=["x"])
+    act(() => result.current.startFreshThread()); // "New chat" → fresh thread + disarm
     // The just-failed row lands in a late refetch; it must NOT be adopted.
     rerender({
       activeChatId: null,