diff --git a/src/App.tsx b/src/App.tsx
index 39064bf..294e795 100644
--- a/src/App.tsx
+++ b/src/App.tsx
@@ -386,6 +386,14 @@ function App() {
    * deferred until Framer Motion finishes the exit transition.
    */
   const requestHideOverlay = useCallback(() => {
+    // Cancel any in-flight work: active Ollama streaming, image processing,
+    // or screen capture. This ensures hiding the overlay (via double-tap
+    // Control, Escape, Cmd+W, or the X button) behaves like pressing Stop.
+    cancel();
+    pendingSubmitRef.current = null;
+    setIsSubmitPending(false);
+    setPendingUserMessage(null);
+
     windowAnchorRef.current = null;
     isPreExpandedRef.current = false;
     /* v8 ignore start -- DOM ref null guard: always set when overlay is visible */
@@ -407,7 +415,7 @@ function App() {
       }
       return 'hiding';
     });
-  }, []);
+  }, [cancel]);
 
   /** Ref attached to the chat-mode history dropdown for click-outside detection. */
   const historyDropdownRef = useRef<HTMLDivElement>(null);
diff --git a/src/hooks/__tests__/useOllama.test.tsx b/src/hooks/__tests__/useOllama.test.tsx
index e2c0b49..9338469 100644
--- a/src/hooks/__tests__/useOllama.test.tsx
+++ b/src/hooks/__tests__/useOllama.test.tsx
@@ -573,6 +573,118 @@ describe('useOllama', () => {
     });
   });
 
+  // ─── Stale channel after reset ───────────────────────────────────────────────
+
+  describe('stale channel after reset', () => {
+    it('ignores Token chunks arriving after reset()', async () => {
+      const { result } = renderHook(() => useOllama());
+
+      await act(async () => {
+        await result.current.ask('hello');
+      });
+
+      const channel = getChannel();
+
+      act(() => {
+        channel!.simulateMessage({ type: 'Token', data: 'Partial' });
+      });
+      expect(result.current.streamingContent).toBe('Partial');
+
+      // Reset clears state and bumps the epoch
+      act(() => {
+        result.current.reset();
+      });
+      expect(result.current.messages).toEqual([]);
+      expect(result.current.streamingContent).toBe('');
+
+      // Old channel sends more chunks after the reset
+      act(() => {
+        channel!.simulateMessage({ type: 'Token', data: ' stale token' });
+      });
+
+      // Should remain empty; the stale token must be discarded
+      expect(result.current.streamingContent).toBe('');
+      expect(result.current.messages).toEqual([]);
+    });
+
+    it('ignores Cancelled chunk arriving after reset()', async () => {
+      const { result } = renderHook(() => useOllama());
+
+      await act(async () => {
+        await result.current.ask('hello');
+      });
+
+      const channel = getChannel();
+
+      act(() => {
+        channel!.simulateMessage({ type: 'Token', data: 'Partial' });
+      });
+
+      act(() => {
+        result.current.reset();
+      });
+
+      // Old channel delivers Cancelled after reset
+      act(() => {
+        channel!.simulateMessage({ type: 'Cancelled' });
+      });
+
+      // The partial content must NOT reappear as a finalized message
+      expect(result.current.messages).toEqual([]);
+      expect(result.current.isGenerating).toBe(false);
+    });
+
+    it('ignores Done chunk arriving after reset()', async () => {
+      const onTurnComplete = vi.fn();
+      const { result } = renderHook(() => useOllama(onTurnComplete));
+
+      await act(async () => {
+        await result.current.ask('hello');
+      });
+
+      const channel = getChannel();
+
+      act(() => {
+        channel!.simulateMessage({ type: 'Token', data: 'Full answer' });
+      });
+
+      act(() => {
+        result.current.reset();
+      });
+
+      // Old channel delivers Done after reset
+      act(() => {
+        channel!.simulateMessage({ type: 'Done' });
+      });
+
+      expect(result.current.messages).toEqual([]);
+      expect(onTurnComplete).not.toHaveBeenCalled();
+    });
+
+    it('ignores Error chunk arriving after reset()', async () => {
+      const { result } = renderHook(() => useOllama());
+
+      await act(async () => {
+        await result.current.ask('hello');
+      });
+
+      const channel = getChannel();
+
+      act(() => {
+        result.current.reset();
+      });
+
+      act(() => {
+        channel!.simulateMessage({
+          type: 'Error',
+          data: { kind: 'Other', message: 'Something went wrong\nHTTP 500' },
+        });
+      });
+
+      expect(result.current.messages).toEqual([]);
+    });
+  });
+
   // ─── onTurnComplete callback ─────────────────────────────────────────────────
 
   describe('onTurnComplete callback', () => {
diff --git a/src/hooks/useOllama.ts b/src/hooks/useOllama.ts
index d0499a6..304ec2f 100644
--- a/src/hooks/useOllama.ts
+++ b/src/hooks/useOllama.ts
@@ -1,4 +1,4 @@
-import { useState, useCallback } from 'react';
+import { useState, useCallback, useRef } from 'react';
 import { invoke, Channel } from '@tauri-apps/api/core';
 
 /** Mirrors the Rust OllamaErrorKind enum sent over IPC. */
@@ -46,6 +46,10 @@ export function useOllama(
   const [streamingContent, setStreamingContent] = useState('');
   const [isGenerating, setIsGenerating] = useState(false);
 
+  // Epoch counter: bumped on every reset so that stale channel callbacks from
+  // a previous generation can detect they are outdated and bail out.
+  const epochRef = useRef(0);
+
   /**
    * Submits a message to the Ollama backend and initiates the streaming response.
    * The backend manages conversation history — only the new user message is sent.
@@ -82,12 +86,19 @@ export function useOllama(
       setStreamingContent('');
       setIsGenerating(true);
 
+      // Snapshot the epoch so this channel's callbacks can detect a reset.
+      const epochAtStart = epochRef.current;
+
       const channel = new Channel<StreamChunk>();
       // Use block-scoped variable to accumulate the stream and occasionally flush to React state,
       // mitigating rendering lag from hundreds of fast chunk events.
       let currentContent = '';
 
       channel.onmessage = (chunk) => {
+        // A reset occurred since this generation started; discard all
+        // remaining chunks so stale content never re-populates the UI.
+        if (epochRef.current !== epochAtStart) return;
+
         if (chunk.type === 'Token') {
           currentContent += chunk.data;
           setStreamingContent(currentContent);
@@ -165,6 +176,7 @@ export function useOllama(
 
   /** Resets all conversation state to prepare for a fresh session. */
   const reset = useCallback(() => {
+    epochRef.current += 1;
     setMessages([]);
     setStreamingContent('');
     setIsGenerating(false);