fix(cli): preserve description in subject-bearing thought chunks (#3691)

When a streamed reasoning chunk arrived with both a parsed subject (from
**Title**) and a description (the body text after \n\n in the same chunk),
the Thought event handler routed only to setThought and discarded the
description. As a result, the first body word that happened to share a
chunk with the closing ** was dropped from the persistent reasoning
display.

Treat subject-only chunks as discrete loading-indicator updates and route
all chunks carrying streamed text through the throttled buffer. The
existing flush merger preserves the subject across batched events.

packages/cli/src/ui/hooks/useGeminiStream.test.tsx

@@ -2734,6 +2734,51 @@ describe('useGeminiStream', () => {
       });
     });
 
+    it('should render descriptions from subject-bearing thought chunks', async () => {
+      mockSendMessageStream.mockReturnValue(
+        (async function* () {
+          yield {
+            type: ServerGeminiEventType.Thought,
+            value: {
+              subject: 'Evaluating installation approach',
+              description: 'The',
+            },
+          };
+          yield {
+            type: ServerGeminiEventType.Thought,
+            value: {
+              subject: '',
+              description: ' user mentioned globally installed qwen,',
+            },
+          };
+          yield {
+            type: ServerGeminiEventType.Finished,
+            value: { reason: 'STOP', usageMetadata: undefined },
+          };
+        })(),
+      );
+
+      const { result } = renderTestHook();
+
+      await act(async () => {
+        await result.current.submitQuery('Streamed thought');
+      });
+
+      await waitFor(() => {
+        expect(mockAddItem).toHaveBeenCalledWith(
+          expect.objectContaining({
+            type: 'gemini_thought',
+            text: 'The user mentioned globally installed qwen,',
+          }),
+          expect.any(Number),
+        );
+      });
+      expect(result.current.thought).toEqual({
+        subject: 'Evaluating installation approach',
+        description: 'The user mentioned globally installed qwen,',
+      });
+    });
+
     it('should show a retry countdown and update pending history over time', async () => {
       vi.useFakeTimers();
       try {

packages/cli/src/ui/hooks/useGeminiStream.ts

@@ -1285,9 +1285,12 @@ export const useGeminiStream = (
           dualOutput?.processEvent(event);
           switch (event.type) {
             case ServerGeminiEventType.Thought:
-              // If the thought has a subject, it's a discrete status update rather than
-              // a streamed textual thought, so we update the thought state directly.
-              if (event.value.subject) {
+              // Subject-only chunks are discrete status updates for the
+              // loading indicator and render immediately. Anything carrying
+              // streamed text (with or without a subject) goes through the
+              // throttled buffer so it batches with adjacent reasoning
+              // chunks; the flush merger preserves the subject.
+              if (event.value.subject && !event.value.description) {
                 flushBufferedStreamEvents();
                 setThought(event.value);
               } else {