fix: handle missing usage metadata on premature Anthropic stream termination

src/strands/models/anthropic.py

@@ -405,12 +405,22 @@ async def stream(
         try:
             async with self.client.messages.stream(**request) as stream:
                 logger.debug("got response from model")
+                event = None
                 async for event in stream:
                     if event.type in AnthropicModel.EVENT_TYPES:
                         yield self.format_chunk(event.model_dump())
 
-                usage = event.message.usage  # type: ignore
-                yield self.format_chunk({"type": "metadata", "usage": usage.model_dump()})
+                usage = getattr(getattr(event, "message", None), "usage", None) if event else None
+                if usage is not None:
+                    yield self.format_chunk({"type": "metadata", "usage": usage.model_dump()})
+                else:
+                    logger.warning("stream ended without usage metadata (possible premature termination)")
+                    yield self.format_chunk(
+                        {
+                            "type": "metadata",
+                            "usage": {"input_tokens": 0, "output_tokens": 0},
+                        }
+                    )
 
         except anthropic.RateLimitError as error:
             raise ModelThrottledException(str(error)) from error

tests/strands/models/test_anthropic.py

@@ -739,7 +739,55 @@ async def test_stream(anthropic_client, model, agenerator, alist):
 
 
 @pytest.mark.asyncio
-async def test_stream_rate_limit_error(anthropic_client, model, alist):
+async def test_stream_premature_termination(anthropic_client, model, agenerator, alist):
+    """Test that stream handles premature termination without crashing.
+
+    When the Anthropic API stream ends before message_stop (e.g. network
+    timeout), event.message.usage may be None. The code must not crash
+    with AttributeError.
+
+    Regression test for #1868.
+    """
+    mock_event_1 = unittest.mock.Mock(
+        type="message_start",
+        model_dump=lambda: {"type": "message_start"},
+    )
+    # Last event has no .message attribute (simulating premature termination)
+    mock_event_2 = unittest.mock.Mock(
+        type="content_block_stop",
+        model_dump=lambda: {"type": "content_block_stop", "index": 0},
+        spec=["type", "model_dump"],
+    )
+
+    mock_context = unittest.mock.AsyncMock()
+    mock_context.__aenter__.return_value = agenerator([mock_event_1, mock_event_2])
+    anthropic_client.messages.stream.return_value = mock_context
+
+    messages = [{"role": "user", "content": [{"text": "hello"}]}]
+    response = model.stream(messages, None, None)
+
+    # Should not raise AttributeError
+    tru_events = await alist(response)
+
+    # Should still yield a metadata event with zero usage
+    assert any("metadata" in str(e) for e in tru_events)
+
+
+@pytest.mark.asyncio
+async def test_stream_empty_no_events(anthropic_client, model, agenerator, alist):
+    """Test that stream handles an empty event sequence without crashing."""
+    mock_context = unittest.mock.AsyncMock()
+    mock_context.__aenter__.return_value = agenerator([])
+    anthropic_client.messages.stream.return_value = mock_context
+
+    messages = [{"role": "user", "content": [{"text": "hello"}]}]
+    response = model.stream(messages, None, None)
+
+    # Should not raise UnboundLocalError or AttributeError
+    tru_events = await alist(response)
+
+    # Should still yield a metadata event with zero usage
+    assert any("metadata" in str(e) for e in tru_events)
     anthropic_client.messages.stream.side_effect = anthropic.RateLimitError(
         "rate limit", response=unittest.mock.Mock(), body=None
     )