Only decode JSON input buffer in Anthropic Claude streaming

CHANGELOG.md

@@ -11,6 +11,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased
 
+### Fixed
+
+- `opentelemetry-instrumentation-botocore`: Handle dict input in _decode_tool_use for Bedrock streaming
+  ([#3875](https://github.com/open-telemetry/opentelemetry-python-contrib/pull/3875))
+
 ## Version 1.38.0/0.59b0 (2025-10-16)
 
 ### Fixed

instrumentation/opentelemetry-instrumentation-botocore/src/opentelemetry/instrumentation/botocore/extensions/bedrock_utils.py

@@ -36,16 +36,6 @@
 _StreamErrorCallableT = Callable[[Exception], None]
 
 
-def _decode_tool_use(tool_use):
-    # input get sent encoded in json
-    if "input" in tool_use:
-        try:
-            tool_use["input"] = json.loads(tool_use["input"])
-        except json.JSONDecodeError:
-            pass
-    return tool_use
-
-
 # pylint: disable=abstract-method
 class ConverseStreamWrapper(ObjectProxy):
     """Wrapper for botocore.eventstream.EventStream"""
@@ -368,10 +358,13 @@ def _process_anthropic_claude_chunk(self, chunk):
         if message_type == "content_block_stop":
             # {'type': 'content_block_stop', 'index': 0}
             if self._tool_json_input_buf:
-                self._content_block["input"] = self._tool_json_input_buf
-            self._message["content"].append(
-                _decode_tool_use(self._content_block)
-            )
+                try:
+                    self._content_block["input"] = json.loads(
+                        self._tool_json_input_buf
+                    )
+                except json.JSONDecodeError:
+                    self._content_block["input"] = self._tool_json_input_buf
+            self._message["content"].append(self._content_block)
             self._content_block = {}
             self._tool_json_input_buf = ""
             return

instrumentation/opentelemetry-instrumentation-botocore/tests/test_botocore_bedrock.py

@@ -2975,6 +2975,82 @@ def test_invoke_model_with_response_stream_invalid_model(
     assert len(logs) == 0
 
 
+@pytest.mark.parametrize(
+    "input_value,expected_output",
+    [
+        ({"location": "Seattle"}, {"location": "Seattle"}),
+        ({}, {}),
+        (None, None),
+    ],
+)
+def test_anthropic_claude_chunk_tool_use_input_handling(
+    input_value, expected_output
+):
+    """Test that _process_anthropic_claude_chunk handles various tool_use input formats."""
+    from opentelemetry.instrumentation.botocore.extensions.bedrock_utils import (
+        InvokeModelWithResponseStreamWrapper,
+    )
+
+    def stream_done_callback(response, ended):
+        pass
+
+    def stream_error_callback(exc, ended):
+        pass
+
+    wrapper = InvokeModelWithResponseStreamWrapper(
+        stream=mock.MagicMock(),
+        stream_done_callback=stream_done_callback,
+        stream_error_callback=stream_error_callback,
+        model_id="anthropic.claude-3-5-sonnet-20240620-v1:0",
+    )
+
+    # Simulate message_start
+    wrapper._process_anthropic_claude_chunk(
+        {
+            "type": "message_start",
+            "message": {
+                "role": "assistant",
+                "content": [],
+            },
+        }
+    )
+
+    # Simulate content_block_start with specified input
+    content_block = {
+        "type": "tool_use",
+        "id": "test_id",
+        "name": "test_tool",
+    }
+    if input_value is not None:
+        content_block["input"] = input_value
+
+    wrapper._process_anthropic_claude_chunk(
+        {
+            "type": "content_block_start",
+            "index": 0,
+            "content_block": content_block,
+        }
+    )
+
+    # Simulate content_block_stop
+    wrapper._process_anthropic_claude_chunk(
+        {"type": "content_block_stop", "index": 0}
+    )
+
+    # Verify the message content
+    assert len(wrapper._message["content"]) == 1
+    tool_block = wrapper._message["content"][0]
+    assert tool_block["type"] == "tool_use"
+    assert tool_block["id"] == "test_id"
+    assert tool_block["name"] == "test_tool"
+
+    if expected_output is not None:
+        assert tool_block["input"] == expected_output
+        assert isinstance(tool_block["input"], dict)
+    else:
+        assert "input" not in tool_block
+
+
 def amazon_nova_messages():
     return [
         {"role": "user", "content": [{"text": "Say this is a test"}]},