Only decode JSON input buffer in Anthropic Claude streaming

_decode_tool_use was only used when _tool_json_input_buf was found, but we were decoding the entire _content_block after adding _tool_json_input_buf to it. The _content_block overall which could contain non-JSON elements (e.g. {}), causing failures. To fix this, we have removed _decode_tool_use helper function and inlined JSON decoding logic directly into content_block_stop handler in _process_anthropic_claude_chunk, where we only use it to decode _tool_json_input_buf before appending to _content_block.

Patch based on open-telemetry/opentelemetry-python-contrib#3875 with code copied directly from https://github.com/open-telemetry/opentelemetry-python-contrib/blob/v0.54b1/instrumentation/opentelemetry-instrumentation-botocore/src/opentelemetry/instrumentation/botocore/extensions/bedrock_utils.py#L289

Author: thpierce

CHANGELOG.md

@@ -13,3 +13,5 @@ If your change does not need a CHANGELOG entry, add the "skip changelog" label t
 ## Unreleased
 - Add botocore instrumentation extension for Bedrock AgentCore services with span attributes
   ([#490](https://github.com/aws-observability/aws-otel-python-instrumentation/pull/490))
+- [PATCH] Only decode JSON input buffer in Anthropic Claude streaming
+    ([#497](https://github.com/aws-observability/aws-otel-python-instrumentation/pull/497))

aws-opentelemetry-distro/src/amazon/opentelemetry/distro/patches/_botocore_patches.py

@@ -334,8 +334,88 @@ def patched_extract_tool_calls(
             tool_calls.append(tool_call)
         return tool_calls
 
+    # TODO: The following code is to patch a bedrock bug that was fixed in
+    # opentelemetry-instrumentation-botocore==0.60b0 in:
+    # https://github.com/open-telemetry/opentelemetry-python-contrib/pull/3875
+    # Remove this code once we've bumped opentelemetry-instrumentation-botocore dependency to 0.60b0
+    def patched_process_anthropic_claude_chunk(self, chunk):
+        # pylint: disable=too-many-return-statements,too-many-branches
+        if not (message_type := chunk.get("type")):
+            return
+
+        if message_type == "message_start":
+            # {'type': 'message_start', 'message': {'id': 'id', 'type': 'message', 'role': 'assistant', 'model': 'claude-2.0', 'content': [], 'stop_reason': None, 'stop_sequence': None, 'usage': {'input_tokens': 18, 'output_tokens': 1}}}
+            if chunk.get("message", {}).get("role") == "assistant":
+                self._record_message = True
+                message = chunk["message"]
+                self._message = {
+                    "role": message["role"],
+                    "content": message.get("content", []),
+                }
+            return
+
+        if message_type == "content_block_start":
+            # {'type': 'content_block_start', 'index': 0, 'content_block': {'type': 'text', 'text': ''}}
+            # {'type': 'content_block_start', 'index': 1, 'content_block': {'type': 'tool_use', 'id': 'id', 'name': 'func_name', 'input': {}}}
+            if self._record_message:
+                block = chunk.get("content_block", {})
+                if block.get("type") == "text":
+                    self._content_block = block
+                elif block.get("type") == "tool_use":
+                    self._content_block = block
+            return
+
+        if message_type == "content_block_delta":
+            # {'type': 'content_block_delta', 'index': 0, 'delta': {'type': 'text_delta', 'text': 'Here'}}
+            # {'type': 'content_block_delta', 'index': 1, 'delta': {'type': 'input_json_delta', 'partial_json': ''}}
+            if self._record_message:
+                delta = chunk.get("delta", {})
+                if delta.get("type") == "text_delta":
+                    self._content_block["text"] += delta.get("text", "")
+                elif delta.get("type") == "input_json_delta":
+                    self._tool_json_input_buf += delta.get("partial_json", "")
+            return
+
+        if message_type == "content_block_stop":
+            # {'type': 'content_block_stop', 'index': 0}
+            if self._tool_json_input_buf:
+                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
+
+        if message_type == "message_delta":
+            # {'type': 'message_delta', 'delta': {'stop_reason': 'end_turn', 'stop_sequence': None}, 'usage': {'output_tokens': 123}}
+            if (
+                stop_reason := chunk.get("delta", {}).get("stop_reason")
+            ) is not None:
+                self._response["stopReason"] = stop_reason
+            return
+
+        if message_type == "message_stop":
+            # {'type': 'message_stop', 'amazon-bedrock-invocationMetrics': {'inputTokenCount': 18, 'outputTokenCount': 123, 'invocationLatency': 5250, 'firstByteLatency': 290}}
+            if invocation_metrics := chunk.get(
+                "amazon-bedrock-invocationMetrics"
+            ):
+                self._process_invocation_metrics(invocation_metrics)
+
+            if self._record_message:
+                self._response["output"] = {"message": self._message}
+                self._record_message = False
+                self._message = None
+
+            self._stream_done_callback(self._response)
+            return
+
     bedrock_utils.ConverseStreamWrapper.__init__ = patched_init
     bedrock_utils.ConverseStreamWrapper._process_event = patched_process_event
+    bedrock_utils.InvokeModelWithResponseStreamWrapper._process_anthropic_claude_chunk = patched_process_anthropic_claude_chunk
     bedrock_utils.extract_tool_calls = patched_extract_tool_calls
 
     # END The OpenTelemetry Authors code

aws-opentelemetry-distro/tests/amazon/opentelemetry/distro/patches/test_instrumentation_patch.py

@@ -120,10 +120,6 @@ def _run_patch_behaviour_tests(self):
         self._test_unpatched_botocore_propagator()
         self._test_unpatched_gevent_instrumentation()
         self._test_unpatched_starlette_instrumentation()
-        # TODO: remove these tests once we bump botocore instrumentation version to 0.56b0
-        # Bedrock Runtime tests
-        self._test_unpatched_converse_stream_wrapper()
-        self._test_unpatched_extract_tool_calls()
 
         # Apply patches
         apply_instrumentation_patches()
@@ -222,6 +218,17 @@ def _test_unpatched_botocore_instrumentation(self):
         # DynamoDB
         self.assertTrue("dynamodb" in _KNOWN_EXTENSIONS, "Upstream has removed a DynamoDB extension")
 
+        # Bedrock Runtime tests
+        # TODO: remove these tests once we bump botocore instrumentation version to 0.56b0
+        self._test_unpatched_converse_stream_wrapper()
+        self._test_unpatched_extract_tool_calls()
+
+        # TODO: remove these tests once we bump botocore instrumentation version to 0.60b0
+        self._test_unpatched_process_anthropic_claude_chunk({"location": "Seattle"}, {"location": "Seattle"})
+        self._test_unpatched_process_anthropic_claude_chunk(None, None)
+        self._test_unpatched_process_anthropic_claude_chunk({}, {})
+
+
     def _test_unpatched_gevent_instrumentation(self):
         self.assertFalse(gevent.monkey.is_module_patched("os"), "gevent os module has been patched")
         self.assertFalse(gevent.monkey.is_module_patched("thread"), "gevent thread module has been patched")
@@ -267,10 +274,14 @@ def _test_patched_botocore_instrumentation(self):
         # Bedrock Agent Operation
         self._test_patched_bedrock_agent_instrumentation()
 
-        # TODO: remove these tests once we bump botocore instrumentation version to 0.56b0
         # Bedrock Runtime
+        # TODO: remove these tests once we bump botocore instrumentation version to 0.56b0
         self._test_patched_converse_stream_wrapper()
         self._test_patched_extract_tool_calls()
+        # TODO: remove these tests once we bump botocore instrumentation version to 0.60b0
+        self._test_patched_process_anthropic_claude_chunk({"location": "Seattle"}, {"location": "Seattle"})
+        self._test_patched_process_anthropic_claude_chunk(None, None)
+        self._test_patched_process_anthropic_claude_chunk({}, {})
 
         # Bedrock Agent Runtime
         self.assertTrue("bedrock-agent-runtime" in _KNOWN_EXTENSIONS)
@@ -679,6 +690,73 @@ def _test_patched_extract_tool_calls(self):
         result = bedrock_utils.extract_tool_calls(message_with_string_content, True)
         self.assertIsNone(result)
 
+    def _test_patched_process_anthropic_claude_chunk(self, input_value: Dict[str, str], expected_output: Dict[str, str]):
+        self._test_process_anthropic_claude_chunk(input_value, expected_output, False)
+
+    def _test_unpatched_process_anthropic_claude_chunk(self, input_value: Dict[str, str], expected_output: Dict[str, str]):
+        self._test_process_anthropic_claude_chunk(input_value, expected_output, True)
+
+    def _test_process_anthropic_claude_chunk(self, input_value: Dict[str, str], expected_output: Dict[str, str], expect_exception: bool):
+        """Test that _process_anthropic_claude_chunk handles various tool_use input formats."""
+        wrapper = bedrock_utils.InvokeModelWithResponseStreamWrapper(
+            stream=MagicMock(),
+            stream_done_callback=MagicMock,
+            stream_error_callback=MagicMock,
+            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
+        try:
+            wrapper._process_anthropic_claude_chunk(
+                {"type": "content_block_stop", "index": 0}
+            )
+        except TypeError:
+            if expect_exception:
+                return
+            else:
+                raise
+
+        # Verify the message content
+        self.assertEqual(len(wrapper._message["content"]), 1)
+        tool_block = wrapper._message["content"][0]
+        self.assertEqual(tool_block["type"], "tool_use")
+        self.assertEqual(tool_block["id"], "test_id")
+        self.assertEqual(tool_block["name"], "test_tool")
+
+        if expected_output is not None:
+            self.assertEqual(tool_block["input"], expected_output)
+            self.assertIsInstance(tool_block["input"], dict)
+        else:
+            self.assertNotIn("input", tool_block)
+
     def _test_patched_bedrock_agent_instrumentation(self):
         """For bedrock-agent service, both extract_attributes and on_success provides attributes,
         the attributes depend on the API being invoked."""