fix: Turn off parallel tool calling for Claude (#2736)

letta/interfaces/anthropic_streaming_interface.py

@@ -91,7 +91,14 @@ class AnthropicStreamingInterface:
     def get_tool_call_object(self) -> ToolCall:
         """Useful for agent loop"""
         # hack for tool rules
-        tool_input = json.loads(self.accumulated_tool_call_args)
+        try:
+            tool_input = json.loads(self.accumulated_tool_call_args)
+        except json.JSONDecodeError as e:
+            logger.warning(
+                f"Failed to decode tool call arguments for tool_call_id={self.tool_call_id}, "
+                f"name={self.tool_call_name}. Raw input: {self.accumulated_tool_call_args!r}. Error: {e}"
+            )
+            raise
         if "id" in tool_input and tool_input["id"].startswith("toolu_") and "function" in tool_input:
             arguments = str(json.dumps(tool_input["function"]["arguments"], indent=2))
         else:

letta/llm_api/anthropic_client.py

@@ -202,7 +202,7 @@ class AnthropicClient(LLMClientBase):
             tool_choice = {"type": "auto", "disable_parallel_tool_use": True}
             tools_for_request = [OpenAITool(function=f) for f in tools]
         elif force_tool_call is not None:
-            tool_choice = {"type": "tool", "name": force_tool_call}
+            tool_choice = {"type": "tool", "name": force_tool_call, "disable_parallel_tool_use": True}
             tools_for_request = [OpenAITool(function=f) for f in tools if f["name"] == force_tool_call]
 
             # need to have this setting to be able to put inner thoughts in kwargs

letta/otel/metrics.py

@@ -1,8 +1,8 @@
 import re
 import time
-from typing import List, Optional
+from typing import List
 
-from fastapi import Depends, FastAPI, Request
+from fastapi import FastAPI, Request
 from opentelemetry import metrics
 from opentelemetry.exporter.otlp.proto.grpc.metric_exporter import OTLPMetricExporter
 from opentelemetry.metrics import NoOpMeter