feat: cleanup cancellation code and add more logging (#6588)

letta/llm_api/anthropic_client.py

@@ -168,7 +168,12 @@ class AnthropicClient(LLMClientBase):
         if hasattr(llm_config, "response_format") and isinstance(llm_config.response_format, JsonSchemaResponseFormat):
             betas.append("structured-outputs-2025-11-13")
 
-        return await client.beta.messages.create(**request_data, betas=betas)
+        # log failed requests
+        try:
+            return await client.beta.messages.create(**request_data, betas=betas)
+        except Exception as e:
+            logger.error(f"Error streaming Anthropic request: {e} with request data: {json.dumps(request_data)}")
+            raise e
 
     @trace_method
     async def send_llm_batch_request_async(

letta/llm_api/google_vertex_client.py

@@ -140,11 +140,16 @@ class GoogleVertexClient(LLMClientBase):
     @trace_method
     async def stream_async(self, request_data: dict, llm_config: LLMConfig) -> AsyncIterator[GenerateContentResponse]:
         client = self._get_client()
-        response = await client.aio.models.generate_content_stream(
-            model=llm_config.model,
-            contents=request_data["contents"],
-            config=request_data["config"],
-        )
+
+        try:
+            response = await client.aio.models.generate_content_stream(
+                model=llm_config.model,
+                contents=request_data["contents"],
+                config=request_data["config"],
+            )
+        except Exception as e:
+            logger.error(f"Error streaming Google Vertex request: {e} with request data: {json.dumps(request_data)}")
+            raise e
         # Direct yield - keeps response alive in generator's local scope throughout iteration
         # This is required because the SDK's connection lifecycle is tied to the response object
         async for chunk in response:

letta/llm_api/openai_client.py

@@ -1,4 +1,5 @@
 import asyncio
+import json
 import os
 import time
 from typing import Any, List, Optional
@@ -762,17 +763,25 @@ class OpenAIClient(LLMClientBase):
 
         # Route based on payload shape: Responses uses 'input', Chat Completions uses 'messages'
         if "input" in request_data and "messages" not in request_data:
-            response_stream: AsyncStream[ResponseStreamEvent] = await client.responses.create(
-                **request_data,
-                stream=True,
-                # stream_options={"include_usage": True},
-            )
+            try:
+                response_stream: AsyncStream[ResponseStreamEvent] = await client.responses.create(
+                    **request_data,
+                    stream=True,
+                    # stream_options={"include_usage": True},
+                )
+            except Exception as e:
+                logger.error(f"Error streaming OpenAI Responses request: {e} with request data: {json.dumps(request_data)}")
+                raise e
         else:
-            response_stream: AsyncStream[ChatCompletionChunk] = await client.chat.completions.create(
-                **request_data,
-                stream=True,
-                stream_options={"include_usage": True},
-            )
+            try:
+                response_stream: AsyncStream[ChatCompletionChunk] = await client.chat.completions.create(
+                    **request_data,
+                    stream=True,
+                    stream_options={"include_usage": True},
+                )
+            except Exception as e:
+                logger.error(f"Error streaming OpenAI Chat Completions request: {e} with request data: {json.dumps(request_data)}")
+                raise e
         return response_stream
 
     @trace_method