fix: use streaming for Anthropic API to handle long responses

- Switch from messages.create() to messages.stream() for both sync and async
- Required for max_tokens > 21333 or long-running requests
- Add warnings for empty responses and truncation (stop_reason == max_tokens)

Author: XiaoConstantine

rlm/clients/anthropic.py

@@ -1,3 +1,4 @@
+import logging
 from collections import defaultdict
 from typing import Any
 
@@ -6,6 +7,8 @@
 from rlm.clients.base_lm import BaseLM
 from rlm.core.types import ModelUsageSummary, UsageSummary
 
+logger = logging.getLogger(__name__)
+
 
 class AnthropicClient(BaseLM):
     """
@@ -42,9 +45,23 @@ def completion(self, prompt: str | list[dict[str, Any]], model: str | None = Non
         if system:
             kwargs["system"] = system
 
-        response = self.client.messages.create(**kwargs)
+        # Use streaming to handle long responses (required for max_tokens > 21333)
+        parts: list[str] = []
+        with self.client.messages.stream(**kwargs) as stream:
+            for text in stream.text_stream:
+                parts.append(text)
+            response = stream.get_final_message()
+
+        full_response = "".join(parts)
+
+        if not full_response:
+            logger.warning("Received empty response from Anthropic API")
+
+        if response.stop_reason == "max_tokens":
+            logger.warning(f"Response truncated at {self.max_tokens} tokens")
+
         self._track_cost(response, model)
-        return response.content[0].text
+        return full_response
 
     async def acompletion(
         self, prompt: str | list[dict[str, Any]], model: str | None = None
@@ -59,9 +76,23 @@ async def acompletion(
         if system:
             kwargs["system"] = system
 
-        response = await self.async_client.messages.create(**kwargs)
+        # Use streaming to handle long responses (required for max_tokens > 21333)
+        parts: list[str] = []
+        async with self.async_client.messages.stream(**kwargs) as stream:
+            async for text in stream.text_stream:
+                parts.append(text)
+            response = await stream.get_final_message()
+
+        full_response = "".join(parts)
+
+        if not full_response:
+            logger.warning("Received empty response from Anthropic API")
+
+        if response.stop_reason == "max_tokens":
+            logger.warning(f"Response truncated at {self.max_tokens} tokens")
+
         self._track_cost(response, model)
-        return response.content[0].text
+        return full_response
 
     def _prepare_messages(
         self, prompt: str | list[dict[str, Any]]