fix: Make OpenAI context window exceeded error more specific (#2624)

letta/agents/letta_agent.py

@@ -14,7 +14,7 @@ from letta.agents.helpers import (
     _prepare_in_context_messages_no_persist_async,
     generate_step_id,
 )
-from letta.errors import LLMContextWindowExceededError
+from letta.errors import ContextWindowExceededError
 from letta.helpers import ToolRulesSolver
 from letta.helpers.datetime_helpers import get_utc_timestamp_ns
 from letta.helpers.tool_execution_helper import enable_strict_mode
@@ -692,7 +692,7 @@ class LettaAgent(BaseAgent):
         llm_config: LLMConfig,
         force: bool,
     ) -> List[Message]:
-        if isinstance(e, LLMContextWindowExceededError):
+        if isinstance(e, ContextWindowExceededError):
             return await self._rebuild_context_window(
                 in_context_messages=in_context_messages, new_letta_messages=new_letta_messages, llm_config=llm_config, force=force
             )

letta/errors.py

@@ -88,10 +88,6 @@ class LLMPermissionDeniedError(LLMError):
     """Error when permission is denied by LLM service"""
 
 
-class LLMContextWindowExceededError(LLMError):
-    """Error when the context length is exceeded."""
-
-
 class LLMNotFoundError(LLMError):
     """Error when requested resource is not found"""
 

letta/llm_api/openai_client.py

@@ -8,11 +8,11 @@ from openai.types.chat.chat_completion_chunk import ChatCompletionChunk
 
 from letta.constants import LETTA_MODEL_ENDPOINT
 from letta.errors import (
+    ContextWindowExceededError,
     ErrorCode,
     LLMAuthenticationError,
     LLMBadRequestError,
     LLMConnectionError,
-    LLMContextWindowExceededError,
     LLMNotFoundError,
     LLMPermissionDeniedError,
     LLMRateLimitError,
@@ -342,11 +342,9 @@ class OpenAIClient(LLMClientBase):
             # Check message content if finer-grained errors are needed
             # Example: if "context_length_exceeded" in str(e): return LLMContextLengthExceededError(...)
             # TODO: This is a super soft check. Not sure if we can do better, needs more investigation.
-            if "context" in str(e):
-                return LLMContextWindowExceededError(
-                    message=f"Bad request to OpenAI (context length exceeded): {str(e)}",
-                    code=ErrorCode.INVALID_ARGUMENT,  # Or more specific if detectable
-                    details=e.body,
+            if "This model's maximum context length is" in str(e):
+                return ContextWindowExceededError(
+                    message=f"Bad request to OpenAI (context window exceeded): {str(e)}",
                 )
             else:
                 return LLMBadRequestError(