Skip to content

armgit / letta

Public Imported from GitHub
0

Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.

fix(core): handle MCP errors and API key whitespace (#9306)

Kian Jones662ec082cf22
diff --git a/letta/llm_api/openai_client.py b/letta/llm_api/openai_client.py
index 991e8d84d..482013620 100644
--- a/letta/llm_api/openai_client.py
+++ b/letta/llm_api/openai_client.py
@@ -620,12 +620,20 @@ class OpenAIClient(LLMClientBase):
client = OpenAI(**self._prepare_client_kwargs(llm_config))
# Route based on payload shape: Responses uses 'input', Chat Completions uses 'messages'
- if "input" in request_data and "messages" not in request_data:
- resp = client.responses.create(**request_data)
- return resp.model_dump()
- else:
- response: ChatCompletion = client.chat.completions.create(**request_data)
- return response.model_dump()
+ try:
+ if "input" in request_data and "messages" not in request_data:
+ resp = client.responses.create(**request_data)
+ return resp.model_dump()
+ else:
+ response: ChatCompletion = client.chat.completions.create(**request_data)
+ return response.model_dump()
+ except json.JSONDecodeError as e:
+ logger.error(f"[OpenAI] Failed to parse API response as JSON: {e}")
+ raise LLMServerError(
+ message=f"OpenAI API returned invalid JSON response (likely an HTML error page): {str(e)}",
+ code=ErrorCode.INTERNAL_SERVER_ERROR,
+ details={"json_error": str(e), "error_position": f"line {e.lineno} column {e.colno}"},
+ )
@trace_method
async def request_async(self, request_data: dict, llm_config: LLMConfig) -> dict:
@@ -638,12 +646,20 @@ class OpenAIClient(LLMClientBase):
kwargs = await self._prepare_client_kwargs_async(llm_config)
client = AsyncOpenAI(**kwargs)
# Route based on payload shape: Responses uses 'input', Chat Completions uses 'messages'
- if "input" in request_data and "messages" not in request_data:
- resp = await client.responses.create(**request_data)
- return resp.model_dump()
- else:
- response: ChatCompletion = await client.chat.completions.create(**request_data)
- return response.model_dump()
+ try:
+ if "input" in request_data and "messages" not in request_data:
+ resp = await client.responses.create(**request_data)
+ return resp.model_dump()
+ else:
+ response: ChatCompletion = await client.chat.completions.create(**request_data)
+ return response.model_dump()
+ except json.JSONDecodeError as e:
+ logger.error(f"[OpenAI] Failed to parse API response as JSON: {e}")
+ raise LLMServerError(
+ message=f"OpenAI API returned invalid JSON response (likely an HTML error page): {str(e)}",
+ code=ErrorCode.INTERNAL_SERVER_ERROR,
+ details={"json_error": str(e), "error_position": f"line {e.lineno} column {e.colno}"},
+ )
def is_reasoning_model(self, llm_config: LLMConfig) -> bool:
return is_openai_reasoning_model(llm_config.model)
diff --git a/letta/schemas/providers/base.py b/letta/schemas/providers/base.py
index 73e4a239e..bad771649 100644
--- a/letta/schemas/providers/base.py
+++ b/letta/schemas/providers/base.py
@@ -263,6 +263,11 @@ class ProviderCreate(ProviderBase):
base_url: str | None = Field(None, description="Base URL used for requests to the provider.")
api_version: str | None = Field(None, description="API version used for requests to the provider.")
+ @field_validator("api_key", "access_key", mode="before")
+ @classmethod
+ def strip_whitespace(cls, v: str | None) -> str | None:
+ return v.strip() if isinstance(v, str) else v
+
class ProviderUpdate(ProviderBase):
api_key: str = Field(..., description="API key or secret key used for requests to the provider.")
@@ -271,6 +276,11 @@ class ProviderUpdate(ProviderBase):
base_url: str | None = Field(None, description="Base URL used for requests to the provider.")
api_version: str | None = Field(None, description="API version used for requests to the provider.")
+ @field_validator("api_key", "access_key", mode="before")
+ @classmethod
+ def strip_whitespace(cls, v: str | None) -> str | None:
+ return v.strip() if isinstance(v, str) else v
+
class ProviderCheck(BaseModel):
provider_type: ProviderType = Field(..., description="The type of the provider.")
@@ -279,3 +289,8 @@ class ProviderCheck(BaseModel):
region: str | None = Field(None, description="Region used for requests to the provider.")
base_url: str | None = Field(None, description="Base URL used for requests to the provider.")
api_version: str | None = Field(None, description="API version used for requests to the provider.")
+
+ @field_validator("api_key", "access_key", mode="before")
+ @classmethod
+ def strip_whitespace(cls, v: str | None) -> str | None:
+ return v.strip() if isinstance(v, str) else v
diff --git a/letta/services/mcp/base_client.py b/letta/services/mcp/base_client.py
index d297e83fd..6b4a0536d 100644
--- a/letta/services/mcp/base_client.py
+++ b/letta/services/mcp/base_client.py
@@ -87,6 +87,9 @@ class AsyncBaseMCPClient:
# Log at debug level to avoid triggering production alerts for expected failures
if e.__class__.__name__ in ("McpError", "ToolError"):
logger.debug(f"MCP tool '{tool_name}' execution failed: {str(e)}")
+ # Return error message with failure status instead of raising to avoid Datadog alerts
+ return str(e), False
+ # Re-raise unexpected errors
raise
parsed_content = []