adding step_exception to provide more signal of why the step failed (#730)

skyvern/forge/agent.py

@@ -583,6 +583,7 @@ class ForgeAgent:
                     actions=actions,
                     action_results=[],
                     actions_and_results=[],
+                    step_exception=None,
                 )
                 return step, detailed_agent_step_output
 
@@ -776,13 +777,14 @@ class ForgeAgent:
                 step_order=step.order,
                 step_retry=step.retry_index,
             )
+            detailed_agent_step_output.step_exception = "CancelledError"
             failed_step = await self.update_step(
                 step=step,
                 status=StepStatus.failed,
                 output=detailed_agent_step_output.to_agent_step_output(),
             )
             return failed_step, detailed_agent_step_output.get_clean_detailed_output()
-        except Exception:
+        except Exception as e:
             LOG.exception(
                 "Unexpected exception in agent_step, marking step as failed",
                 task_id=task.task_id,
@@ -790,6 +792,7 @@ class ForgeAgent:
                 step_order=step.order,
                 step_retry=step.retry_index,
             )
+            detailed_agent_step_output.step_exception = e.__class__.__name__
             failed_step = await self.update_step(
                 step=step,
                 status=StepStatus.failed,
@@ -882,6 +885,7 @@ class ForgeAgent:
             actions=None,
             action_results=None,
             actions_and_results=None,
+            step_exception=None,
         )
         return step, browser_state, detailed_output
 

skyvern/forge/sdk/api/llm/api_handler_factory.py

@@ -14,6 +14,7 @@ from skyvern.forge.sdk.api.llm.exceptions import (
     DuplicateCustomLLMProviderError,
     InvalidLLMConfigError,
     LLMProviderError,
+    LLMProviderErrorRetryableTask,
 )
 from skyvern.forge.sdk.api.llm.models import LLMAPIHandler, LLMConfig, LLMRouterConfig
 from skyvern.forge.sdk.api.llm.utils import llm_messages_builder, parse_api_response
@@ -211,8 +212,8 @@ class LLMAPIHandlerFactory:
                     **active_parameters,
                 )
                 LOG.info("LLM API call successful", llm_key=llm_key, model=llm_config.model_name)
-            except openai.OpenAIError as e:
+            except litellm.exceptions.APIError as e:
-                raise LLMProviderError(llm_key) from e
+                raise LLMProviderErrorRetryableTask(llm_key) from e
             except CancelledError:
                 t_llm_cancelled = time.perf_counter()
                 LOG.error(

skyvern/forge/sdk/api/llm/exceptions.py

@@ -40,6 +40,11 @@ class LLMProviderError(BaseLLMError):
         super().__init__(f"Error while using LLMProvider {llm_key}")
 
 
+class LLMProviderErrorRetryableTask(LLMProviderError):
+    def __init__(self, llm_key: str) -> None:
+        super().__init__(f"Retryable error while using LLMProvider {llm_key}")
+
+
 class NoProviderEnabledError(BaseLLMError):
     def __init__(self) -> None:
         super().__init__(

skyvern/webeye/actions/models.py

@@ -39,6 +39,7 @@ class DetailedAgentStepOutput(BaseModel):
     actions: list[Action] | None
     action_results: list[ActionResult] | None
     actions_and_results: list[tuple[Action, list[ActionResult]]] | None
+    step_exception: str | None = None
 
     class Config:
         exclude = ["scraped_page", "extract_action_prompt"]