Skip to content

Automate browser based workflows with AI

SKY-10322: skip flex LLM router on UI-triggered runs (#6183)

Shuchang Zheng9d7be5d50186
diff --git a/skyvern/forge/agent_functions.py b/skyvern/forge/agent_functions.py
index 9b1265d3e..1d694a27d 100644
--- a/skyvern/forge/agent_functions.py
+++ b/skyvern/forge/agent_functions.py
@@ -463,6 +463,14 @@ class AgentFunction:
"""
return None
+ def get_non_flex_llm_key(self, llm_key: str | None) -> str | None:
+ """Return a non-flex router twin for the given LLM key, or None if no twin exists.
+
+ Cloud overrides this with the flex-router → standard-router mapping.
+ OSS no-op so self-hosted users without flex routers see no behavior change.
+ """
+ return None
+
async def should_use_flex_llm_routing(
self,
*,
diff --git a/skyvern/forge/sdk/api/llm/api_handler_factory.py b/skyvern/forge/sdk/api/llm/api_handler_factory.py
index 134891d4a..bc48efb74 100644
--- a/skyvern/forge/sdk/api/llm/api_handler_factory.py
+++ b/skyvern/forge/sdk/api/llm/api_handler_factory.py
@@ -44,6 +44,7 @@ from skyvern.forge.sdk.artifact.manager import BulkArtifactCreationRequest
from skyvern.forge.sdk.artifact.models import ArtifactType
from skyvern.forge.sdk.core import skyvern_context
from skyvern.forge.sdk.core.skyvern_context import SkyvernContext
+from skyvern.forge.sdk.db.enums import WorkflowRunTriggerType
from skyvern.forge.sdk.models import SpeculativeLLMMetadata, Step
from skyvern.forge.sdk.schemas.ai_suggestions import AISuggestion
from skyvern.forge.sdk.schemas.task_v2 import TaskV2, Thought
@@ -715,6 +716,36 @@ class LLMAPIHandlerFactory:
)
return None
+ @staticmethod
+ def _maybe_get_non_flex_handler(default: LLMAPIHandler) -> LLMAPIHandler | None:
+ """Mirror of `_maybe_get_flex_handler` for the inverse direction: swap a flex
+ default for its non-flex twin when `trigger_type == manual` (UI/MCP)."""
+ ctx = skyvern_context.current()
+ if ctx is None or ctx.trigger_type != WorkflowRunTriggerType.manual:
+ return None
+ effective_key = (
+ getattr(default, "llm_key", None)
+ or getattr(getattr(default, "__self__", None), "llm_key", None)
+ or settings.LLM_KEY
+ )
+ non_flex_key = app.AGENT_FUNCTION.get_non_flex_llm_key(effective_key)
+ if not non_flex_key or not LLMConfigRegistry.is_registered(non_flex_key):
+ return None
+ try:
+ non_flex_handler = LLMAPIHandlerFactory.get_llm_api_handler(non_flex_key)
+ span = otel_trace.get_current_span()
+ span.set_attribute("non_flex_routing_applied", True)
+ span.set_attribute("non_flex_routing_resolved_key", non_flex_key)
+ return non_flex_handler
+ except Exception:
+ LOG.warning(
+ "Failed to get non-flex LLM API handler, falling back to default (flex) resolution.",
+ non_flex_key=non_flex_key,
+ effective_key=effective_key,
+ exc_info=True,
+ )
+ return None
+
@staticmethod
def wrap_for_flex_routing(handler: LLMAPIHandler) -> LLMAPIHandler:
"""Wrap a default LLM handler so it applies flex routing when context says so.
@@ -726,6 +757,9 @@ class LLMAPIHandlerFactory:
"""
async def flex_aware_handler(*args: Any, **kwargs: Any) -> Any:
+ non_flex_handler = LLMAPIHandlerFactory._maybe_get_non_flex_handler(handler)
+ if non_flex_handler is not None:
+ return await non_flex_handler(*args, **kwargs)
flex_handler = LLMAPIHandlerFactory._maybe_get_flex_handler(handler)
target = flex_handler if flex_handler is not None else handler
return await target(*args, **kwargs)
@@ -750,6 +784,9 @@ class LLMAPIHandlerFactory:
# Flex routing is treated as default-path tier optimization, not a model rewriter.
# Treat empty string as "no override" — block models persist `llm_key=""` rather than NULL.
if not override_llm_key:
+ non_flex_handler = LLMAPIHandlerFactory._maybe_get_non_flex_handler(default)
+ if non_flex_handler is not None:
+ return non_flex_handler
flex_handler = LLMAPIHandlerFactory._maybe_get_flex_handler(default)
if flex_handler is not None:
return flex_handler