Automate browser based workflows with AI
Python aiapiautomationbrowserbrowser-automationcomputergptllmplaywrightpowerautomatepuppeteerpythonrpaseleniumvisionworkflow
Add a Gemini API flex-tier config for gemini-3.1-flash-lite (#7949)
Shuchang Zhengab7e2d9f6522
diff --git a/skyvern/forge/sdk/api/llm/config_registry.py b/skyvern/forge/sdk/api/llm/config_registry.pyindex 598b18d7d..4198b281e 100644--- a/skyvern/forge/sdk/api/llm/config_registry.py+++ b/skyvern/forge/sdk/api/llm/config_registry.py@@ -1447,6 +1447,23 @@ if settings.ENABLE_GEMINI:),),)+ # litellm prices the gemini/ flex tier itself (50% of standard), unlike the Vertex flex+ # configs whose discount is applied at the cost site in api_handler_factory.+ LLMConfigRegistry.register_config(+ "GEMINI_3.1_FLASH_LITE_FLEX",+ LLMConfig(+ "gemini/gemini-3.1-flash-lite",+ ["GEMINI_API_KEY"],+ supports_vision=True,+ add_assistant_prefix=False,+ max_completion_tokens=65536,+ litellm_params=LiteLLMParams(+ thinking_level="medium" if settings.GEMINI_INCLUDE_THOUGHT else "minimal",+ service_tier="flex",+ timeout=FLEX_EXECUTION_TIMEOUT_SECONDS,+ ),+ ),+ )LLMConfigRegistry.register_config("GEMINI_3.5_FLASH_LITE",LLMConfig(diff --git a/tests/unit_tests/test_gemini_flex_config.py b/tests/unit_tests/test_gemini_flex_config.pynew file mode 100644index 000000000..f057c8e1e--- /dev/null+++ b/tests/unit_tests/test_gemini_flex_config.py@@ -0,0 +1,68 @@+import dataclasses+import importlib+from collections.abc import Iterator++import pytest++from skyvern import config+from skyvern.config import Settings+from skyvern.forge.sdk.api.llm import config_registry+from skyvern.forge.sdk.settings_manager import SettingsManager+from skyvern.schemas.llm import LLMConfig+++def gemini_registry(monkeypatch: pytest.MonkeyPatch) -> Iterator[None]:+ """ENABLE_GEMINI is off by default, so the Gemini block only registers on a reload with it+ set. The registry dict, config_registry's module-level `settings` binding and SettingsManager+ are all process-global; snapshot and restore the three so the reload can't leak into other+ tests (or be leaked into, depending on execution order).+ """+ previous_settings = SettingsManager.get_settings()+ previous_configs = dict(config_registry.LLMConfigRegistry._configs)+ previous_module_settings = config_registry.settings++ gemini_settings = Settings(ENABLE_GEMINI=True, GEMINI_API_KEY="test-key")+ SettingsManager.set_settings(gemini_settings)+ monkeypatch.setattr(config, "settings", gemini_settings)+ config_registry.LLMConfigRegistry._configs.clear()+ importlib.reload(config_registry)+ try:+ yield+ finally:+ SettingsManager.set_settings(previous_settings)+ config_registry.LLMConfigRegistry._configs.clear()+ config_registry.LLMConfigRegistry._configs.update(previous_configs)+ config_registry.settings = previous_module_settings+++def test_gemini_3_1_flash_lite_flex_requests_the_flex_service_tier(gemini_registry: None) -> None:+ assert config_registry.LLMConfigRegistry.is_registered("GEMINI_3.1_FLASH_LITE_FLEX")++ flex = config_registry.LLMConfigRegistry.get_config("GEMINI_3.1_FLASH_LITE_FLEX")+ assert isinstance(flex, LLMConfig)+ assert flex.model_name == "gemini/gemini-3.1-flash-lite"+ assert flex.required_env_vars == ["GEMINI_API_KEY"]+ assert flex.litellm_params is not None+ assert flex.litellm_params["service_tier"] == "flex"+ assert flex.litellm_params["timeout"] == config_registry.FLEX_EXECUTION_TIMEOUT_SECONDS+++def test_gemini_3_1_flash_lite_flex_differs_from_standard_only_by_tier(gemini_registry: None) -> None:+ """The flex twin must stay a pure tier swap: any other divergence (model id, token ceiling,+ thinking level) would make flex and standard runs silently non-comparable."""+ flex = config_registry.LLMConfigRegistry.get_config("GEMINI_3.1_FLASH_LITE_FLEX")+ standard = config_registry.LLMConfigRegistry.get_config("GEMINI_3.1_FLASH_LITE")+ assert isinstance(flex, LLMConfig)+ assert isinstance(standard, LLMConfig)++ flex_fields = dataclasses.asdict(flex)+ standard_fields = dataclasses.asdict(standard)+ assert flex_fields.pop("litellm_params") is not None+ assert standard_fields.pop("litellm_params") is not None+ assert flex_fields == standard_fields++ assert flex.litellm_params is not None+ assert standard.litellm_params is not None+ tier_only = {k: v for k, v in flex.litellm_params.items() if k not in ("service_tier", "timeout")}+ assert tier_only == dict(standard.litellm_params)