Skip to content

Automate browser based workflows with AI

fix: self-hosted model keys resolve to unregistered fallback configs (#5099)

pedrohsdb1bedb95fcf38
diff --git a/skyvern/config.py b/skyvern/config.py
index 1d7ac4688..8f07cf053 100644
--- a/skyvern/config.py
+++ b/skyvern/config.py
@@ -479,21 +479,29 @@ class Settings(BaseSettings):
Keys are model names available to blocks in the frontend. These map to key names
in LLMConfigRegistry._configs.
"""
- mapping: dict[str, dict[str, str]] = {
- "gemini-2.5-pro-preview-05-06": {"llm_key": "VERTEX_GEMINI_2.5_PRO", "label": "Gemini 2.5 Pro"},
- "gemini-2.5-flash": {
- "llm_key": "VERTEX_GEMINI_2.5_FLASH",
- "label": "Gemini 2.5 Flash",
- },
- "gemini-3-pro-preview": {"llm_key": "VERTEX_GEMINI_3_PRO", "label": "Gemini 3 Pro (Latest)"},
- "gemini-3.0-flash": {"llm_key": "VERTEX_GEMINI_3.0_FLASH", "label": "Gemini 3 Flash"},
- "mercury-2": {"llm_key": "INCEPTION_MERCURY_2", "label": "Inception Mercury 2"},
- "gemini-2.5-flash-lite": {
- "llm_key": "VERTEX_GEMINI_2.5_FLASH_LITE",
- "label": "Gemini 2.5 Flash Lite",
- },
+ mapping: dict[str, dict[str, str]] = {}
+
+ # Gemini models: prefer Vertex when enabled, fall back to direct Gemini API
+ gemini_models = [
+ ("gemini-2.5-pro-preview-05-06", "VERTEX_GEMINI_2.5_PRO", "GEMINI_2.5_PRO", "Gemini 2.5 Pro"),
+ ("gemini-2.5-flash", "VERTEX_GEMINI_2.5_FLASH", "GEMINI_2.5_FLASH", "Gemini 2.5 Flash"),
+ ("gemini-3-pro-preview", "VERTEX_GEMINI_3_PRO", "GEMINI_3_PRO", "Gemini 3 Pro (Latest)"),
+ ("gemini-3.0-flash", "VERTEX_GEMINI_3.0_FLASH", "GEMINI_3.0_FLASH", "Gemini 3 Flash"),
+ ]
+ for model_name, vertex_key, gemini_key, label in gemini_models:
+ mapping[model_name] = {
+ "llm_key": vertex_key if self.ENABLE_VERTEX_AI else gemini_key,
+ "label": label,
+ }
+
+ # Gemini Flash Lite: Vertex-only (no direct Gemini API config exists)
+ mapping["gemini-2.5-flash-lite"] = {
+ "llm_key": "VERTEX_GEMINI_2.5_FLASH_LITE",
+ "label": "Gemini 2.5 Flash Lite",
}
+ mapping["mercury-2"] = {"llm_key": "INCEPTION_MERCURY_2", "label": "Inception Mercury 2"}
+
# GPT models: prefer Azure when enabled, fall back to OpenAI
gpt_models = [
("azure/gpt-4.1", self.ENABLE_AZURE_GPT4_1, "AZURE_OPENAI_GPT4_1", "OPENAI_GPT4_1", "GPT 4.1"),
@@ -537,6 +545,26 @@ class Settings(BaseSettings):
"label": "Anthropic Claude 4.5 Haiku",
}
+ # Anthropic Claude 4.5 Sonnet & Opus
+ if self.ENABLE_BEDROCK_ANTHROPIC:
+ mapping["claude-sonnet-4-5-20250929"] = {
+ "llm_key": "BEDROCK_ANTHROPIC_CLAUDE4.5_SONNET_INFERENCE_PROFILE",
+ "label": "Anthropic Claude 4.5 Sonnet",
+ }
+ mapping["claude-opus-4-5-20251101"] = {
+ "llm_key": "BEDROCK_ANTHROPIC_CLAUDE4.5_OPUS_INFERENCE_PROFILE",
+ "label": "Anthropic Claude 4.5 Opus",
+ }
+ else:
+ mapping["claude-sonnet-4-5-20250929"] = {
+ "llm_key": "ANTHROPIC_CLAUDE4.5_SONNET",
+ "label": "Anthropic Claude 4.5 Sonnet",
+ }
+ mapping["claude-opus-4-5-20251101"] = {
+ "llm_key": "ANTHROPIC_CLAUDE4.5_OPUS",
+ "label": "Anthropic Claude 4.5 Opus",
+ }
+
# Anthropic Claude 4.6 Opus: prefer Bedrock when enabled, fall back to direct API
if self.ENABLE_BEDROCK_ANTHROPIC:
mapping["claude-opus-4-6"] = {
diff --git a/skyvern/forge/sdk/api/llm/api_handler_factory.py b/skyvern/forge/sdk/api/llm/api_handler_factory.py
index e9162fc1d..cea5d4146 100644
--- a/skyvern/forge/sdk/api/llm/api_handler_factory.py
+++ b/skyvern/forge/sdk/api/llm/api_handler_factory.py
@@ -331,6 +331,9 @@ class LLMAPIHandlerFactory:
exc_info=True,
)
+ # Anthropic API requires budget_tokens >= 1024
+ ANTHROPIC_MIN_THINKING_BUDGET = 1024
+
@staticmethod
def _apply_anthropic_thinking_optimization(
parameters: dict[str, Any], new_budget: int, llm_config: LLMConfig | LLMRouterConfig, prompt_name: str
@@ -354,11 +357,13 @@ class LLMAPIHandlerFactory:
model=model_label,
)
else:
- # Use thinking parameter with budget_tokens for Anthropic models
+ # Use thinking parameter with budget_tokens for Anthropic models.
+ # Anthropic API requires budget_tokens >= 1024.
+ clamped_budget = max(new_budget, LLMAPIHandlerFactory.ANTHROPIC_MIN_THINKING_BUDGET)
if "thinking" in parameters and isinstance(parameters["thinking"], dict):
- parameters["thinking"]["budget_tokens"] = new_budget
+ parameters["thinking"]["budget_tokens"] = clamped_budget
else:
- parameters["thinking"] = {"budget_tokens": new_budget, "type": "enabled"}
+ parameters["thinking"] = {"budget_tokens": clamped_budget, "type": "enabled"}
# Get safe model label for logging
model_label = getattr(llm_config, "model_name", None)
if model_label is None and isinstance(llm_config, LLMRouterConfig):
@@ -367,7 +372,7 @@ class LLMAPIHandlerFactory:
LOG.debug(
"Applied thinking budget optimization (thinking)",
prompt_name=prompt_name,
- budget=new_budget,
+ budget=clamped_budget,
model=model_label,
)
diff --git a/tests/unit/test_text_prompt_block.py b/tests/unit/test_text_prompt_block.py
index f1f661cdc..ed87b1aa6 100644
--- a/tests/unit/test_text_prompt_block.py
+++ b/tests/unit/test_text_prompt_block.py
@@ -16,15 +16,13 @@ block_module = sys.modules["skyvern.forge.sdk.workflow.models.block"]
@pytest.mark.asyncio
@pytest.mark.parametrize(
- ("model_name", "expected_llm_key"),
- [
- ("gemini-2.5-flash", "VERTEX_GEMINI_2.5_FLASH"),
- ("gemini-3-pro-preview", "VERTEX_GEMINI_3_PRO"),
- ],
+ "model_name",
+ ["gemini-2.5-flash", "gemini-3-pro-preview"],
)
-async def test_text_prompt_block_uses_selected_model(monkeypatch, model_name, expected_llm_key):
+async def test_text_prompt_block_uses_selected_model(monkeypatch, model_name):
# Reset SettingsManager to base settings so cloud overrides from earlier tests don't leak
monkeypatch.setattr(SettingsManager, "_SettingsManager__instance", base_settings)
+ expected_llm_key = base_settings.get_model_name_to_llm_key()[model_name]["llm_key"]
now = datetime.now(timezone.utc)
output_parameter = OutputParameter(
parameter_type=ParameterType.OUTPUT,
diff --git a/tests/unit/test_text_prompt_block_yaml.py b/tests/unit/test_text_prompt_block_yaml.py
index 6235df0cd..5a67bedba 100644
--- a/tests/unit/test_text_prompt_block_yaml.py
+++ b/tests/unit/test_text_prompt_block_yaml.py
@@ -21,10 +21,13 @@ class TestTextPromptBlockYAMLNormalization:
def test_converts_known_llm_key_to_model(self, monkeypatch) -> None:
monkeypatch.setattr(SettingsManager, "_SettingsManager__instance", base_settings)
+ # Use the actual llm_key that base_settings maps gemini-2.5-flash to
+ # (VERTEX_GEMINI_2.5_FLASH when ENABLE_VERTEX_AI=True, GEMINI_2.5_FLASH otherwise)
+ actual_llm_key = base_settings.get_model_name_to_llm_key()["gemini-2.5-flash"]["llm_key"]
block = TextPromptBlockYAML(
label="summarize",
prompt="Summarize the data.",
- llm_key="VERTEX_GEMINI_2.5_FLASH",
+ llm_key=actual_llm_key,
)
assert block.model == {"model_name": "gemini-2.5-flash"}