setting up a separate handler for parse input or select, for experiment (#3449)

skyvern/config.py

@@ -121,6 +121,7 @@ class Settings(BaseSettings):
     SINGLE_CLICK_AGENT_LLM_KEY: str | None = None
     SINGLE_INPUT_AGENT_LLM_KEY: str | None = None
     PROMPT_BLOCK_LLM_KEY: str | None = None
+    PARSE_SELECT_LLM_KEY: str | None = None
     EXTRACTION_LLM_KEY: str | None = None
     # COMMON
     LLM_CONFIG_TIMEOUT: int = 300

skyvern/forge/app.py

@@ -68,6 +68,11 @@ SINGLE_CLICK_AGENT_LLM_API_HANDLER = LLMAPIHandlerFactory.get_llm_api_handler(
 SINGLE_INPUT_AGENT_LLM_API_HANDLER = LLMAPIHandlerFactory.get_llm_api_handler(
     SETTINGS_MANAGER.SINGLE_INPUT_AGENT_LLM_KEY or SETTINGS_MANAGER.SECONDARY_LLM_KEY or SETTINGS_MANAGER.LLM_KEY
 )
+PARSE_SELECT_LLM_API_HANDLER = (
+    LLMAPIHandlerFactory.get_llm_api_handler(SETTINGS_MANAGER.PARSE_SELECT_LLM_KEY)
+    if SETTINGS_MANAGER.PARSE_SELECT_LLM_KEY
+    else SECONDARY_LLM_API_HANDLER
+)
 EXTRACTION_LLM_API_HANDLER = (
     LLMAPIHandlerFactory.get_llm_api_handler(SETTINGS_MANAGER.EXTRACTION_LLM_KEY)
     if SETTINGS_MANAGER.EXTRACTION_LLM_KEY

skyvern/webeye/actions/handler.py

@@ -3723,7 +3723,8 @@ async def _get_input_or_select_context(
         action_reasoning=action.reasoning,
         element_id=action.element_id,
     )
-    json_response = await app.SECONDARY_LLM_API_HANDLER(
+    # Use centralized parse-select handler (set at init or via scripts)
+    json_response = await app.PARSE_SELECT_LLM_API_HANDLER(
         prompt=prompt, step=step, prompt_name="parse-input-or-select-context"
     )
     json_response["intention"] = action.intention