Speed optimizations: Economy element tree and TOTP context parsing skip (#3936)

skyvern/forge/agent.py

@@ -2066,6 +2066,27 @@ class ForgeAgent:
 
         # If we don't have pre-scraped data, scrape normally
         if scraped_page is None:
+            # Check PostHog for speed optimizations BEFORE scraping
+            # This decision will be used in both:
+            # 1. SVG conversion skip (in agent_functions.py cleanup)
+            # 2. Tree selection (economy vs regular tree)
+            # By checking once and storing in context, we ensure perfect coordination
+            if context:
+                try:
+                    distinct_id = task.workflow_run_id if task.workflow_run_id else task.task_id
+                    context.enable_speed_optimizations = await app.EXPERIMENTATION_PROVIDER.is_feature_enabled_cached(
+                        "ENABLE_SPEED_OPTIMIZATIONS",
+                        distinct_id,
+                        properties={"organization_id": task.organization_id},
+                    )
+                except Exception:
+                    LOG.warning(
+                        "Failed to check ENABLE_SPEED_OPTIMIZATIONS feature flag",
+                        exc_info=True,
+                        task_id=task.task_id,
+                    )
+                    context.enable_speed_optimizations = False
+
             # start the async tasks while running scrape_website
             if engine not in CUA_ENGINES:
                 self.async_operation_pool.run_operation(task.task_id, AgentPhase.scrape)
@@ -2113,7 +2134,51 @@ class ForgeAgent:
         )
         # TODO: we only use HTML element for now, introduce a way to switch in the future
         element_tree_format = ElementTreeFormat.HTML
-        element_tree_in_prompt: str = scraped_page.build_element_tree(element_tree_format)
+
+        # OPTIMIZATION: Use economy tree (skip SVGs) when ENABLE_SPEED_OPTIMIZATIONS is enabled
+        # Economy tree removes all SVG elements from the DOM tree sent to LLM
+        # - SVGs are decorative (icons, logos, graphics) - not needed for action planning
+        # - Even for charts/graphs: LLM sees them in screenshots, not SVG code
+        # - Saves ~8s per SVG x ~15 SVGs = ~120s per workflow (30% speedup!)
+        #
+        # RETRY STRATEGY: Use economy tree on first attempt only
+        # - retry_index 0: Use economy tree (fast, no SVGs)
+        # - retry_index 1+: Use regular tree (SVGs loaded from existing 4-week cache)
+        # Note: SVG conversions are already cached globally with 4-week TTL, so retries are fast
+        #
+        # COORDINATION: The enable_speed_optimizations decision is made ONCE before scraping
+        # and stored in context. Both SVG conversion skip (agent_functions.py) and tree
+        # selection (here) use the SAME value, ensuring perfect coordination.
+        element_tree_in_prompt: str = ""
+
+        # Use the speed optimization decision from context (set before scraping)
+        enable_speed_optimizations = context.enable_speed_optimizations if context else False
+
+        if not enable_speed_optimizations:
+            # Optimization disabled - use regular tree always
+            element_tree_in_prompt = scraped_page.build_element_tree(element_tree_format)
+        elif step.retry_index == 0:
+            # First attempt - use economy tree (fast, no SVG conversion)
+            # Note: SVG conversion was already skipped in cleanup_element_tree_func
+            # based on the same context.enable_speed_optimizations value
+            element_tree_in_prompt = scraped_page.build_economy_elements_tree(element_tree_format)
+            LOG.info(
+                "Speed optimization: Using economy element tree (skipping SVGs)",
+                step_order=step.order,
+                step_retry=step.retry_index,
+                task_id=task.task_id,
+                workflow_run_id=task.workflow_run_id,
+            )
+        else:
+            # Retry 1+ - use regular tree (SVGs will be loaded from existing 4-week cache)
+            element_tree_in_prompt = scraped_page.build_element_tree(element_tree_format)
+            LOG.info(
+                "Speed optimization: Using regular tree on retry (SVGs from global cache)",
+                step_order=step.order,
+                step_retry=step.retry_index,
+                task_id=task.task_id,
+                workflow_run_id=task.workflow_run_id,
+            )
         extract_action_prompt = ""
         if engine not in CUA_ENGINES:
             extract_action_prompt, use_caching = await self._build_extract_action_prompt(

skyvern/forge/agent_functions.py

@@ -571,8 +571,30 @@ class AgentFunction:
                 if "children" in queue_ele:
                     queue.extend(queue_ele["children"])
 
-            # Convert all eligible SVGs in parallel
-            if eligible_svgs:
+            # SPEED OPTIMIZATION: Skip SVG conversion when using economy tree
+            # Economy tree removes SVGs, so no point converting them
+            #
+            # COORDINATION: Use the same enable_speed_optimizations decision from context
+            # that was set in agent.py BEFORE scraping. This ensures SVG conversion skip
+            # is perfectly coordinated with economy tree selection.
+            skip_svg_conversion = False
+            if eligible_svgs and task and step:
+                # Get the optimization decision from context (set before scraping in agent.py)
+                current_context = skyvern_context.current()
+                enable_speed_optimizations = current_context.enable_speed_optimizations if current_context else False
+
+                if enable_speed_optimizations and step.retry_index == 0:
+                    skip_svg_conversion = True
+                    LOG.info(
+                        "Speed optimization: Skipping SVG conversion (will use economy tree)",
+                        step_order=step.order,
+                        step_retry=step.retry_index,
+                        workflow_run_id=task.workflow_run_id,
+                        svg_count=len(eligible_svgs),
+                    )
+
+            # Convert all eligible SVGs in parallel (unless skipped by optimization)
+            if eligible_svgs and not skip_svg_conversion:
                 await asyncio.gather(*[_convert_svg_to_string(element, task, step) for element, frame in eligible_svgs])
 
             return element_tree

skyvern/forge/sdk/core/skyvern_context.py

@@ -37,6 +37,7 @@ class SkyvernContext:
     use_prompt_caching: bool = False
     cached_static_prompt: str | None = None
     vertex_cache_name: str | None = None  # Vertex AI cache resource name for explicit caching
+    enable_speed_optimizations: bool = False
 
     # script run context
     script_id: str | None = None

skyvern/webeye/actions/handler.py

@@ -69,6 +69,7 @@ from skyvern.forge.sdk.api.files import (
 from skyvern.forge.sdk.api.llm.api_handler_factory import LLMAPIHandlerFactory, LLMCallerManager
 from skyvern.forge.sdk.api.llm.exceptions import LLMProviderError
 from skyvern.forge.sdk.core import skyvern_context
+from skyvern.forge.sdk.core.skyvern_context import current as skyvern_current
 from skyvern.forge.sdk.core.skyvern_context import ensure_context
 from skyvern.forge.sdk.models import Step
 from skyvern.forge.sdk.schemas.tasks import Task
@@ -1096,16 +1097,46 @@ async def handle_input_text_action(
     incremental_element: list[dict] = []
     auto_complete_hacky_flag: bool = False
 
-    input_or_select_context = await _get_input_or_select_context(
-        action=action,
-        element_tree_builder=scraped_page,
-        skyvern_element=skyvern_element,
-        step=step,
-    )
+    # OPTIMIZATION: Skip expensive LLM context parsing for TOTP and secret values
+    # TOTP inputs don't need autocomplete detection - we already have the generated code
+    # This saves ~4-5s per TOTP digit (6 digits = ~27s saved for 2FA!)
+    # Gated by ENABLE_SPEED_OPTIMIZATIONS feature flag
+    skip_context_parsing = False
+    if (
+        is_totp_value
+        or is_secret_value
+        or (action.totp_timing_info and action.totp_timing_info.get("is_totp_sequence"))
+    ):
+        try:
+            current_context = skyvern_current()
+            enable_speed_optimizations = current_context.enable_speed_optimizations if current_context else False
+
+            if enable_speed_optimizations:
+                skip_context_parsing = True
+                LOG.info(
+                    "Speed optimization: Skipping input context parsing for TOTP/secret input",
+                    element_id=skyvern_element.get_id(),
+                    is_totp=is_totp_value,
+                    is_secret=is_secret_value,
+                    is_multi_field_totp=bool(action.totp_timing_info),
+                )
+        except Exception:
+            LOG.warning("Failed to read ENABLE_SPEED_OPTIMIZATIONS from context for TOTP optimization", exc_info=True)
+
+    if skip_context_parsing:
+        input_or_select_context = None
+    else:
+        input_or_select_context = await _get_input_or_select_context(
+            action=action,
+            element_tree_builder=scraped_page,
+            skyvern_element=skyvern_element,
+            step=step,
+        )
 
     # check if it's selectable
     if (
-        not input_or_select_context.is_search_bar  # no need to to trigger selection logic for search bar
+        input_or_select_context is not None
+        and not input_or_select_context.is_search_bar  # no need to to trigger selection logic for search bar
         and not is_totp_value
         and not is_secret_value
         and skyvern_element.get_tag_name() == InteractiveElement.INPUT
@@ -1361,7 +1392,8 @@ async def handle_input_text_action(
             return [ActionSuccess()]
 
         if not await skyvern_element.is_raw_input():
-            if await skyvern_element.is_auto_completion_input() or input_or_select_context.is_location_input:
+            is_location_input = input_or_select_context.is_location_input if input_or_select_context else False
+            if input_or_select_context and (await skyvern_element.is_auto_completion_input() or is_location_input):
                 if result := await input_or_auto_complete_input(
                     input_or_select_context=input_or_select_context,
                     scraped_page=scraped_page,