Default Skyvern 2.0 to google.com if generated website fails to load (#2291)

skyvern/config.py

@@ -238,6 +238,8 @@ class Settings(BaseSettings):
     ENABLE_LOG_ARTIFACTS: bool = False
     ENABLE_CODE_BLOCK: bool = False
 
+    TASK_BLOCKED_SITE_FALLBACK_URL: str = "https://www.google.com"
+
     # SkyvernClient Settings
     SKYVERN_BASE_URL: str = "https://api.skyvern.com"
     SKYVERN_API_KEY: str = "PLACEHOLDER"

skyvern/services/task_v2_service.py

@@ -10,7 +10,13 @@ from playwright.async_api import Page
 from sqlalchemy.exc import OperationalError
 
 from skyvern.config import settings
-from skyvern.exceptions import FailedToSendWebhook, TaskTerminationError, TaskV2NotFound, UrlGenerationFailure
+from skyvern.exceptions import (
+    FailedToSendWebhook,
+    MissingBrowserState,
+    TaskTerminationError,
+    TaskV2NotFound,
+    UrlGenerationFailure,
+)
 from skyvern.forge import app
 from skyvern.forge.prompts import prompt_engine
 from skyvern.forge.sdk.artifact.models import ArtifactType
@@ -481,31 +487,86 @@ async def run_task_v2_helper(
         task_history_record: dict[str, Any] = {}
         context = skyvern_context.ensure_context()
 
+        # Always ensure browser_state is available at the start of the loop
         current_url: str | None = None
         page: Page | None = None
-        browser_state = app.BROWSER_MANAGER.get_for_workflow_run(workflow_run_id, workflow_run.parent_workflow_run_id)
-        if browser_state:
+        fallback_url = settings.TASK_BLOCKED_SITE_FALLBACK_URL
+        browser_state = app.BROWSER_MANAGER.get_for_workflow_run(
+            workflow_run_id=workflow_run_id, parent_workflow_run_id=workflow_run.parent_workflow_run_id
+        )
+        if browser_state is None:
+            fallback_occurred = False
+            try:
+                browser_state = await app.BROWSER_MANAGER.get_or_create_for_workflow_run(
+                    workflow_run=workflow_run,
+                    url=url,
+                    browser_session_id=browser_session_id,
+                )
+            except Exception:
+                LOG.warning("Failed to get or create browser state, fallback to Google", exc_info=True, url=url)
+
+                browser_state = await app.BROWSER_MANAGER.get_or_create_for_workflow_run(
+                    workflow_run=workflow_run,
+                    url=fallback_url,
+                    browser_session_id=browser_session_id,
+                )
+
+                fallback_occurred = True
+
+            if browser_state is None:
+                LOG.error("Failed to create browser state even after fallback", workflow_run_id=workflow_run_id)
+                raise MissingBrowserState(workflow_run_id=workflow_run_id)
+
+            page = await browser_state.get_working_page()
+
+            page_loaded = False
+            if page:
+                try:
+                    # Check if the page has a body element to verify it loaded
+                    # page will always be None if browser state failed to load
+                    page_loaded = await browser_state.validate_browser_context(page)
+                except Exception:
+                    page_loaded = False
+                    LOG.warning(
+                        "Page failed to load properly, fallback to Google",
+                        exc_info=True,
+                        url=url,
+                        current_url=current_url,
+                    )
+
+            if not page_loaded:
+                # Page failed to load properly, fallback to Google
+                if page:
+                    try:
+                        await page.goto(fallback_url, timeout=15000)
+                        fallback_occurred = True
+                    except Exception:
+                        LOG.exception("Failed to load Google fallback", exc_info=True, url=url, current_url=current_url)
+        else:
             page = await browser_state.get_working_page()
             if page:
                 current_url = await SkyvernFrame.get_url(page)
 
         if i == 0 and current_url != url:
-            # The first iteration is always a GOTO_URL task
-            task_type = "goto_url"
-            plan = f"Go to this website: {url}"
-            task_history_record = {"type": task_type, "task": plan}
-            block, block_yaml_list, parameter_yaml_list = await _generate_goto_url_task(
-                workflow_id=workflow_id,
-                url=url,
-            )
+            if fallback_occurred:
+                plan = f"Go to Google because the intended website ({url}) failed to load properly."
+                task_type = "goto_url"
+                task_history_record = {"type": task_type, "task": plan}
+                block, block_yaml_list, parameter_yaml_list = await _generate_goto_url_task(
+                    workflow_id=workflow_id,
+                    url=fallback_url,
+                )
+            else:
+                # Page loaded successfully, proceed with original URL
+                plan = f"Go to this website: {url}"
+                task_type = "goto_url"
+                task_history_record = {"type": task_type, "task": plan}
+                block, block_yaml_list, parameter_yaml_list = await _generate_goto_url_task(
+                    workflow_id=workflow_id,
+                    url=url,
+                )
         else:
             try:
-                if browser_state is None:
-                    browser_state = await app.BROWSER_MANAGER.get_or_create_for_workflow_run(
-                        workflow_run=workflow_run,
-                        url=url,
-                        browser_session_id=browser_session_id,
-                    )
                 scraped_page = await scrape_website(
                     browser_state,
                     url,

skyvern/webeye/browser_factory.py

@@ -274,16 +274,6 @@ class BrowserContextFactory:
 
             raise UnknownErrorWhileCreatingBrowserContext(browser_type, e) from e
 
-    @classmethod
-    def set_validate_browser_context(cls, validator: Callable[[Page], Awaitable[bool]]) -> None:
-        cls._validator = validator
-
-    @classmethod
-    async def validate_browser_context(cls, page: Page) -> bool:
-        if cls._validator is None:
-            return True
-        return await cls._validator(page)
-
 
 class VideoArtifact(BaseModel):
     video_path: str | None = None
@@ -603,6 +593,7 @@ class BrowserState:
                         loading_time=end_time - start_time,
                         url=url,
                     )
+                    # Do we need this?
                     await asyncio.sleep(5)
                     LOG.info(f"Successfully went to {url}", url=url, retry_time=retry_time)
                     return
@@ -639,6 +630,30 @@ class BrowserState:
         await self.set_working_page(last_page, len(self.browser_context.pages) - 1)
         return last_page
 
+    async def validate_browser_context(self, page: Page) -> bool:
+        # validate the content
+        try:
+            skyvern_frame = await SkyvernFrame.create_instance(frame=page)
+            html = await skyvern_frame.get_content()
+        except Exception:
+            LOG.error(
+                "Error happened while getting the first page content",
+                exc_info=True,
+            )
+            return False
+
+        if "Bad gateway error" in html:
+            LOG.warning("Bad gateway error on the page, recreate a new browser context with another proxy node")
+            return False
+
+        if "client_connect_forbidden_host" in html:
+            LOG.warning(
+                "capture the client_connect_forbidden_host error on the page, recreate a new browser context with another proxy node"
+            )
+            return False
+
+        return True
+
     async def must_get_working_page(self) -> Page:
         page = await self.get_working_page()
         assert page is not None
@@ -710,7 +725,7 @@ class BrowserState:
             )
         page = await self.__assert_page()
 
-        if not await BrowserContextFactory.validate_browser_context(await self.get_working_page()):
+        if not await self.validate_browser_context(await self.get_working_page()):
             if not await self.close_current_open_page():
                 LOG.warning("Failed to close the current open page, going to skip the browser context validation")
                 return page