Skip to content

Automate browser based workflows with AI

fix: create fallback episodes for code_v2 runs (#5021)

pedrohsdb056005f7c9bf
diff --git a/skyvern/services/script_service.py b/skyvern/services/script_service.py
index c43eb0066..3689aa086 100644
--- a/skyvern/services/script_service.py
+++ b/skyvern/services/script_service.py
@@ -50,7 +50,7 @@ from skyvern.forge.sdk.workflow.models.block import (
ValidationBlock,
)
from skyvern.forge.sdk.workflow.models.parameter import PARAMETER_TYPE, OutputParameter, ParameterType
-from skyvern.forge.sdk.workflow.models.workflow import Workflow
+from skyvern.forge.sdk.workflow.models.workflow import Workflow, is_adaptive_caching
from skyvern.schemas.runs import RunEngine
from skyvern.schemas.scripts import (
CreateScriptResponse,
@@ -1109,6 +1109,80 @@ async def _fallback_to_ai_run(
)
return
+ # Record a fallback episode for adaptive caching (code_v2).
+ # This must happen BEFORE the AI step runs so we capture the page state
+ # at the moment of failure, not after the AI agent has modified the page.
+ # NOTE: This mirrors _record_fallback_episode() in workflow/service.py.
+ # The two implementations should be kept in sync if the episode schema changes.
+ fallback_episode_id: str | None = None
+ form_fields_snapshot: list | None = None
+ if workflow_permanent_id and is_adaptive_caching(workflow, workflow_run):
+ try:
+ # Capture page state at the moment of script failure
+ page_url = None
+ page_text_snapshot = None
+ working_page = None
+ try:
+ browser_state = await app.BROWSER_MANAGER.get_or_create_for_workflow_run(
+ workflow_run=workflow_run,
+ )
+ working_page = await browser_state.get_working_page()
+ if working_page:
+ page_url = working_page.url
+ page_text_snapshot = (await working_page.inner_text("body", timeout=5000))[:5000]
+ except Exception:
+ LOG.debug("Failed to capture page state for fallback episode", exc_info=True)
+
+ # Extract structured form field metadata from the DOM
+ try:
+ if working_page:
+ form_fields_snapshot = await working_page.evaluate("""() => {
+ const fields = [];
+ for (const el of document.querySelectorAll('input, select, textarea')) {
+ if (el.type === 'hidden') continue;
+ const labelEl = el.closest('label')
+ || (el.id && document.querySelector('label[for="' + el.id + '"]'));
+ const label = labelEl ? labelEl.textContent.trim().substring(0, 100) : '';
+ const ariaLabel = el.getAttribute('aria-label') || '';
+ const placeholder = el.getAttribute('placeholder') || '';
+ if (!label && !ariaLabel && !placeholder && !el.name) continue;
+ fields.push({
+ tag: el.tagName.toLowerCase(),
+ type: el.getAttribute('type') || el.tagName.toLowerCase(),
+ label: label,
+ name: el.getAttribute('name') || '',
+ required: el.required || el.getAttribute('aria-required') === 'true',
+ placeholder: placeholder,
+ });
+ }
+ return fields.slice(0, 50);
+ }""")
+ except Exception:
+ LOG.debug("Failed to extract form field metadata for fallback episode", exc_info=True)
+
+ # _fallback_to_ai_run is only called for TaskBlock-style blocks (navigation,
+ # extraction, action, login, download), never for ConditionalBlock, so
+ # fallback_type is always "full_block" here.
+ episode = await app.DATABASE.create_fallback_episode(
+ organization_id=organization_id,
+ workflow_permanent_id=workflow_permanent_id,
+ workflow_run_id=workflow_run_id,
+ block_label=cache_key,
+ fallback_type="full_block",
+ script_revision_id=context.script_revision_id,
+ classify_result=context.last_classify_result,
+ error_message=str(error)[:2000] if error else None,
+ page_url=page_url,
+ page_text_snapshot=page_text_snapshot,
+ )
+ fallback_episode_id = episode.episode_id
+ except Exception:
+ LOG.warning(
+ "Failed to record fallback episode in _fallback_to_ai_run",
+ block_label=cache_key,
+ exc_info=True,
+ )
+
# 2. create a new step for ai run
ai_step = await app.DATABASE.create_step(
task_id=task_id,
@@ -1222,6 +1296,51 @@ async def _fallback_to_ai_run(
except Exception as e:
LOG.warning("Failed to regenerate script block after AI fallback", error=str(e), exc_info=True)
# Don't fail the entire fallback process if script regeneration fails
+
+ # Update fallback episode with AI execution results
+ if fallback_episode_id:
+ try:
+ fallback_succeeded = task.status not in [TaskStatus.terminated, TaskStatus.failed]
+ agent_actions_summary: dict[str, Any] = {
+ "block_status": str(task.status),
+ }
+ if form_fields_snapshot:
+ agent_actions_summary["form_fields"] = form_fields_snapshot
+ if not fallback_succeeded and task.failure_reason:
+ agent_actions_summary["failure_reason"] = str(task.failure_reason)[:2000]
+ # Fetch actions from the AI step.
+ # NOTE: service.py uses _build_action_summary() which includes element
+ # attributes and CSS selectors. We can't import it here (circular import:
+ # service.py imports script_service). Core fields suffice for now.
+ try:
+ actions = await app.DATABASE.get_task_actions(
+ task_id=task_id,
+ organization_id=organization_id,
+ )
+ agent_actions_summary["actions"] = [
+ {
+ "action_type": a.action_type,
+ "intention": a.intention,
+ "reasoning": a.reasoning,
+ "status": a.status,
+ }
+ for a in actions[:20]
+ ]
+ except Exception:
+ LOG.debug("Could not fetch actions for fallback episode", exc_info=True)
+
+ await app.DATABASE.update_fallback_episode(
+ episode_id=fallback_episode_id,
+ organization_id=organization_id,
+ agent_actions=agent_actions_summary,
+ fallback_succeeded=fallback_succeeded,
+ )
+ except Exception:
+ LOG.warning(
+ "Failed to update fallback episode with agent actions",
+ episode_id=fallback_episode_id,
+ exc_info=True,
+ )
except Exception as e:
LOG.warning("Failed to fallback to AI run", cache_key=cache_key, exc_info=True)
# Update block status to failed if workflow block was created