vodorod/Dorod-Sky
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

optimize max steps failure reason for task v2 (#2225)

Browse Source
This commit is contained in:
Shuchang Zheng
2025-04-24 14:46:00 +08:00
committed by GitHub
parent 56cfb55096
commit 9e790cd82c
3 changed files with 83 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
skyvern/forge/prompts/skyvern/task_v2_summarize-max-steps-reason.j2 Normal file
View File

@@ -0,0 +1,20 @@
The user is performing tasks on the webpage to achieve the user goal. The user breaks the goal down into several mini plans to execute step by step. You are here to help the user summarize the main reason why the user goal has not been achieved within {{ block_cnt }} mini plans. This summary should be based on the provided screenshot, the user goal and the results of each mini plan.
Make sure to ONLY return the JSON object in this format with no additional text before or after it:
```json
{
"page_info": str, // Think step by step. Describe all the useful information in the page related to the user goal.
"reasoning": str, // Think step by step. Describe the reason you summarized based on 'page_info', screenshots, user goal and result of mini plans. Keep the reasoning short and to the point.
}
User Goal:
{{ navigation_goal }}
Mini plans and the results:
{% for each_block in history %}{{ each_block }}
{% endfor %}
Current datetime, ISO format:
```
{{ local_datetime }}
```

2
skyvern/forge/sdk/schemas/task_v2.py
View File

@@ -61,11 +61,13 @@ class ThoughtType(StrEnum):
metadata = "metadata"
user_goal_check = "user_goal_check"
internal_plan = "internal_plan"
failure_describe = "failure_describe"
class ThoughtScenario(StrEnum):
generate_plan = "generate_plan"
user_goal_check = "user_goal_check"
failure_describe = "failure_describe"
summarization = "summarization"
generate_metadata = "generate_metadata"
extract_loop_values = "extract_loop_values"

64
skyvern/services/task_v2_service.py
View File

@@ -85,6 +85,62 @@ def _generate_data_extraction_schema_for_loop(loop_values_key: str) -> dict:
}
async def _summarize_max_steps_failure_reason(
task_v2: TaskV2, organization_id: str, browser_state: BrowserState | None
) -> str:
"""
Summarize the failure reason for the task v2.
"""
try:
assert task_v2.workflow_run_id is not None
if browser_state is None:
return "Failed to start browser"
page = await browser_state.get_working_page()
if page is None:
return "Failed to get the current browser page"
screenshots = await SkyvernFrame.take_split_screenshots(page=page, url=str(task_v2.url), draw_boxes=False)
run_blocks = await app.DATABASE.get_workflow_run_blocks(
workflow_run_id=task_v2.workflow_run_id,
organization_id=organization_id,
)
history = [f"{idx + 1}. {block.description} -- {block.status}" for idx, block in enumerate(run_blocks[::-1])]
thought = await app.DATABASE.create_thought(
task_v2_id=task_v2.observer_cruise_id,
organization_id=task_v2.organization_id,
workflow_run_id=task_v2.workflow_run_id,
workflow_id=task_v2.workflow_id,
workflow_permanent_id=task_v2.workflow_permanent_id,
thought_type=ThoughtType.failure_describe,
thought_scenario=ThoughtScenario.failure_describe,
)
context = skyvern_context.ensure_context()
prompt = prompt_engine.load_prompt(
template="task_v2_summarize-max-steps-reason",
block_cnt=len(run_blocks),
navigation_goal=task_v2.prompt,
history=history,
local_datetime=datetime.now(context.tz_info).isoformat(),
)
json_response = await app.LLM_API_HANDLER(
prompt=prompt,
screenshots=screenshots,
prompt_name="task_v2_summarize-max-steps-reason",
thought=thought,
)
return json_response.get("reasoning", "")
except Exception:
LOG.warning("Failed to summarize the failure reason for task v2", exc_info=True)
return ""
async def initialize_task_v2(
organization: Organization,
user_prompt: str,
@@ -386,6 +442,7 @@ async def run_task_v2_helper(
task_history: list[dict] = []
yaml_blocks: list[BLOCK_YAML_TYPES] = []
yaml_parameters: list[PARAMETER_YAML_TYPES] = []
browser_state: BrowserState | None = None
max_steps = int_max_steps_override or settings.MAX_STEPS_PER_TASK_V2
for i in range(DEFAULT_MAX_ITERATIONS):
@@ -733,10 +790,11 @@ async def run_task_v2_helper(
)
if total_step_count >= max_steps:
LOG.info("Task v2 failed - run out of steps", max_steps=max_steps, workflow_run_id=workflow_run_id)
failure_reason = await _summarize_max_steps_failure_reason(task_v2, organization_id, browser_state)
await mark_task_v2_as_failed(
task_v2_id=task_v2_id,
workflow_run_id=workflow_run_id,
failure_reason=f'Reached the max number of {max_steps} steps. If you need more steps, update the "Max Steps Override" configuration when running the task. Or add/update the "x-max-steps-override" header with your desired number of steps in the API request.',
failure_reason=f'Reached the max number of {max_steps} steps. Possible failure reasons: {failure_reason} If you need more steps, update the "Max Steps Override" configuration when running the task. Or add/update the "x-max-steps-override" header with your desired number of steps in the API request.',
organization_id=organization_id,
)
return workflow, workflow_run, task_v2
@@ -747,11 +805,11 @@ async def run_task_v2_helper(
max_steps=max_steps,
workflow_run_id=workflow_run_id,
)
failure_reason = await _summarize_max_steps_failure_reason(task_v2, organization_id, browser_state)
task_v2 = await mark_task_v2_as_failed(
task_v2_id=task_v2_id,
workflow_run_id=workflow_run_id,
# TODO: add a better failure reason with LLM
failure_reason="Max iterations reached",
failure_reason=f"Max iterations reached. Possible failure reasons: {failure_reason}",
organization_id=organization_id,
)