Xingyao Wang ca424ec15d
[agent] Add LLM risk analyzer (#9349)
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: openhands <openhands@all-hands.dev>
Co-authored-by: Graham Neubig <neubig@gmail.com>
Co-authored-by: llamantino <213239228+llamantino@users.noreply.github.com>
Co-authored-by: mamoodi <mamoodiha@gmail.com>
Co-authored-by: Tim O'Farrell <tofarr@gmail.com>
Co-authored-by: Hiep Le <69354317+hieptl@users.noreply.github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Ryan H. Tran <descience.thh10@gmail.com>
Co-authored-by: Neeraj Panwar <49247372+npneeraj@users.noreply.github.com>
Co-authored-by: sp.wack <83104063+amanape@users.noreply.github.com>
Co-authored-by: Insop <1240382+insop@users.noreply.github.com>
Co-authored-by: test <test@test.com>
Co-authored-by: Engel Nyst <enyst@users.noreply.github.com>
Co-authored-by: Zhonghao Jiang <zhonghao.J@outlook.com>
Co-authored-by: Ray Myers <ray.myers@gmail.com>
2025-08-22 14:02:36 +00:00

179 lines
6.2 KiB
Python

from dataclasses import asdict
from datetime import datetime
from enum import Enum
from typing import Any
from pydantic import BaseModel
from openhands.events import Event, EventSource
from openhands.events.serialization.action import action_from_dict
from openhands.events.serialization.observation import observation_from_dict
from openhands.events.serialization.utils import remove_fields
from openhands.events.tool import ToolCallMetadata
from openhands.llm.metrics import Cost, Metrics, ResponseLatency, TokenUsage
# TODO: move `content` into `extras`
TOP_KEYS = [
'id',
'timestamp',
'source',
'message',
'cause',
'action',
'observation',
'tool_call_metadata',
'llm_metrics',
]
UNDERSCORE_KEYS = [
'id',
'timestamp',
'source',
'cause',
'tool_call_metadata',
'llm_metrics',
]
DELETE_FROM_TRAJECTORY_EXTRAS = {
'dom_object',
'axtree_object',
'active_page_index',
'last_browser_action',
'last_browser_action_error',
'focused_element_bid',
'extra_element_properties',
}
DELETE_FROM_TRAJECTORY_EXTRAS_AND_SCREENSHOTS = DELETE_FROM_TRAJECTORY_EXTRAS | {
'screenshot',
'set_of_marks',
}
def event_from_dict(data: dict[str, Any]) -> 'Event':
evt: Event
if 'action' in data:
evt = action_from_dict(data)
elif 'observation' in data:
evt = observation_from_dict(data)
else:
raise ValueError(f'Unknown event type: {data}')
for key in UNDERSCORE_KEYS:
if key in data:
value = data[key]
if key == 'timestamp' and isinstance(value, datetime):
value = value.isoformat()
if key == 'source':
value = EventSource(value)
if key == 'tool_call_metadata':
value = ToolCallMetadata(**value)
if key == 'llm_metrics':
metrics = Metrics()
if isinstance(value, dict):
metrics.accumulated_cost = value.get('accumulated_cost', 0.0)
# Set max_budget_per_task if available
metrics.max_budget_per_task = value.get('max_budget_per_task')
for cost in value.get('costs', []):
metrics._costs.append(Cost(**cost))
metrics.response_latencies = [
ResponseLatency(**latency)
for latency in value.get('response_latencies', [])
]
metrics.token_usages = [
TokenUsage(**usage) for usage in value.get('token_usages', [])
]
# Set accumulated token usage if available
if 'accumulated_token_usage' in value:
metrics._accumulated_token_usage = TokenUsage(
**value.get('accumulated_token_usage', {})
)
value = metrics
setattr(evt, '_' + key, value)
return evt
def _convert_pydantic_to_dict(obj: BaseModel | dict) -> dict:
if isinstance(obj, BaseModel):
return obj.model_dump()
return obj
def event_to_dict(event: 'Event') -> dict:
props = asdict(event)
d = {}
for key in TOP_KEYS:
if hasattr(event, key) and getattr(event, key) is not None:
d[key] = getattr(event, key)
elif hasattr(event, f'_{key}') and getattr(event, f'_{key}') is not None:
d[key] = getattr(event, f'_{key}')
if key == 'id' and d.get('id') == -1:
d.pop('id', None)
if key == 'timestamp' and 'timestamp' in d:
if isinstance(d['timestamp'], datetime):
d['timestamp'] = d['timestamp'].isoformat()
if key == 'source' and 'source' in d:
d['source'] = d['source'].value
if key == 'recall_type' and 'recall_type' in d:
d['recall_type'] = d['recall_type'].value
if key == 'tool_call_metadata' and 'tool_call_metadata' in d:
d['tool_call_metadata'] = d['tool_call_metadata'].model_dump()
if key == 'llm_metrics' and 'llm_metrics' in d:
d['llm_metrics'] = d['llm_metrics'].get()
props.pop(key, None)
if 'security_risk' in props and props['security_risk'] is None:
props.pop('security_risk')
# Remove task_completed from serialization when it's None (backward compatibility)
if 'task_completed' in props and props['task_completed'] is None:
props.pop('task_completed')
if 'action' in d:
# Handle security_risk for actions - include it in args
if 'security_risk' in props:
props['security_risk'] = props['security_risk'].value
d['args'] = props
if event.timeout is not None:
d['timeout'] = event.timeout
elif 'observation' in d:
d['content'] = props.pop('content', '')
# props is a dict whose values can include a complex object like an instance of a BaseModel subclass
# such as CmdOutputMetadata
# we serialize it along with the rest
# we also handle the Enum conversion for RecallObservation
d['extras'] = {
k: (v.value if isinstance(v, Enum) else _convert_pydantic_to_dict(v))
for k, v in props.items()
}
# Include success field for CmdOutputObservation
if hasattr(event, 'success'):
d['success'] = event.success
else:
raise ValueError(f'Event must be either action or observation. has: {event}')
return d
def event_to_trajectory(event: 'Event', include_screenshots: bool = False) -> dict:
d = event_to_dict(event)
if 'extras' in d:
remove_fields(
d['extras'],
DELETE_FROM_TRAJECTORY_EXTRAS
if include_screenshots
else DELETE_FROM_TRAJECTORY_EXTRAS_AND_SCREENSHOTS,
)
return d
def truncate_content(content: str, max_chars: int | None = None) -> str:
"""Truncate the middle of the observation content if it is too long."""
if max_chars is None or len(content) <= max_chars or max_chars < 0:
return content
# truncate the middle and include a message to the LLM about it
half = max_chars // 2
return (
content[:half]
+ '\n[... Observation truncated due to length ...]\n'
+ content[-half:]
)