frontier-agent-demo / workflows /_shared /research /observers /finalize_answer_observer.py
christlurker's picture
FrontierAgent react demo
9d0fd45
Raw
History Blame Contribute Delete
1.96 kB
"""Stops the agent loop on a terminal "I am done" tool call."""
from __future__ import annotations
import logging
from collections.abc import Iterable
from frontier_agent.core.loop_types import (
BaseObserver,
Intervention,
ToolResult,
TurnContext,
)
logger = logging.getLogger(__name__)
FINALIZE_TOOL_NAME = "finalize_answer"
class FinalizeAnswerObserver(BaseObserver):
critical = True
def __init__(
self,
tool_names: Iterable[str] = (FINALIZE_TOOL_NAME,),
content_arg_name: str = "content",
) -> None:
self._tool_names = frozenset(tool_names)
self._content_arg_name = content_arg_name
self._triggered = False
async def on_tool_result(self, ctx: TurnContext, result: ToolResult) -> None:
if result.name not in self._tool_names:
return
if result.is_error:
# Let the model see the error and try again; don't latch.
return
args = result.args if isinstance(result.args, dict) else {}
content = (args.get(self._content_arg_name) or "").strip()
if not content:
return
try:
confidence = max(0.0, min(1.0, float(args.get("confidence", 0.7))))
except (TypeError, ValueError):
confidence = 0.7
self._triggered = True
if isinstance(ctx.metadata, dict):
ctx.metadata["final_answer"] = content
ctx.metadata["final_answer_confidence"] = confidence
logger.info(
"FinalizeAnswerObserver latched via %s: turn=%d, len=%d, conf=%.2f",
result.name, ctx.turn, len(content), confidence,
)
async def on_turn_end(self, ctx: TurnContext) -> Intervention | None:
if not self._triggered:
return None
# One-shot: subsequent on_turn_end calls shouldn't re-emit.
self._triggered = False
return Intervention(stop_reason="final_answer")