aeolusyansheng's picture
Improve Japanese comments and change default language to English
621bde6
Raw History Blame Contribute Delete
5 kB
"""
Nested Agent - LangGraphグラフ定義
Orchestrator -> [Coder -> Tester, Documenter] -> Finalizerのネスト呼び出しを実装
"""
import json
from typing import TypedDict, Optional, List, Dict, Annotated
from langgraph.graph import StateGraph, END, START
from langgraph.types import Send
from nested_agent.agents.orchestrator import orchestrator_node, finalizer_node
from nested_agent.agents.coder import coder_node
from nested_agent.agents.tester import tester_node
from nested_agent.agents.documenter import documenter_node
# Reducer定義
def merge_dicts(left: Optional[dict], right: Optional[dict]) -> dict:
if left is None: left = {}
if right is None: right = {}
return {**left, **right}
# State定義
class NestedState(TypedDict):
requirement: str # 元の要件
plan: Optional[Dict[str, bool]] # Orchestratorの計画結果
plan_reason: Optional[str] # 計画理由
coder_output: Optional[str] # Coder出力
tester_output: Optional[str] # Tester出力
documenter_output: Optional[str] # Documenter出力
final_output: Optional[str] # 最終統合結果
model_used_by: Annotated[dict, merge_dicts]
# ルーティング関数
def route_after_plan(state: NestedState):
"""Orchestrator計画後、Coderフェーズに入るかを決定"""
plan = state.get("plan", {})
if plan.get("coder"):
return "coder_agent"
return "executor_dispatcher"
def executor_dispatcher(state: NestedState) -> List[Send]:
"""
コア調整ロジック:計画に基づいてTesterとDocumenterを並列呼び出し
"""
plan = state.get("plan", {})
sends = []
if plan.get("tester"):
sends.append(Send("tester_agent", state))
if plan.get("documenter"):
sends.append(Send("documenter_agent", state))
return sends if sends else [Send("finalizer_agent", state)]
# StateGraphを構築
def build_graph() -> StateGraph:
workflow = StateGraph(NestedState)
# ノードを追加
workflow.add_node("orchestrator_agent", orchestrator_node)
workflow.add_node("coder_agent", coder_node)
workflow.add_node("tester_agent", tester_node)
workflow.add_node("documenter_agent", documenter_node)
workflow.add_node("finalizer_agent", finalizer_node)
# エントリーポイントを設定
workflow.set_entry_point("orchestrator_agent")
# 1. 計画後のルーティング
workflow.add_conditional_edges(
"orchestrator_agent",
route_after_plan,
{
"coder_agent": "coder_agent",
"executor_dispatcher": "finalizer_agent" # 何も不要な場合は直接最終化へ(論理的にはあり得ないが)
}
)
# 2. Coder完了後に並列タスクを分配(またはorchestratorからcoderをスキップして直接分配)
# 注意:LangGraphはadd_node外でdispatcher関数をSend対象として定義することを直接サポートしない
# executor_dispatcherを実行するための中継ノードが必要
def dispatcher_node(state: NestedState):
return state # 中継のみ
workflow.add_node("dispatcher", dispatcher_node)
workflow.add_edge("coder_agent", "dispatcher")
# orchestratorからcoderが不要な場合もdispatcherへ
# (上記のroute_after_planロジックの対象をdispatcherに変更)
# dispatcherノードに合わせてルーティングロジックを再定義
workflow.add_conditional_edges(
"orchestrator_agent",
lambda state: "coder_agent" if state.get("plan", {}).get("coder") else "dispatcher",
{
"coder_agent": "coder_agent",
"dispatcher": "dispatcher"
}
)
# 3. Dispatcherからfan-out
workflow.add_conditional_edges(
"dispatcher",
executor_dispatcher,
["tester_agent", "documenter_agent", "finalizer_agent"]
)
# 4. fan-in:並列タスク完了後にfinalizerへ集約
workflow.add_edge("tester_agent", "finalizer_agent")
workflow.add_edge("documenter_agent", "finalizer_agent")
# 5. 終了
workflow.add_edge("finalizer_agent", END)
return workflow.compile()
# ストリーミング実行関数
def stream_nested(requirement: str):
"""
ネストエージェントワークフローをストリーミング実行
"""
graph = build_graph()
initial_state = {
"requirement": requirement,
"plan": None,
"plan_reason": None,
"coder_output": None,
"tester_output": None,
"documenter_output": None,
"final_output": None,
"model_used_by": {},
}
for event in graph.stream(initial_state, stream_mode="updates"):
for node_name, state_update in event.items():
yield node_name, state_update