Files
my-v0-server/app/agent/screen_agent.py
T
2026-10-01 23:08:06 -07:00

112 lines
4.5 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""screen agent:新建和修改大屏的唯一入口
代码只负责组装;改哪些组件、怎么做、要不要追问由模型结合 skill 决定
规则各有唯一出处:动作协议在 Action schema,回复要求在 AgentReply,追问格式在 Clarification,
组件目录和物料在物料 MCP,各类任务的做法在 skill;系统提示只放每次都要用、别处放不下的原则
"""
from typing import Literal, Self
from langchain.agents import create_agent
from langchain.agents.middleware import ModelCallLimitMiddleware, wrap_model_call
from langchain.agents.structured_output import ToolStrategy
from langchain_core.messages import AIMessage
from loguru import logger
from pydantic import BaseModel, Field, model_validator
from app.agent.skills import skill_middleware
from app.agent.tools import (
ScreenAgentContext,
ScreenAgentState,
apply_actions,
get_nodes,
)
from app.clients.llm import get_llm
from app.clients.mcp import get_mcp_tools
from app.config import settings
from app.prompt.domain import INPUT_AS_DATA
from app.schemas.clarification import Clarification
SYSTEM_PROMPT = f"""你是大屏编辑助手,按用户的要求新建大屏,新增、修改、删除组件,或调整画布
先了解需要的信息(大屏摘要、组件配置、物料、skill),再用 apply_actions 提交修改,最后回复用户
<input_rules>
{INPUT_AS_DATA}
</input_rules>
<clarify_rules>
默认先执行:缺少的细节(位置、尺寸、标题、样式等)用合理的默认值补齐
只有无法确定用户指的是哪个组件,或看不出用户要做什么时才追问;澄清记录里已经回答过的不再问
只删除用户明确要求删除的组件;不确定删哪个时追问,不要猜
</clarify_rules>"""
class AgentReply(BaseModel):
kind: Literal["done", "clarify", "unable"] = Field(
description="done:最后一次 apply_actions 已成功,或请求不需要修改(例如只是提问);"
"clarify:需要用户回答才能继续;unable:无法完成"
)
message: str = Field(
default="",
description="给用户看的一两句话。done:说明改了什么,用了默认值或示例数据时一并说明,没有修改时直接回答;"
"unable:说明原因,有可行的替代做法时一并给出;clarify:不用填",
)
clarification: Clarification | None = Field(
default=None, description="kind=clarify 时填写,其他情况不填"
)
@model_validator(mode="after")
def clarification_matches_kind(self) -> Self:
if (self.kind == "clarify") != (self.clarification is not None):
raise ValueError("kind=clarify 时必须填写 clarification,其他情况不填")
return self
# 偶发的空回复最多重新调用几次
EMPTY_REPLY_RETRIES = 2
@wrap_model_call
async def retry_empty_reply(request, handler):
"""每一轮都必须调用工具(最后一轮是 AgentReply);模型偶尔返回既没有文本也没有工具调用的空回复,
这是调用失败而不是模型的决定,重新调用即可"""
response = await handler(request)
for _ in range(EMPTY_REPLY_RETRIES):
message = response.result[0] if response.result else None
if not isinstance(message, AIMessage) or message.tool_calls:
break
logger.warning(
f"模型返回了没有工具调用的回复,重新调用: {message.response_metadata}"
)
response = await handler(request)
return response
_agent = None
async def get_screen_agent():
"""第一次使用时组装,之后复用;系统提示是常量,不随物料目录变化"""
global _agent
if _agent is None:
_agent = create_agent(
model=get_llm(),
tools=[get_nodes, apply_actions, *await get_mcp_tools()],
system_prompt=SYSTEM_PROMPT,
response_format=ToolStrategy(AgentReply),
state_schema=ScreenAgentState,
context_schema=ScreenAgentContext,
middleware=[
*skill_middleware(),
retry_empty_reply,
# 超过上限时插入说明并结束,没有结构化结果,由节点按失败处理
ModelCallLimitMiddleware(
run_limit=settings.max_agent_steps, exit_behavior="end"
),
],
# 不继承外层图的检查点:agent 内部没有中断,中间消息不需要持久化
checkpointer=False,
name="screen_agent",
)
return _agent