mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-19 09:22:01 +00:00
587bf7e27c
起因是一次探索运行:Copilot 改节点时把工具绑定整份抹掉,agent 没有工具只能写「假设调用」, 校验节点的自动修复编出 total_count=0 让 schema 通过,协作成员把工具调用标记当文字输出了四轮, 轮数用完后节点把原话当成功结果交出去——运行标成已完成,数据库一次都没查。 平台可信度(先写失败测试) - Copilot 修改节点改为合并:没写的字段保留,null 才删除;协作成员按名字合并。前后端同一套规则, 工具集合变小时自查报警、改图回执列出变化 - 新校验规则:提示词点名了某个工具而节点没有绑定它,判错误,Copilot 自查会交回去修 - 校验修复只许改格式:冒出原文没有的值就判修复无效;修复这次模型调用计入用量 - 识别模型把工具调用写成文字的情况(DSML、<tool_call> 等标记):纠正一次,仍然如此就判失败并说明原因 - 协作团队用完轮数不再按成功交付:默认失败,可选降档,点出一次都没派到的成员 - 审批恢复后,同一节点里已经执行过的工具调用复用结果,不再重复执行 后端 - 查询时限交给数据库执行(Oracle call_timeout、PG statement_timeout、MySQL max_execution_time、 SQLite progress handler),可按数据源配置;超时报错与实际耗时一致 - 等审批的运行可以真正取消;会话列表带最后一轮状态;测连接结果持久化;探查结构可只看不存; 自定义工具可草稿试跑、写坏的参数定义保存时拒绝;MCP 服务改动后作废旧工具缓存; 检索命中带贡献分解;重建索引在后台跑并报进度;报错统一收进 core/errors 前端 - 公共层:管理页共用件、失败说明(lib/explain)、健康状态进公共库;目录按表各自加载并有超时, 一张表卡住不再拖垮整站;上传显示真实字节进度;离线横幅只播报不变的句子 - 画布:卡片只说游标所在的那一次执行;等审批时可放弃运行;团队收尾判定与未完成说清楚; 回放用量取那一刻的值 - 编排页:Copilot 合并语义与工具变化回执;「用完轮数时」选项;正式运行期间锁住所有改图入口 - 回答流 / 问数据:新失败类型说人话并给下一步;画布与右栏互相定位;数据源范围、回收站、重新核对 - 记录页:「航迹」页签复用画布时间轴,可拖动回放、列出关键时刻;列表行带执行 / 等人 / 失败缩略条 - 外壳:统一的离开前确认(data router),导航、⌘K、快捷键、后退都会先问 验证(沙箱):tsc 干净,pytest 838 passed(两遍),14 个检查脚本中 13 个全绿; check-ui 余 1 项旧断言(按旧日期格式找运行行),3c 更新。
209 lines
8.8 KiB
Python
209 lines
8.8 KiB
Python
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import json
|
||
import logging
|
||
import time
|
||
from typing import Any
|
||
|
||
from langchain_core.tools import BaseTool
|
||
from sqlalchemy import select
|
||
|
||
from app.db.base import SessionLocal
|
||
from app.db.models import McpServer
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
def _server_config(row: McpServer) -> dict[str, Any]:
|
||
if row.transport == "http":
|
||
return {"transport": "streamable_http", "url": row.url or ""}
|
||
return {
|
||
"transport": "stdio",
|
||
"command": row.command or "",
|
||
"args": list(row.args or []),
|
||
"env": dict(row.env or {}),
|
||
}
|
||
|
||
|
||
def _config_key(row: McpServer) -> str:
|
||
"""一个服务的连接配置,缓存据此判断手上的工具是不是按现在这份配置建的。"""
|
||
return json.dumps(_server_config(row), sort_keys=True, ensure_ascii=False)
|
||
|
||
|
||
class McpManager:
|
||
"""管理外部 MCP server 的连接与工具缓存。
|
||
|
||
工具在图里用 `mcp:<server>/<tool>` 引用。连接是懒建立的,
|
||
并且按 server 缓存,避免每个节点都去重新拉起一次 stdio 子进程。
|
||
|
||
缓存按服务逐个补、逐个作废。以前只在「一个都没加载过」时整批建一次:改了
|
||
启动命令,运行照样用旧命令建出来的工具;删掉的服务照样调得到;缓存建好之后
|
||
新加的服务,工具列表里一直没有它,除非有人去点刷新。现在每次取工具都先对一遍
|
||
库里启用着的服务:缺的连上,配置变了的重连,删了、停用了的不再交出去。
|
||
"""
|
||
|
||
def __init__(self) -> None:
|
||
self._tools: dict[str, list[BaseTool]] = {} # server name -> tools
|
||
#: 每个服务的工具是按哪份连接配置建的(_config_key)
|
||
self._built_from: dict[str, str] = {}
|
||
self._lock = asyncio.Lock()
|
||
|
||
def invalidate(self, *names: str) -> None:
|
||
"""服务改了、删了:丢掉按旧配置建的工具,下次用到时再连。"""
|
||
for name in names:
|
||
self._tools.pop(name, None)
|
||
self._built_from.pop(name, None)
|
||
|
||
def _fresh(self, row: McpServer) -> bool:
|
||
return row.name in self._tools and self._built_from.get(row.name) == _config_key(row)
|
||
|
||
async def _load_servers(self, names: list[str] | None = None) -> list[McpServer]:
|
||
async with SessionLocal() as session:
|
||
query = select(McpServer).where(McpServer.enabled.is_(True))
|
||
if names is not None:
|
||
query = query.where(McpServer.name.in_(names))
|
||
rows = await session.execute(query)
|
||
return list(rows.scalars())
|
||
|
||
async def _connect(self, rows: list[McpServer]) -> None:
|
||
from langchain_mcp_adapters.client import MultiServerMCPClient
|
||
|
||
for row in rows:
|
||
key = _config_key(row)
|
||
started = time.perf_counter()
|
||
try:
|
||
client = MultiServerMCPClient({row.name: _server_config(row)})
|
||
tools = list(await asyncio.wait_for(client.get_tools(), timeout=30))
|
||
status, error = "ok", None
|
||
except Exception as e: # noqa: BLE001 - 一个 server 挂了不影响其他的
|
||
logger.warning("MCP server %s 连接失败: %s", row.name, e)
|
||
tools, status, error = [], "error", _reason(e, row.command)
|
||
# 连不上也缓存成空:不然每次取工具都要再等一轮 30 秒超时。改配置或刷新时重连
|
||
self._tools[row.name] = tools
|
||
self._built_from[row.name] = key
|
||
await self._mark(row.id, key, status, error, [t.name for t in tools], started)
|
||
|
||
async def _mark(
|
||
self, server_id: str, key: str, status: str, error: str | None, tool_names: list[str],
|
||
started: float,
|
||
) -> None:
|
||
from app.api.health import record
|
||
|
||
async with SessionLocal() as session:
|
||
row = await session.get(McpServer, server_id)
|
||
# 连到一半服务被改了:这次连的是旧配置,它的结论不能盖掉改完之后的「未检查」
|
||
if row and _config_key(row) == key:
|
||
row.status = status
|
||
row.last_error = error
|
||
row.tools_cache = tool_names
|
||
# 加载工具本身就是一次连接:记下来,工具页不用再单独测一次才知道
|
||
row.last_check = record(
|
||
status == "ok", int((time.perf_counter() - started) * 1000), error)
|
||
await session.commit()
|
||
|
||
async def _ensure(self, rows: list[McpServer]) -> None:
|
||
"""rows 里还没连过、或者是按旧配置连的,连一次补上。"""
|
||
if all(self._fresh(r) for r in rows):
|
||
return
|
||
async with self._lock:
|
||
# 等锁的时候别人可能已经连好了
|
||
stale = [r for r in rows if not self._fresh(r)]
|
||
if stale:
|
||
await self._connect(stale)
|
||
|
||
async def refresh(self) -> dict[str, list[str]]:
|
||
"""全部重连一遍(服务端的工具变了,配置没变时用)。"""
|
||
async with self._lock:
|
||
rows = await self._load_servers()
|
||
self.invalidate(*(set(self._tools) - {r.name for r in rows}))
|
||
await self._connect(rows)
|
||
return {r.name: [t.name for t in self._tools.get(r.name, [])] for r in rows}
|
||
|
||
async def ensure_loaded(self) -> None:
|
||
await self._ensure(await self._load_servers())
|
||
|
||
async def list_tools(self) -> list[dict[str, Any]]:
|
||
rows = await self._load_servers()
|
||
await self._ensure(rows)
|
||
out: list[dict[str, Any]] = []
|
||
for row in rows:
|
||
server = row.name
|
||
for tool in self._tools.get(server, []):
|
||
out.append(
|
||
{
|
||
"id": f"mcp:{server}/{tool.name}",
|
||
"name": tool.name,
|
||
"server": server,
|
||
"description": tool.description or "",
|
||
"category": f"MCP · {server}",
|
||
"schema": tool.args_schema.model_json_schema()
|
||
if getattr(tool, "args_schema", None)
|
||
else {},
|
||
}
|
||
)
|
||
return out
|
||
|
||
async def get_tools(self, refs: list[str]) -> list[BaseTool]:
|
||
"""按 `mcp:<server>/<tool>` 取工具;`mcp:<server>/*` 表示该 server 全部工具。
|
||
|
||
只交出库里还在、并且启用着的服务的工具:删了、停用了的,缓存里就算还有也不给。
|
||
"""
|
||
parsed = []
|
||
for ref in refs:
|
||
body = ref[4:] if ref.startswith("mcp:") else ref
|
||
server, _, tool_name = body.partition("/")
|
||
parsed.append((server, tool_name))
|
||
rows = await self._load_servers(sorted({server for server, _ in parsed}))
|
||
await self._ensure(rows)
|
||
live = {r.name for r in rows}
|
||
wanted: list[BaseTool] = []
|
||
for server, tool_name in parsed:
|
||
tools = self._tools.get(server, []) if server in live else []
|
||
if tool_name in ("", "*"):
|
||
wanted.extend(tools)
|
||
continue
|
||
for tool in tools:
|
||
if tool.name == tool_name:
|
||
wanted.append(tool)
|
||
break
|
||
return wanted
|
||
|
||
async def probe(self, row: McpServer) -> dict[str, Any]:
|
||
"""连一次看看,用于设置页的"测试连接"。"""
|
||
from langchain_mcp_adapters.client import MultiServerMCPClient
|
||
|
||
try:
|
||
client = MultiServerMCPClient({row.name: _server_config(row)})
|
||
tools = await asyncio.wait_for(client.get_tools(), timeout=30)
|
||
return {
|
||
"ok": True,
|
||
"tools": [
|
||
{"name": t.name, "description": t.description or ""} for t in tools
|
||
],
|
||
}
|
||
except Exception as e: # noqa: BLE001
|
||
from app.core.errors import raw
|
||
|
||
return {"ok": False, "error": _reason(e, row.command), "detail": raw(e)}
|
||
|
||
|
||
def _reason(e: BaseException, command: str | None = None) -> str:
|
||
"""连不上 MCP 服务的原因,写成工具页上看得懂的话。
|
||
|
||
最常见的两种:命令不存在(没装 npx / uvx)和超时(进程起来了但不讲 MCP)。
|
||
"""
|
||
from app.core.errors import explain
|
||
|
||
if isinstance(e, FileNotFoundError):
|
||
# stdio 客户端抛的这个异常常常不带文件名,只剩「找不到命令 」半句:补上配置的命令
|
||
missing = e.filename or command or ""
|
||
return f"启动不了:找不到命令 {missing}。确认它已安装、并且在后端的 PATH 里"
|
||
if isinstance(e, asyncio.TimeoutError):
|
||
return "连上了但 30 秒内没有应答:确认这个命令启动的是 MCP 服务,参数有没有写错"
|
||
reason, hint = explain(e)
|
||
return f"{reason}。{hint}" if hint else reason
|
||
|
||
|
||
mcp_manager = McpManager()
|