mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-21 14:12:02 +00:00
第三波前两批:六个「看着成功其实没干活」的平台漏洞堵上,记录页能回放,管理页与画布收尾
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。
This commit is contained in:
@@ -10,7 +10,7 @@ from pydantic import BaseModel, Field
|
||||
from sqlalchemy import select
|
||||
from sqlalchemy.ext.asyncio import AsyncSession
|
||||
|
||||
from app.api.errors import explain, raw
|
||||
from app.core.errors import explain, raw
|
||||
from app.core.config import settings
|
||||
from app.db.base import get_session
|
||||
from app.db.models import Document, MemoryItem, Skill
|
||||
@@ -433,20 +433,129 @@ async def probe_embedding(base_url: str) -> dict[str, Any]:
|
||||
return {"base_url": base_url, "models": models}
|
||||
|
||||
|
||||
class _ReindexJob:
|
||||
"""一次重建索引。进度按批更新:kb / store 每提交一批回调一次,不插值。"""
|
||||
|
||||
def __init__(self, collection: str | None, chunks: int, memories: int) -> None:
|
||||
from app.db.base import new_id, utcnow
|
||||
|
||||
self.id = new_id()
|
||||
self.collection = collection
|
||||
self.state = "running" # running | done | failed
|
||||
# chunks → index(倒排重建,没有分批进度)→ memories → done
|
||||
self.phase = "chunks"
|
||||
self.chunks = {"total": chunks, "done": 0}
|
||||
self.memories = {"total": memories, "done": 0}
|
||||
self.started_at = utcnow()
|
||||
self.finished_at: Any = None
|
||||
self.error: str | None = None
|
||||
self.hint: str | None = None
|
||||
self.result: dict[str, Any] | None = None
|
||||
|
||||
def chunk_step(self, done: int, total: int) -> None:
|
||||
self.chunks = {"total": total, "done": done}
|
||||
if done >= total:
|
||||
self.phase = "index"
|
||||
|
||||
def memory_step(self, done: int, total: int) -> None:
|
||||
self.memories = {"total": total, "done": done}
|
||||
|
||||
def view(self) -> dict[str, Any]:
|
||||
return {
|
||||
"id": self.id, "state": self.state, "phase": self.phase,
|
||||
"collection": self.collection,
|
||||
"total": self.chunks["total"] + self.memories["total"],
|
||||
"done": self.chunks["done"] + self.memories["done"],
|
||||
"chunks": dict(self.chunks), "memories": dict(self.memories),
|
||||
"started_at": self.started_at.isoformat(),
|
||||
"finished_at": self.finished_at.isoformat() if self.finished_at else None,
|
||||
"error": self.error, "hint": self.hint, "result": self.result,
|
||||
}
|
||||
|
||||
|
||||
#: 最近一次重建。只在进程里记:重建是幂等的,服务重启丢了进度,再点一次就是
|
||||
_REINDEX: _ReindexJob | None = None
|
||||
#: 后台任务要有人拿着引用,否则跑到一半可能被回收
|
||||
_REINDEX_TASK: Any = None
|
||||
|
||||
|
||||
def reindex_job() -> _ReindexJob | None:
|
||||
return _REINDEX
|
||||
|
||||
|
||||
async def _run_reindex(job: _ReindexJob) -> None:
|
||||
from app.db.base import SessionLocal, utcnow
|
||||
|
||||
try:
|
||||
async with SessionLocal() as session:
|
||||
out = await kb.reindex(session, job.collection, on_progress=job.chunk_step)
|
||||
job.phase = "memories"
|
||||
# 记忆不按 collection 分,重建就是全量
|
||||
out["memories_reindexed"] = await store.reindex(session, on_progress=job.memory_step)
|
||||
job.result, job.state, job.phase = out, "done", "done"
|
||||
except Exception as e: # noqa: BLE001 - 失败要落到任务状态上,不能只进日志
|
||||
logger.warning("重建索引失败:%s", raw(e))
|
||||
reason, hint = explain(e)
|
||||
job.state, job.error, job.hint = "failed", f"重建索引没做完:{reason}", hint or None
|
||||
finally:
|
||||
job.finished_at = utcnow()
|
||||
|
||||
|
||||
@kb_router.post("/reindex")
|
||||
async def reindex(
|
||||
collection: str | None = None, session: AsyncSession = Depends(get_session)
|
||||
) -> dict[str, Any]:
|
||||
collection: str | None = None,
|
||||
background: bool = Query(default=False, description="后台跑,立刻返回任务;进度用 GET /kb/reindex 查"),
|
||||
session: AsyncSession = Depends(get_session),
|
||||
):
|
||||
"""用当前 embedder 重算向量。换了模型之后唯一的恢复手段。
|
||||
|
||||
知识库和长期记忆一起重建:它们共用一个 embedder,换模型时一起失效。
|
||||
分成两个按钮的结果是点了一个、以为好了,另一个还在悄悄退回关键词——
|
||||
真踩过,记忆那边连个提示都没有。
|
||||
|
||||
几千段配远端 embedding 要好几分钟,压在请求里界面只能干转圈。background=true
|
||||
时立刻返回 202 和任务,按批更新进度;不带它仍然等做完再返回,老调用方不受影响。
|
||||
同一时间只跑一个:两次重建交错写同一批向量,谁最后写赢都说不清。
|
||||
"""
|
||||
out = await kb.reindex(session, collection)
|
||||
# 记忆不按 collection 分,重建就是全量
|
||||
out["memories_reindexed"] = await store.reindex(session)
|
||||
return out
|
||||
global _REINDEX, _REINDEX_TASK
|
||||
import asyncio
|
||||
|
||||
from fastapi.responses import JSONResponse
|
||||
from sqlalchemy import func
|
||||
|
||||
from app.db.models import Chunk
|
||||
|
||||
count = select(func.count(Chunk.id))
|
||||
if collection:
|
||||
count = count.where(Chunk.collection == collection)
|
||||
chunks = int((await session.execute(count)).scalar_one() or 0)
|
||||
memories = int((await session.execute(select(func.count(MemoryItem.id)))).scalar_one() or 0)
|
||||
# 先数完再查「有没有在跑」,查和占位之间不能有 await:以前先查后数,两个同时
|
||||
# 到的请求都在计数那里让出去,回来都以为没在跑,于是两个都起了
|
||||
if _REINDEX is not None and _REINDEX.state == "running":
|
||||
running = _REINDEX.view()
|
||||
raise HTTPException(
|
||||
409, f"已经在重建索引了({_REINDEX.collection or '全部集合'},"
|
||||
f"已完成 {running['done']}/{running['total']}),等它做完再点",
|
||||
)
|
||||
job = _REINDEX = _ReindexJob(collection, chunks, memories)
|
||||
if background:
|
||||
_REINDEX_TASK = asyncio.create_task(_run_reindex(job))
|
||||
return JSONResponse(job.view(), status_code=202)
|
||||
|
||||
await _run_reindex(job)
|
||||
if job.state == "failed":
|
||||
raise HTTPException(502, job.error + (f"。{job.hint}" if job.hint else ""))
|
||||
return job.result
|
||||
|
||||
|
||||
@kb_router.get("/reindex")
|
||||
async def reindex_status() -> dict[str, Any]:
|
||||
"""最近一次重建的进度:{state, phase, total, done, chunks, memories, result, error…}。
|
||||
|
||||
还没重建过返回 {state: "idle"}。
|
||||
"""
|
||||
return _REINDEX.view() if _REINDEX is not None else {"state": "idle"}
|
||||
|
||||
|
||||
#: 单段返回多少字。切块本来就在千字量级,详情页是给人扫一眼"切得对不对"的,
|
||||
|
||||
Reference in New Issue
Block a user