mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 10:51:58 +00:00
ThearchyHelios
761117344b
## 每跑一次图就多一台常驻 VM，没有任何自动回收路径 _reap_idle 只在 run() 里调，也就是说没人再跑代码时闲置 VM 永远收不掉， microvm_idle_seconds 形同虚设；而 thread_id 每个 run 都是新的，runner 的 finally 又从不调 cleanup，于是每跑一次带代码节点的图就多一台几百 MB 的常驻 microVM（seatbelt 那边同样积了几十个历史会话目录）。 两处一起补：run 到终态就 cleanup(run_id)（interrupted 除外，人回来还要接着跑）， 外加 lifespan 里起一个后台回收任务定期 _reap_idle——回收本来也不该占着请求 路径的时间。实测连跑三次 live_vms 稳定回到 0。 ## 超时会连人带文件销毁整台 VM 会话契约是"同 session_id 共享文件系统，可以先写文件下一步再读回来"，而 microVM 超时直接 _destroy，工作区连同用户数据一起没了，返回结果里没有任何 提示。另外三个后端超时只 killpg、工作目录照留——换到 microVM 之后同一张图的 行为就变了。现在超时先在 VM 内 pkill 掉跑飞的进程（入口名带唯一后缀，不会 误伤同会话的并发执行），确认 VM 还活着就保住工作区；收拾不动才销毁。 限额变更那条没法避免重建（VM 内存是启动时定死的），但不能不吭声：同一个 run 里两个代码节点填了不同的 memory_mb 就会触发。ExecResult 新增 session_reset， 代码节点在时间线上明确告警，否则下游只会看到一个无从追查的 FileNotFoundError。 ## 新建 scope 写入第一条后永远搜不到 hybrid_rank 的 _norm 在 hi==lo 时把所有分数抹成 0，而候选只有一条时 hi 恒等于 lo，合并分 0.0 会被 min_score 过滤掉。"先写一条、下一轮再 recall"恰恰是 agent 最常见的用法。全都一样分是"都一样好"，不是"都一样差"，改成返回 1.0。 ## 记忆去重误判，写入被静默丢弃且返回不相关的旧记忆 remember 拿 recall 的分数判重，但那是 min-max 排名归一化的结果——排第一的那条 在两路信号上都会被归成 1.0，与它和查询像不像毫无关系。于是 scope 里只要已有 两条记忆，新内容随便撞一个字排到第一就被判成重复丢掉，接口还返回 201 并把那条 不相关的旧记忆当成"刚保存的"交回去。改成直接比未归一化的余弦相似度。 验证：三条互不相关的记忆都能写进去、完全相同的仍被去重、单条 scope 可检索。
504 lines
20 KiB
Python
504 lines
20 KiB
Python
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import contextlib
|
||
import logging
|
||
import time
|
||
from datetime import datetime, timezone
|
||
from typing import Any
|
||
|
||
from langgraph.checkpoint.sqlite.aio import AsyncSqliteSaver
|
||
from langgraph.types import Command
|
||
from sqlalchemy import select, update
|
||
|
||
from app.core.bus import bus
|
||
from app.core.config import settings
|
||
from app.core.events import EventType, RunEventModel
|
||
from app.db.base import SessionLocal
|
||
from app.db.models import Approval, Run, RunEvent, Workflow
|
||
from app.engine.compiler import compile_graph, initial_state
|
||
from app.engine.context import RunContext
|
||
from app.engine.schema import GraphSpec, validate_graph
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
# 流式增量（正文 token、思考 delta）量太大，只做实时推送不落库；
|
||
# 轨迹里正文靠 node.finished 的 preview，思考靠调用结束时的单条 llm.thinking 汇总
|
||
_EPHEMERAL = {EventType.LLM_TOKEN, EventType.LLM_THINKING_DELTA}
|
||
|
||
|
||
class RunManager:
|
||
"""负责一次运行的完整生命周期：启动、串流、中断、恢复、取消。"""
|
||
|
||
def __init__(self) -> None:
|
||
self._tasks: dict[str, asyncio.Task[None]] = {}
|
||
self._checkpointer: AsyncSqliteSaver | None = None
|
||
self._cm: Any = None
|
||
self._semaphore = asyncio.Semaphore(settings.max_concurrent_runs)
|
||
|
||
# ---------------- 生命周期 ----------------
|
||
|
||
async def setup(self) -> None:
|
||
settings.ensure_dirs()
|
||
self._cm = AsyncSqliteSaver.from_conn_string(str(settings.checkpoint_path))
|
||
self._checkpointer = await self._cm.__aenter__()
|
||
# 上次进程被强杀时留下的记录。running 和 queued 要分开处理：
|
||
# running 至少跑过一步、有 checkpoint，可以续；queued 还在排队，
|
||
# 一个 checkpoint 都没写过，"恢复"它只会从 START 用默认值重跑一遍，
|
||
# 还丢掉 run.input——那不是恢复，是伪造一份结果。
|
||
async with SessionLocal() as session:
|
||
await session.execute(
|
||
update(Run)
|
||
.where(Run.status == "running")
|
||
.values(status="interrupted", error="服务重启，运行已挂起，可从断点恢复")
|
||
)
|
||
await session.execute(
|
||
update(Run)
|
||
.where(Run.status == "queued")
|
||
.values(status="failed", error="服务重启时这次运行还在排队，没有可恢复的进度，请重新发起")
|
||
)
|
||
await session.commit()
|
||
|
||
async def shutdown(self) -> None:
|
||
for task in list(self._tasks.values()):
|
||
task.cancel()
|
||
await asyncio.gather(*self._tasks.values(), return_exceptions=True)
|
||
if self._cm is not None:
|
||
with contextlib.suppress(Exception):
|
||
await self._cm.__aexit__(None, None, None)
|
||
|
||
@property
|
||
def checkpointer(self) -> AsyncSqliteSaver:
|
||
if self._checkpointer is None:
|
||
raise RuntimeError("RunManager 还没初始化")
|
||
return self._checkpointer
|
||
|
||
# ---------------- 事件 ----------------
|
||
|
||
async def _emit(
|
||
self,
|
||
run_id: str,
|
||
event_type: EventType | str,
|
||
*,
|
||
node_id: str | None = None,
|
||
data: dict[str, Any] | None = None,
|
||
persist: bool = True,
|
||
) -> None:
|
||
event = RunEventModel(
|
||
run_id=run_id,
|
||
seq=bus.next_seq(run_id),
|
||
type=EventType(str(event_type)),
|
||
node_id=node_id,
|
||
data=data or {},
|
||
)
|
||
await bus.publish(event)
|
||
if persist and event.type not in _EPHEMERAL:
|
||
async with SessionLocal() as session:
|
||
session.add(
|
||
RunEvent(
|
||
run_id=run_id,
|
||
seq=event.seq,
|
||
type=str(event.type),
|
||
node_id=event.node_id,
|
||
ts=event.ts,
|
||
data=event.data,
|
||
)
|
||
)
|
||
await session.execute(
|
||
update(Run).where(Run.id == run_id).values(last_seq=event.seq)
|
||
)
|
||
await session.commit()
|
||
|
||
async def note(
|
||
self, run_id: str, event_type: EventType | str, *, node_id: str | None = None,
|
||
**data: Any,
|
||
) -> None:
|
||
"""给 API 层用的事件入口（比如 formal 启动时记录口径升版处置）。"""
|
||
await self._emit(run_id, event_type, node_id=node_id, data=data)
|
||
|
||
# ---------------- 启动 ----------------
|
||
|
||
async def start(
|
||
self,
|
||
*,
|
||
graph: dict[str, Any],
|
||
input_payload: dict[str, Any],
|
||
workflow_id: str | None = None,
|
||
workflow_name: str = "",
|
||
memory_scope: str = "default",
|
||
collection: str = "default",
|
||
run_class: str = "exploratory",
|
||
version: int | None = None,
|
||
version_hash: str | None = None,
|
||
started_by: str | None = None,
|
||
) -> Run:
|
||
spec = GraphSpec.model_validate(graph)
|
||
report = validate_graph(spec)
|
||
if not report.ok:
|
||
raise ValueError(
|
||
"图校验未通过：" + "；".join(i.message for i in report.issues if i.level == "error")
|
||
)
|
||
|
||
async with SessionLocal() as session:
|
||
run = Run(
|
||
workflow_id=workflow_id,
|
||
workflow_name=workflow_name,
|
||
status="queued",
|
||
graph=graph,
|
||
input=input_payload,
|
||
run_class=run_class,
|
||
version=version,
|
||
version_hash=version_hash,
|
||
started_by=started_by,
|
||
)
|
||
run.thread_id = run.id # 一个 run 一条 checkpoint 线程
|
||
session.add(run)
|
||
await session.commit()
|
||
await session.refresh(run)
|
||
|
||
self._tasks[run.id] = asyncio.create_task(
|
||
self._drive(
|
||
run.id,
|
||
spec,
|
||
initial_state(input_payload),
|
||
workflow_id=workflow_id,
|
||
memory_scope=memory_scope,
|
||
collection=collection,
|
||
)
|
||
)
|
||
return run
|
||
|
||
async def resume(
|
||
self, run_id: str, response: Any, *, approval_id: str | None = None,
|
||
actor: str | None = None,
|
||
) -> Run:
|
||
"""从人工介入的断点继续。
|
||
|
||
两件事以前是错的：
|
||
|
||
一是不看有没有 checkpoint 就发 Command。重启扫描会把 queued（从没跑过、
|
||
没有任何 checkpoint）也标成 interrupted，对这种 run 发 Command，LangGraph
|
||
会拿一个空 state 从 START 重跑整张图——run.input 完全没被注入，最后还落成
|
||
succeeded。跑出来的是一份用默认值算的假结果。
|
||
|
||
二是同一 superstep 里有多个 interrupt 时，只发标量 Command 会让 LangGraph
|
||
直接抛 RuntimeError 打爆整个 run，而所有 pending 审批还都被写成同一个回复。
|
||
现在按 interrupt_id 逐条收集，凑齐了才继续，没凑齐就继续等下一个人回复。
|
||
"""
|
||
async with SessionLocal() as session:
|
||
run = await session.get(Run, run_id)
|
||
if not run:
|
||
raise KeyError(f"找不到运行 {run_id}")
|
||
if run.status not in ("interrupted",):
|
||
raise ValueError(f"当前状态是 {run.status}，无法恢复")
|
||
spec = GraphSpec.model_validate(run.graph)
|
||
workflow_id = run.workflow_id
|
||
|
||
# 先问引擎：这条线程到底停在哪、还等着哪些 interrupt
|
||
run_ctx = RunContext(run_id=run_id, thread_id=run_id, spec=spec, workflow_id=workflow_id)
|
||
app = compile_graph(spec, run_ctx).compile(checkpointer=self.checkpointer)
|
||
config = {"configurable": {"thread_id": run_id}}
|
||
snapshot = await app.aget_state(config)
|
||
waiting = list(getattr(snapshot, "interrupts", None) or [])
|
||
|
||
if not waiting and not getattr(snapshot, "next", None):
|
||
# 没有断点可续。多半是重启时被标成 interrupted 的 queued 运行。
|
||
raise ValueError(
|
||
"这次运行没有可恢复的断点（很可能重启前还没真正开始跑）。"
|
||
"请重新发起一次运行，而不是恢复——继续下去只会用默认值跑出一份假结果。"
|
||
)
|
||
|
||
answers: dict[str, Any] = {}
|
||
async with SessionLocal() as session:
|
||
pending = list((
|
||
await session.execute(
|
||
select(Approval).where(
|
||
Approval.run_id == run_id, Approval.status == "pending"
|
||
)
|
||
)
|
||
).scalars())
|
||
|
||
# 定位这次回复的是哪一条：显式指定优先，否则只在唯一 pending 时才敢猜
|
||
target = None
|
||
if approval_id:
|
||
target = next((a for a in pending if a.id == approval_id), None)
|
||
if target is None:
|
||
raise ValueError(f"审批 {approval_id} 不在待处理列表里")
|
||
elif len(pending) == 1:
|
||
target = pending[0]
|
||
elif len(pending) > 1:
|
||
raise ValueError(
|
||
f"这次运行有 {len(pending)} 条待处理的人工介入，"
|
||
"恢复时必须指定 approval_id 说明回复的是哪一条"
|
||
)
|
||
|
||
now = datetime.now(timezone.utc)
|
||
if target is not None:
|
||
target.status = "answered" # 已回复，但还没提交给引擎
|
||
target.response = (
|
||
_safe(response) if isinstance(response, dict) else {"value": _safe(response)}
|
||
)
|
||
target.resolved_at = now
|
||
if actor:
|
||
target.resolved_by = actor
|
||
await session.commit()
|
||
|
||
# 凑齐没有？引擎还在等的每一个 interrupt 都得有答案
|
||
answered = list((
|
||
await session.execute(
|
||
select(Approval).where(
|
||
Approval.run_id == run_id,
|
||
Approval.status.in_(["answered", "pending"]),
|
||
)
|
||
)
|
||
).scalars())
|
||
by_iid = {a.interrupt_id: a for a in answered if a.interrupt_id}
|
||
missing = []
|
||
for item in waiting:
|
||
iid = getattr(item, "id", None)
|
||
rec = by_iid.get(iid)
|
||
if rec is None or rec.status != "answered":
|
||
missing.append(iid)
|
||
continue
|
||
payload = rec.response or {}
|
||
answers[iid] = payload.get("value", payload) if isinstance(payload, dict) else payload
|
||
|
||
if missing:
|
||
# 还差人没回。保持 interrupted，把已回复的那条留在 answered 上等齐
|
||
await session.commit()
|
||
run = await session.get(Run, run_id)
|
||
return run
|
||
|
||
# 齐了：正式落 resolved，然后驱动引擎
|
||
for rec in answered:
|
||
if rec.status == "answered":
|
||
rec.status = "resolved"
|
||
run = await session.get(Run, run_id)
|
||
bus.set_seq(run_id, run.last_seq)
|
||
run.status = "running"
|
||
run.error = None
|
||
await session.commit()
|
||
await session.refresh(run)
|
||
|
||
# 单个 interrupt 用标量（LangGraph 两种都认），多个必须用 {interrupt_id: value} 映射
|
||
command = (
|
||
Command(resume=response)
|
||
if len(waiting) <= 1
|
||
else Command(resume=answers)
|
||
)
|
||
await self._emit(run_id, EventType.RUN_RESUMED, data={"response": _safe(response)})
|
||
self._tasks[run_id] = asyncio.create_task(
|
||
self._drive(run_id, spec, command, workflow_id=workflow_id)
|
||
)
|
||
return run
|
||
|
||
async def cancel(self, run_id: str) -> bool:
|
||
task = self._tasks.get(run_id)
|
||
if task and not task.done():
|
||
task.cancel()
|
||
return True
|
||
return False
|
||
|
||
def is_active(self, run_id: str) -> bool:
|
||
task = self._tasks.get(run_id)
|
||
return bool(task and not task.done())
|
||
|
||
# ---------------- 执行 ----------------
|
||
|
||
async def _drive(
|
||
self,
|
||
run_id: str,
|
||
spec: GraphSpec,
|
||
payload: Any,
|
||
*,
|
||
workflow_id: str | None = None,
|
||
memory_scope: str = "default",
|
||
collection: str = "default",
|
||
) -> None:
|
||
started = time.perf_counter()
|
||
status = "succeeded"
|
||
error: str | None = None
|
||
final_state: dict[str, Any] = {}
|
||
|
||
async with self._semaphore:
|
||
try:
|
||
async with SessionLocal() as session:
|
||
await session.execute(
|
||
update(Run)
|
||
.where(Run.id == run_id)
|
||
.values(status="running", started_at=datetime.now(timezone.utc))
|
||
)
|
||
await session.commit()
|
||
await self._emit(
|
||
run_id,
|
||
EventType.RUN_STARTED,
|
||
data={"nodes": len(spec.nodes), "resumed": isinstance(payload, Command)},
|
||
)
|
||
|
||
run_ctx = RunContext(
|
||
run_id=run_id,
|
||
thread_id=run_id,
|
||
spec=spec,
|
||
workflow_id=workflow_id,
|
||
memory_scope=memory_scope,
|
||
collection=collection,
|
||
)
|
||
app = compile_graph(spec, run_ctx).compile(checkpointer=self.checkpointer)
|
||
config = {
|
||
"configurable": {"thread_id": run_id},
|
||
"recursion_limit": settings.max_graph_steps,
|
||
}
|
||
|
||
interrupted = False
|
||
async for mode, chunk in app.astream(
|
||
payload, config=config, stream_mode=["custom", "updates"]
|
||
):
|
||
if mode == "custom":
|
||
await self._handle_custom(run_id, chunk)
|
||
elif mode == "updates":
|
||
if await self._handle_updates(run_id, chunk):
|
||
interrupted = True
|
||
|
||
snapshot = await app.aget_state(config)
|
||
final_state = dict(snapshot.values or {})
|
||
if interrupted or snapshot.interrupts:
|
||
status = "interrupted"
|
||
elif snapshot.next:
|
||
status = "interrupted"
|
||
|
||
except asyncio.CancelledError:
|
||
status = "cancelled"
|
||
error = "用户取消"
|
||
await self._emit(run_id, EventType.RUN_CANCELLED, data={})
|
||
raise
|
||
except Exception as e: # noqa: BLE001
|
||
status = "failed"
|
||
error = f"{type(e).__name__}: {e}"
|
||
logger.exception("run %s 失败", run_id)
|
||
await self._emit(run_id, EventType.RUN_FAILED, data={"error": error})
|
||
finally:
|
||
elapsed = int((time.perf_counter() - started) * 1000)
|
||
await self._finalize(run_id, status, error, final_state, elapsed)
|
||
self._tasks.pop(run_id, None)
|
||
# 运行到终态就把沙箱会话收掉。thread_id 每个 run 都是新的，
|
||
# 不收的话每跑一次带代码节点的图就多一台常驻 microVM（几百 MB），
|
||
# 而且没有任何自动回收路径会碰它——闲置回收只在"下次有人执行
|
||
# 代码"时才触发。interrupted 要留着，人回来还得接着跑。
|
||
if status in ("succeeded", "failed", "cancelled"):
|
||
with contextlib.suppress(Exception):
|
||
from app.sandbox.manager import sandbox_manager
|
||
|
||
await sandbox_manager.cleanup(run_id)
|
||
|
||
async def _handle_custom(self, run_id: str, chunk: Any) -> None:
|
||
"""节点通过 get_stream_writer 发来的事件。"""
|
||
if not isinstance(chunk, dict) or not chunk.get("__agentlab__"):
|
||
return
|
||
await self._emit(
|
||
run_id,
|
||
chunk.get("type", EventType.LOG),
|
||
node_id=chunk.get("node_id"),
|
||
data=chunk.get("data") or {},
|
||
)
|
||
|
||
async def _handle_updates(self, run_id: str, chunk: Any) -> bool:
|
||
"""LangGraph 的状态更新。这里只关心中断信号 —— 节点级事件已经由 custom 流覆盖。"""
|
||
if not isinstance(chunk, dict):
|
||
return False
|
||
interrupts = chunk.get("__interrupt__")
|
||
if not interrupts:
|
||
return False
|
||
|
||
for item in interrupts:
|
||
value = getattr(item, "value", item)
|
||
interrupt_id = getattr(item, "id", None)
|
||
payload = value if isinstance(value, dict) else {"value": value}
|
||
async with SessionLocal() as session:
|
||
existing = (
|
||
await session.execute(
|
||
select(Approval).where(
|
||
Approval.run_id == run_id,
|
||
Approval.interrupt_id == interrupt_id,
|
||
Approval.status == "pending",
|
||
)
|
||
)
|
||
).scalar_one_or_none()
|
||
if existing is None:
|
||
session.add(
|
||
Approval(
|
||
run_id=run_id,
|
||
node_id=str(payload.get("node_id") or ""),
|
||
interrupt_id=interrupt_id,
|
||
mode=str(payload.get("mode") or "approve"),
|
||
title=str(payload.get("title") or "需要你确认"),
|
||
payload=payload,
|
||
schema_=payload.get("schema") or {},
|
||
)
|
||
)
|
||
await session.commit()
|
||
await self._emit(
|
||
run_id,
|
||
EventType.RUN_INTERRUPTED,
|
||
node_id=payload.get("node_id"),
|
||
data={"interrupt_id": interrupt_id, "payload": _safe(payload)},
|
||
)
|
||
return True
|
||
|
||
async def _finalize(
|
||
self,
|
||
run_id: str,
|
||
status: str,
|
||
error: str | None,
|
||
state: dict[str, Any],
|
||
elapsed_ms: int,
|
||
) -> None:
|
||
usage = dict(state.get("usage") or {})
|
||
usage["duration_ms"] = elapsed_ms
|
||
output = dict(state.get("output") or {})
|
||
|
||
async with SessionLocal() as session:
|
||
run = await session.get(Run, run_id)
|
||
if run:
|
||
run.status = status
|
||
run.error = error
|
||
run.output = output
|
||
run.usage = usage
|
||
if status != "interrupted":
|
||
run.finished_at = datetime.now(timezone.utc)
|
||
# 终态时对全部已落库事件计算清单哈希：事后任何对事件流的
|
||
# 增删改都会与这个值对不上。中断态不算——事件还会继续追加。
|
||
from app.core.artifact_store import manifest_hash as _manifest
|
||
|
||
rows = await session.execute(
|
||
select(RunEvent.seq, RunEvent.type, RunEvent.node_id, RunEvent.data)
|
||
.where(RunEvent.run_id == run_id)
|
||
.order_by(RunEvent.seq)
|
||
)
|
||
run.manifest_hash = _manifest([tuple(r) for r in rows])
|
||
await session.commit()
|
||
|
||
if status == "succeeded":
|
||
await self._emit(
|
||
run_id,
|
||
EventType.RUN_FINISHED,
|
||
data={"output": _safe(output), "usage": usage, "duration_ms": elapsed_ms},
|
||
)
|
||
if status in ("succeeded", "failed", "cancelled"):
|
||
await bus.close(run_id)
|
||
|
||
|
||
def _safe(value: Any, limit: int = 4000) -> Any:
|
||
"""把任意对象裁剪成能安全放进事件负载的形状。"""
|
||
if isinstance(value, str):
|
||
return value[:limit]
|
||
if isinstance(value, dict):
|
||
return {k: _safe(v, limit // 2) for k, v in list(value.items())[:50]}
|
||
if isinstance(value, list):
|
||
return [_safe(v, limit // 2) for v in value[:50]]
|
||
if isinstance(value, (int, float, bool)) or value is None:
|
||
return value
|
||
return str(value)[:limit]
|
||
|
||
|
||
run_manager = RunManager()
|