mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 07:02:00 +00:00
ThearchyHelios
56ca53ad35
用量（第 1 波遗留）：查清楚了，丢的不是浏览器那边——问数据页一直读到 SSE 结束、usage 那一条收得到——而是服务端生成过程
（api/copilot_gen._iter_ops）读到 done、look、ask 那一行就关掉模型的流，流末尾那一段才报的用量（Anthropic 的输出 token、
OpenAI 兼容系的 stream_usage）一次都拿不到，每次建图都记成「估不出」，这一轮的花费多半是下限。改成交出那一行之后先把流读完
再关（_drain，内容丢掉）：照约定那一行就是最后一行，读完只多收用量那一段；模型在那一行之后还往下写超过 DRAIN_CHARS 字闲话的
当场关流、照实记成没报全（不为闲话付钱，也不当成报全了）。评测的作答（app/eval/answer.Driver）走同一个 answer_ops，
用量跟着报全，和问数据页同一个口径。test_usage_meter：原先断言「输出 token 为 0、记成估不出」的那条改成报全；另加 look 后、
闲话超限、评测和问数据页收到同一份用量三条；去掉修复回跑，三条都抓得住。providers/usage.py 的说明跟着改。
通知：notify_eval 的链接从定时任务的触发历史改成这次评测的详情页 /evals/runs/{id}（评测页这一波上线）；测试设上对外访问地址、
断言链接（原来的断言在没设地址时对空列表恒真）。
评测接口（只加不改，app/api/evals.py）：题集列表和详情带 latest（最近一次评测）、trend（同一变体最近几次评完的准确率，旧的
在前，换了模型、目录的不连成一条线）、schedules（目标是它的定时任务）；评测详情带 actions（停止、只重试没做完的、重新评分、
钉住 / 取消基线能不能做，和接口拒收的条件同一份）和没做完的项数；POST /suites/{id}/cases/preview 写 SQL 定期望的预览
（同期望的 SQL 探针，交回列和前 50 行，遮罩列照保存时那一道拒收、没过守卫的指明是哪一格，都不回 422）；GET /signals
「口径命中与重复度」的判据和分档、窗口趋势（数只出自 ledger/stats.compute_stats）、成员不少于 3 的问题类（只出自
chat/questions，评测运行不进）。为什么放服务端：界面层不放逻辑，统计、能不能做都由服务端给。tests/test_evals_ui_api.py。
289 lines
15 KiB
Python
289 lines
15 KiB
Python
"""评测题集的定时与通知（app/eval/scheduled.py、notify/service.notify_eval；细案 01 做法 15，总方案 X4、X6）。
|
||
|
||
- 定时接口按 target_kind 收评测题集：核对目标交给 app/eval/scheduled.check_target（题集在、没归档、有启用的题）；非工作流
|
||
目标不收工作流、输入模板、Webhook（Webhook 的事件说的是那一次运行，非工作流目标收不到，照实拒）；接口返回目标的种类、id、
|
||
名字；工作流的定时行为一字不变（test_schedules_api.py）；
|
||
- 到点建一次评测：定时跑 1 次，截至时间是计划触发时间，ScheduleFire.eval_run_id 指向它；上一次还没结束就 skipped_overlap；
|
||
错过的按宽限补（同一套调度循环）；
|
||
- 有一道题从通过变成不通过推「定时评测有退步」，没评完推「定时评测没有评完」：只推订阅了的全局渠道，消息里只有题集名、
|
||
通过数、退步题数、链接，不放期望值和答案里的数。手动发起的评测不推（叫法照实带「定时」）。题集名、渠道都是虚构的。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import json
|
||
import uuid
|
||
from datetime import timedelta, timezone
|
||
|
||
import httpx
|
||
import pytest
|
||
from httpx import ASGITransport, AsyncClient
|
||
from sqlalchemy import delete, select
|
||
|
||
from app.core.crypto import encrypt
|
||
from app.db.base import SessionLocal, utcnow
|
||
from app.db.models import NotifyChannel, Schedule
|
||
from app.db.models_eval import EvalResult, EvalRun
|
||
from app.engine.runner import run_manager
|
||
from app.eval import answer as answer_mod
|
||
from app.eval import runner
|
||
from app.main import app
|
||
from app.notify import core
|
||
from app.notify import service as notify
|
||
from app.scheduler import loop, service
|
||
from tests import eval_samples as es
|
||
|
||
UTC = timezone.utc
|
||
URL = "https://hooks.example.com/agentlab/eval-7c2a"
|
||
|
||
|
||
@pytest.fixture(autouse=True)
|
||
async def engine_up():
|
||
await run_manager.setup()
|
||
async with SessionLocal() as session:
|
||
await session.execute(delete(Schedule))
|
||
await session.commit()
|
||
yield
|
||
await runner.manager.shutdown()
|
||
await notify.drain()
|
||
async with SessionLocal() as session:
|
||
await session.execute(delete(Schedule))
|
||
await session.execute(delete(NotifyChannel).where(NotifyChannel.name.like("评测通知-%")))
|
||
await session.commit()
|
||
await es.drop_suites()
|
||
await run_manager.shutdown()
|
||
|
||
|
||
@pytest.fixture
|
||
async def client():
|
||
async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
|
||
yield c
|
||
|
||
|
||
@pytest.fixture
|
||
def received(monkeypatch):
|
||
got: list[httpx.Request] = []
|
||
|
||
def handler(request: httpx.Request) -> httpx.Response:
|
||
got.append(request)
|
||
return httpx.Response(200, json={"ok": True})
|
||
|
||
monkeypatch.setattr(core, "_transport", httpx.MockTransport(handler))
|
||
monkeypatch.setattr(core, "RETRY_DELAYS", (0, 0, 0))
|
||
return got
|
||
|
||
|
||
async def _public_url(url: str):
|
||
"""临时设对外访问地址（通知里的链接按它拼），交回复原的函数：别的测试不受影响。"""
|
||
from app.core import public_url
|
||
from app.db.models import Setting
|
||
|
||
async with SessionLocal() as session:
|
||
row = await session.get(Setting, "notify")
|
||
before = dict(row.value or {}) if row is not None else None
|
||
if row is None:
|
||
session.add(Setting(key="notify", value={"public_base_url": url}))
|
||
else:
|
||
row.value = {**(row.value or {}), "public_base_url": url}
|
||
await session.commit()
|
||
await public_url.refresh(session)
|
||
|
||
async def restore() -> None:
|
||
async with SessionLocal() as session:
|
||
row = await session.get(Setting, "notify")
|
||
if before is None and row is not None:
|
||
await session.delete(row)
|
||
elif row is not None:
|
||
row.value = before
|
||
await session.commit()
|
||
await public_url.refresh(session)
|
||
|
||
return restore
|
||
|
||
|
||
def stub(monkeypatch, *, delay: float = 0.0, path: str = "asked"):
|
||
async def drive(self):
|
||
await asyncio.sleep(delay)
|
||
self.turn.path = path
|
||
return self.turn
|
||
|
||
monkeypatch.setattr(answer_mod.Driver, "drive", drive)
|
||
|
||
|
||
ASK = [{"question": "含糊的问题", "expectations": [{"kind": "ask"}]}]
|
||
|
||
|
||
async def _schedule(client, suite_id: str, **extra) -> dict:
|
||
made = await client.post("/api/schedules", json={"name": "每周回归", "target_kind": "eval_suite",
|
||
"target_id": suite_id, "cron": "0 9 * * 1",
|
||
"timezone": "Asia/Shanghai", **extra})
|
||
assert made.status_code == 201, made.text
|
||
return made.json()
|
||
|
||
|
||
async def test_the_schedule_api_takes_an_eval_suite(client):
|
||
suite_id = await es.make_suite("none", name="虚构回归题集", cases=ASK)
|
||
body = await _schedule(client, suite_id)
|
||
assert (body["target_kind"], body["target_id"], body["target_name"]) == ("eval_suite", suite_id, "虚构回归题集")
|
||
assert body["target_label"] == "评测题集" and body["workflow_id"] == "" and body["webhook"] is None
|
||
assert body["input_template"] == {} and body["target_exists"] is True
|
||
listed = (await client.get("/api/schedules")).json()["items"]
|
||
assert [s["target_name"] for s in listed] == ["虚构回归题集"]
|
||
for bad, field in (({"target_id": "nope"}, "target_id"), ({"target_id": None}, "target_id"),
|
||
({"webhook": {"url": URL}}, "webhook"), ({"input_template": {"a": "b"}}, "input_template"),
|
||
({"workflow_id": "wf"}, "workflow_id"), ({"target_kind": "nope"}, "target_kind")):
|
||
refused = await client.post("/api/schedules", json={"name": "x", "target_kind": "eval_suite",
|
||
"target_id": suite_id, "cron": "0 9 * * 1",
|
||
"timezone": "Asia/Shanghai", **bad})
|
||
assert refused.status_code == 422 and refused.json()["field"] == field, (bad, refused.text)
|
||
# 改名、改时间、停用都照常；目标种类建好不能改
|
||
patched = await client.patch(f"/api/schedules/{body['id']}", json={"name": "每天回归", "cron": "0 9 * * *",
|
||
"enabled": False, "webhook": None,
|
||
"version": body["version"]})
|
||
assert patched.status_code == 200, patched.text
|
||
assert patched.json()["name"] == "每天回归" and patched.json()["next_run_at"] is None
|
||
moved = await client.patch(f"/api/schedules/{body['id']}", json={"target_kind": "workflow",
|
||
"version": patched.json()["version"]})
|
||
assert moved.status_code == 422 and moved.json()["field"] == "target_kind"
|
||
# 题集归档了：照实拒
|
||
archived = await es.make_suite("none", name="归档的题集", cases=ASK)
|
||
async with SessionLocal() as session:
|
||
from app.db.models_eval import EvalSuite
|
||
|
||
(await session.get(EvalSuite, archived)).archived = True
|
||
await session.commit()
|
||
refused = await client.post("/api/schedules", json={"name": "x", "target_kind": "eval_suite", "target_id": archived,
|
||
"cron": "0 9 * * 1", "timezone": "Asia/Shanghai"})
|
||
assert refused.status_code == 422 and "已归档" in refused.json()["detail"]
|
||
|
||
|
||
async def test_firing_starts_one_evaluation_and_skips_on_overlap(client, monkeypatch):
|
||
stub(monkeypatch, delay=0.5)
|
||
suite_id = await es.make_suite("none", cases=ASK, default_variant={"repeats": 3})
|
||
body = await _schedule(client, suite_id)
|
||
now = utcnow().replace(microsecond=0)
|
||
fire = await service.fire_once(body["id"], now, "manual", actor="王芳")
|
||
assert fire.status == "started" and fire.eval_run_id and fire.run_id is None
|
||
async with SessionLocal() as session:
|
||
run = await session.get(EvalRun, fire.eval_run_id)
|
||
assert run.trigger == "schedule" and run.schedule_id == body["id"] and run.repeats == 1, "定时只跑 1 次"
|
||
assert run.as_of.replace(tzinfo=UTC) == now and run.created_by == "王芳"
|
||
again = await service.fire_once(body["id"], now + timedelta(seconds=1), "manual")
|
||
assert again.status == "skipped_overlap" and f"#{run.id[:6]}" in again.reason
|
||
same = await service.fire_once(body["id"], now, "manual")
|
||
assert same.id == fire.id, "同一次触发只发起一次"
|
||
await es.wait_eval(run.id)
|
||
history = (await client.get(f"/api/schedules/{body['id']}/fires")).json()["items"]
|
||
assert {f["eval_run_id"] for f in history if f["status"] == "started"} == {run.id}
|
||
|
||
|
||
async def test_a_missed_fire_within_grace_is_made_up(client, monkeypatch):
|
||
stub(monkeypatch)
|
||
suite_id = await es.make_suite("none", cases=ASK)
|
||
body = await _schedule(client, suite_id)
|
||
due = utcnow().replace(microsecond=0) - timedelta(minutes=20)
|
||
async with SessionLocal() as session:
|
||
await session.execute(Schedule.__table__.update().where(Schedule.id == body["id"]).values(next_run_at=due))
|
||
await session.commit()
|
||
fires = await loop.fire_due(due + timedelta(minutes=20))
|
||
[made_up] = [f for f in fires if f.status == "started"]
|
||
assert made_up.kind == "misfire" and made_up.eval_run_id
|
||
await es.wait_eval(made_up.eval_run_id)
|
||
|
||
|
||
async def _channel(events: list[str]) -> str:
|
||
async with SessionLocal() as session:
|
||
row = NotifyChannel(name=f"评测通知-{uuid.uuid4().hex[:6]}", scope="global", format="generic", url=encrypt(URL),
|
||
events=events)
|
||
session.add(row)
|
||
await session.commit()
|
||
return row.id
|
||
|
||
|
||
async def test_a_regression_is_pushed_without_numbers(client, monkeypatch, received, tmp_path):
|
||
name, sid = await es.make_source(tmp_path)
|
||
es.use_copilot(monkeypatch, es.Copilot(name))
|
||
es.use_report(monkeypatch)
|
||
suite_id = await es.make_suite(sid, name="虚构周回归")
|
||
body = await _schedule(client, suite_id)
|
||
await _channel(["eval.regressed", "eval.failed"])
|
||
first = await service.fire_once(body["id"], utcnow().replace(microsecond=0) - timedelta(minutes=5), "manual")
|
||
await es.wait_eval(first.eval_run_id)
|
||
assert received == [], "第一次没有可比的、也评完了：不推"
|
||
await es.set_notes(sid, {"label": es.NOTES["label"]})
|
||
second = await service.fire_once(body["id"], utcnow().replace(microsecond=0), "manual")
|
||
await es.wait_eval(second.eval_run_id)
|
||
[request] = received
|
||
message = json.loads(request.content)
|
||
assert message["event"] == "eval.regressed" and message["title"] == "定时评测「虚构周回归」有退步"
|
||
text = message["text"]
|
||
assert "通过：0 / 1 题" in text and "退步：1 题" in text
|
||
for secret in ("330", "310", "gmv", "orders", "paid_date"):
|
||
assert secret not in json.dumps(message, ensure_ascii=False), f"通知里不该有 {secret}"
|
||
from app.db.models import NotifyDelivery
|
||
|
||
async with SessionLocal() as session:
|
||
[delivery] = (await session.execute(select(NotifyDelivery).where(
|
||
NotifyDelivery.subject_id == second.eval_run_id))).scalars().all()
|
||
assert delivery.subject_kind == "eval_run" and delivery.status == "ok"
|
||
from tests.fixtures.sources import drop_source
|
||
|
||
await drop_source(sid)
|
||
|
||
|
||
async def test_an_unfinished_scheduled_evaluation_pushes_failed(client, monkeypatch, received):
|
||
async def drive(self):
|
||
self.turn.path, self.turn.copilot_failed, self.turn.reason = "build_failed", True, "助手本轮未完成"
|
||
return self.turn
|
||
|
||
monkeypatch.setattr(answer_mod.Driver, "drive", drive)
|
||
suite_id = await es.make_suite("none", cases=ASK * 4, name="虚构失败题集", budget={"max_parallel": 1})
|
||
body = await _schedule(client, suite_id)
|
||
await _channel(["eval.failed"])
|
||
restore = await _public_url("https://agentlab.example.com")
|
||
try:
|
||
fire = await service.fire_once(body["id"], utcnow().replace(microsecond=0), "manual")
|
||
await es.wait_eval(fire.eval_run_id)
|
||
finally:
|
||
await restore()
|
||
async with SessionLocal() as session:
|
||
run = await session.get(EvalRun, fire.eval_run_id)
|
||
skipped = (await session.execute(select(EvalResult).where(
|
||
EvalResult.eval_run_id == run.id, EvalResult.status == "skipped"))).scalars().all()
|
||
assert run.status == "partial" and run.stop_reason.startswith("连续 3 题") and len(skipped) == 1
|
||
[request] = received
|
||
message = json.loads(request.content)
|
||
assert message["event"] == "eval.failed" and message["title"] == "定时评测「虚构失败题集」没有评完"
|
||
assert "通过：0 / 3 题" in message["text"]
|
||
# 链接去这次评测的详情页（第 4 波评测页上线之前去定时任务的触发历史）；设了对外访问地址才有链接
|
||
assert [link["url"] for link in message["links"]] == [f"https://agentlab.example.com/evals/runs/{fire.eval_run_id}"], \
|
||
message["links"]
|
||
|
||
|
||
def test_the_eval_events_say_they_are_for_scheduled_evaluations():
|
||
"""这两个事件只在定时发起的评测收尾时推（runner._conclude 认 trigger == "schedule"）：手动发起的那次人自己在看，不推。
|
||
叫法得照实说「定时」——只写「评测失败」，人勾上之后手动评测失败了永远收不到，界面上也没处说为什么（复核 L3）；
|
||
eval.failed 也不只是失败：到了上限、止损停下（partial）同样推，所以说「没有评完」。"""
|
||
assert notify.EVENTS["eval.regressed"].label == "定时评测有退步"
|
||
assert notify.EVENTS["eval.failed"].label == "定时评测没有评完"
|
||
|
||
|
||
async def test_a_manual_evaluation_that_stops_short_pushes_nothing(client, monkeypatch, received):
|
||
"""手动发起的评测止损停下：订阅了「定时评测没有评完」的渠道也不推（叫法说的就是定时的）。"""
|
||
async def drive(self):
|
||
self.turn.path, self.turn.copilot_failed, self.turn.reason = "build_failed", True, "助手本轮未完成"
|
||
return self.turn
|
||
|
||
monkeypatch.setattr(answer_mod.Driver, "drive", drive)
|
||
suite_id = await es.make_suite("none", cases=ASK * 4, name="虚构手动题集", budget={"max_parallel": 1})
|
||
await _channel(["eval.regressed", "eval.failed"])
|
||
started = await client.post(f"/api/evals/suites/{suite_id}/evaluations", json={})
|
||
assert started.status_code == 202, started.text
|
||
[ev] = started.json()["evaluations"]
|
||
await es.wait_eval(ev["id"])
|
||
await notify.drain()
|
||
async with SessionLocal() as session:
|
||
run = await session.get(EvalRun, ev["id"])
|
||
assert run.trigger == "manual" and run.status == "partial" and run.stop_reason.startswith("连续 3 题")
|
||
assert received == [], "手动发起的评测不推"
|