AgentLab/backend/tests/test_eval_runner.py

"""评测执行（app/eval/runner.py；细案 01 做法 9，评测-D9）。

作答这一步多半换成桩（Driver.drive）：这里守的是后台任务本身——排队、并发上限（不超过 max_parallel，也不超过全局运行
并发的四分之一）、停止、只重试没做完的、重启标中断不续跑、单次和每日金额上限（评测单独一本账，不占裁判、核查的额度）、
连续几题助手调用失败止损、每题超时会取消正在跑的运行；停在审批上的作答走真的运行（评测不代批，取消掉）。
题目多半是「期望反问」的题：没有期望来源，不用跑期望探针。问法、题集名都是虚构的。
"""
from __future__ import annotations

import asyncio

import pytest
from httpx import ASGITransport, AsyncClient
from sqlalchemy import select

from app.db.base import SessionLocal, utcnow
from app.db.models import Run, Setting
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 tests import eval_samples as es
from tests.fixtures.sources import drop_source


async def _clear_ledger() -> None:
    async with SessionLocal() as session:
        row = await session.get(Setting, "eval_spend")
        if row is not None:
            await session.delete(row)
            await session.commit()


@pytest.fixture(autouse=True)
async def engine_up(monkeypatch):
    # 同一个测试进程里别的测试建的数据源可能超过评测拍目录快照的上限（app/eval/snapshots.MAX_SOURCES），评测会按
    # 设计报「数据源超过 20 个」——这里的评测不测这个上限，分到哪个进程都要照常跑
    from app.eval import snapshots

    monkeypatch.setattr(snapshots, "MAX_SOURCES", 10_000)
    await run_manager.setup()
    await _clear_ledger()
    yield
    await _clear_ledger()
    await runner.manager.shutdown()
    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


async def _until(cond, what: str, *, timeout: float = 30.0) -> None:
    """等到 cond() 为真；等不到就让测试失败、说清在等什么，不让整次测试卡死（作答运行在排队之前就失败时，排队永远等不到）。"""
    loop = asyncio.get_running_loop()
    deadline = loop.time() + timeout
    while not await cond():
        if loop.time() > deadline:
            raise AssertionError(f"等了 {timeout:g} 秒还没等到：{what}")
        await asyncio.sleep(0.02)


def ask_cases(n: int) -> list[dict]:
    return [{"question": f"第 {i} 个含糊的问题", "expectations": [{"kind": "ask"}]} for i in range(1, n + 1)]


def stub(monkeypatch, *, delay: float = 0.0, path: str = "asked", cost: float = 0.0, copilot_failed: bool = False,
         current_run: str | None = None, live: list | None = None):
    """作答换成桩：等 delay 秒，交回 path；cost 记在助手的用量上。live 记下同时在答的有几题。"""
    peak = {"now": 0, "max": 0}

    async def drive(self):
        peak["now"] += 1
        peak["max"] = max(peak["max"], peak["now"])
        try:
            self.turn.current_run = current_run
            await self._phase("build")
            await asyncio.sleep(delay)
            self.turn.path = path
            self.turn.copilot_failed = copilot_failed
            if copilot_failed:
                self.turn.reason = "助手本轮未完成：接入返回了 502"
            self.turn.usage["assistant"] = {"by_model": {}, "input_tokens": 0, "output_tokens": 0, "calls": 1,
                                            "cost_usd": cost, "unpriced_calls": 0}
            return self.turn
        finally:
            peak["now"] -= 1

    monkeypatch.setattr(answer_mod.Driver, "drive", drive)
    return peak


async def start(suite_id: str, variant: dict | None = None, **kw) -> list[EvalRun]:
    from app.db.models_eval import EvalSuite
    from app.eval.variants import normalize_variant

    async with SessionLocal() as session:
        suite = await session.get(EvalSuite, suite_id)
        return await runner.start(session, suite, [normalize_variant(variant or {})], actor="评测测试", **kw)


async def results(run_id: str) -> list[EvalResult]:
    async with SessionLocal() as session:
        return list((await session.execute(select(EvalResult).where(EvalResult.eval_run_id == run_id)
                                           .order_by(EvalResult.seq))).scalars())


async def get(run_id: str) -> EvalRun:
    async with SessionLocal() as session:
        return await session.get(EvalRun, run_id)


async def test_items_run_in_the_background_within_the_parallel_cap(monkeypatch):
    peak = stub(monkeypatch, delay=0.15)
    suite_id = await es.make_suite("none", cases=ask_cases(5), budget={"max_parallel": 3})
    [run] = await start(suite_id)
    assert run.status == "queued" and run.counts["total"] == 5
    await es.wait_eval(run.id)
    row = await get(run.id)
    assert row.status == "succeeded" and row.counts["done"] == 5 and row.counts["passed"] == 5
    assert peak["max"] == 3, "同时最多 max_parallel 题"
    assert all(r.status == "ok" and r.verdict == "pass" and r.current is None for r in await results(run.id))
    assert row.summary["accuracy"]["passed"] == 5 and row.summary["verdicts"]["pass"] == 5


async def test_the_cap_never_exceeds_a_quarter_of_the_global_concurrency(monkeypatch):
    from app.core.config import settings

    monkeypatch.setattr(settings, "max_concurrent_runs", 4)
    peak = stub(monkeypatch, delay=0.1)
    suite_id = await es.make_suite("none", cases=ask_cases(3), budget={"max_parallel": 3})
    [run] = await start(suite_id)
    await es.wait_eval(run.id)
    assert peak["max"] == 1, "全局并发 4 的四分之一是 1"


async def test_a_second_evaluation_of_the_same_suite_waits(monkeypatch):
    stub(monkeypatch, delay=0.3)
    suite_id = await es.make_suite("none", cases=ask_cases(1))
    [run] = await start(suite_id)
    with pytest.raises(runner.EvalRefused) as refused:
        await start(suite_id)
    assert refused.value.code == runner.EVAL_BUSY and refused.value.status == 409
    await es.wait_eval(run.id)


async def test_stop_then_retry_only_the_unfinished(monkeypatch):
    stub(monkeypatch, delay=0.05)
    suite_id = await es.make_suite("none", cases=ask_cases(4), budget={"max_parallel": 1})
    [run] = await start(suite_id)
    # 等第一题评完再停
    for _ in range(100):
        if (await get(run.id)).counts.get("done", 0) >= 1:
            break
        await asyncio.sleep(0.02)
    stub(monkeypatch, delay=5.0)
    await asyncio.sleep(0.1)
    assert await runner.stop(run.id)
    row = await get(run.id)
    # 停止是「停下」不是「出错」：原因记在 stop_reason；没跑的题单独计 not_run，不算进 errors，done 不再等于 total
    assert row.status == "cancelled" and row.stop_reason == runner.STOPPED and row.error is None
    before = {r.id: (r.status, r.finished_at) for r in await results(run.id)}
    done = [rid for rid, (status, _) in before.items() if status == "ok"]
    assert done and any(status == "cancelled" for status, _ in before.values())
    assert row.counts["errors"] == 0 and row.counts["done"] == len(done)
    assert row.counts["not_run"] == 4 - len(done) and row.counts["done"] + row.counts["not_run"] == row.counts["total"]
    assert row.summary["not_run"] == 4 - len(done) and row.summary["errors"] == 0
    assert not await runner.stop(run.id), "已经结束的停不了"

    stub(monkeypatch, delay=0.01)
    await runner.retry(run.id)
    await es.wait_eval(run.id)
    after = {r.id: (r.status, r.finished_at) for r in await results(run.id)}
    assert all(status == "ok" for status, _ in after.values())
    assert all(after[rid] == before[rid] for rid in done), "评完了的不重做"
    assert (await get(run.id)).status == "succeeded"
    with pytest.raises(runner.EvalRefused) as nothing:
        await runner.retry(run.id)
    assert nothing.value.code == runner.NOTHING_TO_RETRY


async def test_a_restart_marks_unfinished_evaluations_interrupted_and_does_not_resume():
    suite_id = await es.make_suite("none", cases=ask_cases(1))
    async with SessionLocal() as session:
        row = EvalRun(suite_id=suite_id, status="running", started_at=utcnow())
        session.add(row)
        await session.flush()
        session.add(EvalResult(eval_run_id=row.id, status="running"))
        await session.commit()
        run_id = row.id
    assert await runner.startup() >= 1
    row = await get(run_id)
    assert row.status == "interrupted" and row.stop_reason == runner.INTERRUPTED and row.error is None
    assert row.counts["not_run"] == 1 and row.counts["running"] == 0 and row.counts["errors"] == 0, row.counts
    [item] = await results(run_id)
    assert item.status == "cancelled"
    assert not runner.manager.running(run_id), "不自动续跑"


async def test_stopping_a_queued_evaluation_counts_everything_as_not_run(monkeypatch):
    """排队中就停掉（还没开始）：每一题都没跑，stop_reason 写停止，计数跟着重算（以前还是建的时候那份：pending 照旧）。"""
    stub(monkeypatch)
    suite_id = await es.make_suite("none", cases=ask_cases(2))
    monkeypatch.setattr(runner.manager, "start", lambda *_a, **_k: None)  # 不让它开跑：停在排队
    [run] = await start(suite_id)
    assert await runner.stop(run.id)
    row = await get(run.id)
    assert row.status == "cancelled" and row.stop_reason == runner.STOPPED
    assert (row.counts["not_run"], row.counts["pending"], row.counts["done"]) == (2, 0, 0), row.counts


async def test_the_per_run_cap_stops_new_items(monkeypatch):
    stub(monkeypatch, cost=1.0)
    suite_id = await es.make_suite("none", cases=ask_cases(3), budget={"max_parallel": 1, "per_run_max_usd": 1.5})
    [run] = await start(suite_id)
    await es.wait_eval(run.id)
    row = await get(run.id)
    items = await results(run.id)
    assert [r.status for r in items] == ["ok", "ok", "skipped"]
    assert row.status == "partial" and row.stop_reason == runner.LIMIT_RUN.format(limit=1.5)
    assert (row.counts["done"], row.counts["not_run"], row.counts["errors"], row.counts["skipped"]) == (2, 1, 0, 0), \
        "到了上限没开评的是没跑，不是跳过（跳过是评了、但期望算不出、不评分）"
    assert "金额上限" in items[2].error and row.summary["cost_usd"] == 2.0


async def test_the_daily_cap_is_its_own_ledger(monkeypatch):
    from app.engine import judge
    from app.engine.spend import _today, eval_spend, verify_spend

    stub(monkeypatch, cost=0.4)
    async with SessionLocal() as session:
        for key in ("eval_spend", "judge_spend", "verify_spend"):
            row = await session.get(Setting, key)
            if row is not None:
                await session.delete(row)
        session.add(Setting(key="eval_spend", value={"date": _today(), "usd": 4.9, "calls": 3, "unpriced_calls": 0}))
        await session.commit()
    judge_before, verify_before = await judge.judge_spend.daily(), await verify_spend.daily()
    suite_id = await es.make_suite("none", cases=ask_cases(3), budget={"max_parallel": 1})
    [first] = await start(suite_id)
    await es.wait_eval(first.id)
    # 第一次没有可估的每题花费：按实际累计，越过每日上限之后不再开新题
    spent = await eval_spend.daily()
    assert spent["usd"] > 5.0 - 1e-9 and spent["calls"] == 3 + 1, spent
    row = await get(first.id)
    assert [r.status for r in await results(first.id)] == ["ok", "skipped", "skipped"]
    assert row.stop_reason == runner.LIMIT_DAILY.format(limit=5.0)
    assert await judge.judge_spend.daily() == judge_before and await verify_spend.daily() == verify_before, \
        "评测不占裁判、核查的额度"


async def test_consecutive_assistant_failures_stop_the_evaluation(monkeypatch):
    stub(monkeypatch, path="build_failed", copilot_failed=True)
    suite_id = await es.make_suite("none", cases=ask_cases(5), budget={"max_parallel": 1})
    [run] = await start(suite_id)
    await es.wait_eval(run.id)
    row = await get(run.id)
    statuses = [r.status for r in await results(run.id)]
    assert statuses[:3] == ["ok", "ok", "ok"] and statuses[3:] == ["skipped", "skipped"]
    assert row.stop_reason.startswith("连续 3 题助手调用失败") and "502" in row.stop_reason


async def test_too_many_timeouts_stop_the_evaluation(monkeypatch):
    async def drive(self):
        self.turn.path, self.turn.timed_out, self.turn.reason = "build_failed", True, "超过每题时限"
        return self.turn

    monkeypatch.setattr(answer_mod.Driver, "drive", drive)
    suite_id = await es.make_suite("none", cases=ask_cases(9), budget={"max_parallel": 1})
    [run] = await start(suite_id)
    await es.wait_eval(run.id)
    row = await get(run.id)
    statuses = [r.status for r in await results(run.id)]
    assert statuses == ["ok"] * runner.STOP_TIMEOUTS + ["skipped"] * (9 - runner.STOP_TIMEOUTS)
    assert row.stop_reason == runner.STOP_TIMEOUT_TEXT.format(window=runner.STOP_WINDOW, n=runner.STOP_TIMEOUTS)
    assert row.summary["not_run"] == 9 - runner.STOP_TIMEOUTS


async def test_a_case_over_its_time_limit_stops_its_run(monkeypatch):
    stopped: list[str | None] = []

    async def fake_stop(run_id, actor=None):
        stopped.append(run_id)

    monkeypatch.setattr(answer_mod, "stop_run", fake_stop)
    stub(monkeypatch, delay=5.0, current_run="run-in-flight")
    original = runner.EvalExecution._prepare

    async def prepare(self):
        units = await original(self)
        self.budget["per_case_timeout_s"] = 0.2
        return units

    monkeypatch.setattr(runner.EvalExecution, "_prepare", prepare)
    suite_id = await es.make_suite("none", cases=ask_cases(1))
    [run] = await start(suite_id)
    await es.wait_eval(run.id)
    [item] = await results(run.id)
    assert stopped == ["run-in-flight"], "超时取消正在跑的那次运行"
    assert item.status == "ok" and item.path == "build_failed" and "每题时限" in item.error


async def test_time_queued_for_a_run_slot_does_not_count_against_the_case_limit(monkeypatch):
    """复核：满载时作答运行排队等名额的那一段不算每题时限（排了 2 秒、时限 1 秒，拿到名额后很快答完：不记超时）。"""
    from app.cluster import slots

    stored = slots.values({"max_total": 2, "interactive_reserve": 1})

    async def fixed(*, fresh: bool = False):  # noqa: ARG001
        return stored

    monkeypatch.setattr(slots, "policy", fixed)
    slots.forget_policy()
    gate = asyncio.Event()

    async def blocker() -> None:
        async with slots.hold("eval-slot-blocker", "eval"):
            await gate.wait()

    async def drive(self):
        self.turn.current_run = "eval-answer-in-queue"
        await self._phase("build")
        async with slots.hold(self.turn.current_run, "eval"):
            await asyncio.sleep(0.05)
        self.turn.path = "asked"
        self.turn.usage["assistant"] = {"by_model": {}, "input_tokens": 0, "output_tokens": 0, "calls": 1,
                                        "cost_usd": 0.0, "unpriced_calls": 0}
        return self.turn

    monkeypatch.setattr(answer_mod.Driver, "drive", drive)
    original = runner.EvalExecution._prepare

    async def prepare(self):
        units = await original(self)
        self.budget["per_case_timeout_s"] = 1.0
        return units

    monkeypatch.setattr(runner.EvalExecution, "_prepare", prepare)
    held = asyncio.create_task(blocker())
    try:
        async def blocking() -> bool:
            return bool((await slots.snapshot())["classes"]["system"]["running"])

        await _until(blocking, "占名额的那一段拿到名额")
        suite_id = await es.make_suite("none", cases=ask_cases(1))
        [run] = await start(suite_id)
        await _until(lambda: slots.waiting("eval-answer-in-queue"), "作答运行排队等名额")
        await asyncio.sleep(2.0)
        gate.set()
        await es.wait_eval(run.id)
    finally:
        gate.set()
        await held
    [item] = await results(run.id)
    assert item.status == "ok" and item.path == "asked" and not item.error, item.error


async def test_within_does_not_count_the_paused_stretches():
    paused = {"now": True}

    async def is_paused() -> bool:
        return paused["now"]

    async def work() -> str:
        await asyncio.sleep(0.5)
        paused["now"] = False
        await asyncio.sleep(0.05)
        return "done"

    assert await runner.within(work(), 0.2, is_paused, tick=0.02) == "done", "暂停的那 0.5 秒不算"

    async def never():
        await asyncio.sleep(10)

    async def awake() -> bool:
        return False

    loop = asyncio.get_running_loop()
    t0 = loop.time()
    with pytest.raises(asyncio.TimeoutError):
        await runner.within(never(), 0.2, awake, tick=0.02)
    assert loop.time() - t0 < 1.0


async def test_a_run_stopped_at_an_approval_is_cancelled_not_approved(client, monkeypatch, tmp_path):
    from app.eval.answer import APPROVAL_REASON

    class Approving(es.Copilot):
        async def astream(self, messages):
            if "上线前自查" in messages[1][1]:
                yield es._line({"op": "done"})
                return
            for op in [{"op": "add_node", "node": {"id": "start", "type": "input", "label": "输入", "config": {}}},
                       {"op": "add_node", "node": {"id": "ok", "type": "human", "label": "确认",
                                                   "config": {"mode": "approve", "title": "请确认"}}},
                       {"op": "add_edge", "edge": {"source": "start", "target": "ok"}},
                       {"op": "add_node", "node": {"id": "out", "type": "output", "label": "答案",
                                                   "config": {"fields": [{"name": "答案", "value": "好"}]}}},
                       {"op": "add_edge", "edge": {"source": "ok", "target": "out", "sourceHandle": "approved"}},
                       {"op": "done", "explanation": "确认一下", "run": True}]:
                yield es._line(op)

    name, sid = await es.make_source(tmp_path)
    try:
        es.use_copilot(monkeypatch, Approving(name))
        suite_id = await es.make_suite(sid, cases=[{"question": "要人确认的问题", "expectations": [{"kind": "ask"}]}])
        [run] = await start(suite_id)
        await es.wait_eval(run.id)
        [item] = await results(run.id)
        assert item.path == "run_failed" and "评测不代批" in item.error
        async with SessionLocal() as session:
            answered = await session.get(Run, item.run_id)
        assert answered.status == "cancelled", "评测不代批：取消掉"
        details = (await client.get(f"/api/evals/evaluations/{run.id}/results/{item.id}")).json()
        assert details["details"]["turn"]["reason"] == APPROVAL_REASON
    finally:
        await drop_source(sid)


async def test_the_api_starts_estimates_and_reports_progress(client, monkeypatch):
    stub(monkeypatch, delay=0.05)
    suite_id = await es.make_suite("none", cases=ask_cases(2))
    estimate = (await client.post(f"/api/evals/suites/{suite_id}/evaluations/estimate", json={})).json()
    assert estimate["cases"] == 2 and estimate["estimate"]["total_usd"] is None and estimate["parallel"] == 2
    assert estimate["budget"]["daily_max_usd"] == 5.0 and estimate["arms"][0]["models"]["assistant"] == "mock-fast"
    started = await client.post(f"/api/evals/suites/{suite_id}/evaluations", json={"variant": {"repeats": 2}})
    assert started.status_code == 202, started.text
    [ev] = started.json()["evaluations"]
    assert ev["counts"]["total"] == 4 and ev["variant"]["repeats"] == 2
    await es.wait_eval(ev["id"])
    got = (await client.get(f"/api/evals/evaluations/{ev['id']}")).json()
    assert got["status"] == "succeeded" and got["current"] == [] and got["summary"]["pass_rates"]
    listed = (await client.get(f"/api/evals/suites/{suite_id}/evaluations")).json()["evaluations"]
    assert [e["id"] for e in listed] == [ev["id"]]
    rows = (await client.get(f"/api/evals/evaluations/{ev['id']}/results")).json()["results"]
    assert len(rows) == 4 and rows[0]["question"] == "第 1 个含糊的问题"
    second = (await client.post(f"/api/evals/suites/{suite_id}/evaluations/estimate", json={})).json()
    assert second["estimate"]["basis"] == ev["id"], "第二次按上一次的每题平均花费估"
    bad = await client.post(f"/api/evals/suites/{suite_id}/evaluations", json={"variant": {"catalog": "nope"}})
    assert bad.status_code == 422 and bad.json()["code"] == "eval_variant_invalid"


async def test_the_estimate_says_it_cannot_price_instead_of_zero():
    """上一次评测的模型都不在价格目录里（价格目录空）：估算以前写「约 $0」，像不花钱。现在一分都没估出来的写估不出；估出了
    一部分的写「至少」、另有几次估不出（老总表里每题平均记成 0、却有估不出的调用的，也按估不出算）。"""
    from datetime import timedelta

    suite_id = await es.make_suite("none", cases=ask_cases(1))
    now = utcnow()
    async with SessionLocal() as session:
        session.add(EvalRun(suite_id=suite_id, status="succeeded", created_at=now - timedelta(minutes=2),
                            summary={"per_case_cost_usd": 0.0, "unpriced_calls": 6}))
        await session.commit()
        got = await runner.estimate(session, suite_id, 3)
    assert got["total_usd"] is None and got["per_item_usd"] is None and got["unpriced_calls"] == 6, got
    assert "估不出" in got["note"] and got["basis"], got
    async with SessionLocal() as session:
        session.add(EvalRun(suite_id=suite_id, status="succeeded", created_at=now - timedelta(minutes=1),
                            summary={"per_case_cost_usd": 0.02, "unpriced_calls": 2}))
        await session.commit()
        got = await runner.estimate(session, suite_id, 3)
    assert got["total_usd"] == 0.06 and got["at_least"] is True and got["unpriced_calls"] == 2, got
    assert "至少" in got["note"] and "2 次" in got["note"], got["note"]
    async with SessionLocal() as session:
        session.add(EvalRun(suite_id=suite_id, status="succeeded", created_at=now,
                            summary={"per_case_cost_usd": 0.0, "unpriced_calls": 0}))
        await session.commit()
        got = await runner.estimate(session, suite_id, 3)
    assert got["total_usd"] == 0.0 and not got.get("at_least") and not got.get("unpriced_calls"), "目录价是 0 的才是真不花钱"


def test_the_judge_role_keeps_its_unpriced_calls():
    """按角色的用量里裁判那一份以前只有金额和次数：裁判模型不在价格目录里时写成 $0。"""
    usage = runner.judge_usage({"claims": {"cost_usd": 0.0, "calls": 2, "priced": False},
                                "text": {"cost_usd": 0.0, "calls": 1, "usage": {"unpriced_calls": 1}}}, 0.0)
    assert usage == {"cost_usd": 0.0, "calls": 3, "unpriced_calls": 3}
    priced = runner.judge_usage({"claims": {"cost_usd": 0.01, "calls": 2, "priced": True}}, 0.01)
    assert priced == {"cost_usd": 0.01, "calls": 2, "unpriced_calls": 0}


async def test_results_give_the_question_actually_asked_and_the_template_apart(client, monkeypatch):
    """逐题结果的 question 是这一题实际问的（变量按这次评测的截至时间渲染好了，和作答时 Driver 收到的一字不差），模板原文
    另放 question_template（下一轮验收：以前 question 是带 {{ schedule.last_week_start }} 的原始模板）。没有变量的题两个一样。"""
    from datetime import datetime, timezone

    from app.db.models_eval import EvalCase
    from app.eval.cases import render, variable_values

    asked: list[str] = []
    real_init = answer_mod.Driver.__init__

    def init(self, question, **kw):
        asked.append(question)
        real_init(self, question, **kw)

    monkeypatch.setattr(answer_mod.Driver, "__init__", init)
    stub(monkeypatch)
    template = "{{ schedule.last_week_start }} 到 {{ schedule.last_week_end }} 那周的问题"
    suite_id = await es.make_suite("none", cases=[{"question": template, "expectations": [{"kind": "ask"}]},
                                                  {"question": "没有变量的问题", "expectations": [{"kind": "ask"}]}])
    as_of = datetime(2026, 10, 14, 2, 0, tzinfo=timezone.utc)
    [run] = await start(suite_id, as_of=as_of)
    await es.wait_eval(run.id)
    rows = (await client.get(f"/api/evals/evaluations/{run.id}/results")).json()["results"]
    expected = render(template, variable_values(as_of, "Asia/Shanghai", None))
    assert "{{" not in expected and rows[0]["question"] == expected, rows[0]["question"]
    assert rows[0]["question"] in asked, "和作答时实际问的一样"
    assert rows[0]["question_template"] == template
    assert rows[1]["question"] == rows[1]["question_template"] == "没有变量的问题"
    # 题目改过（指纹对不上）：现在的模板不一定是当时问的，没有作答运行可取时不编一个，question 为空、模板照给
    async with SessionLocal() as session:
        case = (await session.execute(select(EvalCase).where(EvalCase.suite_id == suite_id, EvalCase.seq == 1))).scalar_one()
        case.question, case.case_hash = "改过的 {{ schedule.today }}", "changed"
        await session.commit()
    rows = (await client.get(f"/api/evals/evaluations/{run.id}/results")).json()["results"]
    assert rows[0]["question"] is None and rows[0]["question_template"] == "改过的 {{ schedule.today }}"


async def test_eval_settings_group_and_spend_endpoint(client):
    got = (await client.get("/api/settings")).json()["eval"]
    assert got == {"max_parallel": 2, "per_case_timeout_s": 900, "per_run_max_usd": 2.0, "daily_max_usd": 5.0,
                   "keep_runs": 20, "judge": False}
    bad = await client.put("/api/settings", json={"values": {"eval": {"keep_runs": 0}}})
    assert bad.status_code == 422 and "保留" in bad.json()["detail"]
    bad = await client.put("/api/settings", json={"values": {"eval": {"per_run_max_usd": 9.0}}})
    assert bad.status_code == 422, "单次上限不能高于每日上限"
    ok = await client.put("/api/settings", json={"values": {"eval": {"daily_max_usd": 3.0, "judge": True},
                                                            "eval_spend": {"usd": 0}}})
    assert ok.status_code == 200 and ok.json()["eval"]["daily_max_usd"] == 3.0 and "eval_spend" not in ok.json()
    spend = (await client.get("/api/settings/eval/spend")).json()
    assert set(spend) == {"date", "usd", "calls", "unpriced_calls", "daily_max_usd"} and spend["daily_max_usd"] == 3.0
    # 题集只能把每日上限往下收
    from app.eval.suites import effective_budget

    assert effective_budget({"daily_max_usd": 9.0}, {"daily_max_usd": 3.0})["daily_max_usd"] == 3.0
    assert effective_budget({"daily_max_usd": 1.0}, {"daily_max_usd": 3.0})["daily_max_usd"] == 1.0
    await client.put("/api/settings", json={"values": {"eval": {}}})


async def test_a_workflow_case_runs_its_pinned_version_as_an_exploratory_run(client, monkeypatch, tmp_path):
    """工作流题：钉住那一版的图按输入模板跑一次探索运行（来源 eval），不发正式运行、不建图、不复核。"""
    from app.core.artifact_store import graph_hash
    from app.db.models import Workflow, WorkflowVersion
    from tests.fixtures import dual_check_shop as shop

    name, sid = await es.make_source(tmp_path)
    try:
        es.use_report(monkeypatch)

        def node(nid, ntype, **config):
            return {"id": nid, "type": ntype, "position": {"x": 0, "y": 0}, "data": {"label": nid, "config": config}}

        nodes = [node("start", "input", fields=[{"name": "week", "required": True}]),
                 node("q", "tool", tool=f"db_query__{name}", args={"sql": shop.RIGHT}),
                 node("write", "report", instructions="写周销售"),
                 node("out", "output", fields=[{"name": "报告", "value": "{{ nodes.write.text }}"}])]
        graph = {"nodes": nodes, "edges": [{"source": a["id"], "target": b["id"]} for a, b in zip(nodes, nodes[1:])]}
        async with SessionLocal() as session:
            wf = Workflow(name="虚构周销售", graph=graph, status="published", published_version=1)
            session.add(wf)
            await session.flush()
            session.add(WorkflowVersion(workflow_id=wf.id, version=1, graph=graph, graph_hash=graph_hash(graph),
                                        level="published"))
            await session.commit()
            wf_id = wf.id
        suite_id = await es.make_suite(sid, cases=[{
            "kind": "workflow", "workflow_id": wf_id, "workflow_version": 1,
            "input_template": {"week": "{{ schedule.last_week }}"},
            "gold": {"kind": "sql", "datasource_id": sid, "sql": shop.RIGHT},
            "expectations": [{"kind": "scalar", "column": "gmv"}]}])
        [run] = await start(suite_id)
        await es.wait_eval(run.id)
        [item] = await results(run.id)
        assert item.path == "ran" and item.verdict == "pass", (item.error, item.scores)
        async with SessionLocal() as session:
            answered = await session.get(Run, item.run_id)
        assert answered.run_class == "exploratory" and answered.origin_kind == "eval" and answered.workflow_id is None
        assert set(answered.input) == {"week"} and answered.input["week"].count("-W") == 1
        # 钉住当基线、取消
        pinned = await client.post(f"/api/evals/evaluations/{run.id}/pin", json={"pinned": True})
        assert pinned.status_code == 200 and pinned.json()["pinned"] is True
        assert (await client.post(f"/api/evals/evaluations/{run.id}/pin", json={"pinned": "yes"})).status_code == 422
        assert (await client.get("/api/evals/pairs/nope")).status_code == 404
    finally:
        await drop_source(sid)


async def test_a_review_asking_for_a_retry_rebuilds_on_the_same_graph(monkeypatch, tmp_path):
    """复核给了 retry_instruction（chat/turn 的规则判过可以重来）：带着这次运行的图、照那句话重来一次，两次运行都记在这一题
    名下、用量都算这一题的；第二次复核不再要重来就收尾。"""
    import app.api.copilot as copilot

    name, sid = await es.make_source(tmp_path)
    try:
        model = es.Copilot(name)
        es.use_copilot(monkeypatch, model)
        es.use_report(monkeypatch)
        reviews: list[int] = []
        real = copilot.review_of_run

        async def review(session, run, *, question, provider=None, model=None, always=False, attempt=0):
            reviews.append(attempt)
            out = await real(session, run, question=question, provider=provider, model=model, always=always,
                             attempt=attempt)
            if attempt == 0:
                out.update(verdict="annotated", severity="broken", retry=True,
                           retry_instruction="这个工作流上一次运行没有成功：步数用满了。改完继续回答原来的问题")
            return out

        monkeypatch.setattr(copilot, "review_of_run", review)
        suite_id = await es.make_suite(sid)
        [run] = await start(suite_id)
        await es.wait_eval(run.id)
        [item] = await results(run.id)
        assert reviews == [0, 1], "第一次复核要重来，第二次（attempt=1）不再要"
        assert item.path == "ran" and item.verdict == "pass", item.error
        retry_call = next(c for c in model.calls if "步数用满了" in str(c[-1][1]) or "步数用满了" in str(c[1][1]))
        assert "这是当前的工作流" in str(retry_call[1][1]), "带着原图交回去改"
        async with SessionLocal() as session:
            answers = (await session.execute(select(Run.id).where(Run.origin_kind == "eval"))).scalars().all()
            mine = [r for r in answers if (await session.get(Run, r)).origin.get("eval_run_id") == run.id
                    and (await session.get(Run, r)).origin.get("role") == "answer"]
        assert len(mine) == 2 and item.run_id in mine
    finally:
        await drop_source(sid)