"""核查 agent 收尾超时之后，反方审稿、双盲复算对外怎么说（期 5 真实模型验收 M1 的复核）。 核查 agent 查完了、自己停下，交结果那次调用超过了时长上限：verify_agent 照实记 limit=timeout_s（以前模型自己停下时是 None，test_verify_rationale_rules 钉住了这一层）。两个节点据此按时长上限说，不再把核查 agent 的原话（「收尾时超过时长 上限，没能交结果」）摆上审批卡、时间线： - 反方审稿：这句计入 challenge_limit（limit=timeout_s，claims_skipped 算上它），不再记成 challenge_unreviewed；审稿记录 里这句的理由是「反查未完成：已达上限（每句时长上限 N 秒）」。时长按剩余时间收紧过的那组带 cause=run_deadline，理由 说运行时长上限将至 - 双盲复算：无法复算的原因是「盲算已达每项时长上限（N 秒）」，发 dual_check_limit；收紧过的说运行时长上限将至，不发 dual_check_limit（它叫人调高每项的上限，调高了只会更快撞线），时间线另起一条带 cause 的 dual_check_unchecked 交结果那次用 slow_final 拖住，收尾宽限调到 0.3 秒；收紧过的那组把 allot() 看到的钟拨到「给下游留出余量、扣掉收尾宽限 之后只剩 0.8 秒」（真的钟照走，核查 agent 按真的 0.8 秒到点）。模型一律是 MockChatModel 剧本，数据源是现造的 SQLite 文件，区域、金额都是虚构的。 """ from __future__ import annotations import pytest from app.core import artifact_store from app.engine import deadline, verify_agent from tests.fixtures.dual_check_shop import RIGHT, Blind from tests.test_challenge_node import ( # noqa: F401 - 夹具按名字引入 Models, engine_up, finish, graph, logs, output_of, shop, ) from tests.test_dual_check_node import clean_settings, dual_output, events, start # noqa: F401 from tests.test_dual_check_node import graph as dual_graph from tests.test_dual_check_node import logs as dual_logs from tests.test_dual_check_node import shop as dual_shop # noqa: F401 from tests.test_verify_rationale_rules import slow_final #: 收紧过的那组：allot() 算出的查库时长 ROOM = 0.8 #: 收紧过的时长到点时，反方审稿每句的理由、双盲复算的无法复算原因（默认 600 秒的一段） SQUEEZED_CLAIM = "反查未完成：已达上限（本次运行即将达到 600 秒的时长上限，反查时长已按剩余时间收紧）" SQUEEZED_BLIND = "本次运行即将达到 600 秒的时长上限，盲算时长已按剩余时间收紧，到时未能完成" @pytest.fixture def settles_slowly(monkeypatch): """交结果那次调用拖 5 秒，收尾宽限只给 0.3 秒：查完了的核查一律在收尾时超时。""" monkeypatch.setattr(verify_agent, "SETTLE_GRACE_S", 0.3) slow_final(monkeypatch) def squeeze(monkeypatch) -> None: """allot() 看到的「现在」：这一段的截止时刻扣掉给下游的余量（下游只有成果节点，一份）、收尾宽限，再往前 ROOM 秒。 每开始一项都一样，所以每一项的时长都收紧到 ROOM。""" def now() -> float: found = deadline.current() return found.deadline - deadline.reserve(found.seconds) - verify_agent.SETTLE_GRACE_S - ROOM - 0.05 monkeypatch.setattr(deadline, "_now", now) # -------------------------------------------------------------------------- # 反方审稿 # -------------------------------------------------------------------------- async def test_a_claim_whose_verdict_runs_out_of_time_counts_as_hitting_the_limit(monkeypatch, shop, settles_slowly): models = Models(monkeypatch, shop) row = await finish(graph(shop, timeout_s=0.5), run_class="exploratory") assert row.status == "succeeded", row.error assert [kind for kind, _, _ in models.calls].count("final") == 0, "交结论那次都没等到回复" [limit] = await logs(row.id, "challenge_limit") assert (limit["limit"], limit["claims_skipped"]) == ("timeout_s", 2) assert "cause" not in limit and "每句时长上限 0.5 秒" in limit["message"] and "2 句未能给出结论" in limit["message"] assert await logs(row.id, "challenge_unreviewed") == [], "触了上限的由 challenge_limit 说，不另记成未能反查" out = await output_of(row.id, "chal") assert out["counts"]["inconclusive"] == 2 and out["unreviewed"] == 2 record = artifact_store.load(out["challenge_artifact"]) for claim in record["claims"]: assert claim["verdict"] == "inconclusive" and claim["rationale_doc"] is None assert claim["rationale"] == "反查未完成：已达上限（每句时长上限 0.5 秒）" assert len(claim["queries"]) == 1, "查到的照记" assert "没能交结果" not in str(record) assert row.output["_issuance"]["tier"] == "formal", "没给出结论的句子不让出具降档" async def test_a_squeezed_claim_whose_verdict_runs_out_of_time_says_the_run_deadline(monkeypatch, shop, settles_slowly): Models(monkeypatch, shop) squeeze(monkeypatch) row = await finish(graph(shop), run_class="exploratory") assert row.status == "succeeded", row.error [limit] = await logs(row.id, "challenge_limit") assert (limit["limit"], limit["claims_skipped"], limit["claims"]) == ("timeout_s", 2, 2) assert limit["cause"] == deadline.RUN_DEADLINE and limit["run_seconds"] == 600 assert "每句时长上限" not in limit["message"], "收紧过的不叫人调高每句的上限" assert await logs(row.id, "challenge_unreviewed") == [] record = artifact_store.load((await output_of(row.id, "chal"))["challenge_artifact"]) assert [c["rationale"] for c in record["claims"]] == [SQUEEZED_CLAIM, SQUEEZED_CLAIM] # -------------------------------------------------------------------------- # 双盲复算 # -------------------------------------------------------------------------- async def test_a_blind_check_whose_result_runs_out_of_time_names_the_time_limit(monkeypatch, dual_shop, settles_slowly): name, _ = dual_shop blind = Blind(monkeypatch, name, {"销售额": (RIGHT, "gmv")}) run = await start(dual_graph(name, RIGHT), options={"verify": {"dual_check": {"timeout_s": 0.5}}}) assert run.status == "succeeded", run.error assert blind.runs == 0, "交结果那次没等到回复" evs = await events(run.id) assert [(d["limit"], d["metric"]) for d in dual_logs(evs, "dual_check_limit")] == [("timeout_s", "gmv")] out = dual_output(run, evs) assert out["status"] == "unchecked" and out["metrics"][0]["status"] == "unchecked" assert out["metrics"][0]["reason"] == "盲算已达每项时长上限（0.5 秒）" assert out["issuance_gaps"] == ["指标「销售额」无法复算：盲算已达每项时长上限（0.5 秒）"] [said] = dual_logs(evs, "dual_check_unchecked") assert "cause" not in said and said["metrics"][0]["reason"] == "盲算已达每项时长上限（0.5 秒）" assert "没能交结果" not in str(out) async def test_a_squeezed_blind_check_whose_result_runs_out_of_time_says_the_run_deadline(monkeypatch, dual_shop, settles_slowly): name, _ = dual_shop Blind(monkeypatch, name, {"销售额": (RIGHT, "gmv")}) squeeze(monkeypatch) run = await start(dual_graph(name, RIGHT)) assert run.status == "succeeded", run.error evs = await events(run.id) assert dual_logs(evs, "dual_check_limit") == [], "收紧过的不叫人调高每项的上限" out = dual_output(run, evs) assert out["metrics"][0]["status"] == "unchecked" and out["metrics"][0]["reason"] == SQUEEZED_BLIND [said] = dual_logs(evs, "dual_check_unchecked") assert said["cause"] == deadline.RUN_DEADLINE and said["run_seconds"] == 600