mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 06:41:58 +00:00
ThearchyHelios
c97390511e
渲染（evidence._see_cuts）： - 括号是两边唯一的分隔、去掉之后数字字母会粘成一串的（「[[v:Q1.r0.销售额]]（[[see:Q1]]）2 月」渲染成「4,919.52 月」，片段 本身没错、月份不算裸数字，哪项核对都查不出来），括号换成一个空格；半角逗号夹着依据、去掉后一个之后同理（「212,5」→ 「212, 5」）。切口可以带替换串，连成一片的按并集换 - 只装着依据的括号去掉之后，整对括号当一个依据看外面：括号前悬着的停顿标点（「华南居首，（[[see:Q1]]）。」）、到句末只隔着 的空白、夹在两个分隔标点中间的后一个，和不带括号的写法收拾得一样 - 依据后面到句末标点只隔着空白的，空白一起去掉（「华南居首 [[see:Q1]] 。」不再剩「华南居首 。」） - 换行不再算句末：列表项末尾的「；」、软换行处连着两个分句的「，」是写作者自己的，依据挂在它后面正合规矩，照旧留着（上一轮 把它们删了） - 没有依据的文档渲染、内容哈希不变（随机两万份和基线逐份比过）；流式分块输出仍等于整篇渲染 - 依据写在「依据」「见」「取自」「来源」这类引出语后面时引出语悬空，渲染时去不掉（引出语在依据前面，流式渲染得攒住每个可能 是引出语的字）：测试里改成严格的 xfail 写明已知限制，不再把验收里的缺陷钉成预期；核查 agent 遇到这种写法记一条日志 核查 agent（verify_agent）： - 有没有给理由按渲染出来的正文判：只写了依据（[[see:Q1]]、「（[[see:Q1]]）」）显示出来是空的，占位词后面挂着依据的显示出来 只剩占位词，都算没给。以前按原文判，标记里的字也算，理由文档照样交回：审批卡画出一片空白，判「被推翻」的因为引用到了 Q1 被当成有出处、出具降档；封存之后的质询回答只剩系统说明却记成回答成功 - 「无」「暂无」「none」只在理由是另一个结果的说明时（双盲复算、反方审稿）算没给；结构里只有理由一项（审批质询）时理由就是 回答，「无。」是一个回答，不再报「模型没有给出回答内容」让提问人重问、再花一次钱 - 收尾超时的原话收成常量 SETTLE_TIMEOUT，调用方据此换成自己的说法 反方审稿：交结论那次超过时长上限的句子，理由写「反查未完成：已达上限（每句时长上限 N 秒）」（收紧过的说运行时长上限将至）， 和时间线上的 challenge_limit 一致，不再把核查 agent 的内部原话摆上审批卡。 测试：补节点级用例钉住收尾超时之后两个节点对外的说法（challenge_limit 的 limit、claims_skipped、cause，双盲复算的无法复算 原因和 dual_check_limit，收紧过的两种）；理由只有依据、占位词挂依据、「无。」作为回答的各层用例。每处改动都做过去掉实现、 测试变红的验证。
140 lines
7.9 KiB
Python
140 lines
7.9 KiB
Python
"""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
|