"""WP-4 AI 兜底起草(recipe_ai)的单元测试(P2-SPEC 9.6 的 WP-4 部分)。 不调用任何真实模型:ai_draft 注入假模型(async ainvoke(messages) -> AIMessage);可用性的用例用独立的内存库建接入, resolve_draft_model 走到 build_chat_model 时 monkeypatch 成假模型。夹具一律合成、假名。 """ from __future__ import annotations import asyncio import datetime as dt import json from typing import Any import pytest from langchain_core.messages import AIMessage, HumanMessage, SystemMessage from openpyxl import Workbook from app.data import recipe_ai as A from app.data import recipe_suggest as S from app.data.recipe_types import Draft, DraftFacts, Extraction, Fact, Grid, GridCell, Problem, Recipe, RecipeProblem from tests.test_recipe_suggest import FLOW, build, flow_book, ok_validate # ========================================================================== # 假模型与桩 # ========================================================================== def reply(content: Any, *, inp: int = 100, out: int = 50, meta: dict[str, Any] | None = None) -> AIMessage: return AIMessage(content=content, usage_metadata={"input_tokens": inp, "output_tokens": out, "total_tokens": inp + out}, response_metadata=meta or {}) class FakeModel: """按预设顺序返回;元素是异常就抛,是协程函数就 await 它。记下每次收到的消息列表。""" def __init__(self, replies: list[Any], name: str = "fake-model") -> None: self.replies = list(replies) self.calls: list[list[Any]] = [] self.model_name = name async def ainvoke(self, messages: list[Any]) -> Any: self.calls.append(list(messages)) r = self.replies.pop(0) if isinstance(r, BaseException): raise r if callable(r): return await r() return r def flow_json(**patch: Any) -> str: data = json.loads(json.dumps(FLOW)) data.update(patch) return json.dumps(data, ensure_ascii=False) def claims_validate(calls: list | None = None): """静态校验桩:pydantic、AI 的 note 必须为空(ai_note_forbidden)、每条事实恰好被认领一次(fact_unclaimed)。""" def _v(data: dict[str, Any], facts: DraftFacts | None, origin: str): if calls is not None: calls.append((data, facts, origin)) try: r = Recipe.model_validate(data) except Exception as e: # noqa: BLE001 return None, [RecipeProblem("/", "schema", str(e)[:200])] probs = [] if origin == "ai": for i, t in enumerate(r.tables): if t.note: probs.append(RecipeProblem(f"/tables/{i}/note", "ai_note_forbidden", f"表「{t.name}」:AI 起草的配方说明必须为空")) claimed = [x.claims for x in r.relations if x.claims] for f in (facts.facts if facts else []): if claimed.count(f.id) != 1: probs.append(RecipeProblem("/relations", "fact_unclaimed", f"系统发现的关系「{f.text}」没有被认领")) return (r if not probs else r), probs return _v FACTS = DraftFacts(facts=[ Fact("F1", "sum_eq", "客流汇总", "第 5 行 = 第 6 行 + 第 7 行(31 列中 31 列成立)", {"segment": "日客流", "total": "全日客流(人次)", "parts": ["分区甲(人次)", "分区乙(人次)"], "rows": [5, 6, 7]}), Fact("F2", "not_equal_sum", "客流汇总", "日间、夜间各行之和 与 第 5 行:31 列中 0 列相等", {"a": ["日间", "夜间"], "b_segment": "日客流", "b": "全日客流(人次)", "equal": 0, "checked": 31}), ], candidates={"日间时段客流(人次)": ["日间", "日间时段客流"], "夜间时段客流(人次)": ["夜间", "夜间时段客流"]}) @pytest.fixture(scope="module") def flow(): wb, cache, nums = flow_book() sc, grids = build(wb, cache) return sc, grids, nums async def run(flow, model, *, validate=None, dry_run=None, rules: Draft | None = None, facts=FACTS, **kw): sc, grids, _ = flow return await A.ai_draft(None, grids=grids, scan=sc, facts=facts, rules_draft=rules, filename="x.xlsx", # type: ignore[arg-type] validate=validate or claims_validate(), dry_run=dry_run, model=model, **kw) # ========================================================================== # 常量 # ========================================================================== def test_constants_match_copilot(): from app.api import copilot assert A.COPILOT_SETTING_KEY == copilot.COPILOT_SETTING_KEY assert A.AI_MAX_TOKENS == copilot.COPILOT_MAX_TOKENS assert A.AI_FALLBACK_MAX_TOKENS == copilot._FALLBACK_MAX_TOKENS assert A._BUDGET_WORDS == copilot._BUDGET_WORDS assert A.AI_MAX_ATTEMPTS == 3 # ========================================================================== # 可用性(独立的内存库) # ========================================================================== class _MemDB: """独立的内存库:不碰测试会话共用的库,也不受别的测试建的接入影响。""" async def __aenter__(self): from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine from sqlalchemy.pool import StaticPool import app.db.models # noqa: F401 - 建表前要先登记全部模型 from app.db.base import Base self.engine = create_async_engine("sqlite+aiosqlite:///:memory:", poolclass=StaticPool) async with self.engine.begin() as conn: await conn.run_sync(Base.metadata.create_all) self.session = async_sessionmaker(self.engine, expire_on_commit=False, class_=AsyncSession)() return self.session async def __aexit__(self, *exc): await self.session.close() await self.engine.dispose() @pytest.fixture async def mem_session(): async with _MemDB() as session: yield session async def _setup(session, providers: list[dict[str, Any]], saved: dict[str, str] | None = None) -> None: from app.core.crypto import encrypt from app.db.models import Provider, Setting for p in providers: p = dict(p) if p.get("api_key"): p["api_key"] = encrypt(p["api_key"]) session.add(Provider(**p)) if saved is not None: session.add(Setting(key=A.COPILOT_SETTING_KEY, value=saved)) await session.commit() @pytest.fixture def no_env_keys(monkeypatch): for k in ("ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN", "OPENAI_API_KEY", "ANTHROPIC_BASE_URL"): monkeypatch.delenv(k, raising=False) CASES = { "none": ([], None), "mock_only": ([{"name": "演示", "kind": "mock", "models": [{"id": "mock-fast"}], "default_model": "mock-fast"}], None), "name_disabled": ([{"name": "甲接入", "kind": "anthropic", "api_key": "k", "models": [{"id": "claude-x"}], "default_model": "claude-x", "enabled": False}], {"provider": "甲接入", "model": ""}), "model_disabled": ([{"name": "乙接入", "kind": "anthropic", "api_key": "k", "models": [{"id": "claude-y"}], "default_model": "claude-y", "enabled": False}], {"provider": "", "model": "claude-y"}), "missing_key": ([{"name": "丙接入", "kind": "anthropic", "api_key": None, "models": [{"id": "claude-z"}], "default_model": "claude-z"}], None), } @pytest.mark.asyncio async def test_unavailable_cases(no_env_keys, monkeypatch, flow): """没有接入、只有演示接入、存的接入名已停用、存的模型名所属接入已停用、缺 key:都不可用,原因各不相同, 假模型一次都没被调。""" from app.providers import factory built: list[FakeModel] = [] real = factory.build_chat_model def spy(provider, spec): real(provider, spec) # 缺 key 时这里就抛 ProviderNotConfigured fake = FakeModel([]) built.append(fake) return fake monkeypatch.setattr(factory, "build_chat_model", spy) sc, grids, _ = flow reasons: dict[str, str] = {} for case, (providers, saved) in CASES.items(): async with _MemDB() as session: await _setup(session, providers, saved) avail = await A.ai_availability(session) assert not avail.available and avail.reason, case with pytest.raises(A.AiUnavailable) as e: await A.ai_draft(session, grids=grids, scan=sc, facts=FACTS, rules_draft=None, filename="x.xlsx", validate=claims_validate(), dry_run=None) assert str(e.value) == avail.reason reasons[case] = avail.reason assert all(not m.calls for m in built) assert len(set(reasons.values())) == len(CASES), reasons assert "已停用" in reasons["name_disabled"] and "甲接入" in reasons["name_disabled"] assert "已停用" in reasons["model_disabled"] and "乙接入" in reasons["model_disabled"] assert "API Key" in reasons["missing_key"] assert "演示" in reasons["mock_only"] assert "未配置模型接入" in reasons["none"] @pytest.mark.asyncio async def test_available_builds_model_without_request(mem_session, no_env_keys, monkeypatch): from app.providers import factory await _setup(mem_session, [{"name": "丁接入", "kind": "anthropic", "api_key": "sk-test", "models": [{"id": "claude-w"}], "default_model": "claude-w"}], {"provider": "丁接入", "model": "claude-w"}) specs: list[Any] = [] def fake_build(provider, spec): specs.append(spec) return FakeModel([], name=spec.model) monkeypatch.setattr(factory, "build_chat_model", fake_build) model, model_id, avail = await A.resolve_draft_model(mem_session) assert avail.available and avail.model == "claude-w" and avail.provider == "丁接入" and model_id == "claude-w" assert [s.max_tokens for s in specs] == [A.AI_MAX_TOKENS, A.AI_FALLBACK_MAX_TOKENS] assert all(s.thinking == "off" and s.temperature == 0 for s in specs) assert not model.model.calls # 只建对象,不发请求 @pytest.mark.asyncio async def test_budget_rejection_falls_back_to_small_budget(mem_session, no_env_keys, monkeypatch, flow): """服务方以 400「max_tokens too large」拒绝 → 退回 8192 重来,之后的调用都用 8192。""" from app.providers import factory await _setup(mem_session, [{"name": "戊接入", "kind": "anthropic", "api_key": "sk-test", "models": [{"id": "claude-v"}], "default_model": "claude-v"}]) class BudgetError(Exception): status_code = 400 models: dict[int, FakeModel] = {} def fake_build(provider, spec): if spec.max_tokens == A.AI_MAX_TOKENS: m = FakeModel([BudgetError("max_tokens: 32768 > 8192, which is the maximum allowed")]) else: m = FakeModel([reply("不是 JSON"), reply(flow_json())]) models[spec.max_tokens] = m return m monkeypatch.setattr(factory, "build_chat_model", fake_build) sc, grids, _ = flow res = await A.ai_draft(mem_session, grids=grids, scan=sc, facts=FACTS, rules_draft=None, filename="x.xlsx", validate=claims_validate(), dry_run=None) assert res.draft is not None and res.attempts == 2 assert len(models[A.AI_MAX_TOKENS].calls) == 1 assert len(models[A.AI_FALLBACK_MAX_TOKENS].calls) == 2 # ========================================================================== # 压缩表示不外泄 # ========================================================================== def _leaky_book(): """一张交叉表 + 一张列表:特征数、像数的文字、带常量的公式、姓名电话、隐藏行列、遮罩列。""" wb = Workbook() ws = wb.active ws.title = "客流汇总" ws["B2"] = "统计时间范围:2026年8月1日至2026年8月8日" for i in range(8): ws.cell(4, 3 + i, f"8月{i + 1}日") ws["B5"] = "全日客流(人次)" ws["B6"] = "分区甲(人次)" ws["B7"] = "分区乙(人次)" ws["B8"] = "秘密指标(人次)" for i in range(8): ws.cell(5, 3 + i, 987654 + i) ws.cell(6, 3 + i, 123457 + i) ws.cell(7, 3 + i, 5555.75) ws.cell(8, 3 + i, 424242) ws["C9"] = "1,234" ws["B9"] = "补录(人次)" ws["D9"] = "=3000+2500" ws["E9"] = "=C5*1.0675" ws["F9"] = '="内部"&B5' ws["G9"] = "='附表'!C5+1" ws.row_dimensions[8].hidden = True ws["L5"] = "隐藏列里的话" ws.column_dimensions["L"].hidden = True lst = wb.create_sheet("名单") lst.append(["姓名", "电话", "金额", "工号"]) people = [("张三", "138-1234-5678", 10, "X-保密-1"), ("李四", "139-8765-4321", 20, "X-保密-2"), ("王五", "137-0000-1111", 30, "X-保密-3"), ("赵六", "136-2222-3333", 40, "X-保密-4"), ("钱七", "135-4444-5555", 50, "X-保密-5")] for p in people: lst.append(list(p)) cache = {"客流汇总": {"D9": 5500, "E9": 1054320.6, "F9": "内部987654", "G9": 99}} return wb, cache def test_compress_does_not_leak_values(): wb, cache = _leaky_book() sc, grids = build(wb, cache) facts = S.detect_facts(grids, sc) c = A.compress(grids, sc, facts, mask_headers={"工号"}) text = c.text for g in grids.values(): for cell in g.cells.values(): # 特征数逐个断言(小数字会和坐标、统计期里的年份撞上,不在这里查) if isinstance(cell.value, (int, float)) and not isinstance(cell.value, bool) and abs(cell.value) >= 5000: assert str(cell.value) not in text, cell.value for secret in ("987654", "123457", "5555.75", "424242", "1,234", "3000", "2500", "1.0675", "内部", "1054320", "张三", "李四", "138-1234-5678", "139-8765-4321", "秘密指标", "隐藏列里的话", "保密"): assert secret not in text, secret assert "<文本数字>" in text assert "=<数>+<数>" in text and "=C5*<数>" in text and "=<文本>&B5" in text assert "G9 ='附表'!<格>+<数>" in text # 跨工作表引用只留表名 assert "隐藏行:第 8 行(文字 1 格、数字 8 格)" in text assert "隐藏列:L 列(文字 1 格、数字 0 格)" in text assert "A 列「姓名」是数据列(只给画像):5 格,不同取值 5 个" in text assert "B 列「电话」是数据列" in text assert "D 列「工号」是遮罩列(只给画像)" in text assert "C9「<文本数字>」" in text # 结构文字照发 for shown in ("全日客流(人次)", "统计时间范围:2026年8月1日至2026年8月8日", "序列「{}月{}日」8 格", "姓名", "金额"): assert shown in text, shown # 同一份网格两次压缩,sha256 相同 again = A.compress(grids, sc, facts, mask_headers={"工号"}) assert again.sha256 == c.sha256 and again.text == text and c.chars == len(text) def test_sanitize_formula(): assert A.sanitize_formula("=3000+2500") == "=<数>+<数>" assert A.sanitize_formula("=C5*1.0675") == "=C5*<数>" assert A.sanitize_formula('=IF(B5>50000,"高","低")') == "=IF(B5><数>,<文本>,<文本>)" assert A.sanitize_formula("='内部'!C5+SUM(C1:C4)") == "='内部'!<格>+SUM(C1:C4)" def test_compress_flow_shape(flow): sc, grids, nums = flow c = A.compress(grids, sc, S.detect_facts(grids, sc)) assert "## 工作表「客流汇总」 已用区域 B2:AG30,非空格 771" in c.text assert "C28:AG28 =SUM(C22:C25) 逐列平移(31 格,均有保存值)" in c.text assert "C10:AG10「·」×31" in c.text assert "C5:AG7 整数" in c.text assert "F1(等式" in c.text and "分段 id「客流汇总_按日」" in c.text assert "「日间时段客流(人次)」:日间、日间时段客流" in c.text assert not any(str(v) in c.text for v in nums["jia"] + nums["yi"]) # 压缩表示只有规格里的四个字段(接口层照此存库) assert set(c.__dataclass_fields__) == {"text", "chars", "truncated", "sha256"} def test_compress_folds_large_sheet(): wb = Workbook() ws = wb.active ws.title = "大表" ws.append(["编号", "金额"]) for i in range(400): ws.append([f"N{i:04d}", 1000 + i]) sc, grids = build(wb) c = A.compress(grids, sc, DraftFacts()) assert c.truncated assert "第 201–381 行(共 181 行)折叠为列画像:" in c.text # 前 200 行、后 20 行照常 assert "N0300" not in c.text @pytest.mark.asyncio async def test_too_large_raises_before_calling_model(flow, monkeypatch): monkeypatch.setattr(A, "COMPRESS_MAX_CHARS_PER_SHEET", 100) model = FakeModel([reply(flow_json())]) with pytest.raises(A.AiTooLarge): await run(flow, model) assert model.calls == [] monkeypatch.setattr(A, "COMPRESS_MAX_CHARS_PER_SHEET", 12_000) monkeypatch.setattr(A, "COMPRESS_MAX_CHARS", 100) with pytest.raises(A.AiTooLarge): await run(flow, model) assert model.calls == [] # ========================================================================== # 起草、修订、失败 # ========================================================================== @pytest.mark.asyncio async def test_invalid_json_then_valid(flow): model = FakeModel([reply("我先想一想……"), reply(flow_json(), inp=200, out=80)]) res = await run(flow, model) assert res.draft is not None and res.draft.origin == "ai" and res.error is None assert res.attempts == 2 and len(res.usage) == 2 assert [u.ok for u in res.usage] == [True, True] assert res.usage[1].input_tokens == 200 and res.usage[1].output_tokens == 80 assert all(u.model == "fake-model" and u.attempt == i + 1 for i, u in enumerate(res.usage)) assert res.draft.cards[0].id == "ai_origin" and res.draft.cards[-1].id == "mode" assert res.draft.complete assert Recipe.model_validate(res.draft.recipe).model_dump(mode="json", exclude_defaults=True) == \ Recipe.model_validate(FLOW).model_dump(mode="json", exclude_defaults=True) assert res.compressed_chars > 0 and len(res.compressed_sha256) == 64 # 修订是同一轮对话:system + 首次 prompt + 上一轮回复(原样)+ repair second = model.calls[1] assert len(second) == 4 assert isinstance(second[0], SystemMessage) and isinstance(second[1], HumanMessage) assert second[0].content == A.RECIPE_DRAFT_SYSTEM assert second[1].content == model.calls[0][1].content assert isinstance(second[2], AIMessage) and second[2].content == "我先想一想……" assert isinstance(second[3], HumanMessage) and "不是合法的 JSON" in second[3].content @pytest.mark.asyncio async def test_block_list_content_is_extracted(flow): content = [{"type": "thinking", "thinking": "先写交叉表 {\"x\": 1}", "signature": "s"}, {"type": "text", "text": "```json\n" + flow_json() + "\n```"}] model = FakeModel([reply(content)]) res = await run(flow, model) assert res.draft is not None and res.attempts == 1 @pytest.mark.asyncio async def test_truncated_reply_is_not_repaired(flow): model = FakeModel([reply(flow_json()[:200], meta={"stop_reason": "max_tokens"}), reply(flow_json())]) res = await run(flow, model) assert res.draft is not None and res.attempts == 2 assert [u.ok for u in res.usage] == [False, True] assert len(model.calls[1]) == 2 # 没有 repair:同样的消息列表重来 assert [m.content for m in model.calls[1]] == [m.content for m in model.calls[0]] # 三次都截断:error 是截断的原因,不是「不是 JSON」 cut = reply("{", meta={"finish_reason": "length"}) res2 = await run(flow, FakeModel([cut, cut, cut])) assert res2.draft is None and "输出被截断" in (res2.error or "") assert [u.ok for u in res2.usage] == [False, False, False] @pytest.mark.asyncio async def test_timeouts_fail_without_raising(flow, monkeypatch): monkeypatch.setattr(A, "AI_TIMEOUT_S", 0.05) async def slow(): await asyncio.sleep(1) return reply(flow_json()) model = FakeModel([slow, slow, slow]) res = await run(flow, model) assert res.draft is None and res.attempts == 3 assert "调用超时" in (res.error or "") assert len(res.usage) == 3 and not any(u.ok for u in res.usage) assert all(len(c) == 2 for c in model.calls) @pytest.mark.asyncio async def test_call_exception_then_success(flow): model = FakeModel([RuntimeError("connection reset"), reply(flow_json())]) res = await run(flow, model) assert res.draft is not None and res.attempts == 2 assert [u.ok for u in res.usage] == [False, True] res2 = await run(flow, FakeModel([RuntimeError("boom")] * 3)) assert res2.draft is None and (res2.error or "").startswith("调用失败:") @pytest.mark.asyncio async def test_note_and_unclaimed_fact_trigger_repair(flow): with_note = json.loads(flow_json()) with_note["tables"][0]["note"] = "这是AI写的说明" unclaimed = json.loads(flow_json()) unclaimed["relations"] = unclaimed["relations"][:1] calls: list = [] model = FakeModel([reply(json.dumps(with_note, ensure_ascii=False)), reply(json.dumps(unclaimed, ensure_ascii=False)), reply(flow_json())]) res = await run(flow, model, validate=claims_validate(calls)) assert res.draft is not None and res.attempts == 3 assert [c[2] for c in calls] == ["ai", "ai", "ai"] assert "/tables/0/note:" in model.calls[1][-1].content and "ai_note_forbidden" not in model.calls[1][-1].content assert "没有被认领" in model.calls[2][-1].content @pytest.mark.asyncio async def test_dry_run_problems_are_fed_back_with_model_message_only(flow): sc, grids, _ = flow grids = dict(grids) g = grids["客流汇总"] cells = dict(g.cells) cells[(31, 3)] = GridCell(987654) cells[(31, 4)] = GridCell("约7777") grids["客流汇总"] = Grid(g.sheet, g.bounds, cells, g.merges, g.hidden_rows, g.hidden_cols) n = {"k": 0} def dry(recipe): n["k"] += 1 if n["k"] == 1: return Extraction(ok=False, problems=[ Problem("outside_number", "structure", "第 31 行 C31 的 987654 在数据区外", ["客流汇总!C31"]), Problem("value_not_number", "structure", "D31 的「约7777」不是数字", ["客流汇总!D31"], model_message="D31 不是数字"), Problem("outside_digits", "confirm", "区域外有含数字的文字"), ]) return Extraction(ok=True) model = FakeModel([reply(flow_json()), reply(flow_json())]) res = await A.ai_draft(None, grids=grids, scan=sc, facts=FACTS, rules_draft=None, filename="x.xlsx", # type: ignore[arg-type] validate=claims_validate(), dry_run=dry, model=model) assert res.draft is not None and res.draft.complete and res.attempts == 2 fed = model.calls[1][-1].content assert "outside_number(客流汇总!C31)" in fed assert "value_not_number:D31 不是数字(客流汇总!D31)" in fed assert "987654" not in fed and "7777" not in fed and "outside_digits" not in fed @pytest.mark.asyncio async def test_cpu_work_runs_off_the_event_loop_and_async_dry_run_is_awaited(flow, monkeypatch): """ai_draft 跑在事件循环上:静态校验、干跑、卡片问题都不能在循环线程上同步执行(一次干跑可能要几秒, 会堵住同进程的所有请求)。协程函数的 dry_run 直接 await(接口层那份自己占解析名额、进线程池); 普通函数的 dry_run 放进线程池。""" import threading loop_thread = threading.get_ident() where: dict[str, list[int]] = {"validate": [], "questions_for": [], "sync_dry": []} inner = claims_validate() def validate(data, facts, origin): where["validate"].append(threading.get_ident()) return inner(data, facts, origin) real_questions = S.questions_for def questions_for(*a, **kw): where["questions_for"].append(threading.get_ident()) return real_questions(*a, **kw) monkeypatch.setattr(S, "questions_for", questions_for) awaited: list[str] = [] async def dry_async(recipe): awaited.append(type(recipe).__name__) await asyncio.sleep(0) return Extraction(ok=True) res = await run(flow, FakeModel([reply(flow_json())]), validate=validate, dry_run=dry_async) assert res.draft is not None and res.draft.complete and awaited == ["Recipe"] assert where["validate"] and all(t != loop_thread for t in where["validate"]) assert where["questions_for"] and all(t != loop_thread for t in where["questions_for"]) def dry_sync(recipe): where["sync_dry"].append(threading.get_ident()) return Extraction(ok=True) res = await run(flow, FakeModel([reply(flow_json())]), dry_run=dry_sync) assert res.draft is not None and where["sync_dry"] and all(t != loop_thread for t in where["sync_dry"]) # 协程版干跑出错:同普通版,变成一条起草失败原因,不让起草整个失败 async def boom(recipe): raise RuntimeError("内部细节") res = await run(flow, FakeModel([reply(flow_json())] * 3), dry_run=boom) assert res.draft is not None and not res.draft.complete assert res.draft.failures == [S._FAIL_TEXT["dry_run"]] and "内部细节" not in json.dumps(res.draft.failures) @pytest.mark.asyncio async def test_last_dry_run_problems_still_return_draft(flow): """次数用完时,静态校验通过但干跑仍有结构问题的配方照样返回(问题随草稿交给用户)。""" prob = Problem("cell_unclaimed", "structure", "第 31 行没有去处", ["客流汇总!B31"], model_message="第 31 行没有去处") model = FakeModel([reply(flow_json())] * 3) res = await run(flow, model, dry_run=lambda r: Extraction(ok=False, problems=[prob])) assert res.attempts == 3 and res.draft is not None and not res.draft.complete assert res.draft.failures == ["第 31 行没有去处"] and res.error is None # 后面一次退化成不合法的 JSON:仍返回之前那份静态校验通过的配方 model2 = FakeModel([reply(flow_json()), reply("坏了"), reply("还是坏的")]) res2 = await run(flow, model2, dry_run=lambda r: Extraction(ok=False, problems=[prob])) assert res2.draft is not None and res2.attempts == 3 @pytest.mark.asyncio async def test_never_valid_returns_none_with_usage(flow): model = FakeModel([reply("x"), reply("{\"recipe_format\": 1}"), reply("y")]) res = await run(flow, model) assert res.draft is None and res.error == "AI 未能起草出合法的配方" assert len(res.usage) == 3 and res.attempts == 3 @pytest.mark.asyncio async def test_prompt_uses_failures_for_model_not_failures(flow): rules = Draft(recipe=None, complete=False, origin="rules", failures=["C12 的「约987654」不是数字"], failures_for_model=["工作表「客流汇总」:C12 不是数字"]) model = FakeModel([reply(flow_json())]) await run(flow, model, rules=rules) prompt = model.calls[0][1].content assert "987654" not in prompt assert "工作表「客流汇总」:C12 不是数字" in prompt assert "# 配方 JSON Schema" in prompt and '"recipe_format"' in prompt assert "人次、人、元" in prompt # 半成品配方也带上(紧凑形式) sc, grids, _ = flow rd = S.draft(sc, grids, "x.xlsx", validate=ok_validate()) model2 = FakeModel([reply(flow_json())]) await run(flow, model2, rules=rd) assert "半成品配方" in model2.calls[0][1].content and "客流汇总_按日" in model2.calls[0][1].content @pytest.mark.asyncio async def test_given_compressed_is_what_is_sent(flow): sc, grids, _ = flow c = A.compress(grids, sc, FACTS) model = FakeModel([reply(flow_json())]) res = await run(flow, model, compressed=c) assert c.text in model.calls[0][1].content assert res.compressed_sha256 == c.sha256 @pytest.mark.asyncio async def test_rules_draft_section_hides_hidden_and_numeric_labels(): wb = Workbook() ws = wb.active ws.title = "表" ws["B2"] = "2026年8月1日至2026年8月8日" for i in range(8): ws.cell(4, 3 + i, f"8月{i + 1}日") for r, lab in ((5, "甲(人次)"), (6, "藏起来的指标"), (7, "2025")): ws.cell(r, 2, lab) for i in range(8): ws.cell(r, 3 + i, 10 + r + i) ws.row_dimensions[6].hidden = True sc, grids = build(wb) rd = S.draft(sc, grids, "x.xlsx", validate=ok_validate()) assert rd.recipe is not None and "藏起来的指标" in json.dumps(rd.recipe, ensure_ascii=False) model = FakeModel([reply(flow_json())]) await A.ai_draft(None, grids=grids, scan=sc, facts=rd.facts, rules_draft=rd, filename="x.xlsx", # type: ignore[arg-type] validate=ok_validate(), dry_run=None, model=model) prompt = model.calls[0][1].content assert "藏起来的指标" not in prompt assert '"2025"' not in prompt # ========================================================================== # 评审修复:数据行不当表头外发、压缩表示只靠 text、日期轴不发值 # ========================================================================== @pytest.mark.asyncio async def test_list_row_with_blank_number_does_not_leak(): """名单第 4 行的金额空着:这一行不能被当成表头,姓名、电话、备注都不出现在压缩表示和提示词里。""" from tests.test_recipe_suggest import _roster for blank in (False, True): sc, grids = build(_roster(blank_before_wang=blank)) rd = S.draft(sc, grids, "x.xlsx", validate=ok_validate()) c = A.compress(grids, sc, rd.facts) model = FakeModel([reply(flow_json())]) await A.ai_draft(None, grids=grids, scan=sc, facts=rd.facts, rules_draft=rd, filename="x.xlsx", # type: ignore[arg-type] validate=ok_validate(), dry_run=None, model=model) prompt = model.calls[0][1].content for secret in ("张三", "王五", "钱七", "138-1234-5678", "137-2222-3333", "137_2222_3333", "转介绍", "老客户"): assert secret not in c.text, (blank, secret) assert secret not in prompt, (blank, secret) assert "A 列「姓名」是数据列(只给画像):5 格" in c.text assert "B 列「电话」是数据列" in c.text and "D 列「备注」是数据列" in c.text @pytest.mark.asyncio async def test_rules_section_scrubs_column_names_made_from_hidden_text(): """万一起草把数据行认成了表头:配方里的列名是 to_sql_name 变过的写法,同样要遮掉。""" from tests.test_recipe_suggest import _roster sc, grids = build(_roster()) rd = S.draft(sc, grids, "x.xlsx", validate=ok_validate()) bad = json.loads(json.dumps(rd.recipe, ensure_ascii=False)) cols = bad["sheets"][0]["blocks"][0]["columns"] cols[0].update(header="王五", name="王五") cols[1].update(header="137-2222-3333", name="c_137_2222_3333") bad["tables"][0]["grain"] = ["王五"] leaky = Draft(recipe=bad, complete=False, origin="rules", failures=[], failures_for_model=[]) model = FakeModel([reply(flow_json())]) await A.ai_draft(None, grids=grids, scan=sc, facts=rd.facts, rules_draft=leaky, filename="x.xlsx", # type: ignore[arg-type] validate=ok_validate(), dry_run=None, model=model) prompt = model.calls[0][1].content assert "半成品配方" in prompt assert "王五" not in prompt and "137" not in prompt @pytest.mark.asyncio async def test_compressed_with_only_the_four_fields_is_sent_once(flow): """接口层把压缩表示按四个字段存库再取回:提示词的前三段就是 text,不重复、不重算。""" sc, grids, _ = flow facts = S.detect_facts(grids, sc) c = A.compress(grids, sc, facts) stored = A.Compressed(text=c.text, chars=c.chars, truncated=c.truncated, sha256=c.sha256) m1, m2 = FakeModel([reply(flow_json())]), FakeModel([reply(flow_json())]) await run(flow, m1, facts=facts, compressed=stored) await run(flow, m2, facts=facts) prompt = m1.calls[0][1].content for head in ("# 工作表压缩表示", "# 系统发现", "# 分段标题候选词", "# 单位词表", "# 规则起草的半成品与失败原因", "# 配方 JSON Schema"): assert prompt.count(head) == 1, head assert prompt.startswith(c.text + "\n\n# 单位词表\n") assert prompt == m2.calls[0][1].content # 与 RECIPE_DRAFT_PROMPT 的写法逐字相同(前三段取压缩表示) assert A.RECIPE_DRAFT_PROMPT == A.RECIPE_SECTIONS_PROMPT + "\n\n" + A.RECIPE_REQUEST_PROMPT def test_date_axis_cells_send_type_only(): """日期轴上是日期格(不是「8月1日」这样的文字)时只发类型签名:表格里没有统计期格时,轴上的年月日就是数据。""" wb = Workbook() ws = wb.active ws.title = "日报" for i in range(8): ws.cell(4, 3 + i, dt.datetime(2026, 8, 1 + i)) for r, lab in ((5, "全日客流(人次)"), (6, "分区甲(人次)")): ws.cell(r, 2, lab) for i in range(8): ws.cell(r, 3 + i, 1000 + r * 10 + i) sc, grids = build(wb) assert S.layout_hints(grids["日报"]).axis_row == 4 c = A.compress(grids, sc, S.detect_facts(grids, sc)) assert "C4:J4 日期" in c.text for leak in ("2026-08", "2026年", "8月1日", "08-01"): assert leak not in c.text, leak assert "B5「全日客流(人次)」" in c.text # ========================================================================== # 整体审查(AU / SE)修复 # ========================================================================== def _hidden_unit_book() -> Workbook: """交叉表:第 7 行隐藏,标签带单位后缀(「绝密分区(人次)」),并且是 F1(5 = 6 + 7)的成员。""" wb = Workbook() ws = wb.active ws.title = "表" ws["B2"] = "统计时间范围:2026年8月1日至2026年8月8日" for i in range(8): ws.cell(4, 3 + i, f"8月{i + 1}日") labels = ((5, "全日客流(人次)"), (6, "分区甲(人次)"), (7, "绝密分区(人次)")) for r, lab in labels: ws.cell(r, 2, lab) for i in range(8): a, b = 100 + i, 30 + i ws.cell(6, 3 + i, a) ws.cell(7, 3 + i, b) ws.cell(5, 3 + i, a + b) ws.row_dimensions[7].hidden = True return wb @pytest.mark.asyncio async def test_hidden_label_with_unit_suffix_never_leaves_in_any_call(): """SE-1、SE-2:隐藏行的标签(含去掉单位后缀的「绝密分区」)不出现在首轮全文、也不出现在任何一轮修订消息里。""" sc, grids = build(_hidden_unit_book()) rd = S.draft(sc, grids, "x.xlsx", validate=ok_validate()) facts = S.detect_facts(grids, sc) assert rd.recipe is not None and "绝密分区" in json.dumps(rd.recipe, ensure_ascii=False) assert any("绝密分区" in json.dumps(f.detail, ensure_ascii=False) for f in facts.facts) c = A.compress(grids, sc, facts) first = A.first_call_text(c, grids=grids, scan=sc, rules_draft=rd) assert "绝密" not in first assert "半成品配方" in first and "人次" in first # 半成品照发,只遮隐藏的部分 # 模型把首轮看到的半成品原样抄回来(<隐藏> 对不上真实标签):静态校验的消息会引用系统发现的标签原文 echoed = json.dumps(A._Outbound(A._secret_texts(grids, sc), grids).json(rd.recipe), ensure_ascii=False) from app.data import recipe as R def validate(data, f, origin): return R.validate_recipe(data, facts=f, origin=origin) model = FakeModel([reply(echoed), reply(echoed), reply(echoed)]) res = await A.ai_draft(None, grids=grids, scan=sc, facts=facts, rules_draft=rd, filename="x.xlsx", # type: ignore[arg-type] validate=validate, dry_run=None, model=model, compressed=c) assert res.attempts == 3 and len(model.calls) == 3 repairs = [m.content for call in model.calls for m in call if isinstance(m, HumanMessage)] assert len(repairs) >= 3 and any("上一次的配方有以下问题" in x for x in repairs) for text in repairs: assert "绝密" not in text def test_hidden_column_header_with_unit_suffix_is_scrubbed(): """SE-2:隐藏列的表头「机密奖金(元)」去掉单位后成了列名「机密奖金」和 units 的键,同样不能外发。""" wb = Workbook() ws = wb.active ws.title = "明细" for j, h in enumerate(["日期", "门店", "机密奖金(元)", "数量(件)"]): ws.cell(1, j + 1, h) for i in range(10): ws.cell(2 + i, 1, dt.datetime(2026, 8, 1 + i)) ws.cell(2 + i, 2, ["店甲", "店乙"][i % 2]) ws.cell(2 + i, 3, 100 + i) ws.cell(2 + i, 4, 5 + i) ws.column_dimensions["C"].hidden = True sc, grids = build(wb) rd = S.draft(sc, grids, "x.xlsx", validate=ok_validate()) assert rd.recipe is not None and "机密奖金" in json.dumps(rd.recipe, ensure_ascii=False) text = A.first_call_text(A.compress(grids, sc, rd.facts), grids=grids, scan=sc, rules_draft=Draft(recipe=rd.recipe, complete=False, origin="rules")) assert "机密" not in text and "半成品配方" in text and "数量" in text def test_outbound_line_scrubs_quoted_numeric_text_and_hidden_variants(): sc, grids = build(_hidden_unit_book()) ob = A._Outbound(A._secret_texts(grids, sc), grids) line = ob.line("应为「全日客流(人次)」=「分区甲(人次)」+「绝密分区(人次)」;列「绝密分区」;表「绝密分区_人次」") assert "绝密" not in line and "全日客流(人次)" in line and "分区甲(人次)" in line def test_total_word_in_middle_of_data_column_is_profiled_not_sent(): """SE-3:备注列中间一格以「累计」开头:它是数据,只进画像,不当合计标签原样外发;最左一列的「合计」照发。""" wb = Workbook() ws = wb.active ws.title = "名单" ws.append(["姓名", "产品", "数量", "金额(元)", "备注"]) rows = [("张一", "甲品", 3, 30, "老客户来电"), ("张二", "甲品", 2, 20, "累计消费12000元"), ("张三", "甲品", 1, 10, "新客户到店"), ("张四", "甲品", 5, 50, "转介绍来的"), ("张五", "甲品", 4, 40, "门店活动价"), ("张六", "甲品", 6, 60, "线上下单的")] for r in rows: ws.append(list(r)) ws.append(["合计", None, 21, 210, None]) sc, grids = build(wb) c = A.compress(grids, sc, S.detect_facts(grids, sc)) assert "累计消费" not in c.text and "12000" not in c.text assert "E 列「备注」是数据列(只给画像):6 格" in c.text assert "A8「合计」" in c.text def test_extract_json_deep_nesting_is_value_error(): """SE-6:深层嵌套(json.loads 会抛 RecursionError)一律按「不是合法的 JSON」处理。""" deep = '{"x":' + "[" * 12000 + "}" with pytest.raises(ValueError): A._extract_json(deep) closed = '{"x":' + "[" * 100 + "]" * 100 + "}" with pytest.raises(ValueError): A._extract_json(closed) assert A._extract_json('{"a": "[[[[{{{{", "b": [[1]]}') == {"a": "[[[[{{{{", "b": [[1]]} @pytest.mark.asyncio async def test_deeply_nested_reply_returns_none_with_usage(flow): """SE-6:回复是约 1.2 万层不闭合的「[」时 ai_draft 不抛异常,返回 draft=None、每次调用的用量都在。""" deep = '{"x":' + "[" * 12000 + "}" model = FakeModel([reply(deep), reply(deep), reply(deep)]) res = await run(flow, model) assert res.draft is None and res.error and len(res.usage) == 3 == res.attempts @pytest.mark.asyncio async def test_exception_after_reply_is_returned_not_raised(flow, monkeypatch): """SE-6:拿到回复之后的处理(这里是 check_recipe 本身)抛了异常:不逃出 ai_draft,用量照记。""" def boom(*a, **kw): raise RecursionError("maximum recursion depth exceeded") monkeypatch.setattr(S, "check_recipe", boom) model = FakeModel([reply(flow_json())]) res = await run(flow, model) assert res.draft is None and len(res.usage) == 1 and res.error @pytest.mark.asyncio async def test_given_model_id_is_recorded_and_budget_wrapper_has_a_name(flow): """AU-7 / SE-4:接口层把同意时解析出的模型和它的 id 交进来,用量记的就是这个 id;_BudgetFallback 的 model_name 是 id,不是底层模型对象的 repr。""" model = FakeModel([reply(flow_json())], name="ignored") res = await run(flow, model, model_id="local-model-A") assert [u.model for u in res.usage] == ["local-model-A"] wrapped = A._BudgetFallback(object(), object(), "cloud-model-B") assert wrapped.model_name == "cloud-model-B"