"""相容性 Skill（``skills.compatibility.skill``）三层契约的集成测试（任务 18）。 覆盖需求 8.2 与平台契约： - extract：规范化 SMILES / 辅料、识别官能团、渲染结构图（RDKit 可选→降级无图）。 - compute：**纯 Python** 确定性风险分级（签名无 svc，需求 2.1）；数据不足优雅拒绝（8.3）。 - explain：复用 professional_analyzer + 相容性 prompts，LLM 仅写文字、不改数值（2.3）。 - 结构图与风险色块经 ChartService / ReportService 呈现。 测试自包含：svc.llm 用 stub；RDKit 相关断言用 importorskip 守护；不触网。 导入路径由 tests/conftest.py 设置（platform 目录与仓库根加入 sys.path）。 """ from __future__ import annotations import importlib.util import inspect import pytest from kernel.services import Services from kernel.skill_base import ComputeResult, ExtractedData, PharmaSkill, RawInput from skills.compatibility import rules from skills.compatibility.skill import SKILL, CompatibilitySkill _RDKIT_AVAILABLE = importlib.util.find_spec("rdkit") is not None # --------------------------------------------------------------------------- # Stub LLM / Services # --------------------------------------------------------------------------- class _StubLLMResult: def __init__(self, ok: bool, content: str = "") -> None: self.ok = ok self.content = content class _StubLLM: """记录调用并返回固定文本的 LLM 桩。""" def __init__(self, ok: bool = True, content: str = "一、机理分析……\n二、[必须] 控制限度……") -> None: self._ok = ok self._content = content self.calls: list[tuple[str, str]] = [] def complete(self, system: str, user: str, *, temperature: float = 0.3, **kwargs): self.calls.append((system, user)) return _StubLLMResult(self._ok, self._content if self._ok else "") def _svc(llm=None, lang: str = "zh") -> Services: return Services(llm=llm, lang=lang) # --------------------------------------------------------------------------- # 契约：compute 签名无 svc（需求 2.1） # --------------------------------------------------------------------------- def test_compute_signature_has_no_service_handle(): """compute(self, data) 的签名不含 svc / llm 句柄（架构级阻断幻觉计算）。""" sig = inspect.signature(CompatibilitySkill.compute) params = list(sig.parameters) assert params == ["self", "data"] assert "svc" not in params and "llm" not in params def test_skill_is_pharmaskill_and_registered_instance(): assert isinstance(SKILL, CompatibilitySkill) assert isinstance(SKILL, PharmaSkill) assert SKILL.meta.id == "compatibility" # 输入类型声明含 SMILES 与 EXCIPIENT。 kinds = {k.value for k in SKILL.meta.input_kinds} assert "smiles" in kinds and "excipient" in kinds # --------------------------------------------------------------------------- # compute：确定性风险分级 # --------------------------------------------------------------------------- def test_compute_grades_risk_from_extracted_data(): """compute 对 extract 产物做确定性分级（伯胺 + 乳糖 → 高风险）。""" data = ExtractedData( payload={ "smiles": "NCCO", "api_name": "Test API", "excipients": ["乳糖"], "functional_groups": [{"id": "primary_amine"}], "structure_image": None, }, method="manual", ) result = SKILL.compute(data) assert isinstance(result, ComputeResult) assert result.can_proceed is True assert result.summary["overall_risk"] == rules.RISK_HIGH assert result.summary["n_high"] >= 1 assert result.trace # 计算追踪非空 def test_compute_refuses_when_no_excipients(): """无辅料时 compute 优雅拒绝（need 8.3）。""" data = ExtractedData(payload={"smiles": "CCO", "excipients": [], "functional_groups": []}) result = SKILL.compute(data) assert result.can_proceed is False assert result.refusal is not None assert "辅料" in result.refusal["reason"] def test_compute_injects_structure_figure_when_available(): """extract 渲染出结构图时，compute 把其放入 figures 供报告注入。""" data = ExtractedData( payload={ "smiles": "CCO", "excipients": ["乳糖"], "functional_groups": [{"id": "primary_amine"}], "structure_image": "data:image/png;base64,AAAA", } ) result = SKILL.compute(data) assert "structure" in result.figures assert result.figures["structure"]["image_base64"].startswith("data:image/png") def test_compute_is_deterministic(): data = ExtractedData( payload={ "smiles": "NCCO", "excipients": ["乳糖", "硬脂酸镁"], "functional_groups": [{"id": "primary_amine"}, {"id": "ester"}], } ) first = SKILL.compute(data).summary again = SKILL.compute(data).summary assert first == again # --------------------------------------------------------------------------- # explain：LLM 仅写文字；风险矩阵 / 结构渲染收归 report_service（需求 3.1） # --------------------------------------------------------------------------- def _assemble_html(result, sections, lang: str = "zh") -> str: """用 ReportService 组装相容性报告 HTML（矩阵/结构/KPI 由 report_service 渲染）。""" from services.report_service import ReportService return ReportService().assemble(SKILL.meta, result, sections, lang=lang) def test_explain_returns_only_mechanism_and_matrix_rendered_by_report_service(): data = ExtractedData( payload={ "smiles": "NCCO", "api_name": "Test API", "excipients": ["乳糖"], "functional_groups": [{"id": "primary_amine"}], "structure_image": None, } ) result = SKILL.compute(data) llm = _StubLLM(ok=True) sections = SKILL.explain(result, _svc(llm)) # LLM 被调用一次（机理 + 处方）。 assert len(llm.calls) == 1 # 机理段来自 LLM 文本。 assert "机理分析" in sections.sections["mechanism"] # 风险矩阵/结构不再由 explain 产出（已收归 report_service）。 assert "risk_matrix" not in sections.sections assert "structure" not in sections.sections # compute 产出结构化矩阵数据；report_service 据此渲染高风险徽章与矩阵。 assert result.summary["risk_matrix_2d"]["rows"] html = _assemble_html(result, sections, "zh") assert "高风险" in html assert "compat-matrix" in html def test_explain_does_not_alter_compute_numbers(): """explain 不得修改 compute 的风险分级（需求 2.3）。""" data = ExtractedData( payload={ "smiles": "NCCO", "excipients": ["乳糖"], "functional_groups": [{"id": "primary_amine"}], } ) result = SKILL.compute(data) overall_before = result.summary["overall_risk"] SKILL.explain(result, _svc(_StubLLM(ok=True))) assert result.summary["overall_risk"] == overall_before def test_explain_degrades_gracefully_without_llm(): """svc.llm 为 None 时，explain 用确定性摘要兜底，不崩溃。""" data = ExtractedData( payload={ "smiles": "NCCO", "excipients": ["乳糖"], "functional_groups": [{"id": "primary_amine"}], } ) result = SKILL.compute(data) sections = SKILL.explain(result, _svc(llm=None)) assert "mechanism" in sections.sections # 兜底摘要应包含确定性依据文本。 assert sections.sections["mechanism"] def test_explain_falls_back_when_llm_fails(): data = ExtractedData( payload={"smiles": "NCCO", "excipients": ["乳糖"], "functional_groups": [{"id": "primary_amine"}]} ) result = SKILL.compute(data) sections = SKILL.explain(result, _svc(_StubLLM(ok=False))) assert sections.sections["mechanism"] # 非空兜底 def test_explain_english_language(): data = ExtractedData( payload={"smiles": "NCCO", "excipients": ["乳糖"], "functional_groups": [{"id": "primary_amine"}]} ) result = SKILL.compute(data) sections = SKILL.explain(result, _svc(_StubLLM(ok=True), lang="en")) # 英文报告：风险矩阵由 report_service 渲染为英文标签。 html = _assemble_html(result, sections, "en") assert "High risk" in html # --------------------------------------------------------------------------- # extract：规范化 + 官能团识别 + 结构图（RDKit 可选） # --------------------------------------------------------------------------- def test_extract_parses_excipients_heuristically(): """无 RDKit 依赖路径：辅料按行解析，payload schema 完整。""" raw = RawInput(smiles="", excipient="乳糖\n硬脂酸镁") data = SKILL.extract(raw, _svc()) assert isinstance(data, ExtractedData) assert "乳糖" in data.payload["excipients"] assert "硬脂酸镁" in data.payload["excipients"] # 无 SMILES 时附提示。 assert "未提供 SMILES" in data.notes @pytest.mark.skipif(not _RDKIT_AVAILABLE, reason="RDKit 未安装，跳过结构识别断言") def test_extract_identifies_functional_groups_with_rdkit(): """RDKit 可用时，乙醇胺（NCCO）应识别出伯胺并渲染结构图。""" raw = RawInput(smiles="NCCO", excipient="乳糖") data = SKILL.extract(raw, _svc()) group_ids = { (g.get("id") if isinstance(g, dict) else g) for g in data.payload["functional_groups"] } assert "primary_amine" in group_ids # 结构图为 data URI。 assert data.payload["structure_image"] is None or data.payload["structure_image"].startswith("data:image") def test_extract_degrades_when_rdkit_unavailable(monkeypatch): """RDKit 不可用时，extract 跳过结构图与官能团识别但不崩溃。""" # 注入一个声称不可用的渲染器。 class _NoRDKit: is_available = False skill = CompatibilitySkill(molecule_renderer=_NoRDKit()) raw = RawInput(smiles="NCCO", excipient="乳糖") data = skill.extract(raw, _svc()) assert data.payload["functional_groups"] == [] assert data.payload["structure_image"] is None assert "RDKit 不可用" in data.notes # --------------------------------------------------------------------------- # 端到端：extract → compute → explain（RDKit 可选） # --------------------------------------------------------------------------- def test_end_to_end_pipeline_smiles_plus_excipient(): raw = RawInput(smiles="NCCO", excipient="乳糖") llm = _StubLLM(ok=True) svc = _svc(llm) data = SKILL.extract(raw, svc) result = SKILL.compute(data) # 注意：compute 不接收 svc sections = SKILL.explain(result, svc) assert result.can_proceed is True assert result.summary["overall_risk"] in ( rules.RISK_NONE, rules.RISK_LOW, rules.RISK_MEDIUM, rules.RISK_HIGH ) # explain 仅产出机理叙述；矩阵/结构由 report_service 渲染。 assert "mechanism" in sections.sections assert "risk_matrix" not in sections.sections def test_can_handle_scoring(): assert SKILL.can_handle(RawInput(smiles="CCO", excipient="乳糖")) == pytest.approx(0.9) assert SKILL.can_handle(RawInput(excipient="乳糖")) == pytest.approx(0.6) assert SKILL.can_handle(RawInput(smiles="CCO")) == pytest.approx(0.4) assert SKILL.can_handle(RawInput()) == pytest.approx(0.0)