c9f69d9d6d
范围（不含 design-story-foundation、docs/、humanization/README.md 等进行中改动）:
1. 新增 harness/ 控制平面
- skill_harness.py 静态审计：32 个运行时 Skill 的 frontmatter/manifest/文档污染，当前 0 问题
- run_selected.py 选择性执行器：manifest 与磁盘一一对账、依赖阻断、
空跑与 skip-only 失败关闭、AST 测试形状门
- manifests/skills.json：32 个 Skill 的合同责任方与协作领域登记
- manifests/test-inventory.json：81 个测试资产登记
- specs/skill-testing.md 与 README.md：测试分层、证据边界与 harness 职责
2. 实现测试从 .claude/skills/*/scripts/ 迁至 tests/skills/<skill>/
- 71 个测试文件迁移并修复项目根与临时目录运行导入
- 数据库触发器测试宽泛异常收窄为 psycopg.errors.RaiseException
- 抽取离线大测试拆出真实 PG smoke（默认阻断，不计入离线通过）
- 抽取 presence 去重边界拆出独立测试：493 + 78 = 571 项检查不变
3. 运行时文档清理
- 13 个 SKILL.md 移除自测/离线验证段落、测试命令与测试文件事实源表述，
只保留运行时合同；业务运行合同、额度、授权与离线模式均保留
4. SoT 同步
- AGENTS.md：新增 Skill 领域索引（7 个合同责任方分组，覆盖 32 个运行时 Skill）
- 领域 07：测试入口改由 harness/manifests/ 登记，SKILL.md 不承载测试命令
- humanization 覆盖矩阵：活动测试路径同步迁移
验证证据: harness 自测 15 项 + runner 自测 13 项通过；静态审计 32 Skill / 0 问题；
73 个非数据库测试通过；8 个集成条目中 6 个 PostgreSQL 项被依赖门明确阻断；
py_compile 与 git diff --check 通过。未连接 PostgreSQL、网络、真实模型或额度。
已知边界: 真正 skill_behavior_eval 仍为 0，尚未验证任何 Skill 自然语言行为；
evaluate-frozen-replay 的 raw 存储边界冲突留待单独治理。
61 lines
2.4 KiB
Python
61 lines
2.4 KiB
Python
#!/usr/bin/env python3
|
|
"""章后抽取的机械绑定测试，不连接数据库、不调用模型。"""
|
|
import pathlib
|
|
import sys
|
|
|
|
PROJECT_ROOT = pathlib.Path(__file__).resolve().parents[3]
|
|
SCRIPT_DIR = PROJECT_ROOT / ".claude" / "skills" / "extract-chapter-knowledge" / "scripts"
|
|
sys.path.insert(0, str(SCRIPT_DIR))
|
|
from extract_knowledge import ExtractionContractError, normalize_extraction, salvage_extraction # noqa: E402
|
|
|
|
|
|
BODY = "林深走进舰桥，何岚把黑色钥匙交给他。"
|
|
|
|
|
|
def test_normalize_binds_evidence_and_aliases():
|
|
result = normalize_extraction({
|
|
"entities": [{"type": "人物", "name": "林深", "brief": "机师", "fields": {}, "evidence": "林深走进舰桥"}],
|
|
"relations": [],
|
|
"state": {"handoff": "继续调查"},
|
|
}, BODY)
|
|
assert result["entities"][0]["type"] == "character"
|
|
assert result["entities"][0]["evidence"] in BODY
|
|
|
|
|
|
def test_normalize_rejects_fabricated_evidence():
|
|
try:
|
|
normalize_extraction({
|
|
"entities": [{"type": "item", "name": "黑钥匙", "brief": "钥匙", "fields": {}, "evidence": "正文没有这句"}],
|
|
"relations": [], "state": {},
|
|
}, BODY)
|
|
except ExtractionContractError as exc:
|
|
assert "证据不在正文" in str(exc)
|
|
else:
|
|
raise AssertionError("编造证据必须失败关闭")
|
|
|
|
|
|
def test_salvage_drops_unbound_relation_and_keeps_exact_name_evidence():
|
|
result = salvage_extraction({
|
|
"entities": [
|
|
{"type": "item", "name": "黑色钥匙", "brief": "钥匙", "fields": {}, "evidence": "模型编造"},
|
|
{"type": "item", "name": "不存在的物件", "brief": "?", "fields": {}, "evidence": "模型编造"},
|
|
],
|
|
"relations": [{"source": "林深", "target": "何岚", "type": "伙伴", "description": "?", "evidence": "模型编造"}],
|
|
"state": {},
|
|
}, "林深走进舰桥，何岚把黑色钥匙交给他。")
|
|
assert [item["name"] for item in result["entities"]] == ["黑色钥匙"]
|
|
assert result["entities"][0]["evidence"] == "黑色钥匙"
|
|
assert result["relations"] == []
|
|
assert result["mechanicalDrops"] == {"entities": 1, "relations": 1}
|
|
|
|
|
|
def main():
|
|
test_normalize_binds_evidence_and_aliases()
|
|
test_normalize_rejects_fabricated_evidence()
|
|
test_salvage_drops_unbound_relation_and_keeps_exact_name_evidence()
|
|
print("全部通过")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|