修复:DSH 自主查询、图片直显与报告稳定性

允许 Agent 在循环中查目录、修正查询和提问,保留实时权限检查及模型原始回答。
统一相同查询证据与图表资源;排名补齐阶段指标,修正 OR 重量筛选,并限制模型明细输入。
增加运行恢复、图表自适应坐标轴与两位小数,修复报告主图选择、重复图及边缘标签。

验证:独立副本通过 162 项 Python、73 项前端及 21 项后端测试,前端构建通过。
线上已用原 8 题复测,并追加两轮双期对比;TOP100 与组合筛选分别核对 100 行和 20 行。
This commit is contained in:
2026-11-12 16:14:20 +08:00
parent f1a2fbd494
commit bc19475b29
35 changed files with 983 additions and 114 deletions
+20 -4
View File
@@ -22,7 +22,7 @@ from setup_dsh_runtime import inspect_runtime, runtime_node
class AuthoredModel:
"""A deterministic provider wire fixture which exercises the actual DSH loop."""
def __init__(self, *, stall=False, length=False, transient=False, malformed=False):
def __init__(self, *, stall=False, length=False, transient=False, malformed=False, repair=False):
self.calls, self.stall, self.length = [], stall, length
self.started = threading.Event()
self.release = threading.Event()
@@ -60,10 +60,12 @@ class AuthoredModel:
if owner.stall:
owner.release.wait(20)
return
if not any(message.get('role') == 'tool' for message in body['messages']):
send({'tool_calls': [{'index': 0, 'id': 'authored-quality-query', 'type': 'function',
if not any(message.get('role') == 'tool' for message in body['messages']) or repair and len(owner.calls)<3:
arguments = ('{"mode":"query","period":{"year":2026},"filters":{"product":"'
+ ('18X15X4' if len(owner.calls)==1 else '18X15X4+T') + '"}}') if repair else '{"mode":"catalog"}'
send({'tool_calls': [{'index': 0, 'id': 'authored-quality-query-'+str(len(owner.calls)), 'type': 'function',
'function': {'name': runtime.PUBLIC_TOOL, 'arguments':
'{"mode":' if malformed and len(owner.calls) == 1 else '{"mode":"catalog"}'}}]})
'{"mode":' if malformed and len(owner.calls) == 1 else arguments}}]})
send({}, 'tool_calls')
else:
send({'content': '已核实'})
@@ -263,6 +265,20 @@ class RealDshTests(unittest.TestCase):
self.assertEqual(acc.tool_calls[0]['status'], 'failed')
self.assertIn('当前范围未授权', acc.tool_calls[0]['outputSummary'])
def test_real_loop_corrects_a_rejected_product_argument_and_continues(self):
from nao_quality_tools import QualityToolError
seen=[]
def query(arguments):
seen.append(arguments['filters']['product'])
if seen[-1]=='18X15X4':
raise QualityToolError('名称未匹配,请使用完整规格18X15X4+T。',code='invalid_arguments')
return {'kind':'quality_metrics','message':'完整规格查询成功'}
acc,status,reason,stopped=self.run_agent(AuthoredModel(repair=True),tool=query)
self.assertEqual((status,reason,stopped),('COMPLETED','stop',True),self.summary)
self.assertEqual(seen,['18X15X4','18X15X4+T'])
self.assertEqual([call['status'] for call in acc.tool_calls],['failed','completed'])
self.assertEqual(acc.answer,'已核实质量指标。')
def test_no_data_is_a_normal_tool_outcome_without_zero_metrics(self):
acc,status,reason,stopped=self.run_agent(AuthoredModel(),tool=lambda args:{
'kind':'quality_no_data','status':'no_data','summary':None,'rows':[],