"""Semantic conversational replies must avoid agent overhead without bypassing evidence work.""" import io,json,sys,threading,unittest from pathlib import Path sys.path.insert(0,str(Path(__file__).resolve().parents[1]/'scripts')) from nao_agent_task import normalize_task,plan_task,TaskUnavailable def direct(**changes): return {'kind':'explain','goal':'回应开场问候并介绍质量分析能力', 'directResponse':{'purpose':'capabilities','answer':'你好,我可以帮你查询质量指标、比较期间变化并分析重点规格。','requiresMemory':False},**changes} class ConversationTests(unittest.TestCase): def test_minimal_semantic_reply_needs_no_full_business_plan(self): task=normalize_task(direct(),'请介绍一下你能做哪些事') self.assertEqual(task['directResponse']['answer'],direct()['directResponse']['answer']) self.assertFalse(task['freshQuery']);self.assertEqual(task['visualPreference'],'none') self.assertEqual(task['dimensions'],[]);self.assertEqual(task['fields'],[]) def test_direct_reply_cannot_request_business_actions_or_hide_memory_writes(self): cases=[{'kind':'metrics'},{'freshQuery':True},{'comparisonRequired':True}, {'verifySource':True},{'fullDetails':True},{'needsClarification':True}, {'fields':['rate']},{'stages':['black']},{'dimensions':['product']}, {'filterUpdates':{'group':'OTHER'}},{'period':{'year':2026,'month':9}}, {'historyMode':'all'},{'visualPreference':'required'}] for changes in cases: with self.subTest(changes=changes),self.assertRaises(TaskUnavailable):normalize_task(direct(**changes),'q') for changes in ({'requiresMemory':True},{'purpose':'business_analysis'}, {'answer':'本月合格率为95.5%'},{'answer':'已完成统计 [Q12345678]'}, {'answer':'https://example.test'},{'answer':'Bearer secretvalue'}): value=direct();value['directResponse'].update(changes) with self.subTest(changes=changes),self.assertRaises(TaskUnavailable):normalize_task(value,'q') def test_plain_chat_preserves_active_business_frame_without_declaring_new_analysis(self): task=normalize_task(direct(),'谢谢你',{'releaseId':'r','period':{'year':2026,'month':9},'filters':{'group':'G'}}) self.assertEqual(task['period'],{'year':2026,'month':9});self.assertEqual(task['filters'],{'group':'G'}) self.assertFalse(task['freshQuery']) def test_rephrased_capability_reply_is_still_plain_conversation(self): value=direct(kind='restate') result=normalize_task(value,'换一种说法介绍一下你能做什么') self.assertEqual(result['kind'],'explain');self.assertIn('directResponse',result) def test_optional_null_direct_reply_does_not_block_real_business_planning(self): from test_nao_agent_task import task_value value=task_value(dimensions=['product'],dimensionStages={'product':['black']},directResponse=None) result=normalize_task(value,'查黑片重点规格') self.assertEqual(result['kind'],'analysis');self.assertTrue(result['freshQuery']);self.assertNotIn('directResponse',result) def test_numbered_capability_lists_do_not_trigger_replanning_or_allow_metric_values(self): value=direct();value['directResponse']['answer']='我能帮助你:\n1. 查询质量指标\n2、比较期间变化\n3) 分析重点规格' task=normalize_task(value,'介绍一下能力') self.assertNotIn('1.',task['directResponse']['answer']);self.assertEqual(task['directResponse']['answer'].count('- '),3) value['directResponse']['answer']='1. 合格率为95.5%' with self.assertRaises(TaskUnavailable):normalize_task(value,'q') value['directResponse']['answer']='我可以帮你筛选TOP5重点规格,也可比较两期指标。' self.assertIn('TOP5',normalize_task(value,'介绍能力')['directResponse']['answer']) for text in ('合格率百分之九十五','报废重量17513','我已记住你的偏好','已经保存了你的设置'): value['directResponse']['answer']=text with self.subTest(text=text),self.assertRaises(TaskUnavailable):normalize_task(value,'q') def test_capability_units_are_not_claimed_measurements(self): value=direct();value['directResponse']['answer']='可查询投入重量(kg)、合格率(%),并按个百分点比较变化。' self.assertIn('kg',normalize_task(value,'介绍能力')['directResponse']['answer']) def test_capabilities_can_describe_query_results_without_claiming_work_completed(self): value=direct();value['directResponse']['answer']='我可以解释查询结果,并把当前数据整理成图表和报告。' self.assertIn('查询结果',normalize_task(value,'能做什么')['directResponse']['answer']) value['directResponse']['answer']='我已经查到最新数据,合格率为95%' with self.assertRaises(TaskUnavailable):normalize_task(value,'q') def test_direct_repair_keeps_minimal_reply_and_points_to_the_actual_field(self): bad=direct();bad['directResponse']['answer']='合格率95%' replies=[bad,direct(),{'aligned':True,'requiresEvidence':False}];payloads=[];errors=[] def opener(req,timeout): payloads.append(json.loads(json.loads(req.data)['messages'][1]['content'])) return io.BytesIO(json.dumps({'choices':[{'finish_reason':'stop','message':{'content':json.dumps(replies.pop(0))}}]}).encode()) result=plan_task(base_url='https://offline.invalid',api_key='synthetic',model='m',question='你好',context={},cancel_event=threading.Event(),opener=opener,on_validation_error=errors.append) self.assertIn('directResponse',result);self.assertEqual(errors[0]['field'],'directResponse.answer') self.assertEqual(errors[0]['receivedType'],'dict') self.assertIn('directResponse',payloads[1]['invalidPlan']);self.assertIn('无需补全',payloads[1]['repairInstructions']) def test_planner_reviews_model_authored_reply(self): bodies=[] def opener(request,timeout): bodies.append(json.loads(request.data));reply=direct() if len(bodies)==1 else {'aligned':True,'requiresEvidence':False} return io.BytesIO(json.dumps({'choices':[{'finish_reason':'stop','message':{'content':json.dumps(reply)}}]}).encode()) task=plan_task(base_url='https://offline.invalid',api_key='synthetic',model='m',question='嗨,先认识一下', context={},cancel_event=threading.Event(),opener=opener) self.assertEqual(len(bodies),2);self.assertIn('directResponse',task) self.assertIn('directResponse',bodies[0]['messages'][0]['content']) def test_exact_json_code_fence_does_not_require_replanning_or_skip_validation(self): calls=[] def opener(request,timeout): calls.append(request) reply=direct() if len(calls)==1 else {'aligned':True,'requiresEvidence':False} return io.BytesIO(json.dumps({'choices':[{'finish_reason':'stop','message':{'content':'```json\n'+json.dumps(reply)+'\n```'}}]}).encode()) result=plan_task(base_url='https://offline.invalid',api_key='synthetic',model='m',question='你好',context={},cancel_event=threading.Event(),opener=opener) self.assertEqual(len(calls),2);self.assertFalse(result['freshQuery']) def test_internal_turn_correlation_marker_is_not_user_visible(self): from nao_quality_runtime import public_progress_text from nao_thinking_stream import public_thinking_text text='原文 quality-turn:12345678-1234-1234-1234-123456789abc 继续答复' for sanitizer in (public_progress_text,public_thinking_text): output=sanitizer(text);self.assertNotIn('quality-turn:',output);self.assertIn('继续答复',output) class AutonomousConversationReviewTests(unittest.TestCase): def test_capability_descriptions_do_not_claim_to_have_run_a_query(self): value=direct();value['directResponse']['answer']='我可以帮你比较不同产业部谁表现更好,并整理已核实的历史结果。' self.assertEqual(normalize_task(value,'你能做什么')['directResponse']['answer'],value['directResponse']['answer']) def test_unsupported_bare_explain_is_repaired_before_protocol_confirmation(self): from test_nao_agent_task import task_value replies=[task_value(kind='explain',freshQuery=False,fields=[],stages=[],dimensions=[],dimensionStages={}),direct(),{'aligned':True,'requiresEvidence':False}] calls=[] def opener(request,timeout): calls.append(json.loads(request.data));reply=replies.pop(0) return io.BytesIO(json.dumps({'choices':[{'finish_reason':'stop','message':{'content':json.dumps(reply)}}]}).encode()) task=plan_task(base_url='https://offline.invalid',api_key='synthetic',model='m',question='介绍一下能做哪些事',context={'_protocol':{'version':1,'autonomousEnabled':True}},cancel_event=threading.Event(),opener=opener) self.assertIn('directResponse',task);self.assertEqual(len(calls),3);self.assertEqual(task['planRecovery']['errors'][0]['field'],'kind.protocol') def test_autonomous_greeting_still_receives_independent_semantic_review(self): bodies=[] def opener(request,timeout): bodies.append(json.loads(request.data));reply=direct() if len(bodies)==1 else {'aligned':True,'requiresEvidence':False} return io.BytesIO(json.dumps({'choices':[{'finish_reason':'stop','message':{'content':json.dumps(reply)}}]}).encode()) task=plan_task(base_url='https://offline.invalid',api_key='synthetic',model='m',question='你好',context={'_protocol':{'version':1,'autonomousEnabled':True}},cancel_event=threading.Event(),opener=opener) self.assertIn('directResponse',task);self.assertEqual(len(bodies),2);self.assertIn('plannedTask',json.loads(bodies[1]['messages'][1]['content'])) def test_autonomous_business_answer_mislabelled_as_conversation_cannot_skip_evidence_review(self): calls=[] def opener(request,timeout): body=json.loads(request.data);calls.append(body) if 'plannedTask' in json.loads(body['messages'][1]['content']):reply={'aligned':True,'requiresEvidence':True} else: reply=direct();reply['directResponse']['answer']='这批比之前差,得调整设备。' return io.BytesIO(json.dumps({'choices':[{'finish_reason':'stop','message':{'content':json.dumps(reply)}}]}).encode()) with self.assertRaises(TaskUnavailable):plan_task(base_url='https://offline.invalid',api_key='synthetic',model='m',question='解释这批次质量问题',context={'_protocol':{'version':1,'autonomousEnabled':True}},cancel_event=threading.Event(),opener=opener) self.assertTrue(any('plannedTask' in json.loads(body['messages'][1]['content']) for body in calls)) def test_unqueried_chinese_values_and_plain_causal_assertions_are_rejected(self): for text in ['黑片合格率下降了,主要原因是设备故障。','上月报废重量为一百。']: value=direct();value['directResponse']['answer']=text with self.subTest(text=text),self.assertRaises(TaskUnavailable):normalize_task(value,'你好') if __name__=='__main__':unittest.main()