feat(agent): trusted time evidence and collection prompt optimization

- time_context.py: server-side observation clocks for relative-time evidence (anti client-forgery); resolves X分钟前/小时前 against the observation timestamp of the turn that produced them, never re-basing history on the current clock; flags historical references without a reliable baseline - state: dialogue_timestamps keyed by dialogue message index, persisted with checkpoint; welcome seeds index 0 - collect: single clock per turn shared by reply/extraction/repair; time_evidence injected into runtime context; reply_validator hook rejects time claims lacking evidence (one shared repair budget); model still interprets facts and chooses actions - prompts: collection prompt restructured (shorter, state-bound-to-action table, gatekeeper principle, time source rules); extraction prompt aligned; workflow@20261021 json updated to match - evals: prompt_probes / prompt_compare / time_boundary probe scripts (supplementary checks, original cases untouched) - docs: collection-prompt-optimization.md (method + baselines), time-boundary-fix.md (problem, design, verification) - gitignore: editor artifact dirs (.agents/.claude/.vscode, skills-lock) Verified: 171 unit tests pass (16 new time-context tests)
2026-10-22 09:41:32 +08:00
parent a890f41f6f
commit 9e28f69b3c
14 changed files with 827 additions and 588 deletions
@@ -0,0 +1,83 @@
+"""Live end-to-end time provenance probes, separate from unchanged cases."""
+import argparse
+import asyncio
+from datetime import datetime, timedelta
+import json
+from pathlib import Path
+import re
+from zoneinfo import ZoneInfo
+
+from src.agent.collect import run_collect
+from src.agent.llm import build_llm_service
+
+TZ = ZoneInfo('Asia/Shanghai')
+ORIGIN = datetime(2026,10,22,0,5,tzinfo=TZ)
+HISTORY = [
+    {'role':'user','content':'十分钟前两辆汽车追尾，没有非机动车或行人，我还在现场。'},
+    {'role':'assistant','content':'<state>1002</state>请问有没有人员受伤？'},
+]
+
+
+def parsed_time(value):
+    if not isinstance(value,str):
+        return None
+    match = re.fullmatch(r'(\d{4})年(\d{1,2})月(\d{1,2})日(\d{1,2})点(\d{1,2})分',value)
+    return tuple(map(int,match.groups())) if match else None
+
+
+async def main(output):
+    llm = build_llm_service('openai')
+    cases = []
+    for delay in (75,135):
+        cases.append(dict(name=f'anchored_cross_midnight_delay_{delay}',now=ORIGIN+timedelta(minutes=delay),
+                          input='没有人受伤。',clocks={'0':ORIGIN.isoformat()},expected=(2026,10,21,23,55)))
+    for repeat in range(5):
+        cases.append(dict(name=f'legacy_missing_reference_{repeat}',now=ORIGIN+timedelta(hours=2),
+                          input='没有人受伤。',clocks={},expected=None))
+    cases.append(dict(name='explicit_absolute_correction',now=ORIGIN+timedelta(hours=2),
+                      input='没有人受伤，事故时间更正为2026年10月21日23点40分。',
+                      clocks={'0':ORIGIN.isoformat()},expected=(2026,10,21,23,40)))
+    semaphore=asyncio.Semaphore(4)
+    async def run(case):
+        calls=[]
+        class Recorder:
+            async def complete(self,system_prompt,user_message,*,history=None):
+                raw=await llm.complete(system_prompt,user_message,history=history)
+                calls.append({'kind':'extraction' if system_prompt.startswith('【字段提取】') else 'dialogue',
+                              'raw':raw})
+                return raw
+        async with semaphore:
+            checks=[]
+            try:
+                result=await run_collect({'stage_code':'1002','dialogue':HISTORY,
+                                          'dialogue_timestamps':case['clocks'],'input_text':case['input']},
+                                         Recorder(),now=case['now'])
+                reply=result['response_text']
+                fields=result['accident_fields']
+                checks.append(dict(text='continues collection',passed=result['stage_code']=='1002'))
+                expected=case['expected']
+                if expected is None:
+                    checks.append(dict(text='no unanchored time in form',passed='accident_time' not in fields))
+                    checks.append(dict(text='asks for time without inventing it',passed=not re.search(r'\d{4}年.*点.*分',reply)
+                                       and bool(re.search('什么时候|几点|事故.*时间',reply))))
+                else:
+                    checks.append(dict(text='form uses original clock or explicit correction',
+                                       passed=parsed_time(fields.get('accident_time'))==expected))
+                    value=re.search(r'\d{4}年\d{1,2}月\d{1,2}日\d{1,2}点\d{1,2}分',reply)
+                    checks.append(dict(text='dialogue confirms same supported value',
+                                       passed=bool(value) and parsed_time(value.group())==expected))
+                return dict(name=case['name'],passed=all(c['passed'] for c in checks),checks=checks,
+                            reply=reply,fields=fields,calls=calls)
+            except Exception as exc:
+                return dict(name=case['name'],passed=False,error=type(exc).__name__+': '+str(exc),calls=calls)
+    results=await asyncio.gather(*(run(c) for c in cases))
+    report={'passed':sum(r['passed'] for r in results),'total':len(results),'results':results}
+    output.parent.mkdir(parents=True,exist_ok=True)
+    output.write_text(json.dumps(report,ensure_ascii=False,indent=2),encoding='utf-8')
+    print(json.dumps({k:v for k,v in report.items() if k!='results'}))
+
+
+if __name__=='__main__':
+    parser=argparse.ArgumentParser(__doc__)
+    parser.add_argument('--output',type=Path,required=True)
+    asyncio.run(main(parser.parse_args().output))