优化:未分类连续维护与柏拉图预汇总,同步已发布修复

This commit is contained in:
2026-11-01 12:40:50 +08:00
parent 572873c6b8
commit 38290f5fec
135 changed files with 17613 additions and 663 deletions
+200 -10
View File
@@ -2,9 +2,14 @@
from __future__ import annotations
import copy
import hashlib
import marshal
import re
import sys
import time
import traceback
from datetime import datetime, timezone
from pathlib import Path
from urllib.error import HTTPError, URLError
from urllib.parse import urlsplit
@@ -37,6 +42,37 @@ def _platform_visual_context(result):
'missingEvidence': copy.deepcopy(result.get('missingEvidence', []))}
def query_evidence_bindings(documents, proofs):
"""Bind public evidence references to this execution's signed API receipts.
Normal evidence references are random and must not be reverse-engineered
from query IDs. The backend independently checks these IDs and chart values.
"""
allowed = {proof.get('queryId') for proof in proofs if isinstance(proof, dict)}
bindings = {}
def query_ids(result):
if not isinstance(result, dict):
return set()
raw = result.get('protocolEvidence')
items = [raw] if isinstance(raw, dict) else raw if isinstance(raw, list) else []
ids = {item['queryId'] for item in items
if isinstance(item, dict) and isinstance(item.get('queryId'), str)}
for side in ('current', 'comparison'):
ids.update(query_ids(result.get(side)))
return ids
for document in documents:
ref = document.get('evidenceRef')
ids = query_ids(document.get('result'))
if not isinstance(ref, str) or not re.fullmatch(r'Q[A-F0-9]{8}', ref) or not ids:
continue
if not ids <= allowed:
raise ValueError('图表证据包含不属于本轮的查询证明')
if ref in bindings and bindings[ref] != ids:
raise ValueError('同一证据编号不能绑定不同正式查询')
bindings[ref] = ids
return [{'evidenceRef': ref, 'queryIds': sorted(ids)} for ref, ids in bindings.items()]
def _confirmed_scalar_selection(request, task):
"""Dispatch an already confirmed scalar choice without rewriting its task."""
resolution = task.get('productResolution') or {}
@@ -67,7 +103,8 @@ def _save_composition(final, actor_root, turn_id):
from nao_quality_runtime import _atomic_json
_atomic_json(actor_root / 'private' / 'answer-composition.json', {
'turnId': turn_id, 'rawModelResponses': final['rawModelResponses'],
'compositionDiagnostics': final.get('compositionDiagnostics', {})})
'compositionDiagnostics': final.get('compositionDiagnostics', {}),
'bindingInput': final.get('_privateBindingInput')})
def _retryable_query_error(error):
@@ -470,6 +507,131 @@ def _safe_private_error_summary(error, task):
return message[:180] or '查询执行时发生内部错误。'
def _private_module_provenance():
"""Fingerprint the code objects actually loaded by this worker."""
modules = {
'nao_protocol_execution': ('_scoped_arguments', '_execute_query_plan'),
'nao_quality_tools': ('QualityTools.quality_data',),
'nao_weight_ranking': ('rank_response',),
'nao_combined_ranking': ('_stage_has_explicit_zero_source', 'combined_response'),
'nao_query_contract': ('bind_query',),
'nao_agent_task': ('dimension_stages',),
}
result = {'python': sys.version.split()[0], 'executable': sys.executable, 'modules': {}}
for name, symbols in modules.items():
module = sys.modules.get(name)
if module is None:
result['modules'][name] = {'loaded': False}
continue
entry = {'loaded': True, 'file': getattr(module, '__file__', None), 'functions': {}}
spec = getattr(module, '__spec__', None)
if spec is not None:
entry['origin'] = getattr(spec, 'origin', None)
source_path = getattr(module, '__file__', None)
if isinstance(source_path, str):
try:
source = Path(source_path)
if source.is_file():
entry['sourceSha256'] = hashlib.sha256(source.read_bytes()).hexdigest()
entry['sourceMtimeNs'] = source.stat().st_mtime_ns
except OSError:
pass
for dotted in symbols:
value = module
try:
for part in dotted.split('.'):
value = getattr(value, part)
code = getattr(value, '__code__', None)
if code is not None:
entry['functions'][dotted] = hashlib.sha256(marshal.dumps(code)).hexdigest()
except (AttributeError, TypeError, ValueError):
continue
result['modules'][name] = entry
return result
def _private_failure_diagnostic(error, diagnostic, task, arguments=None):
"""Retain sanitized first-party validation detail only in private diagnostics."""
from nao_quality_tools import QualityToolError
result = copy.deepcopy(diagnostic)
if isinstance(error, QualityToolError):
message = str(error)
# Never persist SQL, transport addresses, or credential-bearing text,
# even if a future caller wraps an upstream failure in QualityToolError.
unsafe = re.compile(
r'(?is)(?:\b(?:select|insert|update|delete|drop|alter|truncate|create)\b|\bsql\b|https?://|'
r'\b(?:authorization|cookie|set-cookie)\s*[:=]|\bbearer\s+\S+)')
result['errorSummary'] = (
'受控查询校验失败;详细内容已隐藏。'
if unsafe.search(message) else _safe_private_error_summary(error, task))
result['moduleProvenance'] = _private_module_provenance()
rank_trace = getattr(error, '_quality_private_rank_trace', None)
if isinstance(rank_trace, dict):
result['rankResponseTrace'] = copy.deepcopy(rank_trace)
frames = []
for frame in traceback.extract_tb(error.__traceback__)[-16:]:
filename = Path(frame.filename).name
function = frame.name
if (re.fullmatch(r'[A-Za-z0-9_.-]{1,100}', filename)
and re.fullmatch(r'[A-Za-z0-9_.<>-]{1,100}', function)
and type(frame.lineno) is int and frame.lineno > 0):
frames.append({'file': filename, 'function': function, 'line': frame.lineno})
if frames:
result['exceptionFrames'] = frames
if isinstance(arguments, dict):
allowed = {
'mode': {'query', 'compare', 'rank', 'changes', 'investigate', 'defects'},
'dimension': {'product', 'department', 'group', 'defect'},
'metric': {'rate', 'scrapRate', 'totalWeight', 'scrapWeight'},
'stage': {'black', 'bright'},
}
summary = {}
for key, choices in allowed.items():
value = arguments.get(key)
if value in choices or key == 'stage' and value is None:
summary[key] = value
limit = arguments.get('limit')
if type(limit) is int and 1 <= limit <= 100:
summary['limit'] = limit
if summary:
result['queryInput'] = summary
return result
def _controlled_failure_diagnostic(error, *, node_id=None, purpose=None):
"""Return bounded failure classification without copying upstream text or query values."""
from nao_quality_tools import QualityToolError, SourceCoverageUnavailable
from nao_quality_runtime import RuntimeErrorSafe
raw_status = getattr(error, 'http_status', getattr(error, 'code', None))
status = raw_status if type(raw_status) is int and 100 <= raw_status <= 599 else None
retryable = _retryable_query_error(error)
if isinstance(error, SourceCoverageUnavailable):
kind = 'source_coverage'
elif status is not None:
kind = 'transient_http' if retryable else 'http_error'
elif isinstance(error, QualityToolError):
kind = 'formal_query_rejected'
elif isinstance(error, RuntimeErrorSafe) and retryable or isinstance(error, (TimeoutError, ConnectionError, OSError)):
kind = 'transport_unavailable'
elif isinstance(error, ValueError):
kind = 'query_rejected'
else:
kind = 'internal_error'
error_type = type(error).__name__
if not re.fullmatch(r'[A-Za-z][A-Za-z0-9_]{0,63}', error_type):
error_type = 'Exception'
result = {'failureKind': kind, 'errorType': error_type, 'retryable': retryable}
if status is not None:
result['httpStatus'] = status
if isinstance(node_id, str) and re.fullmatch(r'[A-Za-z0-9_-]{1,64}', node_id):
result['nodeId'] = node_id
if purpose in {'metrics', 'ranking', 'trend', 'changes', 'defects', 'source', 'records'}:
result['purpose'] = purpose
return result
def _raw_period_label(value):
from nao_evidence_report import period
try:
@@ -808,9 +970,10 @@ def _scoped_arguments(request, scope, node):
elif node['purpose'] == 'ranking':
dimension = node['dimension']
for stage in dimension_stages(task, dimension, combined=False):
if stage not in node['stages']:
if stage not in node['stages'] and not (stage is None and not node['stages']):
raise ValueError('排名scope请求阶段超出queryPlan节点')
for field in node['fields']:
if stage is None and field == 'rate':continue
item = {'mode': 'rank', 'releaseId': request['releaseId'],
'period': copy.deepcopy(task['period']), 'filters': copy.deepcopy(filters),
'dimension': dimension, 'stage': stage, 'metric': field,
@@ -884,6 +1047,8 @@ def _execute_query_plan(*, request, task, authority, actor_root, delegated, conn
server_gaps = []
calls, proofs, accepted_refs = [], [], set()
query_diagnostics = []
failure_diagnostics = []
private_failure_diagnostics = []
raw_diagnostics = []
query_recovery = []
query_ids = set()
@@ -929,6 +1094,7 @@ def _execute_query_plan(*, request, task, authority, actor_root, delegated, conn
if cancel_event.is_set():raise RuntimeError('cancelled')
node = plan_nodes[scope['nodeId']]
arguments_queue = []
active_arguments = None
scoped_api = None
active_raw_diagnostic = None
try:
@@ -1073,6 +1239,7 @@ def _execute_query_plan(*, request, task, authority, actor_root, delegated, conn
'toolCalls':copy.deepcopy(calls),'toolCount':len(calls)})
continue
for index, arguments in enumerate(arguments_queue):
active_arguments = arguments
if cancel_event.is_set():raise RuntimeError('cancelled')
started = datetime.now(timezone.utc).isoformat()
reads_before = scoped_api.read_count
@@ -1118,9 +1285,13 @@ def _execute_query_plan(*, request, task, authority, actor_root, delegated, conn
emit({'type': 'progress', 'message': f'正式查询 {read_count}/40 已完成。',
'toolCalls': copy.deepcopy(calls), 'toolCount': len(calls)})
except Exception as error:
if isinstance(error, RuntimeError) and (str(error) == 'cancelled'
or not _retryable_query_error(error)):
if isinstance(error, RuntimeError) and str(error) == 'cancelled':
raise
diagnostic = _controlled_failure_diagnostic(
error, node_id=scope.get('nodeId'), purpose=node.get('purpose'))
failure_diagnostics.append(diagnostic)
private_failure_diagnostics.append(
_private_failure_diagnostic(error, diagnostic, scope.get('task'), active_arguments))
recovered = None
if (scoped_api is not None and node['purpose'] == 'metrics' and arguments_queue
and len(arguments_queue) == 1 and arguments_queue[0].get('mode') == 'compare'
@@ -1193,7 +1364,18 @@ def _execute_query_plan(*, request, task, authority, actor_root, delegated, conn
continue
# A child failure remains visible and can never be hidden by a
# successful sibling scope.
server_gaps.append(f"节点{node['nodeId']}范围{scope.get('label') or scope['scopeId']}未完成({type(error).__name__})。")
safe_failure = {
'source_coverage': '所选期间和筛选没有可用的正式来源覆盖',
'formal_query_rejected': '正式查询响应未通过范围或完整性核验',
'transient_http': '正式查询服务暂时不可用,可在稍后重试',
'http_error': '正式查询服务拒绝或无法处理本次请求',
'transport_unavailable': '正式查询连接暂时不可用,可在稍后重试',
'query_rejected': '正式查询未通过已确认的范围或参数校验',
'internal_error': '正式查询执行未完成',
}.get(diagnostic['failureKind'], '正式查询执行未完成')
label = {'metrics':'正式指标','ranking':'正式排名','trend':'正式趋势','changes':'正式变化',
'defects':'正式缺陷','source':'原表核验','records':'原始记录'}.get(node.get('purpose'),'正式查询')
server_gaps.append(f"节点{node['nodeId']}:{label}{safe_failure}。")
if cancel_event.is_set():raise RuntimeError('cancelled')
scopes, pending, gaps, skipped = _protocol_query_snapshot(delegated, include_skipped=True)
@@ -1371,21 +1553,26 @@ def _execute_query_plan(*, request, task, authority, actor_root, delegated, conn
'period': copy.deepcopy(task['period'])}
if task.get('comparisonPeriod') is not None:
agent_context['comparisonPeriod'] = copy.deepcopy(task['comparisonPeriod'])
if any(node.get('purpose')=='changes' for node in task.get('queryPlan',[])):
if not final.get('readingSummary') and any(node.get('purpose')=='changes' for node in task.get('queryPlan',[])):
# The model sees a concise fact set; queried statistical tables must
# still survive in the persisted answer. The existing UI folds these
# sections into its full appendix without further business reads.
# still survive as normal key tables, without exposing audit appendices
# or performing further business reads.
sections=[section for section in scoped_report.get('sections',[])
if isinstance(section,str) and section.strip()]
if isinstance(section,str) and section.strip() and section.strip() not in final['answer']]
if sections:
final['answer'] += '\n\n## 统计数据附录\n\n'+'\n\n'.join(sections)
final['answer'] += '\n\n## 重点数据\n\n'+'\n\n'.join(sections)
result = {'type': 'result', 'engine': 'NAO', 'status': status, 'answer': final['answer'],
'readingSummary': final.get('readingSummary'),
'answerQuality': {'version': 1, 'dataComplete': bool(covered),
'narrativeComplete': bool(final.get('evidenceSufficient')),
'chartsComplete': not missing_visuals,
'missingEvidence': missing_evidence},
'stopReason': 'evidence_finalized' if complete else 'evidence_incomplete',
'finishReason': 'stop', 'completionMethod': 'bounded_scoped_query_plan',
'agentTask': task, 'agentContext': agent_context, 'protocol': identity,
'actualModel': model, 'phaseTimingsMs': timings,
'protocolEvidence': proofs, 'queryEvidenceRefs': [d['evidenceRef'] for d in evidence],
'queryEvidenceBindings': query_evidence_bindings(evidence, proofs),
'toolCalls': calls, 'toolCount': len(calls), 'reportTitle': final.get('reportTitle'),
'evidenceCompletion': {'serverCoverageSufficient': covered,
'modelDeclaredEvidenceSufficient': final.get('modelDeclaredEvidenceSufficient', final.get('evidenceSufficient')),
@@ -1394,6 +1581,7 @@ def _execute_query_plan(*, request, task, authority, actor_root, delegated, conn
'visualSchemaVersion': 1, 'visualState': visuals.get('visualState', 'unavailable'),
'visuals': visuals.get('visuals', []),
'scopeDiagnostics': copy.deepcopy(raw_diagnostics),
'failureDiagnostics': copy.deepcopy(failure_diagnostics),
'queryRecovery': copy.deepcopy(query_recovery),
'scopedReport': {'facts': scoped_report.get('facts', []),
'sections': scoped_report.get('sections', []),
@@ -1404,6 +1592,7 @@ def _execute_query_plan(*, request, task, authority, actor_root, delegated, conn
'phaseTimingsMs': timings, 'queryIds': [proof['queryId'] for proof in proofs],
'protocolEvidence': proofs,
'queryDiagnostics': query_diagnostics,
'failureDiagnostics': private_failure_diagnostics,
'queryRecovery': query_recovery,
'rawScopeDiagnostics': raw_diagnostics,
'completion': {'serverCoverageSufficient': covered,
@@ -1720,6 +1909,7 @@ def execute(*, request, task, authority, actor_root, delegated, connection, canc
'agentTask': task, 'agentContext': agent_context, 'protocol': identity,
'actualModel': model, 'phaseTimingsMs': timings,
'protocolEvidence': proofs, 'queryEvidenceRefs': [d['evidenceRef'] for d in evidence],
'queryEvidenceBindings': query_evidence_bindings(evidence, proofs),
'toolCalls': calls, 'toolCount': len(calls), 'reportTitle': final.get('reportTitle'),
'evidenceCompletion': completion, 'factsComplete': facts_complete, **visuals}
from nao_quality_runtime import _atomic_json