统一:DSH 智能分析、Top100 与图表静默恢复
接入 DSH 专用运行时和对话界面,保留模型原始回答,支持会话分页、连续追问与显式报告。 补齐组合筛选、各维度 Top100、分组查询到排名图的转换,图表故障静默省略并基于保存数据自动恢复。 优化报告风格、图表顺序和表格换行,更新接入说明及构建输入。 验证:独立目录前后端构建通过;118 项 Java、102 项前端、102 项图表、14 项 DSH 应用和 13 项组合筛选检查通过。 仅提交相关源码、锁定依赖与测试;不包含凭据、会话、业务数据或本地产物。
This commit is contained in:
@@ -4,7 +4,9 @@ from __future__ import annotations
|
||||
import contextlib
|
||||
import hmac
|
||||
import json
|
||||
import re
|
||||
import secrets
|
||||
import socket
|
||||
import ssl
|
||||
import threading
|
||||
import time
|
||||
@@ -15,6 +17,8 @@ from urllib.error import HTTPError
|
||||
from urllib.parse import urlsplit
|
||||
from urllib.request import HTTPRedirectHandler, HTTPSHandler, Request, build_opener
|
||||
|
||||
from nao_tool_stream import ToolStreamNormalizer, ToolStreamProtocolError
|
||||
|
||||
|
||||
class _NoRedirect(HTTPRedirectHandler):
|
||||
def redirect_request(self, *_args, **_kwargs):
|
||||
@@ -30,7 +34,9 @@ class ModelGate:
|
||||
|
||||
def __init__(self, *, upstream_base: str, upstream_key: str, model: str,
|
||||
port: int, client_key: str, authorize: Callable[[], None], thinking_options=None, on_thinking=None,
|
||||
native_turn_marker: str | None = None) -> None:
|
||||
native_turn_marker: str | None = None, runtime_engine: str = 'NAO') -> None:
|
||||
if runtime_engine not in ('NAO', 'DSH'):
|
||||
raise ValueError('分析引擎无效。')
|
||||
parsed = urlsplit(upstream_base)
|
||||
if parsed.scheme not in ('http', 'https') or not parsed.hostname or parsed.username or parsed.query or parsed.fragment:
|
||||
raise ValueError('模型接口地址无效。')
|
||||
@@ -46,13 +52,19 @@ class ModelGate:
|
||||
self._on_thinking = on_thinking
|
||||
self._native_turn_marker = native_turn_marker
|
||||
self.native_agent = False
|
||||
self.runtime_engine = runtime_engine
|
||||
self.rejected_stale_requests = 0
|
||||
self.rejected_tool_calls = 0
|
||||
self.rejected_tool_names = []
|
||||
self.tool_stream_failures = []
|
||||
self.tool_stream_diagnostics = []
|
||||
self._request_number = 0
|
||||
self._request_prefix = secrets.token_hex(4)
|
||||
self.request_audit = []
|
||||
self._audit_lock = threading.Lock()
|
||||
self._closed = threading.Event()
|
||||
self._responses = set()
|
||||
self._responses_lock = threading.Lock()
|
||||
bundle = Path('/etc/ssl/cert.pem')
|
||||
trust = ssl.create_default_context(cafile=str(bundle) if bundle.is_file() else None)
|
||||
self._opener = build_opener(_NoRedirect(), HTTPSHandler(context=trust))
|
||||
@@ -99,7 +111,8 @@ class ModelGate:
|
||||
with gate._audit_lock: gate.rejected_stale_requests += 1
|
||||
self.failure(403, '模型请求不属于当前轮次'); return
|
||||
from nao_agent_contract import apply_platform_tools
|
||||
decoded = apply_platform_tools(decoded, native_agent=gate.native_agent)
|
||||
decoded = apply_platform_tools(decoded, native_agent=gate.native_agent,
|
||||
runtime_engine=gate.runtime_engine)
|
||||
decoded = gate.apply_thinking(decoded)
|
||||
offered_tools = {tool['function']['name'] for tool in decoded.get('tools', [])}
|
||||
payload = json.dumps(decoded,ensure_ascii=False).encode()
|
||||
@@ -127,6 +140,7 @@ class ModelGate:
|
||||
'Accept': self.headers.get('Accept', 'application/json')})
|
||||
headers_sent = False
|
||||
thought_ended = False
|
||||
response = None
|
||||
try:
|
||||
from nao_runtime_policy import REQUEST_TIMEOUT_SECONDS
|
||||
remaining = max(0, gate._deadline-time.monotonic()) if gate._deadline is not None else REQUEST_TIMEOUT_SECONDS
|
||||
@@ -140,6 +154,8 @@ class ModelGate:
|
||||
decoded['stream'] = False;decoded.pop('stream_options',None)
|
||||
outgoing.data = json.dumps(decoded,ensure_ascii=False).encode()
|
||||
response = gate._opener.open(outgoing,timeout=min(remaining,REQUEST_TIMEOUT_SECONDS))
|
||||
with gate._responses_lock:
|
||||
gate._responses.add(response)
|
||||
with response:
|
||||
gate._check()
|
||||
from nao_thinking_stream import ProviderStream
|
||||
@@ -149,6 +165,7 @@ class ModelGate:
|
||||
# SSE termination is a transport concern even when no thinking
|
||||
# callback is installed (for example, native stream forwarding).
|
||||
observer = ProviderStream(lambda text:thinking('delta',text=text))
|
||||
tool_stream = ToolStreamNormalizer(offered_tools) if is_sse and gate.runtime_engine == 'DSH' else None
|
||||
raw_json = bytearray()
|
||||
pending_wire = bytearray()
|
||||
if is_sse and not bridge:
|
||||
@@ -170,6 +187,8 @@ class ModelGate:
|
||||
gate._check()
|
||||
checked_at = time.monotonic()
|
||||
if is_sse:
|
||||
if tool_stream is not None:
|
||||
chunk = tool_stream.feed(chunk)
|
||||
observer.feed(chunk)
|
||||
gate.check_returned_tools(observer, offered_tools)
|
||||
if not bridge:
|
||||
@@ -187,7 +206,12 @@ class ModelGate:
|
||||
if is_sse and observer.done:
|
||||
break
|
||||
gate._check()
|
||||
if is_sse:observer.finish()
|
||||
if is_sse:
|
||||
if tool_stream is not None:
|
||||
tail = tool_stream.finish()
|
||||
observer.feed(tail)
|
||||
if not bridge: pending_wire.extend(tail)
|
||||
observer.finish()
|
||||
else:observer.document(json.loads(raw_json))
|
||||
gate.check_returned_tools(observer, offered_tools, final=True)
|
||||
if is_sse and not bridge and pending_wire:
|
||||
@@ -208,6 +232,16 @@ class ModelGate:
|
||||
self.send_response(response.status);self.send_header('Content-Type','application/json')
|
||||
self.send_header('Cache-Control','no-store');self.send_header('Content-Length',str(len(raw_json)))
|
||||
self.end_headers();headers_sent=True;self.wfile.write(raw_json);self.wfile.flush()
|
||||
except ToolStreamProtocolError as error:
|
||||
with gate._audit_lock:
|
||||
if len(gate.tool_stream_failures) < 20:
|
||||
gate.tool_stream_failures.append(error.code)
|
||||
gate.tool_stream_diagnostics.append({'code': error.code, 'phase': phase,
|
||||
'offered': sorted(offered_tools), 'names': error.tool_names})
|
||||
if not headers_sent:self.failure(502,'模型工具响应格式异常,请重试本轮')
|
||||
else:
|
||||
with contextlib.suppress(OSError):
|
||||
self.wfile.write(b'data: {"error":{"message":"Invalid tool stream framing"}}\n\n');self.wfile.flush()
|
||||
except ToolPolicyError:
|
||||
if not headers_sent:self.failure(502,'模型返回了当前任务未开放的工具调用')
|
||||
else:
|
||||
@@ -220,6 +254,9 @@ class ModelGate:
|
||||
if not headers_sent:
|
||||
self.failure(502, '模型接口暂不可用或授权已失效')
|
||||
finally:
|
||||
if response is not None:
|
||||
with gate._responses_lock:
|
||||
gate._responses.discard(response)
|
||||
if not thought_ended:thinking('end',interrupted=True)
|
||||
|
||||
class Server(ThreadingHTTPServer):
|
||||
@@ -239,6 +276,14 @@ class ModelGate:
|
||||
enabled,effort,protocol=options.get('enabled'),options.get('effort'),options.get('protocol')
|
||||
if type(enabled) is not bool or effort not in ('auto','low','medium','high') or protocol not in ('thinking','chat_template'):
|
||||
raise ValueError('思考配置无效')
|
||||
requested_enabled = enabled
|
||||
# User-approved MiMo compatibility: its documented thinking/tool
|
||||
# instability can yield incomplete argument streams. Keep the saved
|
||||
# preference and non-tool requests intact; do not repair model output.
|
||||
tool_thinking_override = bool(enabled and self.native_agent and payload.get('tools')
|
||||
and self._model.lower().startswith('mimo-'))
|
||||
if tool_thinking_override:
|
||||
enabled = False
|
||||
payload.pop('thinking',None);payload.pop('enable_thinking',None);payload.pop('reasoning_effort',None)
|
||||
if protocol=='chat_template':
|
||||
extra=payload.get('chat_template_kwargs')
|
||||
@@ -254,7 +299,8 @@ class ModelGate:
|
||||
payload['chat_template_kwargs']['reasoning_effort']='xhigh' if effort=='high' else effort
|
||||
else:payload['reasoning_effort']=effort
|
||||
with self._audit_lock:
|
||||
if len(self.request_audit)<200:self.request_audit.append({'enabled':enabled,'effort':effort,
|
||||
if len(self.request_audit)<200:self.request_audit.append({'enabled':enabled,'requestedEnabled':requested_enabled,
|
||||
'toolThinkingOverride':tool_thinking_override,'effort':effort,
|
||||
'tools':[tool['function']['name'] for tool in payload.get('tools',[])],
|
||||
'nativeAgent':self.native_agent,
|
||||
'protocol':protocol,'stream':payload.get('stream',False),
|
||||
@@ -271,7 +317,13 @@ class ModelGate:
|
||||
name=tool.get('function',{}).get('name','')
|
||||
valid=(name in offered) if finished else (not name or any(allowed.startswith(name) for allowed in offered))
|
||||
if not valid:
|
||||
with self._audit_lock:self.rejected_tool_calls+=1
|
||||
with self._audit_lock:
|
||||
self.rejected_tool_calls += 1
|
||||
# Names only: never record arguments, provider errors or
|
||||
# arbitrary text masquerading as a function name.
|
||||
safe_name = name if isinstance(name, str) and re.fullmatch(r'[A-Za-z_][A-Za-z0-9_.:-]{0,119}', name) else '<invalid_name>'
|
||||
if safe_name not in self.rejected_tool_names and len(self.rejected_tool_names) < 20:
|
||||
self.rejected_tool_names.append(safe_name)
|
||||
raise ToolPolicyError('Unadvertised tool call')
|
||||
|
||||
def _check(self) -> None:
|
||||
@@ -279,8 +331,24 @@ class ModelGate:
|
||||
raise PermissionError('模型授权已关闭')
|
||||
self._authorize()
|
||||
|
||||
def close(self) -> None:
|
||||
def revoke(self) -> None:
|
||||
"""Close this turn's capability and interrupt in-flight provider reads."""
|
||||
self._closed.set()
|
||||
with self._responses_lock:
|
||||
responses = list(self._responses)
|
||||
for response in responses:
|
||||
# shutdown wakes read1 before HTTPResponse.close waits on its file
|
||||
# lock. No provider address, header or credential is logged here.
|
||||
stream = getattr(response, 'fp', None)
|
||||
raw = getattr(stream, 'raw', stream)
|
||||
channel = getattr(raw, '_sock', None)
|
||||
with contextlib.suppress(OSError, AttributeError):
|
||||
channel.shutdown(socket.SHUT_RDWR)
|
||||
with contextlib.suppress(OSError, ValueError):
|
||||
response.close()
|
||||
|
||||
def close(self) -> None:
|
||||
self.revoke()
|
||||
self._server.shutdown()
|
||||
self._server.server_close()
|
||||
self._thread.join(timeout=1)
|
||||
|
||||
Reference in New Issue
Block a user