修复:智能分析空回答与工具结束条件冲突

统一自主分析工具范围与执行前校验,严格核验正常结束、最终正文和图表恢复状态,并显示空回答及恢复失败原因。

包含两次已上线修复及回归测试,保留已有回答与业务数据。

验证:1101 项 Python、469 项前端、97 项后端测试通过;前后端构建和真实模型、页面验收通过。
This commit is contained in:
2026-11-03 08:56:49 +08:00
parent 7ffb5cdb9c
commit 52d92c95d8
20 changed files with 387 additions and 37 deletions
+17 -8
View File
@@ -62,20 +62,29 @@ def capability_context(task=None):
}
def apply_platform_tools(payload):
"""Do not advertise a native SQL-result renderer for quality API evidence."""
QUALITY_NATIVE_TOOLS = frozenset({'read', 'list', 'grep', 'search', 'mcp_call', 'mcp_connect'})
def apply_platform_tools(payload, *, native_agent=False):
"""Expose tools compatible with platform evidence and answer completion."""
tools = payload.get('tools')
if not isinstance(tools, list): return payload
kept = [tool for tool in tools if not (isinstance(tool, dict)
and isinstance(tool.get('function'), dict)
and tool['function'].get('name') == 'display_chart')]
if len(kept) == len(tools): return payload
allowed = QUALITY_NATIVE_TOOLS if native_agent else QUALITY_NATIVE_TOOLS | {'story'}
# Clarification/warehouse selection are upstream stop conditions with native
# UI replies that this platform cannot consume. SQL, renderers and writers
# also use a different data/artifact lifecycle. Unknown tools stay disabled.
tools = tools if isinstance(tools, list) else []
kept = [tool for tool in tools if isinstance(tool, dict) and tool.get('type') == 'function'
and isinstance(tool.get('function'), dict) and isinstance(tool['function'].get('name'), str)
and tool['function'].get('name') in allowed]
if kept: payload['tools'] = kept
else: payload.pop('tools', None)
choice = payload.get('tool_choice')
if ((isinstance(choice, dict) and (choice.get('function') or {}).get('name') == 'display_chart')
kept_names = {tool['function']['name'] for tool in kept}
choice_name = choice.get('function', {}).get('name') if isinstance(choice, dict) and isinstance(choice.get('function'), dict) else None
if ((isinstance(choice, dict) and choice_name not in kept_names)
or not kept and choice in ('auto', 'required')):
payload.pop('tool_choice', None)
if not kept: payload.pop('parallel_tool_calls', None)
return payload
+50 -13
View File
@@ -21,6 +21,10 @@ class _NoRedirect(HTTPRedirectHandler):
return None
class ToolPolicyError(ValueError):
pass
class ModelGate:
"""A stable per-session endpoint backed only by this turn's live delegation."""
@@ -41,7 +45,9 @@ class ModelGate:
self._thinking = dict(thinking_options) if thinking_options is not None else None
self._on_thinking = on_thinking
self._native_turn_marker = native_turn_marker
self.native_agent = False
self.rejected_stale_requests = 0
self.rejected_tool_calls = 0
self._request_number = 0
self._request_prefix = secrets.token_hex(4)
self.request_audit = []
@@ -92,9 +98,10 @@ class ModelGate:
for message in messages)):
with gate._audit_lock: gate.rejected_stale_requests += 1
self.failure(403, '模型请求不属于当前轮次'); return
decoded = gate.apply_thinking(decoded)
from nao_agent_contract import apply_platform_tools
decoded = apply_platform_tools(decoded)
decoded = apply_platform_tools(decoded, native_agent=gate.native_agent)
decoded = gate.apply_thinking(decoded)
offered_tools = {tool['function']['name'] for tool in decoded.get('tools', [])}
payload = json.dumps(decoded,ensure_ascii=False).encode()
except (ValueError, OSError):
self.failure(400, '模型请求格式无效'); return
@@ -141,9 +148,10 @@ class ModelGate:
thinking('start')
# 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)) if gate._on_thinking or is_sse else None
observer = ProviderStream(lambda text:thinking('delta',text=text))
raw_json = bytearray()
if not bridge:
pending_wire = bytearray()
if is_sse and not bridge:
self.send_response(response.status)
self.send_header('Content-Type', response.headers.get('Content-Type', 'application/json'))
self.send_header('Cache-Control', 'no-store')
@@ -161,21 +169,29 @@ class ModelGate:
if time.monotonic() - checked_at >= 1:
gate._check()
checked_at = time.monotonic()
if observer:
if is_sse:observer.feed(chunk)
else:raw_json.extend(chunk)
if not bridge:
self.wfile.write(chunk)
self.wfile.flush()
if is_sse:
observer.feed(chunk)
gate.check_returned_tools(observer, offered_tools)
if not bridge:
# A partial SSE line cannot be checked yet. Hold
# it until its complete event has passed policy.
pending_wire.extend(chunk)
end=pending_wire.rfind(b'\n')+1
if end:
self.wfile.write(pending_wire[:end]);self.wfile.flush()
del pending_wire[:end]
else:raw_json.extend(chunk)
# A provider may keep the HTTP connection alive with heartbeats
# after its explicit terminal event. Do not wait for EOF, but
# keep reading past finish_reason so the usage tail is preserved.
if is_sse and observer.done:
break
gate._check()
if observer:
if is_sse:observer.finish()
else:observer.document(json.loads(raw_json))
if is_sse: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:
self.wfile.write(pending_wire);self.wfile.flush()
result = observer.completion() if bridge or is_sse else None
stops = [row['finish_reason'] for row in observer.choices.values()] if observer else []
thinking('end',finishReason='length' if 'length' in stops else (stops[-1] if stops else ''),interrupted=bool(observer and not stops))
@@ -188,6 +204,15 @@ class ModelGate:
self.send_response(200);self.send_header('Content-Type','application/json')
self.send_header('Cache-Control','no-store');self.send_header('Content-Length',str(len(encoded)))
self.end_headers();headers_sent=True;self.wfile.write(encoded);self.wfile.flush()
elif not is_sse:
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 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:
with contextlib.suppress(OSError):
self.wfile.write(b'data: {"error":{"message":"Tool call is unavailable for this task"}}\n\n');self.wfile.flush()
except HTTPError as error:
if not headers_sent:
self.failure(error.code if 400 <= error.code <= 599 else 502, '模型接口暂不可用')
@@ -230,6 +255,8 @@ class ModelGate:
else:payload['reasoning_effort']=effort
with self._audit_lock:
if len(self.request_audit)<200:self.request_audit.append({'enabled':enabled,'effort':effort,
'tools':[tool['function']['name'] for tool in payload.get('tools',[])],
'nativeAgent':self.native_agent,
'protocol':protocol,'stream':payload.get('stream',False),
'maxCompletionTokens':payload.get('max_completion_tokens',payload.get('max_tokens')),
'thinkingType':(payload.get('thinking') or {}).get('type'),
@@ -237,6 +264,16 @@ class ModelGate:
'reasoningEffort':payload.get('reasoning_effort',(payload.get('chat_template_kwargs') or {}).get('reasoning_effort'))})
return payload
def check_returned_tools(self, observer, offered, *, final=False):
for choice, tools in observer.tools.items():
finished=final or bool(observer.choices.get(choice,{}).get('finish_reason')) or observer.done
for tool in tools.values():
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
raise ToolPolicyError('Unadvertised tool call')
def _check(self) -> None:
if self._closed.is_set():
raise PermissionError('模型授权已关闭')
+3 -1
View File
@@ -112,7 +112,9 @@ def run_native_exploration(*, request, task, authority, actor_root, delegated, c
and type(progress.get('noProgress')) is int and progress['noProgress'] >= 3)
if controlled_no_progress:
missing.append('连续三次没有取得新证据,调查已停止;其余线索尚未验证。')
natural_finish = 'stop' in accumulator.finish_reasons and stop_succeeded is not False
natural_finish = bool(accumulator.finish_reasons and accumulator.finish_reasons[-1] == 'stop'
and stop_succeeded is not False
and getattr(accumulator, 'final_answer_ready', True))
corrected_tool_errors = (native_reason == 'mcp_error' and natural_finish
and complete_scopes is not None and bool(documents)
and not missing and not progress.get('halted'))
+30 -4
View File
@@ -763,6 +763,17 @@ class ConversationAccumulator:
return value
return ""
@property
def final_answer_ready(self) -> bool:
if not self.answer or any(call.get('status') == 'started' for call in self.tool_calls):
return False
if not self.tool_calls:
return True
last_tool = max((i for i, entry in enumerate(self.timeline)
if entry.get('kind') in ('tool_start', 'tool_result')), default=-1)
return any(entry.get('kind') == 'assistant' and str(entry.get('text') or '').strip()
for entry in self.timeline[last_tool + 1:])
def status(self, *, cancel_cause: str | None = None, stop_succeeded: bool | None = None,
transport_failed: bool = False) -> tuple[str, str | None]:
if cancel_cause == "user":
@@ -781,18 +792,27 @@ class ConversationAccumulator:
return ("PARTIAL" if self.answer else "FAILED", "transport_or_stream_error")
if self.tool_error and (self.unattributed_tool_error or any(call["status"] == "failed" and not call.get("recovered") for call in self.tool_calls)):
return ("PARTIAL" if self.answer else "FAILED", "mcp_error")
if self.answer and (self.done_received or "stop" in self.finish_reasons):
if (self.final_answer_ready and self.finish_reasons[-1] == 'stop'
and stop_succeeded is not False):
return "COMPLETED", "stop"
return ("PARTIAL" if self.answer else "FAILED", "incomplete")
def _contains_error_flag(value: Any) -> bool:
if isinstance(value, dict):
if value.get("isError") is True or value.get("error"):
if (any(value.get(flag) is True for flag in ('isError', 'mcpValidationError', 'mcpAuthRequired'))
or value.get('success') is False or value.get('error')):
return True
return any(_contains_error_flag(item) for item in value.values())
# Inspect only known response envelopes. Rows, source cells, notes and
# arbitrary JSON strings are business data, not nested tool outcomes.
content = value.get('content')
if isinstance(content, list) and any(isinstance(part, dict) and part.get('type') == 'text'
and _contains_error_flag(part.get('text')) for part in content):
return True
return _contains_error_flag(value['structuredContent']) if isinstance(value.get('structuredContent'), dict) else False
if isinstance(value, list):
return any(_contains_error_flag(item) for item in value)
return any(isinstance(part, dict) and part.get('type') == 'text'
and _contains_error_flag(part.get('text')) for part in value)
if isinstance(value, str) and len(value) <= 64 * 1024:
stripped = value.strip()
if stripped.startswith(("{", "[")):
@@ -1922,6 +1942,10 @@ def _sse_request(
ready_state: dict[str, Any] = {"identity": None, "since": None}
def completion_ready() -> bool:
# The watchdog can observe a stale halted exploration state between a
# natural model finish and the stream's trailing [DONE] marker.
if accumulator.done_received or (accumulator.finish_reasons and accumulator.finish_reasons[-1] == 'stop'):
return False
if autonomous:
progress_path = actor_root / 'private' / 'native-exploration-state.json'
if not _require_private_regular_file(progress_path):
@@ -2630,6 +2654,7 @@ def _run_turn(payload: dict[str, Any], *, stdout: Any = sys.stdout, stdin: Any =
from nao_protocol_execution import execute
request['_autonomousAnswer'] = protocol.get('autonomousEnabled') is True
if authority['task'].get('executionMode') == 'native_agent':
model_gate.native_agent = True
from nao_native_run import run_native_exploration
return run_native_exploration(request=request, task=authority['task'], authority=authority,
actor_root=actor_root, delegated=delegated, connection=selected_connection,
@@ -3108,6 +3133,7 @@ def _run_turn(payload: dict[str, Any], *, stdout: Any = sys.stdout, stdin: Any =
if model_gate is not None:
try:
_atomic_json(actor_root/'private'/'last-model-audit.json',{'model':selected_model,'requests':model_gate.request_audit,
'rejectedToolCalls':model_gate.rejected_tool_calls if isinstance(model_gate.rejected_tool_calls,int) else 0,
'rejectedStaleRequests':model_gate.rejected_stale_requests if isinstance(model_gate.rejected_stale_requests,int) else 0})
except (OSError,TypeError,ValueError):pass
model_gate.close()