基座层面的两处缺陷,都属于"静默失败"——排查成本高,且本项目已经各踩过一次。
1. tool_executor.py 的拒绝原因原先无法区分:
- "意图压根没发布白名单"与"白名单里没这个工具"共用一句「工具不在当前意图白名单」,
运维不知道该去补发布配置、还是改白名单内容(客服与风控的意图码都要求三处对齐,
两次都因此多花排查时间);
- 权限与角色两处只说「缺少工具权限」,不说是哪一个。
现在四种情况各有独立 message,各自指向不同的处置动作。
同时把**审计与异常分离**:白名单内容、权限码、角色集属于内部配置,只写进审计;
异常 message 会随 API 响应返回给调用方,保持通用、不泄漏配置。
2. model_gateway.py 的 TASK_CAPABILITY 补齐风控的几处 task_type
(risk_agent_chat / risk_analysis / risk_script / risk_summary / daily_report_suggestion)。
它们要的都是文本生成端点;不登记就会落到"未映射 → 返回全部 active 端点"的分支,
而能否选对端点取决于 model_endpoint_config 的**行顺序**——实测风控能跑通,仅仅因为
deepseek-flash(id=3) 恰好排在 qwen-embedding(id=5) 前面。这个隐式依赖现在消掉了。
未登记的 task_type 仍退回全部端点(保持原有保守策略:让故障表现为调用失败而不是
解析为空),但会记 warning,不再静默。
新增 tests/unit/service/test_tool_executor_denials.py(4 条),锁住"四种拒绝可区分"
与"内部细节只进审计、不进 message"。
ruff / mypy(135 文件) / 607 unit+contract 全绿。
137 lines
5.7 KiB
Python
137 lines
5.7 KiB
Python
import asyncio
|
||
from dataclasses import dataclass
|
||
from datetime import UTC, datetime
|
||
from typing import Any, Protocol
|
||
|
||
from pydantic import BaseModel, ValidationError
|
||
|
||
from app.core.contracts import RequestContext, SourceReference, ToolCallRecord
|
||
from app.core.errors import (
|
||
DependencyUnavailableError,
|
||
ForbiddenAgentError,
|
||
UpstreamTimeoutError,
|
||
ValidationAgentError,
|
||
)
|
||
from app.infrastructure.db import SessionFactory
|
||
from app.model.audit import InteractionAudit
|
||
|
||
|
||
class ToolHandler(Protocol):
|
||
async def __call__(self, arguments: BaseModel, context: RequestContext) -> Any: ...
|
||
|
||
|
||
@dataclass(frozen=True)
|
||
class ToolDefinition:
|
||
name: str
|
||
input_model: type[BaseModel]
|
||
handler: ToolHandler
|
||
required_permission: str
|
||
allowed_roles: tuple[str, ...]
|
||
read_only: bool = True
|
||
timeout_seconds: float = 5
|
||
|
||
|
||
@dataclass(frozen=True)
|
||
class ToolExecution:
|
||
output: Any
|
||
record: ToolCallRecord
|
||
references: tuple[SourceReference, ...] = ()
|
||
|
||
|
||
class ToolRegistry:
|
||
def __init__(self) -> None:
|
||
self._tools: dict[str, ToolDefinition] = {}
|
||
|
||
def register(self, definition: ToolDefinition) -> None:
|
||
if definition.name in self._tools:
|
||
raise ValidationAgentError("工具名称重复")
|
||
if not definition.read_only:
|
||
raise ValidationAgentError("Agent 公共工具仅允许只读")
|
||
self._tools[definition.name] = definition
|
||
|
||
def get(self, name: str) -> ToolDefinition:
|
||
definition = self._tools.get(name)
|
||
if definition is None:
|
||
raise ForbiddenAgentError("工具未注册")
|
||
return definition
|
||
|
||
|
||
class ToolExecutor:
|
||
def __init__(self, registry: ToolRegistry) -> None:
|
||
self.registry = registry
|
||
|
||
async def execute(
|
||
self, *, name: str, arguments: dict[str, Any], intent: str,
|
||
configured_tools: dict[str, tuple[str, ...]], context: RequestContext,
|
||
) -> ToolExecution:
|
||
definition = self.registry.get(name)
|
||
# 把三种"用不了"分开,并且**审计写详细、异常给通用**:
|
||
#
|
||
# 原先前两种情况共用一句"工具不在当前意图白名单",运维无法判断该去补发布配置、
|
||
# 还是该改白名单内容——本项目已经因此踩坑两次(客服、风控的意图码都要求三处对齐,
|
||
# 而缺配置时是静默失败关闭)。权限与角色两处也只说"缺少工具权限",不说是哪一个。
|
||
#
|
||
# 细节只进审计:异常 message 会随 API 响应返回给调用方,
|
||
# 白名单内容、权限码、角色集属于内部配置,不该出现在客户可见的响应里。
|
||
reason = ""
|
||
detail = ""
|
||
if intent not in configured_tools:
|
||
reason = "该意图未配置工具白名单"
|
||
detail = (
|
||
f"意图 {intent!r} 在发布版本里没有任何工具白名单"
|
||
f"(agent_tools / <agent_type>:{intent}),工具失败关闭"
|
||
)
|
||
elif name not in configured_tools[intent]:
|
||
reason = "工具不在当前意图白名单"
|
||
detail = (
|
||
f"工具 {name!r} 不在意图 {intent!r} 的白名单 "
|
||
f"{list(configured_tools[intent])} 内"
|
||
)
|
||
elif definition.required_permission not in context.permissions:
|
||
reason = "缺少工具权限"
|
||
detail = f"缺少权限 {definition.required_permission!r}"
|
||
elif not set(definition.allowed_roles).intersection(context.roles):
|
||
reason = "角色不能使用工具"
|
||
detail = (
|
||
f"角色 {list(context.roles)} 与工具允许的角色 "
|
||
f"{list(definition.allowed_roles)} 无交集"
|
||
)
|
||
if reason:
|
||
await self._audit(name, intent, context, "denied", detail)
|
||
raise ForbiddenAgentError(reason)
|
||
try:
|
||
validated = definition.input_model.model_validate(arguments)
|
||
except ValidationError as exc:
|
||
await self._audit(name, intent, context, "denied", "参数校验失败")
|
||
raise ValidationAgentError("工具参数校验失败") from exc
|
||
try:
|
||
async with asyncio.timeout(definition.timeout_seconds):
|
||
output = await definition.handler(validated, context)
|
||
except TimeoutError as exc:
|
||
await self._audit(name, intent, context, "failed", "timeout")
|
||
raise UpstreamTimeoutError("工具调用超时") from exc
|
||
except Exception as exc:
|
||
await self._audit(name, intent, context, "failed", type(exc).__name__)
|
||
raise DependencyUnavailableError("工具调用失败") from exc
|
||
record = ToolCallRecord(
|
||
tool_name=name, status="succeeded",
|
||
input_summary={key: "[redacted]" for key in arguments},
|
||
output_summary={"result_type": type(output).__name__},
|
||
)
|
||
await self._audit(name, intent, context, "succeeded", "ok")
|
||
reference = SourceReference(source_type="tool", source_id=f"{context.trace_id}:{name}",
|
||
title=name)
|
||
return ToolExecution(output=output, record=record, references=(reference,))
|
||
|
||
async def _audit(
|
||
self, name: str, intent: str, context: RequestContext, status: str, reason: str
|
||
) -> None:
|
||
async with SessionFactory() as session, session.begin():
|
||
session.add(InteractionAudit(
|
||
actor_type="agent", actor_id=int(context.user_id), portal=context.portal,
|
||
action_type="agent.tool_executed", detail={
|
||
"tool_name": name, "intent": intent, "status": status,
|
||
"reason": reason, "trace_id": context.trace_id,
|
||
}, created_at=datetime.now(UTC).replace(tzinfo=None),
|
||
))
|