a207-router-mcp
Provides deterministic intent routing and safety gating for multi-agent systems, enforcing P0 critical word detection, mutually exclusive hard rules (MX-1/2/3), and a permission matrix as pure functions via four MCP tools.
README
a207-router-mcp(M13 · 意图路由与安全闸)
调度 Agent 独占的确定性判定层。安全红线不交给概率模型:P0 危急词、三条互斥硬规则 (MX-1/MX-2/MX-3)、权限矩阵,全部落成可单测的纯函数。
- 技术栈:FastMCP ≥2.0 + pydantic ≥2 + Python 3.13,stdio 传输
- 依据:
docs/01_多智能体与MCP总体需求梳理_v1.md§5/§6/§7、docs/02_Wave1_构建规格_v1.md§2.1 - 零跨包引用:不 import 任何其他
a207-*包,入参由调度 Agent 从 PCP 取值后显式传入
包结构
src/a207_router_mcp/
core.py 安全闸 + 意图分类(纯逻辑,无 fastmcp)
permissions.py 权限矩阵执行点 + 写工具识别(MX-3 落点)
constants.py 角色、13 个 MCP 登记表、权限矩阵、写权策略
lexicon.py P0 危急词、分期词族、意图路由表、槽位词典
slots.py 正则抽槽位
models.py pydantic 请求/响应模型
server.py FastMCP 实例,挂载四个工具
tests/test_tools.py 46 个断言,纯 python 可跑
四个工具
check_safety_gate(intent, caller, target_mcp="")
判定顺序固定:P0 危急词 → 入参登记校验 → MX-1 → MX-2 → MX-3 → 权限矩阵兜底。
返回:{allow, reason, applied_rule, gate, delegate_to, requires_confirmation, matched_keywords}
applied_rule:MX-1 | MX-2 | MX-3 | nonegate:P0_EMERGENCY | INPUT | MX | PERMISSION_MATRIX | PASS,用于区分同为none的拒绝来源
| 规则 | 内容 |
|---|---|
| P0 | 命中危急词(抽搐/少尿/无尿/大量出血/意识不清等 30 余条)即断崖终止,不进入任何业务 MCP |
| MX-1 | 分期/eGFR/几期/肾小球滤过/PEW/生长障碍/血压百分位类问题,caller 为 nutritionist 或 parent_assistant 时拒绝并 delegate_to=doctor_assistant;a207-ckd-clinical-calc-mcp 对非 doctor_assistant 一律关闭 |
| MX-2 | a207-gamification-mcp 只接受 child_companion;child_companion 调 HIS/LIS/营养计算/临床计算/食谱/风险/报告/通知一律拒绝 |
| MX-3 | 写工具 → 允许 caller:upsert_lab_result→doctor_assistant;upsert_food_diary→parent_assistant/child_companion;log_meal_checkin、award_badge→child_companion;push_to_emr→doctor_assistant(requires_confirmation=true);notify_*/trigger_warning_event/close_warning→risk_warning。工具与 target_mcp 不匹配同样拒绝 |
intent 可传用户原文,也可直接传工具名;中文写入说法(写回病历、录入化验、打卡等)已登记别名,避免绕过 MX-3。
classify_intent(text)
规则命中优先,未命中返回 intent="unknown" 且 fallback="llm",由上层软分类后再过一次安全闸。
返回:{intent, mcp, slots, confidence, allowed_callers, matched_keywords, candidates, fallback, note}
| intent | 目标 MCP | 允许 caller |
|---|---|---|
| emergency_red_flag | a207-router-mcp | 全部(强制前置,fallback=halt) |
| patient_profile | a207-his-mcp | 调度/医生/营养师(RL)/家长(RL)/风险预警 |
| clinical_staging | a207-ckd-clinical-calc-mcp | 仅 doctor_assistant |
| lab_result | a207-lis-mcp | 医生/家长(受限视图)/风险预警 |
| nutrition_calc | a207-nutrition-calc-mcp | 医生/营养师/家长 |
| meal_plan | a207-meal-plan-mcp | 医生/营养师 |
| risk_assessment | a207-risk-rules-mcp | 医生/风险预警 |
| nutrition_assessment | a207-nutrition-assessment-mcp | 见权限矩阵,写入按 MX-3 |
| gamification | a207-gamification-mcp | 仅 child_companion |
| notify | a207-notify-mcp | 仅 risk_warning 可写 |
| report | a207-report-mcp | 医生/营养师/家长(RL)/风险预警 |
| knowledge | a207-knowledge-mcp | 全部,按角色切语料 |
| followup | a207-followup-mcp | 医生/营养师/风险预警 |
打分口径:命中关键词长度求和,同分按优先级(P0→P5)取胜;confidence = min(0.95, 0.50 + 0.05 × 分数)。
"排食谱/一周食谱/三餐怎么安排"路由到 M7;裸词"食谱"按构建规格 §2.1 归 M5,由调用方结合上下文决定是否转交。
check_permission(caller, mcp, action="read")
需求文档 §7 权限矩阵的执行点。action 可传 read/write,也可直接传工具名。
返回:{allow, access, resolved_action, reason, write_tool, requires_confirmation, limited_view}
access为矩阵原始格子:R/RL/R/W/-;limited_view=true表示只能给受限视图- 精度说明:命中已登记写工具时,MX-3 映射优先于矩阵格子(v1.2 已用 MX-3 收口写入路径),
因此
parent_assistant + a207-nutrition-assessment-mcp + upsert_food_diary放行,其余读操作仍按矩阵为-。
extract_slots(text)
正则 + 词典抽槽位,抽不到留空,绝不猜测。
返回:{patient_id, food[], analyte[], date[], number[], age_years, weight_kg, height_cm, time_window_days, meal, measure[]}
patient_id匹配 PCP 契约^P[0-9]{4,}$analyte归一到 PCP 字段名(potassium / phosphorus / scr / albumin / hemoglobin …),单位口径以contracts/pcp-schema.json为准:Scr μmol/L,K/P/Ca mmol/L,能量 kcal,蛋白 g,身高 cm,体重 kgnumber保留原始单位;斤自动折算 kg,月龄自动折算岁
验证
python -m py_compile src/a207_router_mcp/*.py
python tests/test_tools.py # total=46 passed=46 failed=0
单测覆盖:MX-1/MX-2/MX-3 越权拒绝与正常放行、P0 断崖、未登记 caller/MCP、矩阵兜底、 意图路由 8 例、权限矩阵 6 组、槽位抽取 6 例,另含"无跨 a207 包引用 + 纯逻辑层不依赖 fastmcp"的静态自查。
启动(需先安装 fastmcp):fastmcp dev src/a207_router_mcp/server.py 或 python -m a207_router_mcp.server。
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.