Claude Tier MacMini-DSR AI实验室层路由v9.1

Mac Mini上Claude CLI的生产级AI编排。

每个任务都通过LangGraph状态机自动路由到最优模型。 克劳德=只有大脑。 Ollama T1模型执行所有代码、文件和bash命令。 T2型号（Gemini/HuggingFace Kimi）只提供分析，从不执行。

______________________________________________________________________

v9.1的新增功能

更改	详细信息
`startup_banner.py`	替换静态回声挂钩——每个屏幕上都有完整的实时状态横幅 `claude` 开始
每个型号的实时状态	每个型号显示 `✅ LIVE (in RAM)` / `⚡ READY (on-demand)` / `✗ NOT PULLED`
LangSmith实时检查	API启动时ping-显示服务器版本+SDK版本+项目名称
LangGraph实时检查	导入+版本检查(`v1.1.3` 确认）
TierEnforcer状态	intercept.py✅ + 数据库行数+server.py存在
22台MCP服务器已验证	启动时检查了所有服务器脚本，如有缺失，请查看 `✗`
12项技能已验证	所有技能 `.md` 已检查文件--标记丢失的文件
HF API修复	使用 `/api/whoami-v2` （已弃用 `/api/whoami` 总是返回401）
HF Pro帐户	显示 `✅ LIVE @DSR07 (Pro)` 带计划+用户名
设置环境读取	`startup_banner.py` 读取 `HF_API_KEY` 直接 `settings.json`
自动预热背景	每次启动时，T1-LOCAL+T1-MID都会加载到背景线程中的RAM中
预热智能防护	仅预热RAM中尚未预热的型号——如果预热则跳过

______________________________________________________________________

现场创业横幅（每个 `claude` 会话）

╔════════════════════════════════════════════════════════════════════════════╗
║             DSR AI-LAB — TIER ROUTING v9  |  FULL LIVE STATUS              ║
╠════════════════════════════════════════════════════════════════════════════╣
║  🧠 Claude      ✅ AUTH  OAuth via macOS Keychain                            ║
║    Bash        NATIVE  (not intercepted)                                   ║
║    Edit/Write  intercept.py → Ollama T1  (auto-routed)                     ║
╠════════════════════════════════════════════════════════════════════════════╣
║                               INFRASTRUCTURE                               ║
║  LangGraph   ✅ LIVE  v1.1.3  8-node pipeline active                        ║
║  LangSmith   ✅ LIVE  server=0.13.32  sdk=0.7.13  project=dsr-ai-lab-tier-v9║
║  TierEnforcer intercept ✅  DB ✅ (N routes logged)  server ✅               ║
╠════════════════════════════════════════════════════════════════════════════╣
║                  EXECUTORS — Ollama  (all code execution)                  ║
║  ⚙ T1-LOCAL  qwen2.5-coder:7b        ✅ LIVE (in RAM)                       ║
║  ⚙ T1-MID    qwen2.5-coder:14b       ✅ LIVE (in RAM)                       ║
║  ⚙ T1-CLOUD  qwen3-coder:480b-cloud  ⚡ READY (on-demand)                   ║
╠════════════════════════════════════════════════════════════════════════════╣
║                ANALYSIS — Gemini / HF  (never execute code)                ║
║  🔍 T2-FLASH  gemini-2.5-flash   ✅ LIVE  v0.33.0                            ║
║  🔍 T2-PRO    gemini-2.5-pro     ✅ LIVE  v0.33.0                            ║
║  🔍 T2-KIMI   Kimi-K2-Instruct   ✅ LIVE  @DSR07 (Pro)                       ║
╠════════════════════════════════════════════════════════════════════════════╣
║              MCP SERVERS (22)  —  ✅ 22 active  ✅ all present               ║
║  ✅tier-enforcer  ✅filesystem  ✅git  ✅memory  ✅github  ✅gdrive  ...        ║
╠════════════════════════════════════════════════════════════════════════════╣
║                        SKILLS (12)  —  ✅ all loaded                        ║
║  ✅aiapp  ✅arch  ✅math  ✅multifile  ✅rca  ✅scope  ✅tier-*  ✅wire          ║
╠════════════════════════════════════════════════════════════════════════════╣
║  🔥 Prewarm   ⏳ Loading 7b+14b → RAM (background)                          ║
║  📦 Pulled    9 Ollama models in library                                    ║
║  🔗 Pipeline  classify→skill→brain→prewarm→execute→escalate→audit           ║
╚════════════════════════════════════════════════════════════════════════════╝

______________________________________________________________________

系统架构

┌────────────────────────────────────────────────────────────────────┐
│                       DSR AI-Lab  Mac Mini                         │
│                                                                    │
│  ┌────────────────────────────────────────────────────────────┐   │
│  │  Claude CLI  —  BRAIN ONLY                                 │   │
│  │  Bash = NATIVE  |  Edit/Write/MultiEdit = BLOCKED          │   │
│  │                                                            │   │
│  │  SessionStart Hook ──► startup_banner.py                   │   │
│  │    Live checks: Ollama + Gemini + HF + LangSmith +         │   │
│  │    LangGraph + TierEnforcer + 22 MCPs + 12 Skills          │   │
│  │    Auto-prewarms T1-LOCAL + T1-MID (background)            │   │
│  │                                                            │   │
│  │  PreToolUse Hook ──► intercept.py ──► Ollama T1            │   │
│  │    Intercepts: Edit | Write | MultiEdit | NotebookEdit     │   │
│  │    Passthrough: Bash (native)                              │   │
│  │                                                            │   │
│  │  tier-enforcer-mcp  (Python / FastMCP 3.1.0)               │   │
│  │  LangGraph 8 nodes: classify → skill_selector →            │   │
│  │  claude_brain → prewarm_check → [t2_analysis|t1_execute]   │   │
│  │  → escalate → audit                                        │   │
│  └────────────────────────────────────────────────────────────┘   │
│                              │                                     │
│          ┌───────────────────┼───────────────────┐                │
│          ▼                   ▼                   ▼                │
│  ┌──────────────┐  ┌───────────────┐  ┌─────────────────┐        │
│  │  T1-LOCAL    │  │   T1-MID      │  │   T1-CLOUD      │        │
│  │ qwen2.5:7b   │  │ qwen2.5:14b   │  │ qwen3:480b-cloud│        │
│  │ Ollama local │  │ Ollama local  │  │  Ollama cloud   │        │
│  │   4.7 GB     │  │   9.0 GB      │  │  (remote GPU)   │        │
│  │  EXECUTES    │  │  EXECUTES     │  │    EXECUTES     │        │
│  └──────────────┘  └───────────────┘  └─────────────────┘        │
│                                                                    │
│  ┌──────────────┐  ┌───────────────┐  ┌─────────────────┐        │
│  │   T2-FLASH   │  │   T2-PRO      │  │    T2-KIMI      │        │
│  │gemini-2.5-   │  │ gemini-2.5-pro│  │ Kimi-K2-Instruct│        │
│  │flash         │  │               │  │  HF Inference   │        │
│  │ ANALYSIS ──► T1-MID executes ◄───────── ANALYSIS     │        │
│  └──────────────┘  └───────────────┘  └─────────────────┘        │
└────────────────────────────────────────────────────────────────────┘

______________________________________________________________________

层级参考

层	型号	角色	主机	RAM
T1-LOCAL	qwen2.5编码器：7b	执行器——简单/快速	Ollama本地主机：11434	4.7 GB
T1-MID	qwen2.5编码器：14b	执行器——复杂代码	Ollama本地主机：11434	9.0 GB
T1-CLOUD	qwen3编码器：480b云	执行器--史诗/绿地	Ollama云	云
T2-FLASH	双生-2.5-lash	分析→ T1-MID执行	Gemini CLI v0.3.0	--
T2-PRO	双子座-2.5-PRO	深度回顾→ T1-MID执行	Gemini CLI v0.3.0	--
T2-KIMI	Qwen/KIMI-K2-指令	数学/算法→ T1-MID执行	HF推理API（Pro）	-

______________________________________________________________________

LangGraph管道（8个节点）

Input Task
    │
    ▼
┌─────────────┐
│  classify   │  Keyword scan → tier assignment
└──────┬──────┘
       ▼
┌──────────────────┐
│  skill_selector  │  Load domain skill file into context
└──────┬───────────┘
       ▼
┌──────────────┐
│ claude_brain │  Claude writes execution plan (runs for EVERY tier)
└──────┬───────┘
       ▼
┌───────────────┐
│ prewarm_check │  Verify T1 models are loaded in Ollama RAM
└──────┬────────┘
       │
   ┌───┴────────────────────────┐
   │                            │
   ▼ (if T2 classified)         ▼ (T1 classified)
┌────────────┐          ┌──────────────┐
│ t2_analysis│          │  t1_execute  │
│ Gemini/Kimi│──────►   │  Ollama T1   │
└────────────┘          └──────┬───────┘
                               ▼
                        ┌─────────────┐
                        │  escalate   │  score < threshold → next tier (max 2x)
                        └──────┬──────┘
                               ▼
                        ┌─────────────┐
                        │    audit    │  Write to routing_log (11 cols)
                        └──────┬──────┘
                               ▼
                              END

______________________________________________________________________

任务分类→ 路由

Incoming Task Text
        │
        ▼ keyword scan (priority order)
        │
        ├─ "debug" / "error" / "failing" / "broken" / "traceback"
        │         └──► T2-FLASH  (gemini-flash analyzes → T1-MID executes)
        │
        ├─ "analyze" / "explain" / "review" / "audit entire"
        │         └──► T2-PRO    (gemini-pro reviews   → T1-MID executes)
        │
        ├─ "algorithm" / "math" / "big-o" / "statistical"
        │         └──► T2-KIMI   (Kimi-K2 reasons      → T1-MID executes)
        │
        ├─ "full platform" / "greenfield" / "complete system" / "end to end"
        │         └──► T1-CLOUD  (qwen3-coder:480b-cloud executes directly)
        │
        ├─ "implement" / "write module" / "refactor" / "integrate"
        │         └──► T1-MID    (qwen2.5-coder:14b executes)
        │
        └─ everything else (default)
                  └──► T1-LOCAL  (qwen2.5-coder:7b executes)

______________________________________________________________________

拦截流（编辑/写入保护）

Claude Brain produces plan
         │
         ▼
  Claude attempts tool call
         │
         ├──[Bash]──────────────────────────► NATIVE EXEC (passthrough)
         │
         └──[Edit | Write | MultiEdit | NotebookEdit]
                    │
                    ▼
             intercept.py  (PreToolUse hook)
                    │
                    ▼
         Route to Ollama T1 model
         POST /api/chat  (keep_alive=-1)
                    │
                    ▼
         Ollama generates + writes file/edit
                    │
                    ▼
         Claude is blocked — did NOT write

______________________________________________________________________

启动序列

Terminal opens
      │
      ▼ (zshrc / Login Item)
claude-ollama-prewarm.sh
  ├─ GET /api/ps → models loaded?
  ├─ If cold: POST /api/chat 7b + 14b  keep_alive=-1  (background)
  └─ Log to ~/.tier-enforcer/prewarm.log

      │
      ▼  user types: claude
Claude CLI authenticates
  ├─ macOS Keychain: "Claude Code-credentials" → OAuth sk-ant-oat01-...
  └─ claude.ai subscription verified

      │
      ▼  settings.json loaded
  ├─ 22 MCP servers spawned via stdio
  ├─ PreToolUse hook: intercept.py registered
  └─ CLAUDE.md brain protocol loaded

      │
      ▼  SessionStart hook fires → startup_banner.py
  Parallel threads (max 6s):
  ├─ Claude OAuth     → macOS Keychain check
  ├─ LangGraph        → import check + version
  ├─ LangSmith        → GET api.smith.langchain.com/info
  ├─ TierEnforcer     → intercept.py + DB + server.py
  ├─ Ollama           → GET /api/tags + /api/ps per-model
  ├─ Gemini CLI       → gemini --version
  ├─ HF API           → GET /api/whoami-v2 (not whoami — that's broken)
  ├─ 22 MCP servers   → each script/command existence
  ├─ 12 Skills        → each .md file existence
  └─ Auto-prewarm bg  → 7b + 14b if not in RAM

  FULL LIVE BANNER printed with actual statuses

      │
      ▼  mandatory MCP calls (CLAUDE.md protocol)
  activate_tier_routing()   → LangGraph 8 nodes compiled
  tier_health_check(ALL)    → live tier status map
  prewarm_models()          → confirm 7b + 14b in Ollama RAM

      │
      ▼
  READY — every task routed through execute_task()

______________________________________________________________________

回退/升级链

T1-LOCAL ──(score<0.45)──► T1-MID ──(score<0.55)──► T1-CLOUD
                                                          │
                                                   (score<0.60)
                                                          ▼
T2-KIMI ◄──(score<0.50)── T2-PRO ◄──(score<0.50)── T2-FLASH
   │
(max 2 fallbacks reached → return best result obtained)

______________________________________________________________________

文件

文件	目的
`tier-enforcer-mcp/server.py`	FastMCP 3.1.0，LangGraph 8节点，SQLite审计
`tier-enforcer-mcp/intercept.py`	PreTool使用钩子--编辑/写入→ 奥拉玛
`tier-enforcer-mcp/startup_banner.py`	新版v9.1 --会话开始时显示完整的实时状态横幅
`tier-enforcer-mcp/langgraph_tier.py`	LangGraph状态+节点定义
`dotfiles/CLAUDE.md`	Brain协议v9——启动调用、层规则
`dotfiles/settings.json`	挂钩+22个MCP服务器+环境变量
`dotfiles/settings.local.json`	会话启动挂钩→ startup_banner.py

______________________________________________________________________

22台MCP服务器

类别	服务器
核心	层执行器、文件系统、git、内存、github、gdrive
开发意图mcp、arch mcp、编码mcp、rca-mcp、集成mcp、aidev mcp、数学mcp
域	预算mcp、上下文mcp、rpa-mcp
平台	移动开发mcp、网络移动开发mcp、网站开发mcp和电子商务mcp
自动化	mac自动化mcp、文件自动化mcp

______________________________________________________________________

12技能(`~/.claude/skills/`)

aiapp · arch · math · multifile · rca · scope · tier-audit · tier-debug · tier-health · tier-report · tier-reset · wire

______________________________________________________________________

环境变量

# Set in ~/.claude/settings.json → mcpServers.tier-enforcer.env
OLLAMA_LOCAL_HOST=http://localhost:11434
OLLAMA_CLOUD_HOST=http://localhost:11434
OLLAMA_TIMEOUT_LOCAL=600
OLLAMA_TIMEOUT_MID=600
OLLAMA_TIMEOUT_CLOUD=600
HF_API_KEY=hf_...                  # HuggingFace read token (whoami-v2 verified)
LANGCHAIN_TRACING_V2=true
LANGCHAIN_PROJECT=dsr-ai-lab-tier-v9

______________________________________________________________________

快速设置

git clone https://github.com/dineshsrivastava07-cell/Claude-Tier-MacMini.git
cd Claude-Tier-MacMini

# Python dependencies
pip install fastmcp langgraph langchain-core huggingface_hub langsmith

# TypeScript tier-router-mcp
npm install && npm run build

# Copy dotfiles
cp dotfiles/CLAUDE.md ~/.claude/CLAUDE.md
cp dotfiles/settings.json ~/.claude/settings.json
cp dotfiles/settings.local.json ~/.claude/settings.local.json

# Copy tier-enforcer-mcp
mkdir -p ~/tier-enforcer-mcp
cp tier-enforcer-mcp/*.py ~/tier-enforcer-mcp/
cp tier-enforcer-mcp/*.sh ~/tier-enforcer-mcp/

# Set API keys in ~/.claude/settings.json
#   HF_API_KEY        = HuggingFace token with read scope
#   LANGCHAIN_API_KEY = LangSmith token (in env, not in settings.json)

# Authenticate Claude
claude auth login   # OAuth → macOS Keychain

# Start
claude
# startup_banner.py fires automatically — full live status shown

______________________________________________________________________

v9 → v9.1更改

特性	v9	v9.1
启动横幅	静态回声单行	`startup_banner.py` --真正的并行实时检查
型号状态	二元Olama✓/✗	每种型号： `✅ LIVE` / `⚡ READY` / `✗ NOT PULLED`
LangSmith	未显示	实时API ping-服务器+SDK版本
LangGraph	未显示	导入+版本
TierEnforcer	未显示	intercept.py+数据库行+服务器.py
MCP状态	未显示	启动时已验证所有22个
技能状态	未显示	启动时已验证全部12项技能
HF终点	`/api/whoami` (401)	`/api/whoami-v2` （正确）
HF密钥源	仅限于环境	读取 `settings.json` 直接
预热	仅外部脚本	背景线程 `startup_banner.py`

______________________________________________________________________

*DSR人工智能实验室——Mac Mini——v9.1-2026-03-22*

Claude Tier Macmini

快速接入

详细介绍