Codex OSS 模式深度解析：接入任意开源大模型

┌─────────────┐     OpenAI Compatible API     ┌──────────────┐
│   Codex     │ ─────────────────────────────→│  任意模型     │
│  Agent Loop │ ←─────────────────────────────│  推理引擎     │
└─────────────┘     JSON Response             └──────────────┘

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# macOS
brew install ollama

# Windows
winget install Ollama.Ollama

# 编程专精模型
ollama pull qwen2.5-coder:32b

# 通用大模型
ollama pull llama3.3:70b

# 轻量快速模型
ollama pull phi4:14b

# 中文优化
ollama pull qwen2.5:32b

// ~/.codex/config.json
{
  "provider": "openai-compatible",
  "model": "qwen2.5-coder:32b",
  "apiBase": "http://localhost:11434/v1",
  "apiKey": "ollama",
  "temperature": 0.7,
  "maxTokens": 8192,
  "streaming": true
}

# 测试 API 是否可用
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen2.5-coder:32b",
    "messages": [{"role": "user", "content": "Hello"}],
    "max_tokens": 50
  }'

# 通过 Codex 测试
codex "写一个 Python 函数计算斐波那契数列"

{
  "provider": "openai-compatible",
  "model": "qwen2.5-coder-32b-instruct",
  "apiBase": "http://localhost:1234/v1",
  "apiKey": "lm-studio",
  "temperature": 0.7
}

# 安装 vLLM
pip install vllm

# 启动 API 服务器
python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-Coder-32B-Instruct \
  --port 8000 \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.9 \
  --max-model-len 32768

{
  "provider": "openai-compatible",
  "model": "Qwen/Qwen2.5-Coder-32B-Instruct",
  "apiBase": "http://your-server:8000/v1",
  "apiKey": "vllm-token",
  "temperature": 0.7,
  "maxTokens": 32768
}

{
  "modelRouting": {
    "default": "qwen2.5-coder:32b",
    "coding": "qwen2.5-coder:32b",
    "reasoning": "deepseek-chat",
    "simple": "phi4:14b",
    "writing": "qwen2.5:32b"
  },
  "providers": {
    "local": {
      "apiBase": "http://localhost:11434/v1",
      "apiKey": "ollama"
    },
    "deepseek": {
      "apiBase": "https://api.deepseek.com/v1",
      "apiKey": "sk-xxx"
    }
  }
}

用户: "写一个二叉树遍历" → 路由到 qwen2.5-coder:32b (本地, 免费)
用户: "分析这段架构设计的问题" → 路由到 deepseek-chat (API, $0.001)
用户: "修正拼写错误" → 路由到 phi4:14b (本地, 快速)

# 生产环境安全清单
security:
  - 模型服务不暴露公网，仅内网访问
  - API 端点添加认证 token
  - 日志记录所有 prompt 和输出
  - 定期审计模型输出内容
  - 网络隔离：模型服务器在独立 VLAN
  
network:
  api_server:
    bind: "0.0.0.0"
    port: 8000
    tls:
      enabled: true
      cert: /etc/ssl/codex.crt
      key: /etc/ssl/codex.key
    auth:
      type: bearer
      tokens:
        - "internal-token-xxx"

# vLLM 监控
python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-Coder-32B-Instruct \
  --port 8000 \
  --disable-log-requests \  # 生产环境关闭请求日志
  --uvicorn-log-level info

# Prometheus 指标
# vLLM 暴露 /metrics 端点
curl http://your-server:8000/metrics | grep vllm

// 降级配置示例
{
  "provider": "openai-compatible",
  "model": "qwen2.5-coder:32b",
  "apiBase": "http://localhost:11434/v1",
  "apiKey": "ollama",
  "fallback": {
    "enabled": true,
    "provider": "openai",
    "model": "gpt-4o-mini",
    "trigger": "function_call_unsupported"
  }
}