这是我实测较为完美的方法，codex不会报错，可以选文件，websearch，自动审查都可以用

Codex → Grok 接入配置

架构概览

Codex → 本地 CPA 网关 (127.0.0.1:8317) → CPA xAI OAuth 账号池 → xAI Grok API

1. Codex 主配置

文件路径： /root/.codex/config.toml

model_provider = "grok"

model = "grok-4.5"

model_catalog_json = "/root/.codex/model_catalog.json"

model_reasoning_effort = "medium"

disable_response_storage = true

personality = "pragmatic"

approvals_reviewer = "auto_review"



[model_providers.grok]

name = "grok"

base_url = "http://127.0.0.1:8317/v1"

wire_api = "responses"

requires_openai_auth = true

2. API Key

文件路径： /root/.codex/auth.json

{

  "OPENAI_API_KEY": "sk-xxxxxxx"

}

说明： 此 key 对应 CPA config.yaml 中的 api-keys，非 xAI 官方 key。

3. 模型元数据

文件路径： /root/.codex/model_catalog.json

这里仅列出需要增加的 grok-4.5 模型，仅供参考，可以让codex自己搜索信息并配置

{

  "models": [

    {

      "slug": "grok-4.5",

      "display_name": "Grok 4.5",

      "description": "Grok 4.5 via custom OpenAI-compatible provider",

      "default_reasoning_level": "medium",

      "supported_reasoning_levels": [

        { "effort": "low", "description": "..." },

        { "effort": "medium", "description": "..." },

        { "effort": "high", "description": "..." },

        { "effort": "xhigh", "description": "..." }

      ],

      "context_window": 256000,

      "max_context_window": 256000,

      "use_responses_lite": false,

      "supports_reasoning_summaries": true,

      "input_modalities": ["text", "image"],

      "base_instructions": "<很长的 Codex system prompt>",

      "model_messages": {

        "instructions_template": "<模板>",

        "instructions_variables": {

          "personality_default": "",

          "personality_friendly": "...",

          "personality_pragmatic": "..."

        },

        "approvals": null

      }

    }

  ]

}

配置要点

关于自动审查，需要在CPA里面 codex-auto-review 映射到 grok-4.5，自己审查自己