OpenClaw Codex GPT-5.4 Enable

Enable `openai-codex/gpt-5.4` in OpenClaw through a config-layer patch instead of rebuilding the app. Use when `openai/gpt-5.4` already works but Codex GPT-5...

Security Scan

The skill's steps to patch ~/.openclaw/openclaw.json are coherent with its purpose, but there are a few inconsistencies and a risky recommendation (an external baseUrl) that you should review before applying the change.

Purpose & Capability

Name and description match the instructions: the skill documents how to add a provider/model/alias/fallback to ~/.openclaw/openclaw.json and verify with openclaw commands. Requiring the openclaw binary only is appropriate.

Instruction Scope

Instructions are scoped to editing the user's OpenClaw config and running openclaw CLI/ session_status. They do not ask to read unrelated system files. However the recommended provider baseUrl (https://chatgpt.com/backend-api) points model traffic to a third-party gateway — that is outside the skill's stated 'enable locally' reassurance and could send user data to an external service.

Install Mechanism

Instruction-only skill with no install steps or code; nothing is written to disk by the skill itself beyond user-applied config edits. This is the lowest install risk.

Credentials

The example JSON contains apiKey: "$OPENAI_API_KEY" but the skill's declared requirements list no environment variables. The skill therefore references a credential (OPENAI_API_KEY) without declaring it — a mismatch users should be aware of. Also, directing requests to a non-OpenAI domain (chatgpt.com) combined with an API key could cause sensitive model traffic or keys to be sent to third parties.

Persistence & Privilege

Skill does not request persistent or elevated privileges and does not set always:true. The workflow edits only the user's ~/.openclaw/openclaw.json (user-level config) and provides a rollback procedure.

What to consider before installing

This skill documents a manual, configuration-layer way to enable openai-codex/gpt-5.4 and is otherwise coherent, but review two things before applying it: (1) the example JSON uses $OPENAI_API_KEY even though the skill doesn't declare any required env vars — ensure you understand which API key will be used and never paste secrets into shared or untrusted files; (2) the suggested baseUrl (https://chatgpt.com/backend-api) is a third-party gateway — unless you explicitly trust that endpoint, point the provider to your trusted gateway (for example an official API host) because model requests and keys could be routed off your system. Always back up ~/.openclaw/openclaw.json, test changes in a separate session, and verify that model traffic is going to a host you control or trust.

OpenClaw Codex GPT-5.4 Enable Skill

把 openai-codex/gpt-5.4 用“配置层补丁”方式接入 OpenClaw，避免直接改安装目录或整包重编译。

适用场景

当你遇到下面这种情况时使用：

• openai/gpt-5.4 已经在 OpenClaw 中可见
• 但 openai-codex/gpt-5.4 仍然显示 not allowed / missing
• 想快速验证 GPT-5.4 Codex 通道，而不是先改 Homebrew 安装产物
• 希望把 provider、模型、别名、fallback 一次补齐

核心思路

优先改 ~/.openclaw/openclaw.json：

1. 在 models.providers 中新增 openai-codex
2. 设置：
   • baseUrl: https://chatgpt.com/backend-api
   • api: openai-codex-responses
3. 为该 provider 补上模型 gpt-5.4
4. 在 agents.defaults.models 和 agents.defaults.model.fallbacks 中补上 openai-codex/gpt-5.4
5. 给它加一个别名，例如 GPT54Codex
6. 用 openclaw models list 和 session_status(model='openai-codex/gpt-5.4') 验证

为什么优先走配置层

• 风险更低：不碰安装目录内打包 JS
• 回滚更简单：只需恢复一份 JSON 配置
• 更适合快速试验和迁移
• 对已经在跑的 OpenClaw 实例更友好

目标文件

~/.openclaw/openclaw.json

建议操作前先备份：

cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.bak.$(date +%Y%m%d-%H%M%S)

参考配置结构

字段名可能随 OpenClaw 版本略有变化，请以你当前 openclaw.json 结构为准，把下面内容合并进去。

{
  "models": {
    "providers": {
      "openai-codex": {
        "apiKey": "$OPENAI_API_KEY",
        "baseUrl": "https://chatgpt.com/backend-api",
        "api": "openai-codex-responses",
        "models": {
          "gpt-5.4": {
            "label": "GPT-5.4 Codex",
            "contextTokens": 400000,
            "maxOutputTokens": 128000,
            "modalities": ["text", "image"],
            "supportsReasoning": true
          }
        }
      }
    }
  },
  "agents": {
    "defaults": {
      "models": {
        "GPT54Codex": "openai-codex/gpt-5.4"
      },
      "model": {
        "fallbacks": [
          "openai-codex/gpt-5.4"
        ]
      }
    }
  }
}

实战流程

1) 先确认当前支持面

openclaw models list --plain | grep 'gpt-5.4'

判断点：

• 如果 openai/gpt-5.4 已经存在，说明基础支持大概率已经在
• 如果只有 OpenAI 版本，没有 Codex 版本，就继续补配置

2) 备份配置

cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.bak.$(date +%Y%m%d-%H%M%S)

3) 编辑 provider

在 models.providers 里新增：

• provider id：openai-codex
• baseUrl: https://chatgpt.com/backend-api
• api: openai-codex-responses

4) 补模型定义

给 openai-codex provider 增加 gpt-5.4 模型定义，至少补：

• label
• context/max tokens
• text/image modality
• reasoning 支持

5) 补 agent 默认映射

在 agents.defaults.models 里补别名：

{
  "GPT54Codex": "openai-codex/gpt-5.4"
}

在 agents.defaults.model.fallbacks 里补：

[
  "openai-codex/gpt-5.4"
]

6) 验证模型注册

openclaw models list --plain | grep 'openai-codex/gpt-5.4'

7) 验证网关是否真正接受

用 OpenClaw 的状态能力做快速实测：

# 在 OpenClaw 会话内
session_status(model='openai-codex/gpt-5.4')

成功标准：

• 不再报 not allowed
• 模型可被选中
• provider 认证链路走 openai-codex:default（视环境展示而定）

经验结论

已验证有效的判断逻辑

1. 先看官方 PR / 改动方向，确认是不是“版本已支持、只是本地没放开”
2. 如果核心 provider 能力已在，优先尝试配置层补齐
3. 用 session_status 验证，比只看静态配置更靠谱
4. 改完后如果上下文上限没有立刻刷新，优先开新会话再测

这套方法特别适合

• 想快速上手新模型
• 不想改安装目录
• 想沉淀成可复用 SOP
• 多机器迁移 OpenClaw 配置

排障

not allowed

• 检查 agents.defaults.models 是否已补别名
• 检查 fallbacks 是否已包含目标模型
• 检查 provider id 是否与模型前缀一致：openai-codex/...

模型存在但调用失败

• 检查 baseUrl 和 api 是否正确
• 检查对应认证是否已就绪
• 检查是否需要重启 OpenClaw Gateway / 新建会话

模型列表看不到

• 检查 JSON 是否写坏
• 检查 provider 是否放在正确层级
• 执行：

openclaw models list --plain | grep codex

回滚

cp ~/.openclaw/openclaw.json.bak.YYYYMMDD-HHMMSS ~/.openclaw/openclaw.json

然后重新启动或刷新相关服务。

推荐验证清单

• openai/gpt-5.4 已可见
• openai-codex provider 已加入
• gpt-5.4 模型定义已加入
• GPT54Codex 别名已加入
• fallbacks 已加入 openai-codex/gpt-5.4
• openclaw models list 能看到目标模型
• session_status(model='openai-codex/gpt-5.4') 成功
• 新会话验证通过

一句话版本

如果 openai/gpt-5.4 已经支持，而 openai-codex/gpt-5.4 还没放开，别急着重编译：先用 ~/.openclaw/openclaw.json 做 provider + model + alias + fallback 的配置层补丁，再用 session_status 验证。"