Back to skills
SKILL.md
Openclaw Codex Oauth Proxy
BSecurityOpenClaw 用 openai-codex(ChatGPT OAuth)跑 GPT-5.x Codex/Spark 的配置与排障(网关进程、代理、模型、thinking、JSON 输出)。
- 57 stars
- 0 votes
- 0 copies
- 0 views
- Added October 17, 2026
Works with
Security analysis
88/100- Exfiltrates credentials via HTTP — exact pattern from Snyk ToxicSkills study
npx -y skills add diegosouzapw/awesome-omni-skill --skill openclaw-codex-oauth-proxy --agent claude-codeAre you the author of Openclaw Codex Oauth Proxy?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/diegosouzapw-openclaw-codex-oauth-proxy)---
name: openclaw-codex-oauth-proxy
description: OpenClaw 用 openai-codex(ChatGPT OAuth)跑 GPT-5.x Codex/Spark 的配置与排障(网关进程、代理、模型、thinking、JSON 输出)。
---
# openclaw-codex-oauth-proxy
## 0) 先确认你走的是 `openai-codex`(不是 API Key 的 `openai`)
```bash
openclaw models status --plain
openclaw models status --probe --probe-provider openai-codex --json
```
看 `status=ok`,并且 provider 是 `openai-codex`。
## 1) “能登录 OAuth,但模型不可用”的常见根因:`api.openai.com` 网络不通
ChatGPT OAuth 的登录主要走 `auth.openai.com`,但模型调用通常仍需要连 `api.openai.com`。
用 **同一台机器**验证(期望返回 `401`,不是超时):
```bash
curl -m 8 -I https://api.openai.com/v1/models
```
如果直连超时,但加代理能通:
```bash
curl -m 8 --proxy http://127.0.0.1:7897 -I https://api.openai.com/v1/models
```
那就要让 **Gateway 常驻进程**也走代理(见下一节)。
## 2) 关键点:Gateway 是常驻服务,不会继承你终端里的代理环境变量
你在终端里能 `curl --proxy ...` 不代表 `openclaw-gateway` 进程也能出网。
把代理写进 `~/.openclaw/openclaw.json` 的 `env.vars`,然后重启网关:
```bash
openclaw config set env.vars.HTTPS_PROXY "http://127.0.0.1:7897"
openclaw config set env.vars.HTTP_PROXY "http://127.0.0.1:7897"
openclaw config set env.vars.ALL_PROXY "http://127.0.0.1:7897"
openclaw config set env.vars.NO_PROXY "127.0.0.1,localhost"
openclaw gateway restart
openclaw gateway status
```
如果没装服务:
```bash
openclaw gateway install
openclaw gateway restart
```
端口冲突/旧进程残留时,用:
```bash
openclaw gateway --force
```
## 3) 设置默认模型 + thinking(例如默认 `medium`)
```bash
openclaw models set openai-codex/gpt-5.3-codex
openclaw config set agents.defaults.thinkingDefault medium
```
切 Spark(如果 models list 里能看到):
```bash
openclaw models set openai-codex/gpt-5.3-codex-spark
```
验证实际生效的模型(看输出里的 `agentMeta.model`):
```bash
openclaw agent --agent main -m "ping" --thinking minimal --json
```
## 4) CLI `--json` 输出前面夹杂插件日志,别直接管道给 `jq`
有些命令会先打印类似 `[plugins] ...`,导致:
- `jq` 解析失败
- 甚至触发 Node CLI `EPIPE`(broken pipe)
稳妥写法:先落盘,再从第一个 `{` 截 JSON。
```bash
openclaw sessions --json > /tmp/sessions.json
python3 - <<'PY'
import json
raw=open("/tmp/sessions.json","r",encoding="utf-8",errors="replace").read()
raw=raw[raw.find("{"):]
print(json.loads(raw).keys())
PY
```
Attribution
Comments
Loading comments…
Newsletter
Get the weekly list of safe new skills, and the ones we pulled.