Codex

unexpected status 401 Unauthorized

First look at the address after url: at the end of the error.

If it is api.openai.com or chatgpt.com: your custom provider isn’t taking effect, and requests still go to the official service. Check each of these:

• Does config.toml have the line model_provider = "sub2api", and does the name match [model_providers.sub2api] exactly (case-sensitive)?
• Are these lines below a [...] section? Lines placed under a section are read as settings of that section. Move them to the very start of the file.
• Is the provider named openai? That is a name built into Codex, and a custom address is ignored for it. Use another name.
• Did you write the config in the project’s .codex/config.toml? A provider can only be defined in ~/.codex/config.toml.
• Is there a leftover OPENAI_BASE_URL in your environment variables?

If it is 2api.dev: the request reached 2api, so the problem is the key.

• ~/.codex/auth.json still holds ChatGPT sign-in information. Back it up, delete it, and write the key again (see Codex CLI).
• The key is wrong, or its group isn’t a GPT group.
• Run codex login status to see which sign-in method is in use.

After you change anything, fully quit Codex and start a new session.

stream disconnected before completion

• Check that base_url is https://2api.dev/v1.
• A dead proxy is set in the terminal. Run env | grep -i proxy to check. Fix the proxy, or add NO_PROXY=2api.dev.
• The connection drops during long thinking: add stream_idle_timeout_ms = 600000 under [model_providers.sub2api].
• A tool such as CC Switch broke the config: write it again as described in Codex CLI.
• If it happens with any model at any time, contact support and include the time.

Every new conversation gets stuck on Reconnecting… 1/5 to 5/5

Codex tries a WebSocket connection first. If your network or proxy doesn’t support it, it retries 5 times before falling back to a normal connection. Delete supports_websockets = true from config.toml, and responses_websockets_v2 = true under [features] (or set both to false). Then fully restart Codex.

Model not supported, or reasoning level not supported

• Model "xxx" is not supported by any configured account in this group: this group doesn’t have the model. Use the model name shown under Use Key (使用密钥) in the console, or look it up with /v1/models.
• Unsupported value: 'xxx' is not supported with the 'yyy' model: the model doesn’t support the reasoning level you set. Change model_reasoning_effort, or delete that line.
• To try another model temporarily: codex -m model-name -c model_reasoning_effort="high"

wire_api = “chat” is no longer supported

New versions of Codex only support the Responses protocol. Change wire_api to "responses".

The VS Code extension says Missing environment variable, or keeps spinning

GUI editors can’t see variables you export in a terminal. Write the key into auth.json (see Codex CLI) and don’t rely on environment variables. Then fully quit the editor and open it again. When the config file has a format error, the extension quietly falls back to default settings. Run codex once in a terminal to see the error with its line number.

Session history disappeared after changing base_url

Session history is tied to the provider name. Keep model_provider = "sub2api", [model_providers.sub2api] and name = "sub2api" identical in all three places. Change only base_url and the key. Don’t rename the provider.

Does signing in to Codex again lose my sessions?

No. As long as you don’t delete ~/.codex/sessions, resetting the sign-in or changing the config doesn’t delete your session history.

Only “exceeded retry limit, last status: 429” is shown

Codex hides the real cause. Run the self-check command to see the full error, then compare it with the kinds of 429 errors.