Codex Stuck on Reconnecting — How to Fix It

Step 1: 30-second diagnosis — network or bug?

Reconnecting means the connection between Codex and OpenAI (a WebSocket / SSE stream) failed to establish or stay alive. Run this quick check before trying anything else:

# Can you reach OpenAI directly? A 401 means the connection works — just no key attached.
$ curl -I https://api.openai.com/v1/models

# Now test through your proxy
$ curl -I -x http://127.0.0.1:7890 https://api.openai.com/v1/models

• Both time out / connection refused → Network problem. Focus on the "Proxy" section below.
• Direct times out, proxy returns 401/200 → The proxy works but Codex isn't using it. See "Proxy not working".
• Both return 401 but codex still reconnects → Likely an auth conflict or version bug. Jump to those sections.

Cause 1: no proxy or wrong proxy (most common outside the US)

Networks that block OpenAI directly will cause Codex to reconnect endlessly the moment it starts. The fix is to route Codex through a proxy — it reads standard proxy environment variables:

export HTTPS_PROXY="http://127.0.0.1:7890"
export HTTP_PROXY="http://127.0.0.1:7890"
export ALL_PROXY="http://127.0.0.1:7890"

After editing, run source ~/.zshrc (or reopen the terminal) and then launch codex. Match the port to your proxy client: Clash commonly uses 7890, some clients use 7897.

Proxy configured but still not working?

• Confirm you set the variable and launched codex in the same terminal session — a new terminal won't inherit a temporary export.
• Check for a lowercase variable (https_proxy) overriding your setting: env | grep -i proxy.
• On corporate networks, make sure a NO_PROXY allowlist isn't excluding the relevant domains.

Cause 2: socks5 proxy (Codex doesn't support it)

A very common trap: Codex does not support socks5 proxies. If your ALL_PROXY is set to socks5://..., connections will fail silently and Codex will loop on Reconnecting. Switch to the HTTP port exposed by your proxy client, or use a tool to convert socks5 to HTTP:

# ✗ Don't use this
export ALL_PROXY="socks5://127.0.0.1:7891"

# ✓ Use the HTTP port instead (most clients expose both)
export HTTPS_PROXY="http://127.0.0.1:7890"

Cause 3: OAuth and API key both present (auth conflict)

When you have both signed in with a ChatGPT account (OAuth) and have OPENAI_API_KEY set in the environment, the two auth mechanisms coexist in an unstable way and can cause erratic reconnects. Clear one of them and re-authenticate:

1. Sign out and clear cached credentials:
   Clear authcopy

   $ codex logout
   $ rm -f ~/.codex/auth.json
   $ unset OPENAI_API_KEY

2. Pick one auth method and sign back in — ChatGPT account: codex login; or API key only: codex login --with-api-key.
3. Run codex again and send a simple message to confirm everything works.

Cause 4: VS Code extension / WSL2 networking

Reconnecting inside an editor or WSL often has a different root cause from the terminal — the two environments use separate network configurations:

VS Code extension keeps Reconnecting

The extension doesn't read your terminal's export — you must configure the proxy in VS Code's own settings: press Ctrl+, to open settings, search for proxy, set Http: Proxy to http://127.0.0.1:7890, then fully restart VS Code.

Windows + WSL2 can't connect

The codex process running inside WSL2 does not automatically use the Windows system proxy. Two approaches:

• Enable mirrored networking in WSL2: this makes WSL share the Windows loopback, so the proxy port becomes reachable. In %USERPROFILE%\.wslconfig add [wsl2] / networkingMode=mirrored, then run wsl --shutdown and reopen WSL.
• Remember that socks5 is unsupported inside WSL2 as well — use the HTTP proxy port there too.

Cause 5: version bug or frozen session

If network and auth are ruled out and Codex still reconnects or freezes on Working, it's likely a known issue in a specific build. Handle in priority order:

1. Close all sessions and restart as a single instance (the most effective community fix): press Ctrl+C to terminate all codex processes, wait about 60 seconds for a full exit, open exactly one new session, and send HELLO to confirm recovery.
2. Upgrade to the latest version:
   Upgradecopy

   $ npm install -g @openai/codex@latest
   $ codex --version
   # Homebrew users: brew upgrade codex

3. Check MCP servers: a misbehaving MCP server can drag down the connection. Temporarily comment out any suspicious [mcp_servers.*] entries in ~/.codex/config.toml, or toggle experimental_use_rmcp_client.
4. Submit diagnostics: run /feedback inside a session — it attaches the request ID, error log, and connection snapshot, helping the team pin down the issue.

Last resort: clean reinstall

If nothing above works, do a clean reinstall (this clears your login — you'll need to sign in again):

$ npm uninstall -g @openai/codex
$ npm install -g @openai/codex@latest
# Optional: wipe config and credentials (use with caution)
$ rm -rf ~/.codex

Frequently asked questions