Codex CLI: stream disconnected before completion: stream closed before...

What it means

The response stream (SSE) from the model disconnected after the HTTP handshake succeeded but before the server emitted response.completed. Per the Codex source, the session loop treats this as a transient error and automatically retries the turn; the TUI shows retry attempts before giving up. The retry count for SSE streaming interruptions is model_providers.<id>.stream_max_retries (default: 5) and the idle timeout is stream_idle_timeout_ms (default: 300000).

Causes

• Since Codex 0.36.0 the OPENAI_API_KEY environment variable is no longer read implicitly; users relying only on the env var hit this error (maintainer aibrahim-oai on #1581: "As of 0.36, we don't support implicit login.").
• Very long sessions: maintainer tibo-openai asked affected users on #3924 whether they were re-using one long session and whether a new chat still reproduced it.
• An over-limit context previously surfaced as this stream error; maintainer etraut-openai closed #3924 saying the "Your input exceeds the context window of this model" error is now properly handled.

How to fix it

1. 1
   Fix 1.

   If you use an API key, log in explicitly so the key is stored in auth.json (maintainer tibo-openai on #1581, and the official auth docs).

   Snippet

   Copy

   printenv OPENAI_API_KEY | codex login --with-api-key

2. 2
   Fix 2.

   Start a new chat or session instead of re-using one very long session (maintainer tibo-openai on #3924).

3. 3
   Fix 3.

   Update to the latest Codex release; maintainers closed #3924 after the context-window error was handled properly, and asked anyone still affected to open a new issue with details.

4. 4
   Fix 4.

   Optional tuning documented in the configuration reference: retry and idle-timeout settings for a model provider.

   TOML

   Copy

   request_max_retries = 4
   stream_max_retries = 10
   stream_idle_timeout_ms = 300000