stream disconnected before completion: stream closed before response.completed

This has been a randomly occurring issue for me as well since around 0.30. i think its a api issue but after about 2-3 retries it finally goes through.

This has been a randomly occurring issue for me as well since around 0.30. i think its a api issue but after about 2-3 retries it finally goes through.

Unfortunately, that's not my case. To this day, on my personal PC, I haven't been able to get it to respond with a single comma—it always gives me the error I mentioned above. And my login process is the classic one through the portal, without using the API or anything like that.

This bug is back in 0.42.0 on MacOs , it was fixed in the last few versions. Same place as well 30% context remaining, then:

⚠️ stream error: stream disconnected before completion: Your input exceeds the context window of this
model. Please adjust your input and try again.; retrying 1/5 in 194ms…

Experiencing this issue since yesterday 9 PM CEST.
VS Code with Codex Extension, Version 0.4.15 & 0.5.15

stream disconnected before completion: Your input exceeds the context window of this model. Please adjust your input and try again.

Changed models (gpt-5-codex-medium to low and gpt-5-minimal), switched mode from Agent to Chat.. nothing works.
Limit is not exceeded

Experiencing this issue since yesterday 9 PM CEST. VS Code with Codex Extension, Version 0.4.15 & 0.5.15 stream disconnected before completion: Your input exceeds the context window of this model. Please adjust your input and try again. Changed models (gpt-5-codex-medium to low and gpt-5-minimal), switched mode from Agent to Chat.. nothing works. Limit is not exceeded

This is a completely separate issue. AFAIK theree isnt a compact command for the extension so when this happens to to I just start a new conversation though now that I think about it you might be able to look for the conversation ID and then use codex-cli to /compact the conversation. I haven't tried but it might be worth experimenting with.

Adding for context, I encountered the same error message as the original post. The first two times I attempted to send through a request, I received the first error:

stream error: stream disconnected before completion: stream closed before response.completed

On the third attempt, it responded with:

stream disconnected before completion: Your input exceeds the context window of this model. Please adjust your input and try again.

A new chat resolved the issue - not sure why it took a 3 attempts to return the actual issue.

Adding for context, I encountered the same error message as the original post. The first two times I attempted to send through a request, I received the first error: stream error: stream disconnected before completion: stream closed before response.completed On the third attempt, it responded with: stream disconnected before completion: Your input exceeds the context window of this model. Please adjust your input and try again. A new chat resolved the issue - not sure why it took a 3 attempts to return the actual issue.

I wish I had your luck, I did another test here and I only get disconnection. Since the launch, I haven’t been able to use Codex even once. When it gets out of beta, I’ll give it another chance. I went back to Claude Code this month.

+1

I ran into this issue when I ran /compact quite close to the context window's limit, and it looked like it succeeded (and showed 100% context available), but then every follow on message had the stream closed error.

Resuming the session again showed the 100% context, but then I tried to run /compact again, and it shot back to 0%.

Resuming the session a second time, then showed 0% context, and the error persisted.

This happened on 0.45 -> 0.50.

Would be nice if we could "clean" the context a bit manually in these situations, for now we can extract some of the "lost" context via ~/.codex/sessions if we're desperate.

Note: _New_ sessions work as expected.

I'm going to close this issue because there's not enough information to repro or diagnose. If this happens again, please use the /feedback slash command to upload the session and post a thread ID.