Repository navigation
“stream disconnected before completion” + “input exceeds context window” loop #4116
Description
What version of Codex is running?
0.40.0
Which model were you using?
gpt-5-codex (medium)
What platform is your computer?
Darwin 24.6.0 arm64 arm
What steps can reproduce the bug?
For me, everytime it completes a yarn test command, it seems to trigger this issue. It likely tried to read the entire output of yarn test and caused a soft lock of the conversation.
What is the expected behavior?
It shouldn't cause a soft lock of the current conversation, as I am not able to run /compact or any other command to avoid losing all context.
What do you see instead?
⚠️ 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 197ms…
⚠️ stream error: stream disconnected before completion: Your input exceeds the context window of this model. Please
adjust your input and try again.; retrying 2/5 in 374ms…
⚠️ stream error: stream disconnected before completion: Your input exceeds the context window of this model. Please
adjust your input and try again.; retrying 3/5 in 869ms…
⚠️ stream error: stream disconnected before completion: Your input exceeds the context window of this model. Please
adjust your input and try again.; retrying 4/5 in 1.634s…
⚠️ stream error: stream disconnected before completion: Your input exceeds the context window of this model. Please
adjust your input and try again.; retrying 5/5 in 3.042s…
■ stream disconnected before completion: Your input exceeds the context window of this model. Please adjust your input
and try again.
Additional information
No response
Activity
robpresult-spec commentedon Sep 24, 2025
I get the same thing - stream error: stream disconnected before completion: Your input exceeds the context window of this model. Please adjust your input and try again.; retrying 2/5 in 434ms…
I try a different model - same thing. Try ChatGPT5 Account - Same thing - Try with Api key - same thing. ouch
scottyleics commentedon Sep 24, 2025
+1 here, same errors, makes the new model pretty much unusable now. Downgrading to GPT5 regular is not a solution as its terrible at coding work.
zjsng commentedon Sep 24, 2025
For anyone else having the same issue, I made a workaround for the problem by saving in
AGENTS.mdan instruction to output the problematic terminal commands into a file and then reading itxiejesses commentedon Sep 25, 2025
+1, This issue is too affecting work.
codedevmk commentedon Sep 27, 2025
same issue here, it says that I've exceeded my context yet when I check rate usage 5h limit 0% used · resets 11:13 PM
and Weekly limit 10% used · resets Oct 1.
Switched to norma and codex thinking modes without success, changed to prerelease and back to latest and there is no difference, this extension is completly locked which makes no sense given my context usage.
roblevine commentedon Sep 28, 2025
+1, model is pretty unusable for detailed work right now.
sebastiananders commentedon Sep 28, 2025
Just encountered this. So while this issue persists I gonna use Claude Code CLI.
raffdp commentedon Sep 28, 2025
It should auto compact before blowing up!
blixt commentedon Oct 2, 2025
I got this despite having a lot of context left, so my guess is the most recent tool call had a result so huge it immediately crossed the limit. Using
/compactdidn't resolve this because it seems the tool result is still included in that flow.So it would be nice to be able to
/undoor something which would take me back one step in the chat history, so I can keep using my 1 hour session by telling the AI how to proceed differently instead of just losing everything that was built up in that time...(I know transcript has "edit prev" which lets me roll back my own messages, but that would roll back the majority of my session since Codex was in control for the most part.)
BWBama85 commentedon Oct 3, 2025
I have not hit my rates and I am getting this:
This is incredibly frustrating and making me want to move outside of ChatGPT.
HenryAlexis commentedon Oct 3, 2025
Same issue here! I haven't reached my rates yet.
stalyndc commentedon Oct 4, 2025
same issued. please fix.
denispcguy commentedon Oct 4, 2025
In my case, pytest output was blowing up all the limits with huge 23k line output. I solved it by adding this to agents.md:
pytest_output.txt(e.g., append> pytest_output.txtor| tee pytest_output.txt) and read results from that file.pytest_output.txtso it only contains fresh output.stalyndc commentedon Oct 4, 2025
Can some conform if this work?
bitflower commentedon Oct 5, 2025
Same here
computa commentedon Oct 6, 2025
rexnetpl commentedon Oct 6, 2025
The same problem
vandocorreia commentedon Oct 6, 2025
same here
fullofcaffeine commentedon Oct 7, 2025
Same here. This is the only thing keeping me from fully switching to Codex. Such a pity: great model, but the UX is too rough right now. Claude Code is miles ahead in that regard.
amittell commentedon Oct 7, 2025
Same problem here, seems unrecoverable once it happens it requires a fresh conversation.
Also reported in issue #2921
wds666666 commentedon Oct 11, 2025
What a stupid bunch openai is, why don't you know the correct answer to the reference Claude?
etraut-openai commentedon Nov 18, 2025
I think this issue is now addressed. If you are still seeing this with the latest version of codex, please open a new issue with repro steps and relevant details.