Context compaction loop leaves ~5% context and stalls session (eventually 0% / interrupted)

Resolved 💬 3 comments Opened Dec 20, 2025 by PauloNRocha Closed Mar 29, 2026

💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

What version of Codex is running?

codex-cli 0.76.0

What subscription do you have?

plus

Which model were you using?

gpt-5.2-codex

What platform is your computer?

Linux 6.8.0-87-generic x86_64 x86_64

What issue are you seeing?

Codex enters a context-compaction loop. After triggering compaction, the session repeatedly shows "Context compacted" and the available context never recovers beyond ~5%. The model stops making progress and eventually the session ends with "Conversation interrupted" and "0% context left".

What steps can reproduce the bug?

Start a longer Codex CLI session inside a git repository.
Iterate until the context becomes large enough to trigger automatic compaction.
Observe repeated "Context compacted" messages.
After each compaction, usable context remains around ~5% and the session cannot proceed.
The session eventually terminates with "Conversation interrupted / 0% context left".

Uploaded thread: 019b038b-ec3e-7dd0-b9ae-814562d486a9

What is the expected behavior?

After a context compaction, the session should restore enough usable context to continue the task, or fail gracefully with a clear instruction to restart. It should not enter a repeated compaction loop with near-zero available context.

Additional information

Codex CLI version in use is 0.76.0 (per codex --version). However, the uploaded rollout metadata reports cli_version: 0.66.0, which might indicate a backend/worker version mismatch. Not sure if related, but flagging in case it helps.
Environment: Linux, bash, network_access=restricted, sandbox_mode=workspace-write.

Impact: session becomes unusable (no progress) once the loop starts.

View original on GitHub ↗

3 Comments

etraut-openai contributor · 8 months ago

Thanks for reporting. If you still have access to that session, please resume it and use the /feedback slash command to upload it and then post the thread ID here.

PauloNRocha · 8 months ago

I was able to access the original session and used /feedback from that session to upload the logs. The same thread ID was reused, since it’s the same session.

Thread ID: 019b038b-ec3e-7dd0-b9ae-814562d486a9

Note: the thread ID is the same as previously mentioned because the upload was done from the same session.

etraut-openai contributor · 5 months ago

This bug report hasn't received any upvotes or additional posts in three months. Closing.