codex resume fails for specific sessions with "stream disconnected before completion", while new sessions work normally

What version of Codex is running?

0.73.0

What subscription do you have?

business

Which model were you using?

gpt-5.2 high

What platform is your computer?

Linux 3.10.0-957.el7.x86_64 x86_64 x86_64

What issue are you seeing?

When using codex resume to continue certain existing sessions, the session fails almost immediately with errors like:

Conversation interrupted - tell the model what to do differently.
stream disconnected before completion: Transport error: error decoding response body

However, starting a brand new session with codex or /new works reliably with the same model, network, proxy, and authentication.

Key observations:

• Only specific existing sessions are affected.
• New sessions created with the same Codex version and model work normally.
• The failure appears during session resume / history replay, not during normal request handling.
• The same environment can successfully make API calls to https://api.openai.com through the configured proxy (verified with curl).
• Killing and restarting local SSH tunnels does not change the behavior.
• Creating a new session and manually copying the relevant context from the old session works as a workaround.

This suggests a bug related to session resume, history replay, or context reconstruction for certain sessions, rather than a networking or authentication issue.

What steps can reproduce the bug?

Uploaded thread: 019b2b81-1f7e-7b23-ad5c-f7d3e5d10776

What is the expected behavior?

codex resume should reliably restore and continue an existing session, or gracefully fall back to a summarized context if the full history cannot be replayed.

At minimum, resuming a session should not fail due to corrupted or partial history data.

Additional information

• The problem is not resolved by restarting the CLI, SSH session, or proxy.
• The problem does not occur for newly created sessions.
• The issue appears consistent with a session-level bug rather than an account- or network-level issue.
• Workaround: start a new session and manually paste relevant context from the old session.

If helpful, I can provide affected session IDs or logs with sensitive information redacted.