Context overflow permanently bricks a conversation; /compact and follow-up messages keep failing

Summary

When a Codex conversation reaches the context limit, the conversation becomes permanently unusable. Every subsequent message returns the same context-overflow error, including /compact.

Reproduction

1. Use a long-running Codex conversation with tool/search output.
2. Continue until Codex reports:

Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying.

1. Send a short follow-up message, such as 繼續.
2. Try /compact.

Actual behavior

• The original conversation remains stuck.
• Short follow-up messages still fail with the same context-overflow error.
• /compact does not recover or compact the conversation.
• In one occurrence, the preceding error was also:

stream disconnected before completion: Incomplete response returned, reason: max_output_tokens

• The only practical workaround is abandoning the conversation and starting a new thread.

Expected behavior

After an overflow, Codex should either:

• automatically compact the history before accepting the next message;
• allow /compact to run against the oversized conversation; or
• create a resumable compacted checkpoint/new thread instead of leaving the original thread permanently dead.

Scope / evidence

This has been observed in multiple conversations (three reported by the user; two matching local session records were independently confirmed). A desktop screenshot shows /compact followed by the same context-overflow error. The issue is reproducible at the conversation-recovery layer and is not caused by the new prompt being long.

Environment

• Codex Desktop / VS Code surface
• Observed July 15-18, 2026
• Local LM Studio/OpenClaw context settings were 60K/64K; changing those did not restore the dead Codex conversation.

No private logs or credentials are included.