Allow the models to retrieve user and assistant messages from the current thread after compaction for richer context

What variant of Codex are you using?

CLI

What feature would you like to see?

After one or more context compactions, earlier user and assistant messages remain stored and visible in the TUI, but the model can no longer directly retrieve exact details that were omitted from its compacted context. This can cause codex to forget decisions or constraints, repeat questions, or act from an incomplete summary.

I'd like codex to give the model bounded, on-demand access to earlier conversation in the current thread.

Retrieval should use the existing canonical persisted rollout rather than injecting the entire history into every inference request or duplicating conversation storage.

For example, if the user selected architecture A before compaction, Codex could search the current thread for that decision and read the exact earlier exchange instead of asking again or relying on an incomplete summary.

Additional information

A working prototype is available here:
https://github.com/shouryamaanjain/codex/pull/1