Regression: auto-compacting triggers too late or fails on 5.5 extra high

Summary

Automatic context compacting appears to have regressed in Codex. It previously seemed fixed, but it is now failing again at least on the model/mode labeled 5.5 extra high.

The suspected issue is that compaction is being attempted too late, around ~95% context usage, instead of starting earlier with enough reserved budget, for example around ~80%.

Observed behavior

• In long-running Codex sessions using 5.5 extra high, compacting does not reliably occur before the session gets critically close to the context limit.
• When the session is already near the limit, compaction either does not work or does not preserve enough room for a clean continuation.
• This can interrupt coding tasks and make the agent lose track of unfinished verification/deployment state.

Expected behavior

• Codex should proactively compact while there is still enough available context to summarize accurately and continue safely.
• A threshold closer to ~80% seems safer than waiting until ~95%, especially for high-context modes with large tool outputs.
• The system should preserve current task state, pending actions, file/repo context, and verification status after compaction.

Scope

• Confirmed/observed by user on: 5.5 extra high.
• Unknown whether lower modes or other reasoning settings are affected.

Suggested fix direction

• Move the automatic compaction trigger earlier, or reserve a larger guaranteed buffer before compaction starts.
• Consider making the threshold model/mode-specific if extra high consumes context differently.
• Add telemetry or diagnostics around compaction trigger point, available remaining context, and whether continuation state was preserved successfully.

Impact

This affects long-running coding workflows because the session can fail or become unreliable exactly when the agent needs to preserve task state across implementation, verification, and follow-up steps.