mirror of
https://github.com/openai/codex.git
synced 2026-05-19 10:43:38 +00:00
Reload the active turn context before each sampling request so nudges and other mid-turn setting changes can take effect without mutating TurnContext in place. Move mid-turn compaction to the start of follow-up requests, skip separate diff injection when compaction re-establishes full context, and persist TurnContextItem baselines only when the active context actually changes. Also refresh the active turn context after override_turn_context, reset the reused model client session on mid-turn model changes, and key model-switch diffs off the persisted reference context so same-turn switch-backs still re-inject model instructions. Co-authored-by: Codex <noreply@openai.com>