mirror of
https://github.com/openai/codex.git
synced 2026-11-06 20:20:34 +00:00
Benjamin Carlsson
a142c90e26
## Why The TUI's cached active turn can be stale when a cloud message is submitted. A lost reply also leaves delivery uncertain, so automatically resending the draft risks submitting it twice. ## What changed - Let the cloud server choose whether an Enter submission starts or steers a turn, and apply its receipt to the matching client message ID. - Keep Tab-queued input waiting until the admitted turn finishes, including before its start notification arrives. - Restore rejected submissions to the composer. Hold duplicate or unconfirmed submissions out of automatic delivery until history confirms them. - Reconcile saved drafts and pending steers by client message ID when resuming sessions, and preserve submissions belonging to other turns during interruption recovery. - Classify cloud preflight failures as invalid parameters so rejected queued input can be restored safely. ## Testing Add regression coverage for cloud receipt outcomes, queued-input ordering, rejected drafts, lost replies, interruption recovery, and session resume with identical message text but distinct client IDs. GitOrigin-RevId: ec4eb581e450488434bc4614ec59273d73ffd0a1