Daily Edition Sources +2

Codex Keeps the Prompt After You Interrupt It

A new Codex TUI change treats an interrupted request as part of the record, then gives the next instruction a clean composer.

Hand-drawn diagram showing an interrupted prompt staying in history, a new blank composer, and evidence cards for interrupt keys, safety buffering, and thread snapshots.
Diagram Punkan interrupt leaves a record, not a forced retry.
repo openai/codex evidence
2 source signals 1 repo 2 linked commits
Evidence: 2 linked commits / July 16, 2026 / Daily Edition
Open Edition Evidence below

Interrupting a coding agent is a small act with large consequences. A person may have stopped it because the plan was wrong, the command looked risky, or a better instruction came to mind. On July 15, openai/codex changed its TUI so an output-free prompt interrupted with Esc or Ctrl-C remains in conversation history while the next composer opens blank.

That is deliberately different from putting the old request back into the editor and rolling the last turn away. The new tests cover both interruption keys, a failed safety-buffer retry, and restoring the matching buffered prompt when a thread snapshot is replayed. The interaction says: the attempt happened, but the next turn is a new decision.

Stopping is information, not cleanup

The change removes the cancelled-turn edit path and separates safety-buffering state so that a retry failure can still restore the right prompt after a thread switch. In a tool that can run commands and make changes, that makes interruption legible rather than pretending an abandoned request never existed.

This also sharpens the boundary from yesterday's Codex history work. The earlier commits made a durable record and a recoverable reading view; this one decides what an interrupted user request means inside that record. A session can show the stop without forcing the user to resume the same wording.

A better next question

This is not a guarantee that an interrupted operation has been cancelled everywhere, nor a promise about all Codex surfaces. It is a tested TUI behavior. The next useful signal is whether interruption status becomes equally explicit for work already delegated to a tool or subagent—where “I stopped asking” and “the work stopped” are not always the same event.

Evidence Trail

Receipts below the story

The article above is the public narrative. This section keeps the source trail and limits on the same page.

Edition
DateJuly 16, 2026
LaneDaily Edition
Confidence68%
Sources2
Reposopenai/codex

Primary Evidence

Evidence Limits

  • These commits establish tested behavior in Codex's TUI path, not a claim that every Codex client or an already-running tool has been cancelled.
  • The public sources do not measure how often users interrupt work or why they do it.
Letters & Corrections

Send a note to the desk

Corrections, missing context, or a follow-up lead.