Windows desktop: chat pause leaves Goal auto-continuing with rapid empty final answers #44735

What version of the Codex App are you using (From “About Codex” dialog)?

Windows desktop app, observed on 2026-09-11. Exact app build not verified; the read-only Appx version query was unavailable in the current PowerShell environment.

What subscription do you have?

Not included in this report. No claim is made about measured token charges or quota exhaustion.

What platform is your computer?

Windows. The current read-only runtime query returned Microsoft Windows NT 10.0.19044.0.
The affected app UI showed GPT-6 Astra with high reasoning.

What issue are you seeing?

After the user instructed a long-running task to pause its work, it stopped substantive work but continued producing rapid empty turns. The user identified that the persistent Goal had not been stopped and was continuing to restart the task.

This report concerns a chat-level pause request, not a claim that a persisted paused Goal was independently verified and ignored.

Independent read-only task inspection observed 10 consecutive completed turns between 2026-09-11 13:45:23 and 13:46:07 Asia/Shanghai (UTC+8), corresponding to 05:45:23–05:46:07 UTC. Each had:

• one agentMessage with phase: "final_answer" and text: "";
• no tool-call items in the returned turn records;
• status: "completed", error: null;
• duration approximately 3–7 seconds.

Exact returned durations, oldest first: 4275, 3905, 4013, 2953, 3348, 3082, 6395, 4016, 5336, 6933 ms.

The first inspection found the task active with a new turn in progress; a subsequent inspection returned idle after the empty turns. The supplied screenshot showed a running Goal, “Thinking”, and no substantive response. Thus, the visible running state did not demonstrate actual task progress.

What steps can reproduce the bug?

Observed workflow (not intentionally rerun after discovery):

1. Run a long-running task with an active persistent Goal.
2. In chat, instruct the agent to pause work and wait for the user.
3. The agent honors the instruction at the work/action level, but the Goal's automatic continuation is not stopped.
4. Observe repeated new turns without user resume; in this incident those turns ended with empty final answers rather than useful work.
5. Inspect recent turn records to distinguish actual execution from a stale “Thinking” indicator.

The initiating pause wording and a before/after persisted Goal-state trace have not been independently extracted for this report. The user confirmed the pause/Goal mismatch; the repeated empty turns were independently read back. Scheduler internals and the precise lower-level cause remain unverified.

What is the expected behavior?

• Honoring an explicit request to pause Goal work should also suspend automatic continuation, preserving the objective and progress until an explicit resume.
• If the agent cannot pause the Goal through its available tools, it should clearly tell the user that the Goal is still active and needs a supported pause action—not silently imply it has stopped.
• Repeated empty final answers with no tool activity should not create an unbounded continuation loop; provide a circuit breaker and a clear actionable UI state.
• Do not require falsely marking an unfinished Goal complete or blocked as a substitute for pause.

Additional information

Related:

• /goal: chat "stop" message causes silent token burn until usage limit (autonomous tick ignores LLM-level self-restraint) #24531 describes the chat-level stop/active Goal mismatch in CLI and is closed as not planned. This report adds a Windows desktop incident with directly observed rapid empty final answers and the misleading running/Thinking presentation.
• Paused Goal still receives automatic continuation turns until marked blocked #37869 describes a different case: continuations after a Goal was independently confirmed persisted-paused. That stronger claim is not being made here.

Impact: repeated model turns without substantive progress while the user intends the work to remain paused; potential wasted usage and confusing status. Token charges and full loop duration were not measured.

No source code, private project paths, full conversation, or unredacted screenshot is attached. No task was resumed, cancelled, or recreated to reproduce this report.