Windows Desktop: scheduled heartbeat answers an older request in a long thread; user also reports UI lag

Summary

A daily heartbeat in a long Windows Codex Desktop conversation started on schedule but ended by answering an older feedback request instead of completing its active scheduled task. This is not a failure to fire the timer.

Observed timeline (UTC)

• 2026-09-07 23:00:35: scheduled task_started (07:00:35 Asia/Shanghai).
• Work on the first item was observed around 23:05–23:06.
• 23:11:47: assistant commentary switched to an older request to record feedback on a previous song.
• 23:12:34: task_complete answered that old feedback; the day's batch remained incomplete.
• Manual recovery later completed the batch without regenerating the existing first item.
  A similar older-request substitution was observed on September 6.

Expected

Preserve the active heartbeat objective across long-thread/context-compaction recovery. An unrelated historical request should not terminate the scheduled task. Actual newer user instructions should still take precedence.

Environment and limits

Windows, Asia/Shanghai, active daily 07:00 heartbeat bound to an existing long-running thread. Exact Desktop version unavailable from the read-only package query. The session log is large (~669 MB). This report does not establish the internal root cause or prove a compaction defect; it documents the wrong-task transition.

Mitigation tried

Reduced the automation prompt to one unambiguous entry point; removed contradictory old download gates; persisted separate current-day production/research states and recovery instructions. Next scheduled run has not yet validated the mitigation. No restart, runtime reinstall or permission weakening was used.

Related but not necessarily the same cause: #39191.

Only a sanitized timeline is included; no raw conversation, credentials, account identifiers or creative files are attached.

Additional user-reported performance symptom

The user also reports noticeable lag/stuttering in Codex Desktop. No CPU, memory, frame-time measurements or reliable reproduction steps have been captured yet. It is unknown whether this performance symptom is related to the long thread or the heartbeat failure; please do not treat that relationship as established.

Start with the sanitized UTC timeline in this issue and investigate the scheduled heartbeat's active-objective handling during long-thread or context-compaction recovery. No source file, test, or entry point is identified, so first establish a reproducible wrong-task transition while preserving precedence for newer instructions. Treat the reported UI lag as a separate symptom unless measurements show a connection.