resume on long thread can ignore newer user redirection and repeat an older conclusion

Summary

After resuming a long Codex thread, Codex can ignore a newer user redirection and repeat an older conclusion from earlier in the same thread.

This did not look like a normal bad answer. It looked like the resumed thread was still anchored to an older internal task state, even though the newer user instruction was present in the visible transcript.

Versions

• Current CLI on this machine: codex-cli 0.136.0
• The affected thread was originally created on: codex-cli 0.134.0

Affected thread

• Thread id: 019e8156-671c-7b10-89c9-f7f265b05c29

Expected behavior

After resume, Codex should follow the latest user instruction and continue from the newest task direction.

Actual behavior

Codex repeated an older review/build-blocker conclusion instead of cleanly following the newer user redirection.

The rough pattern was:

1. We were discussing one UI/data-flow issue.
2. I explicitly redirected the task:

• first clean up the other changes
• then submit
• only come back to that feature later

1. I then clarified the new goal again:

• fix all currently known issues
• do cross review
• only submit after review passes

1. Instead of switching cleanly to that newer direction, Codex repeated the earlier "cannot submit, build is still red / migration is incomplete" conclusion multiple times.

Why this seems like a resume/continuation issue

This was a long-running thread and included:

• aborted turns
• goal continuation
• at least one compaction event
• resume/continue style continuation

The newer user instructions are clearly present in the transcript, but the assistant still falls back to an older conclusion pattern.

Evidence from local transcript

Local transcript file:

~/.codex/sessions/2026/06/01/rollout-2026-06-01T12-59-43-019e8156-671c-7b10-89c9-f7f265b05c29.jsonl

Relevant points from my local copy:

• User redirects task at line 4499
• Chinese text: 嗯， 你先把其他功能都整理下， 提交，然后再开始这个功能

• User clarifies the new goal again at line 4539
• Chinese text: 同步一下， 把这些问题全收了， 收了之后交叉评审，review 过了再提交

• Assistant repeats the older review/build conclusion at:
• line 4553
• line 4555

• Earlier compaction event exists in the same thread at line 2233

Additional note

This is not just “the model gave a bad answer once”.

The issue is that after resume, the thread appears to remain anchored to an older internal task state more strongly than the latest visible user redirection.

Possible contributing factors

This may involve one or more of:

• resume state restoration
• active goal continuation
• compaction summary handoff
• aborted turn recovery
• older thread-state anchoring after resume

Repro quality

I do not yet have a tiny isolated repro, but I do have a full local transcript where the behavior is visible and timestamped.

If useful, I can try to reduce it to a smaller repro based on:

• long thread
• explicit user task redirection
• resume/continue
• compaction or goal continuation
• assistant re-emits an older conclusion instead of following the latest instruction