openai / openai/codex

Preserve authored commentary after tool activity; final answers can be less engaging than the interim prose they replace

Open
#42,017 5 comments 1 reaction 0 assignees View on GitHub

Nobody has claimed this yet.

app bug session
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

Summary

During a single Codex Desktop turn, the assistant sometimes displays a polished, personable commentary message before performing a tool action. When the UI transitions into statuses such as “Planning” or “Reading,” that authored message is replaced or collapsed. The later final answer may preserve the operational result while losing the humor, phrasing, or personality that made the earlier message enjoyable.

This does not generally prevent task completion. It intermittently makes the completed interaction less engaging than what the user already saw during the turn.

Observed behavior
  1. The assistant emits a complete user-facing commentary message.
  2. The message is visible briefly and may contain useful context, humor, or distinctive phrasing.
  3. Tool activity begins and the same area changes to a planning, reading, or similar activity display.
  4. The final answer appears.
  5. The original commentary is no longer readily visible, while the final answer may be flatter or less memorable.

A sanitized local event inspection confirmed that the original prose remained persisted as a completed AgentMessage with phase: commentary, followed by tool activity and a separate final_answer. This suggests a presentation or transcript-retention issue rather than context compaction.

Expected behavior

Completed authored commentary should remain accessible after the turn, perhaps as an expandable progress item. Another possible solution would carry salient user-facing prose into the final answer before collapsing the interim display.

Ephemeral activity labels and intentionally authored commentary should receive different retention treatment.

Impact
  • Enjoyable or unusually apt language can disappear after the user has already read it.
  • The final transcript may underrepresent the quality and personality of the actual interaction.
  • Long-running specialist or personal-assistant workflows feel slightly less continuous.
  • Users may reasonably interpret the replacement as the assistant revising itself downward.

The user described this as an intermittent reduction in cleverness and entertainment value, rather than something that ruins the overall experience.

Environment
  • Codex Desktop 26.825.51511, build 7377
  • macOS 26.6.2, build 25G83
Privacy

No private transcript, health information, names, task identifiers, or local paths are included. Sanitized event-shape evidence can be supplied privately if requested.

Related

Related to phase handling, though distinct from the SDK-specific problem in #30190.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by tracing phase handling and the persisted AgentMessage with phase: commentary through the tool-activity and final_answer flow. Done means authored commentary remains accessible after the turn, or its salient prose is retained, while ephemeral activity labels receive different retention treatment.

Written by the indexing model from the issue text.

Assessment

Domain
desktop
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
48/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.