[Windows app] Work activity converts thought summaries into assistant commentary

App version

Microsoft Store package OpenAI.Codex 26.901.6511.0; desktop application 26.901.51231.

Subscription

Pro 20×

Platform

Windows 11 x64.

Issue

In a Work conversation, user-visible thought summaries appear as ordinary assistant progress text, sometimes joined to an adjacent progress report, instead of staying in the native activity disclosure.

Deterministic reproduction and root cause

The Zm adapter in the distributed viewer converts every reasoning item into an assistant-message with phase: commentary. It ignores the existing presentation: thought versus presentation: preamble distinction and may concatenate the converted content with the previous commentary item.

A synthetic sequence containing an ordinary preamble followed by a thought summary reproduces the conversion in the original function. The repaired adapter converts only explicit preambles; thought summaries retain their original reasoning type, completion state, and content for the native disclosure renderer. Normal progress reports remain visible.

This concerns summaries already supplied to the client for display. It does not request hidden model reasoning, alter CLI events, or infer event duplication from matching CLI footer text. #33033 discusses a related but different ordering symptom; this reproduction does not require a late event after the final answer.

Related finding in the same adapter

The activity group also renders all todo-list updates. Two snapshots can produce adjacent Step 2/3 and Step 3/3 badges. Selecting the last snapshot within that activity group removes the stale badge while preserving original data, other activity order, and plans in separate turns. This is distinct from a plan that never receives a status update, such as the symptom in #35118.

Validation

tests/work-activity.mjs executes the complete original and repaired adapter in twelve cases, including both reasoning presentations, live/completed states, plan updates, separate groups, and restricted/read-only activity containers. No live generation was invoked for these tests.

Reproducible implementation

Complete English source, English release, and upstream evidence guide. The original project is wjtianze/codex-desktop-repair; its main branch remains in Chinese.

The package contains checked patch fragments and synthetic regression cases, not complete proprietary client bundles or account data. Both language editions passed 242 tests: 76 portable cases and 166 cases executing functions extracted from the supported official app. This is a local repair and evidence submission, not an official source merge.

Start with tests/work-activity.mjs and the Zm adapter described in the issue. Run the twelve validation cases, then compare handling of reasoning items with presentation thought versus preamble and the adjacent todo-list snapshots. Done means thought summaries remain native reasoning disclosures, normal progress reports stay visible, and stale plan badges are removed without changing other activity data.