Codex desktop response-selection annotation control disappears mid-session

Summary

The response-selection annotation control in a Codex desktop task worked repeatedly, then disappeared from newer assistant messages during the same session. Response text and user-defined verse markers remained visible. The control has not returned after restoration attempts.

Environment

• Codex desktop app on Windows
• Date observed: 2026-09-08
• Long-running active task with many messages

Expected behavior

Selecting text in an eligible assistant response should expose the annotation control consistently, as it did earlier in the same task.

Actual behavior

The annotation control stopped appearing on newer assistant responses while the response text itself remained selectable/visible.

Recoverable boundary

• Confirmed working: response D9.211 at 4:58:12 PM PDT was successfully returned as a response annotation.
• D9.213 was also subsequently returned as a selected response annotation.
• First explicit absence report occurred after D9.213 and before/at D9.221 (5:09:52 PM PDT).
• Narrow boundary: KNOWN WORKING THROUGH D9.213 → FIRST EXPLICIT ABSENCE REPORT BEFORE/AT D9.221.
• Exact first affected message inside that interval is undetermined.

Repeated observations

At least seven user-reported absence occurrences/clusters are preserved. The first covered an unquantified “last few messages”; later reports explicitly followed test responses including D9.227, D9.233, D9.239–D9.240, D9.257, D9.264, and D9.265.

Recovery attempts

1. Refreshed an open ChatGPT browser conversation (later recognized as the wrong surface for the Codex-native control).
2. Identified and reopened the exact active Codex task through the Codex task navigator.
3. Restored the earlier plain-text response layout and generated harmless selection tests.
4. The user continued to report the control absent.

Impact

The missing control prevents precise selection-based feedback on individual response passages in a long provenance/lineage task and forces manual copying of marker IDs and text.

Evidence boundary

This report establishes user-observed mid-session interface behavior and repeated reproduction. Root cause, deployment timing, developer intent, and whether the element is absent or hidden are unknown.

Suggested diagnostic

Compare a known-working message (D9.213) with a later affected message in the same task, including whether the annotation element is omitted from rendering or present but hidden. Cross-client comparison of the same messages may help distinguish task/message state from client-local rendering state.

No source files, tests, or entry points are named. Start by comparing known-working response D9.213 with affected messages around D9.221 and later, checking whether the annotation element is omitted or hidden and whether the behavior differs across clients. Done means the control remains available for eligible assistant responses throughout a long Codex desktop task.