Codex Desktop wait_threads drops :true from completed assistant JSON while read_thread preserves it

Environment

• Windows, Codex Desktop 26.901.2854.0.
• Rechecked on 2026-09-04 using an existing completed local canary task.
• The same discrepancy was observed on Desktop 26.831.2377.0 on 2026-09-02.
• Affected first-party tools: codex_app.wait_threads and codex_app.read_thread.

Reproduction and observed result

1. Use a completed task whose final assistant message contains a JSON object with two boolean true values.
2. Call read_thread for its latest completed turn and retain the agent message text, message ID and turn ID.
3. Call wait_threads with the same task ID and timeoutMs: 0, without a previously consumed cursor.
4. Compare latestAssistantMessage.text against the message returned by read_thread.

In the observed task, both tools return the same message ID and completed turn ID. The text from read_thread is 812 characters and parses as valid JSON. The text from wait_threads is 802 characters and fails JSON parsing.

The only differences are removal of the two literal :true substrings. Reduced illustration, with keys renamed:

read_thread:  {"flag_one":true,"flag_two":true}
wait_threads: {"flag_one","flag_two"}

This illustration shows the affected fields only. The full comparison verified that all other characters were identical. The outer tool response remains valid JSON; the corruption affects the assistant-message string inside it. The wait response did not mark that string as truncated.

Expected behavior and impact

latestAssistantMessage.text should preserve the stored assistant message, including boolean values. If a summary is returned instead, expose it separately and identify it as a summary.

A coordinator relying on the wait result cannot parse the returned JSON and loses two reported boolean values, although the target task completed successfully. read_thread is a working fallback in this reproduction.

The exact transformation layer is unknown. A regression test comparing both tools' text for the same completed message would help prevent this.

Only sanitized reproduction details are included; no raw logs, private task content, account identifiers or local paths are attached.

Start by reproducing the completed-task comparison between codex_app.wait_threads and codex_app.read_thread with timeoutMs: 0, then trace both tool entry points to find where latestAssistantMessage.text is transformed. Add a regression test using the same completed message and verify that wait_threads preserves both boolean values and produces text matching read_thread.