Codex Desktop reports success when ChatGPT follow-up is silently dropped

Summary

In Codex Desktop, sending a follow-up to an existing ChatGPT conversation through the app thread-management interface reports success, but the message is not persisted and no response run starts.

Environment

• macOS 26.5.2 (25F84)
• Codex Desktop, local host
• Source task kind: chatgpt
• Destination conversation displayed as active with no active flags

Steps to reproduce

1. Identify an existing ChatGPT conversation that has a completed user turn without a final assistant response after a long-running task stalls.
2. Read it through the Codex app thread reader and confirm the latest durable turn and updatedAt value.
3. Call the Codex app follow-up action for that ChatGPT conversation with a normal text prompt.
4. Observe that the action returns a successful result containing the destination thread ID.
5. Read the same conversation again.

Actual behavior

• The follow-up action returns success.
• The submitted user message does not appear in conversation history.
• updatedAt remains unchanged.
• No model run begins and no error is exposed.
• Navigating the desktop app to the conversation and retrying produces the same result.
• The Codex-thread waiting action cannot monitor the destination because it is a ChatGPT conversation, returning that no Codex thread was found.

This occurred on multiple consecutive attempts with both concise and detailed recovery prompts.

Expected behavior

One of the following should occur:

1. The follow-up is durably appended and starts a response run; or
2. The action rejects the request with an explicit unsupported-destination or dispatch error.

It should not acknowledge success while silently dropping the message.

Impact

A long-running research workflow stopped after reporting 19 of 24 completed runs. Because follow-up dispatch silently failed and transient progress was absent from the durable thread reader, Codex could neither resume the task nor reliably recover its checkpoint. The study had to be restarted in a new Codex task.

Privacy note

Conversation and local task identifiers are intentionally omitted from this public report. I can provide them privately to an OpenAI maintainer if needed.