Codex Desktop transcript shows internal XML-like context/control blocks

Summary

Codex Desktop is showing internal structured/context messages directly in the visible chat transcript. These appear as XML-like blocks such as:

<command-name>/exit</command-name>
<command-message>exit</command-message>
<command-args></command-args>

and:

<local-command-stdout>See ya!</local-command-stdout>

It also surfaced larger internal context sections such as workspace instructions and environment context in XML-like wrappers.

Expected behavior

Internal control/context messages should not be visible to the end user in the normal chat transcript. If some diagnostic metadata is needed, it should be hidden behind a debug/verbose view rather than interleaved with the conversation.

Actual behavior

The XML-like control/context blocks are rendered inline in the user-visible conversation, creating noise and making it unclear what is user-facing conversation vs assistant/runtime metadata.

Impact

This is confusing and disruptive during normal use. The user has to mentally filter out internal protocol messages that are not actionable and were not intentionally requested.

Environment

• App: Codex Desktop
• Platform: macOS
• Observed in a normal conversation after exiting an agent/session with /exit

Notes

The examples above are sanitized; no private paths or secrets are included. The issue appears to be transcript/rendering leakage rather than a model behavior problem.