Please confirm raw output source for mixed-script token and malformed support URL

What issue are you seeing?

This is a follow-up to #20054.

In #20054, I reported an anomalous mixed-script fragment in a Codex CLI response. The official reply classified it as likely model behavior and suggested using /feedback for additional model-output reports.

I am now reporting a second observation where the captured assistant output contains both:

1. a Cyrillic word inside an otherwise French sentence, and
2. a malformed OpenAI Help Center URL in the same response context.

Offending sentence:

Ce profil préactive un набор cohérent de postes, puis l’utilisateur ajuste.

The Cyrillic word is:

набор

The malformed URL shown in the captured output begins with the expected Help Center article path:

https://help.openai.com/en/articles/6614161-how-can-i-contact-support

but appears in the rendered/copied assistant response as a much longer percent-encoded URL ending with an image-like fragment, qualitatively of this form:

https://help.openai.com/en/articles/6614161-how-can-i-contact-support%2525252525...%23.avif

Evidence

A screenshot showing both artifacts was added in this comment:

https://github.com/openai/codex/issues/25648#issuecomment-4594765767

The comment includes integrity hashes for the original screenshot, the sanitized PNG copy, and the published SVG wrapper.

Requested clarification

Could maintainers clarify, if this is possible from available logs or telemetry, whether the raw response sent to the client already contained:

1. the Cyrillic token набор, and
2. the malformed OpenAI Help Center URL,

or whether either artifact could have been introduced later by client-side rendering, Markdown/link handling, terminal copy/paste behavior, or another post-processing layer?

If GitHub is not the right place to determine this, please indicate the correct channel and what exact context should be submitted through /feedback or support so the raw response can be checked.

Reproduction status

I do not yet have a deterministic minimal reproduction. The current report is based on the captured output shown in the screenshot and is intended to connect this second observation to #20054 while asking specifically about the raw output source.