I’m seeing a reproducible LaTeX rendering issue in the Codex desktop app.
In my testing, LaTeX written by the user in chat does not render, while assistant-authored LaTeX appears to render more reliably.
Repro steps:
- Open a conversation in the Codex desktop app.
- Send a user message containing one of the following:
$$f(x) \text{ display}$$Inline: $$f(x) \text{ inline}$$\(f(x)\)\[f(x)\]
- Look at how the message appears in the transcript.
Expected:
User-authored LaTeX should render as math, or the supported delimiter set should be documented clearly.
Actual:
The LaTeX appears as raw source in the user message.
What makes this confusing is that assistant-authored math seems to behave differently in the same conversation, so there may be an inconsistency between user-message rendering and assistant-message rendering.
Environment:
- Codex desktop app
- macOS
- Observed on April 6, 2026
- Time zone: Europe/Paris
If anyone else can reproduce this, it would help to compare:
- macOS vs Windows
- desktop app vs browser
- user-authored vs assistant-authored LaTeX
$$...$$vs\(...\)vs\[...\]
I’ve also reported this through OpenAI support, but I’m posting here to see whether others can reproduce it.