macOS Codex app: Korean IME composition breaks into separated jamo after input source switching

Summary

On the Codex macOS desktop app, Korean 2-set IME composition can get stuck in a state where Hangul syllables are committed as separated jamo after switching input sources. The broken state persists until focus leaves Codex, for example by switching to another app and coming back.

Environment

• Codex app: 26.506.31421 (CFBundleVersion 2620)
• Bundled CLI: codex-cli 0.130.0-alpha.5
• Electron Framework: 41.2.0
• OS: macOS 26.4.1 (25E253), arm64
• Input source: Apple Korean 2-Set (com.apple.inputmethod.Korean.2SetKorean); ABC also enabled
• Surface: Codex desktop app message composer

Steps to reproduce

1. Open the Codex desktop app on macOS.
2. Focus the message composer.
3. Use Korean 2-set input, then switch input sources between Korean and English. In my case the bug starts after toggling the input source.
4. Type Korean text in the composer.

Expected behavior

IME preedit/composition should combine jamo into normal Hangul syllables, for example:

이게 영어로 변환하고 나서 글 써지는 거야
그리고 다시 한글로 변환하면 그래도 이렇게 나와

Actual behavior

Input is committed as separated jamo instead of composed Hangul syllables, for example:

ㅇㅣㄱㅔ ㅇㅕㅇㅇㅓㄹㅗ ㅂㅕㄴㅎㅗㅏㄴㅎㅏㄱㅗ ㄴㅏㅅㅓ ㄱㅡㄹ ㅆㅓㅈㅣㄴㅡㄴㄱㅓㅇㅑ...
ㄱㅡㄹㅣㄱㅗ ㄷㅏㅅㅣ ㅎㅏㄴㄱㅡㄹㄹㅗ ㅂㅕㄴㅎㅗㅏㄴㅎㅏㅁㅕㄴ
ㄱㅡㄹㅐㄷㅗ ㅇㅣㄹㅓㅎㄱㅔ ㄴㅏㅇㅗㅏ

The broken state continues even after switching back to Korean input. It clears only after leaving Codex and returning, which looks like the IME focus/composition context is being reset.

Notes

• The macOS input source itself appears normal: Korean 2-Set and ABC are enabled.
• Other apps do not show this exact behavior in the same session.
• This looks like an Electron/renderer composition/preedit handling issue in the composer text field.
• Possibly related class of issue: #4870, but this report is for the macOS desktop app composer and the jamo are actually committed separately rather than only rendered incorrectly during preedit.

• This issue draft was written by Codex on behalf of the user, based on the user's observed reproduction text and local environment checks.