Codex macOS app previews .skill ZIP packages as UTF-8 text mojibake

Bug

Codex macOS app local file preview treats a .skill package as plain text and renders ZIP bytes as mojibake/replacement characters.

This is not file corruption: the .skill file is a valid ZIP archive.

Environment

• App: Codex macOS app
• Bundle id: com.openai.codex
• App version: 26.513.31313
• Build: 2867
• Codex CLI: codex-cli 0.130.0
• OS: macOS 26.3.1

Repro

1. Open a local project containing a .skill package.
2. In Codex macOS app, open/preview a file like:

dist/skills/math-handoff-skill/math-handoff-skill.skill

Actual behavior

The file preview opens a text/code view and shows binary ZIP data as mojibake. The rendered content starts with PK and then many replacement characters / corrupted-looking text.

<img width="575" height="705" alt="Image" src="https://github.com/user-attachments/assets/6835ded7-e04d-49fb-a77e-f3af4c3c9e1d" />

Expected behavior

Codex should not render binary/archive files as UTF-8 text. It should show a binary/archive placeholder, open externally, or expose archive contents.

File validation

The file is valid ZIP data:

• first bytes: 50 4b 03 04
• zipfile.is_zipfile(...) == true
• unzip -t succeeds
• entries include:
• math-handoff-skill/skill-boundary.json
• math-handoff-skill/SKILL.md
• math-handoff-skill/templates/handoff.md

It is invalid UTF-8:

• first UTF-8 decode error at byte 11
• lossy decode produces many U+FFFD replacement chars

Diagnosis

The app appears to classify local workspace files by extension before choosing preview behavior.

The packaged app bundle has an unsupported/archive extension list that includes extensions like zip, jar, tar, gz, rar, etc., but .skill is not included. Therefore .skill falls through to the plain-text preview path.

The read-file handler then decodes bytes as UTF-8 text. Invalid bytes are replaced with U+FFFD, and the file-preview page renders the resulting string.

So the immediate failure seems to be in the Codex macOS app file-preview classification/decoding path, not in the .skill package itself.

Suggested fixes

Minimal:

• classify .skill as an archive/unsupported preview type, same as .zip.

Robust:

• perform binary detection before text rendering:
• strict UTF-8 decode via TextDecoder('utf-8', { fatal: true })
• detect NUL bytes/control-byte ratio
• detect common magic bytes, including ZIP PK\x03\x04
• only pass contents: string to the text renderer when the file is text-safe.

Broader:

• return isBinary / mimeType / encoding metadata from file-read APIs and let the preview page choose the correct renderer.