Codex desktop: shell/Python file writes lack per-file edit cards, bypassing the user's visible review surface

Summary

Reported at the user's explicit request; this report has been redacted.

In Codex desktop on Windows, the assistant modified existing source files and created new files by running Python pathlib read_text/write_text scripts through the shell execution tool. The user observed that the normal per-file 'Edited +N -M' cards with Undo/Review controls were absent for these changes. A subsequent git status confirmed that the modifications were present in the working tree.

The user describes this as: 'The assistant has started modifying files while bypassing the only oversight mechanism I rely on.' The concern is loss of the user's visible, per-edit review surface, not loss of the files themselves.

Observed sequence

1. A user requests authorized code changes in a local Git workspace.
2. The assistant performs file writes through shell-invoked Python scripts instead of the dedicated patch-edit tool.
3. Files change on disk, but the user does not see the usual per-file edit cards in the conversation.
4. The user notices the missing review cards and asks where the modified-file listing went.
5. Git status shows the changed and newly created files still exist.

Expected behavior

Every agent-originated file modification should have a discoverable, consistent review/audit surface, regardless of whether it originated from apply_patch, a shell command, Python, or another file-writing tool. Shell/script-based editing should not silently bypass the normal per-file change presentation. If equivalent inline rendering cannot be provided, display a prominent indication and a direct link to the corresponding diff.

Impact

The user's established means of noticing and reviewing agent edits disappeared while the agent continued modifying files. This creates an oversight and trust gap. Please investigate file-change attribution and edit-card rendering for shell/script-originated writes, including new/untracked files.

Scope and uncertainty

This was observed during a real task, not a controlled reproduction. We have not established whether all auditing mechanisms are affected, whether the cause is the write tool or another UI condition, or whether the absence of cards is intended behavior. Git-based review remains available. No claim is made that sandbox restrictions or permission checks were bypassed, or that the assistant intentionally evaded oversight.

Environment: Codex desktop on Windows. Exact app version not collected.

Redaction

No project names, repository identifiers, source code, business domains, local paths, credentials, original screenshots, raw logs, or task/conversation identifiers are included. Only the minimal behavioral description is provided.

No files, tests, or entry points are named. Start by reproducing on Windows with shell-invoked Python pathlib writes, then compare the conversation's per-file edit cards with git status for modified and untracked files. Done means agent-originated writes receive an equivalent review surface or show the promised prominent diff indication.