Entire

I’ll check the shared agent configuration, symlink targets, and sandbox restrictions to identify what caused that warning. I’m using the development and OpenAI documentation skills for this investigation.

The warning is reproducible: the sandbox fails before any shell command runs, even true from /tmp. It rejects the repository’s .codex symlink while trying to protect that path as read-only. This also prevents me from reading the skill files; I’m checking the official documentation for a compatible directory layout.

The warning is accurate: the repository’s .codex symlink prevents the sandbox from starting.

I reproduced this error:

Even true from /tmp, with shell initialization disabled, fails identically. Consequently, repository inspection, skill reads, tests, and other shell commands cannot start.

Your supplied AGENTS.md says .codex links to .agents. This session protects .codex as read-only inside a writable repository; the sandbox builder cannot enforce that arrangement through this symlink.

The recommended repair is:

Keep .agents/skills/ as the shared source of skills. Codex discovers it directly, so the .codex alias is unnecessary for skill discovery. Official documentation
Replace the top-level .codex symlink with a real directory if Codex-specific configuration is needed, preserving any configuration it currently exposes.
Update AGENTS.md and any setup scripts so they no longer recreate that symlink.
Restart the session and verify shell execution, skill reads, and repository status.

I cannot inspect the configuration to determine what must be preserved, or apply this repair from this session: shell startup is blocked, .codex and .agents are protected, and permission escalation is disabled. The initial filesystem repair needs to happen in your own terminal. No files were changed; the broader workflow audit remains blocked by this startup failure.

I’ll inspect the shared configuration, replace the .codex symlink, and update any instructions that would recreate it. I’m using the development skill to check the agent setup.

The only Codex-specific file exposed through the symlink is Entire’s hook configuration. I’ll move it unchanged into a real .codex/ directory and keep skills in .agents/skills/. I found two instructions prescribing the old symlink layout and no setup scripts that recreate it.

Implemented: .codex/ is now a real directory, with Entire’s hooks preserved unchanged. Skills remain in .agents/skills/, and both instruction files describe the new layout.

Verified that Codex’s workspace-write sandbox starts, reads skills, and runs Git inspection. Configuration checks and git diff --check pass.

Commit both if Entire is part of the team’s shared workflow. They configure different agents:

.agents/settings.json: Claude hooks, accessed through .claude/settings.json.
.codex/hooks.json: Codex hooks.

The versions I inspected contain hook commands, with no credentials or personal paths. They skip execution when Entire isn’t installed, though the session-start hook displays an installation notice.

If Entire is only your personal tooling, ignore both instead.

For the sandbox fix, commit the removal of the .codex symlink and the instruction updates either way. If you exclude .codex/hooks.json, Git won’t retain the empty .codex/ directory—that’s fine.