v1.5 sandbox regression: bwrap loopback fails on GitHub Actions runner

Symptom

When pinning uses: openai/codex-action@v1.5 (instead of @v1), the action posts a comment on the PR saying it cannot read any files:

Не могу выполнить ревью в текущей сессии: любые локальные команды (`git diff`, `cat`, даже `pwd`) падают с ошибкой песочницы:

bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted

So Codex returns a comment but cannot inspect the diff at all → no actual review.

Reproduction

1. Pin openai/codex-action@v1.5 in any workflow
2. Open a PR
3. Codex runs, posts the bwrap error message instead of a real review

@v1 (which currently floats to whatever pre-1.5 release was last published) works fine on the same repo, same runner.

Environment

• GitHub-hosted ubuntu-latest runner
• Triggered on pull_request event
• Repository is private
• Same workflow file works on @v1

Expected

v1.5 should be a non-breaking minor bump — but the bwrap sandbox cannot initialise its loopback interface, so the codex CLI inside cannot exec local commands.

Workaround

Stay on @v1 until v1.6 fixes the sandbox setup.

Related

• Workflow file: standard pull_request trigger with openai/codex-action step
• Problem appeared the moment we changed only the uses: line from @v1 to @v1.5, then disappeared after revert.

Start with the workflow file's pull_request trigger and openai/codex-action step, then reproduce the failure on an ubuntu-latest runner using @v1.5 and compare it with @v1. Inspect the v1.5 sandbox setup and runner permissions; done means the action can execute local commands and produce a real review without requiring the @v1 workaround.