Fix Sandbox macOS Credentials Access

The seach skill in this repository continues to cause issues on codex where it tries to execute the skill within its sandbox which prevents the entire executable from reading its credentials from the local macos secrets store. Can we improve the instructions to prevent that from happening?
I’m using the skill-creator guidance to update the skill safely, and I’ll also read the repository’s search skill itself because that’s the behavior we’re correcting. I’ll trace the current execution instructions, tighten the sandbox/escalation language, and validate the resulting skill.
The failure is visible in the current wording: it says that entire login is required, but never tells Codex that every entire subprocess must run outside the filesystem sandbox. I’m checking nearby skills and repository conventions so the fix uses the established permission terminology and doesn’t accidentally broaden unrelated commands.
I’ve added an explicit first-attempt rule, not merely retry guidance: all entire search and entire checkpoint explain calls must request outside-sandbox execution up front. I also separated genuine authentication failures from the misleading credential error caused by sandboxing. Now I’m validating the skill and reviewing the diff for ambiguity.
The content and whitespace checks pass. The bundled validator itself isn’t directly executable here and its default Python environment lacks PyYAML, so I’m switching to a local interpreter/dependency path check rather than treating that tooling issue as a skill failure.
Updated skills/search/SKILL.md to require:
- Running all
entirecommands outside the Codex sandbox on the first attempt. - Using
sandbox_permissions: "require_escalated"with a credential-access justification. - Reporting authentication errors only after an outside-sandbox attempt.
The change passes frontmatter/body validation and git diff --check. Existing unrelated .entire changes were untouched.
Can you check out a new branch, commit the changes and run pfleidi:pr to create a pull request?
I’m using pfleidi:pr for the branch preparation, commit, push, and pull-request workflow. I’ll keep the pre-existing .entire changes out of the commit and include only the search-skill fix.
The intended change is cleanly isolated, but the worktree also contains unrelated .entire/.gitignore and .entire/logs/ changes that I won’t stage. I’m refreshing origin/main before branching so the new branch starts from the current upstream tip; after committing, the PR workflow may require those unrelated changes to be temporarily stashed before it can perform its clean-tree sync check.
Created branch fix/search-skill-codex-sandbox and commit d6a918f.
PR creation is paused because the worktree contains unrelated uncommitted changes in .entire/.gitignore and .entire/logs/. May I temporarily stash those files, create the PR, then restore them?