Sandbox tar extraction and 12 sandbox tests fail on Windows without symlink privilege #4852

Running the sandbox suite on Windows as a normal, non-elevated user:

tests/sandbox/test_tar_utils.py + test_workspace_paths.py:
12 failed, 93 passed, 15 skipped
All twelve fail with OSError: [WinError 1314] A required privilege is not held by the client. Windows permits symlink creation only for an elevated process or with Developer Mode enabled — neither is the default.

They split into two distinct problems.

1. Product: safe_extract_tarfile cannot extract archives containing symlinks
   Seven of the failures originate in library code, not test setup:

tests/sandbox/test_tar_utils.py:75: in _safe_extract
src/agents/sandbox/util/tar_utils.py:468: in safe_extract_tarfile

E   OSError: [WinError 1314] A required privilege is not held by the client:
    '/usr/local/bin/python3' -> '...\link.txt'
tar_utils.py:468:

    for member in members:
        if not member.issym():
            continue
        rel_path = safe_tar_member_rel_path(member, allow_symlinks=True)
        if rel_path is None:
            continue
        dest = root_resolved / rel_path
        _prepare_replaceable_leaf(dest=dest, rel_path=rel_path, name=member.name)
        os.symlink(member.linkname, dest)

There is no platform handling anywhere in that file. So on an ordinary Windows machine, extracting any archive that contains a symlink member raises, aborting the extraction partway — after the regular files have already been written.

Affected tests, all exercising real extraction behaviour:

test_safe_extract_tarfile_can_rehydrate_existing_leaf_symlink
test_safe_extract_tarfile_can_replace_existing_leaf_file_with_symlink
test_safe_extract_tarfile_can_replace_existing_leaf_symlink_with_file
test_safe_extract_tarfile_can_replace_existing_leaf_symlink_with_directory
test_safe_extract_tarfile_rejects_preexisting_symlink_parent
test_safe_extract_tarfile_rejects_symlink_under_preexisting_symlink_parent
(and one more in the same file)

I have deliberately not proposed a patch for this half. The sandbox has security semantics around symlinks — safe_tar_member_rel_path(..., allow_symlinks=True), escape rejection, _prepare_replaceable_leaf — and the options each carry different meaning:

Skip symlink members on Windows. Extraction completes, but the result silently differs from the archive, which may matter for a sandbox that later validates paths.
Raise a clear, typed error. Honest, but still fails the operation.
Materialise as a copy of the target. Preserves content, changes escape semantics — a copy cannot point outside the root, so some of the existing protections become moot.
Which of those is right is a maintainer decision, not one I should make from outside.

2. Tests: unguarded symlink creation in test setup
   The other five come from the tests themselves, e.g. tests/sandbox/test_workspace_paths.py:798:

os.symlink(target, workspace / "link.txt")
Nothing guards these, and there is no module-level skip or marker.

test_normalize_path_with_symlink_resolution
test_host_io_rejects_write_under_resolved_read_only_extra_path_grant
test_host_io_rejects_extra_path_grant_symlink_to_root
test_host_path_grant_rejects_symlink_to_root
test_host_path_grant_returns_validated_resolved_source

This half is straightforward and I am happy to submit a PR: a small symlink_or_skip() helper used at the affected call sites, so the tests skip with a clear reason where the platform refuses rather than erroring. That is the same approach psf/black uses for its symlink tests.

Why CI doesn't catch either
tests-windows in .github/workflows/tests.yml runs uv run pytest — the full suite, including tests/sandbox/ — on windows-latest, and it is green. GitHub's Windows runners can create symlinks; an ordinary user cannot. So both problems are invisible to CI on the very platform it covers, and appear only on a contributor's or user's machine.

Environment
Windows-11-10.0.26200-SP0
Python 3.13.15
openai-agents-python at HEAD, uv sync dev environment
Non-elevated shell, Developer Mode off

Reproduction

git clone https://github.com/openai/openai-agents-python
cd openai-agents-python
uv sync
uv run pytest tests/sandbox/test_tar_utils.py tests/sandbox/test_workspace_paths.py

Note: a wider tests/sandbox/ run also reports failures in test_docker.py, test_docker_network_mode.py and test_mount_security.py, but those are a missing docker module in my environment and are unrelated to this report.