[Windows][plugin-eval] Absolute paths become C:\C:\..., causing 19/28 evaluator test failures

What version of Codex CLI is running?

codex-cli 0.153.4

What platform are you using?

Windows 10 x64

What issue are you seeing?

The official plugin-eval@openai-curated-remote plugin (Plugin Eval 0.1.2) has a Windows path-normalization failure where an already-absolute Windows path can be prefixed again and become malformed, for example:

C:\C:\Users\...

The Plugin Eval executable itself can run successfully on Windows, but the full evaluator test suite is heavily affected by this path bug.

Reproduction

1. Install/enable plugin-eval@openai-curated-remote on Windows.
2. Run a minimal standalone evaluation fixture.
3. Observe that the fixture succeeds (100/100, exit code 0).
4. Run the full Plugin Eval test suite.
5. Observe failures involving malformed absolute Windows paths with a duplicated drive prefix such as:

C:\C:\Users\...

Actual result

Before updating Codex CLI (0.146.1):

9 passed
19 failed

After updating Codex CLI to stable 0.153.4:

9 passed
19 failed

The failure pattern is unchanged, suggesting this is in Plugin Eval path handling rather than the Codex CLI version.

A simple minimal-skill fixture evaluates successfully at 100/100, so this is not a total Windows incompatibility.

Expected behavior

Absolute Windows paths such as:

C:\Users\<USER>\...

should remain absolute and should never be joined or prefixed into:

C:\C:\Users\...

The evaluator test suite should normalize and resolve native Windows paths correctly.

Impact

• 19/28 Plugin Eval tests fail on Windows.
• Scores for real local skills cannot be treated as fully authoritative while path-related tests are failing.
• Reliable regression testing of Codex skills/plugins on native Windows is blocked.

Additional information

• Codex install method: global npm @openai/codex
• Plugin: plugin-eval@openai-curated-remote
• Plugin Eval version: 0.1.2
• No Plugin Eval source files, manifests, symlinks, or local workarounds were modified.
• A local skill was evaluated statically only; none of its operational capabilities were executed.

The Plugin Eval source currently lives in openai/plugins, but Issues are disabled in that repository, so this report is filed here because the affected plugin is distributed through the official Codex plugin catalog.

If useful, I can provide a sanitized list of the 19 failing test names/path traces without including the local Windows username.

Start by running the full Plugin Eval test suite on native Windows and compare it with the successful minimal standalone fixture. Inspect the Plugin Eval source in openai/plugins for path normalization and resolution, then verify that all 28 tests handle native absolute paths without producing duplicated drive prefixes such as C:\C:\Users....