fix(utils): explain CWD / ROBOVERSE_DATA_DIR mismatches in hf_util instead of reporting traversal

CHANGELOG.md

@@ -38,6 +38,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - `RLTaskEnv.step` publishes the *terminal* observation in `info["observations"]["raw"]["obs"]` instead of the episode's first one (off-policy truncation bootstraps in clean_rl SAC/TD3 and FastTD3 read it).
 - IsaacGym and PyBullet reported `joint_pos_target` in native DoF order while `joint_pos` is in sorted-name order; both now use `get_joint_names(sort=True)` (completes #12).
 - `ParallelSimWrapper`: a worker that died during handler construction or `launch` surfaced as a bare `EOFError`/`ConnectionResetError` from the handshake and left the other workers running; the handshake now raises the worker's own traceback, `close()` tolerates dead workers, and a failed constructor tears the pool down.
+- `hf_util.check_and_download_single`: a `roboverse_data/...` path evaluated from a working directory that is not the parent of `ROBOVERSE_DATA_DIR` was reported as a path-traversal attempt; the error now names the CWD / `ROBOVERSE_DATA_DIR` mismatch and where the asset already is. `test_check_and_download_single_falls_back_to_private_roboverse_data` no longer depends on the caller's `ROBOVERSE_DATA_DIR`.
 - MuJoCo: `<size memory="512M">` is reserved by default; humanoid + mesh scenes no longer die with
   `mj_stackAlloc: out of memory` (get_started/10_mount_camera.py).
 - `hf_util`: a symlinked `roboverse_data` is no longer refused as path traversal; concurrent

metasim/test/test_hf_util.py

@@ -2,6 +2,8 @@
 
 from __future__ import annotations
 
+import os
+
 import pytest
 
 import metasim.utils.hf_util as hf_util
@@ -50,6 +52,9 @@ def fake_file_exists(repo_id, filename, repo_type):
     def fake_hf_hub_download(repo_id, filename, repo_type, local_dir):
         download_calls.append((repo_id, filename, repo_type, local_dir))
 
+    # LOCAL_DIR is resolved from ROBOVERSE_DATA_DIR at import time; pin it to the CWD-relative
+    # default so the test does not depend on the caller's environment.
+    monkeypatch.setattr(hf_util, "LOCAL_DIR", os.path.abspath("roboverse_data"))
     monkeypatch.setattr(hf_util.os.path, "exists", lambda path: False)
     monkeypatch.setattr(hf_util.hf_api, "file_exists", fake_file_exists)
     monkeypatch.setattr(hf_util, "hf_hub_download", fake_hf_hub_download)

metasim/test/test_hf_util_local_dir.py

@@ -166,3 +166,26 @@ def _fake_find(relpath_posix, is_optional_file=False):
         hf_util.check_and_download_single(str(real_dir / "robots" / "g1" / "meshes" / "pelvis.STL"))
     assert "Refusing to fetch" not in str(excinfo.value)
     assert seen["relpath"] == "robots/g1/meshes/pelvis.STL"
+
+
+@pytest.mark.general
+def test_cwd_mismatch_error_names_the_cause(monkeypatch, tmp_path):
+    """A ``roboverse_data/...`` path evaluated from the wrong CWD must explain the CWD /
+    ROBOVERSE_DATA_DIR mismatch (and where the asset already is), not accuse the caller of traversal."""
+    from metasim.utils import hf_util
+
+    data_dir = tmp_path / "shared_data"
+    (data_dir / "robots").mkdir(parents=True)
+    (data_dir / "robots" / "franka.urdf").write_text("<robot/>", encoding="utf-8")
+    elsewhere = tmp_path / "elsewhere"
+    elsewhere.mkdir()
+    monkeypatch.chdir(elsewhere)
+    monkeypatch.setattr(hf_util, "LOCAL_DIR", str(data_dir))
+    monkeypatch.setattr(hf_util.hf_api, "file_exists", lambda *a, **kw: pytest.fail("HfApi reached"))
+
+    with pytest.raises(ValueError) as excinfo:
+        hf_util.check_and_download_single(os.path.join("roboverse_data", "robots", "franka.urdf"))
+    msg = str(excinfo.value)
+    assert "working directory" in msg and str(elsewhere) in msg
+    assert str(data_dir / "robots" / "franka.urdf") in msg
+    assert "ROBOVERSE_DATA_DIR" in msg and "traversal" not in msg

metasim/utils/hf_util.py

@@ -120,6 +120,33 @@ def extract_texture_paths_from_mdl(mdl_file_path: str) -> list[str]:
     return texture_paths
 
 
+_RELATIVE_DATA_PREFIX = "roboverse_data" + os.sep
+
+
+def _outside_local_dir_message(filepath: str, relpath: str) -> str:
+    """Explain why ``filepath`` cannot be fetched into ``LOCAL_DIR``.
+
+    Asset paths in configs are ``roboverse_data/...`` and are opened by the simulator backends
+    relative to the *current working directory*; ``ROBOVERSE_DATA_DIR`` only moves where downloads
+    land. So a relative path that does not resolve under ``LOCAL_DIR`` is almost always a CWD /
+    ``ROBOVERSE_DATA_DIR`` mismatch, not a malicious ``..`` — say so, and point at the asset when it
+    already exists in the configured directory.
+    """
+    if not os.path.isabs(filepath) and filepath.startswith(_RELATIVE_DATA_PREFIX):
+        in_local_dir = os.path.join(LOCAL_DIR, filepath[len(_RELATIVE_DATA_PREFIX) :])
+        found = f" The asset exists at {in_local_dir!r}." if os.path.exists(in_local_dir) else ""
+        return (
+            f"Cannot fetch {filepath!r}: asset paths are relative to the working directory "
+            f"({os.getcwd()!r}), which is not the parent of LOCAL_DIR ({LOCAL_DIR!r}).{found} "
+            f"Run from the repository root, or symlink ./roboverse_data to LOCAL_DIR "
+            f"(ROBOVERSE_DATA_DIR only changes where downloads are stored)."
+        )
+    return (
+        f"Refusing to fetch {filepath!r}: resolves outside LOCAL_DIR ({LOCAL_DIR}) via relative-path "
+        f"traversal ({relpath!r})."
+    )
+
+
 def check_and_download_single(filepath: str):
     """Check if the file exists in the local directory, and download it from the huggingface dataset if it doesn't exist.
 
@@ -145,10 +172,7 @@ def check_and_download_single(filepath: str):
         # outside LOCAL_DIR — refuse it rather than send the escaped path
         # to the HF API. Optional files warn-and-skip; required files raise.
         if relpath.split(os.sep, 1)[0] == "..":
-            msg = (
-                f"Refusing to fetch {filepath!r}: resolves outside LOCAL_DIR "
-                f"({LOCAL_DIR}) via relative-path traversal ({relpath!r})."
-            )
+            msg = _outside_local_dir_message(filepath, relpath)
             if is_optional_file:
                 log.warning(msg + " Skipping optional file.")
                 return