fix(spreadsheet): complete generated-code env isolation

skillopt/envs/spreadsheetbench/codegen_agent.py

@@ -262,6 +262,7 @@ def _build_codex_driver() -> str:
         "import re\n"
         "import subprocess\n"
         "import sys\n\n"
+        "import tempfile\n\n"
         'INPUT_PATH = "input.xlsx"\n'
         'OUTPUT_PATH = "output.xlsx"\n'
         "code = pathlib.Path('solution.py').read_text(encoding='utf-8')\n"
@@ -277,15 +278,21 @@ def _build_codex_driver() -> str:
         "_safe_env = {\n"
         "    'PATH': _os.environ.get('PATH', '/usr/bin:/bin'),\n"
         "    'HOME': str(pathlib.Path('_driver_runner.py').parent.resolve()),\n"
+        "    'TMPDIR': tempfile.gettempdir(),\n"
         "}\n"
         "if _os.name == 'nt':\n"
         "    _safe_env['SYSTEMROOT'] = _os.environ.get('SYSTEMROOT', '')\n"
-        "    _safe_env['TEMP'] = _safe_env['HOME']\n"
-        "    _safe_env['TMP'] = _safe_env['HOME']\n"
+        "    _safe_env['TEMP'] = tempfile.gettempdir()\n"
+        "    _safe_env['TMP'] = tempfile.gettempdir()\n"
         "_safe_env = {k: v for k, v in _safe_env.items() if v}\n"
-        "_res = subprocess.run([sys.executable, str(_patched)], capture_output=True, text=True, env=_safe_env)\n"
+        "try:\n"
+        "    _res = subprocess.run([sys.executable, str(_patched)], capture_output=True, text=True, env=_safe_env)\n"
+        "finally:\n"
+        "    try:\n"
+        "        _patched.unlink()\n"
+        "    except OSError:\n"
+        "        pass\n"
         "if _res.returncode != 0:\n"
-        "    import traceback as _tb\n"
         "    print(_res.stdout, end='')\n"
         "    print(_res.stderr, end='')\n"
         "    sys.exit(2)\n"

skillopt/envs/spreadsheetbench/executor.py

@@ -46,8 +46,9 @@ def run_generated_code(code: str, input_path: str, output_path: str, timeout: in
     with tempfile.NamedTemporaryFile("w", suffix=".py", delete=False) as f:
         f.write(script)
         tmp = f.name
-    # Build a minimal environment so the generated code cannot read API keys,
-    # cloud credentials, or other secrets from the current process environment.
+    # Build a minimal environment so generated code does not directly inherit
+    # API keys, cloud credentials, or other parent-process environment values.
+    # This is environment isolation, not a filesystem or network sandbox.
     import platform as _platform
     _safe_env: dict[str, str] = {
         "PATH": os.environ.get("PATH", "/usr/bin:/bin"),
@@ -58,7 +59,7 @@ def run_generated_code(code: str, input_path: str, output_path: str, timeout: in
         _safe_env["SYSTEMROOT"] = os.environ.get("SYSTEMROOT", "")
         _safe_env["TEMP"] = tempfile.gettempdir()
         _safe_env["TMP"] = tempfile.gettempdir()
-    # Drop empty entries (env dict values must be non-empty strings)
+    # Omit platform-specific entries that are absent in the parent.
     _safe_env = {k: v for k, v in _safe_env.items() if v}
     try:
         proc = subprocess.run(

tests/test_executor_env_isolation.py

@@ -7,8 +7,10 @@ parent process environment. These tests assert that scrubbing behaviour.
 """
 from __future__ import annotations
 
-import os
+import subprocess
+import sys
 
+from skillopt.envs.spreadsheetbench.codegen_agent import _build_codex_driver
 from skillopt.envs.spreadsheetbench.executor import run_generated_code
 
 
@@ -44,3 +46,32 @@ def test_path_still_available_to_generated_code(tmp_path) -> None:
 
     assert ok, err
     assert out.read_text(encoding="utf-8") == "YES"
+
+
+def test_codex_driver_scrubs_env_sets_tempdir_and_cleans_runner(
+    tmp_path, monkeypatch
+) -> None:
+    monkeypatch.setenv("SUPER_SECRET_TOKEN", "do-not-inherit")
+    (tmp_path / "solution.py").write_text(
+        "import os\n"
+        "with open(OUTPUT_PATH, 'w', encoding='utf-8') as f:\n"
+        "    f.write('|'.join([\n"
+        "        os.environ.get('SUPER_SECRET_TOKEN', 'ABSENT'),\n"
+        "        'TMPDIR' if os.environ.get('TMPDIR') else 'NO_TMPDIR',\n"
+        "    ]))\n",
+        encoding="utf-8",
+    )
+    driver = tmp_path / "run_solution.py"
+    driver.write_text(_build_codex_driver(), encoding="utf-8")
+
+    proc = subprocess.run(
+        [sys.executable, str(driver)],
+        cwd=tmp_path,
+        capture_output=True,
+        text=True,
+        timeout=30,
+    )
+
+    assert proc.returncode == 0, proc.stdout + proc.stderr
+    assert (tmp_path / "output.xlsx").read_text(encoding="utf-8") == "ABSENT|TMPDIR"
+    assert not (tmp_path / "_driver_runner.py").exists()