fix: harden security beyond PR #249 — command injection, deps, path injection

plugins/openclaw/slash_sleep.py

@@ -19,6 +19,7 @@
 import json
 import os
 import shutil
+import subprocess
 import sys
 from pathlib import Path
 from datetime import datetime
@@ -104,8 +105,8 @@ def run_category(category: str, *, dry_run: bool = False) -> int:
 
     print(f"=== /sleep run {category}{' (dry-run)' if dry_run else ''} ===")
     print(f"  cmd: {' '.join(cmd)}")
-    rc = os.system(" ".join(f'"{c}"' for c in cmd))
-    return rc
+    result = subprocess.run(cmd)
+    return result.returncode
 
 
 def run_all(*, dry_run: bool = False) -> int:

requirements.txt

@@ -24,7 +24,7 @@ httpx>=0.27.0
 # json_repair>=0.61.0
 
 # ── Optional: WebUI dashboard ────────────────────
-# gradio>=4.0.0
+# gradio>=5.50.0
 
 # ── Optional: Documentation site ─────────────────
 # mkdocs-material>=9.5.0

skillopt/envs/spreadsheetbench/rollout.py

@@ -314,7 +314,8 @@ def process_one(
         # ── Stage 1: run ReAct agent on test case 1 ─────────────────────
         result["phase"] = "agent"
 
-        work_dir = tempfile.mkdtemp(prefix=f"react_{task_id}_")
+        safe_task_id = "".join(c if c.isalnum() or c in "-_" else "_" for c in str(task_id))
+        work_dir = tempfile.mkdtemp(prefix=f"react_{safe_task_id}_")
         try:
             # Copy input so agent works in an isolated directory
             work_input = os.path.join(work_dir, os.path.basename(ip1))

skillopt_webui/app.py

@@ -24,6 +24,38 @@
 
 PROJECT_ROOT = Path(__file__).resolve().parent.parent
 
+
+def _ensure_under_project(path: Path) -> Path:
+    """Resolve *path* and fail closed unless it stays under PROJECT_ROOT."""
+    resolved = path.resolve()
+    try:
+        resolved.relative_to(PROJECT_ROOT.resolve())
+    except ValueError:
+        raise ValueError(f"Path escapes project root: {path}")
+    return resolved
+
+
+def _resolve_project_path(user_path: str, *, subdir: str | None = None) -> Path:
+    """Resolve a UI-supplied path, optionally constrained to a project subdir.
+
+    UI callbacks must never trust Gradio component values (e.g. dropdown text):
+    traversal, absolute paths, and symlinks are all rejected here, at the point
+    the path is about to be consumed.
+    """
+    if not user_path:
+        raise ValueError("Path is empty")
+    candidate = Path(user_path)
+    if not candidate.is_absolute():
+        candidate = PROJECT_ROOT / candidate
+    resolved = _ensure_under_project(candidate)
+    if subdir is not None:
+        allowed = (PROJECT_ROOT / subdir).resolve()
+        try:
+            resolved.relative_to(allowed)
+        except ValueError:
+            raise ValueError(f"Path must be under {subdir!r}: {user_path}")
+    return resolved
+
 # Gradio moved where `theme` lives across versions: <=5 uses `Blocks(theme=...)`,
 # >=6 moved it to `launch()`. Detect the installed major so the WebUI works on
 # any supported version without an ignored-argument warning or a TypeError.
@@ -42,15 +74,86 @@ def discover_configs() -> list[str]:
 
 def load_config(path: str) -> dict:
     """Load a YAML config file."""
-    with open(PROJECT_ROOT / path) as f:
+    config_file = _resolve_project_path(path, subdir="configs")
+    with open(config_file) as f:
         return yaml.safe_load(f)
 
 
+def scan_outputs(out_dir: str) -> list:
+    """Digest experiment results strictly under PROJECT_ROOT.
+
+    The Output Explorer callback. Any path that escapes PROJECT_ROOT is
+    rejected (empty result) at the point data is read, so a traversal arg
+    can never read files outside the project.
+    """
+    rows = []
+    if not out_dir:
+        return rows
+    try:
+        base = _resolve_project_path(out_dir)
+    except ValueError:
+        return rows
+    if not base.exists() or not base.is_dir():
+        return rows
+    for bench_dir in sorted(base.iterdir()):
+        try:
+            bench_dir = _ensure_under_project(bench_dir)
+        except ValueError:
+            continue
+        if not bench_dir.is_dir():
+            continue
+        for run_dir in sorted(bench_dir.iterdir()):
+            try:
+                run_dir = _ensure_under_project(run_dir)
+            except ValueError:
+                continue
+            if not run_dir.is_dir():
+                continue
+            cfg_file = run_dir / "config.yaml"
+            score = "—"
+            steps = "—"
+            if cfg_file.exists():
+                try:
+                    cfg_file = _ensure_under_project(cfg_file)
+                    c = yaml.safe_load(cfg_file.read_text())
+                    steps = str(c.get("train", {}).get("num_steps", "—"))
+                except Exception:
+                    pass
+            # Try to find best score from logs
+            for log_f in run_dir.glob("**/*.jsonl"):
+                try:
+                    log_f = _ensure_under_project(log_f)
+                    with open(log_f) as f:
+                        for line in f:
+                            d = json.loads(line)
+                            if "score" in d:
+                                score = f"{d['score']:.4f}"
+                except Exception:
+                    pass
+            rows.append([
+                run_dir.name,
+                bench_dir.name,
+                score,
+                steps,
+            ])
+    return rows
+
+
 def config_to_display(cfg: dict) -> str:
     """Pretty-print config for display."""
     return yaml.dump(cfg, default_flow_style=False, sort_keys=False)
 
 
+def config_preview(path: str) -> str:
+    """Registered config-preview callback: YAML text for an in-tree config."""
+    if not path:
+        return ""
+    try:
+        return config_to_display(load_config(path))
+    except Exception as exc:
+        return f"Error: {exc}"
+
+
 def _can_connect_to_url(url: str, timeout: float = 0.5) -> bool:
     parsed = urlparse(url)
     host = parsed.hostname
@@ -113,7 +216,8 @@ def validate_training_config(
         if value is not None and value != ""
     ]
     try:
-        cfg = flatten_config(load_merged_config(str(PROJECT_ROOT / config_path), cfg_options))
+        config_file = _resolve_project_path(config_path, subdir="configs")
+        cfg = flatten_config(load_merged_config(str(config_file), cfg_options))
     except Exception as exc:
         return f"❌ Invalid config: {exc}"
 
@@ -490,7 +594,7 @@ def build_ui():
                             label="Config File",
                             value=configs[0] if configs else None,
                         )
-                        config_preview = gr.Code(
+                        config_preview_box = gr.Code(
                             label="Config Preview",
                             language="yaml",
                             interactive=False,
@@ -525,15 +629,7 @@ def build_ui():
 
                         status_text = gr.Textbox(label="Status", interactive=False)
 
-                def on_config_change(path):
-                    if path:
-                        try:
-                            return config_to_display(load_config(path))
-                        except Exception as e:
-                            return f"Error: {e}"
-                    return ""
-
-                config_dropdown.change(on_config_change, config_dropdown, config_preview)
+                config_dropdown.change(config_preview, config_dropdown, config_preview_box)
 
                 def on_launch(cfg_path, lr_val, sched, epochs, batch, workers,
                               slow_update, meta_skill, gate):
@@ -602,46 +698,6 @@ def on_refresh():
                     label="Experiments",
                 )
 
-                def scan_outputs(out_dir):
-                    rows = []
-                    if not out_dir:
-                        return rows
-                    base = PROJECT_ROOT / out_dir
-                    if not base.exists():
-                        return rows
-                    for bench_dir in sorted(base.iterdir()):
-                        if not bench_dir.is_dir():
-                            continue
-                        for run_dir in sorted(bench_dir.iterdir()):
-                            if not run_dir.is_dir():
-                                continue
-                            cfg_file = run_dir / "config.yaml"
-                            score = "—"
-                            steps = "—"
-                            if cfg_file.exists():
-                                try:
-                                    c = yaml.safe_load(cfg_file.read_text())
-                                    steps = str(c.get("train", {}).get("num_steps", "—"))
-                                except Exception:
-                                    pass
-                            # Try to find best score from logs
-                            for log_f in run_dir.glob("**/*.jsonl"):
-                                try:
-                                    with open(log_f) as f:
-                                        for line in f:
-                                            d = json.loads(line)
-                                            if "score" in d:
-                                                score = f"{d['score']:.4f}"
-                                except Exception:
-                                    pass
-                            rows.append([
-                                run_dir.name,
-                                bench_dir.name,
-                                score,
-                                steps,
-                            ])
-                    return rows
-
                 scan_btn.click(scan_outputs, output_dir, results_table)
 
     return app
@@ -668,6 +724,10 @@ def main():
     parser.add_argument("--host", type=str, default="127.0.0.1",
                         help="Server host. Default is localhost; use 0.0.0.0 "
                              "to expose publicly (no auth, use with care).")
+    parser.add_argument("--auth-user", type=str, default=None,
+                        help="Username for basic auth (or set SKILLOPT_WEBUI_USER).")
+    parser.add_argument("--auth-pass", type=str, default=None,
+                        help="Password for basic auth (or set SKILLOPT_WEBUI_PASS).")
     args = parser.parse_args()
 
     if args.host and args.host not in ("127.0.0.1", "localhost", "::1"):
@@ -679,8 +739,35 @@ def main():
             file=sys.stderr,
         )
 
+    if args.share:
+        print(
+            "⚠ warning: --share creates a public tunnel (gradio.live) with no "
+            "authentication by default. Anyone with the URL can start/stop "
+            "training and browse the filesystem via Output Explorer. "
+            "Use --auth-user / --auth-pass (or SKILLOPT_WEBUI_USER / "
+            "SKILLOPT_WEBUI_PASS) to require login.",
+            file=sys.stderr,
+        )
+
+    auth_user = args.auth_user or os.environ.get("SKILLOPT_WEBUI_USER")
+    auth_pass = args.auth_pass or os.environ.get("SKILLOPT_WEBUI_PASS")
+    # Fail-closed: authentication requires BOTH credentials. Supplying only a
+    # username or only a password must not silently launch the UI unauthenticated
+    # (a deployment could expose the training controls without login).
+    if bool(auth_user) != bool(auth_pass):
+        print(
+            "SKILLOPT_WEBUI authentication requires BOTH --auth-user and "
+            "--auth-pass (or SKILLOPT_WEBUI_USER and SKILLOPT_WEBUI_PASS). "
+            "Refusing to start with incomplete credentials.",
+            file=sys.stderr,
+        )
+        sys.exit(1)
+    auth = (auth_user, auth_pass) if auth_user else None
+
     app = build_ui()
     launch_kwargs = build_launch_kwargs(args.host, args.port, args.share)
+    if auth:
+        launch_kwargs["auth"] = auth
     app.launch(**launch_kwargs)
 
 

tests/test_webui_env_preflight.py

@@ -7,15 +7,17 @@
 
 
 def _write_config(tmp_path, model):
-    config_path = tmp_path / "config.yaml"
+    config_dir = tmp_path / "configs"
+    config_dir.mkdir(exist_ok=True)
+    config_path = config_dir / "demo.yaml"
     config_path.write_text(
         yaml.safe_dump({
             "model": model,
             "env": {"name": "searchqa"},
         }),
         encoding="utf-8",
     )
-    return str(config_path)
+    return "configs/demo.yaml"
 
 
 def test_build_training_env_loads_project_dotenv(tmp_path, monkeypatch):
@@ -37,6 +39,7 @@ def test_build_training_env_loads_project_dotenv(tmp_path, monkeypatch):
 
 
 def test_preflight_reports_missing_openai_chat_endpoint(tmp_path, monkeypatch):
+    monkeypatch.setattr(webui_app, "PROJECT_ROOT", tmp_path)
     monkeypatch.delenv("AZURE_OPENAI_ENDPOINT", raising=False)
     monkeypatch.delenv("OPTIMIZER_AZURE_OPENAI_ENDPOINT", raising=False)
     monkeypatch.delenv("TARGET_AZURE_OPENAI_ENDPOINT", raising=False)
@@ -56,6 +59,7 @@ def test_preflight_reports_missing_openai_chat_endpoint(tmp_path, monkeypatch):
 
 
 def test_preflight_reports_unreachable_qwen_endpoint(tmp_path, monkeypatch):
+    monkeypatch.setattr(webui_app, "PROJECT_ROOT", tmp_path)
     monkeypatch.setattr(webui_app, "_can_connect_to_url", lambda _url: False)
     config_path = _write_config(
         tmp_path,
@@ -74,6 +78,7 @@ def test_preflight_reports_unreachable_qwen_endpoint(tmp_path, monkeypatch):
 
 
 def test_preflight_accepts_reachable_qwen_endpoint(tmp_path, monkeypatch):
+    monkeypatch.setattr(webui_app, "PROJECT_ROOT", tmp_path)
     seen_urls = []
     monkeypatch.setattr(webui_app, "_can_connect_to_url", lambda url: seen_urls.append(url) or True)
     config_path = _write_config(