fix(openclaw): validate configured skill paths

adapters/python/openclaw/src/nemo_fabric_adapters/openclaw/adapter.py

@@ -263,6 +263,27 @@ def _channel_secret_env_names(value: Any) -> set[str]:
     return names
 
 
+def _resolve_skill_paths(config: contract.AgentConfig, base_dir: Path) -> list[str]:
+    paths: list[str] = []
+    for index, path in enumerate(config.skills.paths if config.skills else []):
+        try:
+            directory = (base_dir / path).resolve(strict=True)
+        except (OSError, RuntimeError):
+            raise lifecycle.LifecycleError(
+                "openclaw_skill_not_found",
+                "A configured NeMo Fabric skill path does not exist",
+                metadata={"field": f"skills.paths[{index}]"},
+            ) from None
+        if not directory.is_dir() or not (directory / "SKILL.md").is_file():
+            raise lifecycle.LifecycleError(
+                "openclaw_skill_invalid",
+                "NeMo Fabric skill paths must be directories containing a SKILL.md file",
+                metadata={"field": f"skills.paths[{index}]"},
+            )
+        paths.append(str(directory))
+    return paths
+
+
 def _resolve_command(settings: dict[str, Any], base_dir: Path) -> Path:
     configured = settings.get("openclaw_command", "openclaw")
     if not isinstance(configured, str) or not configured.strip():
@@ -468,14 +489,9 @@ def _openclaw_config(
     result["agents"]["entries"] = {agent_id: {}}
     if config.instructions is not None and config.instructions.system is not None:
         result["agents"]["defaults"]["contextInjection"] = "never"
-    if config.skills and config.skills.paths:
-        result["skills"] = {
-            "load": {
-                "extraDirs": [
-                    str((base_dir / path).resolve()) for path in config.skills.paths
-                ]
-            }
-        }
+    skill_paths = _resolve_skill_paths(config, base_dir)
+    if skill_paths:
+        result["skills"] = {"load": {"extraDirs": skill_paths}}
     if config.tools is not None:
         tools: dict[str, list[str]] = {}
         if config.tools.enabled == []:

tests/adapters/test_openclaw.py

@@ -49,7 +49,7 @@ def _context(workspace: Path, runtime_id: str = "openclaw-runtime") -> RuntimeCo
     return RuntimeContext.from_mapping(payload)
 
 
-def _config(command: Path) -> AgentConfig:
+def _config(command: Path, *, with_skills: bool = False) -> AgentConfig:
     settings: dict[str, object] = {"openclaw_command": str(command)}
     return AgentConfig.from_mapping(
         {
@@ -65,7 +65,7 @@ def _config(command: Path) -> AgentConfig:
                     "max_tokens": 64,
                 }
             },
-            "skills": {"paths": ["skills"]},
+            "skills": {"paths": ["skills"]} if with_skills else None,
             "tools": {
                 "enabled": ["browser", "web_search"],
                 "blocked": ["exec"],
@@ -85,7 +85,7 @@ def _config(command: Path) -> AgentConfig:
                         "custom_headers": {"X-Test": "value"},
                         "blocked_tools": ["delete_*"],
                     },
-                }
+                },
             },
         }
     )
@@ -249,6 +249,9 @@ def test_openclaw_gateway_command_uses_setpriv_on_linux_when_available(
 async def test_openclaw_runtime_generates_config_invokes_and_cleans_up(
     mock_openclaw: Path, tmp_path: Path
 ):
+    skills = tmp_path / "skills"
+    skills.mkdir()
+    (skills / "SKILL.md").write_text("# Test skill\n", encoding="utf-8")
     capture = tmp_path / "config.json"
     request_capture = tmp_path / "request.json"
     readonly_capture = tmp_path / "config-readonly.txt"
@@ -264,7 +267,7 @@ async def test_openclaw_runtime_generates_config_invokes_and_cleans_up(
 
     await runtime.start(
         {
-            "config": _config(mock_openclaw),
+            "config": _config(mock_openclaw, with_skills=True),
             "runtime_context": context.to_mapping(),
             "base_dir": str(tmp_path),
         }
@@ -333,6 +336,39 @@ async def test_openclaw_runtime_generates_config_invokes_and_cleans_up(
     assert not state_root.exists()
 
 
+@pytest.mark.parametrize(
+    ("invalid_path", "expected_code"),
+    [
+        ("missing", "openclaw_skill_not_found"),
+        ("file", "openclaw_skill_invalid"),
+        ("missing_manifest", "openclaw_skill_invalid"),
+    ],
+)
+def test_openclaw_rejects_invalid_skill_paths(
+    mock_openclaw: Path,
+    tmp_path: Path,
+    invalid_path: str,
+    expected_code: str,
+):
+    skill_path = tmp_path / "skills"
+    if invalid_path == "file":
+        skill_path.write_text("not a directory", encoding="utf-8")
+    elif invalid_path == "missing_manifest":
+        skill_path.mkdir()
+
+    with pytest.raises(adapter.lifecycle.LifecycleError) as caught:
+        adapter._openclaw_config(
+            _config(mock_openclaw, with_skills=True),
+            _context(tmp_path),
+            base_dir=tmp_path,
+            port=20_000,
+            token_env="OPENCLAW_GATEWAY_TOKEN",
+        )
+
+    assert caught.value.code == expected_code
+    assert caught.value.metadata == {"field": "skills.paths[0]"}
+
+
 def test_openclaw_preserves_context_injection_without_system_instruction(
     mock_openclaw: Path, tmp_path: Path
 ):
@@ -622,7 +658,7 @@ async def test_openclaw_channel_secret_refs_require_environment_variables(
 def test_openclaw_attach_rejects_deployment_owned_configuration(
     mock_openclaw: Path,
 ):
-    config = _config(mock_openclaw)
+    config = _config(mock_openclaw, with_skills=True)
     assert config.harness is not None
     config.harness.settings["channel_config"] = {
         "channels": {"telegram": {"enabled": True}},