Block test_run for Octo

src/octopal/cli/main.py

@@ -2350,7 +2350,7 @@ def _octo_tool_policy_steps() -> list[ToolPolicyPipelineStep]:
         ),
         ToolPolicyPipelineStep(
             label="octo.direct_exec_denylist",
-            policy=ToolPolicy(deny=["exec_run"]),
+            policy=ToolPolicy(deny=["exec_run", "test_run"]),
         ),
     ]
 

src/octopal/runtime/octo/router.py

@@ -1718,7 +1718,7 @@ def _get_octo_tools(octo: Any, chat_id: int) -> tuple[list[ToolSpec], dict[str,
         ),
         ToolPolicyPipelineStep(
             label="octo.direct_exec_denylist",
-            policy=ToolPolicy(deny=["exec_run"]),
+            policy=ToolPolicy(deny=["exec_run", "test_run"]),
         ),
     ]
     all_tools = get_tools(mcp_manager=mcp_manager)

tests/test_cli_tools_resolve.py

@@ -31,6 +31,7 @@ def test_build_tool_resolution_snapshot_for_octo_applies_policy_and_profile() ->
     assert "web_fetch" in blocked_rows
     assert blocked_rows["web_fetch"]["reason"] == "blocked_by_deny:octo.raw_fetch_denylist"
     assert "exec_run" in blocked_rows
+    assert "test_run" in blocked_rows
 
 
 def test_build_tool_resolution_snapshot_for_octo_blocks_direct_exec_without_profile() -> None:
@@ -45,7 +46,9 @@ def test_build_tool_resolution_snapshot_for_octo_blocks_direct_exec_without_prof
     blocked_rows = {row["name"]: row for row in snapshot["blocked"]}
 
     assert "exec_run" not in available_names
+    assert "test_run" not in available_names
     assert blocked_rows["exec_run"]["reason"] == "blocked_by_deny:octo.direct_exec_denylist"
+    assert blocked_rows["test_run"]["reason"] == "blocked_by_deny:octo.direct_exec_denylist"
 
 
 def test_tools_resolve_json_outputs_snapshot() -> None:

tests/test_octo_tool_loop.py

@@ -48,6 +48,7 @@ class DummyOcto:
     tool_specs, ctx = _get_octo_tools(DummyOcto(), 0)
 
     assert "exec_run" not in {tool.name for tool in tool_specs}
+    assert "test_run" not in {tool.name for tool in tool_specs}
 
     async def scenario() -> None:
         result, meta = await _handle_octo_tool_call(
@@ -64,6 +65,29 @@ async def scenario() -> None:
     asyncio.run(scenario())
 
 
+def test_default_octo_tool_policy_blocks_test_run() -> None:
+    class DummyOcto:
+        mcp_manager = None
+
+    tool_specs, ctx = _get_octo_tools(DummyOcto(), 0)
+
+    assert "test_run" not in {tool.name for tool in tool_specs}
+
+    async def scenario() -> None:
+        result, meta = await _handle_octo_tool_call(
+            {"function": {"name": "test_run", "arguments": '{"command":"pytest -q"}'}},
+            tool_specs,
+            ctx,
+        )
+
+        assert result["type"] == "policy_block"
+        assert result["tool"] == "test_run"
+        assert result["reason"] == "blocked_by_deny:octo.direct_exec_denylist"
+        assert meta["error_type"] == "policy_block"
+
+    asyncio.run(scenario())
+
+
 @pytest.mark.asyncio
 async def test_handle_octo_tool_call_captures_tool_exceptions() -> None:
     def _boom(_args, _ctx):

tests/test_router_tool_budget.py

@@ -98,7 +98,9 @@ def fake_get_tools(mcp_manager=None):
     assert "octo_check_update" in names
     assert "octo_update_self" in names
     assert "exec_run" not in names
+    assert "test_run" not in names
     assert "exec_run" not in {spec.name for spec in ctx["tool_resolution_report"].available_tools}
+    assert "test_run" not in {spec.name for spec in ctx["tool_resolution_report"].available_tools}
     assert "mcp_demo_tool_0" not in names
     assert "mcp_demo_tool_0" in all_names
     assert len(tool_specs) < len(all_names)