New web interface, move llm engine to separate process

test/web_agent_test.py

@@ -10,10 +10,9 @@ from sia.web_io_buffer import WebIOBuffer
 from sia.working_memory import WorkingMemory
 from sia.system_metrics import SystemMetrics
 from sia.llm_engine import LlmEngine
-from sia.xml_validator import XMLValidator
 from sia.response_parser import ResponseParser
 from sia.iteration_logger import IterationLogger
-from sia.web_agent import WebAgent, LlmState
+from sia.web_agent import WebAgent, AgentState
 from sia.command import Command
 from sia.command_result import CommandResult
 from sia.entry.reasoning_entry import ReasoningEntry
@@ -22,7 +21,7 @@ from sia.entry import Entry
 from sia.response_buffer import ResponseBuffer
 from pathlib import Path
 
-class MockLlmEngine(LlmEngine):
+class MockLlmEngine:
     """Mock LLM engine for testing."""
     
     def __init__(self, output: str = "<reasoning>test reasoning</reasoning>"):
@@ -67,14 +66,10 @@ class WebAgentTest(unittest.TestCase):
             'default': MockLlmEngine(),
             'alternative': MockLlmEngine("<reasoning>alternative reasoning</reasoning>")
         }
-        self.mock_validator = Mock(spec=XMLValidator)
         self.io_buffer = WebIOBuffer()
         self.working_memory = WorkingMemory()
         self.work_dir = Path("/tmp")  # Use temp directory for tests
         
-        # Mock validator to pass by default
-        self.mock_validator.validate.return_value = None
-        
         # Create parser with IO buffer
         self.parser = ResponseParser(self.work_dir, self.io_buffer)
         
@@ -98,7 +93,6 @@ class WebAgentTest(unittest.TestCase):
             working_memory=self.working_memory,
             metrics=self.mock_metrics,
             llms=self.mock_llms,
-            validator=self.mock_validator,
             parser=self.parser,
             iteration_logger=self.mock_iteration_logger
         )
@@ -107,7 +101,7 @@ class WebAgentTest(unittest.TestCase):
         self.llm_state_changes = []
         self.context_changes = []
     
-    def llm_change_handler(self, llm_name: str, state: LlmState):
+    def llm_change_handler(self, llm_name: str, state: AgentState):
         """Track LLM state changes for verification."""
         self.llm_state_changes.append((llm_name, state))
     
@@ -124,23 +118,11 @@ class WebAgentTest(unittest.TestCase):
         # Check LLM states
         llm_states = self.agent.llms
         self.assertEqual(len(llm_states), 2)
-        self.assertEqual(llm_states['default'], LlmState.IDLE)
-        self.assertEqual(llm_states['alternative'], LlmState.IDLE)
+        self.assertEqual(llm_states['default'], AgentState.IDLE)
+        self.assertEqual(llm_states['alternative'], AgentState.IDLE)
         
         # Check response buffer
         self.assertIsInstance(self.agent.response_buffer, ResponseBuffer)
-        
-    def test_handler_registration(self):
-        """Test adding state and context change handlers."""
-        self.agent.add_llm_change_handler(self.llm_change_handler)
-        self.agent.add_context_change_handler(self.context_change_handler)
-        
-        # Adding same handler twice should not duplicate
-        self.agent.add_llm_change_handler(self.llm_change_handler)
-        self.agent.add_context_change_handler(self.context_change_handler)
-        
-        self.assertEqual(len(self.agent._llm_change_handlers), 1)
-        self.assertEqual(len(self.agent._context_change_handlers), 1)
     
     def test_run_inference(self):
         """Test running inference."""
@@ -153,8 +135,8 @@ class WebAgentTest(unittest.TestCase):
             
             # Check state changes
             self.assertEqual(len(self.llm_state_changes), 2)  # IDLE -> INFERENCE -> IDLE
-            self.assertEqual(self.llm_state_changes[0], ('default', LlmState.INFERENCE))
-            self.assertEqual(self.llm_state_changes[1], ('default', LlmState.IDLE))
+            self.assertEqual(self.llm_state_changes[0], ('default', AgentState.INFERENCE))
+            self.assertEqual(self.llm_state_changes[1], ('default', AgentState.IDLE))
             
             # Verify LLM was called
             self.assertTrue(self.mock_llms['default'].infer_called)
@@ -171,7 +153,7 @@ class WebAgentTest(unittest.TestCase):
     def test_run_inference_busy_llm(self):
         """Test running inference on a busy LLM."""
         # Set LLM state to INFERENCE
-        with patch.object(self.agent, '_llm_states', {'default': LlmState.INFERENCE}):
+        with patch.object(self.agent, '_llm_states', {'default': AgentState.INFERENCE}):
             with self.assertRaises(RuntimeError):
                 self.agent.run_inference('default')
     
@@ -187,7 +169,6 @@ class WebAgentTest(unittest.TestCase):
             working_memory=WorkingMemory(),
             metrics=self.mock_metrics,
             llms={'test': mock_engine},
-            validator=self.mock_validator,
             parser=self.parser,
             iteration_logger=self.mock_iteration_logger
         )
@@ -224,7 +205,6 @@ class WebAgentTest(unittest.TestCase):
             working_memory=WorkingMemory(),
             metrics=self.mock_metrics,
             llms=self.mock_llms,
-            validator=self.mock_validator,
             parser=self.parser,
             iteration_logger=self.mock_iteration_logger
         )
@@ -254,7 +234,6 @@ class WebAgentTest(unittest.TestCase):
             working_memory=WorkingMemory(),
             metrics=self.mock_metrics,
             llms=self.mock_llms,
-            validator=self.mock_validator,
             parser=self.parser,
             iteration_logger=self.mock_iteration_logger
         )
@@ -290,7 +269,7 @@ class WebAgentTest(unittest.TestCase):
             self.assertTrue(mock_command.executed)
             
             # Check command result was stored
-            self.assertEqual(self.agent.command_result, command_result)
+            self.assertEqual(self.agent._command_result, command_result)
             
             # Verify iteration was logged
             self.mock_iteration_logger.log_iteration.assert_called_once()
@@ -329,7 +308,6 @@ class WebAgentTest(unittest.TestCase):
             working_memory=test_memory,
             metrics=self.mock_metrics,
             llms=self.mock_llms,
-            validator=self.mock_validator,
             parser=self.parser,
             iteration_logger=self.mock_iteration_logger
         )