Refactor session runtime persistence

tests/test_agent_runtime.py

@@ -1025,8 +1025,13 @@ class AgentRuntimeTests(unittest.TestCase):
             message for message in result.transcript
             if message.get('metadata', {}).get('kind') == 'compact_boundary'
         ]
-        self.assertEqual(len(transcript_compact_messages), 1)
-        compact_metadata = transcript_compact_messages[0].get('metadata', {})
+        self.assertEqual(len(transcript_compact_messages), 0)
+        model_compact_messages = [
+            message for message in (agent.last_session.messages if agent.last_session else [])
+            if message.metadata.get('kind') == 'compact_boundary'
+        ]
+        self.assertEqual(len(model_compact_messages), 1)
+        compact_metadata = model_compact_messages[0].metadata
         self.assertEqual(compact_metadata.get('compaction_depth'), 1)
         self.assertEqual(compact_metadata.get('nested_compaction_count'), 0)
         self.assertIn('preserved_tail_ids', compact_metadata)

tests/test_compact.py

@@ -255,6 +255,10 @@ class TestCompactConversation(unittest.TestCase):
         self.assertNotIn('<analysis>', result.summary_text)
         # Summary should contain the actual summary content
         self.assertIn('User wanted to test compaction', result.summary_text)
+        # UI display transcript must remain append-only and not be replaced by
+        # compact_summary / compact_boundary model-context messages.
+        display_contents = [m.content for m in agent.last_session.display_messages]
+        self.assertEqual(display_contents, [m.content for m in msgs])
 
     def test_api_error_returns_compaction_error(self) -> None:
         with tempfile.TemporaryDirectory() as tmp_dir:

tests/test_gui_server.py

@@ -483,6 +483,21 @@ class GuiServerTests(unittest.TestCase):
                 },
                 {'role': 'user', 'content': 'continue'},
             ),
+            display_messages=(
+                {'role': 'user', 'content': 'hello'},
+                {
+                    'role': 'assistant',
+                    'content': '上一次任务已中断，后台没有正在执行的进程。',
+                    'state': 'final',
+                    'stop_reason': 'interrupted',
+                    'metadata': {'kind': 'run_status', 'status': 'interrupted'},
+                },
+                {
+                    'role': 'user',
+                    'content': 'pending',
+                    'metadata': {'status': 'running', 'placeholder': True},
+                },
+            ),
             turns=1,
             tool_calls=0,
             usage={},
@@ -498,6 +513,10 @@ class GuiServerTests(unittest.TestCase):
             [message['content'] for message in sanitized.messages],
             ['hello', 'continue'],
         )
+        self.assertEqual(
+            [message['content'] for message in sanitized.display_messages],
+            ['hello'],
+        )
 
     def test_runtime_event_stage_names_long_running_work(self) -> None:
         self.assertEqual(

tests/test_session_store.py

@@ -80,6 +80,7 @@ class TestStoredAgentSessionRoundTrip(unittest.TestCase):
             'user_context': {'lang': 'en'},
             'system_context': {'os': 'linux'},
             'messages': ({'role': 'user', 'content': 'hi'},),
+            'display_messages': ({'role': 'user', 'content': 'visible hi'},),
             'turns': 3,
             'tool_calls': 7,
             'usage': {'input_tokens': 500, 'output_tokens': 300},
@@ -106,6 +107,7 @@ class TestStoredAgentSessionRoundTrip(unittest.TestCase):
         self.assertEqual(loaded.user_context, session.user_context)
         self.assertEqual(loaded.system_context, session.system_context)
         self.assertEqual(loaded.messages, session.messages)
+        self.assertEqual(loaded.display_messages, session.display_messages)
         self.assertEqual(loaded.turns, session.turns)
         self.assertEqual(loaded.tool_calls, session.tool_calls)
         self.assertEqual(loaded.usage, session.usage)
@@ -157,6 +159,30 @@ class TestStoredAgentSessionRoundTrip(unittest.TestCase):
         self.assertEqual(len(loaded.messages), 2)
         self.assertEqual(loaded.messages[0]['role'], 'user')
         self.assertEqual(loaded.messages[1]['role'], 'assistant')
+        self.assertEqual(loaded.display_messages, loaded.messages)
+
+    def test_load_agent_session_falls_back_to_messages_for_display(self) -> None:
+        with tempfile.TemporaryDirectory() as td:
+            directory = Path(td)
+            path = directory / 'legacy.json'
+            data = {
+                'session_id': 'legacy',
+                'model_config': {},
+                'runtime_config': {'cwd': '/'},
+                'system_prompt_parts': [],
+                'user_context': {},
+                'system_context': {},
+                'messages': [{'role': 'user', 'content': 'legacy hi'}],
+                'turns': 0,
+                'tool_calls': 0,
+                'usage': {},
+                'total_cost_usd': 0.0,
+                'file_history': [],
+            }
+            path.write_text(json.dumps(data))
+            loaded = load_agent_session('legacy', directory=directory)
+
+        self.assertEqual(loaded.display_messages, loaded.messages)
 
     def test_load_defaults_for_missing_optional_fields(self) -> None:
         """Missing optional fields get sensible defaults."""