feat: 持久化诊断历史并修复 Agent 命令执行

tests/test_diagnostic_agent.py

@@ -27,8 +27,11 @@ class DiagnosticAgentTest(unittest.TestCase):
                 return await collect_events(run_diagnostic_agent("report", max_steps=3))
 
         events = asyncio.run(run())
-        self.assertEqual([event["type"] for event in events], ["thinking", "command", "result", "thinking", "final"])
-        self.assertEqual(events[1]["mode"], "read")
+        self.assertEqual(
+            [event["type"] for event in events],
+            ["command", "result", "thinking", "command", "result", "thinking", "final"],
+        )
+        self.assertEqual(events[3]["mode"], "read")
         self.assertEqual(events[-1]["content"], "发现一个异常 Pod")
 
     def test_pauses_mutating_command_for_human_approval(self):
@@ -63,13 +66,18 @@ class DiagnosticAgentTest(unittest.TestCase):
         async def fake_completion(messages):
             return '{"action":"command","command":"kubectl delete pod api-0 -n prod","reason":"删除故障 Pod"}'
 
-        async def run():
-            with patch("backend.ai_agent._complete_chat", fake_completion), patch("backend.ai_agent.execute_kubectl") as execute:
-                events = await collect_events(run_diagnostic_agent("report", max_steps=3, execution_mode="ai"))
-                return events, execute
+        executed = []
 
-        events, execute = asyncio.run(run())
-        execute.assert_not_called()
+        async def run():
+            async def fake_execute(command):
+                executed.append(command)
+                return {"command": command, "mode": "read", "returncode": 0, "stdout": "nodes ready", "stderr": ""}
+
+            with patch("backend.ai_agent._complete_chat", fake_completion), patch("backend.ai_agent.execute_kubectl", fake_execute):
+                return await collect_events(run_diagnostic_agent("report", max_steps=3, execution_mode="ai"))
+
+        events = asyncio.run(run())
+        self.assertEqual(executed, ["kubectl get nodes -o wide"])
         self.assertEqual(events[-1]["type"], "approval_required")
 
     def test_rejects_unknown_execution_mode(self):
@@ -80,6 +88,39 @@ class DiagnosticAgentTest(unittest.TestCase):
         self.assertEqual(events[-1]["type"], "error")
         self.assertIn("执行模式", events[-1]["message"])
 
+    def test_ai_mode_forces_initial_read_only_probe_before_model_planning(self):
+        responses = iter([
+            '{"action":"final","analysis":"根据采集结果完成分析"}',
+        ])
+
+        async def fake_completion(messages):
+            return next(responses)
+
+        executed = []
+
+        async def fake_execute(command):
+            executed.append(command)
+            return {"command": command, "mode": "read", "returncode": 0, "stdout": "node-a Ready", "stderr": ""}
+
+        async def run():
+            with patch("backend.ai_agent._complete_chat", fake_completion), patch("backend.ai_agent.execute_kubectl", fake_execute):
+                return await collect_events(run_diagnostic_agent("report", max_steps=3, execution_mode="ai"))
+
+        events = asyncio.run(run())
+        self.assertEqual(executed, ["kubectl get nodes -o wide"])
+        self.assertEqual([event["type"] for event in events[:2]], ["command", "result"])
+        self.assertEqual(events[-1]["type"], "final")
+
+    def test_parse_agent_action_accepts_markdown_json_fence(self):
+        from backend.ai_agent import _parse_agent_action
+
+        action = _parse_agent_action(
+            '我将先检查 Pod。\n```json\n{"action":"command","command":"kubectl get pods -A","reason":"查看 Pod"}\n```'
+        )
+
+        self.assertEqual(action["action"], "command")
+        self.assertEqual(action["command"], "kubectl get pods -A")
+
     def test_stops_when_model_returns_invalid_action(self):
         async def fake_completion(messages):
             return "not-json"