fix: 实现 Agent 修复验证闭环

backend/ai_agent.py

@@ -3,6 +3,7 @@ import json
 import re
 import uuid
 from collections import Counter
+from typing import Optional
 import httpx
 from backend.config import AI_API_BASE, AI_API_KEY, AI_MODEL
 from backend.agent_tools import classify_kubectl_command, execute_kubectl
@@ -40,7 +41,9 @@ AGENT_SYSTEM_PROMPT = """你是 Kubernetes 自主诊断 Agent。你的首要职
 
 每轮必须只返回一个 JSON 对象，不要使用 Markdown 代码块：
 - 继续调查：{"action":"command","command":"kubectl ...","reason":"为什么执行"}
-- 完成分析：{"action":"final","analysis":"中文 Markdown 分析结论"}
+- 完成分析：{"action":"final","analysis":"中文 Markdown 分析结论","verified":true,"remaining_issues":0}
+
+只有完成修复后的全量复检，并确认 remaining_issues 为 0 时才能返回 final。只定位根因、执行命令成功或部分资源恢复都不能结束。
 """
 
 _PENDING_APPROVALS: dict[str, dict] = {}
@@ -72,6 +75,9 @@ def _parse_agent_action(content: str) -> dict:
         raise ValueError("AI 命令动作缺少 command")
     if action["action"] == "final" and not action.get("analysis"):
         raise ValueError("AI 最终动作缺少 analysis")
+    if action["action"] == "final":
+        if action.get("verified") is not True or action.get("remaining_issues") != 0:
+            raise ValueError("AI 最终动作必须确认 verified=true 且 remaining_issues=0")
     return action
 
 
@@ -130,17 +136,58 @@ async def run_diagnostic_agent(
             "content": "AI 执行模式已自动采集第一份现场证据：\n" + json.dumps(initial_result, ensure_ascii=False),
         })
 
-    for step in range(1, max_steps + 1):
+    async for event in _run_agent_loop(
+        messages=messages,
+        execution_mode=execution_mode,
+        max_steps=max_steps,
+        start_step=1,
+        command_counts=command_counts,
+        report=report,
+        question=question,
+    ):
+        yield event
+
+
+async def _run_agent_loop(
+    messages: list[dict],
+    execution_mode: str,
+    max_steps: int,
+    start_step: int,
+    command_counts: Counter,
+    report: str,
+    question: str,
+):
+    for step in range(start_step, max_steps + 1):
         yield {"type": "thinking", "step": step, "message": f"AI 正在规划第 {step} 轮诊断"}
         try:
             content = await _complete_chat(messages)
             action = _parse_agent_action(content)
+        except ValueError as exc:
+            if "最终动作必须确认" in str(exc):
+                messages.extend([
+                    {"role": "assistant", "content": content},
+                    {
+                        "role": "user",
+                        "content": (
+                            "不能结束诊断：尚未通过修复后的全量复检确认所有问题清零。"
+                            "请继续返回 command，检查节点、Pod、Deployment、事件以及原故障资源；"
+                            "只有 verified=true 且 remaining_issues=0 时才可返回 final。"
+                        ),
+                    },
+                ])
+                yield {"type": "verification_required", "message": "尚未确认所有问题清零，Agent 继续执行诊断验证"}
+                continue
+            yield {"type": "error", "message": str(exc)}
+            return
         except Exception as exc:
             yield {"type": "error", "message": str(exc)}
             return
 
         if action["action"] == "final":
-            yield {"type": "final", "content": action["analysis"], "model": AI_MODEL}
+            yield {
+                "type": "final", "content": action["analysis"], "model": AI_MODEL,
+                "verified": True, "remaining_issues": 0,
+            }
             return
 
         command = action["command"]
@@ -191,6 +238,13 @@ async def run_diagnostic_agent(
                 "reason": reason,
                 "created_by": AI_MODEL,
                 "run_id": "",
+                "report": report,
+                "question": question,
+                "execution_mode": execution_mode,
+                "messages": messages + [{"role": "assistant", "content": content}],
+                "next_step": step + 1,
+                "max_steps": max_steps,
+                "command_counts": dict(command_counts),
             }
             yield {
                 "type": "approval_required", "approval_id": approval_id,
@@ -214,6 +268,28 @@ async def run_diagnostic_agent(
     }
 
 
+async def resume_diagnostic_agent(pending: dict, result: dict):
+    """审批命令执行后，把真实结果交回原 Agent 上下文并继续诊断。"""
+    messages = list(pending.get("messages") or [])
+    messages.append({
+        "role": "user",
+        "content": (
+            "用户已批准命令，以下是实际执行结果。请继续自主诊断并验证问题是否解决：\n"
+            + json.dumps(result, ensure_ascii=False)
+        ),
+    })
+    async for event in _run_agent_loop(
+        messages=messages,
+        execution_mode=pending.get("execution_mode", "ai"),
+        max_steps=int(pending.get("max_steps", 30)),
+        start_step=int(pending.get("next_step", 1)),
+        command_counts=Counter(pending.get("command_counts") or {}),
+        report=pending.get("report", ""),
+        question=pending.get("question", ""),
+    ):
+        yield event
+
+
 def take_pending_approval(approval_id: str) -> dict | None:
     return _PENDING_APPROVALS.pop(approval_id, None)