feat(agent-runtime): 补全测试替身编排闭环

2026-09-16 17:31:12 +08:00
parent d3763633c2
commit 1e56a55399
10 changed files with 302 additions and 11 deletions
@@ -0,0 +1,38 @@
+"""Agent 运行配置边界。"""
+
+from dataclasses import dataclass
+
+from django.conf import settings
+
+from common.exceptions import AgentConfigurationError
+
+
+@dataclass(frozen=True)
+class AgentRuntimeConfig:
+    """集中表达真实 SDK 执行所需配置，避免业务服务直接读取环境变量。"""
+
+    api_key: str
+    model: str
+    agent_name: str
+
+    @classmethod
+    def from_settings(cls) -> "AgentRuntimeConfig":
+        """从 Django 设置构建配置；这里只读取，不在日志或异常中回显秘密。"""
+
+        return cls(
+            api_key=str(settings.OPENAI_API_KEY or "").strip(),
+            model=str(settings.OPENAI_MODEL or "").strip(),
+            agent_name=str(settings.AGENT_DEFAULT_NAME or "job_research").strip(),
+        )
+
+    def require_real_execution(self) -> "AgentRuntimeConfig":
+        """真实执行前快速失败；非 Agent 页面无需调用此方法。"""
+
+        missing = []
+        if not self.api_key:
+            missing.append("OPENAI_API_KEY")
+        if not self.model:
+            missing.append("OPENAI_MODEL")
+        if missing:
+            raise AgentConfigurationError(f"真实 Agent 执行缺少配置：{', '.join(missing)}。")
+        return self
@@ -0,0 +1,51 @@
+"""Gateway 执行与本地 Agent Run 持久化之间的编排服务。"""
+
+import logging
+
+from common.logging import sanitize_summary
+
+from .gateway import AgentExecutionRequest, AgentRunnerGateway
+from .models import AgentRun, RunStatus
+from .services import transition_run
+
+logger = logging.getLogger(__name__)
+
+
+def execute_run(run_id, gateway: AgentRunnerGateway) -> AgentRun:
+    """使用注入的 Gateway 驱动一次运行，并确保所有结果都进入稳定终态。
+
+    外部调用发生在两个短事务之间：开始状态先提交，Gateway 返回或抛错后再用
+    独立事务保存成功或失败结果。该函数不重试，避免第一阶段产生重复外部副作用。
+    """
+
+    running = transition_run(run_id, RunStatus.RUNNING)
+    request = AgentExecutionRequest(
+        agent_run_id=str(running.id),
+        user_id=running.owner_id,
+        input_summary=sanitize_summary(running.input_summary),
+    )
+    try:
+        result = gateway.run(request)
+    except Exception as exc:  # Gateway 是外部边界，必须把未知异常转换为可审计失败。
+        # 外部异常消息可能夹带请求参数或认证信息，因此日志和数据库只保留异常类型。
+        error_type = type(exc).__name__
+        logger.error(
+            "Agent Gateway 执行异常，异常类型=%s",
+            error_type,
+            extra={"agent_run_id": str(running.id)},
+        )
+        return transition_run(
+            running.id,
+            RunStatus.FAILED,
+            error_code="gateway_error",
+            error_summary=f"Agent Gateway 执行异常：{error_type}",
+        )
+
+    if result.succeeded:
+        return transition_run(running.id, RunStatus.SUCCEEDED, output=result.output_summary)
+    return transition_run(
+        running.id,
+        RunStatus.FAILED,
+        error_code=result.error_code or "agent_failed",
+        error_summary=result.error_summary,
+    )
@@ -1,11 +1,29 @@
 from django.contrib.auth import get_user_model
-from django.test import TestCase
+from django.db import IntegrityError, transaction
+from django.test import SimpleTestCase, TestCase, override_settings
 from django.urls import reverse

-from common.exceptions import InvalidStateTransition
+from common.exceptions import AgentConfigurationError, InvalidStateTransition, PermissionDenied

-from .models import RunStatus
-from .services import create_run, finish_tool_call, start_tool_call, transition_run
+from .config import AgentRuntimeConfig
+from .gateway import AgentExecutionResult, StubAgentRunnerGateway
+from .models import ApprovalStatus, RunStatus, ToolCallStatus
+from .orchestration import execute_run
+from .services import (
+    create_run,
+    finish_tool_call,
+    request_approval,
+    resolve_approval,
+    start_tool_call,
+    transition_run,
+)
+
+
+class ExplodingGateway:
+    """模拟外部边界抛错，并故意在异常文本中携带不应落库的秘密。"""
+
+    def run(self, request):
+        raise RuntimeError("api_key=should-not-be-stored")


 class AgentRunTests(TestCase):
@@ -32,6 +50,17 @@ class AgentRunTests(TestCase):
        response = self.client.get(reverse("agent_runtime:run-detail", args=(run.id,)))
        self.assertEqual(response.status_code, 404)

+    def test_run_list_only_contains_current_user_data(self):
+        create_run(self.alice, "爱丽丝的运行")
+        create_run(self.bob, "鲍勃的运行")
+        self.client.force_login(self.alice)
+
+        response = self.client.get(reverse("agent_runtime:run-list"))
+
+        self.assertEqual(response.status_code, 200)
+        self.assertContains(response, "爱丽丝的运行")
+        self.assertNotContains(response, "鲍勃的运行")
+
    def test_tool_call_is_recorded_and_sanitized(self):
        run = create_run(self.alice, "工具运行")
        transition_run(run.id, RunStatus.RUNNING)
@@ -40,3 +69,80 @@ class AgentRunTests(TestCase):
        finished = finish_tool_call(call.id, result={"count": 1})
        self.assertEqual(finished.status, "succeeded")
        self.assertEqual(finished.result_summary, {"count": 1})
+
+    def test_tool_call_failure_and_duplicate_finish_are_recorded(self):
+        run = create_run(self.alice, "失败工具运行")
+        transition_run(run.id, RunStatus.RUNNING)
+        call = start_tool_call(run.id, "call-1", "demo_tool", idempotency_key="same-operation")
+        failed = finish_tool_call(call.id, error_code="timeout", error_summary="请求超时")
+        self.assertEqual(failed.status, ToolCallStatus.FAILED)
+        self.assertEqual(failed.error_code, "timeout")
+        with self.assertRaises(InvalidStateTransition):
+            finish_tool_call(call.id, result={"unexpected": True})
+
+    def test_duplicate_tool_idempotency_key_is_rejected(self):
+        run = create_run(self.alice, "幂等工具运行")
+        transition_run(run.id, RunStatus.RUNNING)
+        start_tool_call(run.id, "call-1", "demo_tool", idempotency_key="same-operation")
+        with self.assertRaises(IntegrityError), transaction.atomic():
+            start_tool_call(run.id, "call-2", "demo_tool", idempotency_key="same-operation")
+
+    def test_approval_can_be_approved_once_by_owner(self):
+        run = create_run(self.alice, "确认运行")
+        transition_run(run.id, RunStatus.RUNNING)
+        approval = request_approval(run.id, "approval-1", "external_action", {"token": "secret"})
+        self.assertEqual(approval.request_summary["token"], "***")
+        resolved = resolve_approval(self.alice, approval.id, True, {"reason": "允许"})
+        self.assertEqual(resolved.status, ApprovalStatus.APPROVED)
+        run.refresh_from_db()
+        self.assertEqual(run.status, RunStatus.RUNNING)
+        with self.assertRaises(InvalidStateTransition):
+            resolve_approval(self.alice, approval.id, True)
+
+    def test_approval_rejects_cross_user_and_can_cancel_run(self):
+        run = create_run(self.alice, "拒绝确认运行")
+        transition_run(run.id, RunStatus.RUNNING)
+        approval = request_approval(run.id, "approval-1", "external_action")
+        with self.assertRaises(PermissionDenied):
+            resolve_approval(self.bob, approval.id, False)
+        resolve_approval(self.alice, approval.id, False)
+        run.refresh_from_db()
+        self.assertEqual(run.status, RunStatus.CANCELLED)
+
+    def test_stub_gateway_drives_successful_persistent_run(self):
+        run = create_run(self.alice, "Stub 成功运行", {"query": "Python"})
+        gateway = StubAgentRunnerGateway(AgentExecutionResult(True, {"count": 2}))
+        finished = execute_run(run.id, gateway)
+        self.assertEqual(finished.status, RunStatus.SUCCEEDED)
+        self.assertEqual(finished.output_summary, {"count": 2})
+        self.assertEqual(finished.events.count(), 3)
+
+    def test_stub_gateway_failure_and_exception_reach_failed_state(self):
+        failed_run = create_run(self.alice, "Stub 失败运行")
+        failed_gateway = StubAgentRunnerGateway(
+            AgentExecutionResult(False, error_code="model_error", error_summary="模型失败")
+        )
+        failed = execute_run(failed_run.id, failed_gateway)
+        self.assertEqual(failed.status, RunStatus.FAILED)
+        self.assertEqual(failed.error_code, "model_error")
+
+        exploding_run = create_run(self.alice, "Stub 异常运行")
+        exploded = execute_run(exploding_run.id, ExplodingGateway())
+        self.assertEqual(exploded.status, RunStatus.FAILED)
+        self.assertNotIn("should-not-be-stored", exploded.error_summary)
+
+
+class AgentRuntimeConfigTests(SimpleTestCase):
+    """验证真实执行配置只在调用边界检查，不影响其他 Django 功能。"""
+
+    @override_settings(OPENAI_API_KEY="", OPENAI_MODEL="", AGENT_DEFAULT_NAME="job_research")
+    def test_real_execution_requires_key_and_model(self):
+        with self.assertRaises(AgentConfigurationError):
+            AgentRuntimeConfig.from_settings().require_real_execution()
+
+    @override_settings(
+        OPENAI_API_KEY="test-key", OPENAI_MODEL="test-model", AGENT_DEFAULT_NAME="job_research"
+    )
+    def test_complete_configuration_is_accepted_without_external_call(self):
+        config = AgentRuntimeConfig.from_settings().require_real_execution()
+        self.assertEqual(config.model, "test-model")