refactor(agent): expose split skill files in evaluation

app/agent/middleware/invocation.py

@@ -14,7 +14,7 @@ from pydantic import Field
 
 from app.agent.policy.api import resolve_api_operation
 from app.agent.policy.contracts import ActionEffect, ExecutionOutcome, ToolPolicyContext
-from app.agent.policy.sanitizer import sanitize_for_host
+from app.agent.policy.sanitizer import sanitize_for_host, summarize_error
 from app.agent.tools.base import run_agent_blocking
 from app.agent.tools.impl.api import MoviePilotApiTool
 from app.agent.tools.impl.mcp import McpExternalTool
@@ -223,6 +223,15 @@ class InvocationMiddleware(AgentMiddleware):  # type: ignore[misc]
             )
         except InvocationConflictError:
             return self._message(request, "failed", "同一调用 ID 的工具或参数发生变化，未执行。")
+        except ValueError as error:
+            detail = summarize_error(error, max_chars=240)
+            logger.info(f"工具参数未通过 API 输入合同校验: {detail}")
+            return self._message(
+                request,
+                "failed",
+                "API 工具参数未通过当前 operation 的输入合同校验。"
+                f"校验信息：{detail}。请重新读取对应 Skill 支持文档，只提交已声明的字段后再调用。",
+            )
         except Exception as error:
             logger.warning(f"工具执行认领失败: {type(error).__name__}")
             return self._message(request, "failed", "无法持久记录本次写入，未执行工具。")

scripts/evaluation/runtime.py

@@ -262,6 +262,7 @@ async def _run_isolated(
     from app.agent.contracts import ReplyMode
     from app.agent.memory import MemoryManager
     from app.agent.tools.impl.api import MoviePilotApiTool
+    from app.agent.tools.impl.read_file import ReadFileTool
     from app.db.adapters.invocation import TransactionalInvocationRepository
     from app.db.models.agentinvocation import AgentInvocation
     from app.schemas.types import NotificationChannel
@@ -295,6 +296,17 @@ async def _run_isolated(
             tool.set_agent_context({"is_admin": True, "should_dispatch_reply": False, "require_secret_confirmation": True}
                                    if child else agent._tool_context)
             (agent.evaluation_child_tools if child else agent.evaluation_tools).append(tool)
+            # Skill 主体只返回相对 supporting_files；评测必须提供与生产一致的
+            # read_file 能力，才能验证模型是否按需加载 api/*.md 合同。工具使用
+            # 非管理员上下文，只能读取当前临时 CONFIG_DIR/agent 隔离目录。
+            skill_file_tool = ReadFileTool(session_id=agent.session_id, user_id="1")
+            skill_file_tool.set_message_attr(agent.channel, agent.source, agent.username)
+            skill_file_tool.set_agent_context({
+                "is_admin": False,
+                "should_dispatch_reply": False,
+                "require_secret_confirmation": True,
+            })
+            (agent.evaluation_child_tools if child else agent.evaluation_tools).append(skill_file_tool)
         try:
             result = await agent.process(world.scenario.model_input())
             bundle = agent.evaluation_bundle

tests/test_agent_evaluation_runtime.py

@@ -127,9 +127,9 @@ async def test_complete_production_agent_scenarios(scenario_id, invocation_store
     result = evaluate(world, json.loads(capture["final_text"]))
     assert result.passed, result.violations
     assert capture["tool_catalog_scope"] == "controlled_moviepilot_api_and_production_internal_tools"
-    assert {"moviepilot_api", "read_skill", "update_plan", "get_tool_execution", "task", "subagent_task", "search_tools"} <= set(capture["tool_names"])
+    assert {"moviepilot_api", "read_file", "read_skill", "update_plan", "get_tool_execution", "task", "subagent_task", "search_tools"} <= set(capture["tool_names"])
     assert not {"execute_command", "write_file", "edit_file", "browser", "send_message"} & set(capture["tool_names"])
-    assert capture["child_tool_names"] == ["moviepilot_api"]
+    assert capture["child_tool_names"] == ["moviepilot_api", "read_file"]
     assert capture["task_plan"]["objective"] == "完成用户任务并核验"
     assert any("InvocationMiddleware" in node for node in capture["graph_nodes"])
     assert any("SkillsMiddleware" in node for node in capture["graph_nodes"])

tests/test_agent_invocation_middleware.py

@@ -423,6 +423,7 @@ async def test_api_forbidden_query_keys_do_not_execute_or_claim(invocation_runti
     ])
     result = await _invoke(graph)
     assert _tool_messages(result)[0].status == "error"
+    assert "输入合同校验" in _tool_messages(result)[0].content
     assert records() == []
     run.assert_not_awaited()