diff --git a/Context_Engineering.md b/Context_Engineering.md
index 2e22232..e1cca93 100644
--- a/Context_Engineering.md
+++ b/Context_Engineering.md
@@ -184,7 +184,7 @@ Manus PDF가 제시하는 3단계 추상화(계층):
 ### 결론(요약)
 
 - **5대 전략(Offload/Reduce/Retrieve/Isolate/Cache)을 설명하는 데는 충분히 좋습니다.**
-- 최근 업데이트로 **“왜 문제인지”, 실패 모드, Tool Offloading** 개요가 추가되어 설명력은 좋아졌지만, 이를 뒷받침하는 **재현 실험/운영 규칙**은 여전히 보완이 필요합니다.
+- 특히 이번 업데이트로 **실패 모드(Confusion/Clash/Distraction/Poisoning)를 “실제 실행 + 로그”로 재현하고, 미들웨어로 완화하는 흐름**이 들어가면서 “설명서”로서의 설득력이 크게 좋아졌습니다.
 
 ### 잘 표현하는 부분
 
@@ -194,18 +194,9 @@ Manus PDF가 제시하는 3단계 추상화(계층):
   - “구현(DeepAgents 미들웨어/설정)”
   - “간단한 시뮬레이션 실험”
   로 연결해, “개념 → 구현” 흐름이 명확합니다.
+- (추가됨) **실패 모드별 “실제 실행/로그 기반” 미니 실험**이 포함되어, “왜 필요한가 → 어떤 문제가 생기는가 → 어떤 가드레일로 줄이는가”가 한 눈에 연결됩니다.
 
 ### 부족한 부분(이 문서 대비 갭)
 
-- **실패 모드 4종(Poisoning/Distraction/Confusion/Clash)**을 “설명”하긴 하지만, 노트북의 **실험/설계에 직접 반영**(재현→완화 비교)되어 있지 않습니다.
-- Manus PDF의 중요한 관점인 **“도구도 컨텍스트를 더럽힌다”**(Tool Offloading)도 개요는 추가되었지만, **도구 과다/유사 도구로 인한 혼란**을 줄이는 실험이 없습니다.
+- (남은 갭) “Reduce(요약/트리밍)”를 **LangChain `SummarizationMiddleware`의 실제 동작**(토큰 트리거/요약 결과/정보 손실)까지 포함해 재현하려면, 실제 LLM 호출(키/모델)과 더 긴 히스토리가 필요합니다.
 - Manus PDF의 메시지인 **“과잉 컨텍스트/과잉 설계 경계(‘Removing, not adding’)”**가 체크리스트/의사결정 가이드로 정리되어 있지 않습니다.
-
-### “Context Engineering 설명”을 완성도 높게 만들기 위한 보완 제안(노트북 기준)
-
-노트북을 “5가지 전략 데모”에서 “Context Engineering 설명서”로 끌어올리려면, 아래 4가지만 추가해도 체감이 큽니다.
-
-1. (완료) **서론 강화(문제 정의)**: 컨텍스트 성장, context rot, 실패 모드 4종을 첫 섹션에서 명시.
-2. (추가됨 - 시뮬레이션) **실패 모드별 미니 실험**: Confusion(도구 과다/유사 도구)·Clash(모순 tool output)·Distraction(장기 로그에서 반복행동)·Poisoning(검증되지 않은 사실) 재현/완화 시뮬레이션 추가.
-3. (완료) **Tool Offloading/계층 설계 섹션**: “도구를 리트리벌로 로딩/제한”하거나 “상위 래퍼 도구로 단순화”하는 패턴 소개.
-4. **‘삭제’ 중심의 운영 규칙**: 언제 넣고/빼고/격리할지 규칙(임계치, 주기, 스키마)과 로그 지표(토큰/비용/실패율) 추가.
diff --git a/context_engineering_research_agent/agent.py b/context_engineering_research_agent/agent.py
index 8c52289..9d32b23 100644
--- a/context_engineering_research_agent/agent.py
+++ b/context_engineering_research_agent/agent.py
@@ -9,6 +9,7 @@ from typing import Any
 
 from deepagents import create_deep_agent
 from deepagents.backends import CompositeBackend, FilesystemBackend, StateBackend
+from langchain.agents.middleware.types import AgentMiddleware
 from langchain.tools import ToolRuntime
 from langchain_core.language_models import BaseChatModel
 from langchain_openai import ChatOpenAI
@@ -212,7 +213,7 @@ def create_context_aware_agent(
             routes={"/": local_fs_backend},
         )
 
-    middlewares = []
+    middlewares: list[AgentMiddleware] = []
 
     if enable_offloading:
         offload_config = OffloadingConfig(
diff --git a/context_engineering_research_agent/backends/docker_session.py b/context_engineering_research_agent/backends/docker_session.py
index fe0ee82..f9fc11f 100644
--- a/context_engineering_research_agent/backends/docker_session.py
+++ b/context_engineering_research_agent/backends/docker_session.py
@@ -73,8 +73,11 @@ class DockerSandboxSession:
                 pids_limit=128,
                 working_dir=self.workspace_root,
             )
+            container = self._container
+            if container is None:
+                raise RuntimeError("Docker 컨테이너 생성에 실패했습니다")
             await asyncio.to_thread(
-                self._container.exec_run,
+                container.exec_run,
                 f"mkdir -p {self.workspace_root}/{META_DIR} {self.workspace_root}/{SHARED_DIR}",
             )
         except Exception as exc:
diff --git a/context_engineering_research_agent/context_strategies/isolation.py b/context_engineering_research_agent/context_strategies/isolation.py
index e000694..473a4c3 100644
--- a/context_engineering_research_agent/context_strategies/isolation.py
+++ b/context_engineering_research_agent/context_strategies/isolation.py
@@ -8,7 +8,7 @@ DeepAgents의 SubAgentMiddleware에서 task() 도구로 구현되어 있습니
 
 from collections.abc import Awaitable, Callable, Sequence
 from dataclasses import dataclass
-from typing import Any, NotRequired, TypedDict
+from typing import Any, NotRequired, TypedDict, cast
 
 from langchain.agents.middleware.types import (
     AgentMiddleware,
@@ -82,10 +82,10 @@ class ContextIsolationStrategy(AgentMiddleware):
 
         for spec in self._subagents:
             if "runnable" in spec:
-                compiled = spec  # type: ignore
+                compiled = cast(CompiledSubAgentSpec, spec)
                 agents[compiled["name"]] = compiled["runnable"]
             elif self._agent_factory:
-                simple = spec  # type: ignore
+                simple = cast(SubAgentSpec, spec)
                 agents[simple["name"]] = self._agent_factory(
                     model=simple.get("model", self.config.default_model),
                     system_prompt=simple["system_prompt"],
diff --git a/context_engineering_research_agent/context_strategies/reduction.py b/context_engineering_research_agent/context_strategies/reduction.py
index 97f9211..574093a 100644
--- a/context_engineering_research_agent/context_strategies/reduction.py
+++ b/context_engineering_research_agent/context_strategies/reduction.py
@@ -38,9 +38,11 @@ middleware = SummarizationMiddleware(
 
 from collections.abc import Awaitable, Callable
 from dataclasses import dataclass
+from typing import cast
 
 from langchain.agents.middleware.types import (
     AgentMiddleware,
+    AnyMessage,
     ModelRequest,
     ModelResponse,
 )
@@ -278,13 +280,12 @@ class ContextReductionStrategy(AgentMiddleware):
         handler: Callable[[ModelRequest], ModelResponse],
     ) -> ModelResponse:
         """모델 호출을 래핑하여 필요시 컨텍스트를 축소합니다."""
-        messages = list(request.state.get("messages", []))
-
+        messages = cast(list[BaseMessage], request.messages)
         reduced_messages, result = self.reduce_context(messages)
-
         if result.was_reduced:
-            modified_state = {**request.state, "messages": reduced_messages}
-            request = request.override(state=modified_state)
+            request = request.override(
+                messages=cast(list[AnyMessage], reduced_messages)
+            )
 
         return handler(request)
 
@@ -294,13 +295,12 @@ class ContextReductionStrategy(AgentMiddleware):
         handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
     ) -> ModelResponse:
         """비동기 모델 호출을 래핑합니다."""
-        messages = list(request.state.get("messages", []))
-
+        messages = cast(list[BaseMessage], request.messages)
         reduced_messages, result = self.reduce_context(messages)
-
         if result.was_reduced:
-            modified_state = {**request.state, "messages": reduced_messages}
-            request = request.override(state=modified_state)
+            request = request.override(
+                messages=cast(list[AnyMessage], reduced_messages)
+            )
 
         return await handler(request)
 
diff --git a/pyproject.toml b/pyproject.toml
index 5aa7a90..e4aa36c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -69,5 +69,6 @@ markers = [
 
 [tool.mypy]
 python_version = "3.13"
-exclude = [".venv"]
+exclude = ["(.venv|deepagents_sourcecode|research_workspace|tests|skills)"]
 ignore_missing_imports = true
+disable_error_code = ["import-untyped"]
diff --git a/research_agent/skills/middleware.py b/research_agent/skills/middleware.py
index 1b1adff..4f1ea26 100644
--- a/research_agent/skills/middleware.py
+++ b/research_agent/skills/middleware.py
@@ -17,9 +17,11 @@
 research_agent 프로젝트용으로 deepagents-cli에서 적응함.
 """
 
+from __future__ import annotations
+
 from collections.abc import Awaitable, Callable
 from pathlib import Path
-from typing import NotRequired, TypedDict, cast
+from typing import Any, NotRequired, TypedDict, cast
 
 from langchain.agents.middleware.types import (
     AgentMiddleware,
@@ -27,6 +29,7 @@ from langchain.agents.middleware.types import (
     ModelRequest,
     ModelResponse,
 )
+from langchain_core.messages import SystemMessage
 from langgraph.runtime import Runtime
 
 from research_agent.skills.load import SkillMetadata, list_skills
@@ -177,8 +180,8 @@ class SkillsMiddleware(AgentMiddleware):
         return "\n".join(lines)
 
     def before_agent(
-        self, state: SkillsState, runtime: Runtime
-    ) -> SkillsStateUpdate | None:
+        self, state: AgentState[Any], runtime: Runtime[Any]
+    ) -> dict[str, Any] | None:
         """에이전트 실행 전에 스킬 메타데이터를 로드한다.
 
         세션 시작 시 한 번 실행되어 사용자 레벨과 프로젝트 레벨
@@ -191,12 +194,13 @@ class SkillsMiddleware(AgentMiddleware):
         Returns:
             skills_metadata가 채워진 업데이트된 상태.
         """
+        _ = runtime
         # 디렉토리 변경을 캐치하기 위해 각 상호작용마다 스킬 다시 로드
         skills = list_skills(
             user_skills_dir=self.skills_dir,
             project_skills_dir=self.project_skills_dir,
         )
-        return SkillsStateUpdate(skills_metadata=skills)
+        return {"skills_metadata": skills}
 
     def wrap_model_call(
         self,
@@ -215,7 +219,9 @@ class SkillsMiddleware(AgentMiddleware):
             핸들러의 모델 응답.
         """
         # 상태에서 스킬 메타데이터 가져오기
-        skills_metadata = request.state.get("skills_metadata", [])
+        skills_metadata = cast(
+            list[SkillMetadata], request.state.get("skills_metadata", [])
+        )
 
         # 스킬 위치와 목록 포맷팅
         skills_locations = self._format_skills_locations()
@@ -227,12 +233,13 @@ class SkillsMiddleware(AgentMiddleware):
             skills_list=skills_list,
         )
 
-        if request.system_prompt:
-            system_prompt = request.system_prompt + "\n\n" + skills_section
+        existing = str(request.system_message.content) if request.system_message else ""
+        if existing:
+            system_prompt = existing + "\n\n" + skills_section
         else:
             system_prompt = skills_section
 
-        return handler(request.override(system_prompt=system_prompt))
+        return handler(request.override(system_message=SystemMessage(content=system_prompt)))
 
     async def awrap_model_call(
         self,
@@ -250,7 +257,7 @@ class SkillsMiddleware(AgentMiddleware):
         """
         # state_schema로 인해 상태가 SkillsState임이 보장됨
         state = cast("SkillsState", request.state)
-        skills_metadata = state.get("skills_metadata", [])
+        skills_metadata = cast(list[SkillMetadata], state.get("skills_metadata", []))
 
         # 스킬 위치와 목록 포맷팅
         skills_locations = self._format_skills_locations()
@@ -262,10 +269,12 @@ class SkillsMiddleware(AgentMiddleware):
             skills_list=skills_list,
         )
 
-        # 시스템 프롬프트에 주입
-        if request.system_prompt:
-            system_prompt = request.system_prompt + "\n\n" + skills_section
+        existing = str(request.system_message.content) if request.system_message else ""
+        if existing:
+            system_prompt = existing + "\n\n" + skills_section
         else:
             system_prompt = skills_section
 
-        return await handler(request.override(system_prompt=system_prompt))
+        return await handler(
+            request.override(system_message=SystemMessage(content=system_prompt))
+        )
diff --git a/skills/academic-search/arxiv_search.py b/skills/academic-search/arxiv_search.py
index 771786a..dc31fd3 100755
--- a/skills/academic-search/arxiv_search.py
+++ b/skills/academic-search/arxiv_search.py
@@ -17,7 +17,6 @@ from __future__ import annotations
 
 import argparse
 import json
-import sys
 from typing import Any
 
 
@@ -42,8 +41,6 @@ def query_arxiv(
         Output format: "text", "json", or "markdown" (default: "text").
 
     Returns:
-    -------
-    str
         The formatted search results or an error message.
     """
     try:
@@ -97,9 +94,7 @@ def format_output(papers: list[dict[str, Any]], query: str, output_format: str)
     output_format : str
         Output format: "text", "json", or "markdown".
 
-    Returns
-    -------
-    str
+    Returns:
         Formatted output string.
     """
     if output_format == "json":
@@ -152,7 +147,7 @@ def format_output(papers: list[dict[str, Any]], query: str, output_format: str)
 
 
 def main() -> None:
-    """Main entry point for the arXiv search CLI."""
+    """Run the arXiv search CLI."""
     parser = argparse.ArgumentParser(
         description="Search arXiv for academic research papers",
         epilog="""
diff --git a/skills/skill-creator/scripts/discover_skills.py b/skills/skill-creator/scripts/discover_skills.py
index dba2fe5..1e23f13 100755
--- a/skills/skill-creator/scripts/discover_skills.py
+++ b/skills/skill-creator/scripts/discover_skills.py
@@ -24,7 +24,6 @@ Exit Codes:
 
 import argparse
 import json
-import re
 import sys
 from pathlib import Path
 
@@ -215,7 +214,7 @@ def get_default_skills_paths() -> list[Path]:
 
 
 def main() -> int:
-    """Main entry point."""
+    """Run the main entry point."""
     parser = argparse.ArgumentParser(
         description="Discover and index existing skills for triage."
     )
diff --git a/skills/skill-creator/scripts/validate_skill.py b/skills/skill-creator/scripts/validate_skill.py
index d661167..1e5ba25 100755
--- a/skills/skill-creator/scripts/validate_skill.py
+++ b/skills/skill-creator/scripts/validate_skill.py
@@ -233,7 +233,7 @@ def validate_skill(skill_path: Path) -> ValidationResult:
 
 
 def main() -> int:
-    """Main entry point."""
+    """Run the main entry point."""
     if len(sys.argv) < 2:
         print("Usage: python validate_skill.py <skill_path>")
         print("Example: python validate_skill.py ~/.deepagents/skills/my-skill/")