diff --git a/packages/ai-server/src/bootstrap.py b/packages/ai-server/src/bootstrap.py
index 04dc791e..7be56e0f 100644
--- a/packages/ai-server/src/bootstrap.py
+++ b/packages/ai-server/src/bootstrap.py
@@ -96,12 +96,9 @@ async def health_check():
             }
 
     # #416 — 모든 비즈니스 routes 가 gRPC 로 옮겨짐. router include 없음.
-    # 예외: editorial article chat (#429 Stage 3) — multi-turn / admin UI 호출 패턴.
-    # #429: chat sessions / messages / article load 모두 assets staging.
-    from src.editorial_article_chat.api import create_router as create_chat_router
-
-    assets_db = application.infrastructure().assets_database_manager()
-    app.include_router(create_chat_router(assets_db))
+    # editorial article chat (#446 fixup): 더 이상 HTTP router 없음 — 한 턴 LLM
+    # 실행은 inbound.Queue/RunChatTurn gRPC 로 통일. session/message CRUD 는
+    # api-server 가 소유 (assets DB 직접 접근).
 
     return app
 
diff --git a/packages/ai-server/src/editorial_article_chat/agent.py b/packages/ai-server/src/editorial_article_chat/agent.py
index 449cb9e5..f976d0b4 100644
--- a/packages/ai-server/src/editorial_article_chat/agent.py
+++ b/packages/ai-server/src/editorial_article_chat/agent.py
@@ -1,30 +1,33 @@
-"""Chat agent — Gemini function calling loop.
+"""Chat agent — pure compute (Gemini function calling loop).
+
+#446 fixup: DB 는 만지지 않는다. caller (api-server via gRPC) 가
+  - 현재 layout 과 history 를 input 으로 넘기고
+  - 결과로 events (assistant/tool 메시지 시퀀스) + final_layout 을 받아서
+  - 직접 INSERT / UPDATE 한다.
 
 흐름:
-  1. session 의 기존 messages 로드 → Gemini contents 형태로 변환
-  2. 새 user message append + DB 저장
-  3. Gemini 호출 (tools 바인딩)
-  4. function_call 있으면 → ToolExecutor 로 실행 → tool result 를 Gemini 에 다시
-     보내고 step 반복 (max_steps 까지)
-  5. function_call 없으면 → 최종 assistant text 반환
-  6. 마지막 layout 변경 있으면 DB persist (트랜잭션)
+  1. caller 가 보낸 history 를 Gemini contents 로 변환
+  2. user_message append + 첫 Gemini 호출
+  3. function_call 있으면 → ToolExecutor 로 실행 → tool result 를 Gemini 에 다시
+     보내고 step 반복 (max_steps 까지). 매 step 의 assistant 메시지와 tool
+     메시지를 events 에 누적.
+  4. function_call 없으면 → 최종 assistant text 반환
+  5. 최종 layout 이 입력과 다르면 layout_changed=True
 """
 
 from __future__ import annotations
 
 import logging
-from dataclasses import dataclass
-from typing import Optional
+from dataclasses import dataclass, field
+from typing import Any, Optional
 
 from google import genai
 from google.genai import types as genai_types
 
 from src.editorial_article.models import MagazineLayout
-from src.managers.database import DatabaseManager
 from src.post_editorial.config import get_settings
 from src.post_editorial.gemini_retry import call_gemini_with_fallback
 
-from .repository import ChatMessage, EditorialArticleChatRepository
 from .tools import TOOL_DECLARATIONS, ToolExecutor
 
 
@@ -35,10 +38,16 @@
 
 
 @dataclass
-class TurnResult:
-    assistant_text: str
-    tool_calls_made: int
-    layout_changed: bool
+class RunTurnResult:
+    # caller 가 순서대로 INSERT 할 메시지들. 각 dict 형식:
+    #   assistant: {"role":"assistant", "content":str|None, "tool_calls":[{name,args,id}]|None}
+    #   tool:      {"role":"tool", "tool_call_id":str, "tool_name":str, "tool_result":dict}
+    events: list[dict[str, Any]] = field(default_factory=list)
+    final_text: str = ""
+    final_layout: Optional[MagazineLayout] = None
+    layout_changed: bool = False
+    tool_calls_made: int = 0
+    error_message: str = ""
 
 
 def _system_prompt(article_title: str, sections_brief: str) -> str:
@@ -67,22 +76,29 @@ def _system_prompt(article_title: str, sections_brief: str) -> str:
 
 
 def _build_contents(
-    history: list[ChatMessage], new_user_text: str
+    history: list[dict[str, Any]], new_user_text: str
 ) -> list[genai_types.Content]:
-    """DB 메시지 → Gemini Content[] 변환."""
+    """caller 가 보낸 history (DB 메시지 dict list) → Gemini Content[] 변환.
+
+    각 history dict 의 형식 (caller 측 SeaORM Model 직렬화 결과):
+      {role: "user"|"assistant"|"tool", content?, tool_calls?, tool_call_id?,
+       tool_name?, tool_result?}
+    """
     contents: list[genai_types.Content] = []
     for msg in history:
-        if msg.role == "user" and msg.content:
+        role = msg.get("role")
+        if role == "user" and msg.get("content"):
             contents.append(
                 genai_types.Content(
-                    role="user", parts=[genai_types.Part.from_text(text=msg.content)]
+                    role="user",
+                    parts=[genai_types.Part.from_text(text=msg["content"])],
                 )
             )
-        elif msg.role == "assistant":
+        elif role == "assistant":
             parts: list[genai_types.Part] = []
-            if msg.content:
-                parts.append(genai_types.Part.from_text(text=msg.content))
-            for fc in msg.tool_calls or []:
+            if msg.get("content"):
+                parts.append(genai_types.Part.from_text(text=msg["content"]))
+            for fc in msg.get("tool_calls") or []:
                 parts.append(
                     genai_types.Part.from_function_call(
                         name=fc["name"], args=fc.get("args", {})
@@ -90,14 +106,14 @@ def _build_contents(
                 )
             if parts:
                 contents.append(genai_types.Content(role="model", parts=parts))
-        elif msg.role == "tool":
+        elif role == "tool":
             contents.append(
                 genai_types.Content(
                     role="user",
                     parts=[
                         genai_types.Part.from_function_response(
-                            name=msg.tool_name or "unknown",
-                            response=msg.tool_result or {},
+                            name=msg.get("tool_name") or "unknown",
+                            response=msg.get("tool_result") or {},
                         )
                     ],
                 )
@@ -147,37 +163,31 @@ def _extract_text(response: genai_types.GenerateContentResponse) -> str:
 
 async def run_turn(
     *,
-    db: DatabaseManager,
-    repo: EditorialArticleChatRepository,
-    article_id: str,
-    session_id: str,
+    article_title: str,
+    layout: MagazineLayout,
+    history: list[dict[str, Any]],
     user_text: str,
-) -> TurnResult:
-    """한 턴 실행 — user_text 받아서 모든 tool 호출 후 final text 반환."""
-    layout, meta = await repo.load_article(article_id)
-    if layout is None:
-        raise ValueError(f"article {article_id} has no layout_json")
-
-    history = await repo.list_messages(session_id)
+) -> RunTurnResult:
+    """한 턴 실행 — pure compute. caller 가 결과를 받아 DB 에 persist."""
     contents = _build_contents(history, user_text)
 
     settings = get_settings()
     client = genai.Client(api_key=settings.gemini_api_key)
-    executor = ToolExecutor(db=db, article_id=article_id, layout=layout)
+    executor = ToolExecutor(layout=layout)
+    initial_layout = layout
 
     config = genai_types.GenerateContentConfig(
         temperature=0.4,
         system_instruction=_system_prompt(
-            meta.get("title") or "Untitled", _summarize_sections(layout)
+            article_title or "Untitled", _summarize_sections(layout)
         ),
         tools=[genai_types.Tool(function_declarations=TOOL_DECLARATIONS)],
     )
 
-    # Persist user message immediately so client can poll
-    await repo.insert_user_message(session_id, user_text)
-
+    events: list[dict[str, Any]] = []
     tool_calls_made = 0
     final_text = ""
+
     for step in range(_MAX_STEPS):
         async def _generate(model: str) -> genai_types.GenerateContentResponse:
             return await client.aio.models.generate_content(
@@ -190,28 +200,34 @@ async def _generate(model: str) -> genai_types.GenerateContentResponse:
             )
         except Exception as exc:
             logger.exception("chat agent: Gemini call failed")
-            await repo.insert_assistant_message(
-                session_id, content=f"(에이전트 오류: {exc})"
+            err_text = f"(에이전트 오류: {exc})"
+            events.append({"role": "assistant", "content": err_text})
+            return RunTurnResult(
+                events=events,
+                final_text=err_text,
+                final_layout=executor.layout,
+                layout_changed=False,
+                tool_calls_made=tool_calls_made,
+                error_message=str(exc),
             )
-            await repo.touch_session(session_id)
-            return TurnResult(f"(에이전트 오류: {exc})", tool_calls_made, False)
 
         function_calls = _extract_function_calls(response)
         text_part = _extract_text(response)
 
-        # assistant message 저장 (tool_calls 포함)
+        # assistant 이벤트 누적
         tool_call_payloads = [
             {"name": name, "args": args, "id": cid or f"tc_{step}_{i}"}
             for i, (name, args, cid) in enumerate(function_calls)
         ]
-        await repo.insert_assistant_message(
-            session_id,
-            content=text_part if text_part else None,
-            tool_calls=tool_call_payloads if tool_call_payloads else None,
+        events.append(
+            {
+                "role": "assistant",
+                "content": text_part if text_part else None,
+                "tool_calls": tool_call_payloads if tool_call_payloads else None,
+            }
         )
 
         if not function_calls:
-            # 종료 — 최종 텍스트 반환
             final_text = text_part or ""
             break
 
@@ -225,17 +241,19 @@ async def _generate(model: str) -> genai_types.GenerateContentResponse:
             )
         contents.append(genai_types.Content(role="model", parts=model_parts))
 
-        # 각 tool 실행 + tool message 저장 + Gemini 에 결과 전달
+        # 각 tool 실행 + tool 이벤트 누적 + Gemini 에 결과 전달
         tool_response_parts: list[genai_types.Part] = []
         for i, (name, args, _cid) in enumerate(function_calls):
             tool_calls_made += 1
             result = await executor.execute(name, args)
             tc_id = tool_call_payloads[i]["id"]
-            await repo.insert_tool_message(
-                session_id=session_id,
-                tool_call_id=tc_id,
-                tool_name=name,
-                tool_result=result.to_dict(),
+            events.append(
+                {
+                    "role": "tool",
+                    "tool_call_id": tc_id,
+                    "tool_name": name,
+                    "tool_result": result.to_dict(),
+                }
             )
             tool_response_parts.append(
                 genai_types.Part.from_function_response(
@@ -251,21 +269,14 @@ async def _generate(model: str) -> genai_types.GenerateContentResponse:
             "(편집 사이클 한도 도달 — 안전상 중단. "
             "더 진행하려면 다시 메시지 보내주세요.)"
         )
-        await repo.insert_assistant_message(session_id, content=final_text)
-
-    # layout 변경 있으면 DB persist
-    layout_changed = executor.layout != layout
-    if layout_changed:
-        try:
-            await executor.persist_layout()
-        except Exception:
-            logger.exception("chat agent: persist_layout failed")
-            final_text += "\n\n(주의: DB 저장 중 오류 발생.)"
+        events.append({"role": "assistant", "content": final_text})
 
-    await repo.touch_session(session_id)
+    layout_changed = executor.layout != initial_layout
 
-    return TurnResult(
-        assistant_text=final_text,
-        tool_calls_made=tool_calls_made,
+    return RunTurnResult(
+        events=events,
+        final_text=final_text,
+        final_layout=executor.layout,
         layout_changed=layout_changed,
+        tool_calls_made=tool_calls_made,
     )
diff --git a/packages/ai-server/src/editorial_article_chat/api.py b/packages/ai-server/src/editorial_article_chat/api.py
deleted file mode 100644
index 43bc67e8..00000000
--- a/packages/ai-server/src/editorial_article_chat/api.py
+++ /dev/null
@@ -1,165 +0,0 @@
-"""HTTP endpoints for editorial article chat (FastAPI router).
-
-ai-server 의 다른 routes 는 #416 이후 gRPC 로 옮겨갔지만, chat 은 multi-turn /
-stream / 단순한 admin UI 호출 패턴이라 HTTP 가 더 자연스럽다 (예외).
-
-Routes (모두 prefix `/api/v1/editorial-article-chat`):
-  GET  /sessions/{article_id}                 — 세션 리스트
-  POST /sessions/{article_id}                 — 새 세션 생성
-  GET  /sessions/{session_id}/messages        — 메시지 리스트
-  POST /sessions/{session_id}/messages        — 새 user 메시지 → agent 실행
-"""
-
-from __future__ import annotations
-
-import logging
-from typing import Optional
-
-from fastapi import APIRouter, HTTPException
-from pydantic import BaseModel, Field
-
-from .agent import run_turn
-from .repository import EditorialArticleChatRepository
-from src.managers.database import DatabaseManager
-
-
-logger = logging.getLogger(__name__)
-
-
-# ─────────────────────────────────────────────────────────────────────────────
-# Request / Response models
-# ─────────────────────────────────────────────────────────────────────────────
-
-
-class SessionResponse(BaseModel):
-    id: str
-    article_id: str
-    created_by: Optional[str] = None
-    title: Optional[str] = None
-    last_message_at: Optional[str] = None
-    created_at: str
-
-
-class CreateSessionRequest(BaseModel):
-    created_by: Optional[str] = None
-
-
-class MessageResponse(BaseModel):
-    id: int
-    session_id: str
-    role: str
-    content: Optional[str] = None
-    tool_calls: Optional[list[dict]] = None
-    tool_call_id: Optional[str] = None
-    tool_name: Optional[str] = None
-    tool_result: Optional[dict] = None
-    created_at: str
-
-
-class SendMessageRequest(BaseModel):
-    text: str = Field(min_length=1, max_length=4000)
-
-
-class SendMessageResponse(BaseModel):
-    assistant_text: str
-    tool_calls_made: int
-    layout_changed: bool
-
-
-# ─────────────────────────────────────────────────────────────────────────────
-# Router factory (DI 용 — operation_db / repo 주입)
-# ─────────────────────────────────────────────────────────────────────────────
-
-
-def create_router(db: DatabaseManager) -> APIRouter:
-    repo = EditorialArticleChatRepository(db)
-    router = APIRouter(prefix="/api/v1/editorial-article-chat", tags=["editorial-chat"])
-
-    @router.get("/sessions/{article_id}")
-    async def list_sessions(article_id: str) -> list[SessionResponse]:
-        sessions = await repo.list_sessions(article_id)
-        return [
-            SessionResponse(
-                id=s.id,
-                article_id=s.article_id,
-                created_by=s.created_by,
-                title=s.title,
-                last_message_at=s.last_message_at.isoformat()
-                if s.last_message_at
-                else None,
-                created_at=s.created_at.isoformat(),
-            )
-            for s in sessions
-        ]
-
-    @router.post("/sessions/{article_id}")
-    async def create_session(
-        article_id: str, body: CreateSessionRequest
-    ) -> SessionResponse:
-        # article 존재 확인
-        layout, _ = await repo.load_article(article_id)
-        if layout is None:
-            raise HTTPException(404, f"article {article_id} not found / no layout")
-        sess = await repo.create_session(article_id, body.created_by)
-        return SessionResponse(
-            id=sess.id,
-            article_id=sess.article_id,
-            created_by=sess.created_by,
-            title=sess.title,
-            last_message_at=None,
-            created_at=sess.created_at.isoformat(),
-        )
-
-    @router.get("/sessions/{session_id}/messages")
-    async def list_messages(session_id: str) -> list[MessageResponse]:
-        messages = await repo.list_messages(session_id)
-        return [
-            MessageResponse(
-                id=m.id,
-                session_id=m.session_id,
-                role=m.role,
-                content=m.content,
-                tool_calls=m.tool_calls,
-                tool_call_id=m.tool_call_id,
-                tool_name=m.tool_name,
-                tool_result=m.tool_result,
-                created_at=m.created_at.isoformat(),
-            )
-            for m in messages
-        ]
-
-    @router.post("/sessions/{session_id}/messages")
-    async def send_message(
-        session_id: str, body: SendMessageRequest
-    ) -> SendMessageResponse:
-        # session 의 article_id 확인
-        async with db.acquire() as conn:
-            row = await conn.fetchrow(
-                "SELECT article_id FROM public.editorial_article_chat_sessions WHERE id = $1::uuid",
-                session_id,
-            )
-        if row is None:
-            raise HTTPException(404, f"session {session_id} not found")
-        article_id = str(row["article_id"])
-
-        try:
-            result = await run_turn(
-                db=db,
-                repo=repo,
-                article_id=article_id,
-                session_id=session_id,
-                user_text=body.text,
-            )
-        except ValueError as exc:
-            raise HTTPException(400, str(exc)) from exc
-        except Exception as exc:
-            logger.exception("send_message: agent crashed")
-            raise HTTPException(500, f"agent error: {exc}") from exc
-
-        return SendMessageResponse(
-            assistant_text=result.assistant_text,
-            tool_calls_made=result.tool_calls_made,
-            layout_changed=result.layout_changed,
-        )
-
-    return router
diff --git a/packages/ai-server/src/editorial_article_chat/repository.py b/packages/ai-server/src/editorial_article_chat/repository.py
deleted file mode 100644
index 996e0315..00000000
--- a/packages/ai-server/src/editorial_article_chat/repository.py
+++ /dev/null
@@ -1,272 +0,0 @@
-"""DB access for chat sessions / messages + article layout load."""
-
-from __future__ import annotations
-
-import json
-import logging
-from dataclasses import dataclass
-from datetime import datetime
-from typing import Any, Optional
-from uuid import UUID
-
-from src.editorial_article.models import MagazineLayout
-from src.managers.database import DatabaseManager
-
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass(frozen=True)
-class ChatSession:
-    id: str
-    article_id: str
-    created_by: Optional[str]
-    title: Optional[str]
-    last_message_at: Optional[datetime]
-    created_at: datetime
-
-
-@dataclass(frozen=True)
-class ChatMessage:
-    id: int
-    session_id: str
-    role: str  # user|assistant|tool|system
-    content: Optional[str]
-    tool_calls: Optional[list[dict[str, Any]]]
-    tool_call_id: Optional[str]
-    tool_name: Optional[str]
-    tool_result: Optional[dict[str, Any]]
-    created_at: datetime
-
-
-class EditorialArticleChatRepository:
-    def __init__(self, database_manager: DatabaseManager) -> None:
-        self._db = database_manager
-
-    # ---- Article load ----------------------------------------------------
-
-    async def load_article(
-        self, article_id: str
-    ) -> tuple[Optional[MagazineLayout], dict]:
-        """현재 article 의 layout_json + 메타 (title/status) 로드."""
-        async with self._db.acquire() as conn:
-            row = await conn.fetchrow(
-                """
-                SELECT title, subtitle, hero_image_url, layout_json, status
-                  FROM public.editorial_articles
-                 WHERE id = $1::uuid
-                """,
-                article_id,
-            )
-        if row is None:
-            return None, {}
-        layout_data = row["layout_json"]
-        if isinstance(layout_data, str):
-            try:
-                layout_data = json.loads(layout_data)
-            except Exception:
-                layout_data = None
-        layout = (
-            MagazineLayout.model_validate(layout_data)
-            if isinstance(layout_data, dict)
-            else None
-        )
-        meta = {
-            "title": row["title"],
-            "subtitle": row["subtitle"],
-            "hero_image_url": row["hero_image_url"],
-            "status": row["status"],
-        }
-        return layout, meta
-
-    # ---- Sessions --------------------------------------------------------
-
-    async def create_session(
-        self,
-        article_id: str,
-        created_by: Optional[str] = None,
-    ) -> ChatSession:
-        async with self._db.acquire() as conn:
-            row = await conn.fetchrow(
-                """
-                INSERT INTO public.editorial_article_chat_sessions
-                    (article_id, created_by)
-                VALUES ($1::uuid, $2::uuid)
-                RETURNING id, article_id, created_by, title, last_message_at, created_at
-                """,
-                article_id,
-                UUID(created_by) if created_by else None,
-            )
-        return ChatSession(
-            id=str(row["id"]),
-            article_id=str(row["article_id"]),
-            created_by=str(row["created_by"]) if row["created_by"] else None,
-            title=row["title"],
-            last_message_at=row["last_message_at"],
-            created_at=row["created_at"],
-        )
-
-    async def list_sessions(
-        self, article_id: str, limit: int = 20
-    ) -> list[ChatSession]:
-        async with self._db.acquire() as conn:
-            rows = await conn.fetch(
-                """
-                SELECT id, article_id, created_by, title, last_message_at, created_at
-                  FROM public.editorial_article_chat_sessions
-                 WHERE article_id = $1::uuid
-                 ORDER BY created_at DESC
-                 LIMIT $2
-                """,
-                article_id,
-                limit,
-            )
-        return [
-            ChatSession(
-                id=str(r["id"]),
-                article_id=str(r["article_id"]),
-                created_by=str(r["created_by"]) if r["created_by"] else None,
-                title=r["title"],
-                last_message_at=r["last_message_at"],
-                created_at=r["created_at"],
-            )
-            for r in rows
-        ]
-
-    async def get_or_create_latest_session(
-        self, article_id: str, created_by: Optional[str] = None
-    ) -> ChatSession:
-        sessions = await self.list_sessions(article_id, limit=1)
-        if sessions:
-            return sessions[0]
-        return await self.create_session(article_id, created_by)
-
-    async def touch_session(self, session_id: str) -> None:
-        async with self._db.acquire() as conn:
-            await conn.execute(
-                """
-                UPDATE public.editorial_article_chat_sessions
-                   SET last_message_at = now(), updated_at = now()
-                 WHERE id = $1::uuid
-                """,
-                session_id,
-            )
-
-    # ---- Messages --------------------------------------------------------
-
-    async def list_messages(self, session_id: str) -> list[ChatMessage]:
-        async with self._db.acquire() as conn:
-            rows = await conn.fetch(
-                """
-                SELECT id, session_id, role, content, tool_calls,
-                       tool_call_id, tool_name, tool_result, created_at
-                  FROM public.editorial_article_chat_messages
-                 WHERE session_id = $1::uuid
-                 ORDER BY id ASC
-                """,
-                session_id,
-            )
-        return [_row_to_message(r) for r in rows]
-
-    async def insert_user_message(
-        self, session_id: str, content: str
-    ) -> ChatMessage:
-        return await self._insert(
-            session_id=session_id,
-            role="user",
-            content=content,
-            tool_calls=None,
-            tool_call_id=None,
-            tool_name=None,
-            tool_result=None,
-        )
-
-    async def insert_assistant_message(
-        self,
-        session_id: str,
-        content: Optional[str],
-        tool_calls: Optional[list[dict[str, Any]]] = None,
-    ) -> ChatMessage:
-        return await self._insert(
-            session_id=session_id,
-            role="assistant",
-            content=content,
-            tool_calls=tool_calls,
-            tool_call_id=None,
-            tool_name=None,
-            tool_result=None,
-        )
-
-    async def insert_tool_message(
-        self,
-        session_id: str,
-        tool_call_id: str,
-        tool_name: str,
-        tool_result: dict[str, Any],
-    ) -> ChatMessage:
-        return await self._insert(
-            session_id=session_id,
-            role="tool",
-            content=None,
-            tool_calls=None,
-            tool_call_id=tool_call_id,
-            tool_name=tool_name,
-            tool_result=tool_result,
-        )
-
-    async def _insert(
-        self,
-        *,
-        session_id: str,
-        role: str,
-        content: Optional[str],
-        tool_calls: Optional[list[dict[str, Any]]],
-        tool_call_id: Optional[str],
-        tool_name: Optional[str],
-        tool_result: Optional[dict[str, Any]],
-    ) -> ChatMessage:
-        async with self._db.acquire() as conn:
-            row = await conn.fetchrow(
-                """
-                INSERT INTO public.editorial_article_chat_messages
-                    (session_id, role, content, tool_calls,
-                     tool_call_id, tool_name, tool_result)
-                VALUES ($1::uuid, $2, $3, $4::jsonb, $5, $6, $7::jsonb)
-                RETURNING id, session_id, role, content, tool_calls,
-                          tool_call_id, tool_name, tool_result, created_at
-                """,
-                session_id,
-                role,
-                content,
-                json.dumps(tool_calls) if tool_calls is not None else None,
-                tool_call_id,
-                tool_name,
-                json.dumps(tool_result) if tool_result is not None else None,
-            )
-        return _row_to_message(row)
-
-
-def _row_to_message(row: Any) -> ChatMessage:
-    tool_calls = row["tool_calls"]
-    if isinstance(tool_calls, str):
-        try:
-            tool_calls = json.loads(tool_calls)
-        except Exception:
-            tool_calls = None
-    tool_result = row["tool_result"]
-    if isinstance(tool_result, str):
-        try:
-            tool_result = json.loads(tool_result)
-        except Exception:
-            tool_result = None
-    return ChatMessage(
-        id=row["id"],
-        session_id=str(row["session_id"]),
-        role=row["role"],
-        content=row["content"],
-        tool_calls=tool_calls,
-        tool_call_id=row["tool_call_id"],
-        tool_name=row["tool_name"],
-        tool_result=tool_result,
-        created_at=row["created_at"],
-    )
diff --git a/packages/ai-server/src/editorial_article_chat/tools.py b/packages/ai-server/src/editorial_article_chat/tools.py
index f363703a..025a39d9 100644
--- a/packages/ai-server/src/editorial_article_chat/tools.py
+++ b/packages/ai-server/src/editorial_article_chat/tools.py
@@ -6,13 +6,12 @@
      새 카피를 생성해서 반영 (e.g., regenerate_section_body)
 
 각 tool 의 declaration 은 google-genai 의 FunctionDeclaration 형식.
-실행은 ToolExecutor 가 담당 — DB write 까지 책임지므로 article_id 와 db 를
-context 로 받음.
+실행은 ToolExecutor 가 담당 — pure compute (in-memory layout mutate) 만 수행.
+DB persist 는 호출자 (api-server) 책임 (#446 fixup).
 """
 
 from __future__ import annotations
 
-import json
 import logging
 from dataclasses import dataclass
 from typing import Any, Callable, Optional
@@ -21,7 +20,6 @@
 from google.genai import types as genai_types
 
 from src.editorial_article.models import MagazineLayout, MagazineSection
-from src.managers.database import DatabaseManager
 from src.post_editorial.config import get_settings
 from src.post_editorial.gemini_retry import call_gemini_with_fallback
 
@@ -169,16 +167,13 @@ def to_dict(self) -> dict:
 
 
 class ToolExecutor:
-    """Layout 을 메모리에 들고 있다가 각 tool 호출에 대해 mutate + DB persist."""
-
-    def __init__(
-        self,
-        db: DatabaseManager,
-        article_id: str,
-        layout: MagazineLayout,
-    ) -> None:
-        self._db = db
-        self._article_id = article_id
+    """Layout 을 메모리에 들고 있다가 각 tool 호출에 대해 mutate.
+
+    Pure compute — DB 는 만지지 않는다. 최종 layout 은 caller 가 self.layout
+    프로퍼티로 가져가 persist 한다.
+    """
+
+    def __init__(self, layout: MagazineLayout) -> None:
         self._layout = layout
 
     @property
@@ -199,26 +194,6 @@ async def execute(self, name: str, args: dict) -> ToolResult:
             logger.exception("ToolExecutor: %s crashed", name)
             return ToolResult(False, f"Tool crashed: {exc}", error=str(exc))
 
-    async def persist_layout(self) -> None:
-        layout_json = self._layout.model_dump(mode="json")
-        async with self._db.acquire() as conn:
-            await conn.execute(
-                """
-                UPDATE public.editorial_articles
-                   SET title = $2,
-                       subtitle = $3,
-                       hero_image_url = $4,
-                       layout_json = $5::jsonb,
-                       updated_at = now()
-                 WHERE id = $1::uuid
-                """,
-                self._article_id,
-                self._layout.title or "Untitled",
-                self._layout.subtitle,
-                self._layout.hero_image_url,
-                json.dumps(layout_json),
-            )
-
     # ---- Handlers --------------------------------------------------------
 
     async def _update_title(self, title: str) -> ToolResult:
diff --git a/packages/ai-server/src/grpc/proto/inbound/inbound.proto b/packages/ai-server/src/grpc/proto/inbound/inbound.proto
index 6fa5700e..9bcb8908 100644
--- a/packages/ai-server/src/grpc/proto/inbound/inbound.proto
+++ b/packages/ai-server/src/grpc/proto/inbound/inbound.proto
@@ -28,6 +28,12 @@ service Queue {
   // 없애기 위해 모든 진입을 gRPC 로 통일.
   rpc TriggerSource (TriggerSourceRequest) returns (TriggerSourceResponse);
   rpc ReparseRawPost (ReparseRawPostRequest) returns (ReparseRawPostResponse);
+
+  // #446 fixup — editorial article chat 의 한 턴 LLM 실행만 ai-server 가
+  // 담당. session/message CRUD 와 layout persist 는 api-server 가 소유.
+  // 호출자가 article_id, 현재 layout_json, history_json, user_message 를 넘기면
+  // events_json (assistant/tool 메시지 시퀀스) + final_layout_json 을 반환.
+  rpc RunChatTurn (RunChatTurnRequest) returns (RunChatTurnResponse);
 }
 
 // #214 RawPostsWorker service removed — ai-server schedules itself.
@@ -230,3 +236,32 @@ message ReparseRawPostResponse {
   bool accepted = 1;
   string error_message = 2;
 }
+
+// Editorial article chat — 한 턴 LLM 실행 (#446 fixup).
+//
+// dynamic 페이로드 (layout/history/events/tool_calls/tool_result) 는 모두 JSON
+// string 으로 전달. 이유: 이 RPC 의 입출력이 Pydantic / serde 자유 형식이라
+// proto 의 typed schema 이득이 없음. control-plane RPC (boolean+id) 와 다름.
+message RunChatTurnRequest {
+  string article_id = 1;
+  // MagazineLayout JSON. ai-server 가 model_validate 로 파싱.
+  string layout_json = 2;
+  // [{role, content?, tool_calls?, tool_call_id?, tool_name?, tool_result?}, ...]
+  string history_json = 3;
+  string user_message = 4;
+}
+
+message RunChatTurnResponse {
+  // 호출자가 순서대로 INSERT 해야 하는 메시지들. 각 원소:
+  //   assistant: {role:"assistant", content?, tool_calls?:[{name,args,id}]}
+  //   tool:      {role:"tool", tool_call_id, tool_name, tool_result}
+  string events_json = 1;
+  // 변경 없으면 빈 string. 있으면 전체 layout JSON.
+  string final_layout_json = 2;
+  // 마지막 assistant 텍스트 (UI 즉시 표시용 편의).
+  string final_text = 3;
+  int32 tool_calls_made = 4;
+  bool layout_changed = 5;
+  // 비어있지 않으면 부분/전체 실패. 호출자가 사용자에게 노출.
+  string error_message = 6;
+}
diff --git a/packages/ai-server/src/grpc/proto/inbound/inbound_pb2.py b/packages/ai-server/src/grpc/proto/inbound/inbound_pb2.py
index 9043cbfe..d4b6f55c 100644
--- a/packages/ai-server/src/grpc/proto/inbound/inbound_pb2.py
+++ b/packages/ai-server/src/grpc/proto/inbound/inbound_pb2.py
@@ -24,7 +24,7 @@
 
 
 
-DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n\rinbound.proto\x12\x07inbound\"O\n\x1bProcessPostEditorialRequest\x12\x18\n\x10post_magazine_id\x18\x01 \x01(\t\x12\x16\n\x0epost_data_json\x18\x02 \x01(\t\"R\n\x1cProcessPostEditorialResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\x12\x10\n\x08\x62\x61tch_id\x18\x03 \x01(\t\"X\n\x08\x44\x61taItem\x12\x0f\n\x07item_id\x18\x01 \x01(\t\x12\x0c\n\x04type\x18\x02 \x01(\t\x12\x0b\n\x03url\x18\x03 \x01(\t\x12\x14\n\x07post_id\x18\x04 \x01(\tH\x00\x88\x01\x01\x42\n\n\x08_post_id\";\n\x17ProcessDataBatchRequest\x12 \n\x05items\x18\x01 \x03(\x0b\x32\x11.inbound.DataItem\"N\n\x18ProcessDataBatchResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\x12\x10\n\x08\x62\x61tch_id\x18\x03 \x01(\t\"#\n\x14\x45xtractOGDataRequest\x12\x0b\n\x03url\x18\x01 \x01(\t\"\xf0\x01\n\x0cLinkMetadata\x12\x12\n\x05price\x18\x01 \x01(\tH\x00\x88\x01\x01\x12\x15\n\x08\x63urrency\x18\x02 \x01(\tH\x01\x88\x01\x01\x12\x12\n\x05\x62rand\x18\x03 \x01(\tH\x02\x88\x01\x01\x12\x10\n\x08material\x18\x04 \x03(\t\x12\x13\n\x06origin\x18\x05 \x01(\tH\x03\x88\x01\x01\x12\x15\n\x08\x63\x61tegory\x18\x06 \x01(\tH\x04\x88\x01\x01\x12\x19\n\x0csub_category\x18\x07 \x01(\tH\x05\x88\x01\x01\x42\x08\n\x06_priceB\x0b\n\t_currencyB\x08\n\x06_brandB\t\n\x07_originB\x0b\n\t_categoryB\x0f\n\r_sub_category\"\x92\x01\n\x15\x45xtractOGDataResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0b\n\x03url\x18\x02 \x01(\t\x12\r\n\x05title\x18\x03 \x01(\t\x12\x13\n\x0b\x64\x65scription\x18\x04 \x01(\t\x12\r\n\x05image\x18\x05 \x01(\t\x12\x11\n\tsite_name\x18\x06 \x01(\t\x12\x15\n\rerror_message\x18\x07 \x01(\t\"i\n\x12\x41nalyzeLinkRequest\x12\x0b\n\x03url\x18\x01 \x01(\t\x12\x0f\n\x07post_id\x18\x02 \x01(\t\x12\r\n\x05title\x18\x03 \x01(\t\x12\x13\n\x0b\x64\x65scription\x18\x04 \x01(\t\x12\x11\n\tsite_name\x18\x05 \x01(\t\"I\n\x13\x41nalyzeLinkResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\x12\x10\n\x08\x62\x61tch_id\x18\x03 \x01(\t\"*\n\x06QAPair\x12\x10\n\x08question\x18\x01 \x01(\t\x12\x0e\n\x06\x61nswer\x18\x02 \x01(\t\"\x8c\x01\n\x0fProductMetadata\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x14\n\x0csub_category\x18\x02 \x01(\t\x12\r\n\x05\x62rand\x18\x03 \x01(\t\x12\r\n\x05price\x18\x04 \x01(\t\x12\x10\n\x08\x63urrency\x18\x05 \x01(\t\x12\x11\n\tmaterials\x18\x06 \x03(\t\x12\x0e\n\x06origin\x18\x07 \x01(\t\"\x87\x01\n\x0f\x41rticleMetadata\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x14\n\x0csub_category\x18\x02 \x01(\t\x12\x0e\n\x06\x61uthor\x18\x03 \x01(\t\x12\x16\n\x0epublished_date\x18\x04 \x01(\t\x12\x14\n\x0creading_time\x18\x05 \x01(\t\x12\x0e\n\x06topics\x18\x06 \x03(\t\"\x83\x01\n\rVideoMetadata\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x14\n\x0csub_category\x18\x02 \x01(\t\x12\x0f\n\x07\x63hannel\x18\x03 \x01(\t\x12\x10\n\x08\x64uration\x18\x04 \x01(\t\x12\x12\n\nview_count\x18\x05 \x01(\t\x12\x13\n\x0bupload_date\x18\x06 \x01(\t\"M\n\rOtherMetadata\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x14\n\x0csub_category\x18\x02 \x01(\t\x12\x14\n\x0c\x63ontent_type\x18\x03 \x01(\t\"\xf1\x02\n\x19\x41nalyzeLinkDirectResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\x12\x0f\n\x07summary\x18\x03 \x01(\t\x12\x10\n\x08keywords\x18\x04 \x03(\t\x12\x1c\n\x03qna\x18\x05 \x03(\x0b\x32\x0f.inbound.QAPair\x12\x34\n\x10product_metadata\x18\x06 \x01(\x0b\x32\x18.inbound.ProductMetadataH\x00\x12\x34\n\x10\x61rticle_metadata\x18\x07 \x01(\x0b\x32\x18.inbound.ArticleMetadataH\x00\x12\x30\n\x0evideo_metadata\x18\x08 \x01(\x0b\x32\x16.inbound.VideoMetadataH\x00\x12\x30\n\x0eother_metadata\x18\t \x01(\x0b\x32\x16.inbound.OtherMetadataH\x00\x12\x15\n\rerror_message\x18\n \x01(\tB\n\n\x08metadata\"i\n\x13\x41nalyzeImageRequest\x12\x12\n\nimage_data\x18\x01 \x01(\t\x12\x0f\n\x07item_id\x18\x02 \x01(\t\x12-\n\x0e\x63\x61tegory_rules\x18\x03 \x03(\x0b\x32\x15.inbound.CategoryRule\"8\n\x0c\x43\x61tegoryRule\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x16\n\x0esub_categories\x18\x02 \x03(\t\"o\n\x13ItemWithCoordinates\x12\x14\n\x0csub_category\x18\x01 \x01(\t\x12\x0c\n\x04type\x18\x02 \x01(\t\x12\x10\n\x03top\x18\x03 \x01(\x05H\x00\x88\x01\x01\x12\x11\n\x04left\x18\x04 \x01(\x05H\x01\x88\x01\x01\x42\x06\n\x04_topB\x07\n\x05_left\"7\n\x08ItemList\x12+\n\x05items\x18\x01 \x03(\x0b\x32\x1c.inbound.ItemWithCoordinates\"\xcc\x02\n\x14\x41nalyzeImageResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07subject\x18\x02 \x01(\t\x12\r\n\x05title\x18\x03 \x01(\t\x12\x18\n\x0b\x61rtist_name\x18\x04 \x01(\tH\x00\x88\x01\x01\x12\x17\n\ngroup_name\x18\x05 \x01(\tH\x01\x88\x01\x01\x12\x14\n\x07\x63ontext\x18\x06 \x01(\tH\x02\x88\x01\x01\x12\x37\n\x05items\x18\x07 \x03(\x0b\x32(.inbound.AnalyzeImageResponse.ItemsEntry\x12\x15\n\rerror_message\x18\x08 \x01(\t\x1a?\n\nItemsEntry\x12\x0b\n\x03key\x18\x01 \x01(\t\x12 \n\x05value\x18\x02 \x01(\x0b\x32\x11.inbound.ItemList:\x02\x38\x01\x42\x0e\n\x0c_artist_nameB\r\n\x0b_group_nameB\n\n\x08_context\"?\n\x19\x45xtractPostContextRequest\x12\x0f\n\x07post_id\x18\x01 \x01(\t\x12\x11\n\timage_url\x18\x02 \x01(\t\"\x88\x01\n\x1a\x45xtractPostContextResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07\x63ontext\x18\x02 \x01(\t\x12\x12\n\nstyle_tags\x18\x03 \x03(\t\x12\x0c\n\x04mood\x18\x04 \x01(\t\x12\x0f\n\x07setting\x18\x05 \x01(\t\x12\x15\n\rerror_message\x18\x06 \x01(\t\")\n\x14TriggerSourceRequest\x12\x11\n\tsource_id\x18\x01 \x01(\t\"@\n\x15TriggerSourceResponse\x12\x10\n\x08\x61\x63\x63\x65pted\x18\x01 \x01(\x08\x12\x15\n\rerror_message\x18\x02 \x01(\t\"p\n\x15ReparseRawPostRequest\x12\x13\n\x0braw_post_id\x18\x01 \x01(\t\x12%\n\x18hero_reframe_prompt_hint\x18\x02 \x01(\tH\x00\x88\x01\x01\x42\x1b\n\x19_hero_reframe_prompt_hint\"A\n\x16ReparseRawPostResponse\x12\x10\n\x08\x61\x63\x63\x65pted\x18\x01 \x01(\x08\x12\x15\n\rerror_message\x18\x02 \x01(\t2\x84\x06\n\x05Queue\x12W\n\x10ProcessDataBatch\x12 .inbound.ProcessDataBatchRequest\x1a!.inbound.ProcessDataBatchResponse\x12N\n\rExtractOGData\x12\x1d.inbound.ExtractOGDataRequest\x1a\x1e.inbound.ExtractOGDataResponse\x12H\n\x0b\x41nalyzeLink\x12\x1b.inbound.AnalyzeLinkRequest\x1a\x1c.inbound.AnalyzeLinkResponse\x12T\n\x11\x41nalyzeLinkDirect\x12\x1b.inbound.AnalyzeLinkRequest\x1a\".inbound.AnalyzeLinkDirectResponse\x12K\n\x0c\x41nalyzeImage\x12\x1c.inbound.AnalyzeImageRequest\x1a\x1d.inbound.AnalyzeImageResponse\x12\x63\n\x14ProcessPostEditorial\x12$.inbound.ProcessPostEditorialRequest\x1a%.inbound.ProcessPostEditorialResponse\x12]\n\x12\x45xtractPostContext\x12\".inbound.ExtractPostContextRequest\x1a#.inbound.ExtractPostContextResponse\x12N\n\rTriggerSource\x12\x1d.inbound.TriggerSourceRequest\x1a\x1e.inbound.TriggerSourceResponse\x12Q\n\x0eReparseRawPost\x12\x1e.inbound.ReparseRawPostRequest\x1a\x1f.inbound.ReparseRawPostResponseb\x06proto3')
+DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n\rinbound.proto\x12\x07inbound\"O\n\x1bProcessPostEditorialRequest\x12\x18\n\x10post_magazine_id\x18\x01 \x01(\t\x12\x16\n\x0epost_data_json\x18\x02 \x01(\t\"R\n\x1cProcessPostEditorialResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\x12\x10\n\x08\x62\x61tch_id\x18\x03 \x01(\t\"X\n\x08\x44\x61taItem\x12\x0f\n\x07item_id\x18\x01 \x01(\t\x12\x0c\n\x04type\x18\x02 \x01(\t\x12\x0b\n\x03url\x18\x03 \x01(\t\x12\x14\n\x07post_id\x18\x04 \x01(\tH\x00\x88\x01\x01\x42\n\n\x08_post_id\";\n\x17ProcessDataBatchRequest\x12 \n\x05items\x18\x01 \x03(\x0b\x32\x11.inbound.DataItem\"N\n\x18ProcessDataBatchResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\x12\x10\n\x08\x62\x61tch_id\x18\x03 \x01(\t\"#\n\x14\x45xtractOGDataRequest\x12\x0b\n\x03url\x18\x01 \x01(\t\"\xf0\x01\n\x0cLinkMetadata\x12\x12\n\x05price\x18\x01 \x01(\tH\x00\x88\x01\x01\x12\x15\n\x08\x63urrency\x18\x02 \x01(\tH\x01\x88\x01\x01\x12\x12\n\x05\x62rand\x18\x03 \x01(\tH\x02\x88\x01\x01\x12\x10\n\x08material\x18\x04 \x03(\t\x12\x13\n\x06origin\x18\x05 \x01(\tH\x03\x88\x01\x01\x12\x15\n\x08\x63\x61tegory\x18\x06 \x01(\tH\x04\x88\x01\x01\x12\x19\n\x0csub_category\x18\x07 \x01(\tH\x05\x88\x01\x01\x42\x08\n\x06_priceB\x0b\n\t_currencyB\x08\n\x06_brandB\t\n\x07_originB\x0b\n\t_categoryB\x0f\n\r_sub_category\"\x92\x01\n\x15\x45xtractOGDataResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0b\n\x03url\x18\x02 \x01(\t\x12\r\n\x05title\x18\x03 \x01(\t\x12\x13\n\x0b\x64\x65scription\x18\x04 \x01(\t\x12\r\n\x05image\x18\x05 \x01(\t\x12\x11\n\tsite_name\x18\x06 \x01(\t\x12\x15\n\rerror_message\x18\x07 \x01(\t\"i\n\x12\x41nalyzeLinkRequest\x12\x0b\n\x03url\x18\x01 \x01(\t\x12\x0f\n\x07post_id\x18\x02 \x01(\t\x12\r\n\x05title\x18\x03 \x01(\t\x12\x13\n\x0b\x64\x65scription\x18\x04 \x01(\t\x12\x11\n\tsite_name\x18\x05 \x01(\t\"I\n\x13\x41nalyzeLinkResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\x12\x10\n\x08\x62\x61tch_id\x18\x03 \x01(\t\"*\n\x06QAPair\x12\x10\n\x08question\x18\x01 \x01(\t\x12\x0e\n\x06\x61nswer\x18\x02 \x01(\t\"\x8c\x01\n\x0fProductMetadata\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x14\n\x0csub_category\x18\x02 \x01(\t\x12\r\n\x05\x62rand\x18\x03 \x01(\t\x12\r\n\x05price\x18\x04 \x01(\t\x12\x10\n\x08\x63urrency\x18\x05 \x01(\t\x12\x11\n\tmaterials\x18\x06 \x03(\t\x12\x0e\n\x06origin\x18\x07 \x01(\t\"\x87\x01\n\x0f\x41rticleMetadata\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x14\n\x0csub_category\x18\x02 \x01(\t\x12\x0e\n\x06\x61uthor\x18\x03 \x01(\t\x12\x16\n\x0epublished_date\x18\x04 \x01(\t\x12\x14\n\x0creading_time\x18\x05 \x01(\t\x12\x0e\n\x06topics\x18\x06 \x03(\t\"\x83\x01\n\rVideoMetadata\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x14\n\x0csub_category\x18\x02 \x01(\t\x12\x0f\n\x07\x63hannel\x18\x03 \x01(\t\x12\x10\n\x08\x64uration\x18\x04 \x01(\t\x12\x12\n\nview_count\x18\x05 \x01(\t\x12\x13\n\x0bupload_date\x18\x06 \x01(\t\"M\n\rOtherMetadata\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x14\n\x0csub_category\x18\x02 \x01(\t\x12\x14\n\x0c\x63ontent_type\x18\x03 \x01(\t\"\xf1\x02\n\x19\x41nalyzeLinkDirectResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\x12\x0f\n\x07summary\x18\x03 \x01(\t\x12\x10\n\x08keywords\x18\x04 \x03(\t\x12\x1c\n\x03qna\x18\x05 \x03(\x0b\x32\x0f.inbound.QAPair\x12\x34\n\x10product_metadata\x18\x06 \x01(\x0b\x32\x18.inbound.ProductMetadataH\x00\x12\x34\n\x10\x61rticle_metadata\x18\x07 \x01(\x0b\x32\x18.inbound.ArticleMetadataH\x00\x12\x30\n\x0evideo_metadata\x18\x08 \x01(\x0b\x32\x16.inbound.VideoMetadataH\x00\x12\x30\n\x0eother_metadata\x18\t \x01(\x0b\x32\x16.inbound.OtherMetadataH\x00\x12\x15\n\rerror_message\x18\n \x01(\tB\n\n\x08metadata\"i\n\x13\x41nalyzeImageRequest\x12\x12\n\nimage_data\x18\x01 \x01(\t\x12\x0f\n\x07item_id\x18\x02 \x01(\t\x12-\n\x0e\x63\x61tegory_rules\x18\x03 \x03(\x0b\x32\x15.inbound.CategoryRule\"8\n\x0c\x43\x61tegoryRule\x12\x10\n\x08\x63\x61tegory\x18\x01 \x01(\t\x12\x16\n\x0esub_categories\x18\x02 \x03(\t\"o\n\x13ItemWithCoordinates\x12\x14\n\x0csub_category\x18\x01 \x01(\t\x12\x0c\n\x04type\x18\x02 \x01(\t\x12\x10\n\x03top\x18\x03 \x01(\x05H\x00\x88\x01\x01\x12\x11\n\x04left\x18\x04 \x01(\x05H\x01\x88\x01\x01\x42\x06\n\x04_topB\x07\n\x05_left\"7\n\x08ItemList\x12+\n\x05items\x18\x01 \x03(\x0b\x32\x1c.inbound.ItemWithCoordinates\"\xcc\x02\n\x14\x41nalyzeImageResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07subject\x18\x02 \x01(\t\x12\r\n\x05title\x18\x03 \x01(\t\x12\x18\n\x0b\x61rtist_name\x18\x04 \x01(\tH\x00\x88\x01\x01\x12\x17\n\ngroup_name\x18\x05 \x01(\tH\x01\x88\x01\x01\x12\x14\n\x07\x63ontext\x18\x06 \x01(\tH\x02\x88\x01\x01\x12\x37\n\x05items\x18\x07 \x03(\x0b\x32(.inbound.AnalyzeImageResponse.ItemsEntry\x12\x15\n\rerror_message\x18\x08 \x01(\t\x1a?\n\nItemsEntry\x12\x0b\n\x03key\x18\x01 \x01(\t\x12 \n\x05value\x18\x02 \x01(\x0b\x32\x11.inbound.ItemList:\x02\x38\x01\x42\x0e\n\x0c_artist_nameB\r\n\x0b_group_nameB\n\n\x08_context\"?\n\x19\x45xtractPostContextRequest\x12\x0f\n\x07post_id\x18\x01 \x01(\t\x12\x11\n\timage_url\x18\x02 \x01(\t\"\x88\x01\n\x1a\x45xtractPostContextResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07\x63ontext\x18\x02 \x01(\t\x12\x12\n\nstyle_tags\x18\x03 \x03(\t\x12\x0c\n\x04mood\x18\x04 \x01(\t\x12\x0f\n\x07setting\x18\x05 \x01(\t\x12\x15\n\rerror_message\x18\x06 \x01(\t\")\n\x14TriggerSourceRequest\x12\x11\n\tsource_id\x18\x01 \x01(\t\"@\n\x15TriggerSourceResponse\x12\x10\n\x08\x61\x63\x63\x65pted\x18\x01 \x01(\x08\x12\x15\n\rerror_message\x18\x02 \x01(\t\"p\n\x15ReparseRawPostRequest\x12\x13\n\x0braw_post_id\x18\x01 \x01(\t\x12%\n\x18hero_reframe_prompt_hint\x18\x02 \x01(\tH\x00\x88\x01\x01\x42\x1b\n\x19_hero_reframe_prompt_hint\"A\n\x16ReparseRawPostResponse\x12\x10\n\x08\x61\x63\x63\x65pted\x18\x01 \x01(\x08\x12\x15\n\rerror_message\x18\x02 \x01(\t\"i\n\x12RunChatTurnRequest\x12\x12\n\narticle_id\x18\x01 \x01(\t\x12\x13\n\x0blayout_json\x18\x02 \x01(\t\x12\x14\n\x0chistory_json\x18\x03 \x01(\t\x12\x14\n\x0cuser_message\x18\x04 \x01(\t\"\xa1\x01\n\x13RunChatTurnResponse\x12\x13\n\x0b\x65vents_json\x18\x01 \x01(\t\x12\x19\n\x11\x66inal_layout_json\x18\x02 \x01(\t\x12\x12\n\nfinal_text\x18\x03 \x01(\t\x12\x17\n\x0ftool_calls_made\x18\x04 \x01(\x05\x12\x16\n\x0elayout_changed\x18\x05 \x01(\x08\x12\x15\n\rerror_message\x18\x06 \x01(\t2\xce\x06\n\x05Queue\x12W\n\x10ProcessDataBatch\x12 .inbound.ProcessDataBatchRequest\x1a!.inbound.ProcessDataBatchResponse\x12N\n\rExtractOGData\x12\x1d.inbound.ExtractOGDataRequest\x1a\x1e.inbound.ExtractOGDataResponse\x12H\n\x0b\x41nalyzeLink\x12\x1b.inbound.AnalyzeLinkRequest\x1a\x1c.inbound.AnalyzeLinkResponse\x12T\n\x11\x41nalyzeLinkDirect\x12\x1b.inbound.AnalyzeLinkRequest\x1a\".inbound.AnalyzeLinkDirectResponse\x12K\n\x0c\x41nalyzeImage\x12\x1c.inbound.AnalyzeImageRequest\x1a\x1d.inbound.AnalyzeImageResponse\x12\x63\n\x14ProcessPostEditorial\x12$.inbound.ProcessPostEditorialRequest\x1a%.inbound.ProcessPostEditorialResponse\x12]\n\x12\x45xtractPostContext\x12\".inbound.ExtractPostContextRequest\x1a#.inbound.ExtractPostContextResponse\x12N\n\rTriggerSource\x12\x1d.inbound.TriggerSourceRequest\x1a\x1e.inbound.TriggerSourceResponse\x12Q\n\x0eReparseRawPost\x12\x1e.inbound.ReparseRawPostRequest\x1a\x1f.inbound.ReparseRawPostResponse\x12H\n\x0bRunChatTurn\x12\x1b.inbound.RunChatTurnRequest\x1a\x1c.inbound.RunChatTurnResponseb\x06proto3')
 
 _globals = globals()
 _builder.BuildMessageAndEnumDescriptors(DESCRIPTOR, _globals)
@@ -89,6 +89,10 @@
   _globals['_REPARSERAWPOSTREQUEST']._serialized_end=3038
   _globals['_REPARSERAWPOSTRESPONSE']._serialized_start=3040
   _globals['_REPARSERAWPOSTRESPONSE']._serialized_end=3105
-  _globals['_QUEUE']._serialized_start=3108
-  _globals['_QUEUE']._serialized_end=3880
+  _globals['_RUNCHATTURNREQUEST']._serialized_start=3107
+  _globals['_RUNCHATTURNREQUEST']._serialized_end=3212
+  _globals['_RUNCHATTURNRESPONSE']._serialized_start=3215
+  _globals['_RUNCHATTURNRESPONSE']._serialized_end=3376
+  _globals['_QUEUE']._serialized_start=3379
+  _globals['_QUEUE']._serialized_end=4225
 # @@protoc_insertion_point(module_scope)
diff --git a/packages/ai-server/src/grpc/proto/inbound/inbound_pb2.pyi b/packages/ai-server/src/grpc/proto/inbound/inbound_pb2.pyi
index a6d771a0..07b6f696 100644
--- a/packages/ai-server/src/grpc/proto/inbound/inbound_pb2.pyi
+++ b/packages/ai-server/src/grpc/proto/inbound/inbound_pb2.pyi
@@ -326,3 +326,31 @@ class ReparseRawPostResponse(_message.Message):
     accepted: bool
     error_message: str
     def __init__(self, accepted: bool = ..., error_message: _Optional[str] = ...) -> None: ...
+
+class RunChatTurnRequest(_message.Message):
+    __slots__ = ("article_id", "layout_json", "history_json", "user_message")
+    ARTICLE_ID_FIELD_NUMBER: _ClassVar[int]
+    LAYOUT_JSON_FIELD_NUMBER: _ClassVar[int]
+    HISTORY_JSON_FIELD_NUMBER: _ClassVar[int]
+    USER_MESSAGE_FIELD_NUMBER: _ClassVar[int]
+    article_id: str
+    layout_json: str
+    history_json: str
+    user_message: str
+    def __init__(self, article_id: _Optional[str] = ..., layout_json: _Optional[str] = ..., history_json: _Optional[str] = ..., user_message: _Optional[str] = ...) -> None: ...
+
+class RunChatTurnResponse(_message.Message):
+    __slots__ = ("events_json", "final_layout_json", "final_text", "tool_calls_made", "layout_changed", "error_message")
+    EVENTS_JSON_FIELD_NUMBER: _ClassVar[int]
+    FINAL_LAYOUT_JSON_FIELD_NUMBER: _ClassVar[int]
+    FINAL_TEXT_FIELD_NUMBER: _ClassVar[int]
+    TOOL_CALLS_MADE_FIELD_NUMBER: _ClassVar[int]
+    LAYOUT_CHANGED_FIELD_NUMBER: _ClassVar[int]
+    ERROR_MESSAGE_FIELD_NUMBER: _ClassVar[int]
+    events_json: str
+    final_layout_json: str
+    final_text: str
+    tool_calls_made: int
+    layout_changed: bool
+    error_message: str
+    def __init__(self, events_json: _Optional[str] = ..., final_layout_json: _Optional[str] = ..., final_text: _Optional[str] = ..., tool_calls_made: _Optional[int] = ..., layout_changed: bool = ..., error_message: _Optional[str] = ...) -> None: ...
diff --git a/packages/ai-server/src/grpc/proto/inbound/inbound_pb2_grpc.py b/packages/ai-server/src/grpc/proto/inbound/inbound_pb2_grpc.py
index 86475dd1..fcfeb168 100644
--- a/packages/ai-server/src/grpc/proto/inbound/inbound_pb2_grpc.py
+++ b/packages/ai-server/src/grpc/proto/inbound/inbound_pb2_grpc.py
@@ -79,6 +79,11 @@ def __init__(self, channel):
                 request_serializer=inbound__pb2.ReparseRawPostRequest.SerializeToString,
                 response_deserializer=inbound__pb2.ReparseRawPostResponse.FromString,
                 _registered_method=True)
+        self.RunChatTurn = channel.unary_unary(
+                '/inbound.Queue/RunChatTurn',
+                request_serializer=inbound__pb2.RunChatTurnRequest.SerializeToString,
+                response_deserializer=inbound__pb2.RunChatTurnResponse.FromString,
+                _registered_method=True)
 
 
 class QueueServicer(object):
@@ -148,6 +153,16 @@ def ReparseRawPost(self, request, context):
         context.set_details('Method not implemented!')
         raise NotImplementedError('Method not implemented!')
 
+    def RunChatTurn(self, request, context):
+        """#446 fixup — editorial article chat 의 한 턴 LLM 실행만 ai-server 가
+        담당. session/message CRUD 와 layout persist 는 api-server 가 소유.
+        호출자가 article_id, 현재 layout_json, history_json, user_message 를 넘기면
+        events_json (assistant/tool 메시지 시퀀스) + final_layout_json 을 반환.
+        """
+        context.set_code(grpc.StatusCode.UNIMPLEMENTED)
+        context.set_details('Method not implemented!')
+        raise NotImplementedError('Method not implemented!')
+
 
 def add_QueueServicer_to_server(servicer, server):
     rpc_method_handlers = {
@@ -196,6 +211,11 @@ def add_QueueServicer_to_server(servicer, server):
                     request_deserializer=inbound__pb2.ReparseRawPostRequest.FromString,
                     response_serializer=inbound__pb2.ReparseRawPostResponse.SerializeToString,
             ),
+            'RunChatTurn': grpc.unary_unary_rpc_method_handler(
+                    servicer.RunChatTurn,
+                    request_deserializer=inbound__pb2.RunChatTurnRequest.FromString,
+                    response_serializer=inbound__pb2.RunChatTurnResponse.SerializeToString,
+            ),
     }
     generic_handler = grpc.method_handlers_generic_handler(
             'inbound.Queue', rpc_method_handlers)
@@ -449,3 +469,30 @@ def ReparseRawPost(request,
             timeout,
             metadata,
             _registered_method=True)
+
+    @staticmethod
+    def RunChatTurn(request,
+            target,
+            options=(),
+            channel_credentials=None,
+            call_credentials=None,
+            insecure=False,
+            compression=None,
+            wait_for_ready=None,
+            timeout=None,
+            metadata=None):
+        return grpc.experimental.unary_unary(
+            request,
+            target,
+            '/inbound.Queue/RunChatTurn',
+            inbound__pb2.RunChatTurnRequest.SerializeToString,
+            inbound__pb2.RunChatTurnResponse.FromString,
+            options,
+            channel_credentials,
+            insecure,
+            call_credentials,
+            compression,
+            wait_for_ready,
+            timeout,
+            metadata,
+            _registered_method=True)
diff --git a/packages/ai-server/src/grpc/servicer/metadata_servicer.py b/packages/ai-server/src/grpc/servicer/metadata_servicer.py
index 7d386468..eac7fd67 100644
--- a/packages/ai-server/src/grpc/servicer/metadata_servicer.py
+++ b/packages/ai-server/src/grpc/servicer/metadata_servicer.py
@@ -547,3 +547,69 @@ async def ExtractPostContext(
                 setting="",
                 error_message=str(e),
             )
+
+    async def RunChatTurn(
+        self,
+        request: inbound_pb2.RunChatTurnRequest,
+        context,
+    ) -> inbound_pb2.RunChatTurnResponse:
+        """Editorial article chat — 한 턴 LLM 실행 (#446 fixup).
+
+        Pure compute: caller (api-server) 가 layout/history/user_message 를 보내고
+        events + final_layout 을 받아서 직접 DB persist. ai-server 는 DB 안 만짐.
+        """
+        import json as _json
+
+        from src.editorial_article.models import MagazineLayout
+        from src.editorial_article_chat.agent import run_turn
+
+        try:
+            layout = MagazineLayout.model_validate_json(request.layout_json)
+            history = _json.loads(request.history_json) if request.history_json else []
+            if not isinstance(history, list):
+                raise ValueError("history_json must decode to a list")
+            article_title = layout.title or "Untitled"
+
+            result = await run_turn(
+                article_title=article_title,
+                layout=layout,
+                history=history,
+                user_text=request.user_message,
+            )
+
+            final_layout_json = ""
+            if result.layout_changed and result.final_layout is not None:
+                final_layout_json = result.final_layout.model_dump_json()
+
+            return inbound_pb2.RunChatTurnResponse(
+                events_json=_json.dumps(result.events, ensure_ascii=False),
+                final_layout_json=final_layout_json,
+                final_text=result.final_text,
+                tool_calls_made=result.tool_calls_made,
+                layout_changed=result.layout_changed,
+                error_message=result.error_message,
+            )
+        except ValueError as exc:
+            self.logger.warning(f"RunChatTurn invalid input: {exc}")
+            context.set_code(grpc.StatusCode.INVALID_ARGUMENT)
+            context.set_details(str(exc))
+            return inbound_pb2.RunChatTurnResponse(
+                events_json="[]",
+                final_layout_json="",
+                final_text="",
+                tool_calls_made=0,
+                layout_changed=False,
+                error_message=str(exc),
+            )
+        except Exception as exc:
+            self.logger.error(f"RunChatTurn failed: {exc}", exc_info=True)
+            context.set_code(grpc.StatusCode.INTERNAL)
+            context.set_details(str(exc))
+            return inbound_pb2.RunChatTurnResponse(
+                events_json="[]",
+                final_layout_json="",
+                final_text="",
+                tool_calls_made=0,
+                layout_changed=False,
+                error_message=str(exc),
+            )
diff --git a/packages/api-server/proto/ai.proto b/packages/api-server/proto/ai.proto
index a6688fed..50cf0355 100644
--- a/packages/api-server/proto/ai.proto
+++ b/packages/api-server/proto/ai.proto
@@ -27,6 +27,12 @@ service Queue {
   // /raw-posts/items/{id}/reparse) 를 gRPC 로 통일.
   rpc TriggerSource (TriggerSourceRequest) returns (TriggerSourceResponse);
   rpc ReparseRawPost (ReparseRawPostRequest) returns (ReparseRawPostResponse);
+
+  // #446 fixup — editorial article chat 의 한 턴 LLM 실행만 ai-server 가
+  // 담당. session/message CRUD 와 layout persist 는 api-server 가 소유.
+  // 호출자가 article_id, 현재 layout_json, history_json, user_message 를 넘기면
+  // events_json (assistant/tool 메시지 시퀀스) + final_layout_json 을 반환.
+  rpc RunChatTurn (RunChatTurnRequest) returns (RunChatTurnResponse);
 }
 
 // #214 RawPostsWorker service removed — ai-server now schedules itself.
@@ -228,3 +234,32 @@ message ReparseRawPostResponse {
   bool accepted = 1;
   string error_message = 2;
 }
+
+// Editorial article chat — 한 턴 LLM 실행 (#446 fixup).
+//
+// dynamic 페이로드 (layout/history/events/tool_calls/tool_result) 는 모두 JSON
+// string 으로 전달. 이유: 이 RPC 의 입출력이 Pydantic / serde 자유 형식이라
+// proto 의 typed schema 이득이 없음. control-plane RPC (boolean+id) 와 다름.
+message RunChatTurnRequest {
+  string article_id = 1;
+  // MagazineLayout JSON. ai-server 가 model_validate 로 파싱.
+  string layout_json = 2;
+  // [{role, content?, tool_calls?, tool_call_id?, tool_name?, tool_result?}, ...]
+  string history_json = 3;
+  string user_message = 4;
+}
+
+message RunChatTurnResponse {
+  // 호출자가 순서대로 INSERT 해야 하는 메시지들. 각 원소:
+  //   assistant: {role:"assistant", content?, tool_calls?:[{name,args,id}]}
+  //   tool:      {role:"tool", tool_call_id, tool_name, tool_result}
+  string events_json = 1;
+  // 변경 없으면 빈 string. 있으면 전체 layout JSON.
+  string final_layout_json = 2;
+  // 마지막 assistant 텍스트 (UI 즉시 표시용 편의).
+  string final_text = 3;
+  int32 tool_calls_made = 4;
+  bool layout_changed = 5;
+  // 비어있지 않으면 부분/전체 실패. 호출자가 사용자에게 노출.
+  string error_message = 6;
+}
diff --git a/packages/api-server/src/domains/admin/editorial_article_chat.rs b/packages/api-server/src/domains/admin/editorial_article_chat.rs
new file mode 100644
index 00000000..97f658e4
--- /dev/null
+++ b/packages/api-server/src/domains/admin/editorial_article_chat.rs
@@ -0,0 +1,417 @@
+//! Admin — editorial article chat (#446 fixup).
+//!
+//! Stage 3 매거진 대화형 편집. 이전엔 web 이 ai-server FastAPI 라우터를 직접
+//! 호출했지만 (#446 머지 직후 발견), 모노레포 컨벤션 위반 — ai-server 는
+//! internal compute, web 은 항상 api-server 만 호출. 이 모듈이 새 진입점.
+//!
+//! 책임 분리:
+//!   - api-server (이 모듈): chat session/message CRUD, layout persist, admin
+//!     인증 게이팅
+//!   - ai-server (gRPC `RunChatTurn`): 한 턴 LLM 실행 (Gemini + tool loop) —
+//!     stateless compute, DB 안 만짐
+//!
+//! 라우트:
+//!   GET    /sessions/{article_id}   — 세션 list
+//!   POST   /sessions/{article_id}   — 새 세션 생성
+//!   GET    /messages/{session_id}   — 메시지 list
+//!   POST   /messages/{session_id}   — user 메시지 send + agent 한 턴 실행
+
+use axum::{
+    extract::{Path, State},
+    http::StatusCode,
+    routing::get,
+    Json, Router,
+};
+use sea_orm::{
+    ConnectionTrait, DatabaseBackend, DatabaseConnection, Statement,
+};
+use serde::{Deserialize, Serialize};
+use uuid::Uuid;
+
+use crate::{
+    config::{AppConfig, AppState},
+    error::{AppError, AppResult},
+    middleware::auth::User,
+};
+
+// ─────────────────────────────────────────────────────────────────────────────
+// DTOs
+// ─────────────────────────────────────────────────────────────────────────────
+
+#[derive(Debug, Serialize)]
+pub struct SessionResponse {
+    pub id: Uuid,
+    pub article_id: Uuid,
+    pub created_by: Option<Uuid>,
+    pub title: Option<String>,
+    pub last_message_at: Option<String>,
+    pub created_at: String,
+}
+
+#[derive(Debug, Serialize)]
+pub struct MessageResponse {
+    pub id: i64,
+    pub session_id: Uuid,
+    pub role: String,
+    pub content: Option<String>,
+    pub tool_calls: Option<serde_json::Value>,
+    pub tool_call_id: Option<String>,
+    pub tool_name: Option<String>,
+    pub tool_result: Option<serde_json::Value>,
+    pub created_at: String,
+}
+
+#[derive(Debug, Deserialize)]
+pub struct SendMessageRequest {
+    pub text: String,
+}
+
+#[derive(Debug, Serialize)]
+pub struct SendMessageResponse {
+    pub assistant_text: String,
+    pub tool_calls_made: i32,
+    pub layout_changed: bool,
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Helpers
+// ─────────────────────────────────────────────────────────────────────────────
+
+fn map_session(row: &sea_orm::QueryResult) -> Result<SessionResponse, AppError> {
+    Ok(SessionResponse {
+        id: row.try_get::<Uuid>("", "id").map_err(AppError::DatabaseError)?,
+        article_id: row
+            .try_get::<Uuid>("", "article_id")
+            .map_err(AppError::DatabaseError)?,
+        created_by: row.try_get::<Option<Uuid>>("", "created_by").ok().flatten(),
+        title: row.try_get::<Option<String>>("", "title").ok().flatten(),
+        last_message_at: row
+            .try_get::<Option<chrono::DateTime<chrono::FixedOffset>>>("", "last_message_at")
+            .ok()
+            .flatten()
+            .map(|t| t.to_rfc3339()),
+        created_at: row
+            .try_get::<chrono::DateTime<chrono::FixedOffset>>("", "created_at")
+            .map_err(AppError::DatabaseError)?
+            .to_rfc3339(),
+    })
+}
+
+fn map_message(row: &sea_orm::QueryResult) -> Result<MessageResponse, AppError> {
+    Ok(MessageResponse {
+        id: row.try_get::<i64>("", "id").map_err(AppError::DatabaseError)?,
+        session_id: row
+            .try_get::<Uuid>("", "session_id")
+            .map_err(AppError::DatabaseError)?,
+        role: row
+            .try_get::<String>("", "role")
+            .map_err(AppError::DatabaseError)?,
+        content: row.try_get::<Option<String>>("", "content").ok().flatten(),
+        tool_calls: row
+            .try_get::<Option<serde_json::Value>>("", "tool_calls")
+            .ok()
+            .flatten(),
+        tool_call_id: row
+            .try_get::<Option<String>>("", "tool_call_id")
+            .ok()
+            .flatten(),
+        tool_name: row.try_get::<Option<String>>("", "tool_name").ok().flatten(),
+        tool_result: row
+            .try_get::<Option<serde_json::Value>>("", "tool_result")
+            .ok()
+            .flatten(),
+        created_at: row
+            .try_get::<chrono::DateTime<chrono::FixedOffset>>("", "created_at")
+            .map_err(AppError::DatabaseError)?
+            .to_rfc3339(),
+    })
+}
+
+async fn load_messages(
+    db: &DatabaseConnection,
+    session_id: Uuid,
+) -> AppResult<Vec<MessageResponse>> {
+    let rows = db
+        .query_all(Statement::from_sql_and_values(
+            DatabaseBackend::Postgres,
+            "SELECT id, session_id, role, content, tool_calls, tool_call_id, \
+                    tool_name, tool_result, created_at \
+               FROM public.editorial_article_chat_messages \
+              WHERE session_id = $1::uuid \
+              ORDER BY id ASC",
+            vec![session_id.into()],
+        ))
+        .await
+        .map_err(AppError::DatabaseError)?;
+    rows.iter().map(map_message).collect()
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Handlers
+// ─────────────────────────────────────────────────────────────────────────────
+
+/// GET /sessions/{article_id}
+pub async fn list_sessions(
+    State(state): State<AppState>,
+    _user: axum::Extension<User>,
+    Path(article_id): Path<Uuid>,
+) -> AppResult<Json<Vec<SessionResponse>>> {
+    let rows = state
+        .assets_db
+        .query_all(Statement::from_sql_and_values(
+            DatabaseBackend::Postgres,
+            "SELECT id, article_id, created_by, title, last_message_at, created_at \
+               FROM public.editorial_article_chat_sessions \
+              WHERE article_id = $1::uuid \
+              ORDER BY created_at DESC \
+              LIMIT 20",
+            vec![article_id.into()],
+        ))
+        .await
+        .map_err(AppError::DatabaseError)?;
+    let items: Result<Vec<_>, _> = rows.iter().map(map_session).collect();
+    Ok(Json(items?))
+}
+
+/// POST /sessions/{article_id}
+pub async fn create_session(
+    State(state): State<AppState>,
+    user: axum::Extension<User>,
+    Path(article_id): Path<Uuid>,
+) -> AppResult<(StatusCode, Json<SessionResponse>)> {
+    let row = state
+        .assets_db
+        .query_one(Statement::from_sql_and_values(
+            DatabaseBackend::Postgres,
+            "INSERT INTO public.editorial_article_chat_sessions \
+                    (article_id, created_by) \
+             VALUES ($1::uuid, $2::uuid) \
+             RETURNING id, article_id, created_by, title, last_message_at, created_at",
+            vec![article_id.into(), user.id.into()],
+        ))
+        .await
+        .map_err(AppError::DatabaseError)?
+        .ok_or_else(|| AppError::InternalError("create_session: no row returned".into()))?;
+    Ok((StatusCode::CREATED, Json(map_session(&row)?)))
+}
+
+/// GET /messages/{session_id}
+pub async fn list_messages(
+    State(state): State<AppState>,
+    _user: axum::Extension<User>,
+    Path(session_id): Path<Uuid>,
+) -> AppResult<Json<Vec<MessageResponse>>> {
+    Ok(Json(load_messages(state.assets_db.as_ref(), session_id).await?))
+}
+
+/// POST /messages/{session_id}
+///
+/// 흐름 (오케스트레이션):
+///   1. user 메시지 INSERT (즉시, gRPC 호출 전)
+///   2. article (layout, title) 로드 + history 로드 — 1번 INSERT 포함
+///   3. ai-server gRPC RunChatTurn (deadline 120s) — pure compute
+///   4. 응답의 events_json 을 순서대로 INSERT (assistant + tool 메시지)
+///   5. layout_changed 면 article UPDATE
+///   6. session.last_message_at touch
+pub async fn send_message(
+    State(state): State<AppState>,
+    _user: axum::Extension<User>,
+    Path(session_id): Path<Uuid>,
+    Json(body): Json<SendMessageRequest>,
+) -> AppResult<Json<SendMessageResponse>> {
+    let text = body.text.trim();
+    if text.is_empty() {
+        return Err(AppError::ValidationError(
+            "text 가 비어있습니다.".into(),
+        ));
+    }
+    if text.len() > 4000 {
+        return Err(AppError::ValidationError(
+            "text 가 너무 깁니다 (max 4000).".into(),
+        ));
+    }
+
+    let db = state.assets_db.as_ref();
+
+    // Resolve article_id from session
+    let session_row = db
+        .query_one(Statement::from_sql_and_values(
+            DatabaseBackend::Postgres,
+            "SELECT article_id FROM public.editorial_article_chat_sessions \
+              WHERE id = $1::uuid",
+            vec![session_id.into()],
+        ))
+        .await
+        .map_err(AppError::DatabaseError)?
+        .ok_or_else(|| AppError::NotFound(format!("session {session_id} 없음")))?;
+    let article_id: Uuid = session_row
+        .try_get::<Uuid>("", "article_id")
+        .map_err(AppError::DatabaseError)?;
+
+    // Load article layout + title
+    let article_row = db
+        .query_one(Statement::from_sql_and_values(
+            DatabaseBackend::Postgres,
+            "SELECT title, layout_json \
+               FROM public.editorial_articles \
+              WHERE id = $1::uuid",
+            vec![article_id.into()],
+        ))
+        .await
+        .map_err(AppError::DatabaseError)?
+        .ok_or_else(|| AppError::NotFound(format!("article {article_id} 없음")))?;
+    let layout_value: Option<serde_json::Value> = article_row
+        .try_get::<Option<serde_json::Value>>("", "layout_json")
+        .ok()
+        .flatten();
+    let layout_json = layout_value
+        .ok_or_else(|| AppError::ValidationError("article 의 layout_json 이 없음".into()))?
+        .to_string();
+
+    // INSERT user message (immediate)
+    db.execute(Statement::from_sql_and_values(
+        DatabaseBackend::Postgres,
+        "INSERT INTO public.editorial_article_chat_messages (session_id, role, content) \
+         VALUES ($1::uuid, 'user', $2)",
+        vec![session_id.into(), text.to_string().into()],
+    ))
+    .await
+    .map_err(AppError::DatabaseError)?;
+
+    // Load full history (includes the user msg we just inserted)
+    let history = load_messages(db, session_id).await?;
+    let history_json = serde_json::to_string(&history)
+        .map_err(|e| AppError::InternalError(format!("history serialize failed: {e}")))?;
+
+    // gRPC call to ai-server
+    let resp = state
+        .decoded_ai_client
+        .run_chat_turn(
+            article_id.to_string(),
+            layout_json,
+            history_json,
+            text.to_string(),
+        )
+        .await
+        .map_err(|e| AppError::ExternalService(format!("ai-server: {e}")))?;
+
+    // Iterate events_json and INSERT
+    let events: Vec<serde_json::Value> = serde_json::from_str(&resp.events_json)
+        .map_err(|e| AppError::InternalError(format!("events_json parse failed: {e}")))?;
+    for evt in &events {
+        let role = evt.get("role").and_then(|v| v.as_str()).unwrap_or("");
+        match role {
+            "assistant" => {
+                let content = evt.get("content").and_then(|v| v.as_str());
+                let tool_calls = evt.get("tool_calls").cloned();
+                db.execute(Statement::from_sql_and_values(
+                    DatabaseBackend::Postgres,
+                    "INSERT INTO public.editorial_article_chat_messages \
+                            (session_id, role, content, tool_calls) \
+                     VALUES ($1::uuid, 'assistant', $2, $3::jsonb)",
+                    vec![
+                        session_id.into(),
+                        content.map(|s| s.to_string()).into(),
+                        tool_calls.into(),
+                    ],
+                ))
+                .await
+                .map_err(AppError::DatabaseError)?;
+            }
+            "tool" => {
+                let tool_call_id = evt.get("tool_call_id").and_then(|v| v.as_str());
+                let tool_name = evt.get("tool_name").and_then(|v| v.as_str());
+                let tool_result = evt.get("tool_result").cloned();
+                db.execute(Statement::from_sql_and_values(
+                    DatabaseBackend::Postgres,
+                    "INSERT INTO public.editorial_article_chat_messages \
+                            (session_id, role, tool_call_id, tool_name, tool_result) \
+                     VALUES ($1::uuid, 'tool', $2, $3, $4::jsonb)",
+                    vec![
+                        session_id.into(),
+                        tool_call_id.map(|s| s.to_string()).into(),
+                        tool_name.map(|s| s.to_string()).into(),
+                        tool_result.into(),
+                    ],
+                ))
+                .await
+                .map_err(AppError::DatabaseError)?;
+            }
+            other => {
+                tracing::warn!("RunChatTurn event with unknown role={other:?}, skipping");
+            }
+        }
+    }
+
+    // UPDATE article layout if changed
+    if resp.layout_changed && !resp.final_layout_json.is_empty() {
+        let layout_value: serde_json::Value =
+            serde_json::from_str(&resp.final_layout_json).map_err(|e| {
+                AppError::InternalError(format!("final_layout_json parse failed: {e}"))
+            })?;
+        let title = layout_value
+            .get("title")
+            .and_then(|v| v.as_str())
+            .unwrap_or("Untitled")
+            .to_string();
+        let subtitle = layout_value
+            .get("subtitle")
+            .and_then(|v| v.as_str())
+            .map(|s| s.to_string());
+        let hero = layout_value
+            .get("hero_image_url")
+            .and_then(|v| v.as_str())
+            .map(|s| s.to_string());
+
+        db.execute(Statement::from_sql_and_values(
+            DatabaseBackend::Postgres,
+            "UPDATE public.editorial_articles \
+                SET title = $2, \
+                    subtitle = $3, \
+                    hero_image_url = $4, \
+                    layout_json = $5::jsonb, \
+                    updated_at = now() \
+              WHERE id = $1::uuid",
+            vec![
+                article_id.into(),
+                title.into(),
+                subtitle.into(),
+                hero.into(),
+                layout_value.into(),
+            ],
+        ))
+        .await
+        .map_err(AppError::DatabaseError)?;
+    }
+
+    // Touch session.last_message_at
+    db.execute(Statement::from_sql_and_values(
+        DatabaseBackend::Postgres,
+        "UPDATE public.editorial_article_chat_sessions \
+            SET last_message_at = now(), updated_at = now() \
+          WHERE id = $1::uuid",
+        vec![session_id.into()],
+    ))
+    .await
+    .map_err(AppError::DatabaseError)?;
+
+    Ok(Json(SendMessageResponse {
+        assistant_text: resp.final_text,
+        tool_calls_made: resp.tool_calls_made,
+        layout_changed: resp.layout_changed,
+    }))
+}
+
+pub fn router(state: AppState, app_config: AppConfig) -> Router<AppState> {
+    Router::new()
+        .route("/sessions/{article_id}", get(list_sessions).post(create_session))
+        .route("/messages/{session_id}", get(list_messages).post(send_message))
+        .layer(axum::middleware::from_fn_with_state(
+            state,
+            crate::middleware::admin_db_middleware,
+        ))
+        .layer(axum::middleware::from_fn_with_state(
+            app_config,
+            crate::middleware::auth_middleware,
+        ))
+}
diff --git a/packages/api-server/src/domains/admin/handlers.rs b/packages/api-server/src/domains/admin/handlers.rs
index d2239817..5b1b5c51 100644
--- a/packages/api-server/src/domains/admin/handlers.rs
+++ b/packages/api-server/src/domains/admin/handlers.rs
@@ -7,9 +7,9 @@ use axum::Router;
 use crate::{app_state::AppState, config::AppConfig};
 
 use super::{
-    badges, categories, curations, dashboard, editorial_articles, editorial_candidates,
-    editorial_discovery_settings, editorial_pipeline_settings, editorial_recommendations,
-    magazine_sessions, monitoring, posts, solutions, spots, synonyms,
+    badges, categories, curations, dashboard, editorial_article_chat, editorial_articles,
+    editorial_candidates, editorial_discovery_settings, editorial_pipeline_settings,
+    editorial_recommendations, magazine_sessions, monitoring, posts, solutions, spots, synonyms,
 };
 use crate::domains::reports;
 
@@ -32,6 +32,10 @@ pub fn router(state: AppState, app_config: AppConfig) -> Router<AppState> {
             "/editorial-articles",
             editorial_articles::router(state.clone(), app_config.clone()),
         )
+        .nest(
+            "/editorial-article-chat",
+            editorial_article_chat::router(state.clone(), app_config.clone()),
+        )
         .nest(
             "/editorial-discovery/settings",
             editorial_discovery_settings::router(state.clone(), app_config.clone()),
diff --git a/packages/api-server/src/domains/admin/mod.rs b/packages/api-server/src/domains/admin/mod.rs
index f404ece0..f79112df 100644
--- a/packages/api-server/src/domains/admin/mod.rs
+++ b/packages/api-server/src/domains/admin/mod.rs
@@ -7,6 +7,7 @@ pub mod badges;
 pub mod categories;
 pub mod curations;
 pub mod dashboard;
+pub mod editorial_article_chat;
 pub mod editorial_articles;
 pub mod editorial_candidates;
 pub mod editorial_discovery_settings;
diff --git a/packages/api-server/src/entities/assets_editorial_article_chat_messages.rs b/packages/api-server/src/entities/assets_editorial_article_chat_messages.rs
new file mode 100644
index 00000000..c189a97f
--- /dev/null
+++ b/packages/api-server/src/entities/assets_editorial_article_chat_messages.rs
@@ -0,0 +1,46 @@
+use sea_orm::entity::prelude::*;
+use serde::{Deserialize, Serialize};
+
+/// assets.public.editorial_article_chat_messages (#446 fixup).
+///
+/// 한 chat session 의 메시지 시퀀스. role 은 user/assistant/tool/system.
+/// tool_calls / tool_result 는 자유 형식 JSON (Gemini function calling 페이로드).
+#[derive(Clone, Debug, PartialEq, DeriveEntityModel, Serialize, Deserialize)]
+#[sea_orm(table_name = "editorial_article_chat_messages")]
+pub struct Model {
+    /// bigserial — DB 가 자동 생성.
+    #[sea_orm(primary_key)]
+    pub id: i64,
+
+    /// editorial_article_chat_sessions.id (FK, ON DELETE CASCADE).
+    pub session_id: Uuid,
+
+    /// 'user' | 'assistant' | 'tool' | 'system' (CHECK 제약).
+    pub role: String,
+
+    #[sea_orm(nullable, column_type = "Text")]
+    pub content: Option<String>,
+
+    /// assistant 가 호출한 tool 들. shape: [{name, args, id}].
+    #[sea_orm(nullable, column_type = "JsonBinary")]
+    pub tool_calls: Option<Json>,
+
+    /// role='tool' 일 때, 매칭되는 assistant tool_call 의 id.
+    #[sea_orm(nullable, column_type = "Text")]
+    pub tool_call_id: Option<String>,
+
+    /// role='tool' 일 때, 호출된 tool 이름.
+    #[sea_orm(nullable, column_type = "Text")]
+    pub tool_name: Option<String>,
+
+    /// role='tool' 일 때 tool 실행 결과 (자유 형식 dict).
+    #[sea_orm(nullable, column_type = "JsonBinary")]
+    pub tool_result: Option<Json>,
+
+    pub created_at: DateTimeWithTimeZone,
+}
+
+#[derive(Copy, Clone, Debug, EnumIter, DeriveRelation)]
+pub enum Relation {}
+
+impl ActiveModelBehavior for ActiveModel {}
diff --git a/packages/api-server/src/entities/assets_editorial_article_chat_sessions.rs b/packages/api-server/src/entities/assets_editorial_article_chat_sessions.rs
new file mode 100644
index 00000000..92ebd4ea
--- /dev/null
+++ b/packages/api-server/src/entities/assets_editorial_article_chat_sessions.rs
@@ -0,0 +1,35 @@
+use sea_orm::entity::prelude::*;
+use serde::{Deserialize, Serialize};
+
+/// assets.public.editorial_article_chat_sessions (#446 fixup).
+///
+/// 매거진 article 별 대화형 편집 세션. api-server 가 소유 (이전엔 ai-server
+/// 가 직접 manage 했으나 #446 fixup 으로 이전).
+#[derive(Clone, Debug, PartialEq, DeriveEntityModel, Serialize, Deserialize)]
+#[sea_orm(table_name = "editorial_article_chat_sessions")]
+pub struct Model {
+    #[sea_orm(primary_key, column_type = "Uuid", auto_increment = false)]
+    pub id: Uuid,
+
+    /// assets.editorial_articles.id (FK, ON DELETE CASCADE).
+    pub article_id: Uuid,
+
+    /// admin user id (auth.users.id, FK 없음 — assets ↔ auth 다른 스키마).
+    #[sea_orm(nullable)]
+    pub created_by: Option<Uuid>,
+
+    #[sea_orm(nullable)]
+    pub title: Option<String>,
+
+    /// 마지막 메시지 timestamp — 세션 list ordering 용.
+    #[sea_orm(nullable)]
+    pub last_message_at: Option<DateTimeWithTimeZone>,
+
+    pub created_at: DateTimeWithTimeZone,
+    pub updated_at: DateTimeWithTimeZone,
+}
+
+#[derive(Copy, Clone, Debug, EnumIter, DeriveRelation)]
+pub enum Relation {}
+
+impl ActiveModelBehavior for ActiveModel {}
diff --git a/packages/api-server/src/entities/mod.rs b/packages/api-server/src/entities/mod.rs
index 6edc62f7..2d32ae71 100644
--- a/packages/api-server/src/entities/mod.rs
+++ b/packages/api-server/src/entities/mod.rs
@@ -42,6 +42,8 @@ pub mod votes;
 // raw_posts* 는 신규 assets 프로젝트의 public 스키마로 이동 (assets_* 접두).
 pub mod admin_audit_log;
 pub mod artists;
+pub mod assets_editorial_article_chat_messages;
+pub mod assets_editorial_article_chat_sessions;
 pub mod assets_pipeline_events;
 pub mod assets_pipeline_settings;
 pub mod assets_raw_post_sources;
@@ -196,3 +198,11 @@ pub use assets_pipeline_events::Model as AssetsPipelineEventsModel;
 pub use assets_pipeline_settings::ActiveModel as AssetsPipelineSettingsActiveModel;
 pub use assets_pipeline_settings::Entity as AssetsPipelineSettings;
 pub use assets_pipeline_settings::Model as AssetsPipelineSettingsModel;
+
+pub use assets_editorial_article_chat_sessions::ActiveModel as AssetsEditorialArticleChatSessionsActiveModel;
+pub use assets_editorial_article_chat_sessions::Entity as AssetsEditorialArticleChatSessions;
+pub use assets_editorial_article_chat_sessions::Model as AssetsEditorialArticleChatSessionsModel;
+
+pub use assets_editorial_article_chat_messages::ActiveModel as AssetsEditorialArticleChatMessagesActiveModel;
+pub use assets_editorial_article_chat_messages::Entity as AssetsEditorialArticleChatMessages;
+pub use assets_editorial_article_chat_messages::Model as AssetsEditorialArticleChatMessagesModel;
diff --git a/packages/api-server/src/services/decoded_ai_grpc/client.rs b/packages/api-server/src/services/decoded_ai_grpc/client.rs
index a40a53f8..5295c11a 100644
--- a/packages/api-server/src/services/decoded_ai_grpc/client.rs
+++ b/packages/api-server/src/services/decoded_ai_grpc/client.rs
@@ -10,7 +10,8 @@ use crate::grpc::inbound::{
     AnalyzeLinkDirectResponse, AnalyzeLinkRequest, AnalyzeLinkResponse, CategoryRule,
     ExtractOgDataRequest, ExtractOgDataResponse, ExtractPostContextRequest,
     ExtractPostContextResponse, ProcessPostEditorialRequest, ProcessPostEditorialResponse,
-    ReparseRawPostRequest, ReparseRawPostResponse, TriggerSourceRequest, TriggerSourceResponse,
+    ReparseRawPostRequest, ReparseRawPostResponse, RunChatTurnRequest, RunChatTurnResponse,
+    TriggerSourceRequest, TriggerSourceResponse,
 };
 use crate::observability::grpc::record_decoded_ai_call;
 
@@ -283,6 +284,33 @@ impl DecodedAIGrpcClient {
         record_decoded_ai_call("reparse_raw_post", res.is_ok(), start.elapsed());
         res
     }
+
+    /// #446 fixup — editorial article chat 의 한 턴 LLM 실행.
+    /// LLM tool loop 가 길 수 있으므로 deadline 120s.
+    pub async fn run_chat_turn(
+        &self,
+        article_id: String,
+        layout_json: String,
+        history_json: String,
+        user_message: String,
+    ) -> Result<RunChatTurnResponse, Box<dyn std::error::Error>> {
+        let start = Instant::now();
+        let mut client = self.client.clone();
+        let mut request = tonic::Request::new(RunChatTurnRequest {
+            article_id,
+            layout_json,
+            history_json,
+            user_message,
+        });
+        request.set_timeout(std::time::Duration::from_secs(120));
+        let res = async {
+            let response = client.run_chat_turn(request).await?;
+            Ok::<_, Box<dyn std::error::Error>>(response.into_inner())
+        }
+        .await;
+        record_decoded_ai_call("run_chat_turn", res.is_ok(), start.elapsed());
+        res
+    }
 }
 
 #[cfg(test)]
diff --git a/packages/web/app/api/admin/editorial-article-chat/messages/[sessionId]/route.ts b/packages/web/app/api/admin/editorial-article-chat/messages/[sessionId]/route.ts
index d47fc2f7..41634117 100644
--- a/packages/web/app/api/admin/editorial-article-chat/messages/[sessionId]/route.ts
+++ b/packages/web/app/api/admin/editorial-article-chat/messages/[sessionId]/route.ts
@@ -1,13 +1,14 @@
 import { NextRequest, NextResponse } from "next/server";
 import { createSupabaseServerClient } from "@/lib/supabase/server";
 import { checkIsAdmin } from "@/lib/supabase/admin";
-import { AI_SERVER_HTTP_URL } from "@/lib/server-env";
+import { API_BASE_URL } from "@/lib/server-env";
 
 // GET  /api/admin/editorial-article-chat/messages/{sessionId}  → 메시지 리스트
-// POST /api/admin/editorial-article-chat/messages/{sessionId}  → user 메시지 → agent 실행
-// (Stage 3, issue #429)
+// POST /api/admin/editorial-article-chat/messages/{sessionId}  → user 메시지 → agent 한 턴
+//
+// (#446 fixup) — api-server 가 chat 소유. send_message 는 ~30s 걸릴 수 있음.
 
-async function adminCheck() {
+async function adminSession() {
   const supabase = await createSupabaseServerClient();
   const {
     data: { user },
@@ -16,30 +17,40 @@ async function adminCheck() {
   if (!(await checkIsAdmin(supabase, user.id))) {
     return { error: "Forbidden", status: 403 as const };
   }
-  return { userId: user.id };
+  const {
+    data: { session },
+  } = await supabase.auth.getSession();
+  if (!session?.access_token) {
+    return { error: "No session", status: 401 as const };
+  }
+  return { token: session.access_token };
 }
 
 export async function GET(
   _req: NextRequest,
   context: { params: Promise<{ sessionId: string }> }
 ) {
-  const auth = await adminCheck();
+  const auth = await adminSession();
   if ("error" in auth) {
     return NextResponse.json({ error: auth.error }, { status: auth.status });
   }
-  if (!AI_SERVER_HTTP_URL) {
+  if (!API_BASE_URL) {
     return NextResponse.json(
-      { error: "AI_SERVER_HTTP_URL not configured" },
+      { error: "API_BASE_URL is not configured" },
       { status: 502 }
     );
   }
   const { sessionId } = await context.params;
   try {
     const res = await fetch(
-      `${AI_SERVER_HTTP_URL}/api/v1/editorial-article-chat/sessions/${encodeURIComponent(sessionId)}/messages`
+      `${API_BASE_URL}/api/v1/admin/editorial-article-chat/messages/${encodeURIComponent(sessionId)}`,
+      {
+        headers: { Authorization: `Bearer ${auth.token}` },
+        cache: "no-store",
+      }
     );
-    const text = await res.text();
-    return new NextResponse(text, {
+    const body = await res.text();
+    return new NextResponse(body, {
       status: res.status,
       headers: { "content-type": "application/json" },
     });
@@ -55,29 +66,32 @@ export async function POST(
   req: NextRequest,
   context: { params: Promise<{ sessionId: string }> }
 ) {
-  const auth = await adminCheck();
+  const auth = await adminSession();
   if ("error" in auth) {
     return NextResponse.json({ error: auth.error }, { status: auth.status });
   }
-  if (!AI_SERVER_HTTP_URL) {
+  if (!API_BASE_URL) {
     return NextResponse.json(
-      { error: "AI_SERVER_HTTP_URL not configured" },
+      { error: "API_BASE_URL is not configured" },
       { status: 502 }
     );
   }
   const { sessionId } = await context.params;
-  const body = await req.text();
+  const payload = await req.text();
   try {
     const res = await fetch(
-      `${AI_SERVER_HTTP_URL}/api/v1/editorial-article-chat/sessions/${encodeURIComponent(sessionId)}/messages`,
+      `${API_BASE_URL}/api/v1/admin/editorial-article-chat/messages/${encodeURIComponent(sessionId)}`,
       {
         method: "POST",
-        headers: { "content-type": "application/json" },
-        body,
+        headers: {
+          "content-type": "application/json",
+          Authorization: `Bearer ${auth.token}`,
+        },
+        body: payload,
       }
     );
-    const text = await res.text();
-    return new NextResponse(text, {
+    const body = await res.text();
+    return new NextResponse(body, {
       status: res.status,
       headers: { "content-type": "application/json" },
     });
diff --git a/packages/web/app/api/admin/editorial-article-chat/sessions/[articleId]/route.ts b/packages/web/app/api/admin/editorial-article-chat/sessions/[articleId]/route.ts
index d76ecc4a..37e71c7f 100644
--- a/packages/web/app/api/admin/editorial-article-chat/sessions/[articleId]/route.ts
+++ b/packages/web/app/api/admin/editorial-article-chat/sessions/[articleId]/route.ts
@@ -1,13 +1,14 @@
 import { NextRequest, NextResponse } from "next/server";
 import { createSupabaseServerClient } from "@/lib/supabase/server";
 import { checkIsAdmin } from "@/lib/supabase/admin";
-import { AI_SERVER_HTTP_URL } from "@/lib/server-env";
+import { API_BASE_URL } from "@/lib/server-env";
 
 // GET  /api/admin/editorial-article-chat/sessions/{articleId}  → 세션 리스트
 // POST /api/admin/editorial-article-chat/sessions/{articleId}  → 새 세션 생성
-// (Stage 3, issue #429)
+//
+// (#446 fixup) — api-server 가 chat 소유. 이전엔 ai-server FastAPI 직접 호출.
 
-async function adminCheck() {
+async function adminSession() {
   const supabase = await createSupabaseServerClient();
   const {
     data: { user },
@@ -16,35 +17,40 @@ async function adminCheck() {
   if (!(await checkIsAdmin(supabase, user.id))) {
     return { error: "Forbidden", status: 403 as const };
   }
-  return { userId: user.id };
-}
-
-function aiBase(): string | null {
-  return AI_SERVER_HTTP_URL || null;
+  const {
+    data: { session },
+  } = await supabase.auth.getSession();
+  if (!session?.access_token) {
+    return { error: "No session", status: 401 as const };
+  }
+  return { token: session.access_token };
 }
 
 export async function GET(
   _req: NextRequest,
   context: { params: Promise<{ articleId: string }> }
 ) {
-  const auth = await adminCheck();
+  const auth = await adminSession();
   if ("error" in auth) {
     return NextResponse.json({ error: auth.error }, { status: auth.status });
   }
-  const base = aiBase();
-  if (!base) {
+  if (!API_BASE_URL) {
     return NextResponse.json(
-      { error: "AI_SERVER_HTTP_URL not configured" },
+      { error: "API_BASE_URL is not configured" },
       { status: 502 }
     );
   }
   const { articleId } = await context.params;
   try {
     const res = await fetch(
-      `${base}/api/v1/editorial-article-chat/sessions/${encodeURIComponent(articleId)}`
+      `${API_BASE_URL}/api/v1/admin/editorial-article-chat/sessions/${encodeURIComponent(articleId)}`,
+      {
+        headers: { Authorization: `Bearer ${auth.token}` },
+        cache: "no-store",
+      }
     );
-    const text = await res.text();
-    return new NextResponse(text, {
+    const body = await res.text();
+    return new NextResponse(body, {
       status: res.status,
       headers: { "content-type": "application/json" },
     });
@@ -60,29 +66,30 @@ export async function POST(
   _req: NextRequest,
   context: { params: Promise<{ articleId: string }> }
 ) {
-  const auth = await adminCheck();
+  const auth = await adminSession();
   if ("error" in auth) {
     return NextResponse.json({ error: auth.error }, { status: auth.status });
   }
-  const base = aiBase();
-  if (!base) {
+  if (!API_BASE_URL) {
     return NextResponse.json(
-      { error: "AI_SERVER_HTTP_URL not configured" },
+      { error: "API_BASE_URL is not configured" },
       { status: 502 }
     );
   }
   const { articleId } = await context.params;
   try {
     const res = await fetch(
-      `${base}/api/v1/editorial-article-chat/sessions/${encodeURIComponent(articleId)}`,
+      `${API_BASE_URL}/api/v1/admin/editorial-article-chat/sessions/${encodeURIComponent(articleId)}`,
       {
         method: "POST",
-        headers: { "content-type": "application/json" },
-        body: JSON.stringify({ created_by: auth.userId }),
+        headers: {
+          "content-type": "application/json",
+          Authorization: `Bearer ${auth.token}`,
+        },
       }
     );
-    const text = await res.text();
-    return new NextResponse(text, {
+    const body = await res.text();
+    return new NextResponse(body, {
       status: res.status,
       headers: { "content-type": "application/json" },
     });
diff --git a/packages/web/lib/server-env.ts b/packages/web/lib/server-env.ts
index 57484b4e..d2f57bfd 100644
--- a/packages/web/lib/server-env.ts
+++ b/packages/web/lib/server-env.ts
@@ -8,12 +8,6 @@ function getEnvOrEmpty(name: string): string {
  *
  * 참고 — `AI_SERVER_URL` 은 #416 에서 제거됨. 모든 ai-server 진입은
  * api-server gRPC 통해서. ai-server 는 internal docker network only.
+ * (#446 fixup) — chat 도 동일 원칙. AI_SERVER_HTTP_URL 도 제거됨.
  */
 export const API_BASE_URL = getEnvOrEmpty("API_BASE_URL");
-
-/**
- * AI-server HTTP base URL (#429 Stage 3 chat 예외 — gRPC 아닌 HTTP).
- * dev: http://localhost:10000. prod: ai.decoded.style 또는 internal route.
- * 비어 있으면 chat endpoints 가 502 반환.
- */
-export const AI_SERVER_HTTP_URL = getEnvOrEmpty("AI_SERVER_HTTP_URL");
diff --git a/supabase/migrations/20260505020001_editorial_articles_published.sql b/supabase/migrations/20260505020001_editorial_articles_published.sql
index 91a08259..11b5848d 100644
--- a/supabase/migrations/20260505020001_editorial_articles_published.sql
+++ b/supabase/migrations/20260505020001_editorial_articles_published.sql
@@ -7,6 +7,11 @@
 
 BEGIN;
 
+-- staging editorial_articles 가 같은 이름으로 살아있을 수 있음 (PR #446 머지 후
+-- operation cloud 에 020002 보다 먼저 도달). publish 스키마 (slug 컬럼 등) 와
+-- 충돌하므로 강제 drop 후 재생성. CASCADE 로 chat/events FK 같이 정리.
+DROP TABLE IF EXISTS public.editorial_articles CASCADE;
+
 CREATE TABLE IF NOT EXISTS public.editorial_articles (
     id                  uuid PRIMARY KEY,                      -- assets article id 와 동일
     source_article_id   uuid NOT NULL,                          -- assets 의 원본 (logical, FK 없음)