From 51a6e845d108cae491ea9cf6b0b11b28d1aa8ced Mon Sep 17 00:00:00 2001
From: burnintuna <burningtuna@gmail.com>
Date: Thu, 30 Jul 2026 17:07:02 +0900
Subject: [PATCH] Improve translation reliability and progress feedback

---
 .env.example                          |   7 +
 .gitignore                            |   1 +
 AGENTS.md                             |  28 +-
 README.md                             |  25 +-
 backend/auth.py                       |  59 ++-
 backend/llm_client.py                 |  33 +-
 backend/main.py                       | 543 ++++++++++++++++++++++----
 backend/models.py                     |  17 +
 backend/prompts.py                    |  38 +-
 backend/sessions.py                   |  61 ++-
 frontend/src/api.ts                   |  29 +-
 frontend/src/stores/auth.ts           |  13 +-
 frontend/src/stores/translation.ts    | 241 +++++++++---
 frontend/src/types.ts                 |  23 ++
 frontend/src/views/LoginView.vue      |   7 +-
 frontend/src/views/TranslatorView.vue |  29 +-
 requirements-dev.txt                  |   2 +
 requirements.txt                      |   2 +-
 tests/__init__.py                     |   0
 tests/test_backend.py                 | 252 ++++++++++++
 20 files changed, 1194 insertions(+), 216 deletions(-)
 create mode 100644 requirements-dev.txt
 create mode 100644 tests/__init__.py
 create mode 100644 tests/test_backend.py

diff --git a/.env.example b/.env.example
index 2281e3b..330d923 100644
--- a/.env.example
+++ b/.env.example
@@ -1,6 +1,9 @@
 # JWT 시그니용 키 (변경 권장)
 JWT_SECRET_KEY=my-super-secret-key-change-me
 
+# JWT_SECRET_KEY 미설정 시 생성할 키 파일 (기본: config/.jwt-secret)
+# JWT_SECRET_FILE=config/.jwt-secret
+
 # LLM 설정 파일 경로 (기본: config/llms.json)
 LLMS_CONFIG_PATH=config/llms.json
 
@@ -15,3 +18,7 @@ PORT=8000
 
 # 프론트엔드 빌드 폴더 경로 (기본: frontend/dist)
 FRONTEND_DIST_PATH=frontend/dist
+
+# 개별 LLM 요청 제한 시간(초)과 장문 분할 크기(문자 수)
+LLM_TIMEOUT_SECONDS=180
+TRANSLATION_CHUNK_CHARS=1500
diff --git a/.gitignore b/.gitignore
index 42fa114..18c2811 100644
--- a/.gitignore
+++ b/.gitignore
@@ -12,6 +12,7 @@ frontend/dist/
 
 # Env files
 .env
+config/.jwt-secret
 
 # IDE
 .vscode/
diff --git a/AGENTS.md b/AGENTS.md
index 80e9b9b..818a716 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -19,7 +19,7 @@ Single-server deployment: `uvicorn backend.main:app` serves both API and fronten
 | Auth | JWT (PyJWT), bcrypt (direct via `bcrypt` module) |
 | LLM Client | OpenAI SDK (`openai`) — OpenAI-compatible API format |
 | Frontend | Vue 3 + Composition API `<script setup>`, TypeScript, Pinia |
-| Styling | Tailwind CSS v4 with `darkMode: 'class'` strategy |
+| Styling | Tailwind CSS v3 with `darkMode: 'class'` strategy |
 | Build | Vite (frontend), no bundler for backend |
 
 ## Directory Structure
@@ -55,10 +55,11 @@ LLM-translator/
 │   └── package.json
 ├── config/
 │   ├── llms.json                # LLM server configs [{alias, base_url, api_key, model, context_size}]
-│   └── users.json               # Users [{id, password_hash}] — hashed on first startup from plain-text
+│   └── users.json               # Users [{id, password}] — password_plain is migrated on startup
 ├── requirements.txt             # Python deps: fastapi, uvicorn[standard], pyjwt, bcrypt, openai, python-dotenv
 ├── .env.example                 # Environment variables template
-└── README.md                    # (TBD)
+├── tests/                       # Backend API and session regression tests
+└── README.md                    # Setup and operation guide
 ```
 
 ## Key Patterns & Conventions
@@ -66,9 +67,12 @@ LLM-translator/
 ### Backend
 
 - **No `passlib`** — use `bcrypt` module directly (`bcrypt.hashpw`, `bcrypt.checkpw`). The project uses Python 3.14 which has compatibility issues with passlib's bcrypt wrapper.
-- **LLM client caching** — `llm_client.py` maintains a per-alias `{AsyncOpenAI, LLMConfig}` cache dict keyed by alias string. Clients are lazily instantiated on first use.
-- **Session store** — simple in-memory dict (`sessions.py`). Key is UUID hex (16 chars). TTL 24h from last access. No persistence beyond process lifetime.
+- **LLM client caching** — `llm_client.py` caches an `AsyncOpenAI` client by alias. Clients are lazily instantiated on first use.
+- **Session store** — process-local in-memory dict (`sessions.py`) scoped by authenticated owner. Key is UUID hex (16 chars). TTL is 24h from last access; there is no persistence beyond process lifetime.
 - **JSON extraction** — LLM responses may be wrapped in markdown code blocks. `_extract_json()` strips ```` ```json` ... ```` wrappers before `json.loads()`.
+- **Long-text chunking** — phases 1, 3, and 4 process text sequentially near sentence boundaries. Defaults are 1,500 characters per chunk and a 180-second timeout per LLM call; both are configurable by environment variables.
+- **Progress feedback** — LLM calls stream response deltas into transient session progress. The frontend polls `/api/sessions/{session_id}/progress` and displays the current chunk and a rolling preview.
+- **Malformed Phase 1 JSON** — parsing tolerates fences, surrounding prose, alternate translation keys, missing noun fields, and trailing commas. Remaining failures trigger one JSON-repair call and then a plain-translation fallback.
 
 ### Frontend
 
@@ -82,7 +86,7 @@ LLM-translator/
 Models are defined in `config/llms.json`. Each entry:
 ```json
 {
-  "alias": "Qwen3.6-35B",      // Display name, used as modelPhase1..4 values
+  "alias": "Qwen3.6-35B",      // Display name, used by LLM phases 1, 3, and 4
   "base_url": "http://...",     // OpenAI-compatible base URL (must end with /v1)
   "api_key": "...",             // API key for this server
   "model": "hx370/qwen3.6-35b-a3b",  // Exact model name as used by the LLM provider
@@ -105,7 +109,7 @@ source_text ──► Phase1(LLM) ──► {translated, proper_nouns[], summary
                               phase4_result (final)
 ```
 
-Phase buttons are always freely executable — user can jump to any phase independently. Dependencies are only logical: P2 needs P1 data, P3 needs P1 result, P4 needs P3 result.
+Phase dependencies are enforced: Phase 2 and 3 require Phase 1, and Phase 4 requires Phase 3. Running Phase 3 automatically submits the current Phase 2 edits.
 
 ### Theme / Dark Mode
 
@@ -121,7 +125,7 @@ pip install -r requirements.txt
 cd frontend && npm install && cd ..
 
 # 2. Build frontend (required before serving in production mode)
-cd frontend && npx vite build && cd ..
+cd frontend && npm run build && cd ..
 
 # 3. Configure LLM servers and users
 #    Edit config/llms.json and config/users.json directly
@@ -142,10 +146,12 @@ JWT_SECRET_KEY=your-secret uvicorn backend.main:app --reload
 2. Add API endpoint in `backend/main.py` following the existing pattern:
    ```python
    @app.post("/api/translate/{session_id}/phaseN")
-   async def run_phaseN(session_id: str):
-       session = session_store.get(session_id)
+   async def run_phaseN(session_id: str, user_id: str = Depends(get_current_user)):
+       session = session_store.get(session_id, user_id)
        # ... call chat_complete(alias, system_prompt, user_content)
-       updated = session_store.update(session_id, {"phaseN_result": result_text})
+       updated = session_store.update(
+           session_id, {"phaseN_result": result_text}, user_id
+       )
        return SessionResponse(session_id=session_id, data=updated).model_dump()
    ```
 3. Add frontend component and button in `TranslatorView.vue`
diff --git a/README.md b/README.md
index 3d09c93..755484b 100644
--- a/README.md
+++ b/README.md
@@ -5,7 +5,7 @@
 ## 특징
 
 - **4단계 번역 파이프라인**: 초벌번역 → 고유명사/문체 분석 → 재번역 → 마무리 다듬기
-- **Phase별 별도 LLM 모델 선택**: 각 단계마다 다른 모델을 지정 가능
+- **LLM 단계별 모델 선택**: Phase 1, 3, 4에 서로 다른 모델 지정 가능
 - **고유명사 관리**: 추출된 고유명사를 사용자가 검토·수정 가능
 - **세로 3열 비교 UI**: Phase 1 / Phase 3 / Phase 4 결과를 한눈에 비교
 - **다크모드 지원**: 시스템 설정 자동 감지 + 수동 토글 (localStorage 지속화)
@@ -20,7 +20,7 @@ pip install -r requirements.txt
 cd frontend && npm install && cd ..
 
 # 2. 프론트엔드 빌드 (프로덕션 모드)
-npx vite build --prefix-links=false
+cd frontend && npm run build && cd ..
 
 # 3. 서버 실행
 uvicorn backend.main:app --host 0.0.0.0 --port 8000
@@ -58,11 +58,14 @@ uvicorn backend.main:app --host 0.0.0.0 --port 8000
 
 | 변수 | 설명 | 기본값 |
 |------|------|--------|
-| `JWT_SECRET_KEY` | JWT 서명 키 | `dev-secret-key-change-in-production` |
+| `JWT_SECRET_KEY` | JWT 서명 키 | 미설정 시 키 파일 자동 생성 |
+| `JWT_SECRET_FILE` | 서명 키 자동 생성 파일 | `config/.jwt-secret` |
 | `JWT_EXPIRE_HOURS` | 토큰 만료 시간(시간) | `24` |
 | `LLMS_CONFIG_PATH` | LLM 설정 파일 경로 | `config/llms.json` |
 | `USERS_CONFIG_PATH` | 사용자 설정 파일 경로 | `config/users.json` |
 | `PORT` | 서버 포트 | `8000` |
+| `LLM_TIMEOUT_SECONDS` | 개별 LLM 호출 제한 시간 | `180` |
+| `TRANSLATION_CHUNK_CHARS` | 장문 분할 목표 크기 | `1500` |
 
 ## 프로젝트 구조
 
@@ -92,7 +95,7 @@ cd frontend && npx vite        # http://localhost:5173
 |--------|------|------|------|
 | `POST` | `/api/auth/login` | 로그인 → JWT 토큰 발급 | × |
 | `GET`  | `/api/auth/me`     | 현재 사용자 정보 | ○ (JWT) |
-| `GET`  | `/api/models`      | 사용 가능한 LLM 모델 목록 | × |
+| `GET`  | `/api/models`      | 사용 가능한 LLM 모델 목록 | ○ |
 | `POST` | `/api/translate/create` | 새 번역 세션 생성 | ○ |
 | `GET`  | `/api/sessions/{id}`   | 세션 상태 조회 | ○ |
 | `PATCH`| `/api/sessions/{id}`   | 세션 필드 부분 업데이트 | ○ |
@@ -105,7 +108,7 @@ cd frontend && npx vite        # http://localhost:5173
 
 ```bash
 # 1. 프론트엔드 빌드
-cd frontend && npx vite build && cd ..
+cd frontend && npm run build && cd ..
 
 # 2. 백엔드 서빙 (dist 정적 파일 포함)
 uvicorn backend.main:app --host 0.0.0.0 --port 8000
@@ -113,7 +116,17 @@ uvicorn backend.main:app --host 0.0.0.0 --port 8000
 
 `backend/main.py`는 `frontend/dist/` 폴더를 정적 파일로 마운트하므로, 빌드 후 백엔드 서버 하나만 실행하면 됩니다.
 
-## 테스트 (LLM 서버 연결 확인)
+장문은 문장 또는 줄 경계를 우선하여 기본 1,500자 단위로 나누고 Phase 1, 3, 4에서 순차 처리합니다. 공개 LLM 서버가 혼잡하면 `.env`에서 `LLM_TIMEOUT_SECONDS`를 늘리거나 `TRANSLATION_CHUNK_CHARS`를 줄일 수 있습니다.
+
+LLM 처리 중에는 현재 chunk 번호와 스트리밍으로 수신 중인 최근 내용이 화면에 임시 표시됩니다. Phase 1 JSON이 불완전하면 코드 블록·주변 문구·후행 쉼표를 먼저 보정하고, 실패 시 LLM JSON 복구를 거쳐 일반 번역 결과로 폴백합니다.
+
+## 자동 테스트
+
+```bash
+python -m unittest discover -v
+```
+
+## LLM 서버 연결 확인
 
 ```bash
 curl -s http://btuna.net:45455/v1/chat/completions \
diff --git a/backend/auth.py b/backend/auth.py
index 5061a13..010130c 100644
--- a/backend/auth.py
+++ b/backend/auth.py
@@ -1,21 +1,62 @@
 """JWT-based authentication with config-file user management."""
 
 import os
+import logging
+import secrets
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
 
 import bcrypt
 import jwt
+from dotenv import load_dotenv
 from fastapi import Depends, HTTPException, status
 from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
 
 # ── Config ────────────────────────────────────────────
 
-JWT_SECRET_KEY = os.getenv("JWT_SECRET_KEY", "dev-secret-key-change-in-production")
+PROJECT_ROOT = Path(__file__).parent.parent
+load_dotenv(PROJECT_ROOT / ".env")
+
+
+def _load_jwt_secret() -> str:
+    configured = os.getenv("JWT_SECRET_KEY")
+    if configured:
+        return configured
+
+    secret_path = Path(
+        os.getenv("JWT_SECRET_FILE", str(PROJECT_ROOT / "config" / ".jwt-secret"))
+    )
+    try:
+        secret = secret_path.read_text(encoding="utf-8").strip()
+        if secret:
+            return secret
+    except FileNotFoundError:
+        pass
+
+    secret = secrets.token_urlsafe(48)
+    try:
+        descriptor = os.open(secret_path, os.O_WRONLY | os.O_CREAT | os.O_EXCL, 0o600)
+    except FileExistsError:
+        secret = secret_path.read_text(encoding="utf-8").strip()
+        if secret:
+            return secret
+        raise RuntimeError(f"JWT secret file is empty: {secret_path}")
+    except OSError as exc:
+        raise RuntimeError(
+            "Set JWT_SECRET_KEY because the JWT secret file could not be created: "
+            f"{secret_path}"
+        ) from exc
+
+    with os.fdopen(descriptor, "w", encoding="utf-8") as secret_file:
+        secret_file.write(secret)
+    logging.getLogger(__name__).warning("Created JWT secret file at %s", secret_path)
+    return secret
+
+JWT_SECRET_KEY = _load_jwt_secret()
 JWT_ALGORITHM = "HS256"
 JWT_EXPIRE_HOURS = int(os.getenv("JWT_EXPIRE_HOURS", "24"))
 
-security = HTTPBearer()
+security = HTTPBearer(auto_error=False)
 
 
 # ── User Loading ───────────────────────────────────────
@@ -74,12 +115,20 @@ def decode_access_token(token: str) -> dict | None:
 
 # ── Dependency ───────────────────────────────────────
 
-async def get_current_user(credentials: HTTPAuthorizationCredentials = Depends(security)) -> str:
+async def get_current_user(
+    credentials: HTTPAuthorizationCredentials | None = Depends(security),
+) -> str:
     """Extract and validate JWT, returning the user_id."""
+    if credentials is None:
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail="인증이 필요합니다",
+        )
     payload = decode_access_token(credentials.credentials)
-    if payload is None or "user_id" not in payload:
+    user_id = payload.get("user_id") if payload else None
+    if not isinstance(user_id, str) or get_user_by_username(user_id) is None:
         raise HTTPException(
             status_code=status.HTTP_401_UNAUTHORIZED,
             detail="유효하지 않거나 만료된 토큰입니다",
         )
-    return payload["user_id"]
+    return user_id
diff --git a/backend/llm_client.py b/backend/llm_client.py
index a7dfa1c..232dc20 100644
--- a/backend/llm_client.py
+++ b/backend/llm_client.py
@@ -2,6 +2,7 @@
 
 import os
 from pathlib import Path
+from collections.abc import Awaitable, Callable
 
 from openai import AsyncOpenAI
 
@@ -22,6 +23,7 @@ def _load_llm_configs() -> list[LLMConfig]:
 
 _llm_configs: list[LLMConfig] = []
 _clients: dict[str, AsyncOpenAI] = {}  # alias -> client instance
+LLM_TIMEOUT_SECONDS = float(os.getenv("LLM_TIMEOUT_SECONDS", "180"))
 
 
 def get_llm_configs() -> list[LLMConfig]:
@@ -46,7 +48,8 @@ def get_client_for_alias(alias: str) -> tuple[AsyncOpenAI, LLMConfig]:
         client = AsyncOpenAI(
             api_key=cfg.api_key,
             base_url=cfg.base_url,
-            timeout=120.0,
+            timeout=LLM_TIMEOUT_SECONDS,
+            max_retries=0,
         )
         _clients[alias] = client
 
@@ -58,21 +61,33 @@ async def chat_complete(
     system_prompt: str,
     user_content: str,
     temperature: float = 0.3,
+    on_delta: Callable[[str], Awaitable[None]] | None = None,
 ) -> str:
     """Send a single-turn chat completion and return the assistant message."""
     client, cfg = get_client_for_alias(alias)
-
-    response = await client.chat.completions.create(
-        model=cfg.model,
-        messages=[
+    request = {
+        "model": cfg.model,
+        "messages": [
             {"role": "system", "content": system_prompt},
             {"role": "user", "content": user_content},
         ],
-        temperature=temperature,
-    )
+        "temperature": temperature,
+    }
+    if on_delta is None:
+        response = await client.chat.completions.create(**request)
+        content = response.choices[0].message.content or ""
+        return content.strip()
 
-    content = response.choices[0].message.content or ""
-    return content.strip()
+    stream = await client.chat.completions.create(**request, stream=True)
+    parts: list[str] = []
+    async for event in stream:
+        if not event.choices:
+            continue
+        delta = event.choices[0].delta.content or ""
+        if delta:
+            parts.append(delta)
+            await on_delta(delta)
+    return "".join(parts).strip()
 
 
 def get_context_size(alias: str) -> int | None:
diff --git a/backend/main.py b/backend/main.py
index 94b957f..2206c02 100644
--- a/backend/main.py
+++ b/backend/main.py
@@ -1,12 +1,17 @@
 """FastAPI application — main entry point for the LLM Translator backend."""
 
+import json
 import os
+import re
+from collections import Counter
 from pathlib import Path
+from time import monotonic
 
 from fastapi import Depends, FastAPI, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.staticfiles import StaticFiles
-from pydantic import BaseModel
+from openai import APITimeoutError, OpenAIError
+from pydantic import ValidationError
 
 from backend import auth as auth_mod
 from backend.auth import get_current_user
@@ -24,22 +29,27 @@ from backend.models import (
     ProperNoun,
     SessionResponse,
     TokenResponse,
+    TranslationChunk,
     UserInfo,
     TranslationSession,
 )
 from backend.prompts import (
     SYSTEM_PROMPT_PHASE1,
     SYSTEM_PROMPT_PHASE3,
+    SYSTEM_PROMPT_PHASE4,
     build_phase1_user_prompt,
     build_phase3_user_prompt,
     build_phase4_user_prompt,
 )
-from backend.sessions import session_store
+from backend.sessions import SessionConflictError, session_store
 
 
 # ── App Setup ────────────────────────────────────────
 
 app = FastAPI(title="LLM Translator", version="0.1.0")
+TRANSLATION_CHUNK_CHARS = max(
+    500, int(os.getenv("TRANSLATION_CHUNK_CHARS", "1500"))
+)
 
 # CORS (frontend same-origin by default, allow dev origins)
 app.add_middleware(
@@ -57,9 +67,13 @@ app.add_middleware(
 async def login(req: LoginRequest):
     import bcrypt as _bcrypt
     user = auth_mod.get_user_by_username(req.id)
-    if user is None or not _bcrypt.checkpw(
-        req.password.encode("utf-8"), user["password"].encode("utf-8")
-    ):
+    try:
+        authenticated = user is not None and _bcrypt.checkpw(
+            req.password.encode("utf-8"), user["password"].encode("utf-8")
+        )
+    except (KeyError, ValueError):
+        authenticated = False
+    if not authenticated:
         raise HTTPException(status_code=401, detail="ID 또는 비밀번호가 틀렸습니다")
 
     token = auth_mod.create_access_token({"user_id": user["id"]})
@@ -74,7 +88,7 @@ async def me(user_id: str = Depends(get_current_user)):
 # ── LLM Config Routes ────────────────────────────────
 
 @app.get("/api/models")
-async def list_models():
+async def list_models(_user_id: str = Depends(get_current_user)):
     """Return available LLM model configurations."""
     configs = get_llm_configs()
     return [
@@ -90,7 +104,10 @@ async def list_models():
 # ── Session Routes ───────────────────────────────────
 
 @app.post("/api/translate/create")
-async def create_session(req: CreateSessionRequest):
+async def create_session(
+    req: CreateSessionRequest, user_id: str = Depends(get_current_user)
+):
+    _validate_model_aliases(req.model_dump())
     session_data = TranslationSession(
         source_text=req.source_text,
         source_language=req.source_language,
@@ -100,30 +117,71 @@ async def create_session(req: CreateSessionRequest):
         model_phase3=req.model_phase3,
         model_phase4=req.model_phase4,
     )
-    session_id = session_store.create(session_data)
+    session_id = session_store.create(session_data, user_id)
     return {"session_id": session_id}
 
 
 @app.get("/api/sessions/{session_id}")
-async def get_session(session_id: str):
-    data = session_store.get(session_id)
+async def get_session(session_id: str, user_id: str = Depends(get_current_user)):
+    data = session_store.get(session_id, user_id)
     if data is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
     return SessionResponse(session_id=session_id, data=data).model_dump()
 
 
+@app.get("/api/sessions/{session_id}/progress")
+async def get_session_progress(
+    session_id: str, user_id: str = Depends(get_current_user)
+):
+    data = session_store.get(session_id, user_id)
+    if data is None:
+        raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
+    return {"progress": data.progress.model_dump() if data.progress else None}
+
+
 @app.delete("/api/sessions/{session_id}")
-async def delete_session(session_id: str):
-    deleted = session_store.delete(session_id)
+async def delete_session(session_id: str, user_id: str = Depends(get_current_user)):
+    deleted = session_store.delete(session_id, user_id)
     if not deleted:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
     return {"deleted": True}
 
 
 @app.patch("/api/sessions/{session_id}")
-async def update_session(session_id: str, req: PhaseUpdateRequest):
+async def update_session(
+    session_id: str,
+    req: PhaseUpdateRequest,
+    user_id: str = Depends(get_current_user),
+):
     data = req.model_dump(exclude_none=True)
-    updated = session_store.update(session_id, data)
+    _validate_model_aliases(data)
+    current = session_store.get(session_id, user_id)
+    if current is None:
+        raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
+
+    changed = {key for key, value in data.items() if getattr(current, key) != value}
+    if changed & {"source_text", "source_language", "target_language", "model_phase1"}:
+        data.update({
+            "phase1_result": None,
+            "phase1_chunks": [],
+            "phase2_proper_nouns": [],
+            "phase2_style": "",
+            "phase3_result": "",
+            "phase3_chunks": [],
+            "phase4_result": "",
+            "progress": None,
+        })
+    elif changed & {"model_phase3"}:
+        data.update({
+            "phase3_result": "",
+            "phase3_chunks": [],
+            "phase4_result": "",
+            "progress": None,
+        })
+    elif changed & {"model_phase4"}:
+        data.update({"phase4_result": "", "progress": None})
+
+    updated = session_store.update(session_id, data, user_id)
     if updated is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
     return SessionResponse(session_id=session_id, data=updated).model_dump()
@@ -132,46 +190,69 @@ async def update_session(session_id: str, req: PhaseUpdateRequest):
 # ── Phase Execution Routes ───────────────────────────
 
 @app.post("/api/translate/{session_id}/phase1")
-async def run_phase1(session_id: str):
+async def run_phase1(session_id: str, user_id: str = Depends(get_current_user)):
     """Run Phase 1: rough translation + proper noun extraction + summary + style analysis."""
-    session = session_store.get(session_id)
-    if session is None:
+    snapshot = session_store.get_with_version(session_id, user_id)
+    if snapshot is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
+    session, version = snapshot
 
     alias = session.model_phase1
     if not alias:
         raise HTTPException(status_code=400, detail="Phase 1 LLM 모델을 선택하세요")
+    if not session.source_text.strip():
+        raise HTTPException(status_code=400, detail="원문을 입력하세요")
 
-    # Context size check (rough: source text tokens ~ len/2 for Korean)
-    ctx_size = get_context_size(alias)
-    if ctx_size and len(session.source_text) > ctx_size * 0.7:
-        raise HTTPException(
-            status_code=400,
-            detail=f"원문이 너무 깁니다 (최대 {ctx_size} 컨텍스트의 70% 이내로 줄여주세요)",
+    chunk_results: list[tuple[str, Phase1Result]] = []
+    source_chunks = _split_text(session.source_text)
+    for chunk_index, source_chunk in enumerate(source_chunks, start=1):
+        user_prompt = build_phase1_user_prompt(
+            source_chunk, session.source_language, session.target_language
         )
+        _ensure_context_capacity(alias, f"{SYSTEM_PROMPT_PHASE1}\n{user_prompt}")
+        raw_response = await _chat_complete(
+            alias,
+            SYSTEM_PROMPT_PHASE1,
+            user_prompt,
+            session_id=session_id,
+            user_id=user_id,
+            phase=1,
+            chunk_index=chunk_index,
+            total_chunks=len(source_chunks),
+            status="초벌 번역 수신 중",
+        )
+        chunk_result = await _recover_phase1_response(
+            raw_response,
+            source_chunk,
+            alias,
+            session_id,
+            user_id,
+            chunk_index,
+            len(source_chunks),
+            session.target_language,
+        )
+        chunk_results.append((source_chunk, chunk_result))
 
-    user_prompt = build_phase1_user_prompt(session.source_text, session.target_language)
-    raw_response = await chat_complete(alias, SYSTEM_PROMPT_PHASE1, user_prompt)
-
-    # Parse JSON response — extract from code blocks if wrapped
-    import json
-    cleaned = _extract_json(raw_response)
-    parsed = json.loads(cleaned)
-
-    result = Phase1Result(
-        translated=parsed.get("translated", ""),
-        proper_nouns=[ProperNoun(**pn) for pn in parsed.get("proper_nouns", [])],
-        summary=parsed.get("summary", ""),
-        style=parsed.get("style", ""),
-    )
+    result = _merge_phase1_results([item[1] for item in chunk_results])
 
     # Update session with Phase 1 results and initialize Phase 2 data
     update_data = {
         "phase1_result": result.model_dump(),
+        "phase1_chunks": [
+            TranslationChunk(
+                source_text=source_chunk,
+                translated=chunk_result.translated,
+            ).model_dump()
+            for source_chunk, chunk_result in chunk_results
+        ],
         "phase2_proper_nouns": [pn.model_dump() for pn in result.proper_nouns],
         "phase2_style": result.style,
+        "phase3_result": "",
+        "phase3_chunks": [],
+        "phase4_result": "",
+        "progress": None,
     }
-    updated = session_store.update(session_id, update_data)
+    updated = _update_phase_result(session_id, update_data, user_id, version)
     if updated is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
 
@@ -179,17 +260,25 @@ async def run_phase1(session_id: str):
 
 
 @app.post("/api/translate/{session_id}/phase2")
-async def run_phase2(session_id: str, req: Phase2ConfirmRequest):
+async def run_phase2(
+    session_id: str,
+    req: Phase2ConfirmRequest,
+    user_id: str = Depends(get_current_user),
+):
     """Phase 2: user confirms or edits proper nouns and style."""
-    session = session_store.get(session_id)
+    session = session_store.get(session_id, user_id)
     if session is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
 
     update_data = {
         "phase2_proper_nouns": [pn.model_dump() for pn in req.proper_nouns],
         "phase2_style": req.style,
+        "phase3_result": "",
+        "phase3_chunks": [],
+        "phase4_result": "",
+        "progress": None,
     }
-    updated = session_store.update(session_id, update_data)
+    updated = session_store.update(session_id, update_data, user_id)
     if updated is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
 
@@ -197,27 +286,55 @@ async def run_phase2(session_id: str, req: Phase2ConfirmRequest):
 
 
 @app.post("/api/translate/{session_id}/phase3")
-async def run_phase3(session_id: str):
+async def run_phase3(session_id: str, user_id: str = Depends(get_current_user)):
     """Run Phase 3: re-translation with proper noun and style constraints."""
-    session = session_store.get(session_id)
-    if session is None:
+    snapshot = session_store.get_with_version(session_id, user_id)
+    if snapshot is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
+    session, version = snapshot
 
     if not session.phase1_result:
         raise HTTPException(status_code=400, detail="먼저 Phase 1을 실행하세요")
 
     alias = session.model_phase3 or session.model_phase1
-    user_prompt = build_phase3_user_prompt(
-        phase1_translated=session.phase1_result.translated,
-        proper_nouns=[pn.model_dump() for pn in session.phase2_proper_nouns],
-        style=session.phase2_style or (session.phase1_result.style if session.phase1_result else ""),
-        target_language=session.target_language,
-    )
+    phase1_chunks = session.phase1_chunks or [
+        TranslationChunk(
+            source_text=session.source_text,
+            translated=session.phase1_result.translated,
+        )
+    ]
+    result_chunks: list[str] = []
+    for chunk_index, chunk in enumerate(phase1_chunks, start=1):
+        user_prompt = build_phase3_user_prompt(
+            source_text=chunk.source_text,
+            phase1_translated=chunk.translated,
+            proper_nouns=[pn.model_dump() for pn in session.phase2_proper_nouns],
+            style=session.phase2_style or session.phase1_result.style,
+            target_language=session.target_language,
+        )
+        _ensure_context_capacity(alias, f"{SYSTEM_PROMPT_PHASE3}\n{user_prompt}")
+        chunk_result = await _chat_complete(
+            alias,
+            SYSTEM_PROMPT_PHASE3,
+            user_prompt,
+            session_id=session_id,
+            user_id=user_id,
+            phase=3,
+            chunk_index=chunk_index,
+            total_chunks=len(phase1_chunks),
+            status="재번역 수신 중",
+        )
+        if not chunk_result:
+            raise HTTPException(status_code=502, detail="LLM이 재번역 결과를 반환하지 않았습니다")
+        result_chunks.append(chunk_result)
 
-    result_text = await chat_complete(alias, SYSTEM_PROMPT_PHASE3, user_prompt)
-
-    update_data = {"phase3_result": result_text}
-    updated = session_store.update(session_id, update_data)
+    update_data = {
+        "phase3_result": "\n\n".join(result_chunks),
+        "phase3_chunks": result_chunks,
+        "phase4_result": "",
+        "progress": None,
+    }
+    updated = _update_phase_result(session_id, update_data, user_id, version)
     if updated is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
 
@@ -225,29 +342,50 @@ async def run_phase3(session_id: str):
 
 
 @app.post("/api/translate/{session_id}/phase4")
-async def run_phase4(session_id: str):
+async def run_phase4(session_id: str, user_id: str = Depends(get_current_user)):
     """Run Phase 4: polish for readability."""
-    session = session_store.get(session_id)
-    if session is None:
+    snapshot = session_store.get_with_version(session_id, user_id)
+    if snapshot is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
+    session, version = snapshot
 
     if not session.phase3_result:
         raise HTTPException(status_code=400, detail="먼저 Phase 3을 실행하세요")
 
     alias = session.model_phase4 or session.model_phase3 or session.model_phase1
-
-    user_prompt = build_phase4_user_prompt(
-        phase3_result=session.phase3_result,
-        proper_nouns=[pn.model_dump() for pn in session.phase2_proper_nouns],
-        style=session.phase2_style or (session.phase1_result.style if session.phase1_result else ""),
-    )
-
-    # Patch target_language into system prompt
+    source_chunks = session.phase3_chunks or _split_text(session.phase3_result)
+    result_chunks: list[str] = []
     sys_prompt = SYSTEM_PROMPT_PHASE4.replace("{target_language}", session.target_language)
-    result_text = await chat_complete(alias, sys_prompt, user_prompt)
+    for chunk_index, source_chunk in enumerate(source_chunks, start=1):
+        user_prompt = build_phase4_user_prompt(
+            phase3_result=source_chunk,
+            proper_nouns=[pn.model_dump() for pn in session.phase2_proper_nouns],
+            style=session.phase2_style or (
+                session.phase1_result.style if session.phase1_result else ""
+            ),
+            target_language=session.target_language,
+        )
+        _ensure_context_capacity(alias, f"{sys_prompt}\n{user_prompt}")
+        chunk_result = await _chat_complete(
+            alias,
+            sys_prompt,
+            user_prompt,
+            session_id=session_id,
+            user_id=user_id,
+            phase=4,
+            chunk_index=chunk_index,
+            total_chunks=len(source_chunks),
+            status="마무리 번역 수신 중",
+        )
+        if not chunk_result:
+            raise HTTPException(status_code=502, detail="LLM이 마무리 결과를 반환하지 않았습니다")
+        result_chunks.append(chunk_result)
 
-    update_data = {"phase4_result": result_text}
-    updated = session_store.update(session_id, update_data)
+    update_data = {
+        "phase4_result": "\n\n".join(result_chunks),
+        "progress": None,
+    }
+    updated = _update_phase_result(session_id, update_data, user_id, version)
     if updated is None:
         raise HTTPException(status_code=404, detail="세션을 찾을 수 없습니다")
 
@@ -256,19 +394,276 @@ async def run_phase4(session_id: str):
 
 # ── Helpers ───────────────────────────────────────────
 
+def _split_text(text: str) -> list[str]:
+    """Split long text near sentence or line boundaries for reliable LLM calls."""
+    remaining = text.strip()
+    chunks: list[str] = []
+    boundary_chars = "\n。！？!?；;."
+
+    while len(remaining) > TRANSLATION_CHUNK_CHARS:
+        window = remaining[:TRANSLATION_CHUNK_CHARS]
+        minimum_boundary = TRANSLATION_CHUNK_CHARS // 2
+        split_at = max(window.rfind(char) for char in boundary_chars)
+        if split_at < minimum_boundary:
+            whitespace_at = max(window.rfind(" "), window.rfind("\t"))
+            split_at = (
+                whitespace_at + 1
+                if whitespace_at >= minimum_boundary
+                else TRANSLATION_CHUNK_CHARS
+            )
+        else:
+            split_at += 1
+        chunks.append(remaining[:split_at].strip())
+        remaining = remaining[split_at:].lstrip()
+
+    if remaining:
+        chunks.append(remaining)
+    return chunks
+
+
+def _parse_phase1_response(raw_response: str) -> Phase1Result | None:
+    extracted = _extract_json(raw_response)
+    candidates = [extracted, re.sub(r",\s*([}\]])", r"\1", extracted)]
+    for candidate in candidates:
+        try:
+            parsed = json.loads(candidate)
+            if not isinstance(parsed, dict):
+                continue
+            translated = parsed.get("translated") or parsed.get("translation") or ""
+            proper_nouns = []
+            for item in parsed.get("proper_nouns") or []:
+                if not isinstance(item, dict) or not item.get("original"):
+                    continue
+                proper_nouns.append(
+                    ProperNoun(
+                        original=str(item["original"]),
+                        suggested=str(
+                            item.get("suggested") or item.get("final") or item["original"]
+                        ),
+                        final=str(item.get("final") or ""),
+                    )
+                )
+            result = Phase1Result(
+                translated=str(translated),
+                proper_nouns=proper_nouns,
+                summary=str(parsed.get("summary") or ""),
+                style=str(parsed.get("style") or ""),
+            )
+            if result.translated.strip():
+                return result
+        except (json.JSONDecodeError, TypeError, ValueError, ValidationError):
+            continue
+    return None
+
+
+async def _recover_phase1_response(
+    raw_response: str,
+    source_text: str,
+    alias: str,
+    session_id: str,
+    user_id: str,
+    chunk_index: int,
+    total_chunks: int,
+    target_language: str,
+) -> Phase1Result:
+    parsed = _parse_phase1_response(raw_response)
+    if parsed is not None:
+        return parsed
+
+    repair_system = """\
+당신은 손상된 JSON을 복구하는 도구입니다. 내용은 번역하거나 요약하지 말고, 주어진 응답의 내용을 보존하여 순수 JSON 객체만 반환하세요.\
+"""
+    repair_prompt = f"""\
+다음 응답을 translated, proper_nouns, summary, style 필드를 가진 유효한 JSON으로 복구하세요.
+proper_nouns의 각 항목은 original과 suggested 필드를 가져야 합니다.
+
+손상된 응답:
+---
+{raw_response}
+---"""
+    repaired_response = await _chat_complete(
+        alias,
+        repair_system,
+        repair_prompt,
+        temperature=0.0,
+        session_id=session_id,
+        user_id=user_id,
+        phase=1,
+        chunk_index=chunk_index,
+        total_chunks=total_chunks,
+        status="JSON 응답 복구 중",
+    )
+    parsed = _parse_phase1_response(repaired_response)
+    if parsed is not None:
+        parsed.warnings.append("LLM의 JSON 응답을 자동으로 복구했습니다.")
+        return parsed
+
+    fallback_prompt = f"""\
+다음 원문을 "{target_language}"으로 정확하게 번역하세요.
+설명이나 JSON 없이 번역문만 반환하세요.
+
+원문:
+---
+{source_text}
+---"""
+    translated = await _chat_complete(
+        alias,
+        SYSTEM_PROMPT_PHASE1,
+        fallback_prompt,
+        session_id=session_id,
+        user_id=user_id,
+        phase=1,
+        chunk_index=chunk_index,
+        total_chunks=total_chunks,
+        status="일반 번역으로 전환 중",
+    )
+    if not translated:
+        raise HTTPException(status_code=502, detail="LLM이 번역 결과를 반환하지 않았습니다")
+    return Phase1Result(
+        translated=translated,
+        proper_nouns=[],
+        summary="",
+        style="",
+        warnings=[
+            "JSON 응답을 복구하지 못해 일반 번역으로 대체했습니다. "
+            "이 조각의 고유명사, 요약, 문체 정보는 비어 있습니다."
+        ],
+    )
+
+
+def _merge_phase1_results(results: list[Phase1Result]) -> Phase1Result:
+    proper_nouns: list[ProperNoun] = []
+    seen_originals: set[str] = set()
+    for result in results:
+        for proper_noun in result.proper_nouns:
+            key = proper_noun.original.strip()
+            if key and key not in seen_originals:
+                proper_nouns.append(proper_noun)
+                seen_originals.add(key)
+
+    styles = [result.style for result in results if result.style]
+    style = Counter(styles).most_common(1)[0][0] if styles else ""
+    return Phase1Result(
+        translated="\n\n".join(result.translated for result in results),
+        proper_nouns=proper_nouns,
+        summary="\n".join(result.summary for result in results if result.summary),
+        style=style,
+        warnings=[warning for result in results for warning in result.warnings],
+    )
+
+
+def _validate_model_aliases(data: dict) -> None:
+    available = {config.alias for config in get_llm_configs()}
+    invalid = {
+        value
+        for key, value in data.items()
+        if key.startswith("model_phase") and value and value not in available
+    }
+    if invalid:
+        raise HTTPException(
+            status_code=400,
+            detail=f"알 수 없는 LLM 모델입니다: {', '.join(sorted(invalid))}",
+        )
+
+
+async def _chat_complete(
+    alias: str,
+    system_prompt: str,
+    user_prompt: str,
+    temperature: float = 0.3,
+    *,
+    session_id: str | None = None,
+    user_id: str | None = None,
+    phase: int = 0,
+    chunk_index: int = 1,
+    total_chunks: int = 1,
+    status: str = "응답 수신 중",
+) -> str:
+    parts: list[str] = []
+    last_progress_update = 0.0
+
+    def set_progress(progress_status: str, preview: str = "") -> None:
+        if session_id and user_id:
+            session_store.set_progress(
+                session_id,
+                {
+                    "phase": phase,
+                    "chunk": chunk_index,
+                    "total_chunks": total_chunks,
+                    "status": progress_status,
+                    "preview": preview[-3000:],
+                },
+                user_id,
+            )
+
+    async def on_delta(delta: str) -> None:
+        nonlocal last_progress_update
+        parts.append(delta)
+        now = monotonic()
+        if now - last_progress_update >= 0.25:
+            set_progress(status, "".join(parts))
+            last_progress_update = now
+
+    set_progress("LLM 응답 대기 중")
+    try:
+        result = await chat_complete(
+            alias,
+            system_prompt,
+            user_prompt,
+            temperature=temperature,
+            on_delta=on_delta if session_id and user_id else None,
+        )
+        set_progress(f"{chunk_index}/{total_chunks} 조각 수신 완료", result)
+        return result
+    except APITimeoutError as exc:
+        set_progress("응답 시간 초과", "".join(parts))
+        raise HTTPException(status_code=504, detail="LLM 서버 응답 시간이 초과되었습니다") from exc
+    except ValueError as exc:
+        set_progress("요청 오류", "".join(parts))
+        raise HTTPException(status_code=400, detail=str(exc)) from exc
+    except OpenAIError as exc:
+        set_progress("LLM 서버 요청 실패", "".join(parts))
+        raise HTTPException(status_code=502, detail="LLM 서버 요청에 실패했습니다") from exc
+
+
+def _update_phase_result(
+    session_id: str, data: dict, user_id: str, expected_version: int
+) -> TranslationSession | None:
+    try:
+        return session_store.update(
+            session_id, data, user_id, expected_version=expected_version
+        )
+    except SessionConflictError as exc:
+        raise HTTPException(
+            status_code=409,
+            detail="처리 중 세션이 변경되었습니다. 현재 상태에서 단계를 다시 실행하세요",
+        ) from exc
+
+
+def _ensure_context_capacity(alias: str, content: str) -> None:
+    context_size = get_context_size(alias)
+    if context_size and len(content) > context_size * 0.7:
+        raise HTTPException(
+            status_code=400,
+            detail=f"입력이 너무 깁니다 ({context_size} 컨텍스트의 70% 이내로 줄여주세요)",
+        )
+
+
 def _extract_json(text: str) -> str:
-    """Extract JSON from LLM response, handling markdown code blocks."""
+    """Extract a likely JSON object from code fences or surrounding prose."""
     text = text.strip()
-    # Remove markdown json/code block wrapper
     if "```" in text:
         parts = text.split("```")
         for part in parts:
             part = part.strip()
-            if part.startswith("json"):
+            if part.lower().startswith("json"):
                 part = part[4:].strip()
-            if part.startswith("{"):
-                return part
-    return text
+            if "{" in part and "}" in part:
+                text = part
+                break
+    start = text.find("{")
+    end = text.rfind("}")
+    return text[start : end + 1] if start >= 0 and end > start else text
 
 
 # ── Static File Serving (Frontend) ───────────────────
diff --git a/backend/models.py b/backend/models.py
index bb72f82..68963d2 100644
--- a/backend/models.py
+++ b/backend/models.py
@@ -42,6 +42,20 @@ class Phase1Result(BaseModel):
     proper_nouns: list[ProperNoun]   # 추출된 고유명사 리스트
     summary: str                     # 내용 요약
     style: str                       # 문체 분석 결과
+    warnings: list[str] = []         # JSON 복구/폴백 등 품질 관련 알림
+
+
+class TranslationChunk(BaseModel):
+    source_text: str
+    translated: str
+
+
+class PhaseProgress(BaseModel):
+    phase: int
+    chunk: int
+    total_chunks: int
+    status: str
+    preview: str = ""
 
 
 class TranslationSession(BaseModel):
@@ -55,10 +69,13 @@ class TranslationSession(BaseModel):
     model_phase3: str = ""                        # Phase 3 LLM 모델 alias
     model_phase4: str = ""                        # Phase 4 LLM 모델 alias
     phase1_result: Phase1Result | None = None     # 초벌 번역 결과
+    phase1_chunks: list[TranslationChunk] = []    # 장문 분할 시 원문/초벌 번역 쌍
     phase2_proper_nouns: list[ProperNoun] = []    # Phase 2에서 최종 확정된 고유명사
     phase2_style: str = ""                        # Phase 2에서 최종 확정된 문체
     phase3_result: str = ""                       # 재번역 결과
+    phase3_chunks: list[str] = []                 # 장문 분할 재번역 결과
     phase4_result: str = ""                       # 마무리 다듬기 결과
+    progress: PhaseProgress | None = None          # 장기 작업 진행 상태
 
 
 # ── API Requests ───────────────────────────────────────
diff --git a/backend/prompts.py b/backend/prompts.py
index d52fb60..e6f320d 100644
--- a/backend/prompts.py
+++ b/backend/prompts.py
@@ -5,13 +5,21 @@
 
 SYSTEM_PROMPT_PHASE1 = """\
 당신은 전문 번역가입니다. 사용자의 지시에 따라 원문을 도착 언어로 번역하고, \
-고유명사를 추출하며, 내용 요약을 작성하고, 문체를 분석합니다.\
+고유명사를 추출하며, 내용 요약을 작성하고, 문체를 분석합니다. 번역할 때 원문의 \
+주어와 목적어, 인칭, 시제, 양태, 높임 수준을 정확히 보존합니다.\
 """
 
 
-def build_phase1_user_prompt(source_text: str, target_language: str) -> str:
+def build_phase1_user_prompt(
+    source_text: str, source_language: str, target_language: str
+) -> str:
+    source_instruction = (
+        "원문의 언어를 자동으로 판단하고"
+        if source_language == "auto"
+        else f'원문이 "{source_language}"임을 전제로'
+    )
     return f"""\
-다음 원문을 "{target_language}"으로 번역하세요.
+{source_instruction} 다음 원문을 "{target_language}"으로 번역하세요.
 
 번역 결과를 반드시 아래 JSON 포맷으로만 반환하세요 (마크다운 코드 블록 없이 순수 JSON):
 
@@ -44,13 +52,14 @@ SYSTEM_PROMPT_PHASE3 = """\
 
 
 def build_phase3_user_prompt(
+    source_text: str,
     phase1_translated: str,
     proper_nouns: list[dict],
     style: str,
     target_language: str,
 ) -> str:
     pn_lines = "\n".join(
-        f"  - {pn['original']} → {pn.get('final', pn.get('suggested', ''))}"
+        f"  - {pn['original']} → {pn.get('final') or pn.get('suggested', '')}"
         for pn in proper_nouns if pn.get("original") or pn.get("suggested")
     )
 
@@ -65,10 +74,15 @@ def build_phase3_user_prompt(
 재번역 규칙:
 1. 위 고유명사 매핑을 반드시 준수하세요
 2. 지시된 문체에 맞게 어조를 조정하세요
-3. 초벌 번역의 의미는 유지하되 더 자연스럽고 정확히 다듬으세요
+3. 원문을 기준으로 초벌 번역의 누락이나 오역을 바로잡으세요
 4. 결과만 반환하세요 (설명 없이 번역문만)
 
-초벌 번역:
+원문:
+---
+{source_text}
+---
+
+초벌 번역 참고본:
 ---
 {phase1_translated}
 ---"""
@@ -78,7 +92,8 @@ def build_phase3_user_prompt(
 
 SYSTEM_PROMPT_PHASE4 = """\
 당신은 "{target_language}" 원어민 편집자입니다. 제공된 번역문을 \
-"{target_language}" 화자에게 읽기 편하도록 다듬으세요.\
+"{target_language}" 화자에게 읽기 편하도록 다듬되, 번역문의 의미 관계와 \
+인칭, 시제, 양태, 높임 수준은 변경하지 마세요.\
 """
 
 
@@ -86,9 +101,10 @@ def build_phase4_user_prompt(
     phase3_result: str,
     proper_nouns: list[dict],
     style: str,
+    target_language: str,
 ) -> str:
     pn_lines = "\n".join(
-        f"  - {pn['original']} → {pn.get('final', pn.get('suggested', ''))}"
+        f"  - {pn['original']} → {pn.get('final') or pn.get('suggested', '')}"
         for pn in proper_nouns if pn.get("original") or pn.get("suggested")
     )
 
@@ -102,10 +118,12 @@ def build_phase4_user_prompt(
 
 다듬기 규칙:
 1. 고유명사는 절대 변경하지 마세요
-2. 한국어 화자에게 자연스럽게 읽히도록 어순과 표현을 조정하세요
+2. {target_language} 화자에게 자연스럽게 읽히도록 어순과 표현을 조정하세요
 3. 문장은 간결하고 명확하게 다듬으세요
 4. 의미는 반드시 유지하세요
-5. 결과만 반환하세요 (설명 없이 번역문만)
+5. 주어와 목적어, 화자와 청자의 관계를 뒤바꾸지 마세요
+6. 인칭, 시제, 희망·추측 등의 양태와 높임 수준을 변경하지 마세요
+7. 결과만 반환하세요 (설명 없이 번역문만)
 
 번역문:
 ---
diff --git a/backend/sessions.py b/backend/sessions.py
index 5714034..932e9ca 100644
--- a/backend/sessions.py
+++ b/backend/sessions.py
@@ -7,8 +7,12 @@ from typing import Any
 from backend.models import TranslationSession
 
 
+class SessionConflictError(Exception):
+    """Raised when a long-running phase tries to update a changed session."""
+
+
 class SessionStore:
-    """Thread-safe-ish in-memory dict for translation sessions.
+    """Process-local in-memory store for user-owned translation sessions.
 
     Each session is keyed by a UUID and has a 24-hour TTL.
     Expired sessions are lazily cleaned on access.
@@ -18,18 +22,26 @@ class SessionStore:
         self._store: dict[str, dict[str, Any]] = {}
         self.ttl_hours = ttl_hours
 
-    def create(self, session_data: TranslationSession) -> str:
+    def create(self, session_data: TranslationSession, owner_id: str) -> str:
         session_id = uuid.uuid4().hex[:16]
         self._store[session_id] = {
             "data": session_data.model_dump(),
+            "owner_id": owner_id,
+            "version": 0,
             "created_at": datetime.now(timezone.utc),
             "updated_at": datetime.now(timezone.utc),
         }
         return session_id
 
-    def get(self, session_id: str) -> TranslationSession | None:
+    def get(self, session_id: str, owner_id: str) -> TranslationSession | None:
+        result = self.get_with_version(session_id, owner_id)
+        return result[0] if result else None
+
+    def get_with_version(
+        self, session_id: str, owner_id: str
+    ) -> tuple[TranslationSession, int] | None:
         entry = self._store.get(session_id)
-        if entry is None:
+        if entry is None or entry["owner_id"] != owner_id:
             return None
         # Check TTL
         now = datetime.now(timezone.utc)
@@ -39,22 +51,51 @@ class SessionStore:
             return None
         # Touch (refresh TTL)
         entry["updated_at"] = now
-        return TranslationSession(**entry["data"])
+        return TranslationSession(**entry["data"]), entry["version"]
 
-    def update(self, session_id: str, data: dict[str, Any]) -> TranslationSession | None:
+    def update(
+        self,
+        session_id: str,
+        data: dict[str, Any],
+        owner_id: str,
+        expected_version: int | None = None,
+    ) -> TranslationSession | None:
         entry = self._store.get(session_id)
-        if entry is None:
+        if entry is None or entry["owner_id"] != owner_id:
             return None
+        now = datetime.now(timezone.utc)
+        if now - entry["updated_at"] > timedelta(hours=self.ttl_hours):
+            del self._store[session_id]
+            return None
+        if expected_version is not None and entry["version"] != expected_version:
+            raise SessionConflictError
         entry["data"].update(data)
-        entry["updated_at"] = datetime.now(timezone.utc)
+        entry["version"] += 1
+        entry["updated_at"] = now
         return TranslationSession(**entry["data"])
 
-    def delete(self, session_id: str) -> bool:
-        if session_id in self._store:
+    def delete(self, session_id: str, owner_id: str) -> bool:
+        entry = self._store.get(session_id)
+        if entry is not None and entry["owner_id"] == owner_id:
             del self._store[session_id]
             return True
         return False
 
+    def set_progress(
+        self, session_id: str, progress: dict[str, Any] | None, owner_id: str
+    ) -> bool:
+        """Update transient progress without changing the semantic session version."""
+        entry = self._store.get(session_id)
+        if entry is None or entry["owner_id"] != owner_id:
+            return False
+        now = datetime.now(timezone.utc)
+        if now - entry["updated_at"] > timedelta(hours=self.ttl_hours):
+            del self._store[session_id]
+            return False
+        entry["data"]["progress"] = progress
+        entry["updated_at"] = now
+        return True
+
     def cleanup_expired(self):
         """Remove all expired sessions."""
         now = datetime.now(timezone.utc)
diff --git a/frontend/src/api.ts b/frontend/src/api.ts
index 0632026..6e8255d 100644
--- a/frontend/src/api.ts
+++ b/frontend/src/api.ts
@@ -1,7 +1,9 @@
 /** Axios-based API client with JWT token management */
 
 import axios from 'axios'
-import type { LLMModel, SessionResponse } from './types'
+import type { LLMModel, PhaseProgress, ProperNoun, SessionConfig, SessionResponse } from './types'
+
+export const AUTH_UNAUTHORIZED_EVENT = 'llm-translator:unauthorized'
 
 const api = axios.create({
   baseURL: '/api',
@@ -17,13 +19,13 @@ api.interceptors.request.use((config) => {
   return config
 })
 
-// Auto-redirect on 401
+// Notify the auth store so App.vue can switch views without a router.
 api.interceptors.response.use(
   (res) => res,
   (err) => {
     if (err.response?.status === 401) {
       localStorage.removeItem('jwt_token')
-      window.location.hash = '#/login'
+      window.dispatchEvent(new Event(AUTH_UNAUTHORIZED_EVENT))
     }
     return Promise.reject(err)
   },
@@ -56,15 +58,7 @@ export async function listModels(): Promise<LLMModel[]> {
 
 // ── Sessions ─────────────────────────────────────────
 
-export async function createSession(params: {
-  source_text: string
-  source_language: string
-  target_language: string
-  model_phase1: string
-  model_phase2: string
-  model_phase3: string
-  model_phase4: string
-}): Promise<{ session_id: string }> {
+export async function createSession(params: SessionConfig): Promise<{ session_id: string }> {
   const res = await api.post('/translate/create', params)
   return res.data
 }
@@ -74,13 +68,20 @@ export async function getSession(sessionId: string): Promise<SessionResponse> {
   return res.data
 }
 
+export async function getSessionProgress(
+  sessionId: string,
+): Promise<{ progress: PhaseProgress | null }> {
+  const res = await api.get(`/sessions/${sessionId}/progress`)
+  return res.data
+}
+
 export async function deleteSession(sessionId: string): Promise<void> {
   await api.delete(`/sessions/${sessionId}`)
 }
 
 export async function updateSession(
   sessionId: string,
-  patch: Record<string, any>,
+  patch: Partial<SessionConfig>,
 ): Promise<SessionResponse> {
   const res = await api.patch(`/sessions/${sessionId}`, patch)
   return res.data
@@ -95,7 +96,7 @@ export async function runPhase1(sessionId: string): Promise<SessionResponse> {
 
 export async function runPhase2(
   sessionId: string,
-  proper_nouns: any[],
+  proper_nouns: ProperNoun[],
   style: string,
 ): Promise<SessionResponse> {
   const res = await api.post(`/translate/${sessionId}/phase2`, { proper_nouns, style })
diff --git a/frontend/src/stores/auth.ts b/frontend/src/stores/auth.ts
index ace1220..7869299 100644
--- a/frontend/src/stores/auth.ts
+++ b/frontend/src/stores/auth.ts
@@ -1,6 +1,6 @@
 import { defineStore } from 'pinia'
-import { ref, computed } from 'vue'
-import { me as apiMe, logout as apiLogout } from '../api'
+import { computed, onScopeDispose, ref } from 'vue'
+import { AUTH_UNAUTHORIZED_EVENT, me as apiMe, logout as apiLogout } from '../api'
 
 export const useAuthStore = defineStore('auth', () => {
   const user_id = ref<string | null>(null)
@@ -8,6 +8,14 @@ export const useAuthStore = defineStore('auth', () => {
 
   const isLoggedIn = computed(() => !!user_id.value && loaded.value)
 
+  function handleUnauthorized() {
+    user_id.value = null
+    loaded.value = true
+  }
+
+  window.addEventListener(AUTH_UNAUTHORIZED_EVENT, handleUnauthorized)
+  onScopeDispose(() => window.removeEventListener(AUTH_UNAUTHORIZED_EVENT, handleUnauthorized))
+
   async function check() {
     if (loaded.value) return
     try {
@@ -23,7 +31,6 @@ export const useAuthStore = defineStore('auth', () => {
   async function logout() {
     await apiLogout()
     user_id.value = null
-    window.location.hash = '#/login'
   }
 
   return { user_id, loaded, isLoggedIn, check, logout }
diff --git a/frontend/src/stores/translation.ts b/frontend/src/stores/translation.ts
index 046bf1c..9c77c49 100644
--- a/frontend/src/stores/translation.ts
+++ b/frontend/src/stores/translation.ts
@@ -1,24 +1,32 @@
 import { defineStore } from 'pinia'
-import { ref, computed } from 'vue'
-import type { ProperNoun, TranslationSessionData, HistoryEntry, LLMModel } from '../types'
+import { computed, ref, watch } from 'vue'
+import { isAxiosError } from 'axios'
+import type { HistoryEntry, LLMModel, Phase1Result, PhaseProgress, ProperNoun, SessionConfig, TranslationSessionData } from '../types'
 import * as api from '../api'
+import { useAuthStore } from './auth'
 
 const HISTORY_KEY = 'llm_translator_history'
-const SESSION_KEY_PREFIX = 'llm_translator_session_'
+const MODELS_KEY = 'llm_translator_models'
+
+interface SavedModels {
+  p1?: string
+  p3?: string
+  p4?: string
+}
 
 export const useTranslationStore = defineStore('translation', () => {
+  const auth = useAuthStore()
   // ── State ────────────────────────────────────────
   const sessionId = ref<string | null>(null)
   const sourceText = ref('')
   const sourceLanguage = ref('auto')
   const targetLanguage = ref('한국어')
   const modelPhase1 = ref('')
-  const modelPhase2 = ref('')
   const modelPhase3 = ref('')
   const modelPhase4 = ref('')
 
   // Phase results
-  const phase1Result = ref<{ translated: string; proper_nouns: ProperNoun[]; summary: string; style: string } | null>(null)
+  const phase1Result = ref<Phase1Result | null>(null)
   const phase2ProperNouns = ref<ProperNoun[]>([])
   const phase2Style = ref('')
   const phase3Result = ref('')
@@ -30,22 +38,51 @@ export const useTranslationStore = defineStore('translation', () => {
   const error = ref<string | null>(null)
   const models = ref<LLMModel[]>([])
   const history = ref<HistoryEntry[]>([])
+  const historySnapshot = ref(false)
+  const progress = ref<PhaseProgress | null>(null)
+  let progressTimer: ReturnType<typeof setInterval> | null = null
+  let progressRequestPending = false
 
   // ── Computed ─────────────────────────────────────
   const phase1Done = computed(() => !!phase1Result.value)
-  const phase2Ready = computed(() => phase1Done.value)   // can edit after P1
-  const phase3Ready = computed(() => phase1Done.value && (phase2ProperNouns.value.length > 0 || phase2Style.value))
-  const phase4Ready = computed(() => !!phase3Result.value)
-
-  const activeAbortController = ref<AbortController | null>(null)
+  const phase2Ready = computed(() => phase1Done.value && !historySnapshot.value)
+  const phase3Ready = computed(() => phase1Done.value && !historySnapshot.value)
+  const phase4Ready = computed(() => !!phase3Result.value && !historySnapshot.value)
 
   // ── Actions ──────────────────────────────────────
 
+  function getErrorMessage(error: unknown, fallback: string): string {
+    if (isAxiosError<{ detail?: string }>(error)) {
+      return error.response?.data?.detail || error.message || fallback
+    }
+    return error instanceof Error ? error.message : fallback
+  }
+
+  function loadSavedModels(): SavedModels {
+    try {
+      const parsed: unknown = JSON.parse(localStorage.getItem(MODELS_KEY) || '{}')
+      if (!parsed || typeof parsed !== 'object' || Array.isArray(parsed)) return {}
+      const value = parsed as Record<string, unknown>
+      return {
+        p1: typeof value.p1 === 'string' ? value.p1 : undefined,
+        p3: typeof value.p3 === 'string' ? value.p3 : undefined,
+        p4: typeof value.p4 === 'string' ? value.p4 : undefined,
+      }
+    } catch {
+      return {}
+    }
+  }
+
+  function validModel(alias: string | undefined, fallback: string): string {
+    if (models.value.length === 0) return alias || fallback
+    return alias && models.value.some(model => model.alias === alias) ? alias : fallback
+  }
+
   async function loadModels() {
     try {
       models.value = await api.listModels()
-    } catch (e: any) {
-      console.error('Failed to load models:', e)
+    } catch (error: unknown) {
+      console.error('Failed to load models:', error)
     }
   }
 
@@ -60,24 +97,47 @@ export const useTranslationStore = defineStore('translation', () => {
     phase3Result.value = ''
     phase4Result.value = ''
     error.value = null
+    historySnapshot.value = false
 
-    // Restore model selections from localStorage
-    const savedModels = JSON.parse(localStorage.getItem('llm_translator_models') || '{}')
-    modelPhase1.value = savedModels.p1 || models.value[0]?.alias || ''
-    modelPhase2.value = savedModels.p2 || models.value[0]?.alias || ''
-    modelPhase3.value = savedModels.p3 || models.value[0]?.alias || ''
-    modelPhase4.value = savedModels.p4 || (models.value[1]?.alias || models.value[0]?.alias || '')
+    const savedModels = loadSavedModels()
+    const defaultModel = models.value[0]?.alias || ''
+    modelPhase1.value = validModel(savedModels.p1, defaultModel)
+    modelPhase3.value = validModel(savedModels.p3, defaultModel)
+    modelPhase4.value = validModel(savedModels.p4, models.value[1]?.alias || defaultModel)
   }
 
   function saveModelSelections() {
-    localStorage.setItem('llm_translator_models', JSON.stringify({
-      p1: modelPhase1.value,
-      p2: modelPhase2.value,
-      p3: modelPhase3.value,
-      p4: modelPhase4.value,
-    }))
+    try {
+      localStorage.setItem(MODELS_KEY, JSON.stringify({
+        p1: modelPhase1.value,
+        p3: modelPhase3.value,
+        p4: modelPhase4.value,
+      }))
+    } catch (error) {
+      console.error('Failed to save model selections:', error)
+    }
   }
 
+  watch([modelPhase1, modelPhase3, modelPhase4], saveModelSelections)
+  watch([sourceText, sourceLanguage, targetLanguage, modelPhase1], () => {
+    phase1Result.value = null
+    phase2ProperNouns.value = []
+    phase2Style.value = ''
+    phase3Result.value = ''
+    phase4Result.value = ''
+  }, { flush: 'sync' })
+  watch(modelPhase3, () => {
+    phase3Result.value = ''
+    phase4Result.value = ''
+  }, { flush: 'sync' })
+  watch(modelPhase4, () => {
+    phase4Result.value = ''
+  }, { flush: 'sync' })
+  watch([phase2ProperNouns, phase2Style], () => {
+    phase3Result.value = ''
+    phase4Result.value = ''
+  }, { deep: true, flush: 'sync' })
+
   async function ensureSession() {
     if (!sessionId.value) {
       const res = await api.createSession(getSessionParams())
@@ -85,13 +145,13 @@ export const useTranslationStore = defineStore('translation', () => {
     }
   }
 
-  function getSessionParams() {
+  function getSessionParams(): SessionConfig {
     return {
       source_text: sourceText.value,
       source_language: sourceLanguage.value,
       target_language: targetLanguage.value,
       model_phase1: modelPhase1.value,
-      model_phase2: modelPhase2.value,
+      model_phase2: '',
       model_phase3: modelPhase3.value,
       model_phase4: modelPhase4.value,
     }
@@ -99,10 +159,44 @@ export const useTranslationStore = defineStore('translation', () => {
 
   async function syncToServer() {
     if (!sessionId.value) return
+    await api.updateSession(sessionId.value, getSessionParams())
+  }
+
+  async function submitPhase2() {
+    if (!sessionId.value) throw new Error('세션 생성 실패')
+    const res = await api.runPhase2(
+      sessionId.value,
+      phase2ProperNouns.value.map(pn => ({ ...pn, final: pn.final || pn.suggested })),
+      phase2Style.value,
+    )
+    phase2ProperNouns.value = res.data.phase2_proper_nouns || []
+    phase2Style.value = res.data.phase2_style || ''
+  }
+
+  async function pollProgress() {
+    if (!sessionId.value || progressRequestPending) return
+    progressRequestPending = true
     try {
-      await api.updateSession(sessionId.value, getSessionParams())
-    } catch (e: any) {
-      console.error('Sync failed:', e)
+      const response = await api.getSessionProgress(sessionId.value)
+      progress.value = response.progress
+    } catch {
+      // The phase request handles user-visible errors; polling is best-effort only.
+    } finally {
+      progressRequestPending = false
+    }
+  }
+
+  function startProgressPolling() {
+    stopProgressPolling()
+    progress.value = null
+    void pollProgress()
+    progressTimer = setInterval(() => void pollProgress(), 750)
+  }
+
+  function stopProgressPolling() {
+    if (progressTimer !== null) {
+      clearInterval(progressTimer)
+      progressTimer = null
     }
   }
 
@@ -121,21 +215,25 @@ export const useTranslationStore = defineStore('translation', () => {
     loading.value = true
     currentPhaseLoading.value = 1
     error.value = null
-    activeAbortController.value = new AbortController()
-
+    phase3Result.value = ''
+    phase4Result.value = ''
     try {
       await ensureSession()
       if (!sessionId.value) throw new Error('세션 생성 실패')
+      await syncToServer()
+      startProgressPolling()
       const res = await api.runPhase1(sessionId.value)
       phase1Result.value = res.data.phase1_result
       phase2ProperNouns.value = res.data.phase2_proper_nouns || []
       phase2Style.value = res.data.phase2_style || ''
-    } catch (e: any) {
-      error.value = e.response?.data?.detail || e.message || 'Phase 1 실행 실패'
+      historySnapshot.value = false
+      saveToHistory()
+    } catch (caughtError: unknown) {
+      error.value = getErrorMessage(caughtError, 'Phase 1 실행 실패')
     } finally {
+      stopProgressPolling()
       loading.value = false
       currentPhaseLoading.value = null
-      activeAbortController.value = null
     }
   }
 
@@ -146,18 +244,14 @@ export const useTranslationStore = defineStore('translation', () => {
     loading.value = true
     currentPhaseLoading.value = 2
     error.value = null
+    phase3Result.value = ''
+    phase4Result.value = ''
 
     try {
-      const res = await api.runPhase2(
-        sessionId.value,
-        phase2ProperNouns.value.map(pn => ({ ...pn, final: pn.final || pn.suggested })),
-        phase2Style.value,
-      )
-      // Update local state from server response
-      phase2ProperNouns.value = res.data.phase2_proper_nouns || []
-      phase2Style.value = res.data.phase2_style || ''
-    } catch (e: any) {
-      error.value = e.response?.data?.detail || e.message || 'Phase 2 확인 실패'
+      await submitPhase2()
+      saveToHistory()
+    } catch (caughtError: unknown) {
+      error.value = getErrorMessage(caughtError, 'Phase 2 확인 실패')
     } finally {
       loading.value = false
       currentPhaseLoading.value = null
@@ -177,19 +271,25 @@ export const useTranslationStore = defineStore('translation', () => {
     loading.value = true
     currentPhaseLoading.value = 3
     error.value = null
-    activeAbortController.value = new AbortController()
+    phase3Result.value = ''
+    phase4Result.value = ''
 
     try {
       await ensureSession()
       if (!sessionId.value) throw new Error('세션 생성 실패')
+      await syncToServer()
+      await submitPhase2()
+      saveToHistory()
+      startProgressPolling()
       const res = await api.runPhase3(sessionId.value)
       phase3Result.value = res.data.phase3_result || ''
-    } catch (e: any) {
-      error.value = e.response?.data?.detail || e.message || 'Phase 3 실행 실패'
+      saveToHistory()
+    } catch (caughtError: unknown) {
+      error.value = getErrorMessage(caughtError, 'Phase 3 실행 실패')
     } finally {
+      stopProgressPolling()
       loading.value = false
       currentPhaseLoading.value = null
-      activeAbortController.value = null
     }
   }
 
@@ -202,19 +302,21 @@ export const useTranslationStore = defineStore('translation', () => {
     loading.value = true
     currentPhaseLoading.value = 4
     error.value = null
-    activeAbortController.value = new AbortController()
 
     try {
       await ensureSession()
       if (!sessionId.value) throw new Error('세션 생성 실패')
+      await syncToServer()
+      startProgressPolling()
       const res = await api.runPhase4(sessionId.value)
       phase4Result.value = res.data.phase4_result || ''
-    } catch (e: any) {
-      error.value = e.response?.data?.detail || e.message || 'Phase 4 실행 실패'
+      saveToHistory()
+    } catch (caughtError: unknown) {
+      error.value = getErrorMessage(caughtError, 'Phase 4 실행 실패')
     } finally {
+      stopProgressPolling()
       loading.value = false
       currentPhaseLoading.value = null
-      activeAbortController.value = null
     }
   }
 
@@ -239,32 +341,43 @@ export const useTranslationStore = defineStore('translation', () => {
     // Keep max 50 entries
     if (history.value.length > 50) history.value.pop()
 
-    localStorage.setItem(HISTORY_KEY, JSON.stringify(history.value))
+    try {
+      localStorage.setItem(historyStorageKey(), JSON.stringify(history.value))
+    } catch (error) {
+      console.error('Failed to save translation history:', error)
+    }
+  }
+
+  function historyStorageKey(): string {
+    return `${HISTORY_KEY}:${auth.user_id || 'anonymous'}`
   }
 
   function loadHistory() {
     try {
-      const raw = localStorage.getItem(HISTORY_KEY) || '[]'
+      const key = historyStorageKey()
+      const raw = localStorage.getItem(key) || '[]'
       history.value = JSON.parse(raw) as HistoryEntry[]
+      localStorage.removeItem(HISTORY_KEY)
     } catch {
       history.value = []
     }
   }
 
-  async function restoreFromHistory(entry: HistoryEntry) {
-    sessionId.value = entry.id
+  function restoreFromHistory(entry: HistoryEntry) {
+    sessionId.value = null
     sourceText.value = entry.data.source_text
     sourceLanguage.value = entry.data.source_language
     targetLanguage.value = entry.data.target_language
-    modelPhase1.value = entry.data.model_phase1
-    modelPhase2.value = entry.data.model_phase2
-    modelPhase3.value = entry.data.model_phase3
-    modelPhase4.value = entry.data.model_phase4
+    const defaultModel = models.value[0]?.alias || ''
+    modelPhase1.value = validModel(entry.data.model_phase1, defaultModel)
+    modelPhase3.value = validModel(entry.data.model_phase3, defaultModel)
+    modelPhase4.value = validModel(entry.data.model_phase4, models.value[1]?.alias || defaultModel)
     phase1Result.value = entry.data.phase1_result
     phase2ProperNouns.value = entry.data.phase2_proper_nouns || []
     phase2Style.value = entry.data.phase2_style || ''
     phase3Result.value = entry.data.phase3_result || ''
     phase4Result.value = entry.data.phase4_result || ''
+    historySnapshot.value = true
 
     // Save model selections for future sessions
     saveModelSelections()
@@ -276,7 +389,7 @@ export const useTranslationStore = defineStore('translation', () => {
       source_language: sourceLanguage.value,
       target_language: targetLanguage.value,
       model_phase1: modelPhase1.value,
-      model_phase2: modelPhase2.value,
+      model_phase2: '',
       model_phase3: modelPhase3.value,
       model_phase4: modelPhase4.value,
       phase1_result: phase1Result.value || null,
@@ -309,7 +422,7 @@ export const useTranslationStore = defineStore('translation', () => {
     // Add simple markdown formatting — frontmatter with metadata
     const md = `---
 source_language: ${sourceLanguage.value}
-target_language: ${target_language.value}
+target_language: ${targetLanguage.value}
 timestamp: ${new Date().toISOString()}
 ---\n\n${content}`
     downloadTxt(filename.replace('.txt', '.md'), md)
@@ -318,10 +431,10 @@ timestamp: ${new Date().toISOString()}
   return {
     // State
     sessionId, sourceText, sourceLanguage, targetLanguage,
-    modelPhase1, modelPhase2, modelPhase3, modelPhase4,
+    modelPhase1, modelPhase3, modelPhase4,
     phase1Result, phase2ProperNouns, phase2Style,
     phase3Result, phase4Result,
-    loading, currentPhaseLoading, error, models, history,
+    loading, currentPhaseLoading, error, models, history, progress,
 
     // Computed
     phase1Done, phase2Ready, phase3Ready, phase4Ready,
diff --git a/frontend/src/types.ts b/frontend/src/types.ts
index 3e177bb..1599001 100644
--- a/frontend/src/types.ts
+++ b/frontend/src/types.ts
@@ -17,6 +17,15 @@ export interface Phase1Result {
   proper_nouns: ProperNoun[]
   summary: string
   style: string
+  warnings?: string[]
+}
+
+export interface PhaseProgress {
+  phase: number
+  chunk: number
+  total_chunks: number
+  status: string
+  preview: string
 }
 
 export interface TranslationSessionData {
@@ -28,12 +37,26 @@ export interface TranslationSessionData {
   model_phase3: string
   model_phase4: string
   phase1_result: Phase1Result | null
+  phase1_chunks?: Array<{ source_text: string; translated: string }>
   phase2_proper_nouns: ProperNoun[]
   phase2_style: string
   phase3_result: string
+  phase3_chunks?: string[]
   phase4_result: string
+  progress?: PhaseProgress | null
 }
 
+export type SessionConfig = Pick<
+  TranslationSessionData,
+  | 'source_text'
+  | 'source_language'
+  | 'target_language'
+  | 'model_phase1'
+  | 'model_phase2'
+  | 'model_phase3'
+  | 'model_phase4'
+>
+
 export interface SessionResponse {
   session_id: string
   data: TranslationSessionData
diff --git a/frontend/src/views/LoginView.vue b/frontend/src/views/LoginView.vue
index ba241d9..ff8edcd 100644
--- a/frontend/src/views/LoginView.vue
+++ b/frontend/src/views/LoginView.vue
@@ -1,5 +1,6 @@
 <script setup lang="ts">
 import { ref } from 'vue'
+import { isAxiosError } from 'axios'
 import * as api from '../api'
 
 const id = ref('')
@@ -14,8 +15,10 @@ async function handleLogin() {
   try {
     await api.login(id.value, password.value)
     window.location.reload()
-  } catch (e: any) {
-    errorMessage.value = e.response?.data?.detail || '로그인에 실패했습니다'
+  } catch (error: unknown) {
+    errorMessage.value = isAxiosError<{ detail?: string }>(error)
+      ? error.response?.data?.detail || '로그인에 실패했습니다'
+      : '로그인에 실패했습니다'
   } finally {
     loading.value = false
   }
diff --git a/frontend/src/views/TranslatorView.vue b/frontend/src/views/TranslatorView.vue
index 378cf1a..4c0d4db 100644
--- a/frontend/src/views/TranslatorView.vue
+++ b/frontend/src/views/TranslatorView.vue
@@ -106,7 +106,6 @@ function handleNewTranslation() {
       <section class="bg-white dark:bg-gray-800 rounded-xl border border-gray-200 dark:border-gray-700 p-5">
         <div class="flex flex-wrap items-end gap-4">
           <LLMSelector label="Phase 1 — 초벌번역" v-model:value="store.modelPhase1" :models="store.models" />
-          <LLMSelector label="Phase 2 — 고유명사" v-model:value="store.modelPhase2" :models="store.models" />
           <LLMSelector label="Phase 3 — 재번역" v-model:value="store.modelPhase3" :models="store.models" />
           <LLMSelector label="Phase 4 — 마무리" v-model:value="store.modelPhase4" :models="store.models" />
 
@@ -208,16 +207,32 @@ function handleNewTranslation() {
       </div>
 
       <!-- Loading Indicator -->
-      <div v-if="store.loading && !store.error" class="flex items-center gap-2 text-blue-600 dark:text-blue-400 bg-blue-50 dark:bg-blue-900/20 px-4 py-3 rounded-lg text-sm">
-        <svg class="animate-spin h-4 w-4" xmlns="http://www.w3.org/2000/svg" fill="none" viewBox="0 0 24 24">
-          <circle class="opacity-25" cx="12" cy="12" r="10" stroke="currentColor" stroke-width="4"></circle>
-          <path class="opacity-75" fill="currentColor" d="M4 12a8 8 0 018-8V0C5.373 0 0 5.373 0 12h4z"></path>
-        </svg>
-        {{ phaseLabel(store.currentPhaseLoading || 0) }} 진행 중...
+      <div v-if="store.loading && !store.error" class="text-blue-700 dark:text-blue-300 bg-blue-50 dark:bg-blue-900/20 px-4 py-3 rounded-lg text-sm space-y-3">
+        <div class="flex items-center gap-2">
+          <svg class="animate-spin h-4 w-4" xmlns="http://www.w3.org/2000/svg" fill="none" viewBox="0 0 24 24">
+            <circle class="opacity-25" cx="12" cy="12" r="10" stroke="currentColor" stroke-width="4"></circle>
+            <path class="opacity-75" fill="currentColor" d="M4 12a8 8 0 018-8V0C5.373 0 0 5.373 0 12h4z"></path>
+          </svg>
+          <span>{{ phaseLabel(store.currentPhaseLoading || 0) }} 진행 중</span>
+          <span v-if="store.progress" class="font-medium">
+            · 조각 {{ store.progress.chunk }}/{{ store.progress.total_chunks }} · {{ store.progress.status }}
+          </span>
+          <span v-else class="text-blue-500 dark:text-blue-400">· 서버 응답 대기 중</span>
+        </div>
+        <pre
+          v-if="store.progress?.preview"
+          class="max-h-44 overflow-y-auto whitespace-pre-wrap break-words rounded-lg border border-blue-200 dark:border-blue-800 bg-white/70 dark:bg-gray-900/60 p-3 text-xs leading-relaxed text-gray-600 dark:text-gray-300"
+        >{{ store.progress.preview }}</pre>
       </div>
 
       <!-- Phase 2: Proper Noun Editor -->
       <section v-if="store.phase1Done">
+        <div
+          v-if="store.phase1Result?.warnings?.length"
+          class="mb-4 rounded-xl border border-amber-200 dark:border-amber-800 bg-amber-50 dark:bg-amber-900/20 px-4 py-3 text-sm text-amber-800 dark:text-amber-300"
+        >
+          <p v-for="warning in store.phase1Result.warnings" :key="warning">{{ warning }}</p>
+        </div>
         <ProperNounsEditor
           :proper_nouns="store.phase2ProperNouns"
           :style="store.phase2Style"
diff --git a/requirements-dev.txt b/requirements-dev.txt
new file mode 100644
index 0000000..19a8fa1
--- /dev/null
+++ b/requirements-dev.txt
@@ -0,0 +1,2 @@
+-r requirements.txt
+httpx==0.28.*
diff --git a/requirements.txt b/requirements.txt
index 6738499..8be99d3 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,6 +1,6 @@
 fastapi==0.115.*
 uvicorn[standard]==0.34.*
 pyjwt==2.10.*
-passlib[bcrypt]==1.7.*
+bcrypt>=4.2,<6
 openai==1.66.*
 python-dotenv==1.0.*
diff --git a/tests/__init__.py b/tests/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/tests/test_backend.py b/tests/test_backend.py
new file mode 100644
index 0000000..661a213
--- /dev/null
+++ b/tests/test_backend.py
@@ -0,0 +1,252 @@
+import json
+import unittest
+from unittest.mock import AsyncMock, patch
+
+from fastapi.testclient import TestClient
+
+from backend.auth import create_access_token
+from backend.llm_client import get_llm_configs
+from backend.main import _split_text, app
+from backend.models import TranslationSession
+from backend.sessions import SessionConflictError, SessionStore, session_store
+
+
+class ApiTests(unittest.TestCase):
+    def setUp(self):
+        session_store._store.clear()
+        token = create_access_token({"user_id": "admin"})
+        self.client = TestClient(app, headers={"Authorization": f"Bearer {token}"})
+        self.model = get_llm_configs()[0].alias
+
+    def create_session(self, source_text: str = "Hello") -> str:
+        response = self.client.post(
+            "/api/translate/create",
+            json={
+                "source_text": source_text,
+                "source_language": "영어",
+                "target_language": "한국어",
+                "model_phase1": self.model,
+                "model_phase2": "",
+                "model_phase3": self.model,
+                "model_phase4": self.model,
+            },
+        )
+        self.assertEqual(response.status_code, 200)
+        return response.json()["session_id"]
+
+    def test_translation_routes_require_authentication(self):
+        response = TestClient(app).post(
+            "/api/translate/create", json={"source_text": "Hello"}
+        )
+        self.assertEqual(response.status_code, 401)
+
+    def test_progress_endpoint_returns_only_transient_progress(self):
+        session_id = self.create_session()
+        session_store.set_progress(
+            session_id,
+            {
+                "phase": 1,
+                "chunk": 2,
+                "total_chunks": 3,
+                "status": "수신 중",
+                "preview": "부분 결과",
+            },
+            "admin",
+        )
+
+        response = self.client.get(f"/api/sessions/{session_id}/progress")
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(
+            response.json(),
+            {
+                "progress": {
+                    "phase": 1,
+                    "chunk": 2,
+                    "total_chunks": 3,
+                    "status": "수신 중",
+                    "preview": "부분 결과",
+                }
+            },
+        )
+
+    def test_unknown_model_is_rejected(self):
+        response = self.client.post(
+            "/api/translate/create",
+            json={"source_text": "Hello", "model_phase1": "missing-model"},
+        )
+        self.assertEqual(response.status_code, 400)
+
+    def test_source_update_invalidates_all_results(self):
+        session_id = self.create_session()
+        phase1 = json.dumps(
+            {
+                "translated": "안녕하세요",
+                "proper_nouns": [],
+                "summary": "인사",
+                "style": "중립적",
+            },
+            ensure_ascii=False,
+        )
+        responses = [phase1, "다시 번역", "최종 번역"]
+        with patch("backend.main.chat_complete", new=AsyncMock(side_effect=responses)):
+            self.assertEqual(
+                self.client.post(f"/api/translate/{session_id}/phase1").status_code,
+                200,
+            )
+            self.assertEqual(
+                self.client.post(
+                    f"/api/translate/{session_id}/phase2",
+                    json={"proper_nouns": [], "style": "중립적"},
+                ).status_code,
+                200,
+            )
+            self.assertEqual(
+                self.client.post(f"/api/translate/{session_id}/phase3").status_code,
+                200,
+            )
+            self.assertEqual(
+                self.client.post(f"/api/translate/{session_id}/phase4").status_code,
+                200,
+            )
+
+        response = self.client.patch(
+            f"/api/sessions/{session_id}", json={"source_text": "Changed"}
+        )
+        self.assertEqual(response.status_code, 200)
+        data = response.json()["data"]
+        self.assertIsNone(data["phase1_result"])
+        self.assertEqual(data["phase1_chunks"], [])
+        self.assertEqual(data["phase2_proper_nouns"], [])
+        self.assertEqual(data["phase2_style"], "")
+        self.assertEqual(data["phase3_result"], "")
+        self.assertEqual(data["phase3_chunks"], [])
+        self.assertEqual(data["phase4_result"], "")
+
+    def test_invalid_phase1_json_falls_back_to_plain_translation(self):
+        session_id = self.create_session()
+        mocked = AsyncMock(side_effect=["not json", "still not json", "일반 번역"])
+        with patch("backend.main.chat_complete", new=mocked):
+            response = self.client.post(f"/api/translate/{session_id}/phase1")
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json()["data"]["phase1_result"]["translated"], "일반 번역")
+        self.assertTrue(response.json()["data"]["phase1_result"]["warnings"])
+        self.assertEqual(mocked.await_count, 3)
+
+    def test_phase1_accepts_fenced_json_with_trailing_comma(self):
+        session_id = self.create_session()
+        malformed = """응답입니다.
+```json
+{"translated":"안녕하세요","proper_nouns":[],"summary":"인사","style":"중립적",}
+```
+"""
+        mocked = AsyncMock(return_value=malformed)
+        with patch("backend.main.chat_complete", new=mocked):
+            response = self.client.post(f"/api/translate/{session_id}/phase1")
+
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json()["data"]["phase1_result"]["translated"], "안녕하세요")
+        self.assertEqual(mocked.await_count, 1)
+
+    def test_long_phase1_is_split_into_bounded_chunks(self):
+        source_text = "这是用于测试长文本分割的句子。" * 120
+        expected_chunks = _split_text(source_text)
+        self.assertGreater(len(expected_chunks), 1)
+        self.assertTrue(all(len(chunk) <= 1500 for chunk in expected_chunks))
+        session_id = self.create_session(source_text)
+        phase1 = json.dumps(
+            {
+                "translated": "긴 문장 번역",
+                "proper_nouns": [],
+                "summary": "요약",
+                "style": "중립적",
+            },
+            ensure_ascii=False,
+        )
+        mocked = AsyncMock(side_effect=[phase1] * len(expected_chunks))
+        with patch("backend.main.chat_complete", new=mocked):
+            response = self.client.post(f"/api/translate/{session_id}/phase1")
+
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(mocked.await_count, len(expected_chunks))
+        self.assertEqual(len(response.json()["data"]["phase1_chunks"]), len(expected_chunks))
+
+        self.client.post(
+            f"/api/translate/{session_id}/phase2",
+            json={"proper_nouns": [], "style": "중립적"},
+        )
+        phase3_mock = AsyncMock(side_effect=["재번역 조각"] * len(expected_chunks))
+        with patch("backend.main.chat_complete", new=phase3_mock):
+            response = self.client.post(f"/api/translate/{session_id}/phase3")
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(phase3_mock.await_count, len(expected_chunks))
+        self.assertEqual(len(response.json()["data"]["phase3_chunks"]), len(expected_chunks))
+
+        phase4_mock = AsyncMock(side_effect=["완성 조각"] * len(expected_chunks))
+        with patch("backend.main.chat_complete", new=phase4_mock):
+            response = self.client.post(f"/api/translate/{session_id}/phase4")
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(phase4_mock.await_count, len(expected_chunks))
+        self.assertEqual(
+            response.json()["data"]["phase4_result"].count("완성 조각"),
+            len(expected_chunks),
+        )
+
+    def test_split_text_prefers_english_sentence_boundary(self):
+        sentence = "This is an English sentence with several words. "
+        chunks = _split_text(sentence * 50)
+
+        self.assertGreater(len(chunks), 1)
+        self.assertTrue(all(chunk.endswith(".") for chunk in chunks[:-1]))
+        self.assertTrue(all(len(chunk) <= 1500 for chunk in chunks))
+
+
+class SessionStoreTests(unittest.TestCase):
+    def test_sessions_are_isolated_by_owner(self):
+        store = SessionStore()
+        session_id = store.create(TranslationSession(source_text="secret"), "alice")
+
+        self.assertIsNotNone(store.get(session_id, "alice"))
+        self.assertIsNone(store.get(session_id, "bob"))
+        self.assertIsNone(store.update(session_id, {"source_text": "stolen"}, "bob"))
+        self.assertFalse(store.delete(session_id, "bob"))
+        self.assertEqual(store.get(session_id, "alice").source_text, "secret")
+
+    def test_stale_phase_update_is_rejected(self):
+        store = SessionStore()
+        session_id = store.create(TranslationSession(source_text="first"), "alice")
+        _, version = store.get_with_version(session_id, "alice")
+        store.update(session_id, {"source_text": "second"}, "alice")
+
+        with self.assertRaises(SessionConflictError):
+            store.update(
+                session_id,
+                {"phase3_result": "stale"},
+                "alice",
+                expected_version=version,
+            )
+
+    def test_progress_does_not_change_session_version(self):
+        store = SessionStore()
+        session_id = store.create(TranslationSession(source_text="first"), "alice")
+        _, version = store.get_with_version(session_id, "alice")
+
+        self.assertTrue(
+            store.set_progress(
+                session_id,
+                {
+                    "phase": 1,
+                    "chunk": 1,
+                    "total_chunks": 2,
+                    "status": "수신 중",
+                    "preview": "부분 결과",
+                },
+                "alice",
+            )
+        )
+        session, current_version = store.get_with_version(session_id, "alice")
+        self.assertEqual(current_version, version)
+        self.assertEqual(session.progress.preview, "부분 결과")
+
+
+if __name__ == "__main__":
+    unittest.main()