feat(image-render): main + Dockerfile + compose entry (port 18714)

ai_trade(HIGH, native python :8001) + wsl_docker(NORMAL, WSL2 Ubuntu-24.04
docker compose up). spec의 signal_v2→ai_trade, 22.04→24.04, web-ai-services
→web-ai/services 정정. sudoers NOPASSWD + 재부팅 검증 절차.
Plan-B-Infra Phase 3.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.gitignore

@@ -47,11 +47,11 @@ daily_trade_history.json
 watchlist.json
 bot_ipc.json
 
-# Test (top-level only; signal_v2/tests tracked separately)
+# Test (top-level only; ai_trade/tests tracked separately)
 tests/
 tests/*
-!signal_v2/tests/
+!ai_trade/tests/
-!signal_v2/tests/**
+!ai_trade/tests/**
 
 # System
 Thumbs.db
@@ -63,5 +63,22 @@ KIS_SETUP.md
 .claude/
 
 # Signal V2 runtime data
-signal_v2/data/*.db
+ai_trade/data/*.db
-signal_v2/data/*.db-*
+ai_trade/data/*.db-*
+
+# Plan-B-Insta services 예외 (코드는 추적, .env는 무시 유지)
+!services/
+!services/**/
+!services/**/*.py
+!services/**/Dockerfile
+!services/**/requirements.txt
+!services/**/.env.example
+!services/**/*.j2
+!services/**/*.html
+!services/**/*.css
+!services/**/.gitkeep
+!services/**/pytest.ini
+!services/docker-compose.yml
+# 단 실 .env는 무시 유지
+services/**/.env
+services/.env

CHECK_POINT.md

@@ -0,0 +1,277 @@
+# web-ai CHECK_POINT
+
+> Windows AI Server (192.168.45.59), AMD 9800X3D + RTX 5070 Ti (16GB VRAM).
+> V1(LSTM 레거시) + V2(Chronos-2 signal pipeline) 이중 구조.
+> 2026-05-18 작성.
+
+## 🚀 2026-05-18 박재오 7 결정 — Windows 컴퓨팅 노드 신설 (1주 작업)
+
+박재오 결정 7건 완료. 상세 가이드: `Obsidian Vault/raw/2026-05-18-Windows-NAS-아키텍처-7결정-통합.md`
+
+### 결정 6 — 옵션 4 하이브리드 운영
+
+```
+[Windows AI Server]
+
+🔵 Native Python (NSSM 자동 시작, HIGH priority)
+  ├─ ai_trade (트레이딩 :8001) ⭐ 절대 우선
+  ├─ Ollama qwen3:14b (:11435)
+  └─ MusicGen (:8765)
+
+🟢 WSL2 + Docker Engine (Docker Desktop X, 라이선스·메모리 ↓)
+  ├─ insta-render (:18710) ⭐ NEW
+  ├─ music-render (:8771) ⭐ NEW
+  ├─ video-render (:18712) ⭐ NEW (외부 API 게이트웨이)
+  └─ task-watcher (옵션 d 작업 감지)
+```
+
+### Day 2 — WSL2 + Docker Engine 설치 ⭐ (2시간)
+
+```powershell
+# 관리자 PowerShell
+wsl --install -d Ubuntu-22.04
+# 재부팅 후 Ubuntu 초기 설정
+
+wsl -d Ubuntu-22.04
+# 안에서 Docker Engine 설치
+sudo apt update && sudo apt install -y ca-certificates curl gnupg
+curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo gpg --dearmor -o /etc/apt/keyrings/docker.gpg
+echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/docker.gpg] https://download.docker.com/linux/ubuntu $(lsb_release -cs) stable" | sudo tee /etc/apt/sources.list.d/docker.list
+sudo apt update && sudo apt install -y docker-ce docker-compose-plugin
+sudo usermod -aG docker $USER
+sudo systemctl enable docker
+sudo systemctl start docker
+
+# Tailscale (NAS와 같은 LAN 확인)
+curl -fsSL https://tailscale.com/install.sh | sh
+sudo tailscale up
+```
+
+- [ ] WSL2 + Ubuntu 22.04 설치
+- [ ] Docker Engine 설치 (Desktop X)
+- [ ] Tailscale 가입
+
+### Day 3~4 — insta-render 컨테이너 신설 ⭐ (4시간)
+
+**디렉토리**: `C:\Users\jaeoh\Desktop\workspace\web-ai-services\insta-render\`
+
+**Dockerfile**: Playwright + Python 3.12 (raw 파일에 풀 코드)
+
+**main.py**: Redis 큐 worker + Playwright Browser pool + Semaphore(1)
+
+- [ ] 디렉토리 생성 + Dockerfile + main.py + requirements.txt
+- [ ] `docker compose up -d insta-render`
+- [ ] 테스트: NAS Redis에서 작업 push → Windows에서 처리 확인
+
+### Day 6 — NSSM 자동 시작 ⭐ (1시간)
+
+**NSSM 다운로드**: https://nssm.cc/download
+
+```powershell
+# 관리자 PowerShell
+
+# 트레이딩 자동 시작 (HIGH priority — 결정 b)
+nssm install ai_trade "C:\Python312\python.exe" "-m uvicorn main:app --host 0.0.0.0 --port 8001"
+nssm set ai_trade AppDirectory "C:\Users\jaeoh\Desktop\workspace\web-ai\ai_trade"
+nssm set ai_trade Priority HIGH_PRIORITY_CLASS
+nssm set ai_trade AppStartup AUTO
+
+# WSL2 + Docker 자동 시작
+nssm install wsl_docker "wsl" "-d Ubuntu-22.04 -- sudo service docker start && cd /workspace/web-ai-services && docker compose up -d"
+nssm set wsl_docker AppStartup AUTO
+
+# 시작
+nssm start ai_trade
+nssm start wsl_docker
+```
+
+- [ ] NSSM 다운로드 + 압축 해제 + PATH 추가
+- [ ] ai_trade service 등록 (HIGH priority)
+- [ ] wsl_docker service 등록
+- [ ] 재부팅 후 자동 시작 확인
+
+### Day 7 — 작업 감지 (옵션 d) — task-watcher ⭐ (2시간)
+
+박재오 작업 중 → Redis `queue:paused` 플래그 → 워커 일시정지. 트레이딩은 영향 X.
+
+선택지:
+- **A. 자동 (Python pynput + PowerShell API)** — 마우스·게임 process 감지, 자동 토글
+- **B. 수동 토글** — 박재오님이 작업 시작 시 `redis-cli SET queue:paused 1`, 종료 시 `DEL`
+- **C. NAS frontend에 토글 UI 1개** — 클릭 한 번
+
+→ 시작은 **B 수동 토글** (구현 0, 즉시 가능), 나중에 A 자동화로 진화.
+
+- [ ] B 수동 토글 명령어 확인
+- [ ] 또는 A Python pynput 자동 감지 구현 (선택, 2시간)
+
+### Day 5 — music-render 컨테이너 (선택 — MusicGen 패턴 정착)
+
+기존 NAS music-lab → Windows MusicGen 호출 패턴 이미 운영 중. 표준화만:
+- [ ] Redis 큐 사용으로 전환 (HTTP 직접 호출 X)
+- [ ] Browser pool 같은 패턴 적용 (Suno + MusicGen 동시 1개)
+
+### Day 5 — video-render 컨테이너 (선택 — 영상 생성 결정 4)
+
+외부 영상 API 6개 게이트웨이 (Runway·Sora·Veo·Pika·Kling·Luma):
+- [ ] 박재오 자금·품질 판단 후 1~2개 가입
+- [ ] `.env`에 API 키 추가
+- [ ] video-render Docker 컨테이너 신설
+
+---
+
+## 🔥 2026-05-18 추가 — NAS API 부하 진짜 원인 발견
+
+박재오 발견: 5건 + 중기 2건 적용 후 **web-ai V1+V2 4 process 종료가 NAS CPU 가장 큰 즉시 감소**.
+→ 진짜 병목은 **web-ai → NAS stock(:18500) 인바운드 API 호출 빈도**.
+
+상세: `Obsidian Vault/raw/2026-05-18-NAS-Window-AI-API-부하-해결방안.md`
+
+### 🔴 추가 즉시 작업 (50분으로 70% 부담 감소)
+
+#### A. 캐시 TTL 대폭 증가 ⭐⭐⭐ (10분)
+**파일**: `ai_trade/stock_client.py`
+```python
+# 변경 전 → 변경 후
+PORTFOLIO_TTL = 60   → 180  # 3분
+NEWS_TTL      = 300  → 600  # 10분
+SCREENER_TTL  = 60   → 300  # 5분
+```
+- [ ] 3 TTL 상수 증가
+- 효과: 분당 12 호출 → 3~4 호출 (70% 감소)
+
+#### B. V1·V2 단일화 결정 ⭐⭐ (10분 결정)
+- 동시 운영 시 NAS API 부담 2배 + KIS rate limit 충돌
+- **권장**: V1 폐기 + V2 단독 (Phase 4 자산 활용)
+- 또는 V2 임시 종료 + V1 유지 → Phase 5 진입 시 V1 폐기
+- [ ] 박재오 결정
+- [ ] 선택 안 된 쪽 `legacy/` 또는 `.disabled`
+- [ ] start.bat 한 쪽만
+
+#### C. (NAS 측) stock TTLCache — web-backend CHECK_POINT.md #13 참고
+- 박재오가 web-backend/stock에서 별도 적용
+
+## 🟢 현재 상태
+
+- **V2 Phase 4 완료** (5/17, 56/56 tests pass, main push)
+- Chronos-2 zero-shot 1일 수익률 + 분봉 모멘텀 5단계
+- 09:00~15:30 매 1분 / 16:00 일봉 추론
+- Sell-first 우선순위 (stop_loss · anomaly · take_profit) + 매수 hard gate
+- Confidence = chronos×0.5 + momentum×0.3 + screener×0.2
+
+---
+
+## 🔴 즉시 (이번 주)
+
+### 1. V1 vs V2 KIS rate limit 충돌 해결
+- **현재**: V1 + V2 동시 실행 시 KIS EGW00201 (초당 2회 제한) 충돌
+- **임시 해결**: V2 종료 상태 (현재 V1만 운영 중)
+- **결정 필요**: V1 deprecation 시점 (Phase 6)
+- [ ] 박재오 결정 — V1 폐기 일자 (예: 5/20 / 5/31)
+- [ ] Phase 5 진입 전 V1 정리
+
+### 2. `.venv` 한글 경로 문제 해결
+- 가상환경 한글 경로 깨짐 → 시스템 Python 사용 강제
+- 다른 개발 머신 협업 시 문제
+- [ ] `.venv`를 영문 경로로 이전 또는 시스템 Python으로 통일
+- [ ] start.bat에 경로 명시
+
+### 3. `state.signals` consumer-drain protocol 정의
+- Phase 5 prereq
+- 신호가 누적되기만 하고 소비 로직 미정의
+- [ ] consumer (예: agent-office /signal endpoint)가 처리한 신호 marking
+- [ ] 24h 만료 dedup 외에 *처리됨* 상태 추가
+
+---
+
+## 🟡 중기 (1~2주, Phase 5)
+
+### 4. agent-office `/signal` 엔드포인트 통합
+- web-ai V2가 매수/매도 신호 생성 → agent-office로 push
+- agent-office가 텔레그램 발송 + 사용자 결정 대기
+- [ ] V2에서 agent-office HTTP POST 호출 추가
+- [ ] payload: ticker, action, confidence, reasoning, chronos_quantile
+
+### 5. Ollama Qwen3 14B 통합 (Windows 로컬)
+- 신호 *해석* 레이어 — 단순 규칙 결과 → LLM 자연어 설명
+- 9.3GB VRAM 사용 (Chronos 7GB와 동시 가능, 15.5GB)
+- [ ] Ollama Windows 설치 (이미 실행 중인지 확인)
+- [ ] `state.signals` 큐에서 신호 pop → Qwen3 prompt → 결과 add
+- [ ] 텔레그램 전송 시 LLM 해석 텍스트 포함
+
+### 6. 이중 텔레그램 전송
+- 현재: V1만 텔레그램 발송 (Telegram Bot + KIS 자동주문)
+- Phase 5: V2도 별도 chat_id로 발송 (박재오 본인용 + 검증용)
+- [ ] V2 텔레그램 chat_id 환경변수 (`TELEGRAM_V2_CHAT_ID`)
+- [ ] V1·V2 메시지 톤 차별화 (V1 = 자동주문 / V2 = 신호 알림)
+
+### 7. holidays.json 자동 동기화
+- 현재: NAS에서 수동 copy
+- 한국 휴장일 누락 시 V2 폴링 실수 (휴장일에도 KIS 호출)
+- [ ] NAS realestate-lab 또는 별도 컨테이너에서 휴장일 자동 발급
+- [ ] V2가 부팅 시·매일 00:00에 GET 갱신
+
+---
+
+## 🟢 장기 (1개월+, Phase 6+)
+
+### 8. V1 완전 deprecation
+- LSTM 7-feature 모델 + main_server.py 폐기
+- 모든 자동매매 V2로 통일
+- [ ] V1 종료 일자 박재오 결정
+- [ ] V1 코드 `legacy/` 폴더로 이동 (read-only)
+
+### 9. Chronos-2 모델 미세조정 검토
+- 현재 zero-shot. 한국 주식 데이터로 미세조정 시 정확도 ↑ 가능?
+- 박재오 자체 학습 데이터 (KIS 1년치) → finetune
+- [ ] 데이터 수집·전처리
+- [ ] LoRA 또는 full finetune 결정
+
+### 10. KIS WebSocket 실 운영 검증
+- 현재 코드는 있으나 검증 부족
+- 실시간 호가가 1분 폴링보다 빠른 신호 확보 가능
+- [ ] 1주일 운영 후 latency·드롭 측정
+
+---
+
+## ✅ 최근 완료 (참고)
+
+- 2026-05-17: emit/skip 로깅 추가 (`2aa9f48`)
+- 2026-05-17: signal_generator poll_loop 통합 — Phase 4 완료 (`cc6310d`)
+- 2026-05-16: 코드 리뷰 수정 — sell-first 순서, anomaly 테스트 (`e574074`)
+- 2026-05-16: signal_generator 초안 — 9개 unit test (`b9def06`)
+- 2026-05-15: Foundation — 6개 env 임계값 + state.signals 필드 (`05ab284`)
+- 2026-05-14: FP32 강제 — Chronos FP16 overflow 회피 (`760f914`)
+
+---
+
+## 🔧 운영 커맨드 (Windows PowerShell)
+
+```powershell
+# V2 시작
+cd C:\Users\jaeoh\Desktop\workspace\web-ai\ai_trade
+.\start.bat
+
+# 테스트 실행 (56 tests)
+pytest tests/ -v
+
+# 로그 확인 (V2 실시간)
+Get-Content logs/ai_trade.log -Wait
+
+# Chronos 메모리 사용 확인
+nvidia-smi
+
+# KIS API 헬스 (REST)
+curl http://localhost:8001/health
+```
+
+---
+
+## 📚 참고
+
+- 위키: [[자산-주식-자동매매]] (V3.1 → V2 Phase 4 정정 필요)
+- NAS stock 연동: `192.168.45.54:18500` (X-WebAI-Key 인증)
+- CLAUDE.md (본 디렉토리 루트): Phase 진행도 + 환경변수 명세
+
+## 변경 이력
+
+- 2026-05-18: 페이지 신설. 즉시 3건 (V1/V2 충돌, .venv 한글, consumer protocol) + 중기 4건 (Phase 5) + 장기 3건 (V1 deprecation·Chronos finetune·WebSocket).

CLAUDE.md

@@ -1,24 +1,139 @@
 # web-ai — Workspace 가이드
 
-Windows AI 머신 (AMD 9800X3D + RTX 5070 Ti) 의 두 시그널 파이프라인 컨테이너.
+Windows AI 머신 (AMD 9800X3D + RTX 5070 Ti 16GB) 의 두 신호 파이프라인.
+**Confidence Signal Pipeline V2 의 Windows-side 구현체** (NAS stock 백엔드와 HTTP 연동).
+
+상위 워크스페이스 컨텍스트는 `../CLAUDE.md` 참조.
+
+---
 
 ## 디렉토리 구조
 
-| 경로 | 역할 | 상태 |
+| 경로 | 역할 | 포트 | 상태 |
-|------|------|------|
+|------|------|------|------|
-| `signal_v1/` | V1 자체 자동매매 시스템 (main_server.py + Trading Bot + Telegram Bot + LSTM + Ollama + KIS 자동주문) | 운영 중. Confidence Signal Pipeline V2 Phase 6 에서 deprecation 예정 |
+| `signal_v1/` | ⚠️ **DEPRECATED 2026-05-19** — 레거시 LSTM 봇. 사용 안 함. `legacy/signal_v1/`로 이동 완료 (2026-05-19) | `:8000` | **OFF** |
-| `signal_v2/` | V2 신호 파이프라인 (stock pull worker + Chronos-2 + signal API client) | Phase 2 에서 신설 |
+| `ai_trade/` | 자동매매 메인 (구 `signal_v2` 2026-05-19 rename) — Chronos-bolt + 분봉 모멘텀 + KIS WebSocket + 신호 생성 | `:8001` | **Phase 4 완료 (2026-05-17)**, Phase 5 대기 |
-| `.env` | V1 + V2 환경변수 공유 | KIS_*, TELEGRAM_*, STOCK_API_URL, WEBAI_API_KEY 등 |
+| `legacy/start_v1.bat` | (deprecated) V1 진입점 — root `start.bat`에서 이동됨. 자동 실행 차단 | — | **OFF** |
-| `start.bat` | V1 진입 (signal_v1 디렉토리 안 main_server.py 실행) | V2 별도 start 스크립트는 signal_v2/start.bat |
+| `ai_trade/start.bat` | 자동매매 진입점 | — | `ai_trade/main.py` uvicorn 실행 |
+| `services/` | (예정) NAS↔Windows 분산 worker — insta-render·music-render·video-render·task-watcher | 18710~ | **Plan-B-Insta 작업 중** |
+| `.env` | 환경변수 (`KIS_REAL_*`, `TELEGRAM_*`, `STOCK_API_URL`, `WEBAI_API_KEY`, `LOG_LEVEL`) | — | |
+| `requirements.txt` | 공용 의존성 | — | torch, chronos-forecasting, fastapi, httpx, websockets 등 |
 
-## 운영 가이드
+`.venv` 는 **구조적으로 깨짐**: `pyvenv.cfg` 가 한글 사용자 경로(`C:\Users\박재오\...`) 를 포함하여 콘솔 코드페이지가 roundtrip 못함. 테스트는 시스템 Python 으로 실행: `C:\Users\jaeoh\AppData\Local\Programs\Python\Python312\python.exe -m pytest ai_trade/tests -q`.
 
-- V1 시작: `start.bat` 또는 `cd signal_v1 && python main_server.py`
+---
-- V2 시작 (Phase 2 이후): `cd signal_v2 && python -m uvicorn main:app --port 8001`
+
-- 둘 다 동시 실행 가능 (포트 분리: V1=8000, V2=8001)
+## 서버 시작 방식
+
+### V1 (⚠️ DEPRECATED — 운영 X)
+2026-05-19부터 자동 시작 차단. `legacy/start_v1.bat`에 보존 (참고용만).
+별도 backtest 등 1회성 시 필요 시 박재오 직접 `legacy/start_v1.bat` 실행.
+
+### ai_trade 단독 (smoke/검증)
+```bat
+cd C:\Users\jaeoh\Desktop\workspace\web-ai\ai_trade
+.\start.bat
+```
+기대 로그: `Uvicorn running on http://0.0.0.0:8001`, `poll_loop started`, `[KIS] minute bars ... OK`, `[Chronos] predicted N tickers`, `signal emit XXXXXX buy conf=0.xxx`.
+
+휴장일/장 외 시간엔 `poll_loop` 만 idle. `Application startup complete` 만 보이면 정상.
+
+### V1 + V2 동시 실행 — **권장 안 함**
+**KIS app_key 초당 2회 한도 (EGW00201)** 충돌. V1 cycle + V2 분봉 cron 이 같은 KIS app_key 로 동시 호출하면 rate limit. 채택 해결책: V2 임시 종료 (Phase 3a 결정), Phase 6 V1 deprecation 시 자연 해소. 별도 app_key 발급은 옵션 B.
+
+---
 
 ## Phase 진행 상태 (Confidence Signal Pipeline V2)
 
-`web-ui/docs/superpowers/specs/2026-05-15-confidence-signal-pipeline-v2-architecture.md` 참조.
+| Phase | 내용 | 상태 |
+|-------|------|------|
+| 0 | Architecture & contract spec | ✅ Chronos-2 + Qwen3 14B 채택 |
+| 1 | stock 백엔드 WebAI API 보강 (NAS) | ✅ 102/102 tests, 운영 배포 |
+| 1.5 | V1 → `signal_v1/` rename | ✅ V1 정상 기동 |
+| 2 | ai_trade pull worker + signal API client + scheduler | ✅ 19/19 tests, `:8001` 기동 |
+| 3a | KIS REST 분봉 + WebSocket 호가 + NXT 스케줄 | ✅ 33/33 tests |
+| 3b | Chronos-bolt-base 추론 + 5분봉 모멘텀 분류기 | ✅ 45/45 tests, 실 KIS+Chronos chain 검증 |
+| 4 | Signal Generator (매수/매도 룰) + pull_worker 통합 + 로깅 | ✅ **2026-05-17 완료, 56/56 tests, push 완료** |
+| 5 | agent-office `/signal` + Ollama Qwen3 14B + 이중 텔레그램 | ⏳ 2주 예상 |
+| 6 | signal_v1 deprecation | ⏳ 1주 |
+| 7 | 운영 모니터링 + 4주 IC 검증 | ⏳ 1주 + 4주 |
 
-자세한 V1 가이드는 `signal_v1/CLAUDE.md` 참조.
+상세 spec/plan: `../web-ui/docs/superpowers/specs/` 및 `../web-ui/docs/superpowers/plans/` (web-ui repo 안에 보관됨 — V2 자체 코드와 분리 보관).
+
+---
+
+## ai_trade 디렉토리 내부
+
+| 파일 | 역할 |
+|------|------|
+| `main.py` | FastAPI app + lifespan (StockClient + KISClient + KISWebSocket + ChronosPredictor + SignalDedup 초기화). poll_loop task 생성 |
+| `config.py` | Settings dataclass — 환경변수 로드. Phase 4 추가 6 필드: `stop_loss_pct`, `take_profit_pct`, `chronos_spread_threshold`, `asking_bid_ratio_threshold`, `confidence_threshold`, `min_momentum_for_buy` |
+| `state.py` | PollState (process-wide singleton) — portfolio, screener_preview, news_sentiment, chronos_predictions, minute_bars, asking_price, **signals** (Phase 4) |
+| `stock_client.py` | NAS stock 백엔드 pull (X-WebAI-Key + 메모리 cache 60s/300s/60s + retry) |
+| `kis_client.py` | KIS REST 분봉/호가 — V1 토큰 read-only 공유 (mtime cache) + 초당 2회 throttle + 지수 backoff |
+| `kis_websocket.py` | KIS WebSocket H0STASP0 호가 + approval_key + 재연결 (1→2→4→max 30s) |
+| `chronos_predictor.py` | `amazon/chronos-bolt-base` zero-shot quantile (FP32 강제 — FP16 overflow 회피) |
+| `minute_momentum.py` | 5분봉 → strong_up/weak_up/neutral/weak_down/strong_down 5단계 분류 |
+| `signal_generator.py` | **Phase 4 — 매수/매도 룰 엔진**. `generate_signals(state, dedup, settings)` 진입. sell-first → buy 순서. 신호 emit/skip INFO/DEBUG 로그 |
+| `pull_worker.py` | asyncio cron — 장전 5분 / 장중 1분 / 장후 5분 / NXT / dead zone skip. cycle 끝에 `generate_signals` 호출 |
+| `scheduler.py` | polling window 판정 (KST 캘린더 + 휴장일) |
+| `rate_limit.py` | 초당 N회 token bucket |
+| `dedup.py` | SignalDedup SQLite WAL — `(ticker, action)` PK 24h |
+| `tests/` | 56 tests (pytest + respx HTTP mock + monkeypatch) |
+| `data/` | dedup.db (SQLite WAL) + `holidays.json` (NAS stock 에서 manual copy) |
+| `start.bat` | V2 진입 |
+
+---
+
+## 신호 룰 요약 (Phase 4)
+
+### 매수 (screener Top-N + portfolio, sell 신호 받은 종목은 skip)
+모두 충족:
+1. `chronos.median > 0`
+2. **`chronos.q90 - chronos.q10 < 0.6`** (absolute spread — 2026-05-17 spec amend, 기존 relative formula 가 zero-shot median≈0 빈번에서 모든 신호 거부)
+3. `minute_momentum == strong_up` (env 로 조정 가능)
+4. `asking_price.bid_ratio >= 0.6`
+
+종합 confidence = `chronos_conf * 0.5 + minute_score * 0.3 + screener_norm * 0.2`. `> 0.7` 시 emit.
+
+### 매도 (portfolio only, 우선순위 stop_loss → anomaly → take_profit)
+- **stop_loss**: `pnl_pct < -7%` 즉시 (confidence=1.0)
+- **anomaly**: `chronos.median < -1%` + `strong_down` + `bid_ratio < 0.4` + 종합 conf > 0.7
+- **take_profit**: `pnl_pct > 15%` 검토 (confidence=0.6)
+
+---
+
+## 알려진 함정 / Phase 7 백로그
+
+1. **KIS rate limit (EGW00201)** — V1+V2 동시 실행 시 충돌. Phase 6 자연 해소
+2. **`.venv` 한글 경로 깨짐** — 시스템 Python 사용
+3. **Chronos FP16 overflow** — 한국 주가 5만+ 시 inf. FP32 강제 (`chronos_predictor.py:39-41`)
+4. **`predict_quantiles` positional `inputs`** — ChronosBolt API 새 변경. `try/except TypeError` fallback 처리됨
+5. **`state.signals` consumer-drain protocol 미정의** — Phase 5 prereq. dict 무한 누적 위험 (실제로는 bounded by unique ticker count)
+6. **integration test 가 poll_loop 실제 호출 안 함** — `test_pull_worker.py:test_poll_loop_calls_generate_signals_after_cycle` 가 `generate_signals` 직접 호출. Phase 7 hardening 시 mock-iteration 으로 강화
+7. **KIS WebSocket URL `ws://ops.koreainvestment.com:21000/31000`** — 첫 운영 시 실제 KIS API docs 와 대조 필요
+8. **`_parse_asking_price` 필드 인덱스** — 마지막 2 필드 가정. 실 운영 raw 메시지 캡처 후 매핑 검증 필요
+9. **`holidays.json` 자동 동기화 부재** — NAS stock 의 `holidays.json` 을 수동 copy
+10. **schema rename** — Phase 0 §5.2 의 `lstm_pred_*`, `news_top[]` 는 `chronos_pred_*`, `news_reason(string)` 으로 변경됨. Phase 5 prompt 작성 시 반영
+11. **6개 env 필드가 `.env` 에 미기재** — 기본값으로 동작 가능하나 discoverability 위해 `.env.example` 또는 commented block 추가 권장
+
+---
+
+## 다음 단계 (Phase 5 진입 시 brainstorming 주제)
+
+- `state.signals` consumer 패턴: pop vs leave + Phase 5 자체 dedup
+- agent-office 의 `/signal` endpoint 설계 — POST 페이로드 schema
+- Ollama Qwen3 14B Q4 로컬 호출 — 타임아웃, retry, VRAM 공존 (Chronos + Qwen3 동시 메모리 9.3GB / 15.5GB 가용)
+- 이중 텔레그램 (본인 풀 / 아내 lite) — context augmentation 단일 호출에서 양쪽 메시지 생성
+- LLM 비용: ₩0 목표 유지 (로컬)
+
+---
+
+## 양쪽 디렉토리 (web-ui ↔ web-ai) 작업 시 주의
+
+- **코드**: ai_trade 는 web-ai/, spec/plan/메모리는 web-ui/
+- **커밋**: `web-ai` 와 `web-ui` 는 **별도 Gitea 저장소**. 각각 경로에서만 `git add/commit/push`
+- **메모리**: Claude Code 의 auto-memory 는 디렉토리별 격리. 핵심 reference 는 양쪽에 미러됨 (`./memory-mirror/` 또는 `~/.claude/projects/C--Users-jaeoh-Desktop-workspace-web-ai/memory/`)
+- **spec amendment 발생 시**: 코드는 `web-ai` 에 commit, spec 갱신은 `web-ui/docs/superpowers/specs/` 에 commit (Phase 4 spread formula 변경 사례 = web-ui commit `534ded5`)
+
+자세한 V1 가이드는 `signal_v1/CLAUDE.md` 참조 (있다면).

ai_trade/chronos_predictor.py

@@ -0,0 +1,132 @@
+"""Chronos-2 zero-shot forecaster wrapper."""
+from __future__ import annotations
+import logging
+from dataclasses import dataclass
+from datetime import datetime
+from zoneinfo import ZoneInfo
+
+import numpy as np
+
+logger = logging.getLogger(__name__)
+KST = ZoneInfo("Asia/Seoul")
+
+
+@dataclass
+class ChronosPrediction:
+    median: float
+    q10: float
+    q90: float
+    conf: float
+    as_of: str
+
+
+class ChronosPredictor:
+    """HuggingFace Chronos-2 zero-shot forecaster."""
+
+    def __init__(self, model_name: str = "amazon/chronos-2", device: str | None = None):
+        # BaseChronosPipeline auto-detects model variant (Chronos / ChronosBolt / Chronos-2)
+        # and returns the appropriate sub-pipeline. ChronosPipeline only supports legacy T5.
+        import torch
+        try:
+            from chronos import BaseChronosPipeline
+            pipeline_cls = BaseChronosPipeline
+        except ImportError:
+            from chronos import ChronosPipeline
+            pipeline_cls = ChronosPipeline
+
+        self._device = device or ("cuda" if torch.cuda.is_available() else "cpu")
+        # Always use float32 — Korean stock prices (e.g. 280,000원) exceed FP16 max (~65,504)
+        # causing inf in quantile output. FP32 is safe for typical price magnitudes.
+        dtype = torch.float32
+        logger.info("Loading Chronos pipeline: %s on %s (cls=%s)",
+                    model_name, self._device, pipeline_cls.__name__)
+        # Try `dtype` (newer API) first, fall back to `torch_dtype` (older)
+        try:
+            self._pipeline = pipeline_cls.from_pretrained(
+                model_name, device_map=self._device, dtype=dtype,
+            )
+        except TypeError:
+            self._pipeline = pipeline_cls.from_pretrained(
+                model_name, device_map=self._device, torch_dtype=dtype,
+            )
+        logger.info("Chronos pipeline loaded.")
+
+    def predict_batch(
+        self,
+        daily_ohlcv_dict: dict[str, list[dict]],
+        prediction_length: int = 1,
+        num_samples: int = 100,
+    ) -> dict[str, ChronosPrediction]:
+        """종목별 1-day return 분포 예측.
+
+        ChronosBolt / Chronos-2 등 신모델은 predict_quantiles 사용 (deterministic).
+        Legacy ChronosPipeline (T5) 는 sample-based predict.
+        """
+        import torch
+
+        tickers = list(daily_ohlcv_dict.keys())
+        if not tickers:
+            return {}
+
+        contexts = [
+            torch.tensor([bar["close"] for bar in daily_ohlcv_dict[t]], dtype=torch.float32)
+            for t in tickers
+        ]
+        now_iso = datetime.now(KST).isoformat()
+        results: dict[str, ChronosPrediction] = {}
+
+        # Modern API: predict_quantiles (ChronosBolt / Chronos-2)
+        if hasattr(self._pipeline, "predict_quantiles"):
+            quantile_levels = [0.1, 0.5, 0.9]
+            # ChronosBolt API: positional `inputs` (first arg). Older variants use `context`.
+            try:
+                quantiles_tensor, _ = self._pipeline.predict_quantiles(
+                    contexts,
+                    prediction_length=prediction_length,
+                    quantile_levels=quantile_levels,
+                )
+            except TypeError:
+                quantiles_tensor, _ = self._pipeline.predict_quantiles(
+                    context=contexts,
+                    prediction_length=prediction_length,
+                    quantile_levels=quantile_levels,
+                )
+            quantiles_np = (
+                quantiles_tensor.cpu().numpy()
+                if hasattr(quantiles_tensor, "cpu")
+                else np.asarray(quantiles_tensor)
+            )
+            # shape: [num_series, prediction_length, 3]
+            for i, ticker in enumerate(tickers):
+                q10_price, q50_price, q90_price = quantiles_np[i, 0, :]
+                last_close = daily_ohlcv_dict[ticker][-1]["close"]
+                median = float((q50_price - last_close) / last_close)
+                q10 = float((q10_price - last_close) / last_close)
+                q90 = float((q90_price - last_close) / last_close)
+                spread = (q90 - q10) / max(abs(median), 0.001)
+                conf = float(max(0.0, min(1.0, 1.0 - spread / 2.0)))
+                results[ticker] = ChronosPrediction(
+                    median=median, q10=q10, q90=q90, conf=conf, as_of=now_iso,
+                )
+            return results
+
+        # Legacy API: sample-based predict (ChronosPipeline T5)
+        forecasts = self._pipeline.predict(
+            context=contexts,
+            prediction_length=prediction_length,
+            num_samples=num_samples,
+        )
+        forecasts_np = forecasts.numpy() if hasattr(forecasts, "numpy") else np.asarray(forecasts)
+        for i, ticker in enumerate(tickers):
+            samples = forecasts_np[i, :, 0]
+            last_close = daily_ohlcv_dict[ticker][-1]["close"]
+            returns = (samples - last_close) / last_close
+            median = float(np.quantile(returns, 0.5))
+            q10 = float(np.quantile(returns, 0.1))
+            q90 = float(np.quantile(returns, 0.9))
+            spread = (q90 - q10) / max(abs(median), 0.001)
+            conf = float(max(0.0, min(1.0, 1.0 - spread / 2.0)))
+            results[ticker] = ChronosPrediction(
+                median=median, q10=q10, q90=q90, conf=conf, as_of=now_iso,
+            )
+        return results

ai_trade/config.py

@@ -18,7 +18,7 @@ class Settings:
     )
     port: int = field(default_factory=lambda: int(os.getenv("SIGNAL_V2_PORT", "8001")))
     db_path: Path = field(
-        default_factory=lambda: Path(__file__).parent / "data" / "signal_v2.db"
+        default_factory=lambda: Path(__file__).parent / "data" / "ai_trade.db"
     )
     # KIS — V1 호환 패턴 (KIS_ENV_TYPE virtual/real)
     kis_env_type: str = field(default_factory=lambda: os.getenv("KIS_ENV_TYPE", "virtual").lower())
@@ -34,6 +34,25 @@ class Settings:
                       str(Path(__file__).parent.parent / "signal_v1" / "data" / "kis_token.json"))
         )
     )
+    chronos_model: str = field(default_factory=lambda: os.getenv("CHRONOS_MODEL", "amazon/chronos-2"))
+    stop_loss_pct: float = field(
+        default_factory=lambda: float(os.getenv("STOP_LOSS_PCT", "-0.07"))
+    )
+    take_profit_pct: float = field(
+        default_factory=lambda: float(os.getenv("TAKE_PROFIT_PCT", "0.15"))
+    )
+    chronos_spread_threshold: float = field(
+        default_factory=lambda: float(os.getenv("CHRONOS_SPREAD_THRESHOLD", "0.6"))
+    )
+    asking_bid_ratio_threshold: float = field(
+        default_factory=lambda: float(os.getenv("ASKING_BID_RATIO_THRESHOLD", "0.6"))
+    )
+    confidence_threshold: float = field(
+        default_factory=lambda: float(os.getenv("CONFIDENCE_THRESHOLD", "0.7"))
+    )
+    min_momentum_for_buy: str = field(
+        default_factory=lambda: os.getenv("MIN_MOMENTUM_FOR_BUY", "strong_up")
+    )
 
     @property
     def kis_is_virtual(self) -> bool:

ai_trade/kis_client.py

@@ -4,7 +4,7 @@ import asyncio
 import json
 import logging
 import time
-from datetime import datetime
+from datetime import datetime, timedelta
 from pathlib import Path
 from zoneinfo import ZoneInfo
 
@@ -153,3 +153,41 @@ class KISClient:
             "current_price": current_price,
             "as_of": datetime.now(KST).isoformat(),
         }
+
+    async def get_daily_ohlcv(self, ticker: str, days: int = 60) -> list[dict]:
+        """KRX 일봉 OHLCV (TR_ID FHKST03010100).
+
+        Returns: [{"datetime", "open", "high", "low", "close", "volume"}, ...]
+                 시간 오름차순.
+        """
+        path = "/uapi/domestic-stock/v1/quotations/inquire-daily-itemchartprice"
+        today = datetime.now(KST).strftime("%Y%m%d")
+        start_date = (datetime.now(KST) - timedelta(days=days * 2)).strftime("%Y%m%d")
+        params = {
+            "FID_COND_MRKT_DIV_CODE": "J",
+            "FID_INPUT_ISCD": ticker,
+            "FID_INPUT_DATE_1": start_date,
+            "FID_INPUT_DATE_2": today,
+            "FID_PERIOD_DIV_CODE": "D",
+            "FID_ORG_ADJ_PRC": "1",
+        }
+        raw = await self._request_with_retry(
+            "GET", path, tr_id="FHKST03010100", params=params,
+        )
+        output2 = raw.get("output2", [])
+        bars = []
+        for row in output2:
+            try:
+                date = row["stck_bsop_date"]
+                bars.append({
+                    "datetime": f"{date[:4]}-{date[4:6]}-{date[6:]}",
+                    "open": int(row["stck_oprc"]),
+                    "high": int(row["stck_hgpr"]),
+                    "low": int(row["stck_lwpr"]),
+                    "close": int(row["stck_clpr"]),
+                    "volume": int(row["acml_vol"]),
+                })
+            except (KeyError, ValueError):
+                continue
+        bars.reverse()
+        return bars[-days:]

ai_trade/main.py

@@ -6,13 +6,14 @@ from contextlib import asynccontextmanager
 
 from fastapi import FastAPI
 
-from signal_v2 import state as state_mod
+from ai_trade import state as state_mod
-from signal_v2.config import get_settings
+from ai_trade.chronos_predictor import ChronosPredictor
-from signal_v2.kis_client import KISClient
+from ai_trade.config import get_settings
-from signal_v2.kis_websocket import KISWebSocket
+from ai_trade.kis_client import KISClient
-from signal_v2.pull_worker import poll_loop, make_asking_price_callback
+from ai_trade.kis_websocket import KISWebSocket
-from signal_v2.rate_limit import SignalDedup
+from ai_trade.pull_worker import poll_loop, make_asking_price_callback
-from signal_v2.stock_client import StockClient
+from ai_trade.rate_limit import SignalDedup
+from ai_trade.stock_client import StockClient
 
 logger = logging.getLogger(__name__)
 
@@ -24,6 +25,7 @@ class AppContext:
     poll_task: asyncio.Task | None = None
     kis_client: KISClient | None = None
     kis_ws: KISWebSocket | None = None
+    chronos: ChronosPredictor | None = None
 
 
 _ctx = AppContext()
@@ -69,10 +71,19 @@ async def lifespan(app: FastAPI):
         except Exception:
             logger.exception("KIS WebSocket startup failed — continuing without realtime asking_price")
 
+        # Load Chronos (heavy: ~1GB model download first time)
+        try:
+            _ctx.chronos = ChronosPredictor(model_name=settings.chronos_model)
+        except Exception:
+            logger.exception("ChronosPredictor load failed — continuing without chronos predictions")
+
     _ctx.poll_task = asyncio.create_task(
         poll_loop(
             _ctx.client, state_mod.state, _ctx.shutdown,
             kis_client=_ctx.kis_client,
+            chronos=_ctx.chronos,
+            dedup=_ctx.dedup,
+            settings=settings,
         )
     )
 

ai_trade/momentum_classifier.py

@@ -0,0 +1,69 @@
+"""분봉 OHLCV → 5-level 모멘텀 분류."""
+from __future__ import annotations
+from collections import deque
+
+# 분류 카테고리
+STRONG_UP = "strong_up"
+WEAK_UP = "weak_up"
+NEUTRAL = "neutral"
+WEAK_DOWN = "weak_down"
+STRONG_DOWN = "strong_down"
+
+_BARS_PER_5MIN = 5
+_LOOKBACK_5MIN_BARS = 5
+_VOLUME_AVG_WINDOW = 12  # 60분 = 5분봉 12개
+
+
+def aggregate_1min_to_5min(minute_bars: list[dict]) -> list[dict]:
+    """1분봉 N개 → 5분봉 floor(N/5) 개. 시간 오름차순.
+
+    각 5분봉: open=첫 1분봉 open, high=max, low=min, close=마지막 close, volume=sum.
+    """
+    bars_5min = []
+    chunks = len(minute_bars) // _BARS_PER_5MIN
+    for i in range(chunks):
+        chunk = minute_bars[i * _BARS_PER_5MIN : (i + 1) * _BARS_PER_5MIN]
+        bars_5min.append({
+            "datetime": chunk[0]["datetime"],
+            "open": chunk[0]["open"],
+            "high": max(b["high"] for b in chunk),
+            "low": min(b["low"] for b in chunk),
+            "close": chunk[-1]["close"],
+            "volume": sum(b["volume"] for b in chunk),
+        })
+    return bars_5min
+
+
+def classify_minute_momentum(minute_bars: deque) -> str:
+    """1분봉 deque → 5-level 모멘텀 분류.
+
+    Returns: STRONG_UP / WEAK_UP / NEUTRAL / WEAK_DOWN / STRONG_DOWN
+    """
+    minute_list = list(minute_bars)
+    if len(minute_list) < _BARS_PER_5MIN * _LOOKBACK_5MIN_BARS:
+        return NEUTRAL  # 데이터 부족
+
+    bars_5min = aggregate_1min_to_5min(minute_list)
+    if len(bars_5min) < _LOOKBACK_5MIN_BARS:
+        return NEUTRAL
+
+    recent = bars_5min[-_LOOKBACK_5MIN_BARS:]
+    up_count = sum(1 for b in recent if b["close"] > b["open"])
+
+    # 거래량 multiplier: recent 5 avg vs 60분 avg
+    recent_vol_avg = sum(b["volume"] for b in recent) / len(recent)
+    long_window = bars_5min[-_VOLUME_AVG_WINDOW:]
+    long_vol_avg = sum(b["volume"] for b in long_window) / len(long_window)
+    vol_mult = recent_vol_avg / long_vol_avg if long_vol_avg > 0 else 1.0
+
+    # 5-level 분류
+    if up_count == 5 and vol_mult >= 1.5:
+        return STRONG_UP
+    elif up_count >= 3 and vol_mult >= 1.0:
+        return WEAK_UP
+    elif up_count == 0 and vol_mult >= 1.5:
+        return STRONG_DOWN
+    elif up_count <= 2 and vol_mult < 1.0:
+        return WEAK_DOWN
+    else:
+        return NEUTRAL

ai_trade/pull_worker.py

@@ -5,12 +5,12 @@ import logging
 from collections import deque
 from datetime import datetime
 
-from signal_v2.kis_client import KISClient
+from ai_trade.kis_client import KISClient
-from signal_v2.scheduler import (
+from ai_trade.scheduler import (
-    KST, _is_market_day, _is_polling_window, _next_interval,
+    KST, _is_market_day, _is_polling_window, _next_interval, _is_post_close_trigger,
 )
-from signal_v2.state import PollState
+from ai_trade.state import PollState
-from signal_v2.stock_client import StockClient
+from ai_trade.stock_client import StockClient
 
 logger = logging.getLogger(__name__)
 
@@ -18,6 +18,9 @@ logger = logging.getLogger(__name__)
 async def poll_loop(
     client: StockClient, state: PollState, shutdown: asyncio.Event,
     kis_client: KISClient | None = None,
+    chronos=None,
+    dedup=None,
+    settings=None,
 ) -> None:
     """FastAPI lifespan 에서 asyncio.create_task 로 시작."""
     logger.info("poll_loop started")
@@ -28,6 +31,24 @@ async def poll_loop(
                 await _run_polling_cycle(client, state, kis_client=kis_client)
             except Exception:
                 logger.exception("poll cycle failed")
+            # Minute momentum 갱신 (매 cycle)
+            try:
+                update_minute_momentum_for_all(state)
+            except Exception:
+                logger.exception("minute momentum update failed")
+            # Post-close trigger (16:00 KST)
+            if _is_post_close_trigger(now) and chronos is not None and kis_client is not None:
+                try:
+                    await _run_post_close_cycle(kis_client, chronos, state)
+                except Exception:
+                    logger.exception("post-close cycle failed")
+            # Phase 4: generate signals
+            if dedup is not None and settings is not None:
+                try:
+                    from ai_trade.signal_generator import generate_signals
+                    generate_signals(state, dedup, settings)
+                except Exception:
+                    logger.exception("generate_signals failed")
         interval = _next_interval(now)
         try:
             await asyncio.wait_for(shutdown.wait(), timeout=interval)
@@ -125,3 +146,48 @@ def _screener_tickers(state: PollState) -> list[str]:
     if state.screener_preview is None:
         return []
     return [i["ticker"] for i in state.screener_preview.get("items", []) if "ticker" in i]
+
+
+async def _run_post_close_cycle(kis_client, chronos, state) -> None:
+    """16:00 KST 종가 후 1회: daily fetch + chronos predict."""
+    tickers = list(set(_portfolio_tickers(state)) | set(_screener_tickers(state)))
+    if not tickers:
+        return
+
+    daily_results = await asyncio.gather(*[
+        kis_client.get_daily_ohlcv(t, days=60) for t in tickers
+    ], return_exceptions=True)
+    daily_dict = {}
+    for ticker, result in zip(tickers, daily_results):
+        if isinstance(result, list) and len(result) >= 30:
+            daily_dict[ticker] = result
+            state.daily_ohlcv[ticker] = result
+        elif isinstance(result, Exception):
+            state.fetch_errors[f"daily_ohlcv/{ticker}"] = (
+                state.fetch_errors.get(f"daily_ohlcv/{ticker}", 0) + 1
+            )
+
+    if daily_dict and chronos is not None:
+        try:
+            predictions = chronos.predict_batch(daily_dict)
+        except Exception:
+            logger.exception("chronos predict_batch failed")
+            return
+        for ticker, pred in predictions.items():
+            state.chronos_predictions[ticker] = {
+                "median": pred.median,
+                "q10": pred.q10,
+                "q90": pred.q90,
+                "conf": pred.conf,
+                "as_of": pred.as_of,
+            }
+            state.last_updated[f"chronos/{ticker}"] = pred.as_of
+
+
+def update_minute_momentum_for_all(state) -> None:
+    """매 분봉 cycle 후 호출 — 모든 종목 모멘텀 갱신."""
+    from ai_trade.momentum_classifier import classify_minute_momentum
+    now_iso = datetime.now(KST).isoformat()
+    for ticker, bars in state.minute_bars.items():
+        state.minute_momentum[ticker] = classify_minute_momentum(bars)
+        state.last_updated[f"momentum/{ticker}"] = now_iso

ai_trade/scheduler.py

@@ -76,6 +76,14 @@ def _seconds_until_nxt_or_market_open(now: datetime) -> float:
     return 86400.0
 
 
+def _is_post_close_trigger(now: datetime) -> bool:
+    """16:00 KST ±1분 (post-close cycle 트리거). 평일/영업일만."""
+    if not _is_market_day(now):
+        return False
+    t = now.time()
+    return time(16, 0) <= t < time(16, 1)
+
+
 def _seconds_until_next_market_open(now: datetime) -> float:
     """다음 영업일의 07:00 KST 까지 초수 (휴장일/주말용)."""
     candidate = now.replace(hour=7, minute=0, second=0, microsecond=0)

ai_trade/signal_generator.py

@@ -0,0 +1,228 @@
+"""Phase 4 — 매수/매도 신호 생성.
+
+순수 함수 generate_signals(state, dedup, settings). state 를 mutate.
+"""
+from __future__ import annotations
+import logging
+from datetime import datetime
+from zoneinfo import ZoneInfo
+
+logger = logging.getLogger(__name__)
+KST = ZoneInfo("Asia/Seoul")
+
+MOMENTUM_SCORES = {
+    "strong_up": 1.0,
+    "weak_up": 0.7,
+    "neutral": 0.5,
+    "weak_down": 0.3,
+    "strong_down": 0.0,
+}
+
+
+def generate_signals(state, dedup, settings) -> None:
+    """Phase 4 entry — state-mutating. Evaluation order: sell first (priority), then buy. A ticker receiving a sell signal in this cycle is excluded from buy evaluation to avoid silent overwrite."""
+    _evaluate_sell_signals(state, dedup, settings)
+    _evaluate_buy_signals(state, dedup, settings)
+
+
+# ----- 매수 -----
+
+def _evaluate_buy_signals(state, dedup, settings) -> None:
+    candidates = _buy_candidates(state)
+    for ticker, name, rank in candidates:
+        existing = state.signals.get(ticker)
+        if existing is not None and existing.get("action") == "sell":
+            logger.debug("buy %s skipped: same-cycle sell precedence", ticker)
+            continue
+        if not _check_buy_hard_gate(state, ticker, settings):
+            logger.debug("buy %s skipped: hard gate failed", ticker)
+            continue
+        confidence = _compute_buy_confidence(state, ticker, rank)
+        if confidence <= settings.confidence_threshold:
+            logger.debug("buy %s skipped: confidence %.3f <= %.3f",
+                         ticker, confidence, settings.confidence_threshold)
+            continue
+        if dedup.is_recent(ticker, "buy", within_hours=24):
+            logger.debug("buy %s skipped: dedup 24h", ticker)
+            continue
+        state.signals[ticker] = _build_buy_signal(state, ticker, name, rank, confidence)
+        dedup.record(ticker, "buy", confidence=confidence)
+        logger.info("signal emit %s buy conf=%.3f rank=%s", ticker, confidence, rank)
+
+
+def _buy_candidates(state) -> list[tuple[str, str, int | None]]:
+    """screener Top-N (rank 1..N) + portfolio (rank=None)."""
+    candidates: list[tuple[str, str, int | None]] = []
+    seen: set[str] = set()
+    if state.screener_preview is not None:
+        for i, item in enumerate(state.screener_preview.get("items", [])):
+            ticker = item.get("ticker")
+            if not ticker or ticker in seen:
+                continue
+            seen.add(ticker)
+            name = item.get("name", ticker)
+            candidates.append((ticker, name, i + 1))
+    if state.portfolio is not None:
+        for h in state.portfolio.get("holdings", []):
+            ticker = h.get("ticker")
+            if not ticker or ticker in seen:
+                continue
+            seen.add(ticker)
+            candidates.append((ticker, h.get("name", ticker), None))
+    return candidates
+
+
+def _check_buy_hard_gate(state, ticker: str, settings) -> bool:
+    pred = state.chronos_predictions.get(ticker)
+    if pred is None or pred.get("median", 0) <= 0:
+        return False
+    spread = pred.get("q90", 0) - pred.get("q10", 0)
+    if spread >= settings.chronos_spread_threshold:
+        return False
+    momentum = state.minute_momentum.get(ticker)
+    if momentum != settings.min_momentum_for_buy:
+        return False
+    ap = state.asking_price.get(ticker)
+    if ap is None or ap.get("bid_ratio", 0) < settings.asking_bid_ratio_threshold:
+        return False
+    return True
+
+
+def _compute_buy_confidence(state, ticker: str, rank: int | None) -> float:
+    pred = state.chronos_predictions[ticker]
+    chronos_conf = pred["conf"]
+    minute_score = MOMENTUM_SCORES.get(state.minute_momentum.get(ticker, "neutral"), 0.5)
+    screener_norm = max(0.0, 1 - (rank - 1) / 20) if rank is not None else 0.0
+    return chronos_conf * 0.5 + minute_score * 0.3 + screener_norm * 0.2
+
+
+def _build_buy_signal(state, ticker: str, name: str, rank: int | None, confidence: float) -> dict:
+    ap = state.asking_price[ticker]
+    return {
+        "ticker": ticker,
+        "name": name,
+        "action": "buy",
+        "confidence_webai": confidence,
+        "current_price": ap["current_price"],
+        "avg_price": None,
+        "pnl_pct": None,
+        "context": _build_context(state, ticker, rank),
+        "as_of": datetime.now(KST).isoformat(),
+    }
+
+
+# ----- 매도 -----
+
+def _evaluate_sell_signals(state, dedup, settings) -> None:
+    if state.portfolio is None:
+        return
+    for holding in state.portfolio.get("holdings", []):
+        ticker = holding.get("ticker")
+        if not ticker:
+            continue
+        sell = _try_stop_loss(state, holding, settings)
+        if sell is None:
+            sell = _try_anomaly(state, holding, settings)
+        if sell is None:
+            sell = _try_take_profit(state, holding, settings)
+        if sell is None:
+            continue
+        if dedup.is_recent(ticker, "sell", within_hours=24):
+            logger.debug("sell %s skipped: dedup 24h", ticker)
+            continue
+        state.signals[ticker] = sell
+        dedup.record(ticker, "sell", confidence=sell["confidence_webai"])
+        logger.info("signal emit %s sell conf=%.3f reason=%s",
+                    ticker, sell["confidence_webai"],
+                    sell.get("context", {}).get("sell_reason"))
+
+
+def _try_stop_loss(state, holding: dict, settings) -> dict | None:
+    pnl = holding.get("pnl_pct")
+    if pnl is None or pnl >= settings.stop_loss_pct:
+        return None
+    return _build_sell_signal(state, holding, confidence=1.0, reason="stop_loss")
+
+
+def _try_take_profit(state, holding: dict, settings) -> dict | None:
+    pnl = holding.get("pnl_pct")
+    if pnl is None or pnl <= settings.take_profit_pct:
+        return None
+    return _build_sell_signal(state, holding, confidence=0.6, reason="take_profit")
+
+
+def _try_anomaly(state, holding: dict, settings) -> dict | None:
+    ticker = holding["ticker"]
+    pred = state.chronos_predictions.get(ticker)
+    if pred is None or pred["median"] >= -0.01:
+        return None
+    momentum = state.minute_momentum.get(ticker)
+    if momentum != "strong_down":
+        return None
+    ap = state.asking_price.get(ticker)
+    if ap is None:
+        return None
+    if ap["bid_ratio"] > (1 - settings.asking_bid_ratio_threshold):
+        return None
+    minute_score = 1.0 - MOMENTUM_SCORES.get(momentum, 0.5)
+    confidence = pred["conf"] * 0.5 + minute_score * 0.3 + 1.0 * 0.2
+    if confidence <= settings.confidence_threshold:
+        return None
+    return _build_sell_signal(state, holding, confidence=confidence, reason="anomaly")
+
+
+def _build_sell_signal(state, holding: dict, confidence: float, reason: str) -> dict:
+    ticker = holding["ticker"]
+    return {
+        "ticker": ticker,
+        "name": holding.get("name", ticker),
+        "action": "sell",
+        "confidence_webai": confidence,
+        "current_price": holding.get("current_price"),
+        "avg_price": holding.get("avg_price"),
+        "pnl_pct": holding.get("pnl_pct"),
+        "context": _build_context(state, ticker, rank=None, sell_reason=reason),
+        "as_of": datetime.now(KST).isoformat(),
+    }
+
+
+# ----- Context -----
+
+def _build_context(state, ticker: str, rank: int | None, sell_reason: str | None = None) -> dict:
+    pred = state.chronos_predictions.get(ticker) or {}
+    ap = state.asking_price.get(ticker) or {}
+    news_item = _find_news_sentiment(state, ticker)
+    screener_scores = _find_screener_scores(state, ticker)
+    context: dict = {
+        "chronos_pred_1d": pred.get("median"),
+        "chronos_pred_conf": pred.get("conf"),
+        "chronos_q10": pred.get("q10"),
+        "chronos_q90": pred.get("q90"),
+        "screener_rank": rank,
+        "screener_scores": screener_scores,
+        "minute_momentum": state.minute_momentum.get(ticker),
+        "asking_bid_ratio": ap.get("bid_ratio"),
+        "news_sentiment": news_item.get("score") if news_item else None,
+        "news_reason": news_item.get("reason") if news_item else None,
+    }
+    if sell_reason is not None:
+        context["sell_reason"] = sell_reason
+    return context
+
+
+def _find_news_sentiment(state, ticker: str) -> dict | None:
+    if state.news_sentiment is None:
+        return None
+    for item in state.news_sentiment.get("items", []):
+        if item.get("ticker") == ticker:
+            return item
+    return None
+
+
+def _find_screener_scores(state, ticker: str) -> dict | None:
+    if state.screener_preview is None:
+        return None
+    for item in state.screener_preview.get("items", []):
+        if item.get("ticker") == ticker:
+            return item.get("scores")
+    return None

ai_trade/start.bat

@@ -0,0 +1,3 @@
+@echo off
+cd /d "%~dp0\.."
+python -m uvicorn ai_trade.main:app --host 0.0.0.0 --port 8001

ai_trade/state.py

@@ -8,9 +8,13 @@ class PollState:
     portfolio: dict | None = None
     news_sentiment: dict | None = None
     screener_preview: dict | None = None
-    # Phase 3a additions
     minute_bars: dict[str, deque] = field(default_factory=dict)
     asking_price: dict[str, dict] = field(default_factory=dict)
+    # Phase 3b additions
+    daily_ohlcv: dict[str, list[dict]] = field(default_factory=dict)
+    chronos_predictions: dict[str, dict] = field(default_factory=dict)
+    minute_momentum: dict[str, str] = field(default_factory=dict)
+    signals: dict[str, dict] = field(default_factory=dict)
     last_updated: dict[str, str] = field(default_factory=dict)
     fetch_errors: dict[str, int] = field(default_factory=dict)
 

ai_trade/stock_client.py

@@ -9,11 +9,12 @@ import httpx
 
 logger = logging.getLogger(__name__)
 
-# Cache TTL by endpoint (seconds)
+# Cache TTL by endpoint (seconds).
+# 2026-05-18 — NAS 인바운드 호출 부담 완화 (Plan-A SP-A1).
 _TTL = {
-    "portfolio": 60.0,
+    "portfolio": 180.0,        # 3분 (1분 폴링 시 3 폴링당 1회 실제 fetch)
-    "news-sentiment": 300.0,
+    "news-sentiment": 600.0,   # 10분 (뉴스 sentiment는 자주 안 바뀜)
-    "screener-preview": 60.0,
+    "screener-preview": 300.0, # 5분 (Top-20은 분 단위로 거의 안 바뀜)
 }
 
 # Retry policy

ai_trade/tests/conftest.py

@@ -1,4 +1,4 @@
-"""Pytest fixtures for signal_v2 tests."""
+"""Pytest fixtures for ai_trade tests."""
 from pathlib import Path
 
 import pytest
@@ -8,7 +8,7 @@ import respx
 @pytest.fixture
 def tmp_dedup_db(tmp_path) -> Path:
     """SQLite 단위 테스트용 임시 DB path."""
-    return tmp_path / "test_signal_v2.db"
+    return tmp_path / "test_ai_trade.db"
 
 
 @pytest.fixture

ai_trade/tests/test_chronos_predictor.py

@@ -0,0 +1,92 @@
+"""Tests for ChronosPredictor (model mock)."""
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+import pytest
+
+
+@pytest.fixture
+def mock_pipeline():
+    """Mock BaseChronosPipeline.from_pretrained returning a mock pipeline object."""
+    with patch("chronos.BaseChronosPipeline") as cls:
+        cls.__name__ = "BaseChronosPipeline"
+        instance = MagicMock()
+        # ChronosBolt API: predict_quantiles returns (quantiles_tensor, mean_tensor)
+        # Modern (predict_quantiles) branch will be used since hasattr(MagicMock, "predict_quantiles") is True.
+        cls.from_pretrained.return_value = instance
+        yield instance
+
+
+@pytest.fixture
+def mock_torch_cpu():
+    with patch("torch.cuda.is_available", return_value=False):
+        yield
+
+
+def _daily_ohlcv(close_seq):
+    return [{"datetime": f"2026-05-{i+1:02d}", "open": c, "high": c, "low": c,
+             "close": c, "volume": 1000} for i, c in enumerate(close_seq)]
+
+
+def _mk_quantiles_tensor(q10_price: float, q50_price: float, q90_price: float):
+    """Helper: build predict_quantiles return tensor shape [1, 1, 3]."""
+    import torch
+    return torch.tensor([[[q10_price, q50_price, q90_price]]], dtype=torch.float32)
+
+
+def test_predict_batch_returns_prediction_dict(mock_pipeline, mock_torch_cpu):
+    """mock predict_quantiles → dict[ticker, ChronosPrediction]. last_close=100, q50=102 → median≈+2%."""
+    quantiles = _mk_quantiles_tensor(101.5, 102.0, 102.5)  # narrow around 102
+    mock_pipeline.predict_quantiles.return_value = (quantiles, None)
+
+    from ai_trade.chronos_predictor import ChronosPredictor, ChronosPrediction
+    predictor = ChronosPredictor(model_name="mock-model")
+    daily = {"005930": _daily_ohlcv([100] * 60)}
+    result = predictor.predict_batch(daily)
+    assert "005930" in result
+    pred = result["005930"]
+    assert isinstance(pred, ChronosPrediction)
+    assert abs(pred.median - 0.02) < 0.001
+
+
+def test_conf_high_when_distribution_narrow(mock_pipeline, mock_torch_cpu):
+    """좁은 distribution (q90-q10 작음, median 0 아님) → conf ≈ 1."""
+    # last_close=100, q10=101.99, q50=102.00, q90=102.01
+    # returns: q10=0.0199, q50=0.02, q90=0.0201
+    # spread = (0.0201 - 0.0199) / max(0.02, 0.001) = 0.0002/0.02 = 0.01 → conf = 1 - 0.005 = 0.995
+    quantiles = _mk_quantiles_tensor(101.99, 102.0, 102.01)
+    mock_pipeline.predict_quantiles.return_value = (quantiles, None)
+
+    from ai_trade.chronos_predictor import ChronosPredictor
+    predictor = ChronosPredictor(model_name="mock-model")
+    daily = {"005930": _daily_ohlcv([100] * 60)}
+    result = predictor.predict_batch(daily)
+    assert result["005930"].conf > 0.8
+
+
+def test_conf_low_when_distribution_wide(mock_pipeline, mock_torch_cpu):
+    """넓은 distribution → conf ≈ 0."""
+    # last_close=100, q10=70, q50=100, q90=130
+    # returns: q10=-0.3, q50=0.0, q90=0.3
+    # spread = (0.3 - (-0.3)) / max(0.0, 0.001) = 0.6 / 0.001 = 600 → conf = max(0, 1 - 300) = 0
+    quantiles = _mk_quantiles_tensor(70.0, 100.0, 130.0)
+    mock_pipeline.predict_quantiles.return_value = (quantiles, None)
+
+    from ai_trade.chronos_predictor import ChronosPredictor
+    predictor = ChronosPredictor(model_name="mock-model")
+    daily = {"005930": _daily_ohlcv([100] * 60)}
+    result = predictor.predict_batch(daily)
+    assert result["005930"].conf < 0.3
+
+
+def test_return_computed_from_price_relative_to_last_close(mock_pipeline, mock_torch_cpu):
+    """price 예측 → last_close 대비 return 변환. last_close=100, q50=110 → return ≈ +10%."""
+    quantiles = _mk_quantiles_tensor(109.0, 110.0, 111.0)
+    mock_pipeline.predict_quantiles.return_value = (quantiles, None)
+
+    from ai_trade.chronos_predictor import ChronosPredictor
+    predictor = ChronosPredictor(model_name="mock-model")
+    # last close = 100
+    daily = {"005930": _daily_ohlcv(list(range(41, 101)))}  # last = 100
+    result = predictor.predict_batch(daily)
+    assert abs(result["005930"].median - 0.10) < 0.001

ai_trade/tests/test_kis_client.py

@@ -6,7 +6,7 @@ import httpx
 import pytest
 import respx
 
-from signal_v2.kis_client import KISClient
+from ai_trade.kis_client import KISClient
 
 
 @pytest.fixture
@@ -126,3 +126,36 @@ async def test_get_asking_price_computes_bid_ratio(kis_client_factory):
         assert "as_of" in data
     finally:
         await client.close()
+
+
+@respx.mock
+async def test_get_daily_ohlcv_returns_60_bars(kis_client_factory):
+    """KIS daily endpoint returns 60 ascending bars after parsing."""
+    # Build 60 KIS-format daily bars (descending dates as KIS does)
+    sample_output2 = []
+    for i in range(60):
+        # Generate a fake date 60 days ago, descending
+        day = 60 - i
+        sample_output2.append({
+            "stck_bsop_date": f"2026{(((day-1)//30)+1):02d}{(((day-1)%30)+1):02d}",
+            "stck_oprc": "78000", "stck_hgpr": "78500",
+            "stck_lwpr": "77800", "stck_clpr": str(78000 + i),
+            "acml_vol": "12345",
+        })
+
+    respx.get(
+        "https://openapivts.koreainvestment.com:29443/uapi/domestic-stock/v1/quotations/inquire-daily-itemchartprice"
+    ).mock(return_value=httpx.Response(200, json={"output2": sample_output2}))
+
+    client = kis_client_factory()
+    try:
+        bars = await client.get_daily_ohlcv("005930", days=60)
+        # KIS returns descending; client reverses to ascending
+        assert len(bars) == 60
+        # Ascending order: first item has smaller datetime than last
+        assert bars[0]["datetime"] < bars[-1]["datetime"]
+        assert isinstance(bars[0]["open"], int)
+        assert isinstance(bars[0]["close"], int)
+        assert "datetime" in bars[0]
+    finally:
+        await client.close()

ai_trade/tests/test_kis_websocket.py

@@ -7,7 +7,7 @@ import httpx
 import pytest
 import respx
 
-from signal_v2.kis_websocket import KISWebSocket
+from ai_trade.kis_websocket import KISWebSocket
 
 
 BASE_REST = "https://openapivts.koreainvestment.com:29443"

ai_trade/tests/test_main.py

@@ -10,9 +10,9 @@ def test_health_endpoint_returns_status_online(monkeypatch):
     monkeypatch.setenv("WEBAI_API_KEY", "test-secret")
     # Reload modules so they pick up the new env
     import importlib
-    from signal_v2 import config as cfg
+    from ai_trade import config as cfg
     importlib.reload(cfg)
-    from signal_v2 import main as main_mod
+    from ai_trade import main as main_mod
     importlib.reload(main_mod)
     with TestClient(main_mod.app) as client:
         r = client.get("/health")
@@ -24,17 +24,17 @@ def test_health_endpoint_returns_status_online(monkeypatch):
 
 def test_startup_warns_if_webai_api_key_missing(monkeypatch, caplog):
     # Use setenv with empty string + no-op load_dotenv to defeat .env re-read on reload
-    monkeypatch.setattr("signal_v2.config.load_dotenv", lambda *a, **k: None)
+    monkeypatch.setattr("ai_trade.config.load_dotenv", lambda *a, **k: None)
     monkeypatch.setenv("WEBAI_API_KEY", "")
     monkeypatch.setenv("STOCK_API_URL", "https://test.stock.local")
     import importlib
-    from signal_v2 import config as cfg
+    from ai_trade import config as cfg
     importlib.reload(cfg)
     # After reload, load_dotenv reference is fresh — re-patch
-    monkeypatch.setattr("signal_v2.config.load_dotenv", lambda *a, **k: None)
+    monkeypatch.setattr("ai_trade.config.load_dotenv", lambda *a, **k: None)
-    from signal_v2 import main as main_mod
+    from ai_trade import main as main_mod
     importlib.reload(main_mod)
-    with caplog.at_level(logging.WARNING, logger="signal_v2.main"):
+    with caplog.at_level(logging.WARNING, logger="ai_trade.main"):
         with TestClient(main_mod.app) as client:
             client.get("/health")
         assert any("WEBAI_API_KEY" in rec.message for rec in caplog.records)
@@ -42,7 +42,7 @@ def test_startup_warns_if_webai_api_key_missing(monkeypatch, caplog):
 
 def test_startup_warns_if_kis_app_key_missing(monkeypatch, caplog):
     """KIS app_key 미설정 시 startup WARNING (KIS 호출 disabled) — V1 패턴."""
-    monkeypatch.setattr("signal_v2.config.load_dotenv", lambda *a, **k: None)
+    monkeypatch.setattr("ai_trade.config.load_dotenv", lambda *a, **k: None)
     monkeypatch.setenv("STOCK_API_URL", "https://test.stock.local")
     monkeypatch.setenv("WEBAI_API_KEY", "test-secret")
     # V1 pattern: kis_env_type=virtual, both virtual keys empty
@@ -51,12 +51,12 @@ def test_startup_warns_if_kis_app_key_missing(monkeypatch, caplog):
     monkeypatch.setenv("KIS_REAL_APP_KEY", "")
 
     import importlib
-    from signal_v2 import config as cfg
+    from ai_trade import config as cfg
     importlib.reload(cfg)
-    monkeypatch.setattr("signal_v2.config.load_dotenv", lambda *a, **k: None)
+    monkeypatch.setattr("ai_trade.config.load_dotenv", lambda *a, **k: None)
-    from signal_v2 import main as main_mod
+    from ai_trade import main as main_mod
     importlib.reload(main_mod)
-    with caplog.at_level(logging.WARNING, logger="signal_v2.main"):
+    with caplog.at_level(logging.WARNING, logger="ai_trade.main"):
         with TestClient(main_mod.app) as client:
             client.get("/health")
         assert any("KIS" in rec.message and "app_key" in rec.message.lower() for rec in caplog.records)

ai_trade/tests/test_momentum_classifier.py

@@ -0,0 +1,92 @@
+"""Tests for minute momentum classifier."""
+from collections import deque
+
+from ai_trade.momentum_classifier import (
+    aggregate_1min_to_5min, classify_minute_momentum,
+    STRONG_UP, WEAK_UP, NEUTRAL, WEAK_DOWN, STRONG_DOWN,
+)
+
+
+def _bar(open_, high, low, close, volume):
+    return {
+        "datetime": "2026-05-18T09:00:00+09:00",
+        "open": open_, "high": high, "low": low, "close": close, "volume": volume,
+    }
+
+
+def _make_chunks(num_chunks_up: int, num_chunks_total: int, base_vol: int = 1000):
+    """num_chunks_total 개의 5-bar 청크. num_chunks_up 청크는 양봉, 나머지는 음봉.
+    각 청크는 5개 1분봉. 거래량 = base_vol per bar.
+    """
+    bars = []
+    for i in range(num_chunks_total):
+        is_up = i < num_chunks_up
+        o, c = (100, 110) if is_up else (110, 100)
+        for j in range(5):
+            bars.append(_bar(o, max(o, c) + 5, min(o, c) - 5, c, base_vol))
+    return bars
+
+
+def test_strong_up_5_consecutive_green_with_high_volume():
+    """직전 5개 5분봉 모두 양봉 + 거래량 1.5x → STRONG_UP."""
+    # 60분 (12 5분봉) 데이터: 7 normal + 5 high-vol up
+    older = _make_chunks(num_chunks_up=3, num_chunks_total=7, base_vol=1000)
+    recent = _make_chunks(num_chunks_up=5, num_chunks_total=5, base_vol=2500)
+    minute_bars = deque(older + recent, maxlen=60)
+    assert classify_minute_momentum(minute_bars) == STRONG_UP
+
+
+def test_weak_up_3of5_green_normal_volume():
+    """직전 5개 5분봉 중 3-4개 양봉 + 거래량 ≥ 1.0x → WEAK_UP."""
+    older = _make_chunks(num_chunks_up=3, num_chunks_total=7, base_vol=1000)
+    # 5 chunks: 3 up + 2 down, normal vol
+    recent_up = _make_chunks(num_chunks_up=3, num_chunks_total=3, base_vol=1000)
+    recent_down = _make_chunks(num_chunks_up=0, num_chunks_total=2, base_vol=1000)
+    minute_bars = deque(older + recent_up + recent_down, maxlen=60)
+    assert classify_minute_momentum(minute_bars) == WEAK_UP
+
+
+def test_neutral_mixed():
+    """up_count=2, vol normal → NEUTRAL (rule 미해당)."""
+    older = _make_chunks(num_chunks_up=3, num_chunks_total=7, base_vol=1000)
+    recent_up = _make_chunks(num_chunks_up=2, num_chunks_total=2, base_vol=1000)
+    recent_down = _make_chunks(num_chunks_up=0, num_chunks_total=3, base_vol=1000)
+    minute_bars = deque(older + recent_up + recent_down, maxlen=60)
+    # up_count=2, vol_mult=1.0 → 어느 분기 조건도 만족 안 함 → NEUTRAL
+    assert classify_minute_momentum(minute_bars) == NEUTRAL
+
+
+def test_weak_down_low_green_low_volume():
+    """up_count <= 2 + vol < 1.0 → WEAK_DOWN."""
+    older = _make_chunks(num_chunks_up=3, num_chunks_total=7, base_vol=1000)
+    recent_up = _make_chunks(num_chunks_up=1, num_chunks_total=1, base_vol=500)
+    recent_down = _make_chunks(num_chunks_up=0, num_chunks_total=4, base_vol=500)
+    minute_bars = deque(older + recent_up + recent_down, maxlen=60)
+    # recent 5 chunks avg vol = 500, long 12 avg ≈ (7*1000 + 5*500) / 12 ≈ 791 → vol_mult ≈ 0.63
+    assert classify_minute_momentum(minute_bars) == WEAK_DOWN
+
+
+def test_strong_down_5_consecutive_red_high_volume():
+    """직전 5개 5분봉 모두 음봉 + 거래량 1.5x → STRONG_DOWN."""
+    older = _make_chunks(num_chunks_up=3, num_chunks_total=7, base_vol=1000)
+    recent = _make_chunks(num_chunks_up=0, num_chunks_total=5, base_vol=2500)
+    minute_bars = deque(older + recent, maxlen=60)
+    assert classify_minute_momentum(minute_bars) == STRONG_DOWN
+
+
+def test_aggregate_1min_to_5min_correctness():
+    """5 1분봉 → 1개 5분봉 — open/close/high/low/volume 정확."""
+    bars = [
+        _bar(100, 105, 99, 102, 1000),
+        _bar(102, 108, 101, 107, 1500),
+        _bar(107, 110, 105, 106, 800),
+        _bar(106, 109, 104, 108, 1200),
+        _bar(108, 112, 107, 111, 900),
+    ]
+    result = aggregate_1min_to_5min(bars)
+    assert len(result) == 1
+    assert result[0]["open"] == 100  # 첫 bar
+    assert result[0]["close"] == 111  # 마지막 bar
+    assert result[0]["high"] == 112  # max
+    assert result[0]["low"] == 99    # min
+    assert result[0]["volume"] == 5400  # sum

ai_trade/tests/test_pull_worker.py

@@ -0,0 +1,131 @@
+"""Tests for pull_worker (Phase 3a additions)."""
+from collections import deque
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from ai_trade.state import PollState
+
+
+async def test_minute_polling_cycle_updates_state_minute_bars():
+    """KIS REST mock 의 분봉 데이터가 state.minute_bars[ticker] deque 에 들어간다."""
+    from ai_trade.pull_worker import _run_kis_minute_cycle
+
+    state = PollState()
+    state.portfolio = {"holdings": [{"ticker": "005930"}, {"ticker": "000660"}]}
+    state.screener_preview = {
+        "items": [{"ticker": "005930"}, {"ticker": "035720"}]
+    }
+
+    kis_client_mock = MagicMock()
+    kis_client_mock.get_minute_ohlcv = AsyncMock(side_effect=[
+        [{"datetime": "2026-05-18T09:00:00+09:00", "open": 78000,
+          "high": 78500, "low": 77900, "close": 78300, "volume": 12345}],
+        [{"datetime": "2026-05-18T09:00:00+09:00", "open": 180000,
+          "high": 181000, "low": 179800, "close": 180500, "volume": 5000}],
+        [{"datetime": "2026-05-18T09:00:00+09:00", "open": 51000,
+          "high": 51200, "low": 50800, "close": 51100, "volume": 8000}],
+    ])
+    kis_client_mock.get_asking_price = AsyncMock(return_value={
+        "bid_total": 600, "ask_total": 400, "bid_ratio": 0.6,
+        "current_price": 51100, "as_of": "2026-05-18T09:00:30+09:00",
+    })
+
+    await _run_kis_minute_cycle(kis_client_mock, state)
+
+    # 3 unique tickers (005930, 000660, 035720)
+    assert "005930" in state.minute_bars
+    assert "000660" in state.minute_bars
+    assert "035720" in state.minute_bars
+    assert len(state.minute_bars["005930"]) >= 1
+    # asking_price 만 screener-only ticker (035720) 에 들어가야 함
+    # (portfolio = 005930, 000660 는 WebSocket 으로 들어옴)
+    assert "035720" in state.asking_price
+
+
+def test_websocket_message_updates_state_asking_price():
+    """WebSocket callback factory → state.asking_price 갱신."""
+    from ai_trade.pull_worker import make_asking_price_callback
+
+    state = PollState()
+    cb = make_asking_price_callback(state)
+    cb("005930", {"bid_total": 1000, "ask_total": 800, "bid_ratio": 0.555,
+                  "current_price": 78500, "as_of": "2026-05-18T10:00:00+09:00"})
+    assert state.asking_price["005930"]["bid_total"] == 1000
+    assert "asking_price/005930" in state.last_updated
+
+
+async def test_post_close_cycle_updates_chronos_predictions():
+    """mock kis + mock chronos → state.chronos_predictions + state.daily_ohlcv 갱신."""
+    from unittest.mock import AsyncMock, MagicMock
+    from ai_trade.pull_worker import _run_post_close_cycle
+    from ai_trade.chronos_predictor import ChronosPrediction
+    from ai_trade.state import PollState
+
+    state = PollState()
+    state.portfolio = {"holdings": [{"ticker": "005930"}]}
+    state.screener_preview = {"items": [{"ticker": "000660"}]}
+
+    kis_mock = MagicMock()
+    daily_005930 = [{"datetime": f"2026-05-{i+1:02d}", "open": 100, "high": 105,
+                     "low": 95, "close": 100 + i, "volume": 1000} for i in range(60)]
+    daily_000660 = [{"datetime": f"2026-05-{i+1:02d}", "open": 200, "high": 210,
+                     "low": 190, "close": 200 + i, "volume": 2000} for i in range(60)]
+    # _run_post_close_cycle iterates tickers and calls get_daily_ohlcv per ticker.
+    # Order depends on set() so use side_effect mapping if possible, otherwise list.
+    async def fake_daily(ticker, days=60):
+        if ticker == "005930":
+            return daily_005930
+        if ticker == "000660":
+            return daily_000660
+        return []
+    kis_mock.get_daily_ohlcv = AsyncMock(side_effect=fake_daily)
+
+    chronos_mock = MagicMock()
+    chronos_mock.predict_batch = MagicMock(return_value={
+        "005930": ChronosPrediction(0.02, -0.01, 0.04, 0.85, "2026-05-18T16:00:00+09:00"),
+        "000660": ChronosPrediction(0.03, -0.02, 0.06, 0.75, "2026-05-18T16:00:00+09:00"),
+    })
+
+    await _run_post_close_cycle(kis_mock, chronos_mock, state)
+
+    assert "005930" in state.chronos_predictions
+    assert "000660" in state.chronos_predictions
+    assert state.chronos_predictions["005930"]["median"] == 0.02
+    assert state.chronos_predictions["005930"]["conf"] == 0.85
+    assert "005930" in state.daily_ohlcv
+    assert "chronos/005930" in state.last_updated
+
+
+def test_poll_loop_calls_generate_signals_after_cycle(monkeypatch):
+    """Phase 4: generate_signals 가 cycle 후 state.signals 를 갱신한다."""
+    from unittest.mock import MagicMock
+    from ai_trade.state import PollState
+    from ai_trade.signal_generator import generate_signals
+
+    state = PollState()
+    state.portfolio = {"holdings": [{
+        "ticker": "005930", "name": "삼성전자",
+        "avg_price": 75000, "current_price": 69000,
+        "pnl_pct": -0.08, "profit_rate": -8.0,
+        "quantity": 100, "broker": "키움",
+    }]}
+    state.screener_preview = {"items": []}
+
+    dedup = MagicMock()
+    dedup.is_recent.return_value = False
+
+    settings = MagicMock()
+    settings.stop_loss_pct = -0.07
+    settings.take_profit_pct = 0.15
+    settings.chronos_spread_threshold = 0.6
+    settings.asking_bid_ratio_threshold = 0.6
+    settings.confidence_threshold = 0.7
+    settings.min_momentum_for_buy = "strong_up"
+
+    generate_signals(state, dedup, settings)
+
+    assert "005930" in state.signals
+    assert state.signals["005930"]["action"] == "sell"
+    assert state.signals["005930"]["confidence_webai"] == 1.0
+    dedup.record.assert_called_with("005930", "sell", confidence=1.0)

ai_trade/tests/test_rate_limit.py

@@ -2,7 +2,7 @@
 from datetime import datetime, timedelta
 from zoneinfo import ZoneInfo
 
-from signal_v2.rate_limit import SignalDedup
+from ai_trade.rate_limit import SignalDedup
 
 KST = ZoneInfo("Asia/Seoul")
 
@@ -24,11 +24,11 @@ def test_is_recent_returns_false_after_24h(tmp_dedup_db, monkeypatch):
     now = datetime.now(KST)
     fake_now = now - timedelta(hours=25)
     monkeypatch.setattr(
-        "signal_v2.rate_limit._now_iso", lambda: fake_now.isoformat()
+        "ai_trade.rate_limit._now_iso", lambda: fake_now.isoformat()
     )
     dedup.record("005930", "buy", confidence=0.82)
     # Reset to real now for is_recent check
     monkeypatch.setattr(
-        "signal_v2.rate_limit._now_iso", lambda: now.isoformat()
+        "ai_trade.rate_limit._now_iso", lambda: now.isoformat()
     )
     assert dedup.is_recent("005930", "buy", within_hours=24) is False

ai_trade/tests/test_scheduler.py

@@ -3,7 +3,7 @@ from datetime import datetime
 
 import pytest
 
-from signal_v2.scheduler import _next_interval, _is_market_day, KST
+from ai_trade.scheduler import _next_interval, _is_market_day, KST
 
 
 def _kst(year, month, day, hour, minute=0):

ai_trade/tests/test_signal_generator.py

@@ -0,0 +1,172 @@
+"""Tests for signal_generator."""
+from unittest.mock import MagicMock
+
+import pytest
+
+from ai_trade.signal_generator import generate_signals
+from ai_trade.state import PollState
+
+
+def _settings(**overrides):
+    """Build a Settings-like object for tests (avoid env)."""
+    defaults = dict(
+        stop_loss_pct=-0.07,
+        take_profit_pct=0.15,
+        chronos_spread_threshold=0.6,
+        asking_bid_ratio_threshold=0.6,
+        confidence_threshold=0.7,
+        min_momentum_for_buy="strong_up",
+    )
+    defaults.update(overrides)
+    m = MagicMock()
+    for k, v in defaults.items():
+        setattr(m, k, v)
+    return m
+
+
+def _make_state_with_buy_candidate(
+    ticker="005930", name="삼성전자",
+    chronos_median=0.02, chronos_q10=-0.01, chronos_q90=0.04, chronos_conf=0.85,
+    momentum="strong_up", bid_ratio=0.7, current_price=78500,
+):
+    state = PollState()
+    state.screener_preview = {"items": [{"ticker": ticker, "name": name}]}
+    state.chronos_predictions[ticker] = {
+        "median": chronos_median, "q10": chronos_q10, "q90": chronos_q90,
+        "conf": chronos_conf, "as_of": "2026-05-17T16:00:00+09:00",
+    }
+    state.minute_momentum[ticker] = momentum
+    state.asking_price[ticker] = {
+        "bid_total": int(bid_ratio * 1000),
+        "ask_total": int((1 - bid_ratio) * 1000),
+        "bid_ratio": bid_ratio,
+        "current_price": current_price,
+        "as_of": "2026-05-17T16:00:01+09:00",
+    }
+    return state
+
+
+def _make_state_with_holding(
+    ticker="005930", name="삼성전자",
+    pnl_pct=0.0, avg_price=75000, current_price=75000,
+):
+    state = PollState()
+    state.portfolio = {"holdings": [{
+        "ticker": ticker, "name": name,
+        "avg_price": avg_price, "current_price": current_price,
+        "pnl_pct": pnl_pct, "profit_rate": pnl_pct * 100,
+        "quantity": 100, "broker": "키움",
+    }]}
+    state.screener_preview = {"items": []}
+    return state
+
+
+@pytest.fixture
+def dedup_mock():
+    d = MagicMock()
+    d.is_recent.return_value = False
+    return d
+
+
+def test_buy_signal_when_all_conditions_pass_and_confidence_high(dedup_mock):
+    state = _make_state_with_buy_candidate()
+    generate_signals(state, dedup_mock, _settings())
+    assert "005930" in state.signals
+    sig = state.signals["005930"]
+    assert sig["action"] == "buy"
+    assert sig["confidence_webai"] > 0.7
+    dedup_mock.record.assert_called()
+
+
+def test_silent_when_chronos_median_negative(dedup_mock):
+    state = _make_state_with_buy_candidate(chronos_median=-0.01)
+    generate_signals(state, dedup_mock, _settings())
+    assert "005930" not in state.signals
+
+
+def test_silent_when_distribution_spread_too_wide(dedup_mock):
+    # spread = q90 - q10 = 0.5 - (-0.5) = 1.0 > 0.6 → hard gate fails
+    state = _make_state_with_buy_candidate(
+        chronos_median=0.001, chronos_q10=-0.5, chronos_q90=0.5,
+    )
+    generate_signals(state, dedup_mock, _settings())
+    assert "005930" not in state.signals
+
+
+def test_silent_when_momentum_not_strong_up(dedup_mock):
+    state = _make_state_with_buy_candidate(momentum="weak_up")
+    generate_signals(state, dedup_mock, _settings())
+    assert "005930" not in state.signals
+
+
+def test_silent_when_bid_ratio_below_threshold(dedup_mock):
+    state = _make_state_with_buy_candidate(bid_ratio=0.5)
+    generate_signals(state, dedup_mock, _settings())
+    assert "005930" not in state.signals
+
+
+def test_silent_when_confidence_below_threshold(dedup_mock):
+    # chronos_conf low + rank=20 → confidence < 0.7
+    state = _make_state_with_buy_candidate(chronos_conf=0.3)
+    # add 19 fake items to push 005930 rank to 20
+    state.screener_preview["items"] = (
+        [{"ticker": f"FAKE{i:03d}"} for i in range(19)]
+        + [{"ticker": "005930", "name": "삼성전자"}]
+    )
+    generate_signals(state, dedup_mock, _settings())
+    # confidence_webai = 0.3*0.5 + 1.0*0.3 + 0.05*0.2 = 0.46 < 0.7
+    assert "005930" not in state.signals
+
+
+def test_sell_signal_when_stop_loss_triggered(dedup_mock):
+    state = _make_state_with_holding(pnl_pct=-0.08, current_price=69000, avg_price=75000)
+    generate_signals(state, dedup_mock, _settings())
+    assert "005930" in state.signals
+    sig = state.signals["005930"]
+    assert sig["action"] == "sell"
+    assert sig["confidence_webai"] == 1.0
+    assert sig["pnl_pct"] == -0.08
+
+
+def test_sell_signal_when_take_profit_triggered(dedup_mock):
+    state = _make_state_with_holding(pnl_pct=0.16, current_price=87000, avg_price=75000)
+    generate_signals(state, dedup_mock, _settings())
+    assert "005930" in state.signals
+    sig = state.signals["005930"]
+    assert sig["action"] == "sell"
+    assert sig["confidence_webai"] == 0.6
+
+
+def test_silent_when_dedup_recently_sent(dedup_mock):
+    state = _make_state_with_buy_candidate()
+    dedup_mock.is_recent.return_value = True
+    generate_signals(state, dedup_mock, _settings())
+    assert "005930" not in state.signals
+    dedup_mock.record.assert_not_called()
+
+
+def test_sell_signal_triggers_on_anomaly_path(dedup_mock):
+    """Anomaly sell: median < -1%, momentum strong_down, low bid_ratio, confidence > threshold."""
+    state = PollState()
+    state.portfolio = {"holdings": [{
+        "ticker": "005930", "name": "삼성전자",
+        "avg_price": 75000, "current_price": 70000,
+        "pnl_pct": -0.067,  # within stop_loss tolerance (default -0.07): NOT triggering stop_loss
+        "quantity": 100, "broker": "키움",
+    }]}
+    state.screener_preview = {"items": []}
+    state.chronos_predictions["005930"] = {
+        "median": -0.025, "q10": -0.05, "q90": 0.005, "conf": 0.85,
+    }
+    state.minute_momentum["005930"] = "strong_down"
+    state.asking_price["005930"] = {"current_price": 70000, "bid_ratio": 0.30}
+    # bid_ratio 0.30 < (1 - 0.6) = 0.4 → anomaly bid_ratio gate passes
+    # confidence = 0.85*0.5 + 1.0*0.3 + 1.0*0.2 = 0.425 + 0.3 + 0.2 = 0.925 > 0.7
+
+    generate_signals(state, dedup_mock, _settings())
+
+    assert "005930" in state.signals
+    sig = state.signals["005930"]
+    assert sig["action"] == "sell"
+    assert sig["context"]["sell_reason"] == "anomaly"
+    assert sig["confidence_webai"] > 0.7

ai_trade/tests/test_stock_client.py

@@ -4,7 +4,7 @@ import logging
 import pytest
 import httpx
 
-from signal_v2.stock_client import StockClient
+from ai_trade.stock_client import StockClient
 
 
 BASE_URL = "https://test.stock.local"
@@ -34,7 +34,7 @@ async def test_get_portfolio_normal_returns_dict_with_pnl_pct(mock_stock_api):
 
 
 async def test_get_portfolio_uses_cache_within_ttl(mock_stock_api):
-    """60s TTL 내 두번째 호출 = mock 콜 1회."""
+    """180s TTL 내 두번째 호출 = mock 콜 1회."""
     route = mock_stock_api.get("/api/webai/portfolio").mock(
         return_value=httpx.Response(
             200, json={"holdings": [], "cash": [], "summary": {}}
@@ -56,16 +56,16 @@ async def test_get_portfolio_refetches_after_ttl_expiry(mock_stock_api, monkeypa
             200, json={"holdings": [], "cash": [], "summary": {}}
         )
     )
-    # Fake clock: starts at 0, jumps to 61 between calls
+    # Fake clock: starts at 0, jumps past portfolio TTL (180s) between calls
     fake_time = [0.0]
     monkeypatch.setattr(
-        "signal_v2.stock_client.time.monotonic", lambda: fake_time[0]
+        "ai_trade.stock_client.time.monotonic", lambda: fake_time[0]
     )
 
     client = StockClient(BASE_URL, API_KEY)
     try:
         await client.get_portfolio()
-        fake_time[0] = 61.0  # 60s TTL 만료
+        fake_time[0] = 181.0  # 180s TTL 만료
         await client.get_portfolio()
         assert route.call_count == 2
     finally:
@@ -137,7 +137,7 @@ async def test_get_portfolio_falls_back_to_stale_on_all_failures(
     # Patch time.monotonic BEFORE first call so cached timestamp uses fake clock
     fake_time = [0.0]
     monkeypatch.setattr(
-        "signal_v2.stock_client.time.monotonic", lambda: fake_time[0]
+        "ai_trade.stock_client.time.monotonic", lambda: fake_time[0]
     )
 
     # First call succeeds
@@ -152,13 +152,13 @@ async def test_get_portfolio_falls_back_to_stale_on_all_failures(
         first = await client.get_portfolio()
         assert first["holdings"][0]["ticker"] == "005930"
 
-        # Advance fake clock past TTL (60s) so cache is stale
+        # Advance fake clock past TTL (180s) so cache is stale
-        fake_time[0] = 61.0
+        fake_time[0] = 181.0
 
         # Now mock to return 500s persistently
         route1.mock(return_value=httpx.Response(500, text="server error"))
 
-        with caplog.at_level(logging.WARNING, logger="signal_v2.stock_client"):
+        with caplog.at_level(logging.WARNING, logger="ai_trade.stock_client"):
             result = await client.get_portfolio()
             assert result["holdings"][0]["ticker"] == "005930"  # stale data returned
             assert any(

ai_trade/tests/test_stock_client_ttl.py

@@ -0,0 +1,18 @@
+# tests/test_stock_client_ttl.py
+"""SP-A1 회귀 — _TTL이 NAS 부담 완화를 위한 값으로 설정되어 있어야 함."""
+from ai_trade.stock_client import _TTL
+
+
+def test_portfolio_ttl_is_180s():
+    """portfolio TTL은 180초 이상 (3분 폴링에서 1회 fetch가 3 폴링 커버)."""
+    assert _TTL["portfolio"] >= 180.0
+
+
+def test_news_sentiment_ttl_is_600s():
+    """news-sentiment TTL은 600초 이상 (10분, 뉴스 sentiment는 자주 안 바뀜)."""
+    assert _TTL["news-sentiment"] >= 600.0
+
+
+def test_screener_preview_ttl_is_300s():
+    """screener-preview TTL은 300초 이상 (5분, Top-20은 분 단위로 거의 안 바뀜)."""
+    assert _TTL["screener-preview"] >= 300.0

legacy/signal_v1/DEPRECATED.md

@@ -0,0 +1,26 @@
+# signal_v1 — DEPRECATED
+
+> **2026-05-19부터 사용 안 함.** 신규 작업 금지. 모든 트레이딩은 `web-ai/ai_trade/` (구 `signal_v2`) 에서 진행.
+
+## 폐기 사유
+
+- V2 (`ai_trade`) Phase 4 완료 — Chronos-2 zero-shot 1일 수익률 + 분봉 모멘텀 + 5-state classifier + sell-first 우선순위 + 매수 hard gate가 V1 (LSTM 7-features + Gemini Flash) 보다 정확도·확장성·해석가능성에서 우위
+- V1 + V2 동시 운영 시 KIS API rate limit 충돌
+- NAS 인바운드 polling 부담 (web-ai → NAS API) 의 50% 차지
+
+## 향후 처리
+
+- 디렉토리를 `legacy/signal_v1/`로 이동 예정 (현재 file lock 풀린 후 처리)
+- `start.bat` 진입점은 이미 `legacy/start_v1.bat`으로 이동 → 자동 시작 차단됨
+- DSM Scheduler 등 외부 trigger에 V1 startup 등록되어 있다면 해제 필요 (박재오 확인)
+
+## 활용 (필요 시)
+
+- 코드 참고용 (LSTM 모델 구조, Telegram bot 인터페이스, KIS 자동주문 패턴)
+- 별도 backtest 실행은 가능 (`backtest_runner.py`) — 단 운영 자동 실행 X
+
+## 관련 문서
+
+- 신 운영 가이드: `../ai_trade/`
+- web-ai 통합 가이드: `../CLAUDE.md`
+- V1 vs V2 진단: `../CHECK_POINT.md`

requirements.txt

@@ -7,3 +7,7 @@ pytest>=8.0
 pytest-asyncio>=0.23
 respx>=0.21
 websockets>=12
+# Phase 3b dependencies (Chronos-2 + ML)
+transformers>=4.40
+chronos-forecasting>=1.4
+# torch: typically already installed via V1 venv; if not, install with CUDA support manually

services/docker-compose.yml

@@ -0,0 +1,124 @@
+name: web-ai-services
+
+services:
+  insta-render:
+    build:
+      context: ./insta-render
+    container_name: insta-render
+    restart: unless-stopped
+    ports:
+      - "18710:8000"
+    environment:
+      - TZ=Asia/Seoul
+      - REDIS_URL=${REDIS_URL:-redis://192.168.45.54:6379}
+      - NAS_BASE_URL=${NAS_BASE_URL:-http://192.168.45.54:18700}
+      - INTERNAL_API_KEY=${INTERNAL_API_KEY:-}
+      - INSTA_MEDIA_ROOT=${INSTA_MEDIA_ROOT:-/mnt/nas/webpage/data/insta}
+      - INSTA_MEDIA_URL_PREFIX=${INSTA_MEDIA_URL_PREFIX:-/media/insta}
+      - CARD_TEMPLATE_DIR=/app/templates
+    volumes:
+      - /mnt/nas/webpage/data/insta:/mnt/nas/webpage/data/insta
+    healthcheck:
+      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
+      interval: 60s
+      timeout: 5s
+      retries: 3
+
+  music-render:
+    build:
+      context: ./music-render
+    container_name: music-render
+    restart: unless-stopped
+    ports:
+      - "18711:8000"
+    environment:
+      - TZ=Asia/Seoul
+      - REDIS_URL=${REDIS_URL:-redis://192.168.45.54:6379}
+      - NAS_BASE_URL=${NAS_BASE_URL:-http://192.168.45.54:18600}
+      - INTERNAL_API_KEY=${INTERNAL_API_KEY:-}
+      - SUNO_API_KEY=${SUNO_API_KEY:-}
+      - MUSIC_AI_SERVER_URL=${MUSIC_AI_SERVER_URL:-http://host.docker.internal:8765}
+      - MUSIC_MEDIA_ROOT=${MUSIC_MEDIA_ROOT:-/mnt/nas/webpage/data/music}
+      - MUSIC_MEDIA_URL_PREFIX=${MUSIC_MEDIA_URL_PREFIX:-/media/music}
+    extra_hosts:
+      - "host.docker.internal:host-gateway"
+    volumes:
+      - /mnt/nas/webpage/data/music:/mnt/nas/webpage/data/music
+    healthcheck:
+      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
+      interval: 60s
+      timeout: 5s
+      retries: 3
+
+  video-render:
+    build:
+      context: ./video-render
+    container_name: video-render
+    restart: unless-stopped
+    ports:
+      - "18712:8000"
+    environment:
+      - TZ=Asia/Seoul
+      - REDIS_URL=${REDIS_URL:-redis://192.168.45.54:6379}
+      - NAS_BASE_URL=${NAS_BASE_URL:-http://192.168.45.54:18801}
+      - INTERNAL_API_KEY=${INTERNAL_API_KEY:-}
+      - OPENAI_API_KEY=${OPENAI_API_KEY:-}
+      - GEMINI_API_KEY=${GEMINI_API_KEY:-}
+      - KLING_ACCESS_KEY=${KLING_ACCESS_KEY:-}
+      - KLING_SECRET_KEY=${KLING_SECRET_KEY:-}
+      - SEEDANCE_API_KEY=${SEEDANCE_API_KEY:-}
+      - VIDEO_MEDIA_ROOT=${VIDEO_MEDIA_ROOT:-/mnt/nas/webpage/data/video}
+      - VIDEO_MEDIA_URL_PREFIX=${VIDEO_MEDIA_URL_PREFIX:-/media/video}
+    volumes:
+      - /mnt/nas/webpage/data/video:/mnt/nas/webpage/data/video
+    healthcheck:
+      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
+      interval: 60s
+      timeout: 5s
+      retries: 3
+
+  task-watcher:
+    build:
+      context: ./task-watcher
+    container_name: task-watcher
+    restart: unless-stopped
+    ports:
+      - "18713:8000"
+    environment:
+      - TZ=Asia/Seoul
+      - REDIS_URL=${REDIS_URL:-redis://192.168.45.54:6379}
+      - STOCK_BASE_URL=${STOCK_BASE_URL:-http://192.168.45.54:18500}
+      - TRADING_START=${TRADING_START:-07:00}
+      - TRADING_END=${TRADING_END:-16:30}
+    healthcheck:
+      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
+      interval: 60s
+      timeout: 5s
+      retries: 3
+
+  image-render:
+    build:
+      context: ./image-render
+    container_name: image-render
+    restart: unless-stopped
+    ports:
+      - "18714:8000"
+    environment:
+      - TZ=Asia/Seoul
+      - REDIS_URL=${REDIS_URL:-redis://192.168.45.54:6379}
+      - NAS_BASE_URL=${NAS_BASE_URL:-http://192.168.45.54:18802}
+      - INTERNAL_API_KEY=${INTERNAL_API_KEY:-}
+      - OPENAI_API_KEY=${OPENAI_API_KEY:-}
+      - GEMINI_API_KEY=${GEMINI_API_KEY:-}
+      - COMFYUI_URL=${COMFYUI_URL:-http://host.docker.internal:8188}
+      - FLUX_BLOCK_TRADING_HOURS=${FLUX_BLOCK_TRADING_HOURS:-1}
+      - IMAGE_MEDIA_ROOT=${IMAGE_MEDIA_ROOT:-/mnt/nas/webpage/data/image}
+    extra_hosts:
+      - "host.docker.internal:host-gateway"
+    volumes:
+      - /mnt/nas/webpage/data/image:/mnt/nas/webpage/data/image
+    healthcheck:
+      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
+      interval: 60s
+      timeout: 5s
+      retries: 3

services/image-render/Dockerfile

@@ -0,0 +1,16 @@
+FROM python:3.12-slim-bookworm
+ENV PYTHONUNBUFFERED=1
+
+WORKDIR /app
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ca-certificates \
+ && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir --timeout 600 --retries 5 -r requirements.txt
+
+COPY . .
+
+EXPOSE 8000
+CMD ["python", "-m", "uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000", "--workers", "1"]

services/image-render/env.example

@@ -0,0 +1,18 @@
+# Redis (NAS)
+REDIS_URL=redis://192.168.45.54:6379
+
+# NAS image-lab webhook
+NAS_BASE_URL=http://192.168.45.54:18802
+INTERNAL_API_KEY=replace-me
+
+# API provider keys (worker reports failed if missing)
+OPENAI_API_KEY=
+GEMINI_API_KEY=
+# Seedance key not used by image-render
+
+# FLUX local
+COMFYUI_URL=http://host.docker.internal:8188
+FLUX_BLOCK_TRADING_HOURS=1
+
+# NAS SMB mount target (image-render writes to this, NAS reads via /media/image/)
+IMAGE_MEDIA_ROOT=/mnt/nas/webpage/data/image

services/image-render/main.py

@@ -0,0 +1,36 @@
+"""image-render FastAPI entry — health + lifespan (worker loop spawn)."""
+from __future__ import annotations
+
+import asyncio
+import logging
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI
+
+import worker
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s %(name)s %(levelname)s %(message)s")
+logger = logging.getLogger(__name__)
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    worker_task = asyncio.create_task(worker.worker_loop())
+    logger.info("image-render lifespan 시작")
+    try:
+        yield
+    finally:
+        worker_task.cancel()
+        try:
+            await worker_task
+        except asyncio.CancelledError:
+            pass
+        logger.info("image-render lifespan 종료")
+
+
+app = FastAPI(lifespan=lifespan)
+
+
+@app.get("/health")
+def health():
+    return {"ok": True, "service": "image-render"}

services/image-render/nas_client.py

@@ -0,0 +1,54 @@
+"""NAS webhook 어댑터 — Windows worker → NAS image-lab HTTP 위임.
+
+video-render nas_client 복제 (call-time os.getenv으로 테스트 격리).
+"""
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any, Dict, Optional
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+_TIMEOUT = 10.0
+
+
+def _post(payload: Dict[str, Any]) -> None:
+    nas_base_url = os.getenv("NAS_BASE_URL", "http://192.168.45.54:18802")
+    internal_api_key = os.getenv("INTERNAL_API_KEY", "")
+    url = f"{nas_base_url}/api/internal/image/update"
+    try:
+        r = httpx.post(
+            url,
+            headers={"X-Internal-Key": internal_api_key},
+            json=payload,
+            timeout=_TIMEOUT,
+        )
+        if r.status_code != 200:
+            logger.error("webhook %s returned %d: %s",
+                         payload.get("task_id"), r.status_code, r.text[:200])
+    except Exception:
+        logger.exception("webhook %s 호출 실패", payload.get("task_id"))
+
+
+def webhook_update_task(
+    task_id: str,
+    status: str,
+    progress: int,
+    message: str = "",
+    image_url: Optional[str] = None,
+    error: Optional[str] = None,
+) -> None:
+    payload: Dict[str, Any] = {
+        "task_id": task_id,
+        "status": status,
+        "progress": progress,
+        "message": message,
+    }
+    if image_url is not None:
+        payload["image_url"] = image_url
+    if error is not None:
+        payload["error"] = error
+    _post(payload)

services/image-render/providers/_media.py

@@ -0,0 +1,18 @@
+"""b64 이미지 → NAS SMB 경로 저장 → /media/image URL 반환."""
+from __future__ import annotations
+
+import base64
+import os
+import uuid
+
+IMAGE_MEDIA_ROOT = os.getenv("IMAGE_MEDIA_ROOT", "/mnt/nas/webpage/data/image")
+IMAGE_MEDIA_URL_PREFIX = os.getenv("IMAGE_MEDIA_URL_PREFIX", "/media/image")
+
+
+def save_b64_png(task_id: str, b64_data: str) -> str:
+    os.makedirs(IMAGE_MEDIA_ROOT, exist_ok=True)
+    fname = f"{task_id}-{uuid.uuid4().hex[:8]}.png"
+    path = os.path.join(IMAGE_MEDIA_ROOT, fname)
+    with open(path, "wb") as f:
+        f.write(base64.b64decode(b64_data))
+    return f"{IMAGE_MEDIA_URL_PREFIX}/{fname}"

services/image-render/providers/flux.py

@@ -0,0 +1,79 @@
+"""FLUX 로컬 — ComfyUI HTTP API.
+
+POST {COMFYUI_URL}/prompt  (workflow JSON) → prompt_id
+GET  {COMFYUI_URL}/history/{prompt_id} → outputs → image filename
+GET  {COMFYUI_URL}/view?filename=... → PNG bytes → b64
+
+워크플로우 JSON은 `flux_workflow.json` (ComfyUI UI에서 "Save (API Format)"로 export, CLIPTextEncode 노드 text를 "%PROMPT%"로 수동 치환). 박재오 산출물.
+"""
+from __future__ import annotations
+
+import base64, json, logging, os, time
+from datetime import datetime, timezone, timedelta
+import requests
+
+from nas_client import webhook_update_task
+from providers._media import save_b64_png
+
+logger = logging.getLogger(__name__)
+COMFYUI_URL = os.getenv("COMFYUI_URL", "http://127.0.0.1:8188")
+WORKFLOW_PATH = os.path.join(os.path.dirname(__file__), "flux_workflow.json")
+POLL_INTERVAL = 2
+POLL_MAX = 120
+
+
+def _is_trading_hours() -> bool:
+    kst = timezone(timedelta(hours=9))
+    now = datetime.now(kst)
+    if now.weekday() >= 5:
+        return False
+    return (now.hour, now.minute) >= (9, 0) and (now.hour, now.minute) <= (15, 30)
+
+
+def _load_workflow(prompt: str, size: str) -> dict:
+    with open(WORKFLOW_PATH, encoding="utf-8") as f:
+        wf = json.load(f)
+    # CLIPTextEncode 노드의 text를 prompt로 치환 (workflow에 "%PROMPT%" placeholder 사용)
+    raw = json.dumps(wf).replace("%PROMPT%", prompt.replace('"', "'"))
+    return json.loads(raw)
+
+
+def _submit_prompt(workflow: dict) -> str:
+    r = requests.post(f"{COMFYUI_URL}/prompt", json={"prompt": workflow}, timeout=30)
+    r.raise_for_status()
+    return r.json()["prompt_id"]
+
+
+def _poll_image_b64(prompt_id: str):
+    for _ in range(POLL_MAX):
+        h = requests.get(f"{COMFYUI_URL}/history/{prompt_id}", timeout=10)
+        data = h.json().get(prompt_id)
+        if data and data.get("outputs"):
+            for node_out in data["outputs"].values():
+                for img in node_out.get("images", []):
+                    view = requests.get(f"{COMFYUI_URL}/view",
+                                        params={"filename": img["filename"], "subfolder": img.get("subfolder", ""), "type": img.get("type", "output")},
+                                        timeout=30)
+                    view.raise_for_status()
+                    return base64.b64encode(view.content).decode()
+        time.sleep(POLL_INTERVAL)
+    return None
+
+
+def run_flux_generation(task_id: str, params: dict) -> None:
+    try:
+        if os.getenv("FLUX_BLOCK_TRADING_HOURS") == "1" and _is_trading_hours():
+            webhook_update_task(task_id, "failed", 0, "", error="장중 GPU 보호 — FLUX 거부 (API provider 사용 권장)")
+            return
+        webhook_update_task(task_id, "processing", 10, "FLUX (ComfyUI) 생성 중...")
+        wf = _load_workflow(params["prompt"], params.get("size") or "1024x1024")
+        pid = _submit_prompt(wf)
+        b64 = _poll_image_b64(pid)
+        if not b64:
+            webhook_update_task(task_id, "failed", 0, "", error="ComfyUI 타임아웃 또는 출력 없음")
+            return
+        url = save_b64_png(task_id, b64)
+        webhook_update_task(task_id, "succeeded", 100, "완료", image_url=url)
+    except Exception as e:
+        logger.exception("flux task=%s 실패", task_id)
+        webhook_update_task(task_id, "failed", 0, "", error=str(e))

services/image-render/providers/gpt_image.py

@@ -0,0 +1,47 @@
+"""GPT Image 2.0 — OpenAI Images API.
+
+POST https://api.openai.com/v1/images/generations
+body {model:"gpt-image-1", prompt, size, n:1} → data[0].b64_json
+"""
+from __future__ import annotations
+
+import logging
+import os
+
+import requests
+
+from nas_client import webhook_update_task
+from providers._media import save_b64_png
+
+logger = logging.getLogger(__name__)
+OPENAI_URL = "https://api.openai.com/v1/images/generations"
+DEFAULT_MODEL = "gpt-image-1"
+
+
+def run_gpt_image_generation(task_id: str, params: dict) -> None:
+    try:
+        if not os.getenv("OPENAI_API_KEY"):
+            webhook_update_task(task_id, "failed", 0, "", error="OPENAI_API_KEY 미설정 (Windows .env)")
+            return
+        webhook_update_task(task_id, "processing", 10, "GPT Image 호출 중...")
+        body = {
+            "model": params.get("model") or DEFAULT_MODEL,
+            "prompt": params["prompt"],
+            "size": params.get("size") or "1024x1024",
+            "n": 1,
+        }
+        resp = requests.post(
+            OPENAI_URL,
+            headers={"Authorization": f"Bearer {os.getenv('OPENAI_API_KEY')}", "Content-Type": "application/json"},
+            json=body,
+            timeout=120,
+        )
+        if resp.status_code != 200:
+            webhook_update_task(task_id, "failed", 0, "", error=f"OpenAI {resp.status_code}: {resp.text[:200]}")
+            return
+        b64 = resp.json()["data"][0]["b64_json"]
+        url = save_b64_png(task_id, b64)
+        webhook_update_task(task_id, "succeeded", 100, "완료", image_url=url)
+    except Exception as e:
+        logger.exception("gpt_image task=%s 실패", task_id)
+        webhook_update_task(task_id, "failed", 0, "", error=str(e))

services/image-render/providers/nano_banana.py

@@ -0,0 +1,52 @@
+"""Nano Banana — Gemini 2.5 Flash Image (generativelanguage API).
+
+POST /v1beta/models/{MODEL}:generateContent
+→ candidates[0].content.parts[*].inlineData.data (b64 png)
+"""
+from __future__ import annotations
+
+import logging, os
+import requests
+
+from nas_client import webhook_update_task
+from providers._media import save_b64_png
+
+logger = logging.getLogger(__name__)
+GEMINI_BASE = "https://generativelanguage.googleapis.com/v1beta"
+DEFAULT_MODEL = "gemini-2.5-flash-image"
+
+
+def _extract_b64(data: dict):
+    for cand in data.get("candidates", []):
+        for part in cand.get("content", {}).get("parts", []):
+            inline = part.get("inlineData") or part.get("inline_data")
+            if inline and inline.get("data"):
+                return inline["data"]
+    return None
+
+
+def run_nano_banana_generation(task_id: str, params: dict) -> None:
+    try:
+        if not os.getenv("GEMINI_API_KEY"):
+            webhook_update_task(task_id, "failed", 0, "", error="GEMINI_API_KEY 미설정 (Windows .env)")
+            return
+        webhook_update_task(task_id, "processing", 10, "Nano Banana (Gemini) 호출 중...")
+        model_id = params.get("model") or DEFAULT_MODEL
+        body = {"contents": [{"parts": [{"text": params["prompt"]}]}]}
+        resp = requests.post(
+            f"{GEMINI_BASE}/models/{model_id}:generateContent",
+            headers={"x-goog-api-key": os.getenv("GEMINI_API_KEY"), "Content-Type": "application/json"},
+            json=body, timeout=120,
+        )
+        if resp.status_code != 200:
+            webhook_update_task(task_id, "failed", 0, "", error=f"Gemini {resp.status_code}: {resp.text[:200]}")
+            return
+        b64 = _extract_b64(resp.json())
+        if not b64:
+            webhook_update_task(task_id, "failed", 0, "", error="Gemini 응답에 이미지 없음")
+            return
+        url = save_b64_png(task_id, b64)
+        webhook_update_task(task_id, "succeeded", 100, "완료", image_url=url)
+    except Exception as e:
+        logger.exception("nano_banana task=%s 실패", task_id)
+        webhook_update_task(task_id, "failed", 0, "", error=str(e))

services/image-render/requirements.txt

@@ -0,0 +1,9 @@
+fastapi==0.115.6
+uvicorn[standard]==0.34.0
+requests==2.32.3
+redis>=5.0
+httpx>=0.27
+openai>=1.50.0
+pytest>=8.0
+pytest-asyncio>=0.24
+respx>=0.21

services/image-render/tests/test_flux.py

@@ -0,0 +1,21 @@
+import providers.flux as fx
+
+def test_blocked_during_trading_hours(monkeypatch):
+    monkeypatch.setenv("FLUX_BLOCK_TRADING_HOURS", "1")
+    monkeypatch.setattr(fx, "_is_trading_hours", lambda: True)
+    calls = []
+    monkeypatch.setattr(fx, "webhook_update_task", lambda *a, **k: calls.append((a, k)))
+    fx.run_flux_generation("t1", {"prompt": "a cat"})
+    assert calls[-1][0][1] == "failed"
+    assert "장중" in calls[-1][1]["error"]
+
+def test_success_polls_history_and_saves(monkeypatch):
+    monkeypatch.setattr(fx, "_is_trading_hours", lambda: False)
+    calls = []
+    monkeypatch.setattr(fx, "webhook_update_task", lambda *a, **k: calls.append((a, k)))
+    monkeypatch.setattr(fx, "_load_workflow", lambda prompt, size: {"3": {}})
+    monkeypatch.setattr(fx, "_submit_prompt", lambda wf: "pid-1")
+    monkeypatch.setattr(fx, "_poll_image_b64", lambda pid: "ZmFrZQ==")
+    monkeypatch.setattr(fx, "save_b64_png", lambda tid, b64: "/media/image/t1.png")
+    fx.run_flux_generation("t1", {"prompt": "a cat"})
+    assert [c for c in calls if c[0][1] == "succeeded"]

services/image-render/tests/test_gpt_image.py

@@ -0,0 +1,32 @@
+import providers.gpt_image as gi
+
+
+def test_missing_key_reports_failed(monkeypatch):
+    monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+    calls = []
+    monkeypatch.setattr(gi, "webhook_update_task", lambda *a, **k: calls.append((a, k)))
+    gi.run_gpt_image_generation("t1", {"prompt": "a cat"})
+    # 마지막 호출이 failed
+    assert calls[-1][0][1] == "failed"
+
+
+def test_success_saves_and_reports_url(monkeypatch):
+    monkeypatch.setenv("OPENAI_API_KEY", "sk-test")
+    calls = []
+    monkeypatch.setattr(gi, "webhook_update_task", lambda *a, **k: calls.append((a, k)))
+    monkeypatch.setattr(gi, "save_b64_png", lambda tid, b64: "/media/image/t1.png")
+
+    class FakeResp:
+        status_code = 200
+
+        def json(self):
+            return {"data": [{"b64_json": "ZmFrZQ=="}]}
+
+        def raise_for_status(self):
+            pass
+
+    monkeypatch.setattr(gi.requests, "post", lambda *a, **k: FakeResp())
+
+    gi.run_gpt_image_generation("t1", {"prompt": "a cat"})
+    succeeded = [c for c in calls if c[0][1] == "succeeded"]
+    assert succeeded and succeeded[-1][1]["image_url"] == "/media/image/t1.png"