Add API-efficient daily trading mode for Gemini Free tier compatibility: ## Features - **Batch Decisions**: GeminiClient.decide_batch() analyzes multiple stocks in a single API call using compressed JSON format - **Daily Trading Mode**: run_daily_session() executes N sessions per day at configurable intervals (default: 4 sessions, 6 hours apart) - **Mode Selection**: TRADE_MODE env var switches between daily (batch) and realtime (per-stock) modes - **Requirements Log**: docs/requirements-log.md tracks user feedback chronologically for project evolution ## Configuration - TRADE_MODE: "daily" (default) | "realtime" - DAILY_SESSIONS: 1-10 (default: 4) - SESSION_INTERVAL_HOURS: 1-24 (default: 6) ## API Efficiency - 2 markets × 4 sessions = 8 API calls/day (within Free tier 20 calls) - 3 markets × 4 sessions = 12 API calls/day (within Free tier 20 calls) ## Testing - 9 new batch decision tests (all passing) - All existing tests maintained (298 passed) ## Documentation - docs/architecture.md: Trading Modes section with daily vs realtime - CLAUDE.md: Requirements Management section - docs/requirements-log.md: Initial entries for API efficiency needs Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
11 KiB
System Architecture
Overview
Self-evolving AI trading agent for global stock markets via KIS (Korea Investment & Securities) API. The main loop in src/main.py orchestrates four components across multiple markets with two trading modes: daily (batch API calls) or realtime (per-stock decisions).
Trading Modes
The system supports two trading frequency modes controlled by the TRADE_MODE environment variable:
Daily Mode (default)
Optimized for Gemini Free tier API limits (20 calls/day):
- Batch decisions: 1 API call per market per session
- Fixed schedule: 4 sessions per day at 6-hour intervals (configurable)
- API efficiency: Processes all stocks in a market simultaneously
- Use case: Free tier users, cost-conscious deployments
- Configuration:
TRADE_MODE=daily DAILY_SESSIONS=4 # Sessions per day (1-10) SESSION_INTERVAL_HOURS=6 # Hours between sessions (1-24)
Example: With 2 markets (US, KR) and 4 sessions/day = 8 API calls/day (within 20 call limit)
Realtime Mode
High-frequency trading with individual stock analysis:
- Per-stock decisions: 1 API call per stock per cycle
- 60-second interval: Continuous monitoring
- Use case: Production deployments with Gemini paid tier
- Configuration:
TRADE_MODE=realtime
Note: Realtime mode requires Gemini API subscription due to high call volume.
Core Components
1. Broker (src/broker/)
KISBroker (kis_api.py) — Async KIS API client for domestic Korean market
- Automatic OAuth token refresh (valid for 24 hours)
- Leaky-bucket rate limiter (10 requests per second)
- POST body hash-key signing for order authentication
- Custom SSL context with disabled hostname verification for VTS (virtual trading) endpoint due to known certificate mismatch
OverseasBroker (overseas.py) — KIS overseas stock API wrapper
- Reuses KISBroker infrastructure (session, token, rate limiter) via composition
- Supports 9 global markets: US (NASDAQ/NYSE/AMEX), Japan, Hong Kong, China (Shanghai/Shenzhen), Vietnam (Hanoi/HCM)
- Different API endpoints for overseas price/balance/order operations
Market Schedule (src/markets/schedule.py) — Timezone-aware market management
MarketInfodataclass with timezone, trading hours, lunch breaks- Automatic DST handling via
zoneinfo.ZoneInfo is_market_open()checks weekends, trading hours, lunch breaksget_open_markets()returns currently active marketsget_next_market_open()finds next market to open and when
2. Brain (src/brain/gemini_client.py)
GeminiClient — AI decision engine powered by Google Gemini
- Constructs structured prompts from market data
- Parses JSON responses into
TradeDecisionobjects (action,confidence,rationale) - Forces HOLD when confidence < threshold (default 80)
- Falls back to safe HOLD on any parse/API error
- Handles markdown-wrapped JSON, malformed responses, invalid actions
3. Risk Manager (src/core/risk_manager.py)
RiskManager — Safety circuit breaker and order validation
⚠️ READ-ONLY by policy (see docs/agents.md)
- Circuit Breaker: Halts all trading via
SystemExitwhen daily P&L drops below -3.0%- Threshold may only be made stricter, never relaxed
- Calculated as
(total_eval - purchase_total) / purchase_total * 100
- Fat-Finger Protection: Rejects orders exceeding 30% of available cash
- Must always be enforced, cannot be disabled
4. Notifications (src/notifications/telegram_client.py)
TelegramClient — Real-time event notifications via Telegram Bot API
- Sends alerts for trades, circuit breakers, fat-finger rejections, system events
- Non-blocking: failures are logged but never crash trading
- Rate-limited: 1 message/second default to respect Telegram API limits
- Auto-disabled when credentials missing
- Gracefully handles API errors, network timeouts, invalid tokens
Notification Types:
- Trade execution (BUY/SELL with confidence)
- Circuit breaker trips (critical alert)
- Fat-finger protection triggers (order rejection)
- Market open/close events
- System startup/shutdown status
Setup: See src/notifications/README.md for bot creation and configuration.
5. Evolution (src/evolution/optimizer.py)
StrategyOptimizer — Self-improvement loop
- Analyzes high-confidence losing trades from SQLite
- Asks Gemini to generate new
BaseStrategysubclasses - Validates generated strategies by running full pytest suite
- Simulates PR creation for human review
- Only activates strategies that pass all tests
Data Flow
┌─────────────────────────────────────────────────────────────┐
│ Main Loop (60s cycle per stock, per market) │
└─────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────┐
│ Market Schedule Check │
│ - Get open markets │
│ - Filter by enabled markets │
│ - Wait if all closed │
└──────────────────┬────────────────┘
│
▼
┌──────────────────────────────────┐
│ Broker: Fetch Market Data │
│ - Domestic: orderbook + balance │
│ - Overseas: price + balance │
└──────────────────┬────────────────┘
│
▼
┌──────────────────────────────────┐
│ Calculate P&L │
│ pnl_pct = (eval - cost) / cost │
└──────────────────┬────────────────┘
│
▼
┌──────────────────────────────────┐
│ Brain: Get Decision │
│ - Build prompt with market data │
│ - Call Gemini API │
│ - Parse JSON response │
│ - Return TradeDecision │
└──────────────────┬────────────────┘
│
▼
┌──────────────────────────────────┐
│ Risk Manager: Validate Order │
│ - Check circuit breaker │
│ - Check fat-finger limit │
│ - Raise if validation fails │
└──────────────────┬────────────────┘
│
▼
┌──────────────────────────────────┐
│ Broker: Execute Order │
│ - Domestic: send_order() │
│ - Overseas: send_overseas_order() │
└──────────────────┬────────────────┘
│
▼
┌──────────────────────────────────┐
│ Notifications: Send Alert │
│ - Trade execution notification │
│ - Non-blocking (errors logged) │
│ - Rate-limited to 1/sec │
└──────────────────┬────────────────┘
│
▼
┌──────────────────────────────────┐
│ Database: Log Trade │
│ - SQLite (data/trades.db) │
│ - Track: action, confidence, │
│ rationale, market, exchange │
└───────────────────────────────────┘
Database Schema
SQLite (src/db.py)
CREATE TABLE trades (
id INTEGER PRIMARY KEY AUTOINCREMENT,
timestamp TEXT NOT NULL,
stock_code TEXT NOT NULL,
action TEXT NOT NULL, -- BUY | SELL | HOLD
confidence INTEGER NOT NULL, -- 0-100
rationale TEXT,
quantity INTEGER,
price REAL,
pnl REAL DEFAULT 0.0,
market TEXT DEFAULT 'KR', -- KR | US_NASDAQ | JP | etc.
exchange_code TEXT DEFAULT 'KRX' -- KRX | NASD | NYSE | etc.
);
Auto-migration: Adds market and exchange_code columns if missing for backward compatibility.
Configuration
Pydantic Settings (src/config.py)
Loaded from .env file:
# Required
KIS_APP_KEY=your_app_key
KIS_APP_SECRET=your_app_secret
KIS_ACCOUNT_NO=XXXXXXXX-XX
GEMINI_API_KEY=your_gemini_key
# Optional
MODE=paper # paper | live
DB_PATH=data/trades.db
CONFIDENCE_THRESHOLD=80
MAX_LOSS_PCT=3.0
MAX_ORDER_PCT=30.0
ENABLED_MARKETS=KR,US_NASDAQ # Comma-separated market codes
# Trading Mode (API efficiency)
TRADE_MODE=daily # daily | realtime
DAILY_SESSIONS=4 # Sessions per day (daily mode only)
SESSION_INTERVAL_HOURS=6 # Hours between sessions (daily mode only)
# Telegram Notifications (optional)
TELEGRAM_BOT_TOKEN=1234567890:ABCdefGHIjklMNOpqrsTUVwxyz
TELEGRAM_CHAT_ID=123456789
TELEGRAM_ENABLED=true
Tests use in-memory SQLite (DB_PATH=":memory:") and dummy credentials via tests/conftest.py.
Error Handling
Connection Errors (Broker API)
- Retry with exponential backoff (2^attempt seconds)
- Max 3 retries per stock
- After exhaustion, skip stock and continue with next
API Quota Errors (Gemini)
- Return safe HOLD decision with confidence=0
- Log error but don't crash
- Agent continues trading on next cycle
Circuit Breaker Tripped
- Immediately halt via
SystemExit - Log critical message
- Requires manual intervention to restart
Market Closed
- Wait until next market opens
- Use
get_next_market_open()to calculate wait time - Sleep until market open time
Telegram API Errors
- Log warning but continue trading
- Missing credentials → auto-disable notifications
- Network timeout → skip notification, no retry
- Invalid token → log error, trading unaffected
- Rate limit exceeded → queued via rate limiter
Guarantee: Notification failures never interrupt trading operations.