Commit Graph

6 Commits

Author SHA1 Message Date
agentson
0b20251de0 fix: parse_response에서 missing fields 시 raw 텍스트 보존 (#245)
Some checks failed
CI / test (pull_request) Has been cancelled
pre_market_planner는 prompt_override로 Gemini에 플레이북 JSON을 요청한다.
Gemini가 플레이북 JSON을 반환해도 parse_response가 action/confidence/rationale 키가
없다는 이유로 rationale="Missing required fields"를 반환해 실제 응답이 버려졌다.

이로 인해 플레이북 생성이 항상 실패하고 RSI 기반 기본 폴백이 사용됐으며,
RSI가 없는 해외 시장 데이터와 매칭되지 않아 모든 결정이 HOLD(confidence=0)였다.

수정: missing fields 시 rationale=raw로 설정해 실제 Gemini 응답을 보존한다.
pre_market_planner가 decision.rationale에서 플레이북 JSON을 추출하여 정상 파싱 가능.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-02-25 01:31:54 +09:00
agentson
f7289606fc fix: use prompt_override in gemini_client.decide() for playbook generation
Some checks failed
CI / test (pull_request) Has been cancelled
decide() ignored market_data["prompt_override"], always building a generic
trade-decision prompt. This caused pre_market_planner playbook generation
to fail with JSONDecodeError on every market, falling back to defensive
playbooks. Now prompt_override takes priority over both optimization and
standard prompt building.

Closes #143

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-18 02:02:13 +09:00
agentson
0057de4d12 feat: implement daily trading mode with batch decisions (issue #57)
Some checks failed
CI / test (pull_request) Has been cancelled
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>
2026-02-05 09:28:10 +09:00
agentson
62fd4ff5e1 feat: implement data-driven external data integration (issue #22)
Add objective external data sources to enhance trading decisions beyond
market prices and user input.

## New Modules

### src/data/news_api.py
- News sentiment analysis with Alpha Vantage and NewsAPI support
- Sentiment scoring (-1.0 to +1.0) per article and aggregated
- 5-minute caching to minimize API quota usage
- Graceful degradation when APIs unavailable

### src/data/economic_calendar.py
- Track major economic events (FOMC, GDP, CPI)
- Earnings calendar per stock
- Event proximity checking for high-volatility periods
- Hardcoded major events for 2026 (no API required)

### src/data/market_data.py
- Market sentiment indicators (Fear & Greed equivalent)
- Market breadth (advance/decline ratios)
- Sector performance tracking
- Fear/Greed score calculation

## Integration

Enhanced GeminiClient to seamlessly integrate external data:
- Optional news_api, economic_calendar, and market_data parameters
- Async build_prompt() includes external context when available
- Backward-compatible build_prompt_sync() for existing code
- Graceful fallback when external data unavailable

External data automatically added to AI prompts:
- News sentiment with top articles
- Upcoming high-impact economic events
- Market sentiment and breadth indicators

## Configuration

Added optional settings to config.py:
- NEWS_API_KEY: API key for news provider
- NEWS_API_PROVIDER: "alphavantage" or "newsapi"
- MARKET_DATA_API_KEY: API key for market data

## Testing

Comprehensive test suite with 38 tests:
- NewsAPI caching, sentiment parsing, API integration
- EconomicCalendar event filtering, earnings lookup
- MarketData sentiment and breadth calculations
- GeminiClient integration with external data sources
- All tests use mocks (no real API keys required)
- 81% coverage for src/data module (exceeds 80% requirement)

## Circular Import Fix

Fixed circular dependency between gemini_client.py and cache.py:
- Use TYPE_CHECKING for imports in cache.py
- String annotations for TradeDecision type hints

All 195 existing tests pass. No breaking changes to existing functionality.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-02-04 18:06:34 +09:00
agentson
b26ff0c1b8 feat: implement timezone-based global market auto-selection
Some checks failed
CI / test (pull_request) Has been cancelled
Implement comprehensive multi-market trading system with automatic
market selection based on timezone and trading hours.

## New Features
- Market schedule module with 10 global markets (KR, US, JP, HK, CN, VN)
- Overseas broker for KIS API international stock trading
- Automatic market detection based on current time and timezone
- Next market open waiting logic when all markets closed
- ConnectionError retry with exponential backoff (max 3 attempts)

## Architecture Changes
- Market-aware trading cycle with domestic/overseas broker routing
- Market context in AI prompts for better decision making
- Database schema extended with market and exchange_code columns
- Config setting ENABLED_MARKETS for market selection

## Testing
- 19 new tests for market schedule (timezone, DST, lunch breaks)
- All 54 tests passing
- Lint fixes with ruff

## Files Added
- src/markets/schedule.py - Market schedule and timezone logic
- src/broker/overseas.py - KIS overseas stock API client
- tests/test_market_schedule.py - Market schedule test suite

## Files Modified
- src/main.py - Multi-market main loop with retry logic
- src/config.py - ENABLED_MARKETS setting
- src/db.py - market/exchange_code columns with migration
- src/brain/gemini_client.py - Dynamic market context in prompts

Resolves #5

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-02-04 09:29:25 +09:00
d1750af80f Add complete Ouroboros trading system with TDD test suite
Some checks failed
CI / test (push) Has been cancelled
Implement the full autonomous trading agent architecture:
- KIS broker with async API, token refresh, leaky bucket rate limiter, and hash key signing
- Gemini-powered decision engine with JSON parsing and confidence threshold enforcement
- Risk manager with circuit breaker (-3% P&L) and fat finger protection (30% cap)
- Evolution engine for self-improving strategy generation via failure analysis
- 35 passing tests written TDD-first covering risk, broker, and brain modules
- CI/CD pipeline, Docker multi-stage build, and AI agent context docs

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-02-04 02:08:48 +09:00