🐛 Problem:
- AI returns JSON with fullwidth characters: ［｛
- Regex \[ cannot match fullwidth ［
- extractDecisions() fails with "无法找到JSON数组起始"
🔧 Root Cause:
- fixMissingQuotes() was called AFTER regex matching
- If regex fails to match fullwidth chars, fix function never executes
✅ Solution:
- Call fixMissingQuotes(s) BEFORE regex matching (line 461)
- Convert fullwidth to halfwidth first: ［→[, ｛→{
- Then regex can successfully match the JSON array
📊 Impact:
- Fixes "无法找到JSON数组起始" error
- Supports AI responses with fullwidth JSON characters
- Backward compatible with halfwidth JSON
This fix is identical to z-dev commit 3676cc0
Critical bug: AI can output fullwidth space ( U+3000) between brackets:
Input: ［ ｛"symbol":"BTCUSDT"｝］
↑ ↑ fullwidth space
After previous fix:
[ {"symbol":"BTCUSDT"}]
↑ fullwidth space remained!
Result: validateJSONFormat failed because:
- Checks "[{" (no space) ❌
- Checks "[ {" (halfwidth space U+0020) ❌
- AI output "[ {" (fullwidth space U+3000) ❌
Solution: Replace fullwidth space → halfwidth space
- (U+3000) → space (U+0020)
This allows existing validation logic to work:
strings.HasPrefix(trimmed, "[ {") now matches ✅
Why fullwidth space?
- Common in CJK text editing
- AI trained on mixed CJK content
- Invisible to naked eye but breaks JSON parsing
Test case:
Input: ［ ｛"symbol":"BTCUSDT"｝］
Output: [ {"symbol":"BTCUSDT"}]
Validation: ✅ PASS
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Critical discovery: AI can output different types of "fullwidth" brackets:
- Fullwidth: ［］｛｝(U+FF3B/FF3D/FF5B/FF5D) ← Already handled
- CJK: 【】〔〕(U+3010/3011/3014/3015) ← Was missing!
Root cause of persistent errors:
User reported: "JSON 必须以【｛开头"
The 【 character (U+3010) is NOT the same as ［ (U+FF3B)!
Added CJK punctuation replacements:
- 【 → [ (U+3010 Left Black Lenticular Bracket)
- 】 → ] (U+3011 Right Black Lenticular Bracket)
- 〔 → [ (U+3014 Left Tortoise Shell Bracket)
- 〕 → ] (U+3015 Right Tortoise Shell Bracket)
- 、 → , (U+3001 Ideographic Comma)
Why this was missed:
AI uses different characters in different contexts. CJK brackets (U+3010-3017)
are distinct from Fullwidth Forms (U+FF00-FFEF) in Unicode.
Test case:
Input: 【｛"symbol"："BTCUSDT"】
Output: [{"symbol":"BTCUSDT"}]
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Adds comprehensive JSON validation before parsing to catch common AI output errors:
1. Format validation: Ensures JSON starts with [{ (decision array)
2. Range symbol detection: Rejects ~ symbols (e.g., "leverage: 3~5")
3. Thousands separator detection: Rejects commas in numbers (e.g., "98,000")
Execution order (critical for fullwidth character fix):
1. Extract JSON from response
2. fixMissingQuotes - normalize fullwidth → halfwidth ✅
3. validateJSONFormat - check for common errors ✅
4. Parse JSON
This validation layer provides early error detection and clearer error messages
for debugging AI response issues.
Added helper function:
- min(a, b int) int - returns smaller of two integers
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changes:
- Updated error handling in `GetFullDecision` and `parseFullDecisionResponse` functions to return the decision object even when an error occurs, improving the clarity of error messages.
This ensures that the decision object is consistently returned, allowing for better debugging and handling of errors in the decision-making process.
Changes to decision/engine.go:
- Clean up Context struct field alignment for better readability
- Enhance system prompt to include more technical analysis methods:
* Added: technical resistance levels, Fibonacci, volatility bands
* Changed wording from "you can do X" to "you can do but not limited to X"
to encourage AI to use broader range of analysis techniques
This gives the AI decision engine more explicit guidance on available
technical analysis tools while maintaining flexibility.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Pass leverage config through TraderManager to AutoTrader
- Add BTCETHLeverage and AltcoinLeverage fields to Context and AutoTraderConfig
- Update decision validation to use configured leverage limits
- Display configured leverage in startup message
- Update error messages to show current leverage limits
Changes:
- main.go: Pass leverage config to AddTrader, update startup message
- manager/trader_manager.go: Accept and forward leverage config
- trader/auto_trader.go: Store leverage config, pass to Context
- decision/engine.go: Use dynamic leverage limits in validation
This completes the leverage configuration feature implementation.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Remove prescriptive indicator combinations and let AI freely use all available data.
**Changes**:
- Emphasized AI has access to **raw sequence data** (MidPrices array, 4h candles)
- Listed all available sequences: price, technical (EMA/MACD/RSI), and capital flow (volume/OI)
- Removed hard-coded indicator combinations (e.g., "MACD + RSI + Volume")
- Changed from prescriptive examples to open-ended analysis freedom
- AI can now freely perform trend analysis, pattern recognition, support/resistance calculation
- Reduced minimum close-open interval from 30min to 15min for more flexibility
**Before**:
```
强信号示例:
- 趋势突破 + 多个指标确认（MACD + RSI + 成交量）
- 持仓量暴增 + 价格突破关键位
```
**After**:
```
你拥有的完整数据：
- 📊 原始序列：3分钟价格序列(MidPrices数组) + 4小时K线序列
- 📈 技术序列：EMA20序列、MACD序列、RSI7序列、RSI14序列
- 💰 资金序列：成交量序列、持仓量(OI)序列、资金费率
分析方法（完全由你自主决定）：
- 自由运用序列数据，你可以做趋势分析、形态识别、支撑阻力计算
- 多维度交叉验证（价格+量+OI+指标+序列形态）
- 用你认为最有效的方法发现高确定性机会
```
**Philosophy**: Trust AI to discover effective patterns in raw data rather than constraining it to pre-defined indicator combinations.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Track and display how long each position has been held to help AI make better timing decisions.
**Implementation**:
- Added UpdateTime field to PositionInfo struct (decision/engine.go:26)
- Added positionFirstSeenTime map to AutoTrader for tracking (trader/auto_trader.go:60)
- Record opening time when position is created successfully:
- executeOpenLongWithRecord: Records timestamp for long positions (trader/auto_trader.go:540-541)
- executeOpenShortWithRecord: Records timestamp for short positions (trader/auto_trader.go:593-594)
- Fallback tracking in buildTradingContext for program restart scenarios (trader/auto_trader.go:386-392)
- Auto-cleanup closed positions from tracking map (trader/auto_trader.go:409-414)
- Display duration in user prompt with smart formatting:
- Under 60 min: "持仓时长25分钟"
- Over 60 min: "持仓时长2小时15分钟"
**Example Output**:
```
1. TAOUSDT LONG | 入场价435.5300 当前价433.1900 | 盈亏-0.54% | 杠杆20x | 保证金25 | 强平价418.1528 | 持仓时长2小时15分钟
```
**Benefits**:
- AI can see how long positions have been held
- Helps enforce minimum holding period (30-60 min) from system prompt
- Simple implementation with minimal overhead
- Auto-cleanup prevents memory leaks
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Major improvements to the AI trading decision engine:
**Core Changes in decision/engine.go** (175 lines modified):
1. **Sharpe Ratio Optimization Focus**
- Restructured system prompt to emphasize Sharpe ratio maximization
- Added clear guidance: high-quality trades over frequent trading
- Explained that 3-minute scan interval ≠ trade every cycle
2. **Trading Frequency Controls**
- Defined optimal frequency: 2-4 trades/day (0.1-0.2 trades/hour)
- Over-trading threshold: >2 trades/hour indicates issues
- Minimum holding period: 30-60 minutes per position
3. **Long/Short Balance Incentives**
- Emphasized equal profit potential for long and short positions
- Removed long-bias with explicit short trading encouragement
- Clear guidance: uptrend→long, downtrend→short, sideways→wait
4. **Stricter Entry Signal Standards**
- Strong signals only: confidence ≥75, multi-indicator confirmation
- Weak signals explicitly discouraged (single indicator, unclear trend)
- Self-check mechanism to prevent premature re-entry (<30min)
5. **Enhanced Sharpe Ratio Feedback Loop**
- Sharpe < -0.5: Stop trading for 6+ cycles (18min), deep reflection
- Sharpe -0.5~0: Strict control, confidence >80 only
- Sharpe 0~0.7: Maintain current strategy
- Sharpe >0.7: Consider position size increase
6. **Risk-Reward Ratio Validation**
- Added hard constraint: R:R must be ≥ 3.0:1
- Automatic calculation and validation in `validateDecision()`
- Rejects trades with insufficient risk-reward ratio
7. **Improved Prompt Structure**
- More organized sections with clear headers
- Actionable guidance instead of abstract principles
- Better examples for JSON output format
**Impact**: These changes should significantly improve trading quality,
reduce over-trading, and increase Sharpe ratio through better risk management
and trade selection discipline.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Architecture improvements:
- Extract AI decision engine to dedicated `decision` package
- Create `mcp` package for Model Context Protocol client
- Separate market data structures into `market/data.go`
- Update trader to use new modular structure
New packages:
- `decision/engine.go` - AI decision logic and prompt building
- `mcp/client.go` - Unified AI API client (DeepSeek/Qwen)
- `market/data.go` - Market data type definitions
Benefits:
- Better separation of concerns
- Improved code organization and maintainability
- Easier to test individual components
- More flexible AI provider integration
- Cleaner dependency management
Updated imports:
- trader/auto_trader.go now uses decision and mcp packages
- Consistent API across different AI providers
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
