Dawn Proactive Agent v1.5
An autonomous, self-evolving ETF trading agent architecture designed for the A-share market.
Architecture Overview
┌─────────────────────────────────────────────────────┐
│ Dawn Proactive Agent v1.5 │
├─────────────────────────────────────────────────────┤
│ P0: dawn_reflect.py Hermes-inspired reflection │
│ P1: dawn_state.py LangGraph state machine │
│ P2: dawn_audit.py RagaAI audit trail │
│ P3: dawn_analysis.py ai-berkshire 4D scoring │
│ P4: dawn_guardrails.py OpenAI Agents safety │
├─────────────────────────────────────────────────────┤
│ Active in: 华泰柏瑞杯ETF AI交易巅峰赛 │
│ Period: 2026/6/11 - 2026/7/20 │
│ Portfolio: 8 ETFs, ~¥1M AUM │
└─────────────────────────────────────────────────────┘
Components
P0 - Self-Reflection Engine (dawn_reflect.py)
Hermes-inspired post-trade reflection. After every portfolio adjustment, automatically extracts lessons learned and loads them into the next decision cycle.
- Trigger: post-trade callback
- Output: structured learnings with recall scoring
- Integrates with
.learnings/structured log system
P1 - State Machine (dawn_state.py)
LangGraph-inspired workflow orchestration with checkpoint/resume.
- States: IDLE → ANALYZE → DECIDE → EXECUTE → REFLECT → IDLE
- Checkpoint recovery on timeout/crash
- Timeout downgrade: auto-fallback to safest state
P2 - Audit Trail (dawn_audit.py)
RagaAI-inspired immutable decision logging.
- Every decision recorded: timestamp, reasoning, data sources, signals, outcome
- Queryable by date/strategy/symbol
- Exports as structured JSON for backtesting
P3 - Multi-Dimension Scoring (dawn_analysis.py)
ai-berkshire-inspired composite scoring for ETF selection.
- Technical (30%): momentum, trend, volume
- News Sentiment (25%): real-time financial news analysis
- Capital Flow (25%): sector money flow tracking
- Volume-Price (20%): volume-price divergence detection
P4 - Safety Guardrails (dawn_guardrails.py)
OpenAI Agents SDK-inspired transaction safety checks.
- Blacklist: 科创板(688) and ST stocks blocked
- Position limits: max 40% single ETF
- Daily loss limit: max -5% stop-loss
- Available cash check before execution
Usage
Daily Strategy Run (09:28)
python scripts/dawn_proactive.py --action morning
Post-Market Review (15:05)
python scripts/dawn_proactive.py --action afternoon
Midday Check (11:30)
python scripts/dawn_proactive.py --action midday
Manual Trade
python scripts/dawn_etf_rotator.py --execute
Files
| File | Purpose |
|---|---|
scripts/dawn_proactive.py |
Main orchestrator |
scripts/dawn_etf_rotator.py |
ETF rotation strategy |
scripts/dawn_reflect.py |
P0: Self-reflection |
scripts/dawn_state.py |
P1: State machine |
scripts/dawn_audit.py |
P2: Audit trail |
scripts/dawn_analysis.py |
P3: 4D scoring |
scripts/dawn_guardrails.py |
P4: Safety guardrails |
scripts/dawn_memory_sync.py |
L1↔L3 memory sync |
scripts/dawn_collector.py |
Market data collection |
scripts/dawn_selector.py |
ETF selection logic |
scripts/dawn_monitor.py |
Real-time monitoring |
Results (2026-07-06)
- Cron jobs: 10 automated tasks running daily
- Portfolio: 8 ETFs, 38.2%仓位, ¥1,003,751
- Key wins: 科创板V反守住(+0.69% on 07-03), 全天候自动运行
- Self-healing: 收盘复盘cron timeout自动修复 (agentTurn→command模式)
- Delivery: 飞书推送自动修复 (delivery.to user:前缀)
See samples/proactive_demo.md for a full session trace.
Changelog
v1.5 (2026-07-06)
- P0-P4 framework complete
- Cron self-healing: timeout 120s→300s, agentTurn→command mode
- Delivery fix: feishu user: prefix
- Git cleanup: 380K lines of old archive deleted
v1.0 (2026-06-24)
- Initial dawn agent framework
- LM Studio + DeepSeek-R1 local inference
- Memory sync and knowledge injection