AGENTS.md · diff
git:20260609.4bf2c03 to git:20260630.3d632a3
60 added, 81 removed. Audit A to A.
- # A股量化交易系统 - 重构为AI交易员
+ # AGENTS.md
- ## 项目概述
- A股量化模拟盘系统，正在从"固定规则+打分机器"改造为"全自动AI交易员"。
+ Guidance for AI coding agents working on Quant Pilot.
- ## 技术栈
- - Python 3.12
- - SQLite (本地数据存储)
- - LongBridge OpenAPI (A股数据源)
- - MiMo LLM API (决策引擎)
- - Baostock (备用历史数据)
- - 东方财富 API (涨停板/龙虎榜)
+ Quant Pilot is a research and paper-trading system for A-share market automation.
+ Keep changes conservative, observable, and safe by default.
- ## 项目结构
+ ## Core Principles
+
+ - Preserve `BROKER_MODE=paper` as the default documented mode.
+ - Do not introduce real-money execution paths without explicit design review.
+ - Do not commit secrets, runtime databases, logs, private server details, or account
+ snapshots.
+ - Keep the public dashboard read-only in production.
+ - Prefer standard SDKs and maintained ecosystem packages over custom infrastructure.
+ - Keep implementation style consistent with the existing Python modules and tests.
+ - Add or update tests when changing trading decisions, scheduler behavior, risk,
+ persistence, web security, or deployment contracts.
+
+ ## Useful Local Commands
+
+ ```bash
+ python3 -m compileall -q web scheduler strategy execution data main.py config.py
+ python3 test_web_public_dashboard.py
+ python3 test_agent_driver_contract.py
+ python3 test_auto_control.py
+ python3 test_auto_watchdog.py
+ python3 test_auto_doctor.py
+ python3 test_auto_trader.py
+ python3 test_paper_readiness.py
+ python3 test_paper_bootstrap.py
```
- a-stock-quant/
- ├── config.py # 全局配置
- ├── main.py # 主入口 (--scan/--execute/--review/--full)
- ├── data/ # 数据层
- │ ├── realtime.py # 腾讯实时行情
- │ ├── history.py # Baostock历史数据
- │ ├── eastmoney.py # 东方财富数据
- │ ├── market.py # 市场数据聚合
- │ ├── sentiment.py # 舆情数据
- │ └── database.py # [新建] SQLite存储层
- ├── signals/ # 信号层
- │ ├── technical.py # 技术信号 (Ichimoku/Volume Profile等)
- │ ├── capital.py # 资金面信号
- │ ├── emotion.py # 情绪面信号
- │ ├── fundamental.py # 基本面信号
- │ ├── sentiment.py # 舆情信号 (LLM)
- │ ├── composite.py # 综合信号
- │ └── market_timing.py # 市场择时
- ├── strategy/ # 策略层
- │ ├── stock_picker.py # 选股器
- │ ├── decision.py # 决策引擎
- │ ├── adaptive.py # 自适应引擎
- │ ├── llm_trader.py # [新建] LLM决策引擎
- │ ├── market_regime.py # [新建] 市场环境识别
- │ ├── regime_config.py # [新建] 策略配置中心
- │ └── memory.py # [新建] 交易记忆系统
- ├── execution/ # 执行层
- │ ├── paper_account.py # 模拟账户
- │ └── order.py # 订单管理
- ├── risk/ # 风控层
- │ ├── position.py # 仓位管理
- │ ├── stop_loss.py # 止损管理
- │ └── drawdown.py # 回撤控制
- ├── review/ # 复盘层
- │ ├── daily_review.py # 每日复盘
- │ ├── performance.py # 绩效统计
- │ └── llm_review.py # LLM复盘
- ├── scheduler/ # 调度层
- │ ├── pipeline.py # 管道
- │ ├── market_calendar.py # 交易日历
- │ └── notifier.py # 通知推送
- └── data/ # 数据目录
- ├── paper_account.json # 模拟账户 (将迁移到SQLite)
- ├── signal_cache.json # 信号缓存
- ├── adaptive_state.json # 自适应状态
- └── quant.db # [新建] SQLite数据库
+
+ Some checks require market-data or LLM credentials. If a test cannot run locally,
+ record the reason and the residual risk.
+
+ ## Web and Security Notes
+
+ Production web mode should use:
+
+ ```bash
+ ALPHAPILOT_ENV=production
+ BROKER_MODE=paper
```
- ## 关键命令
- - `python3 main.py --scan` — 扫描信号（选股+打分+决策）
- - `python3 main.py --execute` — 执行交易（风控+模拟盘）
- - `python3 main.py --review` — 每日复盘
- - `python3 main.py --full` — 全链路
+ The production dashboard must not expose:
- ## 数据源配置
- 长桥API凭证在 ~/.hermes/.env:
- - LONGPORT_APP_KEY
- - LONGPORT_APP_SECRET
- - LONGPORT_ACCESS_TOKEN
+ - control buttons
+ - control tokens
+ - raw environment variables
+ - internal file paths
+ - raw tracebacks
+ - shell repair commands
+ - broker credentials
- ## 重构目标
- 将系统从"加权打分"改为"LLM推理决策"，分7个阶段：
- 1. SQLite存储层 + 长桥数据源集成
- 2. 市场环境识别器
- 3. LLM决策引擎
- 4. 交易记忆系统
- 5. 策略配置中心
- 6. 执行层+风控层改造
- 7. 回测引擎改造
+ Control APIs should stay unmounted in production unless explicitly enabled for a
+ private network.
- ## 代码规范
- - 所有中文注释
- - 每个模块有docstring
- - 错误处理完善，API调用有重试
- - 数据库操作用context manager
+ ## Documentation Expectations
+
+ User-facing behavior changes should update README or `docs/`. Risky operational
+ changes should mention rollback and verification steps.
+
+ ## Git Hygiene
+
+ - Keep commits focused.
+ - Do not rewrite unrelated user changes.
+ - Do not remove safety checks or redaction layers to make a test pass.
+ - Avoid committing generated runtime files under `data/`, `logs/`, or `.codex/`.