大时代：自进化多智能体交易系统

📌 Visit the 大时代 website

大时代 is an open-source financial trading agent framework that combines multi-agent collaboration, run-scoped workspaces, and memory to support both backtests and live trading workflows.

The repository name still uses evotraders, but the product-facing branding now follows the 大时代 naming used by the reference branch.

---

Core Features

Multi-agent trading team Six roles collaborate like a real desk: four specialist analysts (fundamentals, technical, sentiment, valuation), one portfolio manager, and one risk manager.

Continuous learning Agents can persist long-term memory with ReMe, reflect after each cycle, and evolve their decision patterns over time.

Backtest and live modes The same runtime model supports historical simulation and live execution with real-time market data.

Operator-facing UI The frontend exposes the trading room, runtime controls, logs, approvals, agent workspaces, and explain/news views.

---

Current Architecture

The repository uses a split-service runtime model for local development and is the default supported path.

Runtime vs Design-Time

• runtime — the active execution layer (scheduler, gateway, pipeline, approvals during a live run)
• run — one concrete execution instance (runs/<run_id>/)
• design-time — configuration and control-plane concepts before a specific runtime is launched
• workspace — the design-time registry exposed by agent_service (workspaces/)

Service Surfaces

Service	Port	Responsibility
backend.apps.agent_service	:8000	Control plane for workspaces, agents, skills, and guard/approval APIs
backend.apps.trading_service	:8001	Read-only trading data APIs
backend.apps.news_service	:8002	Read-only explain/news APIs
backend.apps.runtime_service	:8003	Runtime lifecycle APIs
backend.apps.openclaw_service	:8004	Read-only OpenClaw facade
WebSocket gateway	:8765	Live event/feed channel for the frontend

Active Runtime Path

frontend -> runtime_service/control APIs -> gateway/runtime manager -> market service + pipeline + storage

Runtime state is stored in runs/<run_id>/ — this is the runtime source of truth. The workspaces/ directory is the design-time registry, not the runtime execution path.

Documentation

• docs/current-architecture.md — canonical architecture facts
• services/README.md — service boundaries and migration details
• docs/current-architecture.excalidraw — visual diagram
• docs/development-roadmap.md — next-step execution plan
• docs/terminology.md — consistent terminology guide

---

Quick Start

1. Install

# clone this repository, then:
cd evotraders

# backend runtime dependencies
uv pip install -r requirements.txt

# install package entrypoint in editable mode
uv pip install -e .

# optional
# uv pip install -e ".[dev]"
# pip install -e .

Frontend dependencies:

cd frontend
npm ci
cd ..

Production deployment should prefer requirements.txt for backend and npm ci for frontend so the pulled environment matches the checked-in lockfiles and version pins.

2. Configure environment

cp env.template .env

The root env.template is the canonical local template. A .env.example is also kept in the repo for reference.

Minimum useful variables:

# watchlist
TICKERS=AAPL,MSFT,GOOGL,NVDA,TSLA,META,AMZN

# market data
FIN_DATA_SOURCE=finnhub
FINANCIAL_DATASETS_API_KEY=
FINNHUB_API_KEY=
POLYGON_API_KEY=

# agent model
OPENAI_API_KEY=
OPENAI_BASE_URL=
MODEL_NAME=qwen3-max-preview

# memory (optional unless --enable-memory is used)
MEMORY_API_KEY=

# experimental: switch selected analyst / risk roles to EvoAgent
EVO_AGENT_IDS=

Notes:

• FINNHUB_API_KEY is required for live mode.
• POLYGON_API_KEY enables long-lived market-store ingestion and refresh helpers.
• MEMORY_API_KEY is only required when long-term memory is enabled.
• EVO_AGENT_IDS currently supports analyst roles plus risk_manager and portfolio_manager, and is intended for staged rollout.

Skill Sandbox Security | 技能沙盒安全

Skill scripts can be executed in multiple sandbox modes controlled by SKILL_SANDBOX_MODE:

Mode	Description	Use Case
none	Direct execution, no isolation	Development only (default)
docker	Docker container isolation	Production with Docker
kubernetes	Kubernetes Pod isolation	Enterprise (reserved)

Default configuration (development):

SKILL_SANDBOX_MODE=none

For production with Docker isolation:

SKILL_SANDBOX_MODE=docker
SKILL_SANDBOX_MEMORY_LIMIT=512m
SKILL_SANDBOX_CPU_LIMIT=1.0
SKILL_SANDBOX_NETWORK=none

When running in none mode, a runtime security warning is displayed on first skill execution as a reminder that scripts execute directly without isolation.

Smoke test for a specific staged EvoAgent rollout target:

python3 scripts/smoke_evo_runtime.py --agent-id fundamentals_analyst

This script starts a temporary runtime, verifies the gateway log contains the selected EvoAgent, checks runtime_state.json, validates the approval wake-up path, and then stops the runtime.

You can also include it in the local release check:

./scripts/check-prod-env.sh --smoke-evo

Without EVO_AGENT_IDS, this release check now runs fundamentals_analyst, risk_manager, and portfolio_manager smoke paths by default.

For a production-style local start flow, you can also use:

./start.sh

The checked-in production label in the deploy scripts is only an example run label. It should not be treated as a canonical root-level runtime directory.

3. Start the stack

Recommended local development flow:

./start-dev.sh

This starts:

• agent_service at http://localhost:8000
• trading_service at http://localhost:8001
• news_service at http://localhost:8002
• runtime_service at http://localhost:8003
• gateway WebSocket at ws://localhost:8765

Then start the frontend in another terminal:

cd frontend && npm run dev

Open http://localhost:5173.

You can also run services manually:

python -m uvicorn backend.apps.agent_service:app --host 0.0.0.0 --port 8000 --reload
python -m uvicorn backend.apps.trading_service:app --host 0.0.0.0 --port 8001 --reload
python -m uvicorn backend.apps.news_service:app --host 0.0.0.0 --port 8002 --reload
python -m uvicorn backend.apps.runtime_service:app --host 0.0.0.0 --port 8003 --reload
# compatibility gateway path, not the recommended primary dev entrypoint
python -m backend.main --mode live --host 0.0.0.0 --port 8765

4. Run backtest or live mode

Backtest:

python backend/main.py --mode backtest --config-name smoke_fullstack --start-date 2025-11-01 --end-date 2025-12-01
python backend/main.py --mode backtest --config-name smoke_fullstack --start-date 2025-11-01 --end-date 2025-12-01 --enable-memory

Live:

python backend/main.py --mode live --config-name live
python backend/main.py --mode live --config-name live --enable-memory
python backend/main.py --mode live --config-name live --interval-minutes 60
python backend/main.py --mode live --config-name live --trigger-time 22:30

Help:

python backend/main.py --help

Offline backtest data

If you want a quick backtest demo without external market APIs, download the offline bundle and unzip it into backend/data:

wget "https://agentscope-open.oss-cn-beijing.aliyuncs.com/ret_data.zip"
unzip ret_data.zip -d backend/data

---

Runtime Data Layout

• Long-lived research data lives in data/market_research.db
• Each run writes run-scoped state under runs/<run_id>/
• runs/<run_id>/BOOTSTRAP.md stores run-specific bootstrap values and prompt body
• runs/<run_id>/state/runtime_state.json stores runtime snapshot state
• runs/<run_id>/team_dashboard/*.json is a compatibility/export layer for dashboard consumers, not the primary runtime source of truth
• ENABLE_DASHBOARD_COMPAT_EXPORTS=false can disable those compatibility JSON exports in controlled environments while keeping runtime state persistence intact

Legacy root-level directories such as live/, production/, and backtest/ should be treated as historical compatibility artifacts, not the default runtime location for new work.

Optional retention control:

RUNS_RETENTION_COUNT=20

Only timestamped run folders like YYYYMMDD_HHMMSS are pruned automatically. Named runs such as live, smoke_fullstack, or reload_demo_* are preserved.

---

Frontend Service Routing

The frontend always uses the control plane and runtime APIs, and can optionally call split services directly for read-only data.

Useful frontend env vars:

VITE_CONTROL_API_BASE_URL=http://localhost:8000/api
VITE_RUNTIME_API_BASE_URL=http://localhost:8003/api/runtime
VITE_NEWS_SERVICE_URL=http://localhost:8002
VITE_TRADING_SERVICE_URL=http://localhost:8001
VITE_WS_URL=ws://localhost:8765

If these are not set, the frontend falls back to its local defaults and compatibility paths where available.

---

Decision Flow

Market data -> independent analyst work -> team communication -> portfolio decision ->
risk review -> execution/settlement -> reflection/memory update

The runtime manager also tracks:

• agent registration and status
• pending approvals
• run events
• current session key

---

Custom Configuration

Add or change analyst roles

1. Define the analyst persona in backend/agents/prompts/analyst/personas.yaml
2. Register the role in backend/config/constants.py
3. Optionally add/update the frontend seat metadata in frontend/src/config/constants.js

Example persona entry:

comprehensive_analyst:
  name: "Comprehensive Analyst"
  focus:
    - multi-factor synthesis
  preferred_tools:
    - get_stock_price
    - get_company_financials
  description: |
    A generalist analyst that combines multiple signals.

Configure per-agent models

Model overrides are configured in .env:

AGENT_SENTIMENT_ANALYST_MODEL_NAME=deepseek-v3.2-exp
AGENT_TECHNICAL_ANALYST_MODEL_NAME=glm-4.6
AGENT_FUNDAMENTALS_ANALYST_MODEL_NAME=qwen3-max-preview
AGENT_VALUATION_ANALYST_MODEL_NAME=Moonshot-Kimi-K2-Instruct
AGENT_RISK_MANAGER_MODEL_NAME=qwen3-max-preview
AGENT_PORTFOLIO_MANAGER_MODEL_NAME=qwen3-max-preview

Run-scoped bootstrap config

Each run can override defaults through runs/<run_id>/BOOTSTRAP.md. The front matter is parsed by backend/config/bootstrap_config.py and can define values such as:

tickers:
  - AAPL
  - MSFT
initial_cash: 100000
margin_requirement: 0.5
max_comm_cycles: 2
schedule_mode: daily
trigger_time: "09:30"
enable_memory: false

Run-scoped workspaces are created automatically at runtime. No manual initialization is required.

---

Project Structure

evotraders/
├── backend/
│   ├── agents/      # agent roles, prompts, skills, workspaces
│   ├── api/         # FastAPI routers
│   ├── apps/        # split service surfaces
│   ├── core/        # pipeline, scheduler, state sync
│   ├── runtime/     # runtime manager and agent runtime state
│   └── services/    # gateway, market/storage/db services
├── frontend/        # React + Vite UI
├── shared/          # shared clients and schemas for split services
├── runs/            # run-scoped state and dashboards
├── data/            # long-lived research artifacts
└── services/README.md

---

Testing

Backend tests live under backend/tests and cover service apps, shared clients, domains, routing, enrichment, gateway support, and runtime support.

Typical commands:

pytest
pytest backend/tests/test_runtime_service_app.py
pytest backend/tests/test_trading_service_app.py

Frontend tests:

cd frontend
npm test

---

License and Disclaimer

大时代 is a research and educational project. Review the repository license before redistribution or commercial use.

Risk warning: this project is not investment advice. Test thoroughly before any real-money deployment. Past performance does not guarantee future returns.