🐟 Trade Simulator (MiroFish Architecture)
Multi-agent scenario analysis for traders. Not a spreadsheet — a behavioral simulation. Built on MiroFish's swarm intelligence architecture, adapted from social simulation to market simulation.
MiroFish Integration
This skill implements MiroFish's 5-stage prediction pipeline, replacing social media environments with financial markets:
| MiroFish Stage | Original (Social) | Our Adaptation (Markets) |
|---|---|---|
| 1. Graph Construction | Zep knowledge graph from news/docs | Market State Graph from live Coinglass/HL data |
| 2. Environment Setup | Twitter/Reddit agent profiles | Market participant profiles (Whale, MM, Retail, etc.) |
| 3. Simulation | OASIS dual-platform social interaction | Round-based market interaction with LLM reasoning |
| 4. Report Generation | ReACT report with Zep tools | ReACT report with market data tools |
| 5. Deep Interaction | Interview any social agent | Interview any market participant |
Key MiroFish Patterns Used
- LLM-driven agent reasoning (from
oasis_profile_generator.py) — agents don't use if/else rules. Each agent has a persona prompt and "thinks" each round via LLM call - Simulation config auto-generation (from
simulation_config_generator.py) — describe scenario in natural language, LLM generates agent roster, parameters, event timeline, activity patterns - ReACT report generation (from
report_agent.py) — multi-step reasoning with tool use: plan outline → generate sections → cite evidence → synthesize predictions - Post-simulation interviews (from
zep_tools.pyInterview system) — chat with any agent after simulation to understand their reasoning - Knowledge graph backbone — entities, relationships, and facts structured for agent retrieval (we use in-memory graph instead of Zep Cloud)
What We Don't Use
- ❌ OASIS / camel-ai (social media simulation runtime — irrelevant to markets)
- ❌ Zep Cloud (replaced with local in-memory knowledge graph)
- ❌ Flask frontend (we output to agent conversation)
- ❌ Twitter/Reddit environments (replaced with market environment)
Architecture
skills/trade-simulator/
├── SKILL.md # This file
└── scripts/
├── mirofish_engine.py # Core engine — 5-stage pipeline
├── market_graph.py # Stage 1: Market state graph builder
├── profile_generator.py # Stage 2: LLM agent profile generation
├── simulation_runner.py # Stage 3: Round-based market simulation
├── report_agent.py # Stage 4: ReACT report generation
└── interview.py # Stage 5: Post-sim agent interviewsUsage
Quick Scenario Analysis
Agent: "Run a trade simulation: What happens to my BTC short if ETF inflows spike 500%?"The engine will:
- Build market state graph from live data (OI, funding, liquidations, whale positions)
- Auto-generate 5-8 market participant agents calibrated to current conditions
- Run 6-round simulation where each agent LLM-reasons about their actions
- Generate ReACT analysis report with turning points, cascade analysis, recommendations
- Offer interactive interviews with any simulated agent
Supported Scenarios
- Directional shocks: "What if BTC pumps/dumps 10-20%?"
- Catalyst events: "What if ETF inflows spike?" / "What if Tether depegs?"
- Market structure: "What if funding goes extreme?" / "What if OI doubles?"
- Portfolio stress: "How does my portfolio react to a black swan?"
Interview Mode
After any simulation:
Agent: "Interview the whale agent — why did they cover at round 4?"
Agent: "Ask the market maker about their liquidity decision"Data Sources (Live)
| Data | Tool | What It Feeds |
|---|---|---|
| Open Interest | cg_open_interest() | Market leverage state |
| Funding Rates | funding_rate() | Positioning sentiment |
| Liquidation Levels | cg_liquidations() | Cascade trigger points |
| Whale Positions | cg_hyperliquid_whale_positions() | Whale agent calibration |
| Long/Short Ratios | long_short_ratio() | Crowd positioning |
| Orderbook Depth | hl_orderbook() | MM agent calibration |
| ETF Flows | cg_btc_etf_flows() | Institutional flow context |
| Price/OHLC | cg_ohlc_history() | Price context |
| Social Sentiment | lunar_coin() | Retail agent behavior |
Workflow
- Collect live market data using tools above
- Run simulation:
python3 skills/trade-simulator/scripts/mirofish_engine.py
- Pass market data + scenario + user positions as JSON - Engine runs all 5 MiroFish stages - Returns structured results (agent actions, report, interview-ready state)
- Present results with key insights, PnL impact, risk warnings
- Offer interviews — user can interrogate any agent