config.yaml Reference

config.yaml is the main configuration file for the perpetual futures agent. Copy from the example:

cp config.yaml.example config.yaml

LLM Configuration

llm:
  client_type: langchain_nvidia   # openai | cloudflare | google | litellm | nvidia
  model: deepseek-ai/deepseek-v3.2
  temperature: 0.2                 # low temperature recommended for trading decisions

Supported client_type Values

Value	Provider	Notes
openai	OpenAI	Requires OPENAI_API_KEY
langchain_nvidia	NVIDIA NIM	Requires NVIDIA_API_KEY
google	Google Gemini	Requires GOOGLE_API_KEY
cloudflare	Cloudflare Workers AI	Requires CLOUDFLARE_* vars
litellm	LiteLLM proxy	Flexible multi-provider routing

Temperature

Keep temperature at 0.1–0.3 for trading decisions. Higher values produce more creative but less reliable JSON outputs.

Trading Parameters

trading:
  symbols: [BTC, ETH]       # use simple symbols, NOT pair format (not BTC/USDT)
  max_trade_amount: 100     # maximum USD per single trade
  max_leverage: 10          # maximum leverage multiplier
  limit_order_enabled: false  # use limit orders for entry (vs. market orders)

Symbol Format

Use BTC, ETH — not BTC/USDT or BTC-PERP. Hyperliquid uses simple asset symbols.

Leverage Notes

• Different assets have different maximum leverage on Hyperliquid
• If max_leverage exceeds the asset's limit, orders will be rejected
• Start conservative: max_leverage: 3 or 5 until you're confident in the strategy

Scheduler

scheduler:
  interval_minutes: 3    # fallback polling interval between decisions

The scheduler is the fallback decision loop. When Market Monitor is enabled, volatility spikes trigger decisions immediately without waiting for the interval.

Prompt Strategy

prompt:
  set: nof1-improved    # see table below

Value	Description
default	Standard FinCoT reasoning chain
conservative	Hold on any trend divergence, requires R:R ≥ 2.0
aggressive	Enters on 3 conditions, accepts R:R ≥ 1.2
nof1	Enhanced strategy with full FinCoT integration
nof1-improved	Recommended — complete FinCoT + enhanced data integration
realtime	Prioritizes price action over lagging indicators
realtime-eng	English version of realtime

Enhanced Analysis

enhanced_analysis:
  enabled: true    # enables CEX funding rate, on-chain data collection

This is the base switch for CEX Signals and Regime Adaptive. Other features that depend on this:

enhanced_analysis.enabled: true
  ├── debate.enabled: true         # independent switch
  └── regime_adaptive.enabled: true  # depends on enhanced_analysis

Debate Agent

debate:
  enabled: false    # adds 2 extra LLM calls per decision cycle

See Bull/Bear Debate for details.

Regime Adaptive

regime_adaptive:
  enabled: false
  # Optional parameter overrides:
  # trending:
  #   signal_threshold: 0.5
  #   min_confidence: 0.35
  #   max_leverage: 10
  # ranging:
  #   signal_threshold: 0.75
  #   max_leverage: 5
  # volatile:
  #   signal_threshold: 0.85
  #   max_leverage: 3

See Market Regime Adaptive for details.

Account Protection

account_protection:
  enabled: true
  max_drawdown_pct: 0.10       # pause trading if account drops 10% from peak
  max_daily_loss_pct: 0.05     # pause trading if daily loss exceeds 5%
  max_position_hours: 48       # force-close positions held longer than 48 hours

Always Keep This Enabled

account_protection is a critical safety mechanism. Only disable it if you fully understand the risks.

Market Monitor

market_monitor:
  enabled: false
  check_interval_seconds: 30   # price check frequency
  alert_threshold_pct: 3.0     # HIGH alert threshold
  elevated_threshold_pct: 1.5  # ELEVATED threshold (logged, no trigger)
  extreme_threshold_pct: 5.0   # EXTREME threshold
  cooldown_minutes: 5          # cooldown after a trigger
  reference_window_minutes: 10 # price baseline window

See Market Active Monitoring for alert levels and behavior.

Review Agent

review_agent:
  # 6a: Dual-granularity reflection
  instant_reflection_enabled: false    # after every closed position
  weekly_reflection_enabled: false     # weekly LLM-based strategy review
  weekly_reflection_day: 0             # 0=Monday
  weekly_reflection_hour: 8

  # 6b: Regime-aware memory
  regime_aware_enabled: false
  regime_mismatch_factor: 0.4          # weight penalty for regime mismatch

  # 6c: Confirmation bias protection
  bias_protection_enabled: false
  max_positive_ratio: 0.7              # max ratio of positive lessons in memory
  negative_confidence_boost: 1.15      # boost factor for negative lessons

  # 6d: Fact-subjective split
  fact_subjective_split_enabled: false
  trending_subjective_boost: 1.3
  ranging_factual_boost: 1.3

  # 6e: Prompt meta-reflection
  prompt_meta_reflection_enabled: false
  prompt_optimization_dir: "logs/prompt_optimization"

See Review & Reflection System for all 6 enhancements.

Data Configuration

data:
  timeframe: 1h    # OHLCV candle timeframe for main decisions

Complete Example

llm:
  client_type: langchain_nvidia
  model: deepseek-ai/deepseek-v3.2
  temperature: 0.2

trading:
  symbols: [BTC, ETH]
  max_trade_amount: 100
  max_leverage: 10
  limit_order_enabled: false

scheduler:
  interval_minutes: 3

data:
  timeframe: 1h

prompt:
  set: nof1-improved

enhanced_analysis:
  enabled: true

debate:
  enabled: true

regime_adaptive:
  enabled: true

account_protection:
  enabled: true
  max_drawdown_pct: 0.10
  max_daily_loss_pct: 0.05
  max_position_hours: 48

market_monitor:
  enabled: true
  check_interval_seconds: 30
  alert_threshold_pct: 3.0
  cooldown_minutes: 5

review_agent:
  instant_reflection_enabled: true
  weekly_reflection_enabled: true
  regime_aware_enabled: true
  bias_protection_enabled: true
  fact_subjective_split_enabled: true
  prompt_meta_reflection_enabled: false