메뉴
Centralized configuration management with validation, versioning, and environment-specific overrides for the ML Engine FX trading system.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | ML Engine Configuration Manager |
| version | 1.0.0 |
| author | Kilo Code |
| description | Centralized configuration management with validation, versioning, and environment-specific overrides for the ML Engine FX trading system. |
| tags | ["configuration","validation","yaml","environment","versioning"] |
| category | infrastructure |
| dependencies | [] |
Centralized configuration management system for the ML Engine FX trading ecosystem with validation, environment-specific overrides, versioning, and automated migration.
The Configuration Manager skill provides robust configuration handling for the ML Engine's complex multi-model ensemble system. It addresses configuration drift, validation gaps, and deployment complexity across different environments (development, testing, production).
# Validate a configuration file
python -m kilocode.skills.ml_engine_config_manager validate config/config_improved_H1.yaml
# Validate with detailed output
python -m kilocode.skills.ml_engine_config_manager validate config/config_improved_H1.yaml --verbose
# Check for deprecated fields
python -m kilocode.skills.ml_engine_config_manager validate config/config_improved_H1.yaml --check-deprecated
# Generate from default template
python -m kilocode.skills.ml_engine_config_manager generate --output config/my_config.yaml
# Generate for specific pair
python -m kilocode.skills.ml_engine_config_manager generate --pair EUR_USD --output config/eur_usd_config.yaml
# Generate with custom settings
python -m kilocode.skills.ml_engine_config_manager generate \
--pair GBP_JPY \
--seq-len 120 \
--epochs 200 \
--output config/gbp_jpy_config.yaml
# Apply production overrides
python -m kilocode.skills.ml_engine_config_manager override \
--base config/config_improved_H1.yaml \
--env production \
--output config/production_config.yaml
# Apply custom overrides
python -m kilocode.skills.ml_engine_config_manager override \
--base config/config_improved_H1.yaml \
--env custom \
--overrides '{"training.epochs": 200, "training.batch_size": 128}' \
--output config/custom_config.yaml
# Visual diff between two configs
python -m kilocode.skills.ml_engine_config_manager diff \
config/config_improved_H1.yaml \
config/config_m1_optimized.yaml
# Diff with color output
python -m kilocode.skills.ml_engine_config_manager diff \
config/config_improved_H1.yaml \
config/config_m1_optimized.yaml \
--color
# Diff with structural comparison (ignoring values)
python -m kilocode.skills.ml_engine_config_manager diff \
config/config_improved_H1.yaml \
config/config_m1_optimized.yaml \
--structure-only
# Migrate to latest schema version
python -m kilocode.skills.ml_engine_config_manager migrate \
config/legacy_config.yaml \
--output config/migrated_config.yaml
# Migrate with dry-run (preview changes)
python -m kilocode.skills.ml_engine_config_manager migrate \
config/legacy_config.yaml \
--output config/migrated_config.yaml \
--dry-run
# Migrate with backup
python -m kilocode.skills.ml_engine_config_manager migrate \
config/legacy_config.yaml \
--output config/migrated_config.yaml \
--backup
The ML Engine configuration follows a hierarchical structure with these main sections:
CLI and command interface settings
buddy:
fetch_days: 365 # Days of historical data to fetch
fetch_candles: 3000 # Number of candles for training
instrument: "EUR_USD" # Default trading pair
Training pipeline configuration
training:
epochs: 100 # Training epochs
batch_size: 64 # Batch size
learning_rate: 0.001 # Learning rate
patience: 20 # Early stopping patience
seq_len: 60 # Sequence length for sliding window
use_class_balanced_loss: true # Use class-balanced loss
use_anti_collapse_loss: true # Use anti-collapse loss
Model architecture settings
model:
transformer_d_model: 32 # Transformer model dimension
transformer_num_heads: 4 # Number of attention heads
transformer_num_layers: 2 # Number of encoder layers
transformer_dff: 64 # Feedforward network dimension
transformer_dropout: 0.2 # Dropout rate
Data processing and feature engineering
data:
use_normalized_features: true # Use normalized features
feature_engineering: # Feature engineering settings
adx_period: 14
rsi_period: 14
macd_fast: 12
macd_slow: 26
FX trading settings and guardrails
fx:
session_windows: # Trading session windows
london: ["08:00", "11:30"]
new_york: ["13:00", "17:00"]
daily_circuit_breaker: # Daily loss limit
enabled: true
max_loss_pips: 100
spread_filter: # Spread filter
enabled: true
max_spread_pips: 3.0
# config/overrides/dev.yaml
training:
epochs: 10 # Quick iteration
batch_size: 32 # Smaller batches
model:
transformer_d_model: 16 # Smaller model
fx:
paper_trading: true # Always paper trade in dev
# config/overrides/test.yaml
training:
epochs: 20
use_ema: false # Disable for speed
use_ewc: false # Disable for speed
data:
fetch_days: 90 # Smaller dataset
# config/overrides/prod.yaml
training:
epochs: 100
patience: 20
use_class_balanced_loss: true
use_anti_collapse_loss: true
use_ema: true
use_ewc: true
fx:
paper_trading: false
session_windows:
london: ["08:00", "11:30"]
new_york: ["13:00", "17:00"]
daily_circuit_breaker:
enabled: true
max_loss_pips: 100
Create custom validation rules in references/custom_validators.py:
from pydantic import validator
class CustomConfig(BaseModel):
# Custom validator for sequence length
@validator('seq_len')
def validate_seq_len(cls, v):
if v < 30:
raise ValueError('seq_len must be at least 30')
if v > 200:
raise ValueError('seq_len cannot exceed 200')
if v % 10 != 0:
raise ValueError('seq_len must be multiple of 10')
return v
Use templates in references/config_templates/ for common scenarios:
# List available templates
python -m kilocode.skills.ml_engine_config_manager list-templates
# Generate from template
python -m kilocode.skills.ml_engine_config_manager generate \
--template aggressive \
--pair GBP_JPY \
--output config/aggressive_gbp_jpy.yaml
# Analyze differences between configs
python -m kilocode.skills.ml_engine_config_manager analyze-diff \
config/config_improved_H1.yaml \
config/config_m1_optimized.yaml \
--output reports/config_diff_analysis.md
# Include impact analysis
python -m kilocode.skills.ml_engine_config_manager analyze-diff \
config/config_improved_H1.yaml \
config/config_m1_optimized.yaml \
--impact \
--output reports/config_impact_analysis.md
# Create backup
python -m kilocode.skills.ml_engine_config_manager backup \
config/config_improved_H1.yaml \
--output backups/config_backup_20240129.yaml
# List backups
python -m kilocode.skills.ml_engine_config_manager list-backups
# Restore from backup
python -m kilocode.skills.ml_engine_config_manager restore \
backups/config_backup_20240129.yaml \
--output config/config_improved_H1.yaml
The Configuration Manager provides clear error messages for common issues:
ValidationError: training.epochs must be at least 10
ValidationError: model.transformer_d_model must be multiple of 8
ValidationError: fx.session_windows.london must be in format HH:MM
MigrationError: Cannot migrate from version 1.0 to 3.0 (unsupported)
MigrationError: Field 'legacy_field' is deprecated and has no migration path
OverrideError: Override path 'training.invalid_field' does not exist in base config
OverrideError: Override type mismatch: expected int, got str
Solution: Use --verbose flag to see detailed validation context
python -m kilocode.skills.ml_engine_config_manager validate config.yaml --verbose
Solution: Use --dry-run to preview migration changes
python -m kilocode.skills.ml_engine_config_manager migrate old.yaml --dry-run
Solution: Check override path matches base config structure exactly
# Check structure first
python -m kilocode.skills.ml_engine_config_manager show-structure config.yaml
The Configuration Manager integrates seamlessly with ML Engine:
# In main.py
from kilocode.skills.ml_engine_config_manager import ConfigManager
# Load and validate config
config = ConfigManager.load('config/config_improved_H1.yaml')
# Access config with type hints
epochs = config.training.epochs
seq_len = config.data.seq_len
# Apply environment overrides
config = ConfigManager.apply_overrides(
config,
environment='production'
)
# Validate before training
errors = ConfigManager.validate(config)
if errors:
print(f"Configuration errors: {errors}")
sys.exit(1)
Triggered on runtime errors in the scan loop. Diagnoses the root cause, writes a targeted fix to src/, and reports what changed. The calling harness runs pytest before merging — if tests fail, the fix is reverted automatically.
Post-trade feedback loop system: shaped-reward logging, diagnostics, and self-heal. Use when investigating trade outcomes, checking system health, or understanding automated corrective actions.
Buddy MCP server: 15 tools for system observability, feedback loop control, OANDA state, and knowledge management. Use when querying system health, investigating trades, or triggering diagnostics/self-heal.
Spawned automatically on trade close by ClaudeReflectionHandler. Reads the just-closed trade, compares prediction to outcome, and writes structured learnings / rule drafts / weight proposals that Buddy picks up on the next scan cycle.
Triggered on runtime errors in the scan loop. Diagnoses the root cause, writes a targeted fix to src/, and reports what changed. The calling harness runs pytest before merging — if tests fail, the fix is reverted automatically.
Spawned automatically on trade close by ClaudeReflectionHandler. Reads the just-closed trade, compares prediction to outcome, and writes structured learnings / rule drafts / weight proposals that Buddy picks up on the next scan cycle.