Implement self-improvement learning loop with trade analysis and adaptive config tuning

[Raynergy-svg]

Summary

This PR implements a complete self-improvement learning loop for the Buddy FX trading bot, enabling automatic trade outcome analysis, pattern extraction, rule promotion, and adaptive configuration tuning. The system closes the feedback loop: trades → learnings → rules → config adjustments → better trades.

Key Changes

Core Learning Infrastructure

• StateEngine (src/scanner/automation/state_engine.py): Manages cross-session state persistence in .claude/state.json, enabling continuity between trading sessions with portfolio snapshots and improvement focus tracking.

• LearningEngine (src/scanner/automation/learning_engine.py): Analyzes closed trades and extracts actionable patterns:

  • analyze_trade(): Examines trade outcomes against 5 rules (SL too tight, TP too fast, uncertainty warnings, high consensus, model disagreement)
  • check_promotions(): Identifies patterns with 3+ observations and promotes them to trading rules
  • deep_analyze_loss(): Uses local 3B LLM (Qwen2.5) for detailed analysis of significant losses (>$100)
  • audit() / consolidate(): Manages file growth and archives promoted learnings

• ConfigTuner (src/scanner/automation/config_tuner.py): Reads promoted rules and applies bounded config adjustments:

  • Parses rule patterns and maps them to config fields
  • Enforces safety bounds to prevent runaway tuning (e.g., atr_sl_multiplier: 0.5-2.0)
  • Tracks applied rules to avoid duplicate adjustments

Observation & Metrics

• ObservationLog (src/scanner/automation/observation_log.py): Captures interesting market patterns during scans (regime changes, agent disagreement, near-misses) even when no trade is taken, building a pattern library for future analysis.

• ImprovementTracker (src/scanner/automation/improvement_tracker.py): Records session-level metrics (win rate, P/L, learnings added, rules promoted) to improvement_log.jsonl and computes rolling trends to measure whether the learning system itself is improving.

CLI & Integration

• learn_commands.py (cli/learn_commands.py): New buddy learn CLI command with subcommands:

  • --analyze: Run trade outcome analysis on journal
  • --promote: Check for and promote qualifying patterns
  • --consolidate: Archive and manage file growth
  • --report: Display improvement metrics
  • --status: Show current state and recent learnings

• continuous.py integration: Learning loop runs automatically in _run_smart_loop() after each RL sync, analyzing newly closed trades and applying config tuning before the next scan cycle.

Configuration & Engine Updates

• engine.py: Per-pair adaptive SL/TP multipliers (US-006) read from pair_sl_tp_config.json, allowing pair-specific learning overrides.
• config.py: Added enable_llm_trade_analysis flag to control deep LLM analysis for significant losses.
• argparser.py: Added learn command with --analyze, --promote, --consolidate, --report, --status flags.
• main.py: Dispatch handler for learn subcommand.

Notable Implementation Details

1. Pattern Extraction: Learning entries are formatted as [date] category | pattern_key → action, enabling regex-based pattern counting and promotion when 3+ observations match.

2. Rule Promotion: Patterns are automatically promoted to .claude/rules/trading.md with specific config change recommendations (e.g., "Increase atr_sl_multiplier by 0.1 for EURUSD").

3. Safety Bounds: All config adjustments are clamped to predefined bounds to prevent extreme tuning that could destabilize trading.

4. LLM Integration: Deep loss analysis uses llm_providers.select_buddy_provider_name() to prefer local 3B model (no API tokens required), with graceful fallback if unavailable.

5. **Cross-Session Continu

https://claude.ai/code/session_01AsBRYCkTiUN4K4WXMSXVkb