claude-evolve 1.3.40 → 1.3.42

package/bin/claude-evolve-edit

@@ -25,23 +25,27 @@ USAGE:
 SELECTORS:
   gen01, gen02, etc.  Target specific generation
   all                 Target all generations
-  failed              Target all candidates with failed status
+  failed              Target all candidates with failed status (includes retries)
   complete            Target all candidates with complete status
   pending             Target all candidates with pending status
   running             Target all candidates with running status
 
 ACTIONS:
-  failed    Mark candidates as failed (keeps scores)
-  complete  Mark candidates as complete (keeps scores)  
-  pending   Mark candidates as pending (keeps scores)
-  reboot    Reset completely (delete .py files, clear scores, set pending)
+  failed           Mark candidates as failed (keeps scores)
+  complete         Mark candidates as complete (keeps scores)  
+  pending          Mark candidates as pending (keeps scores)
+  failed-retry1    Mark candidates for retry attempt 1 (bug fixing)
+  failed-retry2    Mark candidates for retry attempt 2 (bug fixing)
+  failed-retry3    Mark candidates for retry attempt 3 (bug fixing)
+  reboot           Reset completely (delete .py files, clear scores, set pending)
 
 EXAMPLES:
-  claude-evolve edit gen03 failed     # Mark all gen03 as failed
-  claude-evolve edit failed pending   # Reset all failed candidates to pending
-  claude-evolve edit complete failed  # Mark all complete as failed for re-run
-  claude-evolve edit all pending      # Mark everything as pending for re-run
-  claude-evolve edit gen02 reboot     # Full reset of gen02 (delete files + clear data)
+  claude-evolve edit gen03 failed        # Mark all gen03 as failed
+  claude-evolve edit failed pending      # Reset all failed candidates to pending
+  claude-evolve edit failed failed-retry1 # Convert failed to retry status (bug fixing)
+  claude-evolve edit complete failed     # Mark all complete as failed for re-run
+  claude-evolve edit all pending         # Mark everything as pending for re-run
+  claude-evolve edit gen02 reboot        # Full reset of gen02 (delete files + clear data)
 
 DESCRIPTION:
   This command helps manage evolution runs when you need to re-evaluate candidates.
@@ -73,9 +77,9 @@ fi
 
 # Validate action
 case "$ACTION" in
-  failed|complete|pending|reboot) ;;
+  failed|complete|pending|failed-retry1|failed-retry2|failed-retry3|reboot) ;;
   *)
-    echo "[ERROR] Action must be one of: failed, complete, pending, reboot" >&2
+    echo "[ERROR] Action must be one of: failed, complete, pending, failed-retry1, failed-retry2, failed-retry3, reboot" >&2
     exit 1
     ;;
 esac
@@ -100,12 +104,14 @@ update_candidates_status() {
 import csv
 import sys
 import os
+import re
 
 csv_file = '$FULL_CSV_PATH'
 selector = '$selector'
 new_status = '$new_status'
 clear_scores = '$clear_scores' == 'true'
 
+
 try:
     # Read CSV
     with open(csv_file, 'r') as f:
@@ -139,6 +145,8 @@ try:
             # Status selector
             if selector == 'pending':
                 matches = current_status == '' or current_status == 'pending'
+            elif selector == 'failed':
+                matches = current_status.startswith('failed')
             else:
                 matches = current_status == selector
         
@@ -210,10 +218,12 @@ delete_evolution_files() {
     candidates_to_delete=$("$PYTHON_CMD" -c "
 import csv
 import sys
+import re
 
 csv_file = '$FULL_CSV_PATH'
 selector = '$selector'
 
+
 try:
     with open(csv_file, 'r') as f:
         reader = csv.reader(f)
@@ -231,6 +241,8 @@ try:
             matches = False
             if selector == 'pending':
                 matches = current_status == '' or current_status == 'pending'
+            elif selector == 'failed':
+                matches = current_status.startswith('failed')
             else:
                 matches = current_status == selector
             
@@ -278,6 +290,15 @@ case "$ACTION" in
   pending)
     update_candidates_status "$SELECTOR" "" "false"  # Empty status means pending
     ;;
+  failed-retry1)
+    update_candidates_status "$SELECTOR" "failed-retry1" "false"
+    ;;
+  failed-retry2)
+    update_candidates_status "$SELECTOR" "failed-retry2" "false"
+    ;;
+  failed-retry3)
+    update_candidates_status "$SELECTOR" "failed-retry3" "false"
+    ;;
   reboot)
     echo "[INFO] Performing full reboot of '$SELECTOR'..."
     delete_evolution_files "$SELECTOR"

package/bin/claude-evolve-ideate

@@ -328,10 +328,25 @@ CRITICAL CSV FORMAT RULES:
   * Are impossible given the codebase structure
   * Would break the algorithm interface requirements
 
+⚠️ AVOID ONLY: Kelly floor/cap adjustments that assume leverage > 1.0 (these get clamped and have no effect)
+
+✅ EXPLORE ANY CREATIVE IDEAS INCLUDING:
+- **Machine Learning**: Neural networks, decision trees, ensemble methods (use train() method properly)
+- **New Indicators**: Custom combinations, alternative calculations, multi-timeframe signals
+- **Market Regime Detection**: VIX patterns, correlation shifts, volume analysis, cross-asset signals
+- **Risk Management**: Dynamic stops, correlation-based position sizing, drawdown protection
+- **Timing**: Time-of-day effects, calendar patterns, volatility timing
+- **Alternative Strategies**: New sub-strategies, momentum scoring, mean reversion variants
+- **Cross-Asset Signals**: Bond yields, sector rotation, crypto correlations
+- **Multi-Timeframe**: Combining 30m/1h/daily signals for confirmation
+- **Advanced Exits**: Profit targets, time-based exits, volatility-based exits
+
+Think outside the box! The codebase supports sophisticated approaches - be creative and ambitious.
+
 Example descriptions:
-- Use ensemble of 3 random forests with different feature subsets
-- Replace neural network with gradient boosting decision trees
-- Implement Monte Carlo tree search for feature selection
+- Train LSTM network on 30-day OHLCV sequences to predict next-day direction probability
+- Add cross-correlation filter that reduces positions when TQQQ correlation with QQQ breaks down
+- Implement intraday momentum using 30-minute data to adjust daily position sizes
 
 Add exactly $count rows to the CSV file now."
 
@@ -391,10 +406,22 @@ CRITICAL CSV FORMAT RULES:
   * What changes made this algorithm successful vs its parent
   * What parameter ranges make sense given the implementation
 
+⚠️ AVOID ONLY: Kelly floor/cap adjustments that assume leverage > 1.0 (these get clamped and have no effect)
+
+✅ EXPLORE PARAMETER TUNING INCLUDING:
+- **Entry/Exit Thresholds**: IBS_BUY_THRESHOLD, LARISSA_LOW_THRESHOLD, RSI levels, etc.
+- **Indicator Periods**: ATR_PERIOD, RSI_PERIOD, moving average lengths, etc.
+- **Strategy Weights**: Emphasize best performers or rebalance for diversification
+- **Risk Parameters**: Stop levels, timeout periods, correlation thresholds
+- **Regime Parameters**: Volatility thresholds, trend detection sensitivity
+- **ML Hyperparameters**: Learning rates, network sizes, ensemble weights (if using ML)
+
+Be creative with parameter combinations and ranges - the system is sophisticated!
+
 Example descriptions:
-- Increase learning rate from 0.001 to 0.01 for faster convergence
-- Reduce batch size from 32 to 16 to improve gradient estimates
-- Set dropout rate to 0.3 instead of 0.1 to prevent overfitting
+- Lower IBS_BUY_THRESHOLD from 0.15 to 0.12 to enter deeper oversold conditions
+- Increase TRS_RSI_PERIOD from 2 to 3 for smoother RSI signals
+- Raise WEIGHT_TDD from 0.38 to 0.42 to emphasize best performing strategy
 
 Add exactly $count parameter tuning rows to the CSV file now."
 
@@ -454,10 +481,23 @@ CRITICAL CSV FORMAT RULES:
   * What architectural decisions led to this algorithm's success
   * Which components are essential vs which can be replaced
 
+⚠️ AVOID ONLY: Kelly floor/cap adjustments that assume leverage > 1.0 (these get clamped and have no effect)
+
+✅ EXPLORE STRUCTURAL INNOVATIONS INCLUDING:
+- **Algorithm Architecture**: Replace sub-strategies, change combination logic, add new layers
+- **Indicator Swaps**: RSI → Stochastic, SMA → Hull MA, Bollinger → Keltner, etc.
+- **Machine Learning Integration**: Add neural networks, decision trees, reinforcement learning
+- **Market Regime Systems**: Multi-regime detection, regime-specific strategies
+- **Risk Management Overhauls**: Portfolio heat, correlation-based sizing, adaptive stops
+- **Multi-Asset Integration**: Cross-asset signals, sector rotation, bond/equity relationships
+- **Time-Based Innovations**: Intraday patterns, calendar effects, volatility timing
+
+The codebase is flexible - think architecturally about major improvements!
+
 Example descriptions:
-- Replace linear layers with convolutional layers for spatial feature learning
-- Use bidirectional LSTM instead of unidirectional for better context
-- Add residual connections between layers to improve gradient flow
+- Replace 2-period RSI with LSTM-predicted momentum scores for TRS strategy
+- Add ensemble voting system where sub-strategies vote on market regime
+- Implement hierarchical risk budgeting with correlation-adjusted position sizing
 
 Add exactly $count structural modification rows to the CSV file now."
 
@@ -517,10 +557,23 @@ CRITICAL CSV FORMAT RULES:
   * Understand which components are compatible for merging
   * Ensure the combined approach is technically feasible in the codebase
 
+⚠️ AVOID ONLY: Kelly floor/cap adjustments that assume leverage > 1.0 (these get clamped and have no effect)
+
+✅ EXPLORE CREATIVE COMBINATIONS INCLUDING:
+- **Strategy Fusion**: Merge successful sub-strategies, combine entry/exit logic
+- **Indicator Blending**: Mix different technical analysis approaches from successful algorithms
+- **Machine Learning Hybrids**: Combine ML predictions with rule-based systems
+- **Multi-Regime Integration**: Blend different regime detection methods
+- **Risk System Combinations**: Merge multiple risk management approaches
+- **Cross-Asset Blends**: Combine internal signals with external market data
+- **Multi-Timeframe Fusion**: Blend signals from different time horizons
+
+Think creatively about what worked in different algorithms and how to combine them!
+
 Example descriptions:
-- Combine ensemble voting from algorithm 3 with feature selection from algorithm 5
-- Use the attention mechanism from algorithm 2 with the optimizer from algorithm 4
-- Merge the preprocessing pipeline from algorithm 1 with the architecture from algorithm 6
+- Combine VIX regime filter from algorithm 3 with LSTM predictions from algorithm 5
+- Merge volatility regime detection with machine learning momentum scoring
+- Integrate multi-timeframe signals with correlation-based position adjustments
 
 Add exactly $count hybrid combination rows to the CSV file now."
 
@@ -591,6 +644,20 @@ CRITICAL CSV FORMAT RULES:
 - Leave performance and status fields completely empty (just commas)
 - Use proper CSV quoting only when descriptions contain commas
 
+⚠️ AVOID ONLY: Kelly floor/cap adjustments that assume leverage > 1.0 (these get clamped and have no effect)
+
+✅ EXPLORE ALL CREATIVE POSSIBILITIES INCLUDING:
+- **Machine Learning**: Neural networks, ensemble methods, reinforcement learning (use train() method)
+- **Advanced Indicators**: Custom combinations, multi-timeframe signals, cross-asset indicators
+- **Market Regime Detection**: VIX patterns, correlation analysis, volatility clustering
+- **Risk Management**: Dynamic stops, portfolio heat, correlation-based position sizing
+- **Alternative Strategies**: New sub-strategies, momentum variants, mean reversion innovations
+- **Multi-Asset Signals**: Sector rotation, bond yields, commodity signals
+- **Time-Based Patterns**: Intraday effects, calendar anomalies, volatility timing
+- **Parameter Optimization**: Entry thresholds, indicator periods, strategy weights
+
+Think outside the box - the system is sophisticated and can handle advanced approaches!
+
 Add exactly $TOTAL_IDEAS algorithm variation rows to the CSV file now."
 
   echo "[INFO] Generating $TOTAL_IDEAS ideas (legacy mode)..."

package/bin/claude-evolve-main

@@ -55,7 +55,7 @@ show_help() {
 claude-evolve - AI-powered algorithm evolution tool
 
 USAGE:
-    claude-evolve [--config=PATH] [COMMAND] [OPTIONS]
+    claude-evolve [--working-dir=PATH] [COMMAND] [OPTIONS]
 
 COMMANDS:
     setup              Initialize evolution workspace
@@ -69,16 +69,16 @@ COMMANDS:
     help               Show this help message
 
 GLOBAL OPTIONS:
-    --config=PATH  Use alternate config file (default: evolution/config.yaml)
-    -h, --help     Show help message
-    -v, --version  Show version
+    --working-dir=PATH  Use alternate working directory (default: evolution/)
+    -h, --help          Show help message
+    -v, --version       Show version
 
 EXAMPLES:
     claude-evolve setup
     claude-evolve ideate 5
     claude-evolve run --timeout 300
     claude-evolve analyze --open
-    claude-evolve --config=evolution-abc/config.yaml run
+    claude-evolve --working-dir=evolution-abc run
 
 For more information, visit: https://github.com/anthropics/claude-evolve
 EOF
@@ -111,19 +111,19 @@ show_menu() {
 }
 
 # Parse global options
-CONFIG_FILE=""
+WORKING_DIR=""
 while [[ $# -gt 0 ]] && [[ "$1" =~ ^-- ]]; do
   case "$1" in
-  --config)
+  --working-dir)
     if [[ -z ${2:-} ]]; then
-      echo -e "${RED}[ERROR] --config requires a file path${NC}" >&2
+      echo -e "${RED}[ERROR] --working-dir requires a directory path${NC}" >&2
       exit 1
     fi
-    CONFIG_FILE="$2"
+    WORKING_DIR="$2"
     shift 2
     ;;
-  --config=*)
-    CONFIG_FILE="${1#*=}"
+  --working-dir=*)
+    WORKING_DIR="${1#*=}"
     shift
     ;;
   *)
@@ -132,9 +132,11 @@ while [[ $# -gt 0 ]] && [[ "$1" =~ ^-- ]]; do
   esac
 done
 
-# Export CONFIG_FILE for subcommands
-if [[ -n $CONFIG_FILE ]]; then
-  export CLAUDE_EVOLVE_CONFIG="$CONFIG_FILE"
+# Export config file path for subcommands
+if [[ -n $WORKING_DIR ]]; then
+  # Remove trailing slash if present
+  WORKING_DIR="${WORKING_DIR%/}"
+  export CLAUDE_EVOLVE_CONFIG="$WORKING_DIR/config.yaml"
 fi
 
 # Check for updates (quick, non-blocking)

package/bin/claude-evolve-run

@@ -236,7 +236,6 @@ cleanup_workers() {
       # Worker finished
       if wait "$pid" 2>/dev/null; then
         echo "[DISPATCHER] Worker $pid completed successfully"
-        consecutive_failures=0  # Reset counter on success
       else
         local exit_code=$?
         if [[ $exit_code -eq 2 ]]; then
@@ -244,24 +243,7 @@ cleanup_workers() {
           # Rate limits don't count as consecutive failures
         else
           echo "[DISPATCHER] Worker $pid failed with exit code $exit_code"
-          ((consecutive_failures++))
-          
-          # Check if we've hit the failure limit
-          if [[ $consecutive_failures -ge $MAX_CONSECUTIVE_FAILURES ]]; then
-            echo "" >&2
-            echo "🚨🚨🚨 EVOLUTION STOPPED: TOO MANY FAILURES 🚨🚨🚨" >&2
-            echo "ERROR: $consecutive_failures consecutive worker failures detected" >&2
-            echo "ERROR: This indicates a systemic problem (Claude API issues, evaluator bugs, etc.)" >&2
-            echo "ERROR: Check logs and fix issues before restarting evolution" >&2
-            echo "🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨" >&2
-            echo "" >&2
-            
-            # Shutdown all workers and exit
-            shutdown_workers
-            exit 1
-          fi
-          
-          echo "[DISPATCHER] Consecutive failures: $consecutive_failures/$MAX_CONSECUTIVE_FAILURES"
+          # With retry mechanism, failures are normal - just keep processing
         fi
       fi
     fi
@@ -406,9 +388,8 @@ else
   echo "[DISPATCHER] No cleanup issues detected - proceeding with run"
 fi
 
-# Consecutive failure tracking
-consecutive_failures=0
-MAX_CONSECUTIVE_FAILURES=5
+# With retry mechanism, we don't need consecutive failure tracking
+# Failures are handled gracefully through the retry system
 
 # Main dispatch loop
 while true; do

package/bin/claude-evolve-status

@@ -105,12 +105,23 @@ fi
 "$PYTHON_CMD" -c "
 import csv
 import sys
+import re
 
 csv_file = '$FULL_CSV_PATH'
 show_brief = '$SHOW_BRIEF' == 'true'
 show_winner_only = '$SHOW_WINNER_ONLY' == 'true'
 evolution_context = '$EVOLUTION_CONTEXT'
 
+def normalize_status(status):
+    '''Convert retry statuses to base status for counting.'''
+    if status.startswith('failed'):
+        return 'failed'
+    return status
+
+def is_retry_status(status):
+    '''Check if status is a retry status.'''
+    return bool(re.match(r'^failed-retry[0-9]+$', status))
+
 try:
     with open(csv_file, 'r') as f:
         reader = csv.reader(f)
@@ -126,6 +137,7 @@ try:
     all_candidates = []
     stats_by_gen = {}
     total_stats = {'pending': 0, 'complete': 0, 'failed': 0, 'running': 0}
+    retry_count = 0
     
     for row in rows[1:]:
         if len(row) >= 1 and row[0]:  # Must have an ID
@@ -139,19 +151,26 @@ try:
                 status = row[4] if len(row) > 4 and row[4] else 'pending'
                 performance = row[3] if len(row) > 3 and row[3] else ''
                 
+                # Normalize status (failed-retry* becomes failed)
+                normalized_status = normalize_status(status)
+                
+                # Count retries
+                if is_retry_status(status):
+                    retry_count += 1
+                
                 # Track by generation
                 if gen not in stats_by_gen:
                     stats_by_gen[gen] = {'pending': 0, 'complete': 0, 'failed': 0, 'running': 0}
                 
-                if status in stats_by_gen[gen]:
-                    stats_by_gen[gen][status] += 1
-                    total_stats[status] += 1
+                if normalized_status in stats_by_gen[gen]:
+                    stats_by_gen[gen][normalized_status] += 1
+                    total_stats[normalized_status] += 1
                 else:
                     stats_by_gen[gen]['pending'] += 1
                     total_stats['pending'] += 1
                 
                 # Collect for winner analysis (only completed with valid scores)
-                if status == 'complete' and performance:
+                if normalized_status == 'complete' and performance:
                     try:
                         score = float(performance)
                         description = row[2] if len(row) > 2 else 'No description'
@@ -183,7 +202,10 @@ try:
         print(f'📊 OVERALL: {total_candidates} total candidates')
         print(f'   • {total_stats[\"pending\"]} pending')
         print(f'   • {total_stats[\"complete\"]} complete') 
-        print(f'   • {total_stats[\"failed\"]} failed')
+        failed_display = f'{total_stats[\"failed\"]} failed'
+        if retry_count > 0:
+            failed_display += f' ({retry_count} retries)'
+        print(f'   • {failed_display}')
         print(f'   • {total_stats[\"running\"]} running')
         print()