claude-evolve 1.5.2 → 1.5.4

package/bin/claude-evolve-ideate

@@ -84,39 +84,76 @@ call_ai_for_ideation() {
     original_csv_count=0
   fi
   
-  # Use centralized AI library
-  local ai_output
-  ai_output=$(call_ai_with_round_robin "$prompt" "ideate" "$gen_num")
-  local ai_exit_code=$?
+  echo "[DEBUG] Original CSV has $original_csv_count data rows" >&2
   
-  # Handle special exit codes
-  # No special handling for exit codes anymore
+  # Get models for ideation
+  local model_list
+  model_list=$(get_models_for_command "ideate")
+  local models=()
+  read -ra models <<< "$model_list"
   
-  if [[ $ai_exit_code -eq 0 ]]; then
-    # For ideation, we need to verify the CSV file was actually modified
+  if [[ ${#models[@]} -eq 0 ]]; then
+    echo "[ERROR] No models configured for ideation" >&2
+    return 1
+  fi
+  
+  # Calculate starting index for round-robin
+  local num_models=${#models[@]}
+  local start_index=$((gen_num % num_models))
+  
+  # Create ordered list based on round-robin
+  local ordered_models=()
+  for ((i=0; i<num_models; i++)); do
+    local idx=$(((start_index + i) % num_models))
+    ordered_models+=("${models[$idx]}")
+  done
+  
+  echo "[AI] Model order for ideate (round-robin): ${ordered_models[*]}" >&2
+  
+  # Try each model until CSV changes
+  for model in "${ordered_models[@]}"; do
+    echo "[AI] Attempting ideate with $model" >&2
+    
+    # Call the model directly
+    local ai_output
+    ai_output=$(call_ai_model_configured "$model" "$prompt")
+    local ai_exit_code=$?
+    
+    echo "[AI] $model completed with exit code $ai_exit_code" >&2
+    
+    # Check if the file was modified - this is ALL that matters
     if [[ -f "$temp_csv_file" ]]; then
       local new_csv_count
       new_csv_count=$(grep -v '^[[:space:]]*$' "$temp_csv_file" | tail -n +2 | wc -l)
       
       if [[ $new_csv_count -gt $original_csv_count ]]; then
-        echo "[INFO] AI completed successfully and modified CSV ($new_csv_count vs $original_csv_count rows)" >&2
+        echo "[INFO] CSV was modified by $model ($new_csv_count vs $original_csv_count rows) - validating format..." >&2
+        
+        # Post-process to ensure all description fields are quoted
+        local fixed_csv_file="${temp_csv_file}.fixed"
+        
+        # Use the CSV fixer script
+        if "$PYTHON_CMD" "$SCRIPT_DIR/../lib/csv_fixer.py" "$temp_csv_file" "$fixed_csv_file"; then
+          mv "$fixed_csv_file" "$temp_csv_file"
+          echo "[INFO] CSV format validated and fixed if needed" >&2
+        else
+          echo "[WARN] CSV format validation failed, using original" >&2
+        fi
+        
         return 0
       else
-        echo "[INFO] $model returned exit code 0 but didn't modify CSV file" >&2
-        echo "[DEBUG] Expected file: $temp_csv_file" >&2
+        echo "[INFO] CSV unchanged after $model (exit code: $ai_exit_code)" >&2
         echo "[DEBUG] Original count: $original_csv_count, New count: $new_csv_count" >&2
-        return 1
+        # Continue to next model
       fi
     else
-      echo "[INFO] Exit code 0 but temp CSV file not found: $temp_csv_file" >&2
-      echo "[DEBUG] Current directory: $(pwd)" >&2
-      echo "[DEBUG] Files matching temp-csv-*.csv:" >&2
-      ls -la temp-csv-*.csv 2>&1 >&2
-      return 1
+      echo "[INFO] Temp CSV file not found after $model: $temp_csv_file" >&2
+      # Continue to next model
     fi
-  fi
+  done
   
-  echo "[INFO] No AI model successfully modified the CSV file" >&2
+  # All models tried, none changed the file
+  echo "[ERROR] All AI models failed to generate ideas" >&2
   return 1
 }
 
@@ -461,8 +498,14 @@ Instructions:
 1. Add exactly $count new rows to the CSV
 2. Use the next available generation numbers (gen$CURRENT_GENERATION-XXX format)
 3. For each idea, create a row with: id,parent_id,description,,pending
-4. For novel ideas: leave parent_id empty
-5. For other idea types: use appropriate parent IDs from these top performers:
+4. CRITICAL CSV FORMATTING RULES:
+   - ALWAYS wrap the description field in double quotes
+   - If the description contains quotes, escape them by doubling them (\" becomes \"\")
+   - Example: gen01-001,gen00-000,\"Implement adaptive RSI thresholds\",,pending
+   - BAD: gen01-001,gen00-000,Implement adaptive RSI thresholds,,pending
+   - NEVER omit quotes - unquoted descriptions cause CSV corruption
+5. For novel ideas: leave parent_id empty
+6. For other idea types: use appropriate parent IDs from these top performers:
 $top_performers
 
 IMPORTANT: Output the complete modified CSV file. Do not add any explanation or other text - just output the CSV."
@@ -839,9 +882,15 @@ CRITICAL INSTRUCTIONS:
   prompt+="
 6. Use the Edit or MultiEdit tool to APPEND exactly $count new rows AT THE END of the CSV file
 7. For each idea, create a row with: id,,description,,pending (empty parent_id for novel ideas)
-8. Each description should be one clear sentence describing a novel algorithmic approach
-9. Focus on creative, ambitious ideas that haven't been tried yet
-10. Consider machine learning, new indicators, regime detection, risk management, etc.
+8. CRITICAL CSV FORMATTING RULES:
+   - ALWAYS wrap the description field in double quotes
+   - If the description contains quotes, escape them by doubling them (\" becomes \"\")
+   - Example: gen01-001,,\"Implement adaptive RSI thresholds based on volatility\",,pending
+   - BAD: gen01-001,,Implement adaptive RSI thresholds based on volatility,,pending
+   - NEVER omit quotes around descriptions - this causes CSV parsing errors
+9. Each description should be one clear sentence describing a novel algorithmic approach
+10. Focus on creative, ambitious ideas that haven't been tried yet
+11. Consider machine learning, new indicators, regime detection, risk management, etc.
 
 IMPORTANT: You must APPEND new rows to the existing CSV file. DO NOT replace the file contents. All existing rows must remain unchanged.
 CRITICAL: You must use your file editing tools (Edit/MultiEdit) to modify the CSV file. DO NOT return CSV text - use your tools to edit the file directly.
@@ -939,8 +988,14 @@ CRITICAL INSTRUCTIONS:
 5. Use the Edit or MultiEdit tool to APPEND exactly $count new rows AT THE END of the CSV file
 6. For each idea, create a row with: id,parent_id,description,,pending
 7. Each parent_id MUST be one of: $valid_parent_ids
-8. Each description should focus on adjusting specific parameters that exist in the parent's source code
-9. Include current and new parameter values (e.g., \"Lower rsi_entry from 21 to 18\")
+8. CRITICAL CSV FORMATTING RULES:
+   - ALWAYS wrap the description field in double quotes
+   - If the description contains quotes, escape them by doubling them (\" becomes \"\")
+   - Example: gen01-001,gen00-000,\"Lower rsi_entry from 21 to 18\",,pending
+   - BAD: gen01-001,gen00-000,Lower rsi_entry from 21 to 18,,pending
+   - NEVER omit quotes around descriptions - this causes CSV parsing errors
+9. Each description should focus on adjusting specific parameters that exist in the parent's source code
+10. Include current and new parameter values - for example: \"Lower rsi_entry from 21 to 18\"
 
 IMPORTANT: You must APPEND new rows to the existing CSV file. DO NOT replace the file contents. All existing rows must remain unchanged.
 CRITICAL: You must use your file editing tools (Edit/MultiEdit) to modify the CSV file. DO NOT return CSV text - use your tools to edit the file directly.
@@ -1020,8 +1075,14 @@ CRITICAL INSTRUCTIONS:
 5. Use the Edit or MultiEdit tool to APPEND exactly $count new rows AT THE END of the CSV file
 6. For each idea, create a row with: id,parent_id,description,,pending
 7. Each parent_id MUST be one of: $valid_parent_ids
-8. Each description should focus on architectural/structural changes based on the parent's actual code
-9. Reference actual components/methods found in the source code
+8. CRITICAL CSV FORMATTING RULES:
+   - ALWAYS wrap the description field in double quotes
+   - If the description contains quotes, escape them by doubling them (\" becomes \"\")
+   - Example: gen01-001,gen00-000,\"Add ML-based regime detection using LSTM\",,pending
+   - BAD: gen01-001,gen00-000,Add ML-based regime detection using LSTM,,pending
+   - NEVER omit quotes around descriptions - this causes CSV parsing errors
+9. Each description should focus on architectural/structural changes based on the parent's actual code
+10. Reference actual components/methods found in the source code
 
 IMPORTANT: You must APPEND new rows to the existing CSV file. DO NOT replace the file contents. All existing rows must remain unchanged.
 CRITICAL: You must use your file editing tools (Edit/MultiEdit) to modify the CSV file. DO NOT return CSV text - use your tools to edit the file directly.
@@ -1101,8 +1162,14 @@ CRITICAL INSTRUCTIONS:
 5. Use the Edit or MultiEdit tool to APPEND exactly $count new rows AT THE END of the CSV file
 6. For each idea, create a row with: id,parent_id,description,,pending
 7. Each parent_id MUST be one of: $valid_parent_ids (choose the primary parent)
-8. Each description should combine actual elements from 2+ algorithms based on their source code
-9. Reference specific components/features found in the actual source code
+8. CRITICAL CSV FORMATTING RULES:
+   - ALWAYS wrap the description field in double quotes
+   - If the description contains quotes, escape them by doubling them (\" becomes \"\")
+   - Example: gen01-001,gen00-000,\"Combine gen01-123's RSI logic with gen01-456's volatility scaling\",,pending
+   - BAD: gen01-001,gen00-000,Combine gen01-123's RSI logic with gen01-456's volatility scaling,,pending
+   - NEVER omit quotes around descriptions - this causes CSV parsing errors
+9. Each description should combine actual elements from 2+ algorithms based on their source code
+10. Reference specific components/features found in the actual source code
 
 IMPORTANT: You must APPEND new rows to the existing CSV file. DO NOT replace the file contents. All existing rows must remain unchanged.
 CRITICAL: You must use your file editing tools (Edit/MultiEdit) to modify the CSV file. DO NOT return CSV text - use your tools to edit the file directly.
@@ -1195,8 +1262,14 @@ CRITICAL INSTRUCTIONS:
 4. If no gen$CURRENT_GENERATION entries exist yet, start with gen$CURRENT_GENERATION-001
 5. Use the Edit or MultiEdit tool to APPEND exactly $TOTAL_IDEAS new rows AT THE END of the CSV file
 6. For each idea, create a row with: id,parent_id,description,,pending
-7. Mix both parameter tuning and structural changes
-8. If building on existing algorithms, use their ID as parent_id, otherwise leave parent_id empty
+7. CRITICAL CSV FORMATTING RULES:
+   - ALWAYS wrap the description field in double quotes
+   - If the description contains quotes, escape them by doubling them (\" becomes \"\")
+   - Example: gen01-001,gen00-000,\"Implement adaptive RSI thresholds based on volatility\",,pending
+   - BAD: gen01-001,gen00-000,Implement adaptive RSI thresholds based on volatility,,pending
+   - NEVER omit quotes around descriptions - this causes CSV parsing errors that corrupt the data
+8. Mix both parameter tuning and structural changes
+9. If building on existing algorithms, use their ID as parent_id, otherwise leave parent_id empty
 
 ⚠️ AVOID ONLY: Kelly floor/cap adjustments that assume leverage > 1.0 (these get clamped and have no effect)
 

package/bin/claude-evolve-run

@@ -478,6 +478,66 @@ ensure_baseline_entry
 # Flag to track API limit status
 api_limit_reached=false
 
+# Check if previous generation has at least one completed item
+check_previous_generation_has_completed() {
+  local csv_path="$1"
+  
+  if [[ ! -f "$csv_path" ]]; then
+    # No CSV file yet - allow first ideation
+    return 0
+  fi
+  
+  # Use Python to find the maximum generation and check if it has completed items
+  "$PYTHON_CMD" -c "
+import csv
+import sys
+
+max_gen = 0
+gen_completed = {}
+
+try:
+    with open('$csv_path', 'r') as f:
+        reader = csv.reader(f)
+        next(reader, None)  # Skip header
+        for row in reader:
+            if row and len(row) >= 5:
+                id_field = row[0].strip()
+                status = row[4].strip() if len(row) > 4 else 'pending'
+                
+                if id_field.startswith('gen') and '-' in id_field:
+                    try:
+                        gen_part = id_field.split('-')[0]  # e.g., 'gen01'
+                        gen_num = int(gen_part[3:])  # Extract number after 'gen'
+                        max_gen = max(max_gen, gen_num)
+                        
+                        if gen_num not in gen_completed:
+                            gen_completed[gen_num] = 0
+                        
+                        if status == 'complete':
+                            gen_completed[gen_num] += 1
+                    except (ValueError, IndexError):
+                        pass
+
+    # If max_gen is 0, no generations exist yet - allow ideation
+    if max_gen == 0:
+        sys.exit(0)
+    
+    # Check if the most recent generation has at least one completed item
+    if gen_completed.get(max_gen, 0) > 0:
+        print(f'[INFO] Generation {max_gen:02d} has {gen_completed[max_gen]} completed items - allowing ideation', file=sys.stderr)
+        sys.exit(0)
+    else:
+        print(f'[INFO] Generation {max_gen:02d} has no completed items - blocking ideation to prevent endless loops', file=sys.stderr)
+        sys.exit(1)
+
+except Exception as e:
+    print(f'[ERROR] Failed to check previous generation: {e}', file=sys.stderr)
+    # On error, allow ideation to maintain backward compatibility
+    sys.exit(0)
+"
+  return $?
+}
+
 # Main dispatch loop
 while true; do
   # Clean up finished workers
@@ -506,7 +566,17 @@ while true; do
     
     # Check if auto ideation is enabled
     if [[ "$AUTO_IDEATE" == "true" || "$AUTO_IDEATE" == "1" ]]; then
-      echo "[DISPATCHER] Auto ideation is enabled. Generating new ideas..."
+      echo "[DISPATCHER] Auto ideation is enabled. Checking prerequisites..."
+      
+      # Check if previous generation has at least one completed item
+      if ! check_previous_generation_has_completed "$FULL_CSV_PATH"; then
+        echo "[DISPATCHER] Evolution complete - previous generation has no completed items."
+        echo "[DISPATCHER] This prevents endless ideation loops when API limits are hit."
+        echo "[DISPATCHER] Wait for current generation to complete, then run 'claude-evolve ideate' manually."
+        break
+      fi
+      
+      echo "[DISPATCHER] Prerequisites met. Generating new ideas..."
       
       # Check if claude-evolve-ideate exists
       ideate_script="$SCRIPT_DIR/claude-evolve-ideate"

package/bin/claude-evolve-status

@@ -244,7 +244,9 @@ try:
     # Show per-generation breakdown (unless brief mode)
     if not show_brief and stats_by_gen:
         print('📈 BY GENERATION:')
-        for gen in sorted(stats_by_gen.keys()):
+        # Sort generations numerically by extracting the number after 'gen'
+        sorted_gens = sorted(stats_by_gen.keys(), key=lambda g: int(g[3:]) if g.startswith('gen') and g[3:].isdigit() else 0)
+        for gen in sorted_gens:
             data = stats_by_gen[gen]
             total = sum(data.values())
             

package/bin/claude-evolve-worker

@@ -14,34 +14,11 @@ TERMINATION_SIGNAL=""
 # Cleanup function to handle termination
 cleanup_on_exit() {
   if [[ -n "$CURRENT_CANDIDATE_ID" ]]; then
-    # Only mark as failed if it was a timeout (SIGTERM from timeout command)
-    # For user interruption (Ctrl-C) or kill, leave it for retry
-    if [[ "$TERMINATION_SIGNAL" == "TERM" ]]; then
-      echo "[WORKER-$$] Timeout detected, marking $CURRENT_CANDIDATE_ID as failed" >&2
-      "$PYTHON_CMD" -c "
-import sys
-sys.path.insert(0, '$SCRIPT_DIR/..')
-from lib.evolution_csv import EvolutionCSV
-try:
-    with EvolutionCSV('$FULL_CSV_PATH') as csv:
-        csv.update_candidate_status('$CURRENT_CANDIDATE_ID', 'failed')
-except:
-    pass  # Best effort cleanup
-" 2>/dev/null || true
-    else
-      echo "[WORKER-$$] Interrupted, leaving $CURRENT_CANDIDATE_ID for retry" >&2
-      # Optionally reset to pending instead of leaving as running
-      "$PYTHON_CMD" -c "
-import sys
-sys.path.insert(0, '$SCRIPT_DIR/..')
-from lib.evolution_csv import EvolutionCSV
-try:
-    with EvolutionCSV('$FULL_CSV_PATH') as csv:
-        csv.update_candidate_status('$CURRENT_CANDIDATE_ID', 'pending')
-except:
-    pass  # Best effort cleanup
-" 2>/dev/null || true
-    fi
+    echo "[WORKER-$$] Worker terminated while processing $CURRENT_CANDIDATE_ID" >&2
+    # If we're interrupted while processing, leave it as "running"
+    # This prevents other workers from picking it up in the same session
+    # A human can manually reset to pending if needed
+    echo "[WORKER-$$] Leaving $CURRENT_CANDIDATE_ID in current state" >&2
   fi
 }
 
@@ -135,6 +112,12 @@ process_candidate() {
   echo "[WORKER-$$] Description: $description"
   echo "[WORKER-$$] Based on ID: $parent_id"
   
+  # Treat "baseline-000" parent ID as empty/baseline
+  if [[ "$parent_id" == "baseline-000" ]]; then
+    parent_id=""
+    echo "[WORKER-$$] Parent ID 'baseline-000' treated as baseline (empty parent)"
+  fi
+  
   # Determine source algorithm
   local source_file
   if [[ -z "$parent_id" ]]; then
@@ -213,16 +196,26 @@ CRITICAL: Do NOT use any git commands (git add, git commit, git reset, etc.). On
       
       # Try AI models with round-robin based on candidate ID
       if ! call_ai_for_evolution "$evolution_prompt" "$candidate_id"; then
-        echo "[WORKER-$$] ERROR: All AI models failed to generate code" >&2
+        echo "[WORKER-$$] ERROR: All AI models failed to generate code - leaving as pending for retry" >&2
         cd "$original_pwd"
         rm -f "$target_file"  # Clean up on failure
-        return 1
+        # Return with special code to indicate AI failure (should remain pending)
+        return 77
       fi
       
       # Restore working directory
       cd "$original_pwd"
       
       echo "[WORKER-$$] Evolution applied successfully"
+      
+      # Check if the generated Python file has syntax errors
+      echo "[WORKER-$$] Checking Python syntax..." >&2
+      if ! "$PYTHON_CMD" -m py_compile "$target_file" 2>&1; then
+        echo "[WORKER-$$] ERROR: Generated Python file has syntax errors!" >&2
+        echo "[WORKER-$$] File: $target_file" >&2
+        # This is still an evaluation failure, not an AI failure
+        return 1
+      fi
     fi
   fi
   
@@ -239,11 +232,22 @@ CRITICAL: Do NOT use any git commands (git add, git commit, git reset, etc.). On
     eval_arg=""
   fi
   local eval_cmd=("$PYTHON_CMD" "$FULL_EVALUATOR_PATH" "$eval_arg")
+  
+  # Add memory limiting if configured
+  if [[ -n "$MEMORY_LIMIT_MB" ]] && [[ "$MEMORY_LIMIT_MB" -gt 0 ]]; then
+    eval_cmd=("$PYTHON_CMD" "$SCRIPT_DIR/../lib/memory_limit_wrapper.py" "$MEMORY_LIMIT_MB" "${eval_cmd[@]}")
+  fi
+  
+  # Add timeout if configured
   [[ -n "$timeout_seconds" ]] && eval_cmd=(timeout "$timeout_seconds" "${eval_cmd[@]}")
   
   # Run evaluation with tee to both display and capture output
   # Use stdbuf to disable buffering for real-time output
-  if stdbuf -o0 -e0 "${eval_cmd[@]}" 2>&1 | tee "$eval_output_file" >&2; then
+  # IMPORTANT: Use PIPESTATUS to get the exit code of the evaluation command, not tee
+  stdbuf -o0 -e0 "${eval_cmd[@]}" 2>&1 | tee "$eval_output_file" >&2
+  local eval_exit_code=${PIPESTATUS[0]}  # Get exit code of first command in pipe
+  
+  if [[ $eval_exit_code -eq 0 ]]; then
     local eval_end=$(date +%s)
     local eval_duration=$((eval_end - eval_start))
     
@@ -353,14 +357,19 @@ with EvolutionCSV('$FULL_CSV_PATH') as csv:
       echo "[WORKER-$$] Output: $eval_output" >&2
       # rm -f "$eval_output_file"  # Keep for debugging
       echo "[WORKER-$$] Evaluation output saved to: $eval_output_file" >&2
+      # Clear CURRENT_CANDIDATE_ID before returning
+      CURRENT_CANDIDATE_ID=""
       return 1
     fi
     
     # Clean up temp file (comment out to keep for debugging)
     # rm -f "$eval_output_file"
     echo "[WORKER-$$] Evaluation output saved to: $eval_output_file" >&2
+    
+    # Clear CURRENT_CANDIDATE_ID on successful completion
+    CURRENT_CANDIDATE_ID=""
   else
-    local exit_code=$?
+    local exit_code=$eval_exit_code
     # Read any output that was captured before failure
     eval_output=$(<"$eval_output_file")
     # rm -f "$eval_output_file"  # Keep for debugging
@@ -370,22 +379,56 @@ with EvolutionCSV('$FULL_CSV_PATH') as csv:
     echo "[WORKER-$$] Output: $eval_output" >&2
     
     # Mark as failed in CSV
-    "$PYTHON_CMD" -c "
+    echo "[WORKER-$$] Marking $candidate_id as failed in CSV" >&2
+    if ! "$PYTHON_CMD" -c "
 import sys
 sys.path.insert(0, '$SCRIPT_DIR/..')
 from lib.evolution_csv import EvolutionCSV
-with EvolutionCSV('$FULL_CSV_PATH') as csv:
-    csv.update_candidate_status('$candidate_id', 'failed')
-"
+try:
+    with EvolutionCSV('$FULL_CSV_PATH') as csv:
+        success = csv.update_candidate_status('$candidate_id', 'failed')
+        if not success:
+            print(f'ERROR: Failed to update status for {candidate_id}', file=sys.stderr)
+            sys.exit(1)
+except Exception as e:
+    print(f'ERROR: Exception updating status: {e}', file=sys.stderr)
+    sys.exit(1)
+" 2>&1; then
+      echo "[WORKER-$$] ERROR: Failed to update CSV status to failed" >&2
+    else
+      echo "[WORKER-$$] Successfully marked $candidate_id as failed" >&2
+    fi
     
+    # Clear CURRENT_CANDIDATE_ID before returning to prevent cleanup handler from resetting it
+    CURRENT_CANDIDATE_ID=""
     return $exit_code
   fi
 }
 
+# Don't reset running candidates on startup - they might be legitimately being processed by another worker
+
 # Main worker loop
 echo "[WORKER-$$] Worker started"
 
 while true; do
+  # Debug: Show current status of all candidates
+  echo "[WORKER-$$] Current candidate statuses:" >&2
+  "$PYTHON_CMD" -c "
+import sys
+sys.path.insert(0, '$SCRIPT_DIR/..')
+from lib.evolution_csv import EvolutionCSV
+with EvolutionCSV('$FULL_CSV_PATH') as csv:
+    rows = csv._read_csv()
+    if rows:
+        start_idx = 1 if rows and rows[0] and rows[0][0].lower() == 'id' else 0
+        status_count = {}
+        for row in rows[start_idx:]:
+            if len(row) > 4:
+                status = row[4].strip() or 'pending'
+                status_count[status] = status_count.get(status, 0) + 1
+        print(f'Status counts: {status_count}', file=sys.stderr)
+" 2>&1 || true
+
   # Try to claim a pending candidate
   candidate_info=$("$PYTHON_CMD" -c "
 import sys
@@ -412,12 +455,26 @@ with EvolutionCSV('$FULL_CSV_PATH') as csv:
   # Set current candidate for cleanup
   CURRENT_CANDIDATE_ID="$candidate_id"
   
-  # Process the candidate
-  if process_candidate "$candidate_id" "$parent_id" "$description"; then
+  # Process the candidate and capture exit code
+  process_candidate "$candidate_id" "$parent_id" "$description"
+  process_exit_code=$?
+  
+  if [[ $process_exit_code -eq 0 ]]; then
     echo "[WORKER-$$] Successfully processed $candidate_id"
+  elif [[ $process_exit_code -eq 77 ]]; then
+    # Special exit code 77 means AI failed to generate code
+    echo "[WORKER-$$] AI generation failed for $candidate_id - marking as failed-ai-retry"
+    # Mark with special status that indicates AI generation failed (not evaluation)
+    "$PYTHON_CMD" -c "
+import sys
+sys.path.insert(0, '$SCRIPT_DIR/..')
+from lib.evolution_csv import EvolutionCSV
+with EvolutionCSV('$FULL_CSV_PATH') as csv:
+    csv.update_candidate_status('$candidate_id', 'failed-ai-retry')
+" 2>/dev/null || true
   else
     echo "[WORKER-$$] Failed to process $candidate_id"
-    # Ensure status is set to failed (might already be done in process_candidate)
+    # Other failures (evaluation errors, etc) mark as failed
     "$PYTHON_CMD" -c "
 import sys
 sys.path.insert(0, '$SCRIPT_DIR/..')

package/lib/ai-cli.sh

@@ -19,6 +19,11 @@ call_ai_model_configured() {
       ai_output=$(timeout 300 claude --dangerously-skip-permissions --model "$model_name" -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
+    gpt-5)
+      local ai_output
+      ai_output=$(timeout 300 codex exec -m gpt-5 --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      local ai_exit_code=$?
+      ;;
     o3)
       local ai_output
       ai_output=$(timeout 300 codex exec -m o3 --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
@@ -100,7 +105,7 @@ clean_ai_output() {
   local model_name="$2"
   
   # Handle codex-specific output format
-  if [[ "$model_name" == "codex" || "$model_name" == "o3" ]]; then
+  if [[ "$model_name" == "codex" || "$model_name" == "o3" || "$model_name" == "gpt-5" ]]; then
     # Clean codex output - extract content between "codex" marker and "tokens used"
     if echo "$output" | grep -q "^\[.*\] codex$"; then
       # Extract content between "codex" line and "tokens used" line
@@ -191,7 +196,7 @@ call_ai_with_round_robin() {
     ai_output=$(call_ai_model_configured "$model" "$prompt")
     local ai_exit_code=$?
     
-    # Just check exit code - no interpretation
+    # Just check exit code
     if [[ $ai_exit_code -eq 0 ]]; then
       # Clean output if needed
       ai_output=$(clean_ai_output "$ai_output" "$model")

package/lib/config.sh

@@ -49,17 +49,13 @@ DEFAULT_AUTO_IDEATE=true
 # Default retry value
 DEFAULT_MAX_RETRIES=3
 
-# Default LLM CLI configuration (using eval for compatibility)
-declare -a DEFAULT_LLM_CLI_KEYS
-declare -a DEFAULT_LLM_CLI_VALUES
-DEFAULT_LLM_CLI_KEYS=(o3 codex gemini opus sonnet)
-DEFAULT_LLM_CLI_VALUES[0]='codex exec -m o3 --dangerously-bypass-approvals-and-sandbox "{{PROMPT}}"'
-DEFAULT_LLM_CLI_VALUES[1]='codex exec --dangerously-bypass-approvals-and-sandbox "{{PROMPT}}"'
-DEFAULT_LLM_CLI_VALUES[2]='gemini -y -p "{{PROMPT}}"'
-DEFAULT_LLM_CLI_VALUES[3]='claude --dangerously-skip-permissions --model opus -p "{{PROMPT}}"'
-DEFAULT_LLM_CLI_VALUES[4]='claude --dangerously-skip-permissions --model sonnet -p "{{PROMPT}}"'
-DEFAULT_LLM_RUN="sonnet gemini"
-DEFAULT_LLM_IDEATE="opus o3"
+# Default memory limit (in MB, 0 means no limit)
+# Set to reasonable limit for ML workloads - about half of available system RAM
+DEFAULT_MEMORY_LIMIT_MB=12288
+
+# Default LLM CLI configuration - use simple variables instead of arrays
+DEFAULT_LLM_RUN="sonnet gpt-5 sonnet gpt-5"
+DEFAULT_LLM_IDEATE="gemini gpt-5 opus"
 
 # Load configuration from config file
 load_config() {
@@ -96,9 +92,13 @@ load_config() {
   # Set retry default
   MAX_RETRIES="$DEFAULT_MAX_RETRIES"
   
+  # Set memory limit default
+  MEMORY_LIMIT_MB="$DEFAULT_MEMORY_LIMIT_MB"
+  
   # Set LLM CLI defaults (compatibility for older bash)
   # Initialize associative array for LLM commands
   # Use simpler approach for compatibility
+  LLM_CLI_gpt_5='codex exec -m gpt-5 --dangerously-bypass-approvals-and-sandbox "{{PROMPT}}"'
   LLM_CLI_o3='codex exec -m o3 --dangerously-bypass-approvals-and-sandbox "{{PROMPT}}"'
   LLM_CLI_codex='codex exec --dangerously-bypass-approvals-and-sandbox "{{PROMPT}}"'
   LLM_CLI_gemini='gemini -y -p "{{PROMPT}}"'
@@ -202,12 +202,14 @@ load_config() {
           # Model definition - key is model name, value is command template
           # Remove single quotes from value if present
           value=$(echo "$value" | sed "s/^'//;s/'$//")
+          # Convert dashes to underscores for bash variable names
+          var_key=$(echo "$key" | sed 's/-/_/g')
           # Debug config loading
           if [[ "${DEBUG_CONFIG:-}" == "true" ]]; then
-            echo "[CONFIG DEBUG] Setting LLM_CLI_${key} = '$value'" >&2
+            echo "[CONFIG DEBUG] Setting LLM_CLI_${var_key} = '$value'" >&2
           fi
           # Use dynamic variable name for compatibility
-          eval "LLM_CLI_${key}=\"$value\""
+          eval "LLM_CLI_${var_key}=\"$value\""
         fi
       else
         # Handle top-level keys
@@ -221,6 +223,7 @@ load_config() {
           python_cmd) PYTHON_CMD="$value" ;;
           auto_ideate) AUTO_IDEATE="$value" ;;
           max_retries) MAX_RETRIES="$value" ;;
+          memory_limit_mb) MEMORY_LIMIT_MB="$value" ;;
           evolution_dir) 
             echo "[WARN] evolution_dir in config is ignored - automatically inferred from config file location" >&2
             ;;
@@ -316,14 +319,17 @@ show_config() {
   echo "  Lock timeout: $LOCK_TIMEOUT"
   echo "  Auto ideate: $AUTO_IDEATE"
   echo "  Max retries: $MAX_RETRIES"
+  echo "  Memory limit: ${MEMORY_LIMIT_MB}MB"
   echo "  LLM configuration:"
   # Show LLM configurations using dynamic variable names
-  for model in o3 codex gemini opus sonnet; do
+  for model in gpt_5 o3 codex gemini opus sonnet; do
     var_name="LLM_CLI_${model}"
     if [[ -n "${!var_name}" ]]; then
-      echo "    $model: ${!var_name}"
+      # Convert underscore back to dash for display
+      display_name=$(echo "$model" | sed 's/_/-/g')
+      echo "    $display_name: ${!var_name}"
     fi
   done
   echo "  LLM for run: $LLM_RUN"
   echo "  LLM for ideate: $LLM_IDEATE"
-}
+}