claude-evolve 1.8.12 → 1.8.14

package/bin/claude-evolve-ideate

@@ -1641,7 +1641,7 @@ echo "[INFO] Starting ideation for generation $CURRENT_GENERATION"
 # Main execution with retry logic and exponential backoff
 retry_count=0
 wait_seconds=300  # Start with 5 minutes
-max_wait_seconds=1800  # Cap at 30 minutes
+max_wait_seconds=300  # Cap at 5 minutes
 
 while true; do
   if [[ $use_strategies == true ]]; then

package/bin/claude-evolve-run

@@ -275,29 +275,34 @@ count_pending_candidates() {
 # Function to get CSV stats
 get_csv_stats() {
   local csv_path="${1:-$FULL_CSV_PATH}"
-  
+
   if [[ ! -f "$csv_path" ]]; then
     echo "[ERROR] CSV not found at: $csv_path" >&2
     echo "0 0 0"
     return
   fi
-  
+
   local total_rows complete_count pending_count
   total_rows=$(wc -l < "$csv_path" | tr -d '[:space:]')
   complete_count=$(grep ',complete' "$csv_path" 2>/dev/null | wc -l | tr -d '[:space:]')
-  
+
   # Count pending using UNIFIED CSV logic
   pending_count=$("$PYTHON_CMD" "$SCRIPT_DIR/../lib/evolution_csv.py" "$csv_path" count)
-  
+
   echo "$total_rows $complete_count $pending_count"
 }
 
-echo "[DISPATCHER] Starting unified evolution engine"
-echo "[DISPATCHER] Configuration: max_workers=$MAX_WORKERS, timeout=${timeout_seconds:-none}"
+# Function to perform full CSV cleanup (duplicates, stuck statuses, missing files, etc.)
+cleanup_csv_full() {
+  if [[ ! -f "$FULL_CSV_PATH" ]]; then
+    echo "[DISPATCHER] No CSV file to clean up" >&2
+    return 0
+  fi
+
+  echo "[DISPATCHER] Performing full CSV cleanup..." >&2
 
-# Clean up any stuck 'running' statuses and duplicates at startup
-if [[ -f "$FULL_CSV_PATH" ]]; then
-  echo "[DISPATCHER] Checking for duplicate candidates..."
+  # Remove duplicate candidates
+  echo "[DISPATCHER] Checking for duplicate candidates..." >&2
   "$PYTHON_CMD" -c "
 import sys
 sys.path.insert(0, '$SCRIPT_DIR/..')
@@ -312,53 +317,49 @@ except Exception as e:
     print(f'[ERROR] Failed to remove duplicates: {e}', file=sys.stderr)
 " 2>&1 || true
 
-  echo "[DISPATCHER] Resetting any stuck 'running' candidates to 'pending'..."
+  # Reset stuck 'running' candidates
+  echo "[DISPATCHER] Resetting any stuck 'running' candidates to 'pending'..." >&2
   if "$SCRIPT_DIR/claude-evolve-edit" running pending >/dev/null 2>&1; then
-    echo "[DISPATCHER] Successfully reset stuck candidates"
+    echo "[DISPATCHER] Successfully reset stuck candidates" >&2
   else
-    echo "[DISPATCHER] No stuck candidates found or edit command not available"
+    echo "[DISPATCHER] No stuck candidates found or edit command not available" >&2
   fi
-fi
 
-# Validate CSV and clean up stuck statuses and duplicates
-if [[ -f "$FULL_CSV_PATH" ]]; then
-  echo "[DISPATCHER] Validating CSV and cleaning up..."
-  
-  # First check for and clean up duplicates
-  echo "[DISPATCHER] Checking for duplicate entries..."
+  # Clean up duplicates with the dedicated cleanup script
+  echo "[DISPATCHER] Checking for duplicate entries..." >&2
   duplicate_check_output=$("$PYTHON_CMD" "$SCRIPT_DIR/claude-evolve-cleanup-duplicates" "$FULL_CSV_PATH" 2>&1)
   if echo "$duplicate_check_output" | grep -q "Found.*duplicate"; then
-    echo "[DISPATCHER] WARNING: Duplicate entries detected in CSV!"
-    echo "$duplicate_check_output"
-    echo "[DISPATCHER] Automatically cleaning up duplicates..."
+    echo "[DISPATCHER] WARNING: Duplicate entries detected in CSV!" >&2
+    echo "$duplicate_check_output" >&2
+    echo "[DISPATCHER] Automatically cleaning up duplicates..." >&2
     if "$PYTHON_CMD" "$SCRIPT_DIR/claude-evolve-cleanup-duplicates" "$FULL_CSV_PATH" --fix; then
-      echo "[DISPATCHER] Duplicates cleaned up successfully"
+      echo "[DISPATCHER] Duplicates cleaned up successfully" >&2
     else
       echo "[ERROR] Failed to clean up duplicates" >&2
-      exit 1
+      return 1
     fi
   else
-    echo "[DISPATCHER] No duplicates found"
+    echo "[DISPATCHER] No duplicates found" >&2
   fi
-  
-  # Check for and clean up invalid entries
-  echo "[DISPATCHER] Checking for invalid entries..."
+
+  # Clean up invalid entries
+  echo "[DISPATCHER] Checking for invalid entries..." >&2
   invalid_check_output=$("$PYTHON_CMD" "$SCRIPT_DIR/claude-evolve-clean-invalid" "$FULL_CSV_PATH" --dry-run 2>&1)
   if echo "$invalid_check_output" | grep -q "Found.*invalid"; then
-    echo "[DISPATCHER] WARNING: Invalid entries detected in CSV!"
-    echo "$invalid_check_output"
-    echo "[DISPATCHER] Automatically cleaning up invalid entries..."
+    echo "[DISPATCHER] WARNING: Invalid entries detected in CSV!" >&2
+    echo "$invalid_check_output" >&2
+    echo "[DISPATCHER] Automatically cleaning up invalid entries..." >&2
     if "$PYTHON_CMD" "$SCRIPT_DIR/claude-evolve-clean-invalid" "$FULL_CSV_PATH"; then
-      echo "[DISPATCHER] Invalid entries cleaned up successfully"
+      echo "[DISPATCHER] Invalid entries cleaned up successfully" >&2
     else
       echo "[ERROR] Failed to clean up invalid entries" >&2
-      exit 1
+      return 1
     fi
   else
-    echo "[DISPATCHER] No invalid entries found"
+    echo "[DISPATCHER] No invalid entries found" >&2
   fi
-  
-  # Then validate and clean stuck statuses
+
+  # Clean stuck statuses and missing files
   if ! "$PYTHON_CMD" -c "
 import csv
 import sys
@@ -366,90 +367,88 @@ import os
 from pathlib import Path
 
 csv_file = '$FULL_CSV_PATH'
-full_output_dir = '$FULL_OUTPUT_DIR' # Pass FULL_OUTPUT_DIR to Python script
-script_dir = '$SCRIPT_DIR' # Pass SCRIPT_DIR for sys.path.append
+full_output_dir = '$FULL_OUTPUT_DIR'
+script_dir = '$SCRIPT_DIR'
 
 try:
-    # Read CSV - let Python's csv module handle all the complexity
     with open(csv_file, 'r') as f:
         reader = csv.reader(f)
         rows = list(reader)
-    
+
     if not rows:
         print('[ERROR] CSV is empty')
         sys.exit(1)
-    
-    # Basic sanity checks
-    header = rows[0]
-    num_fields = len(header)
-    
+
     if len(rows) == 1:
         print('[INFO] CSV has no data rows (only header)')
-    
+
     changed_count = 0
-    
+
     # Clean up any stuck 'running' statuses
     for i in range(1, len(rows)):
         if len(rows[i]) > 4 and rows[i][4] == 'running':
             rows[i][4] = ''
             changed_count += 1
 
-    # Reset failed-parent-missing to pending at startup - give them another chance
+    # Reset failed-parent-missing to pending - give them another chance
     for i in range(1, len(rows)):
         if len(rows[i]) > 4 and rows[i][4] == 'failed-parent-missing':
             rows[i][4] = 'pending'
             changed_count += 1
-    
+
     # Check for missing Python files for completed/failed candidates
     for i in range(1, len(rows)):
         if len(rows[i]) > 4:
             candidate_id = rows[i][0]
             status = rows[i][4]
-            
-            # Only check if status implies a file should exist
+
             if status in ['complete', 'failed', 'failed-ai-retry', 'failed-retry1', 'failed-retry2', 'failed-retry3']:
                 expected_file = Path(full_output_dir) / f'evolution_{candidate_id}.py'
                 if not expected_file.is_file():
                     print(f'[INFO] Detected missing file for {candidate_id} (status: {status}). Resetting to pending.')
-                    rows[i][4] = 'pending' # Reset status to pending
-                    # Clear performance and other fields if desired, for a clean retry
-                    if len(rows[i]) > 3: rows[i][3] = '' # Performance
-                    if len(rows[i]) > 5: rows[i][5] = '' # LLM used for run
+                    rows[i][4] = 'pending'
+                    if len(rows[i]) > 3: rows[i][3] = ''
+                    if len(rows[i]) > 5: rows[i][5] = ''
                     changed_count += 1
-    
+
     if changed_count > 0:
-        # Write back
         with open(csv_file + '.tmp', 'w', newline='') as f:
             writer = csv.writer(f)
             writer.writerows(rows)
         os.rename(csv_file + '.tmp', csv_file)
         print(f'[INFO] Reset {changed_count} candidates (stuck running or missing files) to pending')
-    
-    # Count pending candidates using UNIFIED logic
+
     sys.path.append(script_dir + '/..')
     from lib.evolution_csv import EvolutionCSV
-    
+
     with EvolutionCSV(csv_file) as csv_ops:
-        # Auto-fix any corrupted status fields before counting
         fixed = csv_ops.cleanup_corrupted_status_fields()
         if fixed > 0:
             print(f'[INFO] Auto-fixed {fixed} corrupted status field(s)', file=sys.stderr)
         pending = csv_ops.count_pending_candidates()
 
     print(f'[INFO] CSV loaded: {len(rows)-1} total candidates, {pending} pending')
-    
+
 except csv.Error as e:
     print(f'[ERROR] CSV parsing error: {e}')
-    print('[ERROR] The CSV file appears to be malformed')
     sys.exit(1)
 except Exception as e:
     print(f'[ERROR] Failed to read CSV: {e}')
     sys.exit(1)
 " 2>&1; then
-    echo "[ERROR] CSV validation failed. Please check the error message above."
-    exit 1
+    echo "[ERROR] CSV validation failed during cleanup" >&2
+    return 1
   fi
-fi
+
+  echo "[DISPATCHER] Full CSV cleanup complete" >&2
+  return 0
+}
+
+echo "[DISPATCHER] Starting unified evolution engine"
+echo "[DISPATCHER] Configuration: max_workers=$MAX_WORKERS, timeout=${timeout_seconds:-none}"
+
+# Perform full CSV cleanup at startup
+cleanup_csv_full || exit 1
 
 # Automatic cleanup detection - check for unchanged algorithms and warn user
 echo "[DISPATCHER] Checking for duplicate/unchanged algorithms..."
@@ -585,8 +584,28 @@ while true; do
 
   # Check if API limit was reached
   if [[ "$api_limit_reached" == "true" ]]; then
-    echo "[DISPATCHER] Stopping evolution run due to API usage limits" >&2
-    break
+    echo "[DISPATCHER] All AI models hit usage limits" >&2
+    echo "[DISPATCHER] Waiting 5 minutes before restarting the run process..." >&2
+
+    # Wait 5 minutes with countdown
+    remaining=300
+    while [[ $remaining -gt 0 ]]; do
+      if [[ $((remaining % 60)) -eq 0 ]]; then
+        echo "[DISPATCHER] Restarting in $((remaining / 60)) minutes..." >&2
+      fi
+      sleep 60
+      remaining=$((remaining - 60))
+    done
+
+    echo "[DISPATCHER] Restarting run process to clear stuck states and retry..." >&2
+    # Perform full CSV cleanup to reset stuck states, just like at startup
+    cleanup_csv_full || {
+      echo "[ERROR] CSV cleanup failed after API limit wait" >&2
+      # Continue anyway - better to try than to stop completely
+    }
+    # Clear the flag and continue - this restarts the main loop
+    api_limit_reached=false
+    continue
   fi
 
   # Periodic cleanup of stuck candidates (every 5 iterations, ~25 seconds)

package/bin/claude-evolve-worker

@@ -104,7 +104,7 @@ call_ai_for_evolution() {
   local retry_count=0
   local max_retries=3
   local wait_seconds=300  # Start with 5 minutes
-  local max_wait_seconds=1800  # Cap at 30 minutes
+  local max_wait_seconds=300  # Cap at 5 minutes
 
   while true; do
     # Capture file state before AI call
@@ -407,19 +407,41 @@ with EvolutionCSV('$FULL_CSV_PATH') as csv:
     eval_arg=""
   fi
   local eval_cmd=("$PYTHON_CMD" "$FULL_EVALUATOR_PATH" "$eval_arg")
-  
+
   # Add memory limiting if configured
+  # CRITICAL: Use multiple layers of protection (ulimit + Python wrapper + monitoring)
+  local memory_protection=""
   if [[ -n "$MEMORY_LIMIT_MB" ]] && [[ "$MEMORY_LIMIT_MB" -gt 0 ]]; then
+    # Layer 1: ulimit for hard memory limit (kernel-enforced)
+    # IMPORTANT: Use -m (RSS) not -v (virtual memory) because:
+    # - Neural networks use mmap() which bypasses RLIMIT_AS (-v)
+    # - RSS limit is more reliable for actual memory consumption
+    # Convert MB to KB for ulimit
+    local memory_limit_kb=$((MEMORY_LIMIT_MB * 1024))
+
+    # Try -m first (RSS limit), fall back to -v if not supported
+    if ulimit -m $memory_limit_kb 2>/dev/null; then
+      memory_protection="ulimit -m $memory_limit_kb 2>/dev/null; "
+      echo "[MEMORY] Layer 1: ulimit -m ${memory_limit_kb}KB (RSS limit - catches neural networks)" >&2
+    else
+      memory_protection="ulimit -v $memory_limit_kb 2>/dev/null; "
+      echo "[MEMORY] Layer 1: ulimit -v ${memory_limit_kb}KB (fallback - may not catch neural networks)" >&2
+    fi
+
+    # Layer 2: Python wrapper with PROCESS TREE monitoring (backup protection)
     eval_cmd=("$PYTHON_CMD" "$SCRIPT_DIR/../lib/memory_limit_wrapper.py" "$MEMORY_LIMIT_MB" "${eval_cmd[@]}")
+
+    echo "[MEMORY] Layer 2: Python process tree monitoring (kills entire subprocess tree)" >&2
   fi
-  
+
   # Add timeout if configured
   [[ -n "$timeout_seconds" ]] && eval_cmd=(timeout "$timeout_seconds" "${eval_cmd[@]}")
-  
-  # Run evaluation with tee to both display and capture output
+
+  # Run evaluation with memory protection, tee to both display and capture output
   # Use stdbuf to disable buffering for real-time output
   # 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
+  # The subshell ensures ulimit is applied before the command runs
+  stdbuf -o0 -e0 bash -c "${memory_protection}$(printf '%q ' "${eval_cmd[@]}")" 2>&1 | tee "$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

package/lib/memory_limit_wrapper.py

@@ -4,6 +4,16 @@ Memory-limited execution wrapper for claude-evolve evaluations.
 
 This script runs a command with memory limits to prevent runaway algorithms
 from consuming all system memory and crashing the machine.
+
+CRITICAL: Multi-layer protection approach (both must work together):
+1. ulimit -m (RSS limit) set by calling shell script - kernel-enforced, catches neural networks
+2. This Python wrapper monitors ENTIRE PROCESS TREE every 0.1s and kills if limit exceeded
+
+AIDEV-NOTE: Previous bugs fixed:
+- ulimit -v (virtual memory) doesn't catch neural networks that use mmap()
+- Was only monitoring direct child, not entire process tree (missed grandchildren)
+- Monitoring interval was 0.5s - too slow for fast memory allocations
+- Resource limit failures were silently ignored instead of failing fast
 """
 import sys
 import os
@@ -11,119 +21,197 @@ import subprocess
 import signal
 import time
 import resource
-from typing import Optional
+from typing import Optional, Tuple
 
-def set_memory_limit(limit_mb: int) -> None:
-    """Set memory limit in MB using resource module."""
+def verify_memory_limit_set(limit_mb: int) -> Tuple[bool, str]:
+    """Verify that memory limits are actually enforced."""
+    try:
+        limit_bytes = limit_mb * 1024 * 1024
+
+        # Check RLIMIT_AS (virtual memory)
+        soft_as, hard_as = resource.getrlimit(resource.RLIMIT_AS)
+        if soft_as != resource.RLIM_INFINITY and soft_as <= limit_bytes * 1.1:
+            return True, f"RLIMIT_AS set to {soft_as / (1024*1024):.0f}MB"
+
+        # Check RLIMIT_DATA (data segment)
+        try:
+            soft_data, hard_data = resource.getrlimit(resource.RLIMIT_DATA)
+            if soft_data != resource.RLIM_INFINITY and soft_data <= limit_bytes * 1.1:
+                return True, f"RLIMIT_DATA set to {soft_data / (1024*1024):.0f}MB"
+        except (OSError, ValueError):
+            pass
+
+        return False, "No hard memory limits detected"
+    except Exception as e:
+        return False, f"Error checking limits: {e}"
+
+def set_memory_limit(limit_mb: int) -> bool:
+    """
+    Set memory limit in MB using resource module.
+    Returns True if successful, False otherwise.
+    """
     try:
         # Convert MB to bytes
         limit_bytes = limit_mb * 1024 * 1024
-        
+
         # Set virtual memory limit (address space)
         # On macOS this is the most reliable way to limit memory
         resource.setrlimit(resource.RLIMIT_AS, (limit_bytes, limit_bytes))
-        
+
         # Also try to set data segment limit if available
         try:
             resource.setrlimit(resource.RLIMIT_DATA, (limit_bytes, limit_bytes))
         except (OSError, ValueError):
             # Not available on all systems
             pass
-            
-        print(f"[MEMORY] Set memory limit to {limit_mb}MB", file=sys.stderr)
-        
+
+        # Verify it was actually set
+        is_set, msg = verify_memory_limit_set(limit_mb)
+        if is_set:
+            print(f"[MEMORY] ✓ Hard limit enforced: {msg}", file=sys.stderr)
+            return True
+        else:
+            print(f"[MEMORY] ✗ Hard limit NOT enforced: {msg}", file=sys.stderr)
+            return False
+
     except (OSError, ValueError) as e:
-        print(f"[MEMORY] Warning: Could not set memory limit: {e}", file=sys.stderr)
+        print(f"[MEMORY] ✗ Could not set memory limit: {e}", file=sys.stderr)
+        return False
+
+def get_process_tree_memory_native(pid: int) -> float:
+    """Get total memory usage of entire process group using native ps command."""
+    try:
+        # Get the process group ID
+        pgid = os.getpgid(pid)
+
+        # Get all processes in the process group
+        ps_result = subprocess.run(
+            ["ps", "-o", "rss=", "-g", str(pgid)],
+            capture_output=True,
+            text=True,
+            timeout=1
+        )
+
+        if ps_result.returncode != 0:
+            return 0.0
+
+        # Sum all RSS values from the process group
+        total_rss_kb = 0
+        for line in ps_result.stdout.strip().split('\n'):
+            line = line.strip()
+            if line:
+                try:
+                    total_rss_kb += int(line)
+                except ValueError:
+                    continue
+
+        # Convert KB to MB
+        return total_rss_kb / 1024.0
+    except Exception:
+        return 0.0
 
 def monitor_memory_usage_native(process: subprocess.Popen, limit_mb: int) -> Optional[str]:
-    """Monitor process memory usage using native tools and kill if it exceeds limits."""
-    # print(f"[MEMORY] Starting native monitoring for PID {process.pid} with limit {limit_mb}MB", file=sys.stderr)
-    
+    """Monitor ENTIRE PROCESS TREE memory usage using native tools and kill if it exceeds limits."""
+    print(f"[MEMORY] Monitoring process tree from root PID {process.pid} (limit: {limit_mb}MB)", file=sys.stderr)
+
     while process.poll() is None:
         try:
-            # Use ps command to get memory usage
-            ps_result = subprocess.run(
-                ["ps", "-o", "rss=", "-p", str(process.pid)], 
-                capture_output=True, 
-                text=True, 
-                timeout=1
-            )
-            
-            if ps_result.returncode == 0 and ps_result.stdout.strip():
-                # ps returns RSS in KB, convert to MB
-                memory_kb = int(ps_result.stdout.strip())
-                memory_mb = memory_kb / 1024
-                
-                # print(f"[MEMORY] PID {process.pid} using {memory_mb:.1f}MB (limit: {limit_mb}MB)", file=sys.stderr)
-                
-                if memory_mb > limit_mb:
-                    print(f"[MEMORY] Process exceeded {limit_mb}MB limit (using {memory_mb:.1f}MB), terminating", file=sys.stderr)
-                    # Kill the entire process group - fix race condition
-                    try:
-                        pgid = os.getpgid(process.pid)
-                        os.killpg(pgid, signal.SIGTERM)
-                    except ProcessLookupError:
-                        return f"Memory limit exceeded: {memory_mb:.1f}MB > {limit_mb}MB"
-                    
-                    time.sleep(2)  # Give it time to cleanup
-                    
-                    try:
-                        if process.poll() is None:
-                            pgid = os.getpgid(process.pid) 
-                            os.killpg(pgid, signal.SIGKILL)
-                    except ProcessLookupError:
-                        pass
+            # Get total memory for entire process tree
+            memory_mb = get_process_tree_memory_native(process.pid)
+
+            if memory_mb > limit_mb:
+                print(f"[MEMORY] Process tree exceeded {limit_mb}MB limit (using {memory_mb:.1f}MB), terminating entire tree", file=sys.stderr)
+                # Kill the entire process group
+                try:
+                    pgid = os.getpgid(process.pid)
+                    os.killpg(pgid, signal.SIGTERM)
+                except ProcessLookupError:
                     return f"Memory limit exceeded: {memory_mb:.1f}MB > {limit_mb}MB"
-            
-            time.sleep(0.5)  # Check every 500ms
-            
+
+                time.sleep(2)  # Give it time to cleanup
+
+                try:
+                    if process.poll() is None:
+                        pgid = os.getpgid(process.pid)
+                        os.killpg(pgid, signal.SIGKILL)
+                except ProcessLookupError:
+                    pass
+                return f"Memory limit exceeded: {memory_mb:.1f}MB > {limit_mb}MB"
+
+            time.sleep(0.1)  # Check every 100ms for faster response
+
         except (subprocess.TimeoutExpired, ValueError, ProcessLookupError):
             # Process might have terminated or ps command failed
-            time.sleep(0.5)
+            time.sleep(0.1)
             continue
-    
-    # print(f"[MEMORY] Monitoring stopped for PID {process.pid}", file=sys.stderr)
+
     return None
 
+def get_process_tree_memory_psutil(ps_process) -> float:
+    """Get total memory usage of entire process tree using psutil."""
+    try:
+        import psutil
+        total_mb = 0.0
+
+        # Get memory of root process
+        try:
+            total_mb += ps_process.memory_info().rss / (1024 * 1024)
+        except (psutil.NoSuchProcess, psutil.AccessDenied):
+            return 0.0
+
+        # Get memory of all children (recursive)
+        try:
+            for child in ps_process.children(recursive=True):
+                try:
+                    total_mb += child.memory_info().rss / (1024 * 1024)
+                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                    continue
+        except (psutil.NoSuchProcess, psutil.AccessDenied):
+            pass
+
+        return total_mb
+    except ImportError:
+        return 0.0
+
 def monitor_memory_usage(process: subprocess.Popen, limit_mb: int) -> Optional[str]:
-    """Monitor process memory usage and kill if it exceeds limits."""
+    """Monitor ENTIRE PROCESS TREE memory usage and kill if it exceeds limits."""
     try:
         import psutil
         ps_process = psutil.Process(process.pid)
-        
+        print(f"[MEMORY] Monitoring process tree from root PID {process.pid} (limit: {limit_mb}MB, using psutil)", file=sys.stderr)
+
         while process.poll() is None:
             try:
-                # Get memory usage in MB
-                memory_info = ps_process.memory_info()
-                memory_mb = memory_info.rss / (1024 * 1024)
-                
+                # Get total memory for entire process tree
+                memory_mb = get_process_tree_memory_psutil(ps_process)
+
                 if memory_mb > limit_mb:
-                    print(f"[MEMORY] Process exceeded {limit_mb}MB limit (using {memory_mb:.1f}MB), terminating", file=sys.stderr)
-                    # Kill the entire process group - fix race condition
+                    print(f"[MEMORY] Process tree exceeded {limit_mb}MB limit (using {memory_mb:.1f}MB), terminating entire tree", file=sys.stderr)
+                    # Kill the entire process group
                     try:
                         pgid = os.getpgid(process.pid)
                         os.killpg(pgid, signal.SIGTERM)
                     except ProcessLookupError:
                         return f"Memory limit exceeded: {memory_mb:.1f}MB > {limit_mb}MB"
-                    
+
                     time.sleep(2)  # Give it time to cleanup
-                    
+
                     try:
                         if process.poll() is None:
-                            pgid = os.getpgid(process.pid) 
+                            pgid = os.getpgid(process.pid)
                             os.killpg(pgid, signal.SIGKILL)
                     except ProcessLookupError:
                         pass
                     return f"Memory limit exceeded: {memory_mb:.1f}MB > {limit_mb}MB"
-                
-                time.sleep(0.5)  # Check every 500ms
+
+                time.sleep(0.1)  # Check every 100ms for faster response
             except (psutil.NoSuchProcess, psutil.AccessDenied):
                 # Process already terminated
                 break
     except ImportError:
         # psutil not available, use native monitoring
         return monitor_memory_usage_native(process, limit_mb)
-    
+
     return None
 
 def validate_memory_limit(limit_mb: int) -> bool:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-evolve",
-  "version": "1.8.12",
+  "version": "1.8.14",
   "bin": {
     "claude-evolve": "./bin/claude-evolve",
     "claude-evolve-main": "./bin/claude-evolve-main",