claude-evolve 1.5.3 → 1.5.6

package/bin/claude-evolve-worker

@@ -7,6 +7,17 @@ source "$SCRIPT_DIR/../lib/config.sh"
 source "$SCRIPT_DIR/../lib/csv-lock.sh"
 source "$SCRIPT_DIR/../lib/ai-cli.sh"
 
+# Setup logging to file
+if [[ -n "${FULL_EVOLUTION_DIR:-}" ]]; then
+  LOG_DIR="$FULL_EVOLUTION_DIR/logs"
+  mkdir -p "$LOG_DIR"
+  LOG_FILE="$LOG_DIR/worker-$$-$(date +%Y%m%d-%H%M%S).log"
+  
+  # Log to both terminal and file with timestamps
+  exec > >(while IFS= read -r line; do echo "$(date '+%Y-%m-%d %H:%M:%S'): $line"; done | tee -a "$LOG_FILE") 2>&1
+  echo "[WORKER-$$] Logging to: $LOG_FILE"
+fi
+
 # Track current candidate for cleanup
 CURRENT_CANDIDATE_ID=""
 TERMINATION_SIGNAL=""
@@ -14,34 +25,11 @@ TERMINATION_SIGNAL=""
 # Cleanup function to handle termination
 cleanup_on_exit() {
   if [[ -n "$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
 }
 
@@ -85,6 +73,17 @@ call_ai_for_evolution() {
   local prompt="$1"
   local candidate_id="$2"
   
+  # Get target file path from worker context
+  local target_file="$FULL_OUTPUT_DIR/evolution_${candidate_id}.py"
+  
+  # Capture file state before AI call
+  local file_hash_before=""
+  local file_mtime_before=""
+  if [[ -f "$target_file" ]]; then
+    file_hash_before=$(shasum -a 256 "$target_file" 2>/dev/null | cut -d' ' -f1)
+    file_mtime_before=$(stat -f %m "$target_file" 2>/dev/null || stat -c %Y "$target_file" 2>/dev/null)
+  fi
+  
   # Extract generation and ID numbers for round-robin calculation
   local gen_num=0
   local id_num=0
@@ -108,14 +107,32 @@ call_ai_for_evolution() {
     exit 3
   fi
   
-  if [[ $ai_exit_code -eq 0 ]]; then
-    echo "[WORKER-$$] AI succeeded" >&2
+  # Check if the target file was actually modified
+  local file_was_modified=false
+  if [[ -f "$target_file" ]]; then
+    local file_hash_after
+    local file_mtime_after
+    file_hash_after=$(shasum -a 256 "$target_file" 2>/dev/null | cut -d' ' -f1)
+    file_mtime_after=$(stat -f %m "$target_file" 2>/dev/null || stat -c %Y "$target_file" 2>/dev/null)
+    
+    if [[ "$file_hash_before" != "$file_hash_after" ]] || [[ "$file_mtime_before" != "$file_mtime_after" ]]; then
+      file_was_modified=true
+    fi
+  fi
+  
+  # Success if file was modified OR exit code is 0 (for cases where file validation isn't applicable)
+  if [[ "$file_was_modified" == "true" ]] || [[ $ai_exit_code -eq 0 ]]; then
+    if [[ "$file_was_modified" == "true" ]]; then
+      echo "[WORKER-$$] AI successfully modified $target_file (exit code: $ai_exit_code)" >&2
+    else
+      echo "[WORKER-$$] AI succeeded with exit code 0" >&2
+    fi
     # Output the result for the worker to use
     echo "$ai_output"
     return 0
   fi
   
-  echo "[WORKER-$$] All AI models failed" >&2
+  echo "[WORKER-$$] AI failed: exit code $ai_exit_code, no file changes detected" >&2
   return 1
 }
 
@@ -135,6 +152,12 @@ 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 +236,38 @@ 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-$$] Evolution applied successfully"
+      
+      # Record which AI model generated the code (regardless of evaluation outcome)
+      if [[ -n "${SUCCESSFUL_RUN_MODEL:-}" ]]; then
+        echo "[WORKER-$$] Recording that $SUCCESSFUL_RUN_MODEL generated the code" >&2
+        "$PYTHON_CMD" -c "
+import sys
+sys.path.insert(0, '$SCRIPT_DIR/..')
+from lib.evolution_csv import EvolutionCSV
+with EvolutionCSV('$FULL_CSV_PATH') as csv:
+    csv.update_candidate_field('$candidate_id', 'run-LLM', '$SUCCESSFUL_RUN_MODEL')
+" || echo "[WORKER-$$] Warning: Failed to record run-LLM field" >&2
+      fi
+      
+      # Check if the generated Python file has syntax errors
+      echo "[WORKER-$$] 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 +284,22 @@ 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 +409,19 @@ 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-$$] Evaluation output saved to: $eval_output_file" >&2
+    
+    # Clear CURRENT_CANDIDATE_ID on successful completion
+    CURRENT_CANDIDATE_ID=""
   else
-    local exit_code=$?
+    local exit_code=$eval_exit_code
     # Read any output that was captured before failure
     eval_output=$(<"$eval_output_file")
     # rm -f "$eval_output_file"  # Keep for debugging
@@ -370,22 +431,56 @@ with EvolutionCSV('$FULL_CSV_PATH') as csv:
     echo "[WORKER-$$] 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('$FULL_CSV_PATH') as csv:
+    rows = csv._read_csv()
+    if rows:
+        start_idx = 1 if rows and rows[0] and rows[0][0].lower() == 'id' else 0
+        status_count = {}
+        for row in rows[start_idx:]:
+            if len(row) > 4:
+                status = row[4].strip() or 'pending'
+                status_count[status] = status_count.get(status, 0) + 1
+        print(f'Status counts: {status_count}', file=sys.stderr)
+" 2>&1 || true
+
   # Try to claim a pending candidate
   candidate_info=$("$PYTHON_CMD" -c "
 import sys
@@ -412,12 +507,26 @@ with EvolutionCSV('$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

@@ -12,31 +12,50 @@ call_ai_model_configured() {
   local model_name="$1"
   local prompt="$2"
   
+  # Record start time
+  local start_time=$(date +%s)
+  
   # Build command directly based on model
   case "$model_name" in
     opus|sonnet)
       local ai_output
-      ai_output=$(timeout 300 claude --dangerously-skip-permissions --model "$model_name" -p "$prompt" 2>&1)
+      ai_output=$(timeout 180 claude --dangerously-skip-permissions --model "$model_name" -p "$prompt" 2>&1)
+      local ai_exit_code=$?
+      ;;
+    gpt5high)
+      local ai_output
+      ai_output=$(timeout 420 codex exec --profile gpt5high --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
-    o3)
+    o3high)
       local ai_output
-      ai_output=$(timeout 300 codex exec -m o3 --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      ai_output=$(timeout 500 codex exec --profile o3high --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     codex)
       local ai_output
-      ai_output=$(timeout 300 codex exec --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      ai_output=$(timeout 420 codex exec --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     gemini)
       # Debug: Show exact command
-      echo "[DEBUG] Running: timeout 300 gemini -y -p <prompt>" >&2
+      echo "[DEBUG] Running: timeout 1200 gemini -y -p <prompt>" >&2
       echo "[DEBUG] Working directory: $(pwd)" >&2
       echo "[DEBUG] Files in current dir:" >&2
       ls -la temp-csv-*.csv 2>&1 | head -5 >&2
       local ai_output
-      ai_output=$(timeout 300 gemini -y -p "$prompt" 2>&1)
+      # Gemini needs longer timeout as it streams output while working (20 minutes)
+      ai_output=$(timeout 1200 gemini -y -p "$prompt" 2>&1)
+      local ai_exit_code=$?
+      ;;
+    cursor-sonnet)
+      local ai_output
+      ai_output=$(timeout 180 cursor-agent sonnet -p "$prompt" 2>&1)
+      local ai_exit_code=$?
+      ;;
+    cursor-opus)
+      local ai_output
+      ai_output=$(timeout 300 cursor-agent opus -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     *)
@@ -48,8 +67,12 @@ call_ai_model_configured() {
   # Debug: log model and prompt size
   echo "[DEBUG] Calling $model_name with prompt of ${#prompt} characters" >&2
   
-  # Always log basic info
-  echo "[AI] $model_name exit code: $ai_exit_code, output length: ${#ai_output} chars" >&2
+  # Calculate duration
+  local end_time=$(date +%s)
+  local duration=$((end_time - start_time))
+  
+  # Always log basic info with timing
+  echo "[AI] $model_name exit code: $ai_exit_code, output length: ${#ai_output} chars, duration: ${duration}s" >&2
   
   # Show detailed output if verbose or if there was an error
   if [[ "${VERBOSE_AI_OUTPUT:-false}" == "true" ]] || [[ $ai_exit_code -ne 0 ]]; then
@@ -100,7 +123,7 @@ 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" == "o3high" || "$model_name" == "gpt5high" ]]; then
     # Clean codex output - extract content between "codex" marker and "tokens used"
     if echo "$output" | grep -q "^\[.*\] codex$"; then
       # Extract content between "codex" line and "tokens used" line
@@ -191,11 +214,19 @@ call_ai_with_round_robin() {
     ai_output=$(call_ai_model_configured "$model" "$prompt")
     local ai_exit_code=$?
     
-    # Just check exit code - no interpretation
-    if [[ $ai_exit_code -eq 0 ]]; then
-      # Clean output if needed
-      ai_output=$(clean_ai_output "$ai_output" "$model")
-      echo "[AI] $model returned exit code 0" >&2
+    # Clean output if needed
+    ai_output=$(clean_ai_output "$ai_output" "$model")
+    
+    # Success if exit code is 0, or if it's just a timeout (124)
+    # Timeout doesn't mean the AI failed - it may have completed the task
+    if [[ $ai_exit_code -eq 0 ]] || [[ $ai_exit_code -eq 124 ]]; then
+      if [[ $ai_exit_code -eq 124 ]]; then
+        echo "[AI] $model timed out but continuing (exit code: 124)" >&2
+      else
+        echo "[AI] $model returned exit code 0" >&2
+      fi
+      # Export the successful model for tracking (used by worker)
+      export SUCCESSFUL_RUN_MODEL="$model"
       # Debug: log what AI returned on success
       if [[ "${DEBUG_AI_SUCCESS:-}" == "true" ]]; then
         echo "[AI] $model success output preview:" >&2

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"
-DEFAULT_LLM_IDEATE="gemini o3 opus"
+# 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 cursor-sonnet"
+DEFAULT_LLM_IDEATE="gemini opus gpt5high o3high cursor-opus"
 
 # Load configuration from config file
 load_config() {
@@ -96,14 +92,20 @@ load_config() {
   # Set retry default
   MAX_RETRIES="$DEFAULT_MAX_RETRIES"
   
+  # Set memory limit default
+  MEMORY_LIMIT_MB="$DEFAULT_MEMORY_LIMIT_MB"
+  
   # Set LLM CLI defaults (compatibility for older bash)
   # Initialize associative array for LLM commands
   # Use simpler approach for compatibility
-  LLM_CLI_o3='codex exec -m o3 --dangerously-bypass-approvals-and-sandbox "{{PROMPT}}"'
+  LLM_CLI_gpt5high='codex exec --profile gpt5high --dangerously-bypass-approvals-and-sandbox "{{PROMPT}}"'
+  LLM_CLI_o3high='codex exec --profile o3high --dangerously-bypass-approvals-and-sandbox "{{PROMPT}}"'
   LLM_CLI_codex='codex exec --dangerously-bypass-approvals-and-sandbox "{{PROMPT}}"'
   LLM_CLI_gemini='gemini -y -p "{{PROMPT}}"'
   LLM_CLI_opus='claude --dangerously-skip-permissions --model opus -p "{{PROMPT}}"'
   LLM_CLI_sonnet='claude --dangerously-skip-permissions --model sonnet -p "{{PROMPT}}"'
+  LLM_CLI_cursor_sonnet='cursor-agent sonnet -p "{{PROMPT}}"'
+  LLM_CLI_cursor_opus='cursor-agent opus -p "{{PROMPT}}"'
   LLM_RUN="$DEFAULT_LLM_RUN"
   LLM_IDEATE="$DEFAULT_LLM_IDEATE"
   
@@ -202,12 +204,14 @@ 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 +225,7 @@ 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 +321,18 @@ show_config() {
   echo "  Lock timeout: $LOCK_TIMEOUT"
   echo "  Auto ideate: $AUTO_IDEATE"
   echo "  Max retries: $MAX_RETRIES"
+  echo "  Memory limit: ${MEMORY_LIMIT_MB}MB"
   echo "  LLM configuration:"
   # Show LLM configurations using dynamic variable names
-  for model in o3 codex gemini opus sonnet; do
+  for model in gpt5high o3high codex gemini opus sonnet cursor_sonnet cursor_opus; do
     var_name="LLM_CLI_${model}"
-    if [[ -n "${!var_name}" ]]; then
-      echo "    $model: ${!var_name}"
+    var_value=$(eval echo "\$$var_name")
+    if [[ -n "$var_value" ]]; then
+      # Convert underscore back to dash for display
+      display_name=$(echo "$model" | sed 's/_/-/g')
+      echo "    $display_name: $var_value"
     fi
   done
   echo "  LLM for run: $LLM_RUN"
   echo "  LLM for ideate: $LLM_IDEATE"
-}
+}

package/lib/csv_fixer.py

@@ -0,0 +1,35 @@
+#!/usr/bin/env python3
+"""
+CSV format fixer for claude-evolve
+Ensures proper quoting of CSV fields, especially descriptions
+"""
+
+import csv
+import sys
+
+def fix_csv_format(input_file, output_file):
+    """
+    Read a CSV file and ensure all fields are properly quoted.
+    The csv module handles quoting automatically based on content.
+    """
+    with open(input_file, 'r') as infile:
+        reader = csv.reader(infile)
+        rows = list(reader)
+    
+    with open(output_file, 'w', newline='') as outfile:
+        writer = csv.writer(outfile, quoting=csv.QUOTE_NONNUMERIC)
+        
+        # Write all rows - csv.writer handles quoting automatically
+        for row in rows:
+            writer.writerow(row)
+
+if __name__ == "__main__":
+    if len(sys.argv) != 3:
+        print("Usage: csv_fixer.py <input_file> <output_file>", file=sys.stderr)
+        sys.exit(1)
+    
+    try:
+        fix_csv_format(sys.argv[1], sys.argv[2])
+    except Exception as e:
+        print(f"Error fixing CSV: {e}", file=sys.stderr)
+        sys.exit(1)

package/lib/evolution_csv.py

@@ -199,7 +199,7 @@ class EvolutionCSV:
         for i in range(start_idx, len(rows)):
             row = rows[i]
             
-            if self.is_valid_candidate_row(row) and row[0].strip() == candidate_id:
+            if self.is_valid_candidate_row(row) and row[0].strip().strip('"') == candidate_id.strip().strip('"'):
                 # Ensure row has at least 5 columns
                 while len(row) < 5:
                     row.append('')
@@ -227,7 +227,7 @@ class EvolutionCSV:
         for i in range(start_idx, len(rows)):
             row = rows[i]
             
-            if self.is_valid_candidate_row(row) and row[0].strip() == candidate_id:
+            if self.is_valid_candidate_row(row) and row[0].strip().strip('"') == candidate_id.strip().strip('"'):
                 # Ensure row has at least 4 columns
                 while len(row) < 4:
                     row.append('')
@@ -263,10 +263,14 @@ class EvolutionCSV:
                     field_index = i
                     break
             
-            # If field doesn't exist, add it to header
+            # If field doesn't exist, add it to header and extend all rows
             if field_index is None:
                 field_index = len(header_row)
                 header_row.append(field_name)
+                # Extend all data rows with empty values for the new column
+                for i in range(1, len(rows)):
+                    while len(rows[i]) <= field_index:
+                        rows[i].append('')
         else:
             # No header - we'll use predefined positions for known fields
             field_map = {
@@ -274,7 +278,9 @@ class EvolutionCSV:
                 'basedonid': 1,
                 'description': 2,
                 'performance': 3,
-                'status': 4
+                'status': 4,
+                'idea-llm': 5,
+                'run-llm': 6
             }
             field_index = field_map.get(field_name.lower())
             if field_index is None:
@@ -287,7 +293,10 @@ class EvolutionCSV:
         
         for i in range(start_idx, len(rows)):
             row = rows[i]
-            if self.is_valid_candidate_row(row) and row[0].strip() == candidate_id:
+            # Strip quotes from both stored ID and search ID for comparison
+            stored_id = row[0].strip().strip('"') if len(row) > 0 else ''
+            search_id = candidate_id.strip().strip('"')
+            if self.is_valid_candidate_row(row) and stored_id == search_id:
                 # Ensure row has enough columns
                 while len(row) <= field_index:
                     row.append('')
@@ -311,7 +320,7 @@ class EvolutionCSV:
         start_idx = 1 if rows and rows[0] and rows[0][0].lower() == 'id' else 0
         
         for row in rows[start_idx:]:
-            if self.is_valid_candidate_row(row) and row[0].strip() == candidate_id:
+            if self.is_valid_candidate_row(row) and row[0].strip().strip('"') == candidate_id.strip().strip('"'):
                 return {
                     'id': row[0].strip() if len(row) > 0 else '',
                     'basedOnId': row[1].strip() if len(row) > 1 else '',
@@ -344,7 +353,7 @@ class EvolutionCSV:
             
         for i in range(start_idx, len(rows)):
             row = rows[i]
-            if self.is_valid_candidate_row(row) and row[0].strip() == candidate_id:
+            if self.is_valid_candidate_row(row) and row[0].strip().strip('"') == candidate_id.strip().strip('"'):
                 deleted = True
                 # Skip this row (delete it)
                 continue
@@ -371,6 +380,7 @@ def main():
         print("  update <id> <status>    - Update candidate status")
         print("  perf <id> <performance> - Update candidate performance")
         print("  info <id>               - Get candidate info")
+        print("  field <id> <field> <val>- Update specific field")
         print("  check                   - Check if has pending work")
         sys.exit(1)
         
@@ -430,6 +440,17 @@ def main():
                 has_work = csv_ops.has_pending_work()
                 print("yes" if has_work else "no")
                 
+            elif command == 'field' and len(sys.argv) >= 5:
+                candidate_id = sys.argv[3]
+                field_name = sys.argv[4]
+                value = sys.argv[5] if len(sys.argv) >= 6 else ''
+                success = csv_ops.update_candidate_field(candidate_id, field_name, value)
+                if success:
+                    print(f"Updated {candidate_id} field {field_name} to {value}")
+                else:
+                    print(f"Failed to update {candidate_id} field {field_name}")
+                    sys.exit(1)
+                
             else:
                 print(f"Unknown command: {command}")
                 sys.exit(1)