claude-evolve 1.4.13 → 1.5.1

package/bin/claude-evolve-ideate

@@ -81,47 +81,16 @@ call_ai_with_limit_check() {
   if [[ "$preferred_model" == "o3" ]] && command -v codex >/dev/null 2>&1; then
     echo "[INFO] Using codex o3 for ideation" >&2
     
-    # Call codex with o3 model using -q flag and --full-auto
+    # Call codex with o3 model using exec subcommand
     local ai_output
-    ai_output=$(codex -m o3 --full-auto -q "$prompt" 2>&1)
+    # Pass prompt as argument, not via stdin
+    ai_output=$(timeout 300 codex exec -m o3 --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
     local ai_exit_code=$?
     
     if [[ $ai_exit_code -eq 0 ]]; then
-      # Clean o3 output - it may be JSON with the response in a field
-      local cleaned_output
-      # Try to extract content from JSON response if present
-      if echo "$ai_output" | grep -q '"content"'; then
-        # Attempt to extract content field from JSON
-        cleaned_output=$(echo "$ai_output" | python3 -c "
-import sys
-import json
-try:
-    data = json.load(sys.stdin)
-    if 'content' in data:
-        print(data['content'])
-    elif 'response' in data:
-        print(data['response'])
-    elif 'text' in data:
-        print(data['text'])
-    else:
-        # If no known field, print the whole thing
-        print(json.dumps(data))
-except:
-    # If not valid JSON, print as-is
-    print(sys.stdin.read())
-" 2>/dev/null || echo "$ai_output")
-      else
-        cleaned_output="$ai_output"
-      fi
-      
-      # Validate the output is not empty and doesn't contain error messages
-      if [[ -n "$cleaned_output" ]] && ! echo "$cleaned_output" | grep -q "error\|failed\|exception"; then
-        echo "$cleaned_output"
-        return 0
-      else
-        echo "[WARN] Codex o3 returned invalid output, falling back to Claude Opus" >&2
-        preferred_model="opus"
-      fi
+      # For ideation, AI modifies files directly - just return success
+      echo "[INFO] Codex o3 succeeded" >&2
+      return 0
     else
       echo "[WARN] Codex o3 failed with exit code $ai_exit_code, falling back to Claude Opus" >&2
       preferred_model="opus"
@@ -131,21 +100,13 @@ except:
     
     # Call gemini with -y and -p flags
     local ai_output
-    ai_output=$(gemini -y -p "$prompt" 2>&1)
+    ai_output=$(echo "$prompt" | timeout 300 gemini -y -p 2>&1)
     local ai_exit_code=$?
     
     if [[ $ai_exit_code -eq 0 ]]; then
-      # Check for authentication messages or other non-response content
-      if echo "$ai_output" | grep -q "Attempting to authenticate\|Authenticating\|Loading\|Initializing"; then
-        echo "[WARN] Gemini is still authenticating, falling back to Claude Opus" >&2
-        preferred_model="opus"
-      elif [[ -z "$ai_output" ]] || [[ $(echo "$ai_output" | wc -l) -lt 2 ]]; then
-        echo "[WARN] Gemini returned insufficient output, falling back to Claude Opus" >&2
-        preferred_model="opus"
-      else
-        echo "$ai_output"
-        return 0
-      fi
+      # For ideation, AI modifies files directly - just return success
+      echo "[INFO] Gemini succeeded" >&2
+      return 0
     else
       echo "[WARN] Gemini failed with exit code $ai_exit_code, falling back to Claude Opus" >&2
       preferred_model="opus"
@@ -157,7 +118,7 @@ except:
   
   # Call Claude and capture output
   local claude_output
-  claude_output=$(echo "$prompt" | claude --dangerously-skip-permissions --model "$preferred_model" -p 2>&1)
+  claude_output=$(echo "$prompt" | timeout 300 claude --dangerously-skip-permissions --model "$preferred_model" -p 2>&1)
   local claude_exit_code=$?
   
   # Check for usage limit
@@ -181,17 +142,13 @@ except:
     exit 1
   fi
   
-  # Validate output doesn't contain shell constructs that could corrupt CSV
-  if echo "$claude_output" | grep -E "EOF.*<.*null|<<.*EOF|<.*dev.*null" >/dev/null 2>&1; then
-    echo "[ERROR] AI output contains invalid shell constructs" >&2
-    echo "[DEBUG] Problematic output: $claude_output" | head -5 >&2
-    return 1
+  if [[ $claude_exit_code -eq 0 ]]; then
+    # For ideation, AI modifies files directly - just return success
+    echo "[INFO] Claude $preferred_model succeeded" >&2
+    return 0
+  else
+    return $claude_exit_code
   fi
-  
-  # Output Claude's response
-  echo "$claude_output"
-  
-  return $claude_exit_code
 }
 
 # Backward compatibility alias
@@ -204,68 +161,63 @@ call_ai_with_fallbacks() {
   local prompt="$1"
   local generation="${2:-01}"
   
-  # List of models to try in order
-  local models=()
+  # Determine which model to use based on generation (round-robin)
+  local gen_num
+  if [[ $generation =~ ^0*([0-9]+)$ ]]; then
+    gen_num=$((10#${BASH_REMATCH[1]}))
+  else
+    gen_num=1
+  fi
   
   # Check which AI tools are available
+  local available_models=()
   if command -v codex >/dev/null 2>&1; then
-    models+=("o3")
+    available_models+=("o3")
   fi
   if command -v gemini >/dev/null 2>&1; then
-    models+=("gemini")
+    available_models+=("gemini")
   fi
-  # Claude is always available (fallback)
-  models+=("opus" "sonnet" "haiku")
+  available_models+=("opus")  # Claude Opus always available
+  
+  # Create ordered list based on round-robin for this generation
+  local num_models=${#available_models[@]}
+  local start_index=$((gen_num % num_models))
+  local models=()
+  
+  # Add models in round-robin order starting from the calculated index
+  for ((i=0; i<num_models; i++)); do
+    local idx=$(((start_index + i) % num_models))
+    models+=("${available_models[$idx]}")
+  done
+  
+  echo "[INFO] Model order for generation $generation (round-robin): ${models[*]}" >&2
   
-  # Try each model in sequence
+  # Try each model in the ordered sequence
   for model in "${models[@]}"; do
-    echo "[INFO] Trying $model for ideation (fallback attempt)" >&2
+    echo "[INFO] Trying $model for ideation" >&2
     
     local ai_output
     local ai_exit_code
     
     if [[ "$model" == "o3" ]] && command -v codex >/dev/null 2>&1; then
-      ai_output=$(codex -m o3 --full-auto -q "$prompt" 2>&1)
+      # Pass prompt as argument, not via stdin
+    ai_output=$(timeout 300 codex exec -m o3 --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       ai_exit_code=$?
       
       if [[ $ai_exit_code -eq 0 ]]; then
-        # Clean o3 output like in the original function
-        if echo "$ai_output" | grep -q '"content"'; then
-          ai_output=$(echo "$ai_output" | python3 -c "
-import sys
-import json
-try:
-    data = json.load(sys.stdin)
-    if 'content' in data:
-        print(data['content'])
-    elif 'response' in data:
-        print(data['response'])
-    elif 'text' in data:
-        print(data['text'])
-    else:
-        print(json.dumps(data))
-except:
-    print(sys.stdin.read())
-" 2>/dev/null || echo "$ai_output")
-        fi
-        
-        if [[ -n "$ai_output" ]] && ! echo "$ai_output" | grep -q "error\|failed\|exception"; then
-          echo "$ai_output"
-          return 0
-        fi
+        # For ideation, we don't care about output content - AI modifies files directly
+        echo "[INFO] o3 completed successfully (exit code 0)" >&2
+        return 0
       fi
       
     elif [[ "$model" == "gemini" ]] && command -v gemini >/dev/null 2>&1; then
-      ai_output=$(gemini -y -p "$prompt" 2>&1)
+      ai_output=$(echo "$prompt" | timeout 300 gemini -y -p 2>&1)
       ai_exit_code=$?
       
       if [[ $ai_exit_code -eq 0 ]]; then
-        if ! echo "$ai_output" | grep -q "Attempting to authenticate\|Authenticating\|Loading\|Initializing"; then
-          if [[ -n "$ai_output" ]] && [[ $(echo "$ai_output" | wc -l) -ge 2 ]]; then
-            echo "$ai_output"
-            return 0
-          fi
-        fi
+        # For ideation, we don't care about output content - AI modifies files directly
+        echo "[INFO] gemini completed successfully (exit code 0)" >&2
+        return 0
       fi
       
     else
@@ -275,16 +227,20 @@ except:
       
       if [[ $ai_exit_code -eq 0 ]]; then
         # Check for usage limits
-        if ! echo "$ai_output" | grep -q "Claude AI usage limit reached"; then
-          if ! echo "$ai_output" | grep -E "EOF.*<.*null|<<.*EOF|<.*dev.*null" >/dev/null 2>&1; then
-            echo "$ai_output"
-            return 0
-          fi
+        if echo "$ai_output" | grep -q "Claude AI usage limit reached"; then
+          echo "[INFO] Claude hit usage limit" >&2
+        else
+          # For ideation, we don't care about output content - AI modifies files directly
+          echo "[INFO] $model completed successfully (exit code 0)" >&2
+          return 0
         fi
       fi
     fi
     
     echo "[WARN] $model failed or returned unusable output, trying next model..." >&2
+    echo "[DEBUG] $model exit code: ${ai_exit_code:-unknown}" >&2
+    echo "[DEBUG] $model output length: ${#ai_output} chars" >&2
+    echo "[DEBUG] $model output preview: $(echo "$ai_output" | head -3 | tr '\n' ' ')" >&2
   done
   
   echo "[ERROR] All AI models failed to generate usable output" >&2
@@ -422,17 +378,18 @@ validate_direct_csv_modification() {
   fi
   
   # Validate the modified CSV has more entries than original
+  # Count actual data rows (excluding header and empty lines)
   local original_count
-  original_count=$(wc -l < "$FULL_CSV_PATH")
+  original_count=$(grep -v '^[[:space:]]*$' "$FULL_CSV_PATH" | tail -n +2 | wc -l)
   local new_count
-  new_count=$(wc -l < "$temp_csv")
+  new_count=$(grep -v '^[[:space:]]*$' "$temp_csv" | tail -n +2 | wc -l)
   
-  echo "[DEBUG] Original CSV line count: $original_count" >&2
-  echo "[DEBUG] Modified CSV line count: $new_count" >&2
+  echo "[DEBUG] Original CSV data rows: $original_count" >&2
+  echo "[DEBUG] Modified CSV data rows: $new_count" >&2
   echo "[DEBUG] Expected to add: $expected_count ideas" >&2
   
   if [[ $new_count -le $original_count ]]; then
-    echo "[ERROR] CSV file wasn't modified - same number of lines ($new_count <= $original_count)" >&2
+    echo "[ERROR] CSV file wasn't modified - same number of data rows ($new_count <= $original_count)" >&2
     echo "[DEBUG] First 10 lines of CSV after AI attempt:" >&2
     head -10 "$temp_csv" >&2
     return 1
@@ -529,8 +486,8 @@ process_ai_ideas_direct_old() {
   local top_performers="${3:-}"  # Optional, for non-novel ideas
   local ai_output="$4"  # The AI's response with ideas
   
-  # Create temporary CSV copy
-  local temp_csv="/tmp/claude-evolve-temp-csv-$$.csv"
+  # Create temporary CSV copy in evolution directory (so AI can access it)
+  local temp_csv="$FULL_EVOLUTION_DIR/temp-csv-$$.csv"
   cp "$FULL_CSV_PATH" "$temp_csv"
   
   echo "[DEBUG] Starting CSV modification for $count $idea_type ideas" >&2
@@ -787,42 +744,58 @@ ideate_ai_strategies() {
 generate_novel_ideas_direct() {
   local count="$1"
   
-  # Create temporary CSV copy
-  local temp_csv="/tmp/claude-evolve-temp-csv-$$.csv"
+  # Create temporary CSV copy in evolution directory (so AI can access it)
+  local temp_csv="$FULL_EVOLUTION_DIR/temp-csv-$$.csv"
   cp "$FULL_CSV_PATH" "$temp_csv"
   
   echo "[INFO] Generating $count novel exploration ideas..."
-  echo "[DEBUG] Original CSV has $(wc -l < "$FULL_CSV_PATH") lines" >&2
+  local data_rows=$(grep -v '^[[:space:]]*$' "$FULL_CSV_PATH" | tail -n +2 | wc -l)
+  echo "[DEBUG] Original CSV has $data_rows data rows" >&2
   
-  local prompt="Please add exactly $count novel algorithmic ideas directly to the CSV file $temp_csv.
+  # Use relative paths and change to evolution directory so AI can access files
+  local temp_csv_basename=$(basename "$temp_csv")
+  
+  local prompt="I need you to use your file editing capabilities to add exactly $count novel algorithmic ideas to the CSV file: $temp_csv_basename
 
 Current evolution context:
 - Generation: $CURRENT_GENERATION
-- Algorithm: $FULL_ALGORITHM_PATH
+- Algorithm: algorithm.py (base algorithm)
 - Brief: $(head -20 "$FULL_BRIEF_PATH")
 
 Instructions:
-1. Read the current CSV file to see existing entries
-2. Add exactly $count new rows to the CSV file
-3. Use the next available generation numbers (gen$CURRENT_GENERATION-XXX format)
-4. For each idea, create a row with: id,,description,,pending (empty parent_id for novel ideas)
-5. Each description should be one clear sentence describing a novel algorithmic approach
-6. Focus on creative, ambitious ideas that haven't been tried yet
-7. Consider machine learning, new indicators, regime detection, risk management, etc.
+1. Use the Read tool to examine the current CSV file
+2. Find the highest ID number for generation $CURRENT_GENERATION (e.g., if gen$CURRENT_GENERATION-003 exists, next should be gen$CURRENT_GENERATION-004)
+3. If no gen$CURRENT_GENERATION entries exist yet, start with gen$CURRENT_GENERATION-001
+4. Use the Edit or MultiEdit tool to add exactly $count new rows to the CSV file
+5. For each idea, create a row with: id,,description,,pending (empty parent_id for novel ideas)
+6. Each description should be one clear sentence describing a novel algorithmic approach
+7. Focus on creative, ambitious ideas that haven't been tried yet
+8. Consider machine learning, new indicators, regime detection, risk management, etc.
 
-IMPORTANT: Edit the CSV file directly. Do not return or print CSV data - just modify the file in place."
+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."
 
+  # Change to evolution directory so AI can access files
+  local original_pwd=$(pwd)
+  cd "$FULL_EVOLUTION_DIR"
+  
   # Get AI to directly edit the CSV file
   local ai_response
-  local stderr_file="/tmp/claude-evolve-stderr-$$.txt"
+  local stderr_file="stderr-$$.txt"
   if ! ai_response=$(call_ai_with_fallbacks "$prompt" "$CURRENT_GENERATION" 2>"$stderr_file"); then
     echo "[ERROR] All AI models failed to generate novel ideas" >&2
+    echo "[DEBUG] Stderr output from AI calls:" >&2
     cat "$stderr_file" >&2
+    echo "[DEBUG] Temp CSV location: $temp_csv" >&2
+    echo "[DEBUG] Working directory: $(pwd)" >&2
+    cd "$original_pwd"
     rm -f "$temp_csv" "$stderr_file"
     return 1
   fi
   rm -f "$stderr_file"
   
+  # Restore working directory
+  cd "$original_pwd"
+  
   echo "[DEBUG] AI response: $ai_response" >&2
   
   # Validate that the CSV file was actually modified
@@ -840,18 +813,22 @@ generate_hill_climbing_direct() {
   local count="$1"
   local top_performers="$2"
   
-  # Create temporary CSV copy
-  local temp_csv="/tmp/claude-evolve-temp-csv-$$.csv"
+  # Create temporary CSV copy in evolution directory (so AI can access it)
+  local temp_csv="$FULL_EVOLUTION_DIR/temp-csv-$$.csv"
   cp "$FULL_CSV_PATH" "$temp_csv"
   
   echo "[INFO] Generating $count hill climbing ideas..."
-  echo "[DEBUG] Original CSV has $(wc -l < "$FULL_CSV_PATH") lines" >&2
+  local data_rows=$(grep -v '^[[:space:]]*$' "$FULL_CSV_PATH" | tail -n +2 | wc -l)
+  echo "[DEBUG] Original CSV has $data_rows data rows" >&2
   
   # Extract just the IDs from top performers for clarity
   local valid_parent_ids
   valid_parent_ids=$(echo "$top_performers" | cut -d',' -f1 | paste -sd ',' -)
   
-  local prompt="Please add exactly $count parameter tuning ideas directly to the CSV file $temp_csv.
+  # Use relative paths and change to evolution directory so AI can access files
+  local temp_csv_basename=$(basename "$temp_csv")
+  
+  local prompt="I need you to use your file editing capabilities to add exactly $count parameter tuning ideas to the CSV file: $temp_csv_basename
 
 IMPORTANT: You MUST use one of these exact parent IDs: $valid_parent_ids
 
@@ -859,31 +836,40 @@ Successful algorithms to tune:
 $top_performers
 
 CRITICAL INSTRUCTION: Before generating parameter tuning ideas, you MUST read the source code of the parent algorithms.
-Algorithm source files are located at: $FULL_OUTPUT_DIR/evolution_<PARENT_ID>.py
-For example: $FULL_OUTPUT_DIR/evolution_gen01-251.py
+Algorithm source files are located at: evolution_<PARENT_ID>.py
+For example: evolution_gen01-251.py
 
 Instructions:
-1. Read the current CSV file to see existing entries
-2. Add exactly $count new rows to the CSV file
-3. Use the next available generation numbers (gen$CURRENT_GENERATION-XXX format)
-4. For each idea, create a row with: id,parent_id,description,,pending
-5. Each parent_id MUST be one of: $valid_parent_ids
-6. Each description should focus on adjusting specific parameters that exist in the parent's source code
-7. Include current and new parameter values (e.g., \"Lower rsi_entry from 21 to 18\")
+1. Use the Read tool to examine the current CSV file
+2. Find the highest ID number for generation $CURRENT_GENERATION (e.g., if gen$CURRENT_GENERATION-003 exists, next should be gen$CURRENT_GENERATION-004)
+3. If no gen$CURRENT_GENERATION entries exist yet, start with gen$CURRENT_GENERATION-001
+4. Use the Edit or MultiEdit tool to add exactly $count new rows to the CSV file
+5. For each idea, create a row with: id,parent_id,description,,pending
+6. Each parent_id MUST be one of: $valid_parent_ids
+7. Each description should focus on adjusting specific parameters that exist in the parent's source code
+8. Include current and new parameter values (e.g., \"Lower rsi_entry from 21 to 18\")
 
-IMPORTANT: Edit the CSV file directly. Do not return or print CSV data - just modify the file in place."
+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."
 
+  # Change to evolution directory so AI can access files
+  local original_pwd=$(pwd)
+  cd "$FULL_EVOLUTION_DIR"
+  
   # Get AI to directly edit the CSV file
   local ai_response
-  local stderr_file="/tmp/claude-evolve-stderr-$$.txt"
+  local stderr_file="stderr-$$.txt"
   if ! ai_response=$(call_ai_with_fallbacks "$prompt" "$CURRENT_GENERATION" 2>"$stderr_file"); then
     echo "[ERROR] All AI models failed to generate hill climbing ideas" >&2
     cat "$stderr_file" >&2
+    cd "$original_pwd"
     rm -f "$temp_csv" "$stderr_file"
     return 1
   fi
   rm -f "$stderr_file"
   
+  # Restore working directory
+  cd "$original_pwd"
+  
   echo "[DEBUG] AI response: $ai_response" >&2
   
   # Validate that the CSV file was actually modified
@@ -901,18 +887,22 @@ generate_structural_mutation_direct() {
   local count="$1"
   local top_performers="$2"
   
-  # Create temporary CSV copy
-  local temp_csv="/tmp/claude-evolve-temp-csv-$$.csv"
+  # Create temporary CSV copy in evolution directory (so AI can access it)
+  local temp_csv="$FULL_EVOLUTION_DIR/temp-csv-$$.csv"
   cp "$FULL_CSV_PATH" "$temp_csv"
   
   echo "[INFO] Generating $count structural mutation ideas..."
-  echo "[DEBUG] Original CSV has $(wc -l < "$FULL_CSV_PATH") lines" >&2
+  local data_rows=$(grep -v '^[[:space:]]*$' "$FULL_CSV_PATH" | tail -n +2 | wc -l)
+  echo "[DEBUG] Original CSV has $data_rows data rows" >&2
   
   # Extract just the IDs from top performers for clarity
   local valid_parent_ids
   valid_parent_ids=$(echo "$top_performers" | cut -d',' -f1 | paste -sd ',' -)
   
-  local prompt="Please add exactly $count structural modification ideas directly to the CSV file $temp_csv.
+  # Use relative paths and change to evolution directory so AI can access files
+  local temp_csv_basename=$(basename "$temp_csv")
+  
+  local prompt="I need you to use your file editing capabilities to add exactly $count structural modification ideas to the CSV file: $temp_csv_basename
 
 IMPORTANT: You MUST use one of these exact parent IDs: $valid_parent_ids
 
@@ -920,31 +910,40 @@ Successful algorithms to modify structurally:
 $top_performers
 
 CRITICAL INSTRUCTION: Before generating structural modification ideas, you MUST read the source code of the parent algorithms.
-Algorithm source files are located at: $FULL_OUTPUT_DIR/evolution_<PARENT_ID>.py
-For example: $FULL_OUTPUT_DIR/evolution_gen01-251.py
+Algorithm source files are located at: evolution_<PARENT_ID>.py
+For example: evolution_gen01-251.py
 
 Instructions:
-1. Read the current CSV file to see existing entries
-2. Add exactly $count new rows to the CSV file
-3. Use the next available generation numbers (gen$CURRENT_GENERATION-XXX format)
-4. For each idea, create a row with: id,parent_id,description,,pending
-5. Each parent_id MUST be one of: $valid_parent_ids
-6. Each description should focus on architectural/structural changes based on the parent's actual code
-7. Reference actual components/methods found in the source code
+1. Use the Read tool to examine the current CSV file
+2. Find the highest ID number for generation $CURRENT_GENERATION (e.g., if gen$CURRENT_GENERATION-003 exists, next should be gen$CURRENT_GENERATION-004)
+3. If no gen$CURRENT_GENERATION entries exist yet, start with gen$CURRENT_GENERATION-001
+4. Use the Edit or MultiEdit tool to add exactly $count new rows to the CSV file
+5. For each idea, create a row with: id,parent_id,description,,pending
+6. Each parent_id MUST be one of: $valid_parent_ids
+7. Each description should focus on architectural/structural changes based on the parent's actual code
+8. Reference actual components/methods found in the source code
 
-IMPORTANT: Edit the CSV file directly. Do not return or print CSV data - just modify the file in place."
+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."
 
+  # Change to evolution directory so AI can access files
+  local original_pwd=$(pwd)
+  cd "$FULL_EVOLUTION_DIR"
+  
   # Get AI to directly edit the CSV file
   local ai_response
-  local stderr_file="/tmp/claude-evolve-stderr-$$.txt"
+  local stderr_file="stderr-$$.txt"
   if ! ai_response=$(call_ai_with_fallbacks "$prompt" "$CURRENT_GENERATION" 2>"$stderr_file"); then
     echo "[ERROR] All AI models failed to generate structural mutation ideas" >&2
     cat "$stderr_file" >&2
+    cd "$original_pwd"
     rm -f "$temp_csv" "$stderr_file"
     return 1
   fi
   rm -f "$stderr_file"
   
+  # Restore working directory
+  cd "$original_pwd"
+  
   echo "[DEBUG] AI response: $ai_response" >&2
   
   # Validate that the CSV file was actually modified
@@ -962,18 +961,22 @@ generate_crossover_direct() {
   local count="$1"
   local top_performers="$2"
   
-  # Create temporary CSV copy
-  local temp_csv="/tmp/claude-evolve-temp-csv-$$.csv"
+  # Create temporary CSV copy in evolution directory (so AI can access it)
+  local temp_csv="$FULL_EVOLUTION_DIR/temp-csv-$$.csv"
   cp "$FULL_CSV_PATH" "$temp_csv"
   
   echo "[INFO] Generating $count crossover hybrid ideas..."
-  echo "[DEBUG] Original CSV has $(wc -l < "$FULL_CSV_PATH") lines" >&2
+  local data_rows=$(grep -v '^[[:space:]]*$' "$FULL_CSV_PATH" | tail -n +2 | wc -l)
+  echo "[DEBUG] Original CSV has $data_rows data rows" >&2
   
   # Extract just the IDs from top performers for clarity
   local valid_parent_ids
   valid_parent_ids=$(echo "$top_performers" | cut -d',' -f1 | paste -sd ',' -)
   
-  local prompt="Please add exactly $count hybrid combination ideas directly to the CSV file $temp_csv.
+  # Use relative paths and change to evolution directory so AI can access files
+  local temp_csv_basename=$(basename "$temp_csv")
+  
+  local prompt="I need you to use your file editing capabilities to add exactly $count hybrid combination ideas to the CSV file: $temp_csv_basename
 
 IMPORTANT: You MUST use ONLY these exact parent IDs: $valid_parent_ids
 
@@ -981,31 +984,40 @@ Top performers to combine (reference at least 2 in each idea):
 $top_performers
 
 CRITICAL INSTRUCTION: Before generating hybrid combination ideas, you MUST read the source code of the parent algorithms.
-Algorithm source files are located at: $FULL_OUTPUT_DIR/evolution_<PARENT_ID>.py
-For example: $FULL_OUTPUT_DIR/evolution_gen01-251.py
+Algorithm source files are located at: evolution_<PARENT_ID>.py
+For example: evolution_gen01-251.py
 
 Instructions:
-1. Read the current CSV file to see existing entries
-2. Add exactly $count new rows to the CSV file
-3. Use the next available generation numbers (gen$CURRENT_GENERATION-XXX format)
-4. For each idea, create a row with: id,parent_id,description,,pending
-5. Each parent_id MUST be one of: $valid_parent_ids (choose the primary parent)
-6. Each description should combine actual elements from 2+ algorithms based on their source code
-7. Reference specific components/features found in the actual source code
+1. Use the Read tool to examine the current CSV file
+2. Find the highest ID number for generation $CURRENT_GENERATION (e.g., if gen$CURRENT_GENERATION-003 exists, next should be gen$CURRENT_GENERATION-004)
+3. If no gen$CURRENT_GENERATION entries exist yet, start with gen$CURRENT_GENERATION-001
+4. Use the Edit or MultiEdit tool to add exactly $count new rows to the CSV file
+5. For each idea, create a row with: id,parent_id,description,,pending
+6. Each parent_id MUST be one of: $valid_parent_ids (choose the primary parent)
+7. Each description should combine actual elements from 2+ algorithms based on their source code
+8. Reference specific components/features found in the actual source code
 
-IMPORTANT: Edit the CSV file directly. Do not return or print CSV data - just modify the file in place."
+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."
 
+  # Change to evolution directory so AI can access files
+  local original_pwd=$(pwd)
+  cd "$FULL_EVOLUTION_DIR"
+  
   # Get AI to directly edit the CSV file
   local ai_response
-  local stderr_file="/tmp/claude-evolve-stderr-$$.txt"
+  local stderr_file="stderr-$$.txt"
   if ! ai_response=$(call_ai_with_fallbacks "$prompt" "$CURRENT_GENERATION" 2>"$stderr_file"); then
     echo "[ERROR] All AI models failed to generate crossover hybrid ideas" >&2
     cat "$stderr_file" >&2
+    cd "$original_pwd"
     rm -f "$temp_csv" "$stderr_file"
     return 1
   fi
   rm -f "$stderr_file"
   
+  # Restore working directory
+  cd "$original_pwd"
+  
   echo "[DEBUG] AI response: $ai_response" >&2
   
   # Validate that the CSV file was actually modified
@@ -1025,8 +1037,8 @@ ideate_ai_legacy() {
     exit 1
   fi
 
-  # Create temporary CSV copy
-  local temp_csv="/tmp/claude-evolve-temp-csv-$$.csv"
+  # Create temporary CSV copy in evolution directory (so AI can access it)
+  local temp_csv="$FULL_EVOLUTION_DIR/temp-csv-$$.csv"
   cp "$FULL_CSV_PATH" "$temp_csv"
   
   echo "[INFO] Generating $TOTAL_IDEAS ideas (legacy mode)..."
@@ -1040,15 +1052,18 @@ ideate_ai_legacy() {
   fi
 
   # Build prompt for direct CSV modification
-  local prompt="Please add exactly $TOTAL_IDEAS algorithmic ideas directly to the CSV file $temp_csv.
+  # Use relative paths and change to evolution directory so AI can access files
+  local temp_csv_basename=$(basename "$temp_csv")
+  
+  local prompt="I need you to use your file editing capabilities to add exactly $TOTAL_IDEAS algorithmic ideas to the CSV file: $temp_csv_basename
 
 Algorithm files for context:
-- Base algorithm: $FULL_ALGORITHM_PATH  
-- Evolved algorithms: $FULL_OUTPUT_DIR/evolution_*.py
+- Base algorithm: algorithm.py  
+- Evolved algorithms: evolution_*.py
 
 IMPORTANT: Before generating ideas, you should:
-1. Read the base algorithm to understand the codebase structure and possibilities
-2. Read ALL existing evolution_*.py files to see what modifications have been attempted
+1. Use the Read tool to examine the base algorithm to understand the codebase structure and possibilities
+2. Use the Read tool to examine ALL existing evolution_*.py files to see what modifications have been attempted
 3. Consider which approaches might work well
 
 Project Brief:
@@ -1064,12 +1079,13 @@ $top_performers"
   prompt+="
 
 Instructions:
-1. Read the current CSV file to see existing entries
-2. Add exactly $TOTAL_IDEAS new rows to the CSV file
-3. Use the next available generation numbers (gen$CURRENT_GENERATION-XXX format)
-4. For each idea, create a row with: id,parent_id,description,,pending
-5. Mix both parameter tuning and structural changes
-6. If building on existing algorithms, use their ID as parent_id, otherwise leave parent_id empty
+1. Use the Read tool to examine the current CSV file
+2. Find the highest ID number for generation $CURRENT_GENERATION (e.g., if gen$CURRENT_GENERATION-003 exists, next should be gen$CURRENT_GENERATION-004)
+3. If no gen$CURRENT_GENERATION entries exist yet, start with gen$CURRENT_GENERATION-001
+4. Use the Edit or MultiEdit tool to add exactly $TOTAL_IDEAS new rows to the CSV file
+5. For each idea, create a row with: id,parent_id,description,,pending
+6. Mix both parameter tuning and structural changes
+7. 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)
 
@@ -1083,19 +1099,27 @@ Instructions:
 - Time-Based Patterns: Intraday effects, calendar anomalies, volatility timing
 - Parameter Optimization: Entry thresholds, indicator periods, strategy weights
 
-IMPORTANT: Edit the CSV file directly. Do not return or print CSV data - just modify the file in place."
+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."
 
+  # Change to evolution directory so AI can access files
+  local original_pwd=$(pwd)
+  cd "$FULL_EVOLUTION_DIR"
+  
   # Get AI to directly edit the CSV file
   local ai_response
-  local stderr_file="/tmp/claude-evolve-stderr-$$.txt"
+  local stderr_file="stderr-$$.txt"
   if ! ai_response=$(call_ai_with_fallbacks "$prompt" "$CURRENT_GENERATION" 2>"$stderr_file"); then
     echo "[ERROR] All AI models failed to generate ideas" >&2
     cat "$stderr_file" >&2
+    cd "$original_pwd"
     rm -f "$temp_csv" "$stderr_file"
     return 1
   fi
   rm -f "$stderr_file"
   
+  # Restore working directory
+  cd "$original_pwd"
+  
   echo "[DEBUG] AI response: $ai_response" >&2
   
   # Validate that the CSV file was actually modified

package/bin/claude-evolve-worker

@@ -79,6 +79,244 @@ else
   load_config
 fi
 
+# Call an AI model with a prompt - handles model-specific invocation
+call_ai_model() {
+  local model="$1"
+  local prompt="$2"
+  local ai_output
+  local ai_exit_code
+  
+  case "$model" in
+    "claude")
+      # Pass prompt as argument, not via stdin
+      ai_output=$(timeout 300 claude --dangerously-skip-permissions -p "$prompt" 2>&1)
+      ai_exit_code=$?
+      ;;
+      
+    "gemini")
+      # Pass prompt as argument, not via stdin
+      ai_output=$(timeout 300 gemini -y -p "$prompt" 2>&1)
+      ai_exit_code=$?
+      ;;
+      
+    "codex")
+      # Pass prompt as argument, not via stdin
+      ai_output=$(timeout 300 codex exec --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      ai_exit_code=$?
+      ;;
+      
+    *)
+      echo "[WORKER-$$] ERROR: Unknown AI model: $model" >&2
+      return 1
+      ;;
+  esac
+  
+  # Return output via stdout
+  echo "$ai_output"
+  return $ai_exit_code
+}
+
+# Check if AI output indicates a usage limit was hit
+is_usage_limit_error() {
+  local output="$1"
+  local model="$2"
+  
+  case "$model" in
+    "claude")
+      echo "$output" | grep -q "Claude AI usage limit reached"
+      ;;
+    "gemini")
+      echo "$output" | grep -q "Quota exceeded.*Gemini"
+      ;;
+    "codex")
+      # Add codex-specific limit patterns if they exist
+      false
+      ;;
+    *)
+      false
+      ;;
+  esac
+}
+
+# Validate if AI output is successful
+is_valid_ai_output() {
+  local output="$1"
+  local model="$2"
+  local exit_code="$3"
+  
+  # First check exit code
+  [[ $exit_code -ne 0 ]] && return 1
+  
+  # Model-specific validation
+  case "$model" in
+    "claude")
+      # Claude is straightforward - exit code 0 means success
+      return 0
+      ;;
+      
+    "gemini")
+      # Gemini needs extra validation for auth messages
+      if echo "$output" | grep -q "Attempting to authenticate\|Authenticating\|Loading\|Initializing"; then
+        return 1
+      fi
+      # Also check for minimal output
+      if [[ -z "$output" ]] || [[ $(echo "$output" | wc -l) -lt 2 ]]; then
+        return 1
+      fi
+      return 0
+      ;;
+      
+    "codex")
+      # Codex might return JSON that needs extraction
+      if echo "$output" | grep -q '"content"'; then
+        # Will be cleaned later, just check it's not an error
+        if echo "$output" | grep -q "error\|failed\|exception"; then
+          return 1
+        fi
+      fi
+      [[ -n "$output" ]]
+      ;;
+      
+    *)
+      return 1
+      ;;
+  esac
+}
+
+# Clean AI output if needed (e.g., extract from JSON)
+clean_ai_output() {
+  local output="$1"
+  local model="$2"
+  
+  case "$model" in
+    "codex")
+      # Clean codex output - extract content between "codex" marker and "tokens used"
+      if echo "$output" | grep -q "^\[.*\] codex$"; then
+        # Extract content between "codex" line and "tokens used" line
+        output=$(echo "$output" | awk '/\] codex$/{flag=1;next}/\] tokens used/{flag=0}flag')
+      elif echo "$output" | grep -q '"content"'; then
+        # Old JSON format
+        output=$(echo "$output" | python3 -c "
+import sys
+import json
+try:
+    data = json.load(sys.stdin)
+    if 'content' in data:
+        print(data['content'])
+    elif 'response' in data:
+        print(data['response'])
+    elif 'text' in data:
+        print(data['text'])
+    else:
+        print(json.dumps(data))
+except:
+    print(sys.stdin.read())
+" 2>/dev/null || echo "$output")
+      fi
+      # Trim whitespace
+      output=$(echo "$output" | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')
+      ;;
+  esac
+  
+  echo "$output"
+}
+
+# AI round-robin with fallback function for code evolution
+call_ai_for_evolution() {
+  local prompt="$1"
+  local candidate_id="$2"
+  
+  # Extract generation and ID numbers for round-robin calculation
+  local gen_num=0
+  local id_num=0
+  if [[ $candidate_id =~ ^gen([0-9]+)-([0-9]+)$ ]]; then
+    gen_num=$((10#${BASH_REMATCH[1]}))
+    id_num=$((10#${BASH_REMATCH[2]}))
+  fi
+  
+  # Calculate hash for round-robin (combine generation and ID)
+  local hash_value=$((gen_num * 1000 + id_num))
+  
+  # Check which AI tools are available
+  local available_models=()
+  if command -v claude >/dev/null 2>&1; then
+    available_models+=("claude")
+  fi
+  if command -v gemini >/dev/null 2>&1; then
+    available_models+=("gemini")
+  fi
+  if command -v codex >/dev/null 2>&1; then
+    available_models+=("codex")
+  fi
+  
+  if [[ ${#available_models[@]} -eq 0 ]]; then
+    echo "[WORKER-$$] ERROR: No AI models available!" >&2
+    return 1
+  fi
+  
+  # Create ordered list based on round-robin for this candidate
+  local num_models=${#available_models[@]}
+  local start_index=$((hash_value % num_models))
+  local models=()
+  
+  # Add models in round-robin order starting from the calculated index
+  for ((i=0; i<num_models; i++)); do
+    local idx=$(((start_index + i) % num_models))
+    models+=("${available_models[$idx]}")
+  done
+  
+  echo "[WORKER-$$] Model order for $candidate_id (round-robin): ${models[*]}" >&2
+  
+  # Track if any model hit usage limits
+  local hit_usage_limit=false
+  local limited_models=()
+  
+  # Try each model in the ordered sequence
+  for model in "${models[@]}"; do
+    echo "[WORKER-$$] Attempting code evolution with $model" >&2
+    
+    # Call the AI model
+    local ai_output
+    ai_output=$(call_ai_model "$model" "$prompt")
+    local ai_exit_code=$?
+    
+    # Check for usage limits
+    if is_usage_limit_error "$ai_output" "$model"; then
+      echo "[WORKER-$$] $model hit usage limit - trying next model" >&2
+      hit_usage_limit=true
+      limited_models+=("$model")
+      continue
+    fi
+    
+    # Validate output
+    if is_valid_ai_output "$ai_output" "$model" "$ai_exit_code"; then
+      # Clean output if needed
+      ai_output=$(clean_ai_output "$ai_output" "$model")
+      echo "[WORKER-$$] $model succeeded" >&2
+      # Output the cleaned result for the worker to use
+      echo "$ai_output"
+      return 0
+    fi
+    
+    echo "[WORKER-$$] $model failed (exit code $ai_exit_code), trying next model..." >&2
+    if [[ -n "$ai_output" ]]; then
+      echo "[WORKER-$$] $model error: $(echo "$ai_output" | head -5)" >&2
+    fi
+  done
+  
+  # All models have been tried
+  echo "[WORKER-$$] All AI models failed for code evolution" >&2
+  
+  # If any model hit usage limits and we couldn't complete the task
+  if [[ "$hit_usage_limit" == "true" ]]; then
+    echo "[WORKER-$$] Models hit usage limits: ${limited_models[*]}" >&2
+    echo "[WORKER-$$] Unable to complete evolution due to API limits" >&2
+    exit 3
+  fi
+  
+  return 1
+}
+
 # Validate paths
 if [[ ! -f "$FULL_CSV_PATH" ]]; then
   echo "[WORKER-$$] CSV file not found: $FULL_CSV_PATH" >&2
@@ -153,42 +391,40 @@ with EvolutionCSV('$FULL_CSV_PATH') as csv:
     echo "[WORKER-$$] Copying $source_file to $target_file"
     cp "$source_file" "$target_file"
     
-    # Apply evolution using Claude
-    echo "[WORKER-$$] Applying evolution with Claude..."
-    local evolution_prompt="Modify the algorithm in $target_file based on this description: $description
+    # Apply evolution using AI
+    echo "[WORKER-$$] Applying evolution..."
+    
+    # Use relative path for AI prompt
+    local target_basename=$(basename "$target_file")
+    local evolution_prompt="Modify the algorithm in $target_basename based on this description: $description
     
 The modification should be substantial and follow the description exactly. Make sure the algorithm still follows all interface requirements and can run properly.
 
 Important: Make meaningful changes that match the description. Don't just add comments or make trivial adjustments."
     
     if [[ "$is_baseline" != "true" ]]; then
-      local claude_output
-      claude_output=$(echo "$evolution_prompt" | claude --dangerously-skip-permissions -p 2>&1)
-      local claude_exit_code=$?
+      # Change to evolution directory so AI can access files
+      local original_pwd=$(pwd)
+      cd "$FULL_EVOLUTION_DIR"
       
-      # Check for usage limits
-      if echo "$claude_output" | grep -q "Claude AI usage limit reached"; then
-        echo "[WORKER-$$] ERROR: Claude AI usage limit reached!" >&2
-        echo "[WORKER-$$] ERROR: Cannot continue processing - API limits exceeded" >&2
-        rm -f "$target_file"  # Clean up on failure
-        # Exit with special code 3 to indicate API limit
-        exit 3
-      fi
-      
-      if [[ $claude_exit_code -ne 0 ]]; then
-        echo "[WORKER-$$] ERROR: Claude evolution failed with exit code $claude_exit_code" >&2
-        echo "[WORKER-$$] ERROR: Claude output: $claude_output" >&2
+      # Try AI models with round-robin based on candidate ID
+      if ! call_ai_for_evolution "$evolution_prompt" "$candidate_id"; then
+        echo "[WORKER-$$] ERROR: All AI models failed to generate code" >&2
+        cd "$original_pwd"
         rm -f "$target_file"  # Clean up on failure
         return 1
       fi
       
+      # Restore working directory
+      cd "$original_pwd"
+      
       echo "[WORKER-$$] Evolution applied successfully"
     fi
   fi
   
   # Run evaluation
   echo "[WORKER-$$] Evaluating algorithm..."
-  local eval_output_file="/tmp/claude-evolve-eval-$$-$candidate_id.out"
+  local eval_output_file="$FULL_EVOLUTION_DIR/temp-eval-$$-$candidate_id.out"
   local eval_start=$(date +%s)
   
   # Prepare evaluation command

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-evolve",
-  "version": "1.4.13",
+  "version": "1.5.1",
   "bin": {
     "claude-evolve": "./bin/claude-evolve",
     "claude-evolve-main": "./bin/claude-evolve-main",