claude-evolve 1.3.40 → 1.3.42

package/bin/claude-evolve-worker

@@ -18,6 +18,59 @@ cleanup_temp() {
 # Set trap to clean up temp files on exit
 trap cleanup_temp EXIT INT TERM
 
+# Function to handle failures with retry logic
+handle_failure() {
+  local candidate_id="$1"
+  local current_status="$2"
+  local performance="${3:-0}"
+  
+  # If this is already a retry, increment the retry count
+  if [[ $current_status =~ ^failed-retry([0-9]+)$ ]]; then
+    local retry_num=${BASH_REMATCH[1]}
+    local new_retry_num=$((retry_num + 1))
+    
+    if [[ $new_retry_num -le $MAX_RETRIES ]]; then
+      local new_status="failed-retry${new_retry_num}"
+      update_csv_row_with_lock "$candidate_id" "status" "$new_status"
+      update_csv_row_with_lock "$candidate_id" "performance" "$performance"
+      echo "[WORKER-$$] ✗ Retry $retry_num failed, marked as $new_status"
+      exit 1
+    else
+      update_csv_row_with_lock "$candidate_id" "status" "failed"
+      update_csv_row_with_lock "$candidate_id" "performance" "$performance"
+      echo "[WORKER-$$] ✗ Max retries ($MAX_RETRIES) exceeded, marking as permanently failed"
+      exit 1
+    fi
+  elif [[ $current_status == "failed" ]]; then
+    # Initial failure, convert to retry1 if retries are enabled
+    if [[ $MAX_RETRIES -gt 0 ]]; then
+      update_csv_row_with_lock "$candidate_id" "status" "failed-retry1"
+      update_csv_row_with_lock "$candidate_id" "performance" "$performance"
+      echo "[WORKER-$$] ✗ Initial failure, marked as failed-retry1 for retry"
+      exit 1
+    else
+      update_csv_row_with_lock "$candidate_id" "status" "failed"
+      update_csv_row_with_lock "$candidate_id" "performance" "$performance"
+      echo "[WORKER-$$] ✗ Failed (retries disabled)"
+      # Use exit code 1 - systemic issue since retries are disabled
+      exit 1
+    fi
+  else
+    # Not a failure scenario, convert to retry1 if retries enabled
+    if [[ $MAX_RETRIES -gt 0 ]]; then
+      update_csv_row_with_lock "$candidate_id" "status" "failed-retry1"
+      update_csv_row_with_lock "$candidate_id" "performance" "$performance"
+      echo "[WORKER-$$] ✗ Evaluation failed, marked as failed-retry1 for retry"
+      exit 1
+    else
+      update_csv_row_with_lock "$candidate_id" "status" "failed"
+      update_csv_row_with_lock "$candidate_id" "performance" "$performance"
+      echo "[WORKER-$$] ✗ Evaluation failed (retries disabled)"
+      exit 1
+    fi
+  fi
+}
+
 # Load configuration
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 # shellcheck source=../lib/config.sh
@@ -59,14 +112,25 @@ done
 
 # If no ID provided, find next pending
 if [[ -z $candidate_id ]]; then
-  candidate_id=$(find_next_pending_with_lock)
-  if [[ -z $candidate_id ]]; then
+  candidate_result=$(find_next_pending_with_lock)
+  if [[ -z $candidate_result ]]; then
     echo "[DEBUG] No pending candidates found" >&2
     exit 0
   fi
+  
+  # Parse candidate_id|original_status format
+  if [[ $candidate_result == *"|"* ]]; then
+    candidate_id="${candidate_result%|*}"  # Everything before |
+    original_candidate_status="${candidate_result#*|}"  # Everything after |
+  else
+    # Fallback for old format (shouldn't happen)
+    candidate_id="$candidate_result"
+    original_candidate_status=""
+  fi
 else
   # Mark specified candidate as running
   update_csv_row_with_lock "$candidate_id" "status" "running"
+  original_candidate_status=""  # Unknown for manually specified candidates
 fi
 
 echo "[WORKER-$$] Processing candidate ID: $candidate_id"
@@ -124,6 +188,23 @@ fi
 echo "[WORKER-$$] Description: $description"
 echo "[WORKER-$$] Based on ID: $based_on_id"
 
+# AIDEV-NOTE: Retry logic - detect if this is a retry attempt
+is_retry=false
+retry_count=0
+# Use original_candidate_status for retry detection (if available), otherwise fall back to CSV status
+retry_status="$original_candidate_status"
+if [[ -z "$retry_status" ]]; then
+  retry_status="$status"
+fi
+
+if [[ $retry_status =~ ^failed-retry([0-9]+)$ ]]; then
+  is_retry=true
+  retry_count=${BASH_REMATCH[1]}
+  echo "[WORKER-$$] 🔄 Processing retry attempt #$retry_count"
+elif [[ $retry_status == "failed" ]]; then
+  echo "[WORKER-$$] ⚠️ Initial failure detected - this should be converted to failed-retry1 to enable retries"
+fi
+
 # AIDEV-NOTE: Using common evolution processor logic for consistent handling
 # Determine parent algorithm
 if [[ -z $based_on_id || $based_on_id == "0" || $based_on_id == '""' ]]; then
@@ -139,7 +220,7 @@ else
   
   if [[ ! -f $parent_file ]]; then
     echo "[ERROR] Parent algorithm not found: $parent_file" >&2
-    update_csv_row_with_lock "$candidate_id" "status" "failed"
+    handle_failure "$candidate_id" "$retry_status" "0"
     exit 1
   fi
 fi
@@ -155,11 +236,28 @@ fi
 temp_file="${output_file}.tmp$$"
 
 # Check if processing should be skipped using common logic
+# Set environment variable for retry detection
+if [[ $is_retry == "true" ]]; then
+  export RETRY_CANDIDATE=true
+else
+  export RETRY_CANDIDATE=false
+fi
+
 eval "$("$PYTHON_CMD" "$SCRIPT_DIR/../lib/evolution_processor.py" "$id" "$based_on_id" "$FULL_OUTPUT_DIR" "$ROOT_DIR" "$parent_file" "$output_file")"
 
 # Handle copy operation to temp file
 if [[ "$skip_copy" == "True" ]]; then
   echo "[WORKER-$$] ⚠️  Skipping copy - $reason"
+elif [[ $is_retry == "true" ]]; then
+  # For retries, edit the existing failed algorithm in-place
+  if [[ -f "$output_file" ]]; then
+    cp "$output_file" "$temp_file"
+    echo "[WORKER-$$] 🔄 Copied existing algorithm for retry: $temp_file"
+  else
+    # Fallback to parent if existing file doesn't exist
+    cp "$parent_file" "$temp_file"
+    echo "[WORKER-$$] ⚠️ Existing algorithm not found, using parent: $temp_file"
+  fi
 else
   cp "$parent_file" "$temp_file"
   echo "[WORKER-$$] Copied parent to temp file: $temp_file"
@@ -178,15 +276,28 @@ else
   claude_cmd="${CLAUDE_CMD:-claude}"
   if ! command -v "$claude_cmd" >/dev/null 2>&1; then
     echo "[ERROR] Claude CLI not found" >&2
-    update_csv_row_with_lock "$candidate_id" "status" "failed"
+    handle_failure "$candidate_id" "$retry_status" "0"
     exit 1
   fi
   
   CLAUDE_MODEL="sonnet"
   echo "[WORKER-$$] Using Claude $CLAUDE_MODEL for mutation"
   
-  # Create mutation prompt
-  prompt="Edit the file $temp_file to implement this specific change: $description
+  # Create mutation prompt (different for retries vs initial attempts)
+  if [[ $is_retry == "true" ]]; then
+    prompt="Fix the bugs in the file $temp_file. This algorithm was attempting to implement: $description
+
+The algorithm failed during evaluation. Please:
+- Analyze the code for potential bugs (syntax errors, logical issues, missing imports, etc.)
+- Fix any problems you find
+- Ensure the code runs without errors
+- Make sure it still implements the intended change: $description
+- Add appropriate error handling and validation
+- If possible, suggest a simple way to test this fix
+
+This is retry attempt #$retry_count. Focus on making the code robust and correct."
+  else
+    prompt="Edit the file $temp_file to implement this specific change: $description
 
 Requirements:
 - Edit the file directly (don't just provide comments or suggestions)
@@ -196,6 +307,7 @@ Requirements:
 - Add proper error handling if needed
 
 The file currently contains the parent algorithm. Modify it according to the description above."
+  fi
   
   # Log prompt
   {
@@ -213,7 +325,9 @@ The file currently contains the parent algorithm. Modify it according to the des
   
   # Check for rate limit (multiple possible messages)
   if echo "$claude_output" | grep -q -E "(usage limit|rate limit|limit reached|too many requests)"; then
-    echo "[ERROR] Claude API rate limit reached" >&2
+    echo "⚠️  Claude API rate limit reached" >&2
+    echo "⚠️  Claude output:" >&2
+    echo "$claude_output" >&2
     # Clean up the temp file
     if [[ -f "$temp_file" ]]; then
       rm "$temp_file"
@@ -225,30 +339,29 @@ The file currently contains the parent algorithm. Modify it according to the des
   fi
   
   if [[ $claude_exit_code -ne 0 ]]; then
-    echo "[ERROR] Claude failed to mutate algorithm" >&2
+    echo "⚠️  Claude failed to mutate algorithm (exit code: $claude_exit_code)" >&2
+    echo "⚠️  Claude output:" >&2
+    echo "$claude_output" >&2
     # Clean up the temp file
     if [[ -f "$temp_file" ]]; then
       rm "$temp_file"
       echo "[WORKER-$$] Cleaned up temp file due to Claude failure" >&2
     fi
-    update_csv_row_with_lock "$candidate_id" "status" "failed"
+    handle_failure "$candidate_id" "$retry_status" "0"
     exit 1
   fi
   
   # Verify that Claude actually modified the file
   if [[ -f "$temp_file" && -f "$parent_file" ]]; then
     if cmp -s "$temp_file" "$parent_file"; then
-      echo "" >&2
-      echo "🚨🚨🚨 RED ALERT: UNCHANGED ALGORITHM DETECTED 🚨🚨🚨" >&2
-      echo "ERROR: Temp file is IDENTICAL to parent algorithm!" >&2
-      echo "ERROR: Claude failed to make any changes" >&2
-      echo "ERROR: Marking as failed - no evaluation will run" >&2
-      echo "🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨" >&2
-      echo "" >&2
+      echo "⚠️  Unchanged algorithm detected - Claude didn't modify the file" >&2
+      echo "⚠️  Description was: $description" >&2
+      echo "⚠️  Claude's response:" >&2
+      echo "$claude_output" >&2
       
       # Clean up temp file and mark as failed
       rm "$temp_file"
-      update_csv_row_with_lock "$candidate_id" "status" "failed"
+      handle_failure "$candidate_id" "$retry_status" "0"
       exit 1
     else
       # Changes were made - move temp file to final location
@@ -269,28 +382,41 @@ echo "[WORKER-$$] Running evaluation..."
 eval_output=""
 eval_exit_code=0
 
+# Stream evaluator output in real-time while capturing it
+eval_tempfile=$(mktemp)
 if [[ -n $timeout_seconds ]]; then
   echo "[WORKER-$$] Evaluation timeout: ${timeout_seconds}s"
-  # For Modal compatibility, don't capture stderr
-  if eval_output=$(EXPERIMENT_ID="$id" timeout "$timeout_seconds" "$PYTHON_CMD" "$FULL_EVALUATOR_PATH" "$output_file"); then
-    eval_exit_code=0
+  # Stream output to both log and temp file
+  if EXPERIMENT_ID="$id" timeout "$timeout_seconds" "$PYTHON_CMD" "$FULL_EVALUATOR_PATH" "$output_file" 2>&1 | tee "$eval_tempfile" | while IFS= read -r line; do
+    echo "[EVALUATOR] $line" >> "$LOGFILE"
+    echo "[EVALUATOR] $line" >&2
+  done; then
+    eval_exit_code=${PIPESTATUS[0]}
   else
     eval_exit_code=$?
     if [[ $eval_exit_code -eq 124 ]]; then
       echo "[ERROR] Evaluation timed out" >&2
       update_csv_row_with_lock "$candidate_id" "status" "timeout"
+      rm -f "$eval_tempfile"
       exit 1
     fi
   fi
 else
-  # For Modal compatibility, don't capture stderr
-  if eval_output=$(EXPERIMENT_ID="$id" "$PYTHON_CMD" "$FULL_EVALUATOR_PATH" "$output_file"); then
-    eval_exit_code=0
+  # Stream output to both log and temp file
+  if EXPERIMENT_ID="$id" "$PYTHON_CMD" "$FULL_EVALUATOR_PATH" "$output_file" 2>&1 | tee "$eval_tempfile" | while IFS= read -r line; do
+    echo "[EVALUATOR] $line" >> "$LOGFILE"
+    echo "[EVALUATOR] $line" >&2
+  done; then
+    eval_exit_code=${PIPESTATUS[0]}
   else
     eval_exit_code=$?
   fi
 fi
 
+# Read the complete output from temp file
+eval_output=$(cat "$eval_tempfile")
+rm -f "$eval_tempfile"
+
 # Log evaluator output
 {
   echo "=== WORKER $$ - EVALUATOR OUTPUT ==="
@@ -305,9 +431,7 @@ if [[ $eval_exit_code -eq 0 ]]; then
   if [[ $eval_output =~ ^[[:space:]]*-?[0-9]+\.?[0-9]*[[:space:]]*$ ]]; then
     score=$(echo "$eval_output" | tr -d ' ')
     if [[ $(echo "$score == 0" | bc -l) == "1" ]]; then
-      update_csv_row_with_lock "$candidate_id" "status" "failed"
-      update_csv_row_with_lock "$candidate_id" "performance" "$score"
-      echo "[WORKER-$$] ✗ Evaluation failed with score 0"
+      handle_failure "$candidate_id" "$retry_status" "$score"
       exit 1
     else
       update_csv_row_with_lock "$candidate_id" "performance" "$score"
@@ -347,9 +471,7 @@ if [[ $eval_exit_code -eq 0 ]]; then
   if score=$(echo "$eval_output" | grep -o '"score"[[:space:]]*:[[:space:]]*[0-9.]*' | cut -d: -f2 | tr -d ' '); then
     if [[ -n $score ]]; then
       if [[ $(echo "$score == 0" | bc -l) == "1" ]]; then
-        update_csv_row_with_lock "$candidate_id" "status" "failed"
-        update_csv_row_with_lock "$candidate_id" "performance" "$score"
-        echo "[WORKER-$$] ✗ Evaluation failed with score 0"
+        handle_failure "$candidate_id" "$retry_status" "$score"
         exit 1
       else
         update_csv_row_with_lock "$candidate_id" "performance" "$score"
@@ -364,9 +486,7 @@ if [[ $eval_exit_code -eq 0 ]]; then
   if score=$(echo "$eval_output" | grep -o '"performance"[[:space:]]*:[[:space:]]*[0-9.]*' | cut -d: -f2 | tr -d ' '); then
     if [[ -n $score ]]; then
       if [[ $(echo "$score == 0" | bc -l) == "1" ]]; then
-        update_csv_row_with_lock "$candidate_id" "status" "failed"
-        update_csv_row_with_lock "$candidate_id" "performance" "$score"
-        echo "[WORKER-$$] ✗ Evaluation failed with score 0"
+        handle_failure "$candidate_id" "$retry_status" "$score"
         exit 1
       else
         update_csv_row_with_lock "$candidate_id" "performance" "$score"
@@ -381,10 +501,10 @@ if [[ $eval_exit_code -eq 0 ]]; then
   echo "[ERROR] Expected: plain number (e.g., 1.23) or JSON with 'score' or 'performance' field" >&2
   echo "[ERROR] Actual evaluator output was:" >&2
   echo "$eval_output" >&2
-  update_csv_row_with_lock "$candidate_id" "status" "failed"
+  handle_failure "$candidate_id" "$retry_status" "0"
   exit 1
 else
   echo "[ERROR] Evaluator failed with exit code $eval_exit_code" >&2
-  update_csv_row_with_lock "$candidate_id" "status" "failed"
+  handle_failure "$candidate_id" "$retry_status" "0"
   exit 1
 fi

package/lib/config.sh

@@ -45,6 +45,9 @@ DEFAULT_LOCK_TIMEOUT=10
 # Default auto ideation value
 DEFAULT_AUTO_IDEATE=true
 
+# Default retry value
+DEFAULT_MAX_RETRIES=3
+
 # Load configuration from config file
 load_config() {
   # Accept config file path as parameter
@@ -76,6 +79,9 @@ load_config() {
   # Set auto ideation default
   AUTO_IDEATE="$DEFAULT_AUTO_IDEATE"
   
+  # Set retry default
+  MAX_RETRIES="$DEFAULT_MAX_RETRIES"
+  
   # Load config if found
   if [[ -f "$config_file" ]]; then
     echo "[DEBUG] Loading configuration from: $config_file" >&2
@@ -151,6 +157,7 @@ load_config() {
           parent_selection) PARENT_SELECTION="$value" ;;
           python_cmd) PYTHON_CMD="$value" ;;
           auto_ideate) AUTO_IDEATE="$value" ;;
+          max_retries) MAX_RETRIES="$value" ;;
           evolution_dir) 
             echo "[WARN] evolution_dir in config is ignored - automatically inferred from config file location" >&2
             ;;
@@ -245,4 +252,5 @@ show_config() {
   echo "  Max workers: $MAX_WORKERS"
   echo "  Lock timeout: $LOCK_TIMEOUT"
   echo "  Auto ideate: $AUTO_IDEATE"
+  echo "  Max retries: $MAX_RETRIES"
 }

package/lib/csv-lock.sh

@@ -230,29 +230,39 @@ find_next_pending_with_lock() {
         return 1
     fi
     
-    # Find oldest pending candidate and update to running using Python
+    # Find oldest pending candidate (including retries) and update to running using Python
     local candidate=$("$PYTHON_CMD" -c "
 import csv
 import sys
+import re
+
+def is_pending_retry(status):
+    '''Check if status is pending (empty, pending, or retry status).'''
+    if not status or status == 'pending':
+        return True
+    return status.startswith('failed-retry')
 
 # Read CSV
 with open('$csv_file', 'r') as f:
     reader = csv.reader(f)
     rows = list(reader)
 
-# Find first pending candidate
+# Find first pending candidate (including retries)
 candidate_id = None
+original_status = None
 for i in range(1, len(rows)):
     # If row has fewer than 5 fields, it's pending
     if len(rows[i]) < 5:
         candidate_id = rows[i][0]
+        original_status = ''  # Empty status means pending
         # Ensure row has 5 fields before setting status
         while len(rows[i]) < 5:
             rows[i].append('')
         rows[i][4] = 'running'  # Update status
         break
-    elif len(rows[i]) >= 5 and (rows[i][4] == 'pending' or rows[i][4] == ''):
+    elif len(rows[i]) >= 5 and is_pending_retry(rows[i][4]):
         candidate_id = rows[i][0]
+        original_status = rows[i][4]  # Save original status before overwriting
         rows[i][4] = 'running'  # Update status
         break
 
@@ -261,7 +271,7 @@ if candidate_id:
     with open('${csv_file}.tmp', 'w', newline='') as f:
         writer = csv.writer(f)
         writer.writerows(rows)
-    print(candidate_id)
+    print(f'{candidate_id}|{original_status}')  # Return both ID and original status
 ")
     
     if [ -n "$candidate" ]; then

package/lib/csv_helper.py

@@ -8,6 +8,7 @@ import csv
 import json
 import sys
 import os
+import re
 from typing import Dict, List, Any
 
 
@@ -50,6 +51,40 @@ def ensure_columns(headers: list[str], rows: list[list[str]], new_fields: dict)
     return headers, rows
 
 
+def parse_retry_status(status: str) -> tuple[str, int]:
+    """Parse retry status and return (base_status, retry_count).
+    
+    Examples:
+        'failed' -> ('failed', 0)
+        'failed-retry1' -> ('failed', 1)
+        'failed-retry3' -> ('failed', 3)
+        'complete' -> ('complete', 0)
+    """
+    if not status:
+        return ('', 0)
+    
+    match = re.match(r'^(.*)-retry(\d+)$', status)
+    if match:
+        base_status = match.group(1)
+        retry_count = int(match.group(2))
+        return (base_status, retry_count)
+    else:
+        return (status, 0)
+
+
+def is_retry_candidate(status: str) -> bool:
+    """Check if a status represents a retry candidate."""
+    base_status, _ = parse_retry_status(status)
+    return base_status == 'failed' and status.startswith('failed-retry')
+
+
+def is_pending_retry(status: str) -> bool:
+    """Check if status is pending (empty, 'pending', or retry status)."""
+    if not status or status == 'pending':
+        return True
+    return is_retry_candidate(status)
+
+
 def update_row_with_fields(headers: list[str], rows: list[list[str]], target_id: str, fields: dict):
     """Update a specific row with multiple fields."""
     # Find column indices
@@ -162,9 +197,12 @@ def main():
         try:
             headers, rows = read_csv(csv_file)
             
-            # Find first row with empty status or status == "pending"
+            # Find first row with empty status, "pending", or retry status
             for i, row in enumerate(rows, start=2):  # Start at 2 (1-indexed, skip header)
-                if len(row) < 5 or row[4] == '' or row[4] == 'pending':
+                if len(row) < 5:
+                    print(i)
+                    sys.exit(0)
+                elif len(row) >= 5 and is_pending_retry(row[4]):
                     print(i)
                     sys.exit(0)
             

package/lib/evolution_processor.py

@@ -14,7 +14,8 @@ def should_skip_processing(id_val, based_on_id, parent_file, output_file):
     """
     Determine if evolution processing should be skipped.
     
-    Simple rule: If file exists, skip everything. This handles all edge cases cleanly.
+    Simple rule: If file exists, skip everything UNLESS this is a retry candidate.
+    For retry candidates, we want Claude to process the existing file to fix bugs.
     
     Returns tuple: (skip_copy, skip_claude, reason)
     """
@@ -23,9 +24,15 @@ def should_skip_processing(id_val, based_on_id, parent_file, output_file):
         return True, True, "Baseline algorithm - no processing needed"
     
     # File existence check - if file exists, skip both copy and Claude
-    # This automatically handles self-parent cases and re-runs
+    # EXCEPT for retry candidates which need Claude to fix the existing file
     if os.path.exists(output_file):
-        return True, True, "File already exists - skipping all processing"
+        # Check if this might be a retry candidate by looking for retry status in environment
+        # The worker sets RETRY_CANDIDATE=true for retry processing
+        retry_env = os.environ.get('RETRY_CANDIDATE')
+        if retry_env == 'true':
+            return True, False, "Retry candidate - skip copy but run Claude for bug fixing"
+        else:
+            return True, True, "File already exists - skipping all processing"
     
     # File doesn't exist - proceed with copy and Claude
     return False, False, None

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-evolve",
-  "version": "1.3.40",
+  "version": "1.3.42",
   "bin": {
     "claude-evolve": "./bin/claude-evolve",
     "claude-evolve-main": "./bin/claude-evolve-main",

package/templates/config.yaml

@@ -42,6 +42,10 @@ python_cmd: "python3"
 # When true, automatically generate new ideas when no pending candidates remain
 auto_ideate: true
 
+# Retry configuration
+# Maximum number of retries for failed candidates before marking as permanently failed
+max_retries: 3
+
 # Parallel execution configuration
 parallel:
   # Enable parallel execution of evolution candidates