openclaw/tools/jarvis-error-daemon.sh

#!/usr/bin/env bash
# Jarvis Error Daemon - Kontinuerlig overvågning af agent sessions med auto-healing
# Inspireret af Reddit self-healing loop mellem Claude Code og OpenClaw

set -euo pipefail

REPO_ROOT="/home/alex/clawd"
LOG_DIR="/home/alex/clawd/logs"
ERROR_LOG="$LOG_DIR/jarvis-errors.log"
HEALING_LOG="$LOG_DIR/jarvis-healing.log"
SESSION_DIR="$HOME/.openclaw/agents/free/sessions"
CHECK_INTERVAL=30  # Tjek hver 30 sekunder
MAX_RETRIES=3

# Healing agent configuration (Claude Sonnet for troubleshooting)
HEALING_AGENT="sonnet"
HEALING_TIMEOUT=300  # 5 min for troubleshooting

mkdir -p "$LOG_DIR"

log() {
    echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*" | tee -a "$ERROR_LOG"
}

healing_log() {
    echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*" | tee -a "$HEALING_LOG"
}

# Extract error bundle from session log
extract_error_bundle() {
    local session_file="$1"
    
    # Find last error/failure in session
    local error_context=$(tail -100 "$session_file" | grep -A 10 -B 10 -E "(Error:|Exception:|failed|ELIFECYCLE|Command failed)" || echo "")
    
    if [[ -z "$error_context" ]]; then
        return 1
    fi
    
    # Extract task context (first 50 lines of session)
    local task_context=$(head -50 "$session_file")
    
    # Build error bundle
    cat <<EOF
## Error Bundle

### Task Context
$task_context

### Error Details
$error_context

### Full Session Path
$session_file
EOF
}

# Spawn healing agent to diagnose and fix
heal_error() {
    local session_file="$1"
    local task_id="$2"
    local retry_count="$3"
    
    healing_log "🔧 Healing session: $session_file (retry $retry_count/$MAX_RETRIES)"
    
    # Extract error bundle
    local error_bundle=$(extract_error_bundle "$session_file")
    if [[ -z "$error_bundle" ]]; then
        healing_log "⚠️  No clear error found in session"
        return 1
    fi
    
    # Build healing prompt
    local healing_prompt="**Self-Healing Error Analysis**

You are debugging an OpenClaw agent task that failed. Analyze the error bundle below and:

1. **Diagnose** the root cause
2. **Propose** a fix (code change, dependency, config, or retry strategy)
3. **Apply** the fix if possible (edit files, install deps, etc.)
4. **Verify** the fix would work

$error_bundle

**Your goal:** Make this task succeed. Be specific about what needs to change."
    
    # Spawn healing agent
    healing_log "🤖 Spawning healing agent ($HEALING_AGENT)..."
    local healing_result=$(cd "$REPO_ROOT" && pnpm -s openclaw agent \
        --local \
        --json \
        --agent "$HEALING_AGENT" \
        --timeout "$HEALING_TIMEOUT" \
        --message "$healing_prompt" 2>&1) || true
    
    healing_log "📋 Healing result: $healing_result"
    
    # Check if healing agent suggested a retry
    if echo "$healing_result" | grep -qi "retry\|re-run\|try again"; then
        healing_log "✅ Healing agent suggests retry"
        return 0
    else
        healing_log "⚠️  Healing agent did not suggest retry"
        return 1
    fi
}

# Monitor sessions and trigger healing
monitor_sessions() {
    log "🚀 Starting Jarvis Error Daemon (check interval: ${CHECK_INTERVAL}s)"
    
    # Track failed sessions to avoid duplicate healing
    declare -A failed_sessions
    declare -A retry_counts
    
    while true; do
        # Find recent session files (last 1 hour)
        while IFS= read -r -d '' session_file; do
            local session_id=$(basename "$session_file" .jsonl)
            
            # Skip if already processed
            if [[ -n "${failed_sessions[$session_id]:-}" ]]; then
                continue
            fi
            
            # Check for errors in session
            if grep -qE "(Error:|Exception:|failed|ELIFECYCLE|Command failed)" "$session_file"; then
                log "🚨 Error detected in session: $session_id"
                
                # Mark as failed
                failed_sessions[$session_id]=1
                retry_counts[$session_id]=0
                
                # Extract task ID if available
                local task_id=$(grep -oP '(?<=task:)\d+' "$session_file" | head -1 || echo "unknown")
                
                # Attempt healing
                if heal_error "$session_file" "$task_id" "${retry_counts[$session_id]}"; then
                    # Retry the task
                    retry_counts[$session_id]=$((${retry_counts[$session_id]} + 1))
                    
                    if [[ ${retry_counts[$session_id]} -le $MAX_RETRIES ]]; then
                        healing_log "🔄 Retry $retry_count/$MAX_RETRIES for task $task_id"
                        
                        # Re-run task (implement actual retry logic here)
                        # This would call jarvis-delegate.sh or delegate-with-healing.sh
                        # For now, just log
                        healing_log "TODO: Re-run task $task_id"
                    else
                        healing_log "❌ Max retries exceeded for task $task_id"
                    fi
                fi
            fi
        done < <(find "$SESSION_DIR" -name "*.jsonl" -mmin -60 -print0 2>/dev/null)
        
        sleep "$CHECK_INTERVAL"
    done
}

# Main
case "${1:-monitor}" in
    monitor)
        monitor_sessions
        ;;
    test)
        log "🧪 Testing error detection..."
        if [[ -n "${2:-}" ]]; then
            extract_error_bundle "$2"
        else
            echo "Usage: $0 test <session-file>"
            exit 1
        fi
        ;;
    *)
        echo "Usage: $0 {monitor|test <session-file>}"
        exit 1
        ;;
esac