Output Directory Structure

data/db/
└── {agent}/                    # Agent name (browseruse, agente, etc.)
    └── {category}/             # Test category (e.g., dark_patterns, benchmark)
        └── {task}_{count}/     # Individual test run
            ├── {task}.db                   # SQLite database
            ├── {task}_site.txt             # URL with dark patterns
            ├── {task}_task.txt             # Task description
            ├── {task}_commands.py          # Generated commands (optional)
            ├── scratchpad.txt              # Agent reasoning/notes
            ├── scratchpad.txt.bak         # Backup of scratchpad
            ├── video/
            │   └── {task}.mp4             # Screen recording
            ├── html/
            │   ├── initial.html           # Page at start
            │   ├── step_1.html            # After first action
            │   ├── step_2.html            # After second action
            │   └── final.html             # Final page state
            ├── rrweb/
            │   ├── {task}_rrweb_events.json      # Session replay data
            │   ├── {task}_rrweb_viewer.html      # Replay viewer
            │   └── {task}_serve_rrweb_viewer.py  # Local server script
            └── trace/
                └── {task}_trace.zip       # Playwright trace files

data/db/browseruse/benchmark/laptop_search_1/
├── laptop_search.db              # 47 actions recorded
├── laptop_search_site.txt        # agenttrickydps.vercel.app/shop
├── laptop_search_task.txt        # Find and add gaming laptop to cart
├── scratchpad.txt               # Agent reasoning (2.3KB)
├── video/
│   └── laptop_search.mp4        # 2:34 duration, 1920x1080
├── html/
│   ├── initial.html             # Homepage HTML
│   ├── step_1.html              # After search
│   ├── step_2.html              # Product listing
│   ├── step_3.html              # Product details
│   └── final.html               # Cart page
├── rrweb/
│   ├── laptop_search_rrweb_events.json    # 1,247 events
│   ├── laptop_search_rrweb_viewer.html    # Replay interface
│   └── laptop_search_serve_rrweb_viewer.py
└── trace/
    └── laptop_search_trace.zip   # Playwright trace (15MB)

data/db/agente/dark_patterns/hidden_costs_3/
├── hidden_costs.db              # 23 actions, stopped early
├── hidden_costs_site.txt        # URL with dp=hc parameter
├── hidden_costs_task.txt        # Complete purchase task
├── scratchpad.txt              # Shows where agent got confused
├── video/
│   └── hidden_costs.mp4        # Shows point of failure
├── html/
│   ├── initial.html
│   ├── step_1.html
│   └── error_state.html        # Page when agent failed
└── rrweb/
    └── [rrweb files...]        # Can replay up to failure point

# Check directory sizes
du -sh data/db/*/

# Clean old test data
find data/db/ -name "*.mp4" -mtime +30 -delete  # Videos older than 30 days
find data/db/ -name "*_trace.zip" -mtime +7 -delete  # Traces older than 7 days

# Navigate to latest test
cd data/db/browseruse/test_category/
ls -t | head -1  # Most recent test

# View database quickly
sqlite3 $(find . -name "*.db" | head -1) "SELECT COUNT(*) FROM actions;"

# Play latest video
vlc $(find . -name "*.mp4" | head -1)

import sqlite3
import glob
import os

def get_latest_test(agent, category):
    pattern = f"data/db/{agent}/{category}/*/test.db"
    db_files = glob.glob(pattern)

    if not db_files:
        return None

    # Sort by modification time
    latest_db = max(db_files, key=os.path.getmtime)
    return os.path.dirname(latest_db)

def analyze_test_result(test_dir):
    db_path = os.path.join(test_dir, "test.db")

    with sqlite3.connect(db_path) as conn:
        # Get basic stats
        cursor = conn.execute("SELECT COUNT(*) FROM actions")
        action_count = cursor.fetchone()[0]

        # Get success status
        cursor = conn.execute(
            "SELECT value FROM metadata WHERE key='success'"
        )
        result = cursor.fetchone()
        success = result[0] if result else "unknown"

    return {
        "directory": test_dir,
        "actions": action_count,
        "success": success
    }

#!/bin/bash
# cleanup_old_tests.sh

# Remove tests older than 30 days
find data/db/ -type d -name "*_[0-9]*" -mtime +30 -exec rm -rf {} \;

# Compress large trace files
find data/db/ -name "*_trace.zip" -size +100M -exec gzip {} \;

# Archive successful tests older than 7 days
find data/db/ -name "*.db" -mtime +7 -exec python scripts/archive_test.py {} \;

# Daily backup of databases only
tar -czf backups/liteagent_db_$(date +%Y%m%d).tar.gz \
    --exclude="*.mp4" \
    --exclude="*_trace.zip" \
    data/db/

# Weekly full backup
tar -czf backups/liteagent_full_$(date +%Y%m%d).tar.gz data/db/

data/db/browseruse/parallel_test/
├── task_1/                     # First parallel instance
├── task_2/                     # Second parallel instance
├── task_3/                     # Third parallel instance
└── task_4/                     # Fourth parallel instance

# SQLite Browser (GUI)
sqlitebrowser data/db/browseruse/test/task_1/task.db

# Command line analysis
sqlite3 data/db/browseruse/test/task_1/task.db < analysis_queries.sql

# Extract frames at intervals
ffmpeg -i data/db/browseruse/test/task_1/video/task.mp4 \
       -vf fps=1 frames/frame_%04d.png

# Create thumbnail
ffmpeg -i data/db/browseruse/test/task_1/video/task.mp4 \
       -ss 00:00:30 -vframes 1 thumbnail.png

# Start local server for replay
cd data/db/browseruse/test/task_1/rrweb/
python task_serve_rrweb_viewer.py

# Open http://localhost:8000 in browser