openclaw/NIGHT-WARDEN.md

🌙 Thomas Wraith - Night Warden Agent

Overview

Thomas Wraith is your autonomous night agent using the local Ollama model. Perfect for long-running tasks that can run overnight without worrying about API costs.

Configuration

Agent Profile

• ID: night-thomas
• Name: 🌙 Thomas Wraith
• Primary Model: ollama/qwen2.5:0.5b (local, free, patient)
• Fallbacks: GitHub Copilot GPT-5-mini → Google Gemini Flash
• Theme: Night warden - autonomous overnight task executor

Team Membership

• Wardens Team (Night Operations)
  • night-thomas (free, local, patient)
  • ops-harry (high-power troubleshooting)
  • ops-bob (automation expert)

Capabilities

Perfect For:

• ✅ Long-running refactoring (4+ hours, broken into chunks)
• ✅ Batch processing (50+ small tasks overnight)
• ✅ Documentation generation (analyze codebase, write docs)
• ✅ Test generation (write tests for entire modules)
• ✅ Code analysis (security audits, dependency checks)
• ✅ Data processing (transform/analyze large datasets)

Performance Characteristics:

• Speed: ~10s per response (slow but steady)
• Cost: $0 forever (local model)
• Reliability: Always available (no API limits)
• Context: 32k tokens with auto-compaction
• Chunks: Can handle 50+ chunks for multi-hour tasks

Usage Examples

Simple Task

openclaw agent --agent night-thomas --message "analyze the src/ directory and list all TODOs" --local

Multi-Hour Task (50 chunks)

# Start before bed, let it run overnight
openclaw agent --agent night-thomas --message "
Task: Refactor the entire authentication system
- Break into ~50 small chunks
- Document each change
- Run tests after each chunk
- Commit incrementally
- Take your time, I'm sleeping
" --local --thinking low

Batch Processing

openclaw agent --agent night-thomas --message "
Generate unit tests for all functions in src/utils/
- One test file per source file
- Aim for 80%+ coverage
- Use existing test patterns
- No rush, work through the night
" --local

Workflow Tips

Before Bed:

1. Start Thomas on a chunked task
2. Go to sleep
3. Wake up to completed work

Task Structure:

Long task (4 hours) = 50 chunks × ~5 minutes each

Best Practices:

• Break tasks into atomic chunks
• Let Thomas work at his own pace
• Don't interrupt (let him finish overnight)
• Review results in the morning

Model Behavior

Speed Expectations:

Task Type	Chunks	Est. Time
Simple Q&A	1	~10s
Function refactor	1	~30s
File analysis	5	~3 min
Module refactor	20	~15 min
Full codebase audit	50	~60-240 min

Resource Usage:

• CPU: 100-400% (uses multiple cores)
• RAM: ~500MB-1GB (model loaded)
• Disk: 0 (read-only)
• Network: 0 (fully local)

Integration

Delegate to Thomas:

From any agent:

openclaw agent --message "@night-thomas analyze this codebase overnight" --local

Schedule Night Work:

Use cron or systemd timer:

# crontab -e
0 22 * * * openclaw agent --agent night-thomas --message "nightly code audit" --local

Fallback Chain

If local model fails (shouldn't happen):

1. Primary: ollama/qwen2.5:0.5b (local, free)
2. Fallback 1: github-copilot/gpt-5-mini (subscription)
3. Fallback 2: google/gemini-3-flash-preview (API)

Comparison with Other Agents

Agent	Speed	Cost	Use Case
Claude Sonnet 4.5	Fast	$$$	Production, urgent
GPT-5.3	Fast	$$	Complex reasoning
night-thomas	Slow	FREE	Overnight batch work

Monitoring

Check Thomas's work:

# View session
tail -f ~/.openclaw/agents/night-thomas/agent/sessions/*.jsonl

# Check memory usage
free -h

# See ollama status
systemctl status ollama

Created

• Date: 2026-02-15
• Purpose: Free overnight task executor
• Model: Qwen2.5-0.5B (local)
• Status: Active and ready for night work! 🌙