🌙 AI Night Shift

A multi-agent autonomous framework that lets your AI assistants work while you sleep.

AI Night Shift is an open-source framework for running multiple AI agents (Claude Code, Gemini, and more) in coordinated autonomous sessions during off-hours. Born from 30+ real production night shifts, this isn't theoretical — it's battle-tested.

What Makes This Different

Most "autonomous agent" tools run a single agent in isolation. AI Night Shift orchestrates multiple heterogeneous AI agents working together:

Agent	Engine	Role	Mode
Developer	Claude Code	Coding, debugging, deploying	Continuous (hours)
Researcher	Gemini CLI	Research, data gathering, triage	Periodic (minutes)
Coordinator	Any LLM	Task routing, monitoring	Heartbeat (30min)

They communicate through shared protocols — a file-based message queue, shared chat log, and task board integration.

Architecture

┌─────────────────────────────────────────────┐
│              AI Night Shift                  │
│                                              │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  │
│  │  Claude   │  │  Gemini  │  │Heartbeat │  │
│  │  Code     │  │  CLI     │  │  Agent   │  │
│  │          │  │          │  │          │  │
│  │ night_   │  │ patrol.  │  │ heartbeat│  │
│  │ shift.sh │  │ sh       │  │ _config  │  │
│  └────┬─────┘  └────┬─────┘  └────┬─────┘  │
│       │              │              │        │
│       └──────┬───────┴──────┬───────┘        │
│              │              │                │
│       ┌──────▼──────┐ ┌────▼─────┐          │
│       │ night_chat  │ │ bot_inbox│          │
│       │    .md      │ │  (JSON)  │          │
│       └─────────────┘ └──────────┘          │
│                                              │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  │
│  │ Plugins  │  │Dashboard │  │Templates │  │
│  └──────────┘  └──────────┘  └──────────┘  │
└─────────────────────────────────────────────┘

Quick Start

1. Install

git clone https://github.com/judyailab/ai-night-shift.git
cd ai-night-shift
bash install.sh

2. Configure

# Copy the example config and edit your settings
cp config.env.example config.env
nano config.env

# Customize the night shift prompt
nano claude-code/prompt_template.txt

3. Test

# Run a single round to verify setup
bash claude-code/night_shift.sh --max-rounds 1

4. Schedule

# The installer adds cron jobs automatically, or set up manually:
crontab -e
# Add: 0 1 * * * cd ~/ai-night-shift && bash claude-code/wrapper.sh

Modules

Module	Description	Docs
Claude Code	Continuous developer sessions	README
Gemini	Periodic patrol and research	README
OpenClaw	Heartbeat coordinator pattern	README
Protocols	Inter-agent communication	README
Plugins	Extensible pre/post/task hooks	README
Dashboard	Visual monitoring interface	Open dashboard/index.html
Templates	Prompt templates by use case	4 templates included

Prompt Templates

Template	Use Case
development.txt	Coding, testing, debugging
research.txt	Data gathering, analysis
content.txt	Writing, translation, SEO
maintenance.txt	System admin, monitoring

Plugin System

Extend your night shift with pre-built or custom plugins:

# Enable a plugin
ln -s plugins/examples/system_health.sh plugins/enabled/

# List all plugins
bash plugins/plugin_loader.sh --list

Built-in plugins: System Health, Backup, Git Commit Summary, Morning Report, De-Sloppify

Dashboard

Open dashboard/index.html in a browser. Drag and drop your report files to visualize:

• Agent activity and status
• Round-by-round timeline
• Night chat messages
• System health metrics

Agent Adapters

The night shift runner is agent-agnostic. Switch agents with one config change:

# In config.env
AGENT_ADAPTER=claude-code   # default
# AGENT_ADAPTER=codex-cli   # OpenAI Codex CLI
# AGENT_ADAPTER=aider       # Aider
# AGENT_ADAPTER=custom      # Your own (copy adapters/custom.sh)

Or via CLI flag:

bash claude-code/night_shift.sh --adapter codex-cli

Create your own adapter: copy adapters/custom.sh, implement 5 functions, done. See adapters/ for details.

Prompt Design for Autonomy

The most common pitfall: your agent stops mid-task and waits for confirmation that never comes.

All included templates have an Autonomy Rules block that prevents this:

• Never ask for confirmation — decide and execute
• Never wait for user input — choose the safest option and proceed
• Never use interactive commands
• If stuck 3 times, log and move on

See docs/advanced.md for the full prompt design guide.

Advanced Features

• Completion Signal — agents can say "I'm done" to end the shift early
• Shared Task Notes — cross-round context memory bridge
• De-Sloppify Pattern — separate cleanup pass for code quality
• Anti-Pattern Guide — avoid common autonomous loop pitfalls

Requirements

• Bash 4+ and Python 3.6+
• At least one AI CLI tool:
  • Claude Code (npm install -g @anthropic-ai/claude-code)
  • Gemini CLI (npm install -g @google/gemini-cli)
• A Linux/macOS system with cron and timeout (GNU coreutils; macOS: brew install coreutils)

Safety & Security

• PID locking prevents concurrent runs
• Time windows ensure shifts end on schedule
• Rate limit handling with automatic retry
• No secrets in code — all credentials via environment variables
• Append-only communication — agents can't delete each other's messages
• Plugin timeout — max 5 minutes per plugin execution

Contributing

See CONTRIBUTING.md for guidelines.

License

MIT — Judy AI Lab

---

Built with real-world experience from 30+ autonomous night shifts. If your AI works harder while you sleep, you're doing it right. 🌙