opencode-autopilot 0.1.0b2

Autonomous overnight engineer for OpenCode projects. Bootstrap, build, and improve while you sleep.

opencode-autopilot

Autonomous overnight engineer for OpenCode projects.

Why I built this

I'm mystic9t — I work a day job and build hobby projects in whatever time I have at night. One of those is Vibes, an astrology and wellness web app I genuinely enjoy but rarely have time to iterate on.

I kept seeing people talk about OpenClaw — an autonomous agent loop built on top of Claude Code. It looked powerful, but every time I tried to make sense of it, I hit a wall. The setup was confusing, the derivatives didn't click for me either, and I wanted something that worked without a steep learning curve.

So I did what any developer does: I went down a rabbit hole. I read what other people were doing — the loops, the session triggers, the heartbeat memory patterns — and I stitched together my own version built specifically for OpenCode and its default model, Big Pickle.

The first night I ran it on Vibes, I woke up to two fully functional new features, committed and working. They weren't perfect — I ran a dedicated cleanup session — but the core work was done while I slept. That felt like something worth packaging up.

opencode-autopilot is the result. A structured way to give OpenCode a persistent memory, a blueprint system, and a session loop so it keeps working instead of stopping after 10 minutes waiting for you.

---

Roadmap

v0.1.0 (Current Beta)

• ✅ Kilocode support — Auto-detect and use Kilocode as an alternative to OpenCode
• ✅ Smart tool switching — Automatically fallback to the other tool when one hits rate limits
• ✅ Rate limit resume — Pause overnight runs when both tools are rate-limited, auto-resume when limits reset

Coming Soon

• Interactive mode for step-by-step approval
• Session resume with --resume flag improvements
• Plugin system for custom agents

---

Commands

Command	What it does
opencode-autopilot	Show help and available commands
opencode-autopilot run	Run autonomous improvement sessions on existing projects
opencode-autopilot run --gg [topic]	Full trust mode — agent researches, decides what to build, and builds it
opencode-autopilot config	Set persistent defaults for model/agent

---

Requirements

• OpenCode or Kilocode installed and in PATH
• Python 3.12+

---

Installation

pip install opencode-autopilot

Or run directly with uvx:

uvx opencode-autopilot gg

---

Default model

Defaults to opencode/big-pickle — OpenCode's built-in free model, available to every OpenCode user with no setup.

Switch models per-run or permanently:

# Per-run
opencode-autopilot run --model anthropic/claude-sonnet-4-5

# Set project default
opencode-autopilot config --model anthropic/claude-sonnet-4-5

# Set global default (all projects)
opencode-autopilot config --model anthropic/claude-sonnet-4-5 --global

# Check what's active
opencode-autopilot config --show

---

Usage

# Existing project — agent improves what's there
opencode-autopilot run

# Full trust — agent researches, decides, and builds with no input from you
opencode-autopilot run --gg

# Full trust with a loose nudge
opencode-autopilot run --gg "something for people who read too much"

# Resume an interrupted cycle from session 6
opencode-autopilot run --resume 6

# Fewer sessions, shorter intervals
opencode-autopilot run --sessions 6 --interval 15

# Show help
opencode-autopilot --help
opencode-autopilot run --help

---

How the memory system works

The agent writes to HEARTBEAT/ — never committed to git, always local. It tracks:

• What was done each session and the build status
• Settled Decisions — things tried and abandoned so it never repeats them
• Paid feature ideas logged separately for you to review
• Plans written during exploration sessions

---

License

MIT © mystic9t