Multi-agent orchestration for Claude Code, OpenCode, and Codex CLI.
superharness sits on top of your AI coding agent and gives it the ability to spawn parallel worker agents, manage them in tmux, handle permissions, detect stalls, and clean up — all autonomously. You just run superharness in your project instead of your agent directly.
cargo install superharness
cd your-project
superharnessThat's it. You get the normal agent interface you already know — but now the AI can spawn parallel workers, coordinate them across isolated git worktrees, auto-approve safe operations, detect stuck workers, and clean up — all without you lifting a finger.
- Install —
cargo install superharness. Requires tmux and an AI coding agent (Claude Code, OpenCode, or Codex CLI). - Run —
superharnessin your project directory. - Work — The AI orchestrator spawns parallel worker agents when the task needs them. Each worker runs in its own git worktree. The orchestrator reviews permission prompts, auto-approves safe operations, detects stalled workers, merges results, and cleans up — without you doing anything.
- Optional: Away Mode — Step away and superharness keeps working autonomously, queuing uncertain decisions for your return.
curl -fsSL https://superharness.dev/install.sh | shsudo add-apt-repository ppa:borodutch/superharness
sudo apt update
sudo apt install superharnessyay -S superharnessbrew install backmeupplz/superharness/superharnesscargo install superharnessRequires: tmux · an AI coding agent (Claude Code, OpenCode, or Codex CLI)
superharness has an away mode for when you step out. The AI keeps workers running and handles safe operations, but queues any real decision — architecture choices, destructive operations, anything it isn't sure about — for your return. Full debrief when you're back.
Multi-Harness Support
- Works seamlessly with Claude Code, OpenCode, and Codex CLI
- Unified orchestration across different AI agents
- Automatic busy/idle detection per harness
Tmux Management
- Spawns workers in isolated panes and git worktrees
- Auto-surfaces workers needing attention
- Hides idle workers, keeps main window clean
- Permission prompts are auto-approved for safe operations (edits, git, builds, tests)
Task Decomposition & Parallelization
- Breaks work into independent task units
- Spawns multiple workers simultaneously
- Tracks progress across all tasks
- Respawns crashed workers with context
Away Mode
- Run
superharnessand step away - AI keeps workers running autonomously
- Safe operations auto-approved, uncertain decisions queued
- Get a full debrief when you return
Heartbeat System
- Detects stuck/idle workers
- Wakes the orchestrator immediately when work is done
- No polling, no artificial sleep — event-driven throughout
Intelligent Detection
- Detects busy/idle state per harness (spinners, permission prompts, keybind hints)
- Automatically re-runs blocked tasks when dependencies finish
- Cleans up worktrees on exit
Full command reference for power users and scripting. You don't need any of these day-to-day — just superharness.
| Command | Description |
|---|---|
superharness |
Start opencode in orchestrator mode |
spawn |
Create a new worker pane |
list |
List all active panes (JSON) |
workers |
List workers in human-readable format (F4 popup) |
status-human |
Human-readable status + worker health (F3 popup) |
read --pane %ID |
Read recent output from a pane |
send --pane %ID --text "..." |
Send input to a pane |
ask --pane %ID |
Detect if a worker is asking a question |
kill --pane %ID |
Kill a pane |
hide / show |
Move pane to background tab or surface it |
resize / layout |
Adjust pane geometry |
compact |
Move small/excess panes to background tabs |
surface --pane %ID |
Bring a background pane back to main window |
git-check --dir /path |
Verify repo is clean before creating a worktree |
respawn --pane %ID |
Kill crashed worker and respawn with crash context |
tasks |
List pending (dependency-gated) tasks |
run-pending |
Spawn tasks whose dependencies have finished |
monitor |
Continuous stall detection and auto-recovery |
watch |
Auto-manage all panes: cleanup, approve safe prompts, nudge stalls |
healthcheck |
One-shot structured health snapshot |
status |
Show current mode and queued decisions |
queue-decision |
Queue a decision for human review |
clear-decisions |
Clear resolved decisions |
loop-status |
Show loop detection state for pane(s) |
loop-clear |
Reset loop history after intervention |
checkpoint |
Snapshot a pane's output |
checkpoints |
List saved checkpoints |
resume |
Spawn a new worker from a checkpoint |
memory |
Store or list key-value facts for a pane |
MIT — see LICENSE
Website: superharness.dev · GitHub: backmeupplz/superharness