Experimental · not released yet

One task, one bench.

A bench is one task's git worktree, kept warm, fenced and checkpointed between agents' turns. SuperBench creates them, leases each to one holder at a time, and checks the work is on the remote before the next one picks it up.

It is an experimental prototype for one machine and trusted repositories. Leases are cooperative, and nothing in it isolates untrusted code.

Checkouts pile up, and handoffs lose work

Give every task its own checkout and they multiply: one per card, per agent, per retry. Nobody cleans them up, because nobody is sure which are safe to delete, until a disk fills and the work in flight stops with it.

And when one agent hands a task to another, whatever the first left uncommitted, or committed and never pushed, is easy to lose.

How a bench works

A command-line tool, bench, and a local daemon take each bench through four steps.

  1. bench create

    Create

    A bench is a git worktree for one task, on its own branch.

  2. bench lease

    Lease

    One holder at a time. Of many simultaneous lease requests, one wins; a restart keeps the lease, and a wrong or old token is refused.

  3. bench checkpoint

    Checkpoint

    Pushes the bench's existing HEAD to its branch, never with force, then confirms the remote has exactly that commit. It never makes a commit for you: a dirty or untracked tree is refused.

  4. bench release

    Release

    Release needs a checkpoint. The next holder finds the work where the last one left it.

Leases are cooperative. A lease stops other bench acquisitions; it does not stop direct edits, other shells, or other processes running as the same user. Leases never expire on their own, so a crashed holder's lease waits for someone to inspect it.

A quota for each bench

Each bench gets its own byte and inode caps on a dedicated XFS mount, so a bench that runs out of room is refused writes while the shared store stays writable.

Managed Linux hosts only. Quotas need a dedicated XFS mount, a root-owned policy and a privileged helper, set up by an operator. On macOS and on other hosts a quota is recorded, not enforced.

Shared mirrors and caches

Benches start from shared git mirrors, and each repository chooses shared or checkout-local caches, so a new bench starts warm instead of cloning and installing from scratch.

Limits: the shared package store is part of the managed XFS setup above and shares its limits. If refreshing the shared mirror fails, the lease is refused. Collecting old shared caches is not built yet.

Archive and reap only after review

Only a released, clean bench whose checkpoint the remote has confirmed can be archived, as a read-only tree. Deleting it waits out a grace period, seven days by default, and a person approves the exact list of files first.

Local and reviewed. Archive and reap are administrative commands on the machine itself. A repository can opt in to archiving after a merged checkpoint, but nothing is deleted without a reviewed plan. Automatic log retention is off until an operator turns it on, and retiring a bench's quota during reap applies only on managed Linux hosts.

Supervised runs, fenced

bench run puts a process tree in its own cgroup and stops all of it, detached children too, before the bench is checkpointed and released. A run that fails keeps its lease.

Linux only, and fencing, not isolation. It needs a delegated cgroup v2. It does not run the task as a separate user or isolate its mounts, so use it with trusted repositories only.

Standalone first

SuperBench is a CLI and a local daemon that talk over a Unix socket. The daemon checks who is connecting with the kernel's peer credentials, and it opens no TCP listener. That last part is a reading of the source today; no test asserts it yet.

It depends on no other suite product, and an architecture test fails any import outside its allowlist. When the suite integrates it, the suite calls SuperBench; SuperBench never calls the suite. That integration is not built yet.

What's not done yet

  • Payload isolation. Running a task's payload as a separate user is in progress. Today there is no user or mount isolation, so SuperBench is for trusted repositories only.
  • One host. Leases are coordinated on a single machine. Moving a bench to another host is a reviewed administrative step, not network coordination.
  • No release. There are no tags, no binaries and no installer yet.
  • Linux for supervised runs and quotas. Fenced runs need Linux with cgroup v2, and enforced quotas need a managed Linux host with XFS.
  • No automatic lease expiry, no remote authentication, no web interface and no suite integration.

Follow along

There is nothing to install yet. Watch or star the repository on GitHub, or subscribe to its release feed, and you will hear when the first release is out.

MIT licensed. The source is on GitHub.