enjambre

Problem

I have a PC and a Mac mini that are idle most of the time, and unattended coding agents tend to run with too many permissions or leave nothing reviewable behind. I wanted to spread work across both with minimal permissions and a result I can audit.

Approach

A Windows PC orchestrates and an Apple Silicon Mac mini runs the Claude Code jobs headless. The PC validates a task contract, picks the machine from a table (routing.toml) plus live state, and sends the job over SSH. The node creates a worktree on its own branch, runs the agent inside tmux and leaves a result.json. The PC fetches the branch, scans it for secrets and pushes it; the jobs.jsonl ledger records node, branch, times, exit code and usage.

The visor is a local dashboard that reads each node’s telemetry (CPU, memory pressure, swap, temperature, network, processes) and the job ledger. Its demo mode runs the same code over a synthetic swarm, so it can be shown without exposing any real machine.

Limits: Mac and Windows only today; unattended jobs need the macOS keychain unlocked after a reboot; exit codes do not check the done criteria, so reviewing the branch is still a human step. Code under MIT; the visor’s fonts under SIL OFL 1.1.