Manage workers

Workers run the child sessions of attempts. Choose where they run, configure their connections, and verify their capabilities before assigning work.

Remote child recovery

On Linux workers, Claude Code and Codex run under a shared process owner. An SSH interruption reconnects to the running child and replays unacknowledged input. The owner deduplicates that input and keeps output in journals beside the checkout. An environment change replaces the retained child after confirming that it stopped.

Child environment values, including gateway caller tokens, travel over SSH stdin. SSH and container command arguments exclude those values during launch, reconnection, and cleanup. The worker stores only a hash of the exported environment for reattachment checks.

Orchestrator heartbeats renew a two-minute lease. When the lease expires, the worker stops the child with TERM, then KILL if it has not exited within five seconds. Boot IDs and process start times prevent stale PID files from identifying an unrelated process after a worker restart. The worker programs ship with usine; workers need Node.js and their harness binaries.

Add capacity

1. Add a local worker
2. Add an SSH worker
3. Add a Boat worker
4. Add a boxd worker
5. Add a Hetzner worker
6. Run attempts in Cloudflare sandboxes

Connect and maintain workers

1. Configure a gateway
2. Connect Boat to a gateway
3. Run worker preflight
4. Pause a worker
5. Remove a worker

The scheduler gives each new attempt to the first worker in workers order with a free slot. To send attempts to one worker, such as a worker you just added, pause the workers before it.

Memory for repository checks

The usine repository gate needs at least 6 GiB per attempt (memory_mib: 6144). On a worker with about 7.7 GiB of RAM, use one attempt slot; a 6656 MiB limit leaves about 1 GiB for the host. pnpm check, pnpm build, and pnpm test run Turborepo tasks one at a time with --concurrency=1; web Vitest uses at most two workers. Preserve those limits when running the full gate. Memory-backed /tmp and /dev/shm also count against the container limit.

Memory kills appear below the attempt outcome even when the attempt succeeds. To collect process details, the worker’s SSH user needs access to the host kernel journal or dmesg. Preflight checks this access on workers with a container runtime and lists a missing one under Preflight findings on the worker’s page; attempts still run. On a systemd host, an administrator can run sudo usermod -aG systemd-journal <ssh-user>, then reconnect SSH and run preflight again. New Hetzner workers include this membership. Kernel diagnostics report task names and host PIDs, with no command-line arguments. A record keeps the first 20 processes; the rest, like unavailable details, are counted separately. A container started again, for example after a test stopped its init, still records the processes killed before the start.