Case study · Reliability
Exactly-once channels
Implemented · tested (documented) · runs on the operator's workstation
When two programs hand work to each other through files and a socket, a crash or a retry can make the same message go out twice. These channels make every request claimable exactly once and resumable after a crash, and the message-sending sequence is proven to send at most once.
Overview
The browser-automation system exposes file channels for machine-to-machine use. Each request is claimed exactly once, and a request interrupted by a crash resumes rather than repeats. Socket ownership is generation-numbered, so a stale process cannot act for a newer one. Writers publish atomically and readers act only on completed files.
A message-submission sequence is proven by mutation tests to send at most once: recovery starts from a durable anchor recorded before submission, so a possibly-sent message is never sent twice. An append-only, hash-chained context log lets the system resume after a crash with its history intact.