dsh-maestro-supervisor

When the DSH Web process crashes, any in-flight agent session is lost, the browser goes blank, and the user must manually refresh and restart the conversation. For long-running, multi-turn agent workloads this is a painful and fragile experience.

This plugin addresses the problem with a three-layer defense. An outer systemd daemon lives outside the pnpm → sh → node process tree, polling the health port every 3 seconds; on crash it auto-rolls-back to the latest known-good snapshot (SHA-256 verified, 3-way rotation) and writes a diagnostic report covering health status, git diff, and log tail, with optional Telegram notification. The inner Host plugin scans for interrupted sessions and dangling open turns 8 seconds after boot, recovers provider and model context from persisted session logs, and calls agents.resume to pick the conversation back up. The innermost Client plugin watches for WebSocket close, offline events, and visibility changes, then starts a 1-second HEAD poll that triggers an automatic page reload the moment the server is back online — no manual refresh required.

Built for developers and operators who run DSH Web locally or on a server and need agent sessions to stay alive across restarts. If you want zero-touch recovery with a full audit trail of what went wrong and how it was fixed, this is the resilience layer you are missing.