dsh-doctor

A single bad plugin update can bring your dsh web down entirely, and fixing it means SSHing in, reading logs, editing YAML, and restarting by hand. dsh-doctor turns that into an automated watchdog: it runs as a standalone Node process (LaunchAgent on macOS, systemd user unit on Linux, Task Scheduler on Windows), probes a health endpoint every 30 seconds, and after consecutive failures enters triage, identifies the offending bundle within a hard 60-second budget, writes a safe-mode bypass patch, restarts, and verifies recovery. Your original config file is never mutated.

Beyond boot recovery, dsh-doctor subscribes to the tools/* event waterfall in-process, classifies every failed tool call (transient, agent, business) and records or defers it per policy, so the AI can pull queued errors at a quiet moment via dsh_doctor_drain_deferred. It also maintains a per-session state machine: if a turn goes idle beyond a threshold with no new events, the doctor injects a continue message to nudge the stuck conversation back to life, up to three times per session, then gives up rather than looping forever.

dsh-doctor is built for users who run multi-plugin profiles that a single bad update can take down, and for engineers running long unattended agent tasks that occasionally stall on a flaky network call or an interrupted LLM stream. It is fully independent of dsh-daemon, so both can coexist as layered protection. If you have a single-plugin setup you can fix by hand, or if you want an AI that auto-debugges your entire DSH install, this is not for you. It is a watchdog, not an AI.