dsh-auto-continue

When running long-lived agent tasks in DeepSeek Harness Web, interruptions from network jitter, timeouts, or upstream 5xx errors frequently break the session, forcing users to stay by the screen and manually type “continue” — a drain on both focus and effort. dsh-auto-continue eliminates this repetitive chore: it lives inside the host process, listens to the session event stream in real time, and automatically sends a configurable continue message whenever it detects a non-human interruption, letting the agent seamlessly resume its work with no manual intervention.

This plugin goes far beyond sending a single phrase. It features smart error classification that distinguishes transient failures (network timeouts, 429) from permanent ones (auth errors, insufficient balance, unknown model) — only the former get auto-resumed, while the latter are skipped and notified to avoid futile retries. An adaptive backoff strategy gradually lengthens the wait after consecutive failures, protecting struggling upstreams. An idempotency guard inspects the last tool call before resuming and instructs the model to verify state first, preventing duplicate execution of operations that may have already succeeded. A loop guard watches running turns and restarts them with a prompt to try a different approach when it detects repetitive output or spinning behavior.

The plugin also offers rich customization: continue text supports placeholders to carry failure context, a dedicated template handles max-token interruptions, global and per-session pause controls give you back command at any time, and browser notifications plus a live stats panel keep you informed of every auto-recovery. For developers, researchers, and automation-heavy users who need long-running agent tasks without being derailed by network hiccups, dsh-auto-continue is a reliable and thoughtful guardian.