dsh-web-relay

Handing complex engineering work to a main agent inside DeepSeek Harness typically means coordination lives in chat, review is improvised, and a host restart erases all context. dsh-web-relay replaces that ad-hoc flow with a repeatable, machine-readable protocol covering the entire loop: user intent, external-AI planning, evidence-backed execution by the main agent, multi-tier review, and final release.

Three subsystems carry the weight. The Step List state machine tracks task dependencies as a DAG, enables atomic rollback, restructure, and AutoIteration version gates, and resolves multi-option conflicts via scored alternatives adjudication. A five-tier review chain—external Gemini, Web Gemini, local Claude Code, dialog, then manual—catches issues progressively, stamping every verdict with reviewedBy, provider label, and fallback reason for a fully queryable audit trail; Claude Code runs under a strict tool whitelist and gracefully degrades into the three-party chain if it fails. For recovery, a restart-event trigger paired with a 15-minute heartbeat provides dual-fault-tolerant resume, underpinned by a self-healing watchdog (Windows Task Scheduler or Linux systemd), so long-running tasks survive host restarts with zero manual intervention.

Past a single session, the plugin retains 37 incident-retrospective lessons, 17 capability registry entries, and reusable skill definitions so agent experience compounds rather than evaporates. It is aimed at dsh power users who need multiple AIs collaborating on real engineering work with auditable review, fault-tolerant recovery, and cross-session capability retention.