dsh-web-relay
在终端中运行以下命令:
dsh plugin install victormshan/dsh-web-relay
将以下提示词粘贴到 DeepSeek Harness 对话框中:
在 DeepSeek Harness 中执行 dsh plugin install victormshan/dsh-web-relay 即可安装该插件,项目完整源码位于 https://github.com/victormshan/dsh-web-relay
插件介绍
When you hand complex engineering tasks to a main agent inside DeepSeek Harness, the usual pain points are the same: coordination lives in chat messages, review is ad hoc, and a host restart wipes the context. dsh-web-relay turns the full cycle—user intent, external-AI planning, evidence-backed execution by the main agent, multi-tier review, and final release—into a reusable, executable protocol with machine-readable state at every step.
Three layers do the heavy lifting. On the process side, a Step List state machine manages task dependencies as a DAG, supports atomic rollback, restructure, and AutoIteration version gates, and resolves multi-option conflicts through scored alternatives adjudication. On the review side, a five-tier chain (external Gemini, Web Gemini, local Claude Code, dialog, manual) catches issues progressively; every verdict is stamped with reviewedBy, provider label, and fallback reason so the full audit trail is queryable. Claude Code runs under a strict tool whitelist and degrades back into the three-party chain on failure. On the recovery side, a restart-event trigger plus a 15-minute heartbeat form a two-fault-tolerant resume mechanism, backed by a self-healing watchdog on both Windows Task Scheduler and Linux systemd, so long-running tasks survive host restarts without manual intervention.
Beyond a single run, the plugin persists 37 incident-retrospective lessons, 17 capability registry entries, and reusable skill definitions across sessions, so agent experience compounds rather than evaporating. It is built for dsh power users who need multiple AIs collaborating on engineering work with auditable review, resilient recovery, and cross-session capability retention.
使用场景
- 多 AI 协同完成复杂工程任务时,用五级审核链逐级把关并全程留痕
- 长周期任务遭遇宿主重启后,事件触发加心跳双保险自动续跑不丢状态
- 跨会话沉淀 Agent 事故教训与技能定义,避免重复踩坑
适合人员
- 深度使用 DeepSeek Harness 进行多 AI 工程协作的开发者
- 需要可审计审核链路和故障自愈能力的技术团队
- 追求 Agent 经验跨会话持久化的 dsh 高级用户