rocker2018-droid/dsh-longtask-orchestrator

Plugin插件 Native原生 ⭐ 2 MIT Subagents & Orchestration子代理与编排

Project Overview项目介绍

This repository is a native plugin built specifically for DeepSeek Harness (DSH) designed to orchestrate complex long-running tasks across multiple large language models. It follows a tiered architecture that assigns specialized roles to different models: Codex acts as the management layer for planning, scoring, and final review, DSH itself acts as the execution layer running concrete sub-tasks, and Kimi acts as a supplementary layer for visual acceptance, summarization, and cross-validation. The plugin registers nine global tools that support the entire workflow from task initialization to final report output. To install it, you run a DSH CLI command to add the local plugin path, then restart the dsh web process for changes to take effect.

A typical workflow with this plugin starts with initializing a long task via the longtask_begin tool. Codex first breaks the high-level goal down into a list of sub-tasks that include acceptance criteria and dependency relationships, then DSH executes each sub-task one by one using its own tools or sub-agents. After each sub-task completes, Codex scores the result on a 0-10 scale, and any result below a 7/10 passing grade is redone per Codex’s suggestions, with a maximum of three retries allowed. Once all sub-tasks pass, the workflow proceeds to Kimi supplementation and Codex’s final review before outputting the final structured result.

This plugin is released under the open-source MIT license, so it is free to use and modify. It supports custom configuration for the Codex binary path, system proxy settings, Kimi API credentials, and the task state storage directory via environment variables. Codex API calls consume your existing ChatGPT subscription quota, and it shares the login state with Codex++, so the plugin author recommends reducing Codex++ usage while a long task is running. If Codex becomes unavailable, the task pauses and shows a prompt instead of automatically switching to another model for planning or review.

这是一个专为 DeepSeek Harness(DSH)开发的原生长任务编排插件,采用分层架构处理复杂长任务,将不同大模型分配到对应层级完成工作。它将Codex作为管理层,负责任务规划、打分和终审,DSH本身作为执行层运行具体子任务,Kimi作为补充层完成视觉验收、结果摘要和交叉验证。插件一共注册了9个全局工具,覆盖从任务初始化到最终结果输出的完整流程。

典型工作流从初始化长任务开始,先由Codex拆解出带验收标准和依赖关系的子任务清单,再交由DSH调用自有工具或子Agent逐个执行。每个子任务完成后会返回给Codex打分,未达及格线会根据修改建议重做(最多重试3次),全部通过后再进入Kimi补充环节和Codex终审,最终输出结构化结果。适合需要处理多步骤复杂长任务的DSH用户使用。

该插件遵循MIT开源许可,无使用成本,安装需要通过DSH的CLI命令添加本地插件路径,重启dsh web后生效。配置支持自定义Codex二进制路径、代理设置、Kimi API信息和状态存储目录,Codex调用消耗ChatGPT订阅额度,长任务运行期间建议减少Codex++使用,如果Codex不可用任务会暂停提示,不会自动替换模型。

Pre-install check安装前体检Compatibility · Security兼容性 · 安全性 1 warning1 项注意
  • Only 2 stars - very few users, little community feedback星标只有 2,几乎没人在用,遇到问题缺少社区反馈
DSH walks through these 9 checksDSH 会逐条核对这 9 项

Compatibility兼容性

  • DSH, Node, OS and profile requirementsDSH 版本 / Node 版本 / 操作系统 / profile 是否满足要求
  • External dependencies and runtimes (Electron / Python / Docker, ...)外部依赖与运行时(Electron / Python / Docker 等)是否齐备
  • Conflicts with installed plugins: command names, skill / tool names, ports, duplicate MCP registration与已装插件是否冲突:命令名、skill / tool 重名、端口占用、重复 MCP 注册

Security安全性

  • Repo matches the facts registered here; archived or abandoned?仓库是否与页面登记一致,是否归档或长期停更
  • Safety of preinstall / install / postinstall and install.sh / setup.ps1preinstall / install / postinstall 与 install.sh、setup.ps1 是否安全
  • curl|bash, download-then-execute, obfuscation, unrelated domains → stop immediatelycurl|bash、下载即执行、混淆代码、无关域名 → 立刻停止
  • Typosquatting or unmaintained packages among the new dependencies新增依赖里有没有 typosquatting 或无人维护的包
  • Requested permissions vs. what the feature actually needs申请了哪些权限、是否超出功能所需(filesystem / network / shell / clipboard)
  • Any sudo / admin requirement, plus uninstall and rollback是否要求 sudo / 管理员权限,以及卸载与回滚方式

Anything uncertain must be marked unknown with a note on how to confirm it. This site's signal screen is a static snapshot, not a security audit.拿不准的必须标「未知」并说明要我怎么确认。本站的信号筛查是静态快照,不能替代安全审计。

Or use CLI install (for developers)或使用命令行安装(适合开发者)

CLI Install命令行安装

dsh plugin --profile web add github:rocker2018-droid/dsh-longtask-orchestrator

把 rocker2018-droid/dsh-longtask-orchestrator 加入你的 DSH 配置(web profile)即可启用。

READMEREADME

dsh-longtask-orchestrator

长任务编排插件:Codex 当管理层(规划/打分/审核),DSH 当执行层,Kimi 当补充层(视觉验收/摘要/交叉验证)。

架构

长任务 ──► Codex 规划(codex_plan)──► 子任务清单(含验收标准/依赖)
              │
              ▼
          DSH 执行(agent 自己的工具 / subagent)
              │
              ▼
        Codex 打分(codex_score,7/10 及格)── 不过 → 按建议重做(最多 3 次)
              │ 通过
              ▼
        Kimi 补充(kimi_summarize 压缩 / kimi_review 交叉验证 / kimi_call 视觉验收)
              │
              ▼
        Codex 终审(codex_review)──► Kimi 摘要 ──► 最终报告

工具(9 个,全局注册)

工具 作用
codex_call(prompt, cwd?, model?) 底层 Codex 桥(只读沙箱、自动代理、串行队列)
kimi_call(prompt, image_path?, model?) 底层 Kimi 桥(文本/视觉)
codex_plan(task, context?) Codex 拆解计划 → JSON 子任务清单
codex_score(task, criteria, result) Codex 打分 0-10 + 修改建议
codex_review(plan, results) Codex 终审
kimi_summarize(text, max_chars?) 大结果摘要压缩
kimi_review(result, criteria) Kimi 交叉验证(第二意见)
orchestrate_state(action, state?) 状态文件读写(read/save/clear)
longtask_begin(goal, context?) 初始化长任务:Codex 规划 + 写入 state.json

其他组成部分

  • skill longtask-orchestrator:循环模板,长任务时自动触发
  • 模型组「DeepSeek + 长任务编排」:delegating adapter,选中后每轮注入编排提示

配置

项 默认 说明
codex 二进制 /Applications/ChatGPT.app/Contents/Resources/codex(不存在则回退 PATH 的 codex) ORCHESTRATE_CODEX_BIN 可覆盖
代理 自动检测 macOS 系统代理(scutil) 子进程会带上 HTTPS_PROXY/HTTP_PROXY
Kimi 复用 ~/.dsh/vision.env(VISION_API_KEY/BASE_URL/MODEL) ORCHESTRATE_KIMI_* 可覆盖
状态目录 <工作区>/.orchestrate ORCHESTRATE_STATE_DIR 可覆盖

安装

dsh plugin --profile web add link:/绝对路径/dsh-longtask-orchestrator

重启 dsh web 生效。

使用

  1. 新会话,切到「DeepSeek + 长任务编排」模型组(或直接提出长任务,skill 自动触发)
  2. longtask_begin(goal) 初始化 → 按计划的子任务逐个执行 → codex_score 打分 → 迭代
  3. 全部通过 → codex_review 终审 → 报告

注意事项

  • Codex 走 ChatGPT 订阅额度,与 Codex++ 共用登录态 → codex 调用已串行化,长任务期间尽量少用 Codex++
  • Codex 不可用时任务会暂停并提示,不会用 Kimi 顶替规划/评审
← 上一个 Prev dsh-browser 下一个 Next dsh-hotreload-plugin-manager →