pipeline-kernel

DeepSeek Harness (DSH) 插件（Cordis）｜ 多 Agent 管线管理内核：任务板 / 台账 / 部署 / 注册表 / 巡检 / 打包复用 / 控制面板 A DSH (DeepSeek Harness) Cordis plugin — the management kernel for multi-agent pipelines: task board / ledger / deployment / registry / watchdog / packaging & reuse / control panel.

pipeline-kernel 是管线无关的管理内核：管线 = 配置数据（pipelines 表一行，用 pipeline_define 定义），内核不绑定任何具体业务。默认无预置管线，由使用者自行定义与部署。

The kernel is pipeline-agnostic: a pipeline is just a config row (defined via pipeline_define); no business is baked in. There is no preset pipeline — define and deploy your own.

English · 中文

中文

定位

• 内核 = 管线无关的管理能力（通讯 / 台账 / UI 三大职责，见 docs/）；管线 = 数据（配置行）
• 业务能力（生图、视频等）是管线内部设置，不进内核
• 多语言：UI 与文档自动跟随 DSH 语言设置（中文 / English）

安装

// C:\Users\<you>\.dsh\profiles\web\package.json → dependencies
"dsh-pipeline-kernel": "link:<本包路径>"
// bundles 列表加入 "dsh-pipeline-kernel"

或 dsh plugin --profile web add <本包路径>。重启 dsh web 后，新会话带 pipeline_* 工具（存量会话工具目录固定，需重建会话）。

快速开始

1. 定义管线：pipeline_define {id:"my-pipe", roles:{supervisor:{preset:"pipeline-supervisor"}, ...}, chain:["supervisor",...], gates:{...}}
2. 起链：pipeline_deploy {pipelineId:"my-pipe"}
3. 播种角色会话：pipeline_seed {pipelineId:"my-pipe"}（按定义给未注册角色创建会话，自动写注册表）
4. 投任务：pipeline_push {title, description, tags:["target:<角色>","task:<编号>","attempt:1"]}
5. 流转：pipeline_list → pipeline_claim → pipeline_done {result:"done route:<角色|UID> 产物:..."}（内核自动路由下一环）
6. 看全景：pipeline_status

控制面板（Web UI）

右下角 pipeline 胶囊按钮 → 展开右侧全高侧栏：

• 活动：各管线运行摘要（角色状态 / 进行中任务 / 最近完成 / 失败置顶）
• 管线：管线一等实体（Graph 缩略图 / KPI / 打包产物）；点击任务展开 Task Inspector
• 新建：定义新管线（工作文件夹必选，走系统目录选择器）
• 归档：冷存储——归档不删除数据，可恢复；永久删除仅在归档箱内、需输入管线 id 确认

巡检（watchdog）

v0.8 起 默认开启（无需配置）：每 5 分钟扫描 open 任务补唤醒（同任务 10 分钟冷却）、done 但路由未完成（僵尸）自愈补投、全链静止 30 分钟汇报主管。无管线/无任务时零噪音。如需调整或关闭，在 cordis.patch.yml 覆写：

# c:\users\<you>\.dsh\profiles\web\cordis.patch.yml
pipeline-kernel:
  watchdogEnabled: false          # 关闭巡检（默认 true）
  watchdogIntervalMs: 300000      # 巡检周期（默认 5 分钟）
  watchdogStallMs: 1800000        # 全链静止判定（默认 30 分钟）
  watchdogWakeCooldownMs: 600000  # 同任务补唤醒冷却（默认 10 分钟）
  watchdogReportCooldownMs: 1800000

文档索引

License

MIT

English

Positioning

• Kernel = pipeline-agnostic management (communication / ledger / UI — see docs/); pipeline = data (a config row)
• Business capabilities (image generation, video, …) belong inside the pipeline, not the kernel
• Bilingual: the UI and docs follow the DSH language setting (中文 / English)

Install

// C:\Users\<you>\.dsh\profiles\web\package.json → dependencies
"dsh-pipeline-kernel": "link:<path-to-this-package>"
// add "dsh-pipeline-kernel" to the bundles list

Or dsh plugin --profile web add <path-to-this-package>. After restarting dsh web, new sessions get the pipeline_* tools (existing sessions keep their fixed tool catalog — recreate them).

Quick start

1. Define a pipeline: pipeline_define {id:"my-pipe", roles:{supervisor:{preset:"pipeline-supervisor"}, ...}, chain:["supervisor",...], gates:{...}}
2. Deploy: pipeline_deploy {pipelineId:"my-pipe"}
3. Seed role sessions: pipeline_seed {pipelineId:"my-pipe"} (creates sessions for unregistered roles, writes the registry)
4. Push a task: pipeline_push {title, description, tags:["target:<role>","task:<id>","attempt:1"]}
5. Flow: pipeline_list → pipeline_claim → pipeline_done {result:"done route:<role|UID> artifacts:..."} (auto-routes to the next link)
6. Overview: pipeline_status

Control panel (Web UI)

The pipeline pill at the bottom-right expands into a full-height sidebar:

• Activity: per-pipeline run summary (role states / in-progress tasks / recently completed / failures on top)
• Pipelines: pipelines as first-class entities (graph thumbnail / KPI / packs); clicking a task opens the Task Inspector
• New: define a new pipeline (working folder required, picked via the system directory picker)
• Archive: cold storage — archiving never deletes data and is reversible; permanent deletion lives only in the archive, requiring the pipeline id to confirm

Watchdog

On by default since v0.8 (no config needed): every 5 minutes it re-wakes open tasks (10-minute cooldown per task), self-heals zombie tasks (done but routing never finished), and reports stalled chains to the supervisor after 30 minutes. Zero noise with no pipelines/tasks. Tune or disable in cordis.patch.yml:

# c:\users\<you>\.dsh\profiles\web\cordis.patch.yml
pipeline-kernel:
  watchdogEnabled: false          # turn watchdog off (default true)
  watchdogIntervalMs: 300000      # tick interval (default 5 min)
  watchdogStallMs: 1800000        # chain-stall threshold (default 30 min)
  watchdogWakeCooldownMs: 600000  # per-task wake cooldown (default 10 min)
  watchdogReportCooldownMs: 1800000

Docs

License

MIT