DeepSeek Harness 插件

dsh-plugins-ceelog

DSH web plugin: per-project scheduled tasks with prompts, executed as headless agent sessions in the project directory, with durable run history(英文原文)

跳到安装方式

来源信息

GitHub 仓库
Ceelog/dsh-plugins
最近更新
2026年8月19日
分类
自动化与任务
GitHub stars
8
载体类型
plugin
包路径
src/plugins/dsh-plugin-scheduled-tasks
目录证据
上游声明已找到 dsh.bundle
证据路径
src/plugins/dsh-plugin-scheduled-tasks/package.json#dsh.bundle
核对版本
0.1.0-rc.8
上游核对日期
2026-08-20

该证据由上游目录提供。本站没有安装、运行或安全审核这个插件。

安装

默认先复制一段 Prompt,让 Agent 读 GitHub 仓库和源码;需要自己装时再切到命令。

复制这段 Prompt,发给 DSH、Codex 或其他 Agent,让它先读 GitHub 仓库和源码。

请先不要安装或执行任何命令。阅读这个插件的 GitHub 仓库、README 和关键源码,然后用清楚、直接的方式回答以下问题,帮助我判断它是否适合我的需求:

1. 这个插件是什么,解决什么问题;
2. 适合哪些用户和典型使用场景;
3. 安装后如何使用,并给出一个最小使用示例;
4. 有哪些已知限制,以及隐私、安全、兼容性或维护风险;
5. 给出“推荐 / 有条件推荐 / 不推荐”的明确建议和理由。

请区分仓库明确说明、根据源码推断和未知信息。证据不足时请明确说明,不要猜测或照抄 README。

GitHub:https://github.com/Ceelog/dsh-plugins/tree/HEAD/src/plugins/dsh-plugin-scheduled-tasks
插件名:dsh-plugins-ceelog
作者:Ceelog

检查来源文件

安装前先看这个插件目录里的 README 和其他文件。

文件资源管理器3 个文件
README.md来源说明 · 只读预览

@opendsh/dsh-plugin-scheduled-tasks

DSH web plugin: per-project scheduled tasks with prompts. Create a task in the sidebar (⏰ 定时任务), give it a prompt, and the plugin runs that prompt on schedule — as a fresh headless agent session in the project directory — then records the outcome as durable run history.

![](./docs/demo.png)

What it does

  • Project-scoped tasks — each task belongs to a project directory (a DSH

workspace). The panel lists tasks for the selected workspace.

  • Three schedule kinds

- at — run once at an absolute time (strict RFC 3339 instant, or a local date/time with an explicit IANA time zone; DST gaps are rejected, overlaps pick the earlier instant). - every — run on a fixed creation-anchored interval (≥ 5 minutes); missed intervals are never replayed, only the latest due occurrence runs. - cron — run on a cron calendar rule (five/six/seven-field expression, e.g. 0 9 1-5, evaluated in an explicit IANA time zone, DST-aware).

  • Prompt execution — on schedule, a brand-new agent session is created in

the project directory with the task's prompt and driven to quiescence (the same drive pattern as dsh --profile headless). The run session appears in the project's conversation list with a pinned title (⏰ <task name>) and stays resumable; the final assistant text is also captured into the run record.

  • Per-task model override — each task may pin an explicit

provider/model selection, picked from the grouped provider catalog in the panel (the same groups the DSH model selector renders, with the current default selection shown as the first option). Runs then use that model instead of the deployment default; the effective model of every run is recorded in its history. Leaving the picker empty follows the default selection.

  • Run history — status (running / completed / failed), start/finish

times, output (truncated at 20 KB), error messages; newest first, capped at 20 records per task (configurable).

  • Manual run-now — runs a task immediately without touching its schedule.
  • Lifecycle semantics

- Runs only while the DSH web process is alive. After a restart, overdue one-shots run once and are marked overdue; overdue every tasks run only their latest due occurrence. - One-shot at tasks finish after their single run; every tasks stay active and advance to the next anchor-aligned target.

Architecture

A dual-face npm package installed into the web profile:

HalfEntryRole
Serverlib/index.jsCordis plugin: opens the scheduled_tasks storage domain (ctx.storageDomain), mounts the task store, the scheduler (bounded timers, clock-rollback-safe wakes), the headless executor, and the ctx.tasks typert service.
Protocollib/typert.jsHost TYPERT face (tasks/* endpoints), auto-registered by dsh-typert-loader.
Browserlib/client.jsReact panel mounted into the sidebar.footer.action slot; calls the host through the installed remote.tasks namespace.

The client↔server channel is the DSH typert protocol (the same mechanism dsh-commands uses): strict zod codecs validate every argument and result on both sides, and no session needs to exist for the panel to work.

Install

dsh plugin --profile web add @opendsh/dsh-plugin-scheduled-tasks

Development

pnpm --dir src/plugins/dsh-plugin-scheduled-tasks typecheck   # tsc (TypeScript 7)
pnpm --dir src/plugins/dsh-plugin-scheduled-tasks test        # vitest
pnpm --dir src/plugins/dsh-plugin-scheduled-tasks build       # tsc emit + tsdown client bundle + loader wrapper

After changing sources, rebuild and re-sync the profile copy (pnpm install inside the profile re-copies file: dependencies), then restart dsh web — the plugin has no HMR channel.

Configuration

KeyDefaultMeaning
maxConcurrentRuns2Maximum concurrently running agent sessions across all tasks.
keepRunsPerTask20Run-history records retained per task (oldest pruned beyond the cap).

Troubleshooting

Tool calls that fail with “Interrupted: interrupted” (underlying error Cannot read properties of undefined (reading 'prepare')) are almost always a duplicate @deepseek-ai/dsh-tools copy caused by declaring DSH packages in dependencies instead of peerDependencies. See [docs/troubleshooting.md](docs/troubleshooting.md) for the full write-up.

Limitations

  • Schedules fire only while the web process is running (same posture as

dsh-schedule); there is no external wake-up when the process is down.

  • Each run consumes model tokens with the task's pinned model, or the current

default model when the task pins none — the panel says so explicitly.

  • Run history is refreshed by polling while the panel is open (10 s interval);

push updates are deferred work.