@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.

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:

Half	Entry	Role
Server	lib/index.js	Cordis 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.
Protocol	lib/typert.js	Host TYPERT face (tasks/* endpoints), auto-registered by dsh-typert-loader.
Browser	lib/client.js	React 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

▼
sh
复制代码
dsh plugin --profile web add @opendsh/dsh-plugin-scheduled-tasks

Development

▼
sh
复制代码
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

Key	Default	Meaning
maxConcurrentRuns	2	Maximum concurrently running agent sessions across all tasks.
keepRunsPerTask	20	Run-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 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.