✨ Highlights

• 🗣️ Create naturally — tell an Agent what to do and when.
• 🗓️ Schedule precisely — one-time instants, RFC 5545 recurrence, and IANA time zones.
• 🧼 Start fresh — every run uses a new session with no previous chat history.
• 👀 Stay informed — see result summaries, duration, errors, history, and session links in one place.
• 🛡️ Limit access — choose any permission preset advertised by the DSH Host for each task.
• 🤖 Choose execution — optionally pin an Agent preset, provider/model pair, and ordered skills.
• 📌 Continue a conversation (MVP) — automation_create may pin only the current Agent conversation to a persisted session after explicit target and creation confirmation. The target is checked by the Host persistence API; REST target changes and arbitrary session pickers are unsupported. Pinned runs resume exactly, never create a replacement, and show an Open conversation link when available.
• 🧭 Start with guidance — use outcome-focused templates or a guided creation conversation.
• 🎛️ Stay in control — edit, run, pause, resume, or delete tasks from the UI.

🚀 Quick start

1. Install

dsh plugin --profile web add @alpacachen/dsh-automation

2. Restart DSH

Restart dsh web so the plugin can load.

3. Ask an Agent

Tomorrow at 9:00 AM in Asia/Shanghai, review this workspace for release blockers and create a one-time automation.

Or:

Every weekday at 6:00 PM, summarize today's changes and open work. Report only; do not modify files.

After creation, open Automations in the sidebar to manage the task.

Use New automation at any time for a guided setup, or choose a result-focused template in the empty state. Before creation, the Agent previews the name, schedule and time zone, workspace, Agent preset, provider/model, selected skills, exact Host permission, notifications, and failure-pause policy, then waits for confirmation.

🎛️ Manage and edit

Each task card lets you:

• change the name, prompt, or schedule;
• run now, pause, resume, or resume and run;
• open the latest session or any recent run;
• choose failure-only, every-completion, or no sidebar notifications;
• choose a saved Agent preset, provider/model pair, and ordered skills, or follow Host defaults;
• see or change any Host permission preset for future runs; actual permission changes require confirmation;
• retry failed runs and optionally pause after three consecutive failures;
• delete future scheduling while keeping existing sessions.

Visual recurrence editor

Common schedules do not require writing RRULE by hand:

Settings	Meaning
Daily · every 1 day	Run every day
Weekly · every 3 weeks · Monday–Friday	Every third week, run Monday through Friday
Monthly · every 2 months · day 15	Run on day 15 every two months

You can also stop after a number of runs or on a date. Use Advanced RRULE for uncommon rules.

⏱️ How runs work

Schedule due → Global queue → Fresh or pinned persisted session → Agent runs → Result and session link recorded

• Automations run globally one at a time and never overlap.
• Runs time out after one hour by default; set maxRunDurationMs in plugin config to change the limit.
• Queued or running work can be stopped without moving its future schedule.
• A run active during an unexpected restart is marked outcome unknown instead of being reported as failed or retried.
• Notification-worthy results add a persistent unread badge to the Automations sidebar action.
• Successful runs clear the consecutive-failure count; tasks can optionally auto-pause after three failures or timeouts.
• DSH must be running when work is due; after restart, only the latest missed occurrence runs.
• Transient scheduler failures retry automatically with bounded exponential backoff.
• Paused occurrences are skipped; Resume & run does not move the original schedule.
• Selected skill names load their current definitions before every run. A removed/broken preset, model, skill, or permission fails the run instead of falling back.
• Permission presets with approval: ask remain interactive: unattended runs never auto-approve and may wait until the run timeout.
• Existing state stays at version 1. Older tasks are normalized in memory without a startup rewrite and retain their saved execution and permission settings.

🗓️ Schedule formats

Type	Parameters	Example
One-time	once_at	2026-09-01T01:00:00.000Z
Recurring	rrule + time_zone + start_at	FREQ=WEEKLY;BYDAY=MO,WE,FR

• once_at is an RFC 3339 UTC instant.
• rrule is one RFC 5545 rule without DTSTART.
• time_zone is an IANA zone such as Asia/Shanghai.
• start_at is local wall-clock time in YYYY-MM-DDTHH:mm:ss form.

🧰 Agent tools

Tool	Purpose
automation_options	List current Host Agent, model, skill, and permission options
automation_create	Create after confirming optional agent_preset, provider/model, ordered skills, permission, and policies
automation_update	Set or clear execution overrides and replace skills; permission changes require confirmation
automation_list	List tasks and current run state
automation_run	Queue one immediate run without changing the schedule
automation_pause	Pause future scheduling
automation_resume	Resume, optionally with one immediate run
automation_delete	Delete the task and cancel future scheduling

💡 Example requests

“Every Monday at 9:30 AM, check dependencies for important updates.”

“Move the release check to Friday at 4:00 PM.”

“Resume the daily handoff and run it once now.”

“List all automations and pause the dependency check.”

The built-in release readiness, dependency watch, and daily handoff templates remain editable drafts; selecting one never sends it automatically.

License

MIT