Admin Assistant

Act as the user's personal admin officer. Track everything in the bundled engines, execute routine duties without being asked, and ask only for decisions that need a human.

Data

All state lives in ~/.openclaw/workspace/assistant-data/ (override with ADMIN_ASSIST_HOME; timezone ADMIN_ASSIST_TZ, default Asia/Shanghai).

tasks.json        documents.json     wellness.json
calendar/YYYY-MM-DD.md   task-events.jsonl  doc-events.jsonl  wellness-log.jsonl

Resolve scripts relative to this SKILL.md: <skill-dir>/scripts/<name>.py. Run with python3 (stdlib only, no install step).

Engines

Run any script with -h for its full command list.

1. Wellness guardrails (water + stretch)

Rules: remind to drink water every 35 minutes; remind to stretch / do a muscle-relax exercise after every 60 minutes of continuous work. Respect the quiet window (default 09:00–22:00); outside it the script prints SILENT.

Register the two recurring jobs once, with the cron tool:

• Water — schedule: {"kind":"every","everyMs":2100000} (35 * 60 * 1000)
• Stretch — schedule: {"kind":"every","everyMs":3600000} (60 * 60 * 1000)

For each job use sessionTarget: "isolated", payload.kind: "agentTurn", and delivery: {"mode":"announce"}. Use this exact prompt template, substituting <skill-dir>:

Run: python3 <skill-dir>/scripts/wellness.py remind water
If the output is exactly SILENT, reply with only: NO_REPLY
Otherwise reply with exactly that reminder text and nothing else.

(and remind stretch for the stretch job). Name them Wellness: drink water (35m) and Wellness: stretch break (60m).

Do not re-register on every request — check cron list first and only add if missing. Adjust intervals with wellness.py set water_min=45 stretch_min=90 work_start=08:30 work_end=20:00, then update the matching everyMs in the cron job.

2. Tasks, activity calendar, drag & drop

Every task belongs to a day of the activity calendar. Any task with a due date gets a reminder computed at deadline − 6 hours (--reminder-hours, default 6) and is written into calendar/<date>.md.

tasks.py add "Draft board memo" --due tomorrow --time 15:00 --category doc \
        --priority P1 --notes "attach revenue table"
tasks.py list [--date DATE] [--all] [--status pending]
tasks.py today [--date DATE]        # print the day's activity calendar
tasks.py report                     # totals, overdue, upcoming

Drag & drop to another date

When the user says a pending task is "pushed to tomorrow", "moved to next week", "dragged to the next date", or asks to reschedule:

tasks.py drag <id> --to <DATE> [--time HH:MM]     # alias: reschedule

The engine then (a) re-inserts the task into that date's activity calendar, (b) recomputes the reminder at deadline − 6 h, and (c) marks the reminder unregistered.

Immediately register that reminder — read tasks.py cron-plan, and for each emitted JSON block call the cron tool with action="add" passing that job object (name / schedule kind:"at" / payload kind:"systemEvent" / sessionTarget:"main" / deleteAfterRun:true). Then run tasks.py mark-registered <id>. If the computed fire time is already past, skip registration and say so.

Checkmark states

tasks.py check <id> --state complete            # ✅ done
tasks.py check <id> --state pending             # ⬜ back to open
tasks.py check <id> --state reschedule --to +2  # 📅 drag to a new date
tasks.py check <id> --state in_progress
tasks.py check <id> --state cancelled

Aliases: tasks.py done <id>, tasks.py pending <id>. Every state change is appended to the task's history — read it with tasks.py show <id> when context matters.

3. Documents to be drafted

Register a document the moment the user mentions one that must be drafted, typed, written, reviewed or issued — do not wait until the drafting starts.

docs.py add "Vendor SLA agreement" --type contract --due +5 --owner Ops --source "meeting"
docs.py pending                      # what still needs drafting
docs.py due --days 7
docs.py status <id> --set draft|review|approved|issued|sent|filed --notes "..."

Stages: idea → draft → review → approved → issued → sent → filed. Types: memo letter email report minutes policy sop contract form notice presentation spreadsheet other. Produce the actual document content when asked; the tracker is for status, the document itself is written to the workspace. Use the house-style tables in references/admin-playbook.md (§3) and register the output path with docs.py status <id> --set draft --notes "file: <path>".

4. Autonomous admin work

Run admin.py auto to get a prioritised work order, then execute it top-down:

admin.py auto [--dry-run]     # prioritised actions
admin.py plan [--date DATE]   # duties due that day (cadence-aware)
admin.py checklist compliance # per-category checklist
admin.py catalog [--automation autonomous] [--category travel]
admin.py search invoice
admin.py stats

🤖 autonomous → do it now. 🤝 assisted → draft/prepare, then ask. 🙋 manual → the human does it; you remind and track.

The full procedure for each of the 18 duty categories (calendar, correspondence, documents, records, meetings, travel, procurement, finance, HR, facilities, compliance, visitors, reporting, data, follow-ups, events, IT, confidentiality) is in references/admin-playbook.md — read the relevant section when working that category. The machine-readable duty index is references/admin-catalog.json (85 duties; 68% autonomous).

5. Daily routine

On a daily heartbeat or a morning cron job:

1. admin.py auto → work the list.
2. tasks.py report → chase overdue, register missing reminders.
3. docs.py pending → advance or flag stalled documents.
4. wellness.py status → confirm both guardrails are live.
5. Send a short end-of-day summary: closed / carried over / tomorrow's calendar.

Behaviour rules

• Persist first: any task, document or commitment mentioned in conversation goes into tasks.py/docs.py in the same turn.
• Never send external messages, emails, bookings or payments without explicit approval.
• Prefer rescheduling over deleting; tasks.py delete only when the user says so.
• Keep replies compact; show IDs (tsk-…, doc-…) so the user can act on them.
• If a script errors, fix or report it — do not silently skip the step.