Each collection target can announce new items at once or in a daily digest (off by default) in collect.notifyChannel, as titles and links only, with no LLM; each post is recorded in the activity log. The brain gets a read-only collectedNews tool (every target's newest items by default, fenced as untrusted web content), offered with the app's collection file and the collection feature, and available to scheduled tasks. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
3.0 KiB
schedule tool spec (v2)
Reminders, timers and scheduled tasks of the person asking. Product rules: docs/v2/proactive.md
(제안 1). The app keeps the schedules, fires them and delivers them (app/src/admin.spec.md,
"Schedules"); this tool only asks it to add, list or delete.
Principles
- Only in v2, only the asker's own. The tool exists for a turn only when the v2 voice worker's
brain put a schedule backend in the turn state (
turn_state.set_schedule). Without one (v1, evals without a backend, and scheduled tasks themselves) it is not offered and refuses. The backend acts for the person asking only: they list and delete only what they scheduled themselves. - Never claimed without the tool. The description tells the model to call it for reminders,
timers, alarms and things to check at a time or regularly, and never to say something was
scheduled unless the tool said so (the "5분 타이머" confabulation,
docs/v2/decisions.md). - Tasks only look things up. A
taskruns at its time with nobody there to ask, so it gets the read-only tools only (voice workerTASK_TOOLS: weather, web search, reading a page, the asker's meals, the collected news, the tool search) and no MCP tools. It cannot send, delete, schedule or operate anything. - Raw data. The tool returns plain text for the reply model; no LLM call inside.
Schema
| action | needs | does |
|---|---|---|
add |
text, at or in_minutes (kind, repeat) |
schedules it; returns its id, kind, next time and repeat |
list |
the asker's schedules, soonest first | |
delete |
id |
deletes one of the asker's schedules |
kind:remind(default) tellstextat the time;taskdoestextat the time and tells the result.at: local time of this PC,YYYY-MM-DD HH:MM(aTbetween and seconds are accepted).in_minutes: minutes from now (above 0), used instead ofatwhen given; it spares the model the arithmetic for timers.repeat:none(default),daily,weekdays(Monday to Friday),weekly(the weekday ofat). A time that has passed is refused unless it repeats (the app then takes the next one).text: whitespace collapsed, 1-500 characters.action,kindandrepeatare read as their first known word, so planner forms likeadd_reminderwork. A refused add never reaches the app.- Lines:
- id <id>: <kind> at YYYY-MM-DD HH:MM (<weekday>)[, repeats <repeat>]: <text>.
Backend
turn_state.get_schedule() returns an object with act(action, **fields) -> dict (add: kind,
text, at, repeat; delete: schedule, the schedule's id): {"ok": true, "item": {...}} for add and delete, {"ok": true, "items": [...]} for list, {"ok": false, "error": "..."} on failure. Items carry id, kind, text, next (Unix time) and repeat. The voice
worker implements it (voice_worker/app_bridge.py): each action goes to the app as
schedule_action with the asker and where they asked (voice_worker.spec.md, "Schedules").
Failures come back as Error: ... text with success=False.