Conversation-First Scheduling API: Build the No-Calendar-Sync Fallback

Build the optional-sync fallback in eight steps

Test with three required people and one optional participant. Two calendars are connected, one participant answers conversationally, and one participant starts with an expired provider connection. The meeting should use one coordination record and one completion path.

1. 1

   Parse one bounded meeting instruction

   Create a typed brief with meeting request ID, revision, organizer, purpose, duration, date window, deadline, meeting time zone, working-hour policy, location, required participants, optional participants, approval mode, and permitted participant channels. Keep the source instruction for audit, but do not make later workers reread prose to recover quorum or policy.
   • Reject missing organizer, duration, date window, deadline, or required participant.
   • Use IANA time-zone names for the meeting and each known participant.
   • A change to people, window, duration, or authority creates a new revision.
2. 2

   Resolve authority and contact consent per person

   Record who supplied each participant, whether the organizer may contact them for this meeting, the allowed channel, the identity to use, reminder limit, response deadline, and opt-out state. Calendar-read permission, participant-outreach authority, event-write permission, and cancellation authority are separate capabilities.
   • Never infer outreach authority from possession of an email address.
   • The participant can decline, opt out, or request human contact.
   • Sensitive operations require explicit approval under the organizer’s policy.
3. 3

   Read the calendars that are connected

   Query approved Google and Microsoft free/busy inside the exact date window. Normalize intervals into UTC while preserving source time zone, provider, retrieval time, permission class, and error state. Request the narrowest access that supports the job. Event titles, descriptions, locations, and attendee lists are unnecessary for overlap.
   • A partial or failed provider response is missing evidence, not free time.
   • Classify denied access, expired access, throttling, and provider errors separately.
   • Attach source freshness so a proposal can be invalidated before booking.
4. 4

   Send a bounded availability request for what is missing

   Tell the participant who is organizing, the meeting purpose, duration, bounded date window, displayed time zone, response deadline, and how their answer will be used. Ask for usable windows, not a calendar dump. A private response page can be one conversation surface, but it must return to the same meeting record rather than opening a second polling workflow.
   • Do not expose other participants’ votes, event details, or private constraints.
   • Use one outreach idempotency key per participant, meeting revision, and message purpose.
   • Give the participant a clear decline and correction path.
5. 5

   Parse replies into evidence, not guesses

   Convert ‘Tuesday after 2’ into intervals only when the date context and time zone are known. Store the original reply in a restricted record, then save normalized windows, exclusions, source, received time, parser confidence, and unresolved fields. Ask one narrow clarification when a zone, date, or boundary is ambiguous.
   • Never invent a time zone from an email domain or phone number.
   • Treat ‘probably’ and ‘should work’ according to explicit confidence policy.
   • A correction supersedes old evidence without erasing the audit trail.
6. 6

   Run reminders and recovery as durable work

   Persist outreach state outside the model transcript. Schedule reminders by participant local time, organizer policy, response deadline, and previous contact. Retry transport failures without duplicating a delivered message. Stop on decline, opt-out, reminder cap, deadline, organizer cancellation, or human reclaim.
   • A worker restart resumes the same participant state.
   • Silence from a required participant blocks completion rather than changing the quorum.
   • Every stopped path returns a reason and one safe next action.
7. 7

   Find overlap across both evidence sources

   Intersect every required participant’s current intervals whether they came from connected free/busy or a response. Score survivors for optional attendance, working hours, buffers, organizer preferences, time-zone fairness, and deadline. Save the participant evidence versions behind each proposal and rebuild when any required source changes.
   • Do not rank a slot that excludes a required person.
   • Show which optional people can attend before approval.
   • A no-overlap result names the smallest duration, date-window, role, or preference change an authorized person can make.
8. 8

   Commit and verify one canonical meeting

   Bind approval to the current brief revision, selected slot, attendee set, organizer calendar, and evidence versions. Re-check connected required calendars immediately before the write and confirm that conversational windows remain inside their stated freshness policy. Use one booking key, reconcile uncertain writes, persist the provider event ID, and verify invitations.
   • A repeated command returns the original event rather than creating a duplicate.
   • Completion includes start, end, time zone, required attendees, optional attendees, and dispatch state.
   • A conflict after approval becomes a recoverable exception, not a false success.

Compare fallback approaches by whether they finish the meeting

The no-sync path is not an edge case in cross-company scheduling. Compare how much coordination your application still owns after one participant cannot or will not connect a calendar.