samgu-nrx/caret/meeting-scheduler

Meeting scheduler (Jev edition)

Jev does not generate text. It evaluates typed questions (choice, score, noul) against a JSON state and returns structured answers with probabilities and confidence. This skill therefore splits the work:

State the questions see

{
  "skill": { "name": "meeting-scheduler", "rules": ["..."] },
  "email_thread": { "user_email": "", "subject": "", "messages": [{"from": "", "to": [], "date": "", "body": ""}] },
  "computer_history": { "entries": [{"ts": "", "app": "", "window_title": "", "text": "", "events": []}] },
  "candidates": [{"id": "A", "date": "", "meeting": {"start": "", "end": ""}, "outbound": {}, "return": {}}]
}

candidates is only present in the rank pass. The whole state stays well under Jev's roughly 32k-token budget.

Gates

• intent != schedule_meeting or agreement_in_principle < 0.5: stop after step 1 and report why.
• A candidate whose outbound or return is not in the timetable's coverage is dropped, with the reason listed under dropped.
• An option_<id>_acceptable below 0.5 removes the hold for that option.
• Nothing is sent. Holds are tentative and local (plus an optional webhook the operator configures).

Packaging inputs for Jev

npm run package-input writes jev-input.json: the exact request body for the extract pass (model, state, questions). Paste it into the TypeSafe playground or POST it to https://api.typesafe.ai/v1/systemone with Authorization: Bearer $TYPESAFE_API_KEY.