Skip to content

Assistant: structured time windows — model interprets human time, app does arithmetic, no phrase grammar #265

Description

@pliablepixels

Refs #246, follows #262/#264. Live transcript: "summarize last week" was triaged CHAT (the English keyword overrule has no "week"), and even correctly routed, the free-text when contract depends on an app-side English phrase grammar (resolveWhen) that only understands the phrasings someone has hardcoded regexes for. Every new human variant ("past fortnight", "previous month", German "letzte Woche") is another regex.

Inversion: the model interprets human time words into flat structured fields on the event tools (lastCount+lastUnit for rolling windows, daysAgo, weekday, date, fromTime/toTime for day narrowing); the app converts fields to exact start/stop datetimes in the profile timezone and calls the API. Units and weekdays are closed universal sets, so no word lists survive anywhere: interpretation lives in the model (any language), arithmetic lives in code (exact).

Also: a tool-less (triage CHAT/ACTION) turn where the model attempts a real registry tool flips to tools-enabled instead of refusing, making triage advisory and letting the English live-data keyword regexes be deleted. All gated on the eval with new variant-phrasing cases.


This issue was filed by Claude assisting @pliablepixels.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions