Probabilistic planning · 2026 · Solo — trend research, ensemble and time-zone architecture, test-first implementation, accessibility, security hardening, production verification
Weather Window Booker
A forecast-to-calendar operator that finds continuous daylight windows from live ensemble uncertainty, then leaves the final save with the person doing the work.
The problem
Outdoor work is easy to describe but awkward to schedule against changing weather. A single forecast icon hides model disagreement, partial-hour timing, daylight, and the risk that a plausible slot no longer satisfies the operator's bounds. Weather Window Booker geocodes one real place, reads live Google WeatherNext 2 ensemble members through Open-Meteo, and converts explicit temperature, wet-member, wind, daylight, and duration constraints into one reviewable Google Calendar handoff. It is planning guidance, not a severe-weather or safety service.
Architecture
Key decisions
Align evidence to the booked interval
Open-Meteo stamps precipitation at the end of the preceding hour. The aggregator deliberately shifts that member set back onto the interval it describes and drops the final unverifiable slot, so a two-hour booking evaluates both hours rather than missing the last one.
Bind asynchronous work to one place
Every search and forecast advances a monotonic request sequence, while forecast completion also checks the selected place identity. Starting a new search clears stale rail, ticket, path count, and retrieval metadata; a late response cannot silently mix one city's model data with another city's calendar location.
Make uncertainty and accessibility part of the output
The rail reports each hour's temperature, wet-member percentage, wind bound, daylight, and Fits, Out, or Selected state in visible text and an accessible label. The production matrix covers desktop, mobile, short height, reduced motion, opposite color preference, no JavaScript, keyboard focus, contrast, overflow, and zero console or page errors.
Publish exact release evidence
Functional SHA 8eed40601582cf7e70cb5f063e4adb7a02901955 passed 27 unit and release-contract tests, lint, TypeScript, English-only, production build, zero-vulnerability audit, eight production-preview browser checks, and an independent review after two blocker-fix cycles. Push-trigger evidence SHA 754c6b928112e40e696791783b23e0b1d39ce56b received successful GitHub deployment 5846082740 and Ready Vercel deployment dpl_2SHZUvvR88P718pa84VJ4ZjinNXv; the same eight checks then passed against production.
Translate demand without cloning
Secure radar run 20260810T214614-0700-4bb330d82cfd observed PrimeIntellect-ai/prime-agent at daily rank 1 with 2,642 stars today and 13,291 total on 2026-08-10. WeatherNext ranked ninth with 325 stars today and provides secondary forecast-feed inspiration. The product translates demand for bounded autonomous completion into a weather-to-calendar operator without copying either source's code, architecture, interface, assets, or branding.