I built a deterministic "research & generation sequencing" pipeline in the services/generation folder that orchestrates eight steps from requirement extraction to a printable one‑pager, with progress reported to the frontend via onProgress callbacks.
Each step's onProgress callback feeds the progress array on the Kit document, which the frontend polls and renders as a step list.
Minutes per question are calculated as 15 * difficulty, yielding 15 min for difficulty 1, 30 min for difficulty 2, and 45 min for difficulty 3.
Schedule chunks are allocated with the formula Math.floor((day - 1) * total / daysAvailable), ensuring each of the days_available days receives a contiguous slice of the sorted question list.
When a user edits or adds a question, its _meta.pinned flag is set to true; regeneration (regenerateQuestionCategory) then excludes any pinned items from the model request, sending only uncovered requirements.
A two‑pass coverage check runs after the initial generation; a second narrowly‑scoped generation call runs only if any must‑priority requirement is still uncovered, and the loop is capped at two passes to prevent endless retries.