Three weeks of work (225 commits) never made it onto the roadmap. This page catches you up in plain language, then asks you to make five calls and set the ship order.
All four security-hardening plans were actually built back on June 12. But nobody ever ran the final check to prove they work on the live site, and you never signed off. New problem found this week: right now, if a word gets deleted or re-processed in the background system, it can silently delete a real learner's saved notes and study progress along with it — because the database is wired to cascade the delete automatically. This is cheap to fix (flip a database setting) and expensive to ignore (a real user could lose their progress with zero warning).
Big improvement burst here. Rating a card used to take 4–5 seconds to register — now it's instant (about 1/50th of a second). We removed 16 leftover settings that didn't do anything. Study timers now only count time you're actually active, not idle time. Hint counts are now honest (they used to lie). The little "you're learning this" badges work. What's left: one plan that needs your go-ahead (a redesigned end-of-session summary), the hint-image experiment (see #4), and you saying "yes, this is done."
The big goal — the pipeline running by itself in the background with no babysitting — has been stuck since mid-May. But: we found your own answers to a checkpoint from 2026-07-02 sitting in the database, never read into planning. You said clearly: yes, build the "top up any field for a group of words" tool; keep it on this MacBook Pro only for now (you're traveling); use your Claude subscription, not a paid API; try free Google AI first before anything that costs money. About 10 of your answers on that same checkpoint said "too technical, explain again" — so the next write-up on this topic comes with plain language and a picture of how it flows, no exceptions.
The old plan (using a tool called Ollama to make images) is dead — it broke back in April and was never fixed. The new plan is already running: a different tool (Draw Things, with a model called Z-Image Turbo) is generating sample hint pictures on your Mac right now. It's stuck on two clicks you need to make on your machine to turn on a setting and download the model.
This is genuinely new territory. We built the plumbing that lets a "card" be something other than a vocabulary word — like a grammar rule. Nine test grammar cards are already live on your own profile (they cover one small pattern in Indonesian). Mechanically they worked, but you told us afterward "it didn't actually help us learn" — so we're already redesigning them (third attempt, happening in a separate work session right now). Separately, we mapped out the whole Indonesian grammar curriculum: 126 topics, grouped into 19 families, checked against 4 grammar textbooks. 45 of those topics (all about verb prefixes) are ready to write cards for today.
A separate work session is redesigning the page where you browse all your cards — better handling for paused/retired cards (with automatic un-pausing), and better filters. Nothing written down yet; it'll get its own page here once there's something to show you.
What do you want to see on a regular basis — a weekly one-page roadmap summary like this one? Something shorter? And separately: what do agents keep getting wrong at the planning level (not the code level) that you'd want caught earlier?