@FedShved

I set out to work through a fixing plan for my German study bot. The streak
needed to require a real daily minimum instead of counting a single card, I
wanted timed nudges that push an actual recall task into the chat at set times
so the studying comes to me, and the review screen had a backlog of problems:
multiple-choice options pulled from other cards’ notes so the four answers were
incoherent, a cloze card that printed the answer while you were supposed to be
recalling it, and rating buttons labelled so tersely they meant nothing.

I built all of it. The streak now counts how many cards you answer each day and
only advances once you reach a goal you set with /goal, and it shows today’s
progress everywhere. Nudges are a new module that registers one scheduled job
per time you choose; when one fires it takes a due card, turns it into a
translate or fill-in-the-blank task, sends it inline, and grades whatever you
type back. Multiple-choice options are now filtered down to clean English
meanings, the answer no longer shows before you try, “Again” became “Forgot”
with a short legend, and I added a plain word-only review mode alongside the
existing mixed ones. The change set stayed under tests the whole way, 54 up
to 95.

The real problem only showed up once I used the bot: every card took roughly
half a minute to appear. Nothing looked broken, so I read the logs, ruled out
my own code, and timed the Gemini calls directly. One enrichment call was
taking 27 seconds. I benchmarked the configured model against a few
alternatives on the same request, and gemini-3.5-flash came back around 24
seconds while gemini-2.5-flash answered in about one. The slowness was never in
the bot, it was the model set in the config, and changing that single line took
review from unusable to about a second a card. A smaller fix came out of the
same session: a forgotten card used to reappear the same day, which is pointless
since you just saw it, so it now schedules for tomorrow, and because that made
two of the new-card buttons read the same one-day gap I pushed “Hard” out to
two days.

I set out to make the keyboard actually look like the design I’d settled on.Everything worked by then but it still wore a generic light theme, and the wholepitch is a keyboard that looks and types like Samsung’s, so the look had to land. Ipulled every color and a few sizing numbers into one theme object and repaintedthrough it: a black keyboard, near-black keys, white letters, and red used for onething only, the key you’re pressing. On top of that I added the pieces the mockupshowed — a floating preview of the letter above your finger, a permanent numberrow, a dim symbol in the corner of every key that you get by long-pressing, atwo-page ?123 layer, a blue enter key that turns into a magnifier in search boxes, arow of toolbar icons up top, and holding the spacebar to slide the cursor throughtext.
The part I was wary of was the number row, because it doesn’t just sit on top, itgoes into the same grid the autocorrect uses to read where your taps land. Adding a fifth row shifts every letter’s position, and the worry was that it would quietlythrow off the typo-fixing and start mangling words. It turned out not to matter: thedecoder measures distances in key-widths rather than pixels, so adding oneevenly-sized row leaves the spacing between letters exactly what it was, and thedigits aren’t letters so they never enter the word search at all. The offline testthat runs thousands of fake typos came back with the same numbers as before thechange, false corrections still at zero, which is what let me put the row in the reallayout instead of bolting it on as a separate strip.
The screenshot shows the new black keyboard with the number row on top and a pressedkey lit red with its letter popped up above it.

Note, found while testing the redesign and not caused by it: two things in the decoder are worth fixing later. The bar offers “дом ас” when you type “домас”,because the run-together splitter will break a word in two as long as both halves are real, and “ас” happens to be a real word, so it peels it off even though that’s almost never what you want. And “домасний” won’t correct to “домашний”: that’s aс-for-ш mistake, but с and ш sit on opposite rows of the keyboard, and the auto correct only ever considers the key you hit and its close neighbors, so it never even weighsш as a possibility. The first is a small tuning fix; the second would mean teaching it to consider letter swaps beyond fat-finger range, which is a real feature and has to be done without breaking the zero-false-correction rule.
( the toolbar icons along the top aren’t wired up yet. The language one switches keyboards, but emoji, clipboard, settings, and the overflow menu are drawn for now and don’t do anything — they’re waiting for future wiring)

I wanted to close a gap that felt wrong in real use: autocorrect only ever looked at the word being typed right now. If a typo had already landed in the text and you tapped back into it, the keyboard just sat there and wouldn’t re-check the word or offer the right spelling. An earlier pass had made tapping into a word safe so it no longer mangled it, but the part you actually expect, tap the broken word and get the fix, wasn’t there yet. So now when the cursor lands inside a finished word the keyboard adopts it: it works out where the word starts and ends and, since the original taps are long gone, recreates them by dropping one tap dead-center on each letter’s key, then runs the same autocorrect search it uses while typing. The correct spelling shows up in the bar and one tap swaps it in where the word sits, without disturbing the spaces or words around it, and you can also just backspace and retype and it re-checks live. It never changes the word on its own - tapping in only offers a fix -which keeps the rule the whole project lives by, never correct a word that was already fine, intact at zero false corrections. The tricky part was telling apart the two times the cursor moves - when I move it, since right after committing a word the system reports the cursor now sitting after it, versus when you tap somewhere new. Treating my own cursor updates as a fresh tap made the keyboard re-grab and re-offer the word it had just finished, in a loop. I fixed it by remembering exactly where I last put the cursor and only reacting when it turns up somewhere I didn’t put it. The other half was being conservative about which words to touch: only plain Russian words, never names, numbers, or anything with characters that don’t exist on the keyboard, and never a half-typed fragment.

Finally I was also able to fix a bug word-tearing. At first I had marked it done, but the exact thing that happens when you edit a word still broke. Type a word, press space, backspace into it, keep typing, and the keyboard treats the new letters as a separate new word and splits the old one in half. So I reproduced it on the emulator in the Settings search box, the field from the screenshots, and changed how typing into existing text works.
Now when you start typing with the cursor inside a word the keyboard adopts that whole word and edits it as one unit instead of starting a loose fragment beside it, and I gave it a cursor index within the word so an inserted or deleted letter lands where the cursor actually is instead of always at the end. What slowed me down was that clean scripted taps would not tear the word no matter what I tried, which is the same trap I have fell into earlier. The bug was never in the path I kept testing. It only shows up once space has committed the word, and my reproductions weren’t pressing space first.

Only after many many tries I was able to finally understand the logic of the bug. Space, then backspace, then type. With space, the word is committed text, so backspacing into it does plain deletes and the old code then composed only the new letter as its own fragment. Replacing that “cursor sits inside text” path with whole-word adoption fixed it, and false corrections stayed at zero the whole time since an adopted word is only ever offered, never swapped on its own.

What I set out to do. Build the core of the keyboard: a full Russian ЙЦУКЕН layout that doesn’t just type letters, but fixes fat-finger typos as you go. The whole reason for this project is that Samsung’s Russian autocorrect is rough, so this part had to genuinely work, not just look the part.

Drew the keyboard as a custom Canvas view instead of a UI framework, so it stays small and opens fast on old phones. Then wrote the autocorrect engine - it treats every tap as a point on the screen, not a fixed key, and runs a search over a word dictionary that scores how well your taps match each word against how common that word is. On space it decides whether to swap your typo for the best match. Never touch a word that’s already spelled right.

At first the system said the keyboard was visible, the process was running, no crash in the logs, just a thin blank strip at the bottom. I lost a good while on this. Turned out Android measures the keyboard once with a width of zero before the real layout pass, and I was computing the keyboard’s height from its width, so zero width meant zero height and it never recovered. Fixed it by falling back to the actual screen width when the measured width comes back as zero. Keyboard popped straight up after that.

(to be honest there are still problems with this system which i aim to soon fix, like for example starting mid word, or not spaced word. But at least I was able to get it to work in the most ideal conditions)

I wanted people to actually be able to control the mod. Right now if you want to turn off auto-disconnect you have to dig into a .toml file, which nobody wants to do mid-session. So the goal was an in-game window you open with a key.

What I built: press K and a small window pops up with toggles for the main settings (auto-disconnect, screenshots, alerts). Flip one and it applies right away, no restart. I also added two warnings while I was in there: one that tells you auto-disconnect can get you flagged for combat-logging on PvP servers, and one that warns you screenshots can come out black if you’re in Fullscreen.

I wired up the keybind and pressed K, and nothing happened. Took me a while to realize Minecraft registers keybinds on a completely separate event channel than the one I was using for everything else, and the shortcut for it was deprecated on top of that. Once I moved the registration to the right place it worked first try.

Furthermore, added precautions is the user of the mod will go on a server (there might be problem due to pvp mods banning people after disconnecting after a hit)

Next idea that I might soon working on is covering support for more versions and also expanding the potential places for the bots to work with (discord)