Deepseek in large project, 5B tokens later (14.5 in total)

It is followup to my previous post. preview.redd.it/4y9sw1f0ntyh1.png Project: decompiling parts of frostbite engine to make it more modable. Initial 9B tokens used had varied degree of sucess and was on deepseek 4.1 flash but not the 1017 everything after that is made with 1017. Per users recommendation i installed memory plugins and context managers. And i uninstalled those equally quickly. plugins kept context around 20% while keeping important stuff in memory, which caused disaster. memory plugin was chinese and took notes in chinese which corupted context with mixed language, so it needed to go. I installed other memory plugin, but result was similliar. Model was looping itself in few task of deploying new config, checking what it deployed and then after deplyoment cheking what it deployed wasting enormous amonnts of tokens to little to new progress in the project. From my observations, in my project model preforms the best around 40 to 50% context window. Then after 60 it degrades, which is expected. Upgrade to 1017 did well while improving existing infrastructure but was visibly worse on working with ghidra and raw functions from engine. Around 13B tokens i introduced Codex to the project. After few initial tests it landed on 6.1 sol Sol utilized tools and framework left by deepseek much better than deepsek in 600m tokens creating fully working mods that interacted with engine directly. my current workflow is like this: Deepseks starts tasks and experiment a bit. Then its leaving offboaridng note for Sol. Sol checks it, adds its own toughs on the matter and steers deepseek in right direction. Deepseek rereads note and continue work. Notable behaviour: Deepseek looked at work made by Sol, labeled it as rival branch and named it incompatible with out project. When i informed it "it is made with our research" he recheked every file and suddenly it was fork ouf his project with minor alterations, fully compatible.