little progress report :
the Codex App (now ChatGPT App but fuck it) is just an Electron app
first thing I did was run it in bundled mode : just provide the original app.asar to the exact used Electron version + provide all external assets (codex.exe, computer use stuff, …) this is what yielded the previous screenshots, as the debug stuff is just included in the final payload lol (I have the past 70+ versions backed up just in case it gets removed later, so we could restore it)
second step was running in unbundled mode : extract all assets from the app.asar and run from those files directly. this unlock proper dev and hot-reloading later on
but Rolldown-minified JS files aren't really useful (although might stay used for later patching with shims just pointing at the resulting function/variables aliases… will figure that out later)
next step was properly reverse all that shit into readable code
a custom classifier have been written (thanks the Oxc team for providing such fast tools 🙏) to separate resolvable dependencies from actual OpenAI-written code. this drops the number of .js files to process from 19k+ to 4.5k+
2 tools are then used to speed up the process (takes around an hour to run on all those assets). I might run one of them in a more aggressive mode and the second with a proper sandbox to obtain better results
a third tool is then used to transform minified code into proper function and variable names (well, a custom reimplementation of a tool I randomly found out). this uses a shit ton of LLM subagents to do the work. if you need an estimate, with the current modules splitting between 5.6 luna, terra and sol (a more economic split wasn't done by an abundance of caution), you're looking at roughly 60 million tokens spent, which equates to roughly 250$ at current pricing (without counting the validation that will happen afterwards)
although I'm also working on a way to make that cost massively reduced (<1$) for subsequent updates by just processing the delta of recovered files. messing around with using the new DeepSeek V4 Flash could also lower the pricing tbh
anyway I'm currently halfway through this last step (which accumulated lasts for 50+ hours non-stop) and will then be able to work on the actually modding SDK 😈
l already have multiple mods I wanna make, including a performance optimization one and a multi-profile one
I'm also working aside for @juznem, making a super cool bot. it's in the last stages of development and I can't wait to share more about it later 😋