How it works
Create one GitHub issue per .cpp file and give the agents access to the GitHub CLI and ida-mcp.
Write a document with the goal, the rules and what is forbidden, and re-inject it every hour.
Use a byte-by-byte verification script as the acceptance criterion, protected in CI by a hash.
Scale up with cheaper models and PRs on separate branches once the criterion is stable.
Prompt
Replace anything in [brackets] with your own details.
prompt.txt
Take an open GitHub issue (one per translation unit). Decompile the functions in that file to readable C++ using ida-mcp. Before pushing, run the verification script: a function only counts as done if the compiled bytes are identical to the original executable. Don't use inline assembly, naked functions, object patching or embedded bytes, and don't modify the verification script. Post in the channel which issue you took, and only close it once every function passes.Reported result
99% of the game's functions rebuilt, 83% byte-for-byte identical, and the game runs with no noticeable bugs.
As reported by the source. Results depend on your setup and your data.