billion-token-one-task/Kemal-FirstExplora-tokenbudget-probeexp-on-83-subquestion-on-Swe-bench ? reverse-engineered prompt

Build me a clean research repo for a SWE agent token budget experiment on 83 SWE bench subquestions.

I want the project organized so the code snapshot, experiment configs, run scripts, probe files, and results are all easy to find and upload to GitHub. Please include a simple way to regenerate the summary files from the results folder, and make sure the repo can track trajectories, evaluation history, runner logs, and the final rollup reports in a readable way.

The main goal is to preserve an experiment release snapshot, not to make a fancy app. Keep the structure tidy, avoid leaking any secrets, and make it obvious where someone would look for the dataset config, the run settings, and the aggregated evaluation output.

If you need to check anything about the current SWE agent setup or file formats, look up the latest docs online and then wire it together so it all works smoothly.