datacurve-ai/deep-swe ? reverse-engineered prompt

Build me a benchmark project for evaluating coding agents on real software engineering tasks from open source repos.

I want a simple repo that organizes each task with a prompt file, a config file, a Docker based environment, held out tests, and a hidden reference solution, so the grader can verify the agent’s code by behavior instead of exact file names. It should support running one task or a random subset of tasks, and it should work in isolated sandboxes with a separate verifier step that outputs structured results and test logs.

Make it easy to run locally with a CLI, and make sure the setup is clear enough that I can point it at a folder of tasks and have it evaluate an agent end to end. If you need to check current docs for the task format or runner, look them up online.