yuyaoge/doc2harness is the official code repository for the Doc2Harness paper and evaluates whether a model can build the harness of a coding agent itself. A harness is the execution setup wrapped around an agent; in this repository it is precisely the thing the model has to build by itself. The repository is a partially open-source release that publishes 94 scenarios, and running it requires Python 3.11+ and Docker.
The key point of this repository is that the evaluation axis itself is new. It measures not how well a model uses a patch, but how well it builds the harness of a coding agent. Step 1 runs OpenCode in an isolated container and has the model implement a harness on top of a skeleton that contains only development documentation and interfaces. Step 2 freezes the finished harness and then checks whether a scripted mock model runs end-to-end in a separate container. The observation channels are explicitly specified as four: model requests, files on disk, exit codes, and command output, and each scenario comes with a control scenario that filters out empty passes.
That said, the boundaries are clear. The license is marked NOASSERTION, so no formal license can be confirmed. The figures for the 19 models on the leaderboard cannot be reproduced from the publicly released portion alone, and the README itself states that the remaining cases are not public. As an axis for examining a model's ability to build a harness, it is worth consulting for now, but it is too early to argue about model rankings based on the leaderboard numbers.