mupt-ai/self-bench is a project that benchmarks coding agents and models with a private, Harbor-style evaluation built from a repository's merged pull requests. The production release prod-20261015-44ef8ca fixed an issue where creating new jobs failed in environments where no image builder was configured. According to the release notes, it now re-validates the source build in those environments as well. This change restores the path that previously made it impossible to create jobs without an image builder.
The project's evaluations are built from changes merged in the user's repository, and they are a private evaluation set rather than a public one. So the focus is on evaluating agents and models against the change history of a specific repository, rather than on a general test of ordinary coding ability. This fix also does not change the evaluation method or how scores are computed; it re-enables the source build validation path required for job creation. The release notes describe new jobs failing with a message stating that a configured Modal image builder is required.
The confirmed scope is limited to restoring job creation in environments where no image builder is configured. Because the release summary does not explain whether other build environments or evaluation results changed, this fix alone cannot be taken to mean that benchmark performance or the scope of the evaluation changed. The repository description states that evaluations are generated from merged pull requests, but the change covered by this release is not the evaluation content but the build path needed to run it. Therefore, it is more accurate to view this release as restoring job creation that had been blocked by the absence of an image builder, rather than as adding new evaluation features.