startrail-org/leann is a RAG application that runs on personal devices. RAG is a method that uses retrieval to find external knowledge and add it to the model's answers. The repository description states that it saves 97% of storage space and runs 100% locally, and links the paper as arXiv 2506.08276. The latest commit of this project as of October 28, 2026 (07796bd, https://github.com/StarTrail-org/LEANN/commit/07796bd7e40f2d5f3a4be9e45adf1f62a68481fd) did not change library code, but changed only one release automation workflow.
The file modified was .github/workflows/release-manual.yml. This workflow, which starts a release by hand, queries the run records of the build-and-publish workflow based on the commit SHA of the main branch, checks whether that run ended successfully, and then proceeds with the deployment. What changed is that this check is no longer performed only once but is repeated. Previously, if the relevant CI run record had not yet been created or was still in progress, it failed immediately; now it waits up to 4 hours (14400 seconds). If there is no run record, it queries again every 15 seconds; if the run is in progress, every 30 seconds. When it completes successfully it moves on to the next step, and when it completes with a failure it terminates. If it does not end successfully within 4 hours, it is treated as a failure. The change is limited to this single file, with 38 lines added and 25 lines deleted.
What this commit touches is only the waiting behavior of the release procedure. There are no feature additions, no changes to library behavior, and no changes to the version number, and the version deployed this time is not given. The 97% reduction and the MLsys2026 Best Paper notation in the repository description are claims by the project side, and are not figures independently confirmed here. There is no material on how long main CI usually takes to finish, nor on how much this waiting actually reduced release failures. It is reasonable to read this as a fix that changed the procedure in a stable way in a situation where deployment automation was intermittently failing.