Changes to the ManagedCode.SkillOpt .NET port are documented here. This file covers the NuGet package only; the original Microsoft Python project's history remains in the upstream repository.
- Reject negative usage and progress counters in externally supplied resume state before any model or evaluator call, preventing corrupted state from understating spend or replaying completed work.
- Correct the packaged README installation example to reference the released 0.1.3 package.
- Preserve the official evaluator's typed interpretation-failed flag in per-case metric evidence so application quality gates can reject failed interpretations.
- Keep checkpoints limited to scalar optimizer progress; metric evidence remains in selection and split reports.
- Add an optional typed
TargetMessageFactorythat receives frozen case messages and candidate skill text separately for every target rollout.
- Initial .NET package release of the in-process text-space SkillOpt optimizer.