Repository navigation
Releases
v0.4.0: Safe local decision model support
Decision model support
- Adds FoundryLocalDecisionAdapter with default and per-call model selection, including cached KEV models.
- Resolves the Foundry Local decision/typed-decision task mismatch through a namespaced, version-preserving local registration. Registration references the existing cached assets without copying them.
- Rejects public/local identity collisions, conflicting paths/tasks and package alias overrides; supports exact-registration reuse and concurrent registration.
- Adds optional ModelCacheDir to reuse the Foundry CLI cache.
Compatibility
Retains .NET 10 and .NET 11 on Windows x64, Microsoft.AI.Foundry.Local 2.4.1 and Microsoft.Extensions.AI 11.0.0 from 0.3.0. Chat and embedding public contracts remain unchanged. The SDK's existing in-process chat/embedding API deprecations remain documented.
Verification
31 tests passed on each target framework; PR and merged-main build/test/pack CI passed. Cached KEV CPU inference was verified locally, including shared model paths and unload after use. GPU/NPU inference was not verified. Existing registrations with uncertain ownership were preserved.
NuGet publication uses the repository's official Publish to NuGet workflow and OIDC trusted publishing.
v0.3.0
Supports .NET 10 and .NET 11 with Microsoft.AI.Foundry.Local 2.4.1 and Microsoft.Extensions.AI 11.0.0. The adapter remains Windows x64; the SDK in-process OpenAI clients are deprecated and execution-provider registration is explicit. Embedding failures are now reported without an unhandled exception.