開發工具與代理平台
Foundry Toolkit 1.6.13 Adds Support for Non-ACR Private Container Registries and Changes the Windows ML Runtime
The new VS Code toolkit can deploy hosted agents from private container registries configured with a connection, while also adding an AMD NPU build of Fara 7B and a quantized Aion adapter. Windows ML has been upgraded to `windowsml` 2.3, requiring existing projects to review their execution providers and dependencies.
Microsoft released Foundry Toolkit for VS Code 1.6.13 on September 16, updating several paths involving agent deployment, on-device models, and fine-tuning artifacts. The change with the greatest deployment impact is that hosted agents can now use prebuilt private images stored outside Azure Container Registry, provided that the Foundry project has a matching registry connection. Microsoft’s documentation for external registries describes using OIDC to exchange short-lived tokens, avoiding the need to store usernames, passwords, or long-lived access tokens in configuration.
This does not mean that any private image can run without modification. Foundry hosted agents still require `linux/amd64` containers, while network isolation, image-pull permissions, and RBAC for the project identity must be validated separately. For regulated teams, the value of the new path is that they can retain existing supply chains based on JFrog Artifactory, self-hosted Docker Distribution, or similar systems without copying images to ACR solely for the agent platform. However, CI pipelines must add tests for OIDC, image signatures, and failed image pulls.
For on-device workloads, the release adds a VitisAI INT4 build of Fara 7B with support for Windows x64, AMD NPUs, and image input. The Aion fine-tuning workflow can now download `quantized_adapter.safetensors` while retaining support for the legacy `adapter_model.safetensors`. The former also requires the package to include the corresponding template, so moving only the adapter file may not be sufficient to reconstruct the workflow.
Windows ML requires upgrade testing: the toolchain now uses `windowsml` 2.3 and updated execution providers, while removing the required dependency on the Windows App SDK. The release notes explicitly state that existing projects require changes, rather than describing the upgrade as a completely seamless replacement. Version 1.6.13 also fixes issues in Agent Inspector that caused output or errors to disappear after a response completed, as well as problems in which tool-approval and sign-in steps interrupted continuation. Engineering teams should now pin the extension version and separately run regression tests for local NPU execution, adapter downloads, and private-image deployment. Microsoft has not yet published end-to-end performance or failure-rate data for these new paths.