Foundry initialization improvements

• Initialize a client directly from an endpoint, deployment name, and explicit
  TokenCredential, including Azure CLI and managed identity credentials.
• Use a Foundry resource root or project endpoint for SystemOne. Both resolve to
  /providers/microsoft/v1/systemone, with or without a trailing slash.
• Preserve custom invocation paths and query strings. Azure OpenAI Decisions
  retains its existing routing and rejects unsupported Foundry project endpoints.
• Keep options-based construction, DI registration, and caller-managed HTTP
  transports, with documented ownership and disposal.
• Lead quickstarts with endpoint-first initialization and verify the new API in
  both offline unit tests and exact-version NuGet package consumers.

All four packages are synchronized at 0.8.0:

Validation and limitations

The immutable release source is
dc46fc812629723ea2f20ab785bd73ae29a67780. The protected
publication workflow
builds and tests on Linux and Windows, enforces formatting and coverage gates,
runs every offline sample, and checks all four packages, symbols, Source Link,
and isolated exact-version consumers. Local validation passed 615 tests; five
opt-in live tests remained skipped.

All four packages and adjacent symbol packages were accepted by NuGet.
Exact-version restores from the public feed passed in fresh isolated caches,
including offline consumer execution and matching the validated assembly hashes.

Tentative early access: this release adds no new live service verification.
Live Jev compatibility remains unverified. Offline checks do not establish model
accuracy, calibration, or production readiness. NuGet classifies the
suffix-free version as stable; GitHub marks this release as a prerelease to
communicate its tentative status.

Fixes #7.

Adds Azure OpenAI Decisions support for gpt-6-luna while preserving Microsoft-Decision-1's existing SystemOne API and the provider-reported token usage introduced in 0.6.3.

What's new

• Explicit FoundryDecisionApi.OpenAIDecisions mode for /openai/v1/decisions. The default remains SystemOne.
• Typed DecideAsync requests with text or mixed text/image messages and 1-400 uniquely named Predicate, Choice, and Score questions.
• Typed answers correlated by question name, complete distributions, per-question refusals, optional model metadata and token usage.
• Existing IDecisionClient methods remain text-only and one question per call. OpenAI convenience calls preserve shared DecisionUsage.
• Expanded runnable sample with separate deployment configuration, both offline flows, batch/image examples, distributions, refusals, and usage.
• Step-by-step setup, image data URL instructions, migration guidance, and opt-in live-test commands.

Validation

All three Foundry integration tests passed against live deployments using Azure CLI/Entra authentication on October 26, 2026: Microsoft-Decision-1 Choice/Assess/Score and gpt-6-luna multi-question Predicate/Choice/Score, token usage, and synthetic PNG image input.

Linux and Windows publication validation checks builds, formatting, offline tests, coverage, samples, package metadata, symbols, Source Link, and exact-version isolated consumers. Publication additionally verifies exact-version restores from public NuGet.

Publication and verification run

Packages

The four packages use synchronized version 0.7.0:

• ElBruno.AI.Decisions
• ElBruno.AI.Decisions.Foundry
• ElBruno.AI.Decisions.Ollama
• ElBruno.AI.Decisions.Jev

Early-access limitations

Live Jev compatibility remains unverified. Live Foundry checks establish protocol compatibility, not calibration, semantic accuracy, load behavior, or production readiness. Live refusals and external HTTPS image downloads remain unverified. NuGet classifies 0.7.0 as stable because it has no prerelease suffix; this GitHub release remains marked prerelease to retain the tentative early-access notice.

Foundry walkthrough | Testing scope

Adds provider-reported token usage to Choice, Score and Assessment results across Foundry, Ollama and Jev. Existing constructors remain available; missing counts remain null, reported zero is preserved, and totals are never inferred.

NuGet 0.6.3: core, Foundry, Ollama, Jev.

Publication and exact-version public-feed consumer verification succeeded in run 38644542842. Linux and Windows validation includes tests, formatting, coverage, samples, package metadata, symbols and Source Link.

Tentative early access: live Jev and local Ollama compatibility remains unverified. Token counts across models do not imply equivalent computation or cost. NuGet classifies 0.6.3 as stable because it has no prerelease suffix; this GitHub release is marked prerelease to retain the early-access notice.

Fix: #5. Addresses #4.

Local Laya inference

• Adds explicit loopback-only local Laya configuration with optional bearer authentication.
• Omits an unspecified model so Laya can auto-route; model discovery now fails explicitly because Laya does not expose it.
• Adds local Laya wire-contract and configuration tests plus deployment guidance.

Tentative early-access release

Live Jev service compatibility remains unverified. This release is explicitly
authorized for early evaluation while live testing is deferred. Offline tests
and package checks do not establish production readiness or model accuracy.
APIs may change before 1.0.

The package version is exactly 0.5.0. NuGet therefore lists it in the stable
version channel, despite its tentative status. This GitHub release is marked
prerelease to make the early-access intent visible.

Install

dotnet add package ElBruno.AI.Jev --version 0.5.0

NuGet package

Included

• .NET 10 native Choice, Score, and Noul decisions, structured JSON, typed question handles, model discovery, and native metadata.
• Microsoft.Extensions.AI function tools, routing, and assessments around caller-supplied chat providers.
• Bounded HTTP responses, cancellation, conservative retries, explicit ownership, and per-request credentials.
• Ten runnable samples with explicit synthetic offline mode.
• Package README, XML documentation, MIT license, icon, symbols, and immutable Source Link.
• README badges for NuGet version/downloads, CI, release workflow, .NET, license, and tentative/live-validation status.

Jev is a typed decision service, not a native chat, embedding, image, or speech
generator. This community SDK targets the official TypeSafe API, not the
independent jevtypesafeai.com proxy.

Verification

355 offline tests passed. GitHub-hosted Windows and Linux checks also ran every
offline sample and an isolated exact-version package consumer. macOS passed
build and unit-test smoke checks. Live service tests stayed disabled.

• Cross-platform CI
• NuGet publishing workflow
• Successful public-package verification

NuGet accepted the package and symbols successfully. Initial install-feed
indexing exceeded the publication workflow's short wait window. Once indexed,
an isolated public restore was verified against the original published DLL and
symbols, both locally and in the recovery workflow. Nothing was republished.