.NET: [Feature]: Expose the MEAI IEvaluator -> IAgentEvaluator bridge (AsAgentEvaluator)

[ANcpLua]

Description

Microsoft.Agents.AI already contains exactly the right bridge between the Microsoft.Extensions.AI.Evaluation ecosystem and agent evals — MeaiEvaluatorAdapter (dotnet/src/Microsoft.Agents.AI/Evaluation/MeaiEvaluatorAdapter.cs) wraps any MEAI IEvaluator into an IAgentEvaluator, runs it per item, handles the RawResponse-vs-bare-text shapes, and aggregates results. But it is internal (verified on main @ 7b6d25798), so every consumer who wants to run an MEAI evaluator (the bundled quality evaluators, the safety evaluators, or their own IEvaluator) through LocalEvaluator-style agent pipelines must re-implement the same ~40-line adapter verbatim.

• What problem does it solve? Removes a copy-paste seam between the two first-party evaluation ecosystems. MEAI has a rich, growing IEvaluator catalog; agent-framework's eval pipeline speaks IAgentEvaluator. The bridge exists, is tested by usage, and is the obvious connector — it just isn't reachable.
• Expected behavior: any MEAI IEvaluator becomes an IAgentEvaluator in one call.
• Alternatives considered: keep copying the adapter into every consumer (status quo — drift risk per copy); make the class public without an extension method (works, but an As* extension matches the framework's existing AsAIAgent / MEAI AsChatClient idiom).

Proposed API

namespace Microsoft.Agents.AI;

public static class AgentEvaluatorExtensions
{
    /// <summary>Wraps an MEAI evaluator so it can run in agent evaluation pipelines.</summary>
    public static IAgentEvaluator AsAgentEvaluator(
        this IEvaluator evaluator,
        ChatConfiguration chatConfiguration);
}

Implementation is a one-line delegation to the existing MeaiEvaluatorAdapter, which can stay internal. Purely additive — no behavior change for existing callers. Happy to send the PR (extension class + XML docs + one adapter test; no changes to MeaiEvaluatorAdapter itself) if the shape is acceptable.

Code Sample

using Microsoft.Agents.AI;
using Microsoft.Extensions.AI.Evaluation;
using Microsoft.Extensions.AI.Evaluation.Quality;

IAgentEvaluator quality = new CoherenceEvaluator()
    .AsAgentEvaluator(new ChatConfiguration(judgeChatClient));

AgentEvaluationResults results = await quality.EvaluateAsync(items, "quality-gate");

[ANcpLua]

Two data points that surfaced after filing, for whoever triages this:

Prior art / demand evidence. AgentEvalHQ/AgentEval#66 (merged) ships exactly this extension in userland — AgentEvaluatorExtensions.AsAgentEvaluator(chatConfiguration) — because the in-box adapter is unreachable. Same name, same shape, built outside the framework because it couldn't be imported.

One design note worth considering when exposing the adapter. With the default LastTurn splitter, MeaiEvaluatorAdapter.EvaluateAsync forwards only the query side of the conversation (var (queryMessages, _) = item.Split();) and the fallback bare-text response keeps only assistant text turns — so unless RawResponse is set, tool-call content never reaches the wrapped MEAI evaluator, and tool-selection-style evaluators read 0%. The AgentEval PR above hit this in practice and worked around it by forwarding the full conversation in its own IAgentEvaluator. Exposing the adapter as-is is still useful for text-quality evaluators; either an option to forward the full conversation or a doc note on the splitter behavior would round it out.