[danielgerlag]

What and why

Adds a new package, WorkflowCore.AI, with one step body, ChatClientStep. It calls a model through IChatClient from Microsoft.Extensions.AI, resolved from DI, so any provider with an IChatClient (OpenAI, Azure OpenAI, Ollama and others) can be used from a workflow. Temporal and Microsoft Agent Framework already offer this pattern for .NET; Workflow Core only had the Azure Foundry specific steps.

services.AddChatClient(myChatClient);
services.AddWorkflowChatClientStep();

builder
    .StartWith<ChatClientStep>()
        .Input(step => step.SystemPrompt, data => "Summarize in one sentence.")
        .Input(step => step.Prompt, data => data.Ticket)
        .Output(data => data.Summary, step => step.ResponseText);

Design choices

• Inputs: Prompt, Messages (IList<ChatMessage>), SystemPrompt, Options (ChatOptions). The request is sent as system prompt, then messages, then prompt as a user message. At least one of Prompt or Messages is required, otherwise the step throws before calling the model.
• Outputs: ResponseText and the full ChatResponse as Response.
• Separate project in src/extensions/WorkflowCore.AI, netstandard2.0 like the other extensions, so WorkflowCore gets no new dependency. It depends only on Microsoft.Extensions.AI.Abstractions 10.10.1.
• No custom fluent builder. The standard .Input / .Output mapping already covers it, so AddWorkflowChatClientStep() (registers the step in DI) is the only helper.
• README in the project folder, linked from the main README.

How it was tested

New test/WorkflowCore.AI.Tests (net6.0, net8.0, net10.0) with a fake IChatClient, no network or keys:

• message order and reply mapping, ChatOptions pass through, messages only, missing input throws without calling the model, null client rejected
• one end to end scenario using WorkflowCore.Testing: a real workflow host resolves the step and the client from DI and maps the reply into workflow data

Results:

• dotnet build WorkflowCore.sln: succeeded, 0 errors
• WorkflowCore.AI.Tests: 6/6 on each of net6.0, net8.0, net10.0
• WorkflowCore.AI.AzureFoundry.Tests: 14/14
• WorkflowCore.UnitTests: 62 per framework; should_run_pre_middlewares (SyncWorkflowRunnerTests) failed on net8.0 and net10.0 in this run. This PR does not touch core, and the same test also fails intermittently on master, so it is pre-existing and unrelated.

Limits

• One request and one response per step. No streaming, no tool calling loop, no conversation store.
• Only one IChatClient per container is used; keyed clients are not supported yet.
• If Response is mapped to workflow data it is persisted with the instance, which can be large.
• A failed model call throws and follows normal step retry, which sends the request again.
• Not tested against a real model provider.