[Feature request] Support decision/classification models (e.g. Microsoft-Decision-1 on Foundry) as a lightweight engine or pre-agent step #74832

Several workflows use a full agent run only to classify an input, such as triaging an issue into one of N labels. In dotnet/aspnetcore, issue-triage-agent runs a full agent job plus a detection job for every new issue. A recent run consumed about 22 AI credits, roughly $0.22, and took about 4.5 minutes. Most of that was the agent job writing a ~70k-token prompt to the cache, which holds the area reference.

Decision models are built for this kind of work. They take a state and a fixed set of typed questions (yes/no, choice, score) and return calibrated probabilities in one pass, without generating text. For example, Microsoft-Decision-1 on Microsoft Foundry costs $0.042 per million input tokens, and output is free. In a quick test, classifying real aspnetcore issues with their full title and body took 340-610 input tokens, about $0.00002 per call, in under a second.

Current workaround

I can call the model from a custom jobs: entry or steps: block with curl and a repo secret. That works, but every repo has to reimplement it. It also has these gaps:

• No first-class way to feed the probabilities into the prompt, safe-outputs, or an if: condition.
• No built-in confidence-threshold gate that skips the agent job.
• No usage or cost reporting. The Copilot engine's token-usage.jsonl and AI credit accounting don't cover these calls.
• Authentication and firewall setup are manual (a Foundry key or Entra OIDC, and the Foundry domain).

Proposal

Any of these would help, in rough order of effort:

1. Frontmatter for a decision step, for example:

   decisions:
     provider: azure-foundry
     endpoint: ${{ secrets.FOUNDRY_ENDPOINT }}
     deployment: decision1
     questions:
       area:
         type: choice
         instructions: Which area owns this issue?
         criteria:
           area-blazor: "..."
           area-mvc: "..."

   with typed outputs (needs.decision.outputs.area, .confidence, .probabilities).

2. A confidence gate such as skip-agent-if: area.confidence >= 0.95, so that high-confidence cases can go straight to safe-outputs without an LLM run.

3. A non-generative engine type, or provider support for OpenAI-compatible and Foundry endpoints, with usage reporting in the run summary.