ADR-0128: Foundry non-streaming chat uses the existing AEP runtime path

Status: Accepted — 2026-09-17

Context

ADR-0126 introduced an isolated Foundry AEP extension with deployment discovery but no inference. FR-446 must use a selected Model Profile through Agentstration's durable Runtime and Flow paths. The configured inference endpoint may be resource-routed (/openai/v1) or project-routed (/api/projects/{project}/openai/v1), and these routes require different Microsoft Entra token audiences. Runtime Runs default to automatic streaming, whereas this increment cannot stream.

Decision

• Implement bounded OpenAI/v1 Chat Completions in the Foundry extension's AEP ChatAsync contribution. The central Model Provider, IChatClient, Runtime and Flow boundaries remain provider-neutral.
• Send the selected deployment name as model. Support text system, user and assistant messages, successful tool-result history, and the canonical generation options supported by this increment. Reject explicit unsupported options before making an external request; do not silently drop them.
• Use the configured resource or project inference route without automatic redirects. Project discovery and project-routed inference request https://ai.azure.com/.default; resource-routed inference requests https://cognitiveservices.azure.com/.default. API-key mode continues to use the temporary extension-local key pending FR-439/FR-452.
• Normalize text, token usage and finish reason into AEP; keep the selected deployment name as the returned AEP model identity even if Foundry reports an underlying model ID. Treat content-filter and unsupported tool-call responses as explicit failures. Do not expose provider response bodies in errors.
• When Runtime streaming is Automatic and the effective provider/model capability is unsupported, execute non-streaming. Explicitly requested streaming still fails compatibility validation. This rule is provider-neutral.
• Keep request/response sizes and elapsed time bounded. The default offline test suite uses fake HTTP responses and does not call Azure.

Consequences

Foundry can answer an existing text Agent or Flow through the governed AEP/Runtime path without introducing a Foundry-specific execution API. A Model Profile that requests tools, structured output or explicit reasoning still cannot use those features until their later increments. Local live inference is opt-in and may incur charges. FR-449 remains the security and observability gate before a broader orchestrated or remote preview.