ADR-0013: Model-provider boundary and local Ollama adapter

Status

Accepted — 2026-08-01

Context

Agentstration needs a first executable local model-provider integration while preserving the distinction between technical agent execution and model connectivity. Ollama must remain optional, must not leak into Management or Runtime abstractions, and must not replace the offline deterministic default.

Decision

• Agentstration.ModelProviders contains the provider-neutral IModelProvider and resolver contracts.
• Agentstration.ModelProviders.Ollama owns OllamaSharp registration and exposes the selected local model as the standard Microsoft.Extensions.AI.IChatClient.
• Agentstration.Web is the composition root. It activates the Ollama adapter only when AI:Provider is Ollama.
• Agentstration.Runtime.AgentFramework continues to consume only IChatClient; it does not reference Ollama or Aspire.
• Agentstration.AppHost alone provisions the Ollama container, persistent model volume, and development model through Aspire.
• Direct Web startup remains deterministic by default and can target an independently running Ollama instance through configuration.
• A development-only HTTP diagnostic validates connectivity. Normal agent validation uses the existing Agent Runner and Runtime Run path.

Consequences

The model provider and runtime adapter can evolve independently, tests remain offline, and local Docker orchestration is optional. The first implementation intentionally supports one configured Ollama model per host; multi-model catalogs and Management model-profile resources remain separate future work.