ADR-0016: Real model invocation from Agent Runner

Status

Accepted — 2026-08-02

Context

The Runtime Run vertical already delegated execution to Microsoft Agent Framework, but the Blazor Runner could still resolve simulated clients. Model-profile defaults and console overrides were not propagated to ChatOptions, and the Ollama adapter rejected every model except the host startup default. Mutable profile changes could also leave an already provisioned runtime bound to stale model resolution.

Decision

• Agent Runner uses dedicated canonical HTTP clients even when unrelated dashboard projections use simulated data.
• Runtime exposes readiness and an explicit idempotent local preparation operation for an exact agent generation.
• Preparation creates or reuses the exact immutable revision and a generation-specific local deployment, then reconciles it synchronously.
• AgentFrameworkRuntime resolves the current model profile for every execution and constructs the reconstructible ChatClientAgent at the execution boundary.
• Profile temperature and maximum-output-token defaults flow through model metadata into MAF ChatOptions.
• Runtime overrides are limited to temperature and maxOutputTokens, validated at both UI and Runtime boundaries, and take precedence over profile defaults.
• The Ollama adapter pins each request to the deployment model through ChatOptions.ModelId; it no longer requires that model to equal the host default.
• Runs record the actual provider, model, and effective supported options. Prompts and responses remain excluded from logs.

Consequences

The console Run button now exercises the durable Runtime API and the actual configured provider. Profile changes take effect on the next Run without persisting provider-specific agents. Dashboard simulation remains available, provider-native token streaming remains a later increment, and remote evaluation still requires an explicitly configured provider.