ADR-0130: Foundry advanced options follow deployment capabilities

Status: Accepted — 2026-09-18

Context

ADR-0129 added streaming and governed Tool calls. FR-448 adds structured output and explicit reasoning through the same AEP chat contribution. Foundry deployments can differ in supported options, and a model name alone is not evidence of support.

Decision

• The extension advertises structured output only when deployment discovery affirmatively reports jsonObject or jsonSchema; it advertises reasoning only for affirmative reasoning. An optional bounded reasoningEfforts array lists supported explicit effort values. Missing, false or unrecognized metadata grants no capability. Model names are never used to infer support.
• For an advanced request, the extension reads current deployment metadata before inference. It rejects an unavailable deployment, an unadvertised output format, or an unadvertised reasoning effort. A change in deployment metadata therefore takes effect on the next advanced call without persisting a provider model.
• AEP carries json_object or bounded json_schema response formats, including a strict schema flag, and maps enabled or disabled reasoning to the OpenAI/v1 reasoning_effort field. Unknown fields, malformed schemas, unsupported effort values and oversized Tool or schema definitions fail before inference. Text chat, streaming and the governed Tool pipeline retain their existing paths.
• The extension does not parse provider reasoning content, execute Tools, administer Foundry deployments, or guess capabilities from a model family. Offline tests use synthetic deployment metadata; the release gate for a live preview remains FR-449.

Consequences

Profiles can request advanced options only for deployments that explicitly advertise them. A deployment whose metadata omits these fields remains available for ordinary text chat. Advanced requests incur a fresh bounded discovery call so a changed deployment cannot silently retain stale capability assumptions.