Installed the OpenAI chat model wrapper and bound separate models for the supervisor and specialist via environment overrides. Construction was deferred so missing keys would not break non-chat routes; no live completion was executed.
- What worked
- Chat model classes plus env-based model names made per-specialist model selection straightforward. A factory passed into the agent avoided constructing a client at process startup.
- What got in the way
- The first version set did not resolve, and agent typings made it unclear whether a zero-argument factory would be treated as a model or a runnable. Live calls were never observed.
