Send every model call that can change a trip to one LLM provider, with a server-chosen model per task
Amends ADR-0096: the in-app agent runs on OpenAI alone, so the effect schema has one provider to satisfy.
Decision
Make every model call whose output can change a trip through OpenAI's Responses API, and keep no second provider behind it. Neighbourhood guides fall outside this: Exa's answer endpoint writes them from the trip's destinations and Wikivoyage grounding, and their text is display-only (ADR-0038).
Pick the model tier per task on the server. Turning a typed request into effects uses the smallest model, extracting bookings from an uploaded document the mid tier, and the in-app agent, which runs several rounds of tool calls and web search per turn, the full model. Reasoning effort is set per task in the same place. No request from a client names a model, a reasoning effort or a verbosity.
Why
ADR-0096 kept the effect schema inside the subset both OpenAI strict mode and Anthropic structured outputs accept, and noted that Claude's cap on optional and union-typed parameters would force effects to be split across tools. The agent that shipped uses the Responses API's built-in web search and runs statelessly by resending the model's encrypted reasoning, both specific to that API. A second provider would mean another client, more recorded cassettes and a schema held to the narrower of two dialects.
The tasks differ in cost and difficulty. Intent parsing runs on short text often and cheaply. Extraction reads a whole PDF or image and needs more reasoning. The agent plans across tools. Model choice sets the cost of a request, so it belongs with the server's spend limits.
Rejected alternatives
- Keep Anthropic as a second provider. It doubles the client, the cassettes and the schema constraints for a fallback no outage has called for.
- One model for every task. Either intent parsing pays for the large model, or the agent runs on one too small to plan.
- Let the client choose the model. Any editor could then spend the most expensive configuration on every request.
Consequences
An OpenAI outage or a breaking API change stops every model feature except neighbourhood guides at once. Retiring a model family means re-recording the model tests for each task. Model spans reach Langfuse through the node collector of ADR-0053, so prompts and completions, including trip text, go to Langfuse Cloud as well as to OpenAI.