#Model providers

OpenAI and DeepSeek share a Responses-based conversation-and-tool runner. Provider adapters handle endpoints, authentication, and request differences.

ResponsesRunner assembles history, instructions, and tools. ResponsesClient handles HTTP and streamed response validation. Their boundary uses Responses data contracts.

#Configuration and selection

Provider definitions describe endpoints, credentials, and available model IDs. They are persisted on the server; credentials stay out of browser responses. Each thread selects a provider snapshot and model. Updating the catalog takes effect in an existing thread when that definition is selected again.

#Change at a step boundary

Validated selection → Pending client → Next step → Active client

The current request and its tools finish with captured settings. The next step uses the latest accepted selection. Thread history and identity stay in place; other conversations retain their own selections.

When configurations differ, replay preserves visible text and tool calls/results while omitting foreign provider-private items. Saving provider definitions does not persist conversations.

Implementation: responses.rs.