Model provider request

A model provider request is a single round trip to the model provider: the harness ships the entire context window (system prompt, conversation, tool definitions, every tool result so far) and gets back one response, which is either a message for you or a tool call. Because the model is stateless, nothing is kept on the provider side between requests; the full context is re-sent every time.

A single turn usually contains many of these. You ask for a fix; the harness makes one request; the model asks to read a file; the harness reads it and makes a second request with the file contents appended; the model asks to run tests; a third request. A turn with fifteen tool calls is sixteen requests, and each one re-reads everything that came before. That's why cost per turn climbs as a session grows, and why the prefix cache matters so much.

You'll rarely see requests directly, but you feel them: each is a pause while the model reads and thinks, and each is a line on the bill. Fewer, more useful tool calls beat many small ones. A well-targeted search that returns the right file in one go is cheaper than five reads that hunt for it.