src/engine/bare/llm_turn.rs

This file builds your request and runs the two turn paths. Read it when you want to know exactly what leaves the process on a turn.

Key items

build_turn_request(transient) puts extras first, then machine.full_history(), then your system prompt, then tool schemas, which are None when the registry is empty.

Since 0.4.1 the schema set is the selection. turn_tool_schemas consults your configured tool selector, sync or async, exactly once per turn ahead of the payload estimate, so the request you send and the estimate share one selection. A response_format turn still suppresses the schemas entirely.

turn_request_options() combines your request_options with the fallback routed_model() override, which wins over your model field when routing is active.

do_turn runs the cancel guard, then either do_create_message, a biased select against cancel plus turn timeout, or do_stream through the StreamHandler.

In streaming, AttemptReset wipes the accumulator, and a Fallback is adopted wholesale.

routed_model() and routed_or_client_model() tell you which model serves this turn.

record_turn_success and record_turn_failure are your fallback-breaker glue. Rate-limit escalations count as RateLimit, trips fire on_fallback, and the chain advances on failed fallbacks.

Behavior notes

Your user’s input reaches the model only via the scratchpad. turn_input exists for memory retrieval and observers, never for the request.

Overhead, meaning system plus schemas, is measured once with your configured counter. Since 0.4.0 a response_format turn no longer reserves the tool schemas its request suppresses.

Reported usage feeds your calibration ratio and anchor, as covered in context accounting.

Detection feeds only on terminal replies, so turns with tool calls never count.

Deep dive: The LLM turn.