LLMs and providers

Overview

Coevolved models LLM interaction as a Step:

1. build a prompt from state
2. call a provider (LLMProvider)
3. attach a structured response (LLMResponse) back onto state

This keeps LLM calls observable and composable like everything else.

LLMConfig, requests, and responses

Core types:

• LLMConfig: model + generation parameters + available tools
• LLMRequest: prompt + config
• LLMResponse: text + tool calls + usage metadata

llm_step(...) handles the request lifecycle and emits prompt/response events.

Providers

A provider implements the LLMProvider protocol:

• complete(request) -> LLMResponse

Coevolved includes an OpenAI provider that maps PromptPayload and ToolSpec into OpenAI chat completion requests.

Streaming (optional)

If your provider supports streaming, implement StreamingLLMProvider:

• stream(request) -> Iterator[LLMStreamChunk]

Streaming is useful for UIs (token-by-token rendering) and long outputs, but it adds complexity around tool call deltas and partial usage reporting.

Next steps

OpenAI provider

Configure and use the built-in OpenAI provider.

core.llm reference

llm_step parameters, parsing, and event emission.