core.llm

llm_step

llm_step(...) creates a Step that:

1. builds a prompt from state (prompt_builder)
2. coerces it into a PromptPayload
3. calls provider.complete(...)
4. emits PromptEvent and LLMEvent (when invocation context exists)
5. attaches the result to state (default key: llm_response)

Minimal usage:

from coevolved.core import LLMConfig, llm_step

step = llm_step(
    prompt_builder=lambda s: "Say hi",
    provider=provider,
    config=LLMConfig(model="gpt-4o-mini"),
)

Prompt builders

prompt_builder can return:

• str
• Prompt
• PromptPayload
• dict that validates as PromptPayload

This makes it easy to start simple and evolve to structured prompts and message-based tool calling.

Parsing responses

If you pass parser, Coevolved applies it to the LLMResponse before returning/attaching.

Two common patterns:

• Keep the raw LLMResponse in state for agent loops
• Parse into a structured domain object for downstream deterministic steps

Next steps

LLMs and providers

How providers, configs, requests, and responses fit together.

core.events reference

Prompt and LLM events emitted during execution.