Build a Provider Engine

Provider engines still own final request construction. For benchmark-native runs, Themis renders prompt templates before infer(...) is called.

Use the Prepared Prompt

class MyEngine:
    def infer(self, trial, context, runtime):
        messages = [message.model_dump(mode="json") for message in trial.prompt.messages]
        ...

What Themis Preserves

• trial.prompt.messages contains the rendered messages sent to the model
• trial.prompt.id, trial.prompt.family, and trial.prompt.variables stay available for routing and logging
• trial.task.dimensions, trial.task.slice_id, and trial.task.benchmark_id stay available for request metadata and reporting