marola.llm

Loads a DSPy-compiled prompt artifact (dspy/compile_recommendation_prompt.py produces two: recommendation_prompt.json for the summarizer, review_prompt.json for Reviewer — see ARCHITECTURE.md §5a) and turns it into a plain chat message list any LlmClient can replay.

One backend for "turn a BestHour into a natural-language summary" (see Recommender's query synthesis step, ARCHITECTURE.md §5).

The reviewer/critic pass (FUTURE-WORK.md §4.2, now built rather than just proposed): replays a second DSPy-compiled artifact (review_prompt.json, compiled from dspy/compile_recommendation_prompt.py's ReviewSwimSummary signature) against the summarizer's own draft output, before either reaches a user.

MIP-0010 task 6: the decorator that gives every LlmClient.complete one span — llm. with the GenAI semantic-convention attributes MLflow ingests (gen_ai.operation.name, gen_ai.request.model) plus marola's own sizes (message count, prompt/completion characters).