Module engine 

Evolution engine: the glue that turns primitives into a working A/B loop.

Called from cmd_record (after each session is inserted) and from cmd_roll (manual challenger generation).

Functions

collect_scores_for_config

Load all signals tied to sessions that ran under config_id, group by session, and collapse each session’s signals into a single 0..=1 score.

evaluate_promotion

Evaluate the running experiment (if any) against the promotion threshold. Returns the experiment + decision, or None if no experiment is running.

generate_challenger

Convenience wrapper around generate_challenger_with_picker that uses the default LLM-aware picker. Callers that may not have an LLM should call generate_challenger_with_picker with picker_for_environment(false).

generate_challenger_with_picker

Generate a challenger from the current champion using one mutator, persist it as an AgentConfig row with role=Challenger, start a new Experiment with traffic_share=1.0 (v0.2.0 deploys the challenger full-time and compares against the historical champion’s session population), and apply the challenger config to disk via the adapter.

picker_for_environment

Build the mutator picker, omitting LLM-dependent mutators if no LLM is reachable. Without this, the default 50%-LLM-rewrite weight means roughly half of all challenger generations would silently fail to mutate anything when the user has no Anthropic key and no local Ollama.

promote_challenger

Promote the challenger: mark experiment as Promoted, swap project’s champion pointer, and re-apply the new champion to disk via the adapter.

resolve_active_deployment

Figure out which variant + config_id a new session should be tagged with. If an experiment is running: challenger variant on the challenger config. Otherwise: champion variant on the project’s champion config.

should_evolve

Default scheduler: trigger challenger generation when enough sessions have accumulated since the last champion change. Skips if an experiment is already running.